HARNESSING MACHINE LEARNING TECHNIQUES FOR LARGE-SCALE MAPPING OF INLAND AQUACULTURE WATERBODIES IN BANGLADESH By Hannah Ferriby A THESIS Submitted to Michigan State University in partial fulfillment of the requirements for the degree of Biosystems Engineering – Master of Science 2021 ABSTRACT HARNESSING MACHINE LEARNING TECHNIQUES FOR LARGE-SCALE MAPPING OF INLAND AQUACULTURE WATERBODIES IN BANGLADESH By Hannah Ferriby Aquaculture in Bangladesh has grown dramatically in an unplanned manner in the past few decades, becoming a major contributor to the rural economy in many parts of the country. National systems for the collection of statistics have been unable to keep pace with these rapid changes, and more accurate, up to date information is needed to inform policymakers. Using Sentinel-2 Top of Atmosphere Reflectance images within Google Earth Engine and ArcGIS platforms, we proposed six strategies for improving fishpond detection as the existing techniques seem unreliable. The study area is comprised of seven districts in south-west and south-central Bangladesh. The tested strategies include: 1) identification of the best period for image collection, 2) testing the buffer size for threshold optimization, 3) determining the best combination of image reducer and water- identifying indices, 4) introduction of a convolution filter to enhance edge-detection, 5) evaluating the impact of ground-truthing data on machine learning algorithm training, and 6) identifying the best machine learning classifier. Each enhancement builds on the previous one to develop a comprehensive improvement strategy called the Enhanced Method for fishpond detection. We compared the results of each improvement strategy to the known ground-truthing fishponds as the metric of success. We compared the precision, recall, and F1 score for machine learning classifiers to determine the quality of results. Among the studied methods, the Classification and Regression Trees performed the best. Overall, the proposed strategies enhanced fishpond area detection in all districts within the study area. Copyright by HANNAH FERRIBY 2021 ACKNOWLEDGMENTS I am extremely grateful to Dr. Pouyan Nejadhashemi who guided me through both my research and my program overall. Dr. Nejadhashemi not only led this research and advised me academically, but he also has been a great mentor. I thank him for his encouragement and patience with me throughout this project. In addition, I would like to thank my committee, Dr. Nathan Moore and Dr. Narendra Das for their continual support this past year. Many thanks to both Dr. Ben Belton and Dr. Mohammad Mahfujul Haque for their great expertise and crucial work for this project. I would also like to thank all of my lab mates for always supporting me and challenging me to be better; Anna Raschke, Dr. J. Sebastian Hernandez-Suarez, Josué Kpodo, Ian Kropp, Enid Banda, Mervis Chikafa, Kieron Moller, Shashank Mohan, Nicolas Fernandez, and Dr. Eeswaran Rasu. You all made my time at MSU really enjoyable and exciting – even over Zoom! I cannot begin to express how thankful I am for the support of my family, especially my mother. You three always keep me sane. I would also like to thank Matt for always encouraging me to do my best. You have really been my rock the last year and a half. This thesis is made possible by the support of the United States Agency for International Development (USAID) Grant No. 193900.312455.12B. iv TABLE OF CONTENTS LIST OF TABLES ........................................................................................................................ vii LIST OF FIGURES ....................................................................................................................... xi KEY TO ABBREVIATIONS ...................................................................................................... xiii 1 INTRODUCTION ....................................................................................................................... 1 2 LITERATURE REVIEW ............................................................................................................ 3 2.1 Global Aquaculture ............................................................................................................... 3 2.1.1 Capture Fisheries versus Aquaculture ............................................................................ 3 2.1.2 Aquaculture by Water Type ........................................................................................... 4 2.1.3 Global Aquaculture Market ............................................................................................ 5 2.1.4 Aquaculture and the Environment .................................................................................. 6 2.2 Aquaculture in Bangladesh ................................................................................................... 7 2.2.1 Aquaculture Output in Bangladesh................................................................................. 7 2.2.2 Aquaculture Types and Technologies ............................................................................ 8 2.3 Introduction to Remote Sensing .......................................................................................... 10 2.4 Space-borne Sensors ........................................................................................................... 14 2.4.1 Visible Light Remote Sensing ...................................................................................... 14 2.4.2 Infrared Remote Sensing .............................................................................................. 15 2.4.3 Microwave Remote Sensing ......................................................................................... 15 2.5 Remote Sensing for Aquaculture ........................................................................................ 16 2.5.1 Synthetic Aperture Radar ............................................................................................. 17 2.5.2 Multispectral and Panchromatic Imagery ..................................................................... 20 2.6 Remote Sensing Water Indices for Small Waterbody Location ......................................... 20 2.7 Remote Sensing Aquaculture Applications......................................................................... 25 2.7.1 SAR Applications ......................................................................................................... 25 2.7.2 Multispectral and Panchromatic Imaging Applications ............................................... 28 2.7.3 Table of Relevant Studies on SAR and Optical Remote Sensing ................................ 29 2.8 Basic Machine Learning and its Applications in Detecting Aquaculture Farms ................ 34 2.9 Limitations of Remote Sensing Techniques in Detection of Small Waterbodies ............... 37 2.10 Goals.................................................................................................................................. 38 3 HARNESSING MACHINE LEARNING TECHNIQUES FOR LARGE-SCALE MAPPING OF INLAND AQUACULTURE WATERBODIES IN BANGLADESH .................................. 39 3.1 Introduction ......................................................................................................................... 39 3.2 Materials and Methods ........................................................................................................ 43 3.2.1 Study Area .................................................................................................................... 43 3.2.2 Overview of the Base Method ...................................................................................... 45 3.2.3 Proposed Improvements ............................................................................................... 49 3.2.4 Evaluation Criteria for Comparing Fishpond Detection Algorithms ........................... 56 3.3 Results and Discussion ........................................................................................................ 58 v 3.3.1 Base Method Performance Evaluation within the Study Area in Detecting Waterbodies and Fishponds ........................................................................................................................ 58 3.3.2 Identifying the Best Period of Image Collection for Detecting Fishponds (Improvement 1) .................................................................................................................... 61 3.3.3 Testing the Buffer Size for Threshold Optimization (Improvement 2) ........................ 65 3.3.4 Determining the Combination of Image Reducer and Water-Identifying Index to Improve Waterbody and Fishpond Detection (Improvement 3) ........................................... 67 3.3.5 Implementing Edge Detection with a Convolution Filter to Improve Fishpond Boundary Detection (Improvement 4) ................................................................................... 68 3.3.6 Evaluating the Impact of Ground-Truthing Data on Machine Learning Training (Improvement 5) .................................................................................................................... 71 3.3.7 Adding Random Forest and Support Vector Machine Classifiers to Determine the Best Classifier for the Data (Improvement 6)................................................................................ 72 3.3.8 The Overall Enhancement in Fishpond Detection as Results of the Proposed Improvements based on Medium-Resolution and High-Resolution Imagery ....................... 75 3.3.9 Land Use and Ground Truth Fishpond Characteristics to Explain Trends in Results .. 78 3.4 Conclusion........................................................................................................................... 79 4 OVERALL CONCLUSION ...................................................................................................... 82 5 FUTURE RESEARCH RECOMMENDATIONS .................................................................... 84 APPENDIX ................................................................................................................................... 85 BIBLIOGRAPHY ....................................................................................................................... 135 vi LIST OF TABLES Table 1. Global aquaculture production in million tons of all aquatic organisms by environment type from 2010 to 2018 (FAO, 2020a). .......................................................................................... 5 Table 2. Aquaculture production in 2015 by region (adapted from Subasinghe 2017) .................. 6 Table 3. Descriptions of the five different types of waterbodies used in aquaculture in Bangladesh, adapted from (Jahan et al., 2015). ............................................................................ 10 Table 4. List of active space-borne platforms and their sensor types. .......................................... 12 Table 5. List of different passive and active sensors (NASA EarthData, 2020). ......................... 13 Table 6. Examples of active satellites and their respective spatial resolution and sensor(s) (Ottinger et al., 2016). ................................................................................................................... 14 Table 7. Frequently used microwave wavelength bands, frequency bands, and their common names (Gade & Stoffelen, 2019). ................................................................................................. 16 Table 8. List of relevant studies for fishpond detection using radar and optical satellite imagery. ....................................................................................................................................................... 30 Table 9. Common supervised algorithms and their applications (MathWorks, 2016a). .............. 35 Table 10. Common unsupervised algorithms and their applications (MathWorks, 2016b). ........ 36 Table 11. Base Method assumptions, shortcomings, and alternative approaches ........................ 50 Table 12. Average size, shape, and period with water for gher with and without rice, commercial ponds, and homestead ponds......................................................................................................... 55 Table 13. The percentage of ground-truthing fishponds area and total ground-truthing fishponds correctly identified in all districts using different lengths for image processing (i.e., the entire year, one month, one day). ............................................................................................................ 62 Table 14. Comparing statistics from Yu et al. (2020) machine learning training with historical imagery to our ground-truthing data training................................................................................ 72 Table 15. Comparing the training performance of the four classifiers on the validation ground- truthing data. ................................................................................................................................. 74 vii Table 16. The lower and upper limits for the mean relative difference of the Based and Enhanced Methods performances before and after a machine learning classifier......................................... 76 Table A1. Land cover as percentage of area for each district within the study region (China Ministry of Natural Resources, 2020) ........................................................................................... 88 Table A2. Numbers of ground-truthing data fishponds and non-fishponds by district. ............... 89 Table A3. Base method comparison for all seven districts. .......................................................... 90 Table A4. Time period comparison for Bagerhat. ........................................................................ 91 Table A5. Time period comparison for Barisal. ........................................................................... 92 Table A6. Time period comparison for Bhola. ............................................................................. 93 Table A7. Time period comparison for Gopaglanj. ...................................................................... 94 Table A8. Time period comparison for Jessore. ........................................................................... 95 Table A9. Time period comparison for Khulna. ........................................................................... 96 Table A10. Time period comparison for Satkhira. ....................................................................... 97 Table A11. Buffer size comparison for Bagerhat for both month and single day. ....................... 98 Table A12. Buffer size comparison for Barisal for both month and single day. .......................... 99 Table A13. Buffer size comparison for Bhola for both month and single day. .......................... 100 Table A14. Buffer size comparison for Gopalganj for both month and single day. ................... 101 Table A15. Buffer size comparison for Jessore for both month and single day. ........................ 102 Table A16. Buffer size comparison for Khulna for both month and single day. ........................ 103 Table A17. Buffer size comparison for Satkhira for both month and single day. ...................... 104 Table A18. Mode reducer applied for one month period in Bagerhat comparing water-identifying index combinations. .................................................................................................................... 105 Table A19. Mode reducer applied for one month period in Barisal comparing water-identifying index combinations. .................................................................................................................... 106 viii Table A20. Mode reducer applied for one month period in Bhola comparing water-identifying index combinations. .................................................................................................................... 107 Table A21. Mode reducer applied for one month period in Gopalganj comparing water- identifying index combinations................................................................................................... 108 Table A22. Mode reducer applied for one month period in Jessore comparing water-identifying index combinations. .................................................................................................................... 109 Table A23. Mode reducer applied for one month period in Khulna comparing water-identifying index combinations. .................................................................................................................... 110 Table A24. Mode reducer applied for one month period in Satkhira comparing water-identifying index combinations. .................................................................................................................... 111 Table A25. allNonZero reducer applied for one month period in Bagerhat comparing water- identifying index combinations................................................................................................... 112 Table A26. allNonZero reducer applied for one month period in Barisal comparing water- identifying index combinations................................................................................................... 113 Table A27. allNonZero reducer applied for one month period in Bhola comparing water- identifying index combinations................................................................................................... 114 Table A28. allNonZero reducer applied for one month period in Gopalganj comparing water- identifying index combinations................................................................................................... 115 Table A29. allNonZero reducer applied for one month period in Jessore comparing water- identifying index combinations................................................................................................... 116 Table A30. allNonZero reducer applied for one month period in Khulna comparing water- identifying index combinations................................................................................................... 117 Table A31. allNonZero reducer applied for one month period in Satkhira comparing water- identifying index combinations................................................................................................... 118 Table A32. Water-identifying index combination comparison for Bagerhat using single day time period. ......................................................................................................................................... 119 Table A33. Water-identifying index combination comparison for Barisal using single day time period. ......................................................................................................................................... 120 Table A34. Water-identifying index combination comparison for Bhola using single day time period. ......................................................................................................................................... 121 ix Table A35. Water-identifying index combination comparison for Gopalganj using single day time period. ................................................................................................................................. 122 Table A36. Water-identifying index combination comparison for Jessore using single day time period. ......................................................................................................................................... 123 Table A37. Water-identifying index combination comparison for Khulna using single day time period. ......................................................................................................................................... 124 Table A38. Water-identifying index combination comparison for Satkhira using single day time period. ......................................................................................................................................... 125 Table A39. Best results of the index tests for each of the seven districts. .................................. 126 Table A40. Results from adding the Laplacian 5×5 convolution filter for each district. ........... 127 Table A41. Comparison of LR, CART, RF, and SVM results when applied to ground-truthing fishpond data for Bagerhat, Barisal, Bhola, and Gopalganj during one day with that district’s best performing water-identifying index. ........................................................................................... 128 Table A42. Comparison of LR, CART, RF, and SVM results when applied to ground-truthing fishpond data for Jessore, Khulna, and Satkhira during one day with that district’s best performing water-identifying index. ........................................................................................... 129 Table A43. Land use as percent of total area for 10-meter buffer around ground truth fishpond locations for each district. ........................................................................................................... 131 Table A44. Land use as percent of total area for 50-meter buffer around ground truth fishpond locations for each district. ........................................................................................................... 132 Table A45. Land use as percent of total area for 100-meter buffer around ground truth fishpond locations for each district. ........................................................................................................... 133 Table A46. Median ground truth fishpond size for each district. ............................................... 134 x LIST OF FIGURES Figure 1. Global production in million tons of aquaculture and capture fisheries from 2000 to 2018 (FAO, 2020c). ........................................................................................................................ 4 Figure 2. Aquaculture production in Bangladesh by environment type from 1980 to 2018 (FAO, 2020b) ............................................................................................................................................. 8 Figure 3. Illustration of the SAR Doppler technique (adapted from Gade & Stoffelen, 2019)). . 18 Figure 4. SAR backscattered Doppler shift (adapted from Gade & Stoffelen, 2019)). ................ 19 Figure 5. Study area is broken into seven individual districts. ..................................................... 44 Figure 6. Flowchart outlining the major steps of the Yu et al. (2020) for fishpond identification. ....................................................................................................................................................... 46 Figure 7. Percentage of the ground-truthing area that were correctly identified as waterbodies, classified as fishpond using logistic regression method, and classified using the classification and regression trees method for different districts within the study area. ........................................... 59 Figure 8. Example of the Base Method results in Bagerhat. Orange polygons represent ground- truthing fishponds. Blue polygons are classified as fishponds. Red polygons are classified as non-fishponds. (A) Ground-truthing fishponds. (B) Logistic Regression classification results. (C) Classification and Regression Trees classification results. (Centroid of the pictures above: 22° 35' 59.4096" N, 89° 34' 57.414" E)............................................................................................... 60 Figure 9. Visual comparison of the three-time periods in southern Jessore and southern Bhola. (A) Jessore ground-truthing fishponds (orange). (B) Jessore ground-truthing fishponds and full Figure 9 (cont’d) year water (red). (C) Jessore ground-truthing fishponds and one-month water (purple). (D) Jessore ground-truthing fishponds and single-day water (blue). (Centroid: 22° 57' 49.9248" N, 89° 16' 46.3188" E). (E) Bhola ground-truthing fishponds (orange). (F) Bhola ground-truthing Fishponds and full year water (red). (G) Bhola ground-truthing Fishponds and one-month water (purple). (G) Bhola ground-truthing Fishponds and single-day water (blue). (Centroid: 22° 10' 7.0464" N, 90° 41' 27.546" E)......................................................................... 64 Figure 10. NDWI reflectance values for MM in Khulna for a single-day image. Top left to bottom right: 0-pixel buffer, 1-pixel buffer, 3-pixel buffer, 5-pixel buffer .................................. 66 Figure 11. Ground-truthing (GT) fishpond area identified pre-classifier per district for the combined index, only AWEI, only MNDWI, and only NDWI results using the single day images ....................................................................................................................................................... 68 xi Figure 12. NDWI image in western Barisal (A) before and (B) after the Laplacian 5×5 convolution filter was applied (Centroid: 22° 51' 38.37" N, 90° 5' 53.28" E) .............................. 70 Figure 13. Ground-truthing fishpond area percentage identified by each classifier type. ............ 73 Figure 14. Comparison example of classified fishponds in eastern Gopalganj using (A) Logistic Regression, (B) Classification and Regression Trees, (C) Random Forest, and (D) Support Vector Machine. Ground truthing fishponds are in orange and classified fishponds are in blue. (Centroid: 23° 0' 17.9136'' N, 90° 3' 46.8936'' E). ........................................................................ 75 Figure 15. Impacts of high-resolution imagery in Jamalnagar, Satkhira. (A) High-resolution boundary identification, (B) Base Method water identification, (C) Enhanced Method water Figure 15 (cont’d) identification, and (D) Intersect of Enhanced Method and high-resolution boundaries. (Centroid: 22° 34' 57.6624'' N, 89° 12' 12.1536'' E) ................................................. 77 Figure A1. Land use map of the study region (adapted from region (China Ministry of Natural Resources, 2020)) ......................................................................................................................... 86 Figure A2. Distribution of Ground Truthing Data locations by type. .......................................... 87 xii KEY TO ABBREVIATIONS AOI Areas of Interest AWEI Automated Water Extraction Index CART Classification and Regression Trees CRESDA China Center for Resources Satellite Data and Application EM Electromagnetic energy FAO Food and Agriculture Organization of the United Nations FtF Feed the Future GDP Gross domestic product GEE Google Earth Engine GIS Geographic Information Systems GPS Global Position Systems GRDH Ground range detected high-resolution GT Ground Truth(ing) H Horizontal polarization IP Image processing IW Interferometric wide-swath JRC Joint Research Center k-NN k-Nearest Neighbor LIFDCs Low-income food-deficit countries LR Logistic Regression LST Land surface temperature LWIR Long-wave infrared xiii MM MNDWI Mask MNDWI Modified Normalized Difference Water Index MWIR Mid-wave infrared NDWI Normalized Difference Water Index NIR Near infrared OBF Object-based Feature OLI Operational Land Imager RF Random Forest SAR Synthetic aperture radar SNAP Sentinel Application Platform SVM Support Vector Machine SWIR Short-wave infrared TOA Top of Atmosphere TOPSAR Terrain observation with progressive scans SAR USAID United States Agency for International Development USGS United States Geological Survey UV Ultraviolet V Vertical polarization WPF Water Presence Frequency WRI Water Ratio Index WSA Water Surface Area xiv 1 INTRODUCTION There is an increasing importance placed on inland aquaculture and fisheries due to a growing demand for fish and a stagnating production in capture fishing (Ottinger et al., 2016). Fish is a main source of nutrients for many people around the world and especially in Bangladesh (Heck et al., 2010). The increase in demand for fish is expanding aquaculture, putting pressure on croplands necessary to produce rice, Bangladesh's primary food source, creating a tension for space that needs to be addressed (Yu et al., 2018; Hashem et al., 2014). Meanwhile, there are little to no studies on the current production and potential of aquaculture in Bangladesh to inform policymakers and researchers (Shamsuzzaman et al., 2017). Bangladesh was the fifth-highest aquaculture-producing country in the world (Subasinghe, 2017). However, aquaculture in Bangladesh has evolved dramatically in the last thirty years (Hernandez et al., 2018). A few highlights of this period include: 1) the amount of fish produced skyrocketed with 94% of the production for Bangladeshi consumption, 2) the farmed fish market expanded from 124,000 tons to 2 million tons, and 3) the number of people involved in the aquaculture value chain tripled, and 4) freshwater aquaculture production increased by 167% from 2001 to 2017, surpassing capture fishery (Hernandez et al., 2018; Sattar, 2019). Therefore, there is a need to better understand this evolving industry. There are two primary methods for determining the extent of fishponds. The first method utilizes ground-truthing surveys. The surveys would provide accurate information on fishponds throughout the region of interest, but would be costly and time-consuming (Rhodes et al., 2015). The alternative to ground-truthing surveys is remote sensing to identify the specific locations in which aquaculture occurs. Remote sensing has been used to help with aquaculture management, 1 environmental monitoring, and aquaculture extent (Anand et al., 2020, Ottinger et al., 2016). Remote sensing can analyze large areas in much less time than it would take to conduct a survey. However, due to the complexity and variety in types of aquacultural production in Bangladesh, identifying a reliable technique that can be applied for large-scale studies can be challenging as previous studies only focused on small regions (Hashem et al., 2014, Yu et al., 2020, Huda et al., 2010). This study aims to identify bodies of water throughout the region of interest in Bangladesh and to determine if they are being used for aquaculture using satellite imagery. This process can lead to creating an aquaculture identification method that does not rely on ground-truthing surveys. The specific objectives of this study area to:  Identify the best methods for locating waterbodies throughout the study region.  Examine the existing methods for identifying fishponds and propose new strategies to improve the detection techniques.  Evaluate the extent of aquaculture productions throughout the region of interest. 2 2 LITERATURE REVIEW 2.1 Global Aquaculture 2.1.1 Capture Fisheries versus Aquaculture Aquaculture is comprised of breeding, rearing, and harvesting of fish and aquatic products within freshwater, sea water, brackish water, or inland saline water (FAO, 2020c). Aquaculture can be divided into three main categories: marine, brackish, and freshwater (US Department of Commerce, n.d.). Aquaculture systems can also be divided into different farming styles: water- based systems (e.g., cages and pens, onshore/offshore), land-based systems (e.g., rainfed ponds, irrigated or flow-through systems, tanks, and raceways), recycling systems (e.g., high control enclosed systems), and integrated farming systems (e.g., livestock-fish, agriculture, and fish dual- use aquaculture and irrigation ponds) (Funge-Smith, Simon; Phillips, 2001). Capture fisheries encompass all the aquatic animals killed, caught, trapped or collected in freshwater, brackish water, and/or marine water (United Nations et al., 2003). Due to the accessibility of aquaculture, especially inland, its continual rise in production makes up for the decline in the availability of capture fisheries (Ottinger et al., 2018). Aquaculture contributed 42% of the global fish supply in 2012 and will eclipse capture fishery production by 2030 (Lam, 2016). Despite the global rise in production from aquaculture, a large portion of the fish produced and traded in low-income regions still comes from capture fisheries (Thilsted et al., 2016). Together, these two sectors support the livelihoods of 10-12% of the world population (Lam, 2016). Global food security, income distribution, and ecological sustainability all rely on the balance between capture fisheries and aquaculture (Lam, 2016). Figure 1 shows the trends in global aquaculture, both marine and inland, and capture fisheries, both marine and inland. 3 120 Global Production (million tons) 100 80 60 40 20 0 2000 2003 2006 2009 2012 2015 2018 Year Aquaculture Capture Fisheries Figure 1. Global production in million tons of aquaculture and capture fisheries from 2000 to 2018 (FAO, 2020c). 2.1.2 Aquaculture by Water Type There are three main types of environments used in aquaculture: freshwater (fishponds, fish pens, fish cages), brackish water (fishponds in coastal areas), and marine (fish cages or substrates) (FAO, 1989). Most of the world’s aquaculture occurs in freshwater systems (Ross et al., 2013). Inland aquaculture produces nearly double what marine aquaculture does. In 2016, inland aquaculture produced 51.4 million tons of product, while coastal produced 28.7 million tons, excluding aquatic mammals, crocodiles, alligators, and aquatic plants (FAO, 2018). The gap in fish production between inland and coastal (marine) aquaculture has been growing significantly each year (FAO, 2018). It is important to note that inland fisheries are more accessible to even the poorest of people (FAO, 2018). Taking into account all aquatic organisms, freshwater, and marine habitats contribute nearly the same to global aquaculture production (FAO, 2014). In 2014, freshwater aquaculture 4 produced 46.3 million tons, while marine aquaculture produced 47.4 million (FAO, 2014). Table 1 looks at the increasing trends in all aquaculture from 2010 to 2014. Table 1. Global aquaculture production in million tons of all aquatic organisms by environment type from 2010 to 2018 (FAO, 2020a). Year Environment 2010 2012 2014 2016 2018 Freshwater 35.4 38.9 43.6 47.1 50.4 Brackishwater 5.4 6.4 7.6 8.7 8.6 Marine 37.1 42.8 48.4 52.4 55.6 2.1.3 Global Aquaculture Market Global demand for fish, crustaceans, and mollusks is rising yearly (Ottinger et al., 2018). Global aquaculture in 2015 produced 106 million tons, worth approximately US$163 billion (Subasinghe, 2017). Aquaculture’s role in global fish production increased from 13% to 45% between 1990 and 2015 (Ottinger et al., 2018). According to the Food and Agriculture Organization of the United Nations (FAO), fish and fish production products, such as aquatic plants, sponges, fats and oils, are the most traded food goods in the world (Subasinghe, 2017). At least 75% of the global fish and fishery products are entering international markets (Subasinghe, 2017). Aquaculture products contribute a critical role in the global food system, providing approximately 3 billion people with a minimum of 15% of their animal protein intake (Charles et al., 2010). 19.3 million people were involved in aquaculture during 2016, making it a substantial line of employment (FAO, 2018). During the period between 1961 and 2013, the growth in the global supply of fish is larger than the rate of population growth for the same period (Ahmed et al., 2019). Fish consumption 5 increased per capita from 9.9 kg in the 1960s to 18.9 kg in 2010 (Ottinger et al., 2016). Despite this massive average growth, the growth in low-income food-deficit countries (LIFDCs) was much smaller, from 3.5 to 7.6 kg during the same period (Subasinghe, 2017). LIFDCs make up 15 of the 21 countries with the highest per capita inland fish production (FAO, 2018). With a rapidly growing world population and an annual loss of 5-7 million ha of farmland, there is a pressure on aquaculture to fill the gap in food production (Ahmed et al., 2019). Capture fisheries around the world see stagnating production, yet with growing population and demand, obtaining aquatic food is a major concern for global food security (Ottinger et al., 2016). The current largest region of aquaculture production is Asia, followed distantly by the Americas, Europe, and Africa (Subasinghe, 2017). Table 2 breaks up global aquaculture production data from 2015 to the different regions of the world. Table 2. Aquaculture production in 2015 by region (adapted from Subasinghe 2017) Production Percentage of Region (tons) World Production Africa 1,772,391 2.3% America 3,273,376 4.3% Asia 68,432,034 89.4% Europe 2,875,159 3.8% Oceania 188,066 0.2% 2.1.4 Aquaculture and the Environment While the benefits to individual livelihoods and potentially global food security are great, aquaculture is a cause of environmental degradation and biodiversity loss (Ottinger et al., 2017). The environmental harm from aquaculture is due to the use of pesticides and other chemicals, as well as the discharging of untreated wastewater (Ottinger et al., 2017). Sustainability certification programs were created to address the negative environmental impacts and potentially increase 6 efficiency, but the programs were aimed at higher-value sectors (Charles et al., 2010). If aquaculture’s production keeps growing at its current pace, the environmental effects could be detrimental (Ahmed et al., 2019). 2.2 Aquaculture in Bangladesh 2.2.1 Aquaculture Output in Bangladesh Aquaculture in Bangladesh has shifted dramatically in the last three decades (Hernandez et al., 2018). A few major changes occurred during that time period: the amount of fish produced skyrocketed with 94% of the production for domestic consumption, the farmed fish market grew from 124,000 tons to almost 2 million tons, and the number of people involved in the aquaculture value chain tripled (Hernandez et al., 2018). Bangladesh was the fifth-highest aquaculture producing country in the world in 2015 with 2.1 million tons and a growth rate of 5.3% from the previous year (Subasinghe, 2017). Inland waters in Bangladesh alone produced over 1 million tons in 2015 (FAO, 2018). Aquaculture produced 55% of the country’s fish production in 2014 (Jahan et al., 2015). Freshwater aquaculture production increased by 167% from 2001 to 2017; surpassing capture fishery as the main source of fish in Bangladesh (Sattar, 2019). Figure 2 shows the change in aquaculture production in freshwater and brackishwater from 1980 to 2018. 7 2.5 2 Aquaculture Production (million tons) 1.5 1 0.5 0 1980 1985 1990 1995 2000 2005 2010 2015 Year Brackishwater Freshwater Figure 2. Aquaculture production in Bangladesh by environment type from 1980 to 2018 (FAO, 2020b) Fish makes up two-thirds of the animal protein consumed and one-quarter of the agricultural gross domestic product (GDP) of Bangladesh (Sattar, 2019). Approximately 12 million people are connected with fisheries in Bangladesh, of which 1.4 million people are purely involved in fishery work (Shah, 2003). 2.2.2 Aquaculture Types and Technologies There are two types of aquaculture in Bangladesh: freshwater and coastal, with inland pond culture being the mainstay (Islam et al., 2019). Bangladesh has 4 million ha of inland open waterbodies (rivers, lakes, ponds) and 0.7 million ha of closed waterbodies (man-made ponds) (Sattar, 2019). Pond farming in Bangladesh produces carp, cichlids, and catfish, which contributes 80% of the total recorded aquaculture in the country (Islam et al., 2019). Coastal aquaculture in Bangladesh is comprised of mostly shrimp and prawn farming in ghers (Islam et al., 2019). 8 Integrating fishponds with crops have many advantages including increased diversification, intensification, improved natural resource efficiency, increased productivity, and increased sustainability (Karim & Little, 2018). That is why, pond aquaculture has experienced the fastest growth in Bangladesh (Rashid et al., 2019). There are approximately 1.3 million fishponds in Bangladesh, covering 0.151 million ha (FAO, 2020b). Of the 0.151 million ha, 55% is cultured, 29% is culturable, and 16% is unused (FAO, 2020b). Most of the waterbodies used for aquaculture are operated by a single individual and contain water year-round (Jahan et al., 2015). There are different waterbody types used in Bangladesh: homestead pond, gher, commercial pond, beel, and rice-fish plot (Jahan et al., 2015). The descriptions of these waterbodies are presented in Table 3. For simplification, all types of aquaculture waterbodies will be referred to as fishponds. 9 Table 3. Descriptions of the five different types of waterbodies used in aquaculture in Bangladesh, adapted from (Jahan et al., 2015). Aquaculture Waterbody Type Definition A pond located near a homestead that is used for numerous domestic purposes. These are usually small in size. Homestead Pond A type of rice field in southern Bangladesh that is deepened to hold fish and/or crustaceans. The excess soil is used to Gher created dykes around the pond that can also be used to grow crops. Rice is not necessarily grown concurrently with the fish. Commercial Pond A pond dedicated to the year-round production of fish. Most of the fish grown are sold. These are formed in low-lying lands after floodwaters have recessed or after heavy rains. These ponds are large and can Beel be made suitable for fish by enclosing it with dykes. A type of rice field in northern Bangladesh that is deepened to hold fish. The excess soil is used to create dykes to prevent Rice-Fish Plot fish from escaping. Rice is grown with fish or in consecutive seasons. 2.3 Introduction to Remote Sensing Remote sensing is the acquisition and measurement of data on specific properties of phenomena, objects, or materials through the use of a recording device not in physical contact with what is being observed (Khorram et al., 2012). Remote sensing if often grouped with image processing (IP), geographic information systems (GIS), and Global Positioning System (GPS) to create geospatial science (Khorram et al., 2016). Remote sensing, in an environmental context, 10 involves sensors recording electromagnetic (EM) energy that emanates from the areas or objects being observed on Earth (Khorram et al., 2012). Remote sensing uses wavelengths along the electromagnetic spectrum to view targets. Remote sensing can be small-scale, individual projects looking only at one small section, to large-scale, remote sensing via planes or space-borne sensors. In the following sections, we will discuss the different EM spectrum sections that sensors use, different types of sensors, and the types of resolutions used to describe sensors. Space-borne sensors are broken up into three groups; visible, infrared, and microwave (JianCheng et al., 2012). Remote sensing in the visible wavelength range is commonly used for aerial photographs (Natural Resources Canada, 2015). Infrared remote sensing detects infrared radiation in the form of heat (CRISP & National University of Singapore, 2001). Microwave remote sensing relies on the different physical parameters that govern the microwave range to obtain new, useful data on targets (Woodhouse, 2017). All three are discussed further in sections 1.4.1 to 1.4.3. Table 4 provides examples of active satellites and their sensor types. 11 Table 4. List of active space-borne platforms and their sensor types. Satellite Name Sensor Type(s) GeoEye-1 Optical - Panchromatic and multispectral imaging (1.5.2) (UCSB, 2020) WorldView-1 Optical - Panchromatic and multispectral imaging (1.5.2) (ESA, 2020d) ALOS-2 Radar - SAR (1.5.1) (JAXA, 2020) Landsat-8 Optical - Thermal infrared sensor, radiometer for imaging (USGS, 2020) CBERS-4 Optical - Panchromatic, multispectral imaging (both visual and infrared) (ESA, 2020a) Sentinel-1A/1B Radar - SAR (1.5.1) (ESA, 2020b) Terra Optical - Radiometer, spectroradiometer (NASA, 2020c) There are two different modes of sensing: passive and active. Passive sensors measure the energy that is naturally available, while active sensors provide a source of energy for illumination of the target to be investigated (Natural Resources Canada, 2015). Lists of different passive and active sensors are in Table 5. 12 Table 5. List of different passive and active sensors (NASA EarthData, 2020). Type of Sensor Specific Sensor Description Laser altimeter Measures altitude from Earth’s surface using lidar Lidar Calculates distance by timing laser pulses to reflected light from pulses Radar Calculates distance by timing backscattered radiation from microwave radiation emissions Active Ranging Instrument Measures distance between sensor and target Scatterometer Uses radar to determine surface wind speed and direction Sounder Measures vertical distributions of atmospheric conditions (precipitation, temperature, humidity, cloud composition) Accelerometer Measures acceleration (translational or angular) Hyperspectral Detects narrow spectral bands in visible, near-infrared, Radiometer and mid-infrared sections of electromagnetic spectrum Imaging Radiometer Provides two-dimensional array of pixels Radiometer Measures intensity of electromagnetic radiation in Passive some bands within the spectrum Sounder Measures vertical distributions of atmospheric conditions (temperature, pressure, and composition) Spectrometer Measures and analyzes incident electromagnetic radiation Spectroradiometer Intensity of radiation in multiple wavelength bands The data from remote sensing can be described by four types of resolution: spatial, spectral, temporal, and radiometric. Spatial resolution refers to the ground area captured by a single pixel in an image. Spectral resolution refers to a sensor’s ability to store and detect different wavelengths and is represented by the width of wavelength interval or number of spectral channels. The temporal resolution is the amount of time for a sensor to revisit the same geographic location. 13 Lastly, the radiometric resolution is the sensor sensitivity to brightness (Khorram et al., 2012). Table 6 lists examples of different active satellites with their spatial and temporal resolutions. Table 6. Examples of active satellites and their respective spatial resolution and sensor(s) (Ottinger et al., 2016). Satellite Spatial Resolution Temporal Resolution Name (m) (day) GeoEye-1 0.41 – 1.65 2.1 – 8.3 WorldView-1 0.5 – 1.8 3.7 ALOS-2 1 – 100 14 Landsat-8 15 – 30 16 CBERS-4 20 – 64 5 Sentinel-1A/1B 5 – 80 12 Terra 15 – 90 16 2.4 Space-borne Sensors 2.4.1 Visible Light Remote Sensing Remote sensing using the visible light range in the EM spectrum mainly refers to satellite imagery. This is the simplest and oldest remote sensing technique. Images taken in the visible light range are variants of blue, green, and red light – just as human eyes do. Visible light images show the target the same way our eyes would (Natural Resources Canada, 2015). Objects that appear “whiter” have high albedo, meaning it reflects more radiation than “darker” objects (NSIDC, 2020). Cloud cover and daytime hours can impact what the sensor is able to detect. Disaster 14 management, weather forecasting, urban planning, archeology, and environmental assessment are all examples of applications of satellite imagery (PBS, 2007). 2.4.2 Infrared Remote Sensing The infrared region of the EM spectrum covers from 0.7 µm to 100 µm, making the range 100 times larger than the visible portion. Infrared can be broken up into two different categories: reflected infrared and thermal or emitted infrared. The reflected infrared is used for remote sensing in ways very similar to the visible section. Reflected infrared covers 0.7 µm to 3.0 µm. Thermal infrared is essentially the radiation that is emitted from the target, or the Earth’s surface (Natural Resources Canada, 2015). The wavelength emitted depends on its temperature (CRISP & National University of Singapore, 2001). Clouds also emit and reflect infrared radiation, preventing satellites from obtaining data from the surface of the Earth (NSIDC, 2020). 2.4.3 Microwave Remote Sensing Microwave interactions are typically governed by different physical parameters than other wavelengths along the EM spectrum (Woodhouse, 2017). They range from 1mm to 1 m in length (Gade & Stoffelen, 2019). Microwave sensing can see through obstructive weather conditions (Ellowitz, 1992) and pass through the top layer of soil (Woodhouse, 2017) in a way that other wavelengths cannot due to their length. Thermal emission can also be observed by passive microwave sensors, meaning that they do not rely on background sources like the Sun (Woodhouse, 2017). Microwave sensors are very responsive to different forms of water. They are able to make observations on soil moisture, vegetation water content, and snow cover (JianCheng et al., 2012). Table 7 shows the different bands of microwave sensing that are used and their respective characteristics. When discussing microwave sensing, the band is used to describe the wave attributes. 15 Table 7. Frequently used microwave wavelength bands, frequency bands, and their common names (Gade & Stoffelen, 2019). Band Wavelength band (cm) Frequency band (GHz) L 15.0 - 30.0 1.0 - 2.0 S 7.5 - 15.0 2.0 - 4.0 C 3.8 - 7.5 4.0 - 8.0 X 2.5 - 3.8 8.0 - 12.0 Ku 1.6 - 2.5 12.0 - 18.5 K 1.3 - 1.6 18.5 - 24.0 Ka 0.8 - 1.3 24.0 - 40.0 V 0.4 - 0.8 40.0 - 75.0 W 0.3 - 0.4 75.0 - 110.0 Common microwave sensors include radiometers (passive), altimeters, scatterometers, and synthetic aperture radar (SAR) (active). A few examples of satellites with microwave sensors include TerraSAR-X, ALOS/PALSAR, RADARSAT-2, and COSMO-SkyMed (JianCheng et al., 2012). There are also disadvantages to microwave sensing, such as long antennas are needed to obtain spatial resolutions appropriate for large areas due to long microwave wavelengths (Woodhouse, 2017). Active microwave setups, such as SAR, tend to be the heaviest, largest, and most power-consuming Earth-observing satellites (Woodhouse, 2017). 2.5 Remote Sensing for Aquaculture Remote sensing has the potential to provide data for aquaculture management, including site selection, mapping (Anand et al., 2020), environmental monitoring, and aquaculture inventory (Ottinger et al., 2016). The first step to the outcomes above is using data from remote sensing to locate under-utilized waterbodies to promote aquaculture production (Anand et al., 2020). 16 Generally, three types of products are used for aquaculture structures detection: SAR imagery, medium-resolution multispectral imagery, and high- and very-high-resolution multispectral and panchromatic imagery (Ottinger et al., 2016). There is a wide range of space-borne sensors and instruments that can be used for aquaculture observation. GeoEye-1 is the satellite with the lowest spatial resolution at 0.41 – 1.65 m, followed by WorldView-1 and WorldView-2/3, which each have 0.5 and 1.8 m range, respectively. Some satellites, such as ALOS, Envisat, and ERS-1/2, would work for aquaculture observation, but lack the appropriate spatial resolution needed for small-scale details. These satellites have spatial resolutions of 30 – 1000 m (Ottinger et al., 2016). 2.5.1 Synthetic Aperture Radar SAR is a technique for creating fine-resolution images from a radar system (NASA, 2020a). The term “synthetic” refers to the processing method of backscattered waves to improve the azimuthal resolution, which allows for smaller spatial resolutions (Woodhouse, 2017). The wavelength range for SAR is from a few to tens of centimeters (NASA, 2020a). The resolution of an imaging radar depends heavily on the antenna length; longer antennas provide finer resolution. This creates issues when using SAR on a satellite because an antenna would need to be kilometers long. To address this, the radar’s real aperture is synthetically enhanced by using the Doppler shift of the backscattered signal (Gade & Stoffelen, 2019). As the SAR moves along its path, it sweeps the antenna’s calculated length across the ground while continuously transmitting pulses and receiving the backscattered signal (NASA, 2020a). Both Figure 3 and Figure 4 depict the Doppler shift but as two different representations. At point “A” in Figure 3, the backscattered signal causes a positive Doppler shift at “a”, a zero Doppler shift at “b”, and a negative Doppler shift at “c”. This phenomenon can be referred to as 17 the Doppler History, which the end result of is a much longer aperture (Gade & Stoffelen, 2019). SAR systems use pulse compression, a technique where the echo of a pulse is matched with its original signal creating a shortened pulse of higher energy (Gade & Stoffelen, 2019). The graphical representation of this can be seen in Figure 4. Equation (1 shows how to calculate the beam size on the ground/target (NASA, 2020b): Wavelength Distance to Target × = Beam Width at Target (1) Antenna Dimension Figure 3. Illustration of the SAR Doppler technique (adapted from Gade & Stoffelen, 2019)). 18 Figure 4. SAR backscattered Doppler shift (adapted from Gade & Stoffelen, 2019)). Spaceborne SAR systems will typically have a wavelength in the C or X bands. A SAR system sends either horizontally (H) or vertically (V) polarized pulses. They then receive either H or V polarized return signals. SAR systems are characterized by how they send and receive signals. For example, a SAR system that sends horizontally polarized signals and receives horizontally polarized signals would be named HH (Travaglia et al., 2004). There are three SAR acquisition modes: Stripmap, ScanSAR, and Spotlight (Gade & Stoffelen, 2019). Stripmap is the classic acquisition mode in which the radar stays at a constant angle. ScanSAR mode involves multiple parallel swaths being scanned by the radar, resulting in one wide swath. Lastly, spotlight mode continuously changes the direction of the radar so that a single spot on Earth’s surface is imaged for a longer time (Gade & Stoffelen, 2019). The unique SAR system does have some complications. SAR images are impacted by noise, or speckling, created by constructive and destructive interference between the backscattered energy. This causes the value of the pixel to either increase or decrease, creating random bright and dark spots. It can be useful to apply speckle reducing procedures to better SAR images (Travaglia et al., 2004). 19 2.5.2 Multispectral and Panchromatic Imagery Panchromatic images are taken using the UV (100 nm – 400 nm) and visible portions (400 nm – 700 nm) of the EM spectrum (Natural Resources Canada, 2015). The resulting images are in black-and-white (Khorram et al., 2012). Multispectral imaging involves simultaneously imaging in multiple wavelength bands to analyze targets (Coffey, 2012). A multispectral system normally provides a mixture of visible, near infrared (NIRS), short-wave infrared (SWIR), mid-wave infrared (MWIER), and/or long-wave infrared (LWIR) bands (Coffey, 2012). Multispectral imaging can be thought of as layers of panchromatic images, where each layer corresponds to a specific portion of the EM spectrum (Khorram et al., 2012). A benefit of multispectral imagery is much less subjective than aerial photography due to the higher information content (Chu, 2020). 2.6 Remote Sensing Water Indices for Small Waterbody Location Inland waterbodies play a major role in contributing both food and livelihood security in rural areas (Anand et al., 2020). Small waterbodies include ponds, small lakes, low-order streams, ditches, and springs, but it is an ambiguous term, nonetheless. Ponds are standing waters varying from 2 to 5 ha in area and can be permanent or seasonal. Small lakes constitute standing water greater than 5 ha, but smaller than 100 ha (Biggs et al., 2017). There is no agreed-upon definition for a small stream, but the width of a small stream could vary from <3 to 6 m. Ditches are man- made channels which typically (a) are horizontal, (b) have linear boundaries and often turn at right angles, and (c) show little influence of the natural landscape contours. Lastly, springs are fixed places where groundwater emerges to the surface (Biggs et al., 2017). In order to discover impacts of the number and extent of ponds on climate variability, Al Sayah et al. (2020) researched the French Claise watershed. The study used LANDSAT images, varying in time, to derive a Normalized Difference Water Index (NDWI). An NDWI map is used 20 to find and delineate water surfaces on the visible green light and near-infrared radiations. NDWI maps are widespread and have been used in applications as fine as mapping swimming pools. The NDWI map that was created had an accuracy of 85.74% for pond count and 75% accuracy for pond spatial allocation. In addition to the NDWI, the land surface temperature (LST) index was extracted from each LANDSAT map to show how pond numbers impacted the climate of the region. An LST index can be used to study the temperature measured on Earth’s surface and is known for being one of the most suitable accurate indicators for studying climatic variabilities. The study reached the conclusion that pond-less zones had higher average temperatures than zones with ponds. In addition to NDWI, the Automated Water Extraction Index (AWEI), Modified Normalized Difference Water Index (MNDWI), and Water Ratio Index (WRI) (Zeng et al., 2019) are also widely used indices for detecting waterbodies. Each of these indices is a way of extracting or pinpointing water in a target area. Each index requires different bands of the EM spectrum. The AWEI is an index formulated to effectively eliminate all non-water pixels in an image and improve the accuracy by removing shadow pixels (Feyisa et al., 2014). The MNDWI is advantageous in using to reduce and potentially remove built-up land noise as well as pinpoint water. Because of this, it is suitable for extracting water in an area with a background dominated by built-up land areas (Xu, 2006). The WRI was created to compensate for water’s high reflectance in green and red bands compared to NIR and medium infrared (Gautam et al., 2015). Anand et al. (2020) used high-resolution Cartosat 1 PAN and IRS ResourceSAT LISS IV merged imagery as well as Sentinel 2 Multi-Spectral Imagery to determine the Spatio-temporal water spread and effective water spread area for aquaculture in Chhattisgarh, India. The study used data from the years 2016 to 2018. The ortho-corrected images provided radiometric measurements 21 in top of atmosphere (TOA) reflectance. The study used an NDWI to create a Water Surface Area (WSA) map from each individual image. Equation (2) shows how the green (XGreen) and NIR (XNRI) bands of the Sentinel 2A/2B MSI images are used to create the NDWI (McFeeters, 1995, Gautam et al., 2015). Green and NIR were chosen because of how they highlight the difference between water and land. Aquatic vegetation within a waterbody is differentiated from water. This also can happen in silty waters. 𝑋 −𝑋 (2) NDWI = 𝑋 +𝑋 To aggregate the NDWI water bitmaps, the study used Water Presence Frequency (WPF). The WPF is represented in Equation (3) where WPFj is WFP of jth pixels in a time period; Ij is jth pixel having water in the selected NDWI images; n is the number of images (Anand et al., 2020). The WPF values range from 0 to 100%, where below 66% WPF concluded dry bed area and over 66% concluded water. ∑ 𝐼 (3) WPF = ∗ 100 𝑛 The satellite data (e.g., Cartostat 1 PAN, IRS ResourceSAT LISS IV, and Sentinel-2 MSI) provided an approach for establishing the number of small waterbodies required for a certain production potential in standard production conditions. The study first created a waterbody boundary layer vector file, chose the appropriate time interval (in this case June to September due to monsoons), then extracted the surface water using NDWI and WPF before generating a composite seasonal water surface area map. To obtain accuracy, the study used field verification. Chhattisgarh is a landlocked state in central India which has large numbers of waterbodies. The study mapped over 120,000 waterbodies, creating a total area of 202,000 ha. Of the waterbodies 22 mapped, 97% had an area less than 5 ha. The study concluded that the methods used could be replicated in other areas with poor in-situ data. SAR data is beneficial in addition to other forms of remote sensing (David Ballester- Berman et al. 2018). The Travaglia et al. (2004) study used SAR data because of its all-weather capabilities and because the backscatter allows for identification and separation of features. The aim of the study was to map and inventory coastal aquaculture and fisheries in Lingayen Gulf, the Philippines, using data from ERS-2 and RADARSAT-1. The ERS-2 satellite has a quasi-polar orbit, meaning that the descending orbit is approximately opposite to the ascending orbit. The angle between the scanning directions of two ERS SAR images is 152.7 degrees. Rough water conditions cause the mean backscattering coefficient of fishponds to increase on C-band images acquired in VV polarization than in HH. The ERS SAR has VV polarization while the RADARSAT SAR has HH. Using both SAR systems, a more accurate image is created. An analysis of the two images identifies all features in the target area. The results of the SAR imaging were compared to a database of topographic maps to define the accuracy of the study. The comparison concluded the study had an accuracy of 95% for identifying fishponds. Platforms like Google Earth Engine (GEE) make mapping aquaculture ponds on a national scale much more efficient (Duan, Li, Zhang, Chen, et al., 2020). Duan et al. (2020) integrated spectral, spatial characteristics and morphological operations to create a decision tree classifier. The decision tree was used to extract the aquaculture pond regions along the Chinese coastal zone with an accuracy of 96%. The data came from the Landsat 8 Surface Reflectance Tier 1 dataset from GEE. The decision tree classifier has five steps: creating boundaries of potential aquaculture pond regions, extraction of waterbodies, extraction of intensive ponds, extraction of extensive ponds, and post-processing. Understanding the typical environment for ponds, e.g., low-lying 23 planes of coastal areas, helps determine where potential aquaculture pond regions are. The AWEI distinguished between non-water and water range the low-lying coastal zones, extracting the large waterbodies (ocean, lakes, rivers, extensive ponds). Equation (4 shows how the AWEI is determined, where ρ represents the reflectance value of spectral bands of Landsat 8: band 3 (green), band 5 (NIR), band 6 (SWIR 1), and band 7 (SWIR 2) (Feyisa et al., 2014, Wicaksono & Wicaksono, 2019). AWEI sh = 4 ∗ (𝜌 −𝜌 ) − (0.25 ∗ 𝜌 + 2.75 ∗ 𝜌 ) (4) Feyisa et al. (2014) introduced an additional Equation (5 to the AWEI that removes shadow pixels that were not removed in the previous equation where band 1 refers to the blue section of the EM spectrum. AWEIsh = 𝜌 + 2.5 ∗ 𝜌 − 1.5 ∗ (𝜌 +𝜌 ) − 0.25 ∗ 𝜌 (5) To extract intensive aquaculture ponds and small waterbodies from what was identified by the AWEI, Duan et al. (2020) used an MNDWI proposed by Xu (2006). The MNDWI is shown in Equation (6, where the middle infrared (MIR) band replaces the NIR band. 𝑋 −𝑋 (6) MNDWI = 𝑋 +𝑋 The MNDWI produces three results: water will have greater positive values than what the NDWI results in as it absorbs more MIR light than NIR, built-up land will have negative values, and vegetation and soil will have negative values as soil reflects MIR light more than NIR (Xu, 2006). The WRI depends on the spectral reflectance of the green, red, NIR, and MIR bands. Equation (7 shows the WRI equation (Mohsen et al., 2018). A WRI value greater than 1 represents water. 24 𝑋 +𝑋 (7) WRI = 𝑋 +𝑋 Mohsen et al. (2018) used a combination of the NDWI and WRI to extract water features from satellite images of Lake Burullus in Egypt. WRI requires the MIR band, a band that many satellites do not measure (Mukherjee & Samuel, 2016). The Mohsen et al. (2018) study did not report the accuracy of the methods but instead conducted a statistical analysis using the Mann- Kendall test on whether Lake Burullus was decreasing in size from the year 1972 to 2015. Using the water data from the WRI and NDWI, Mohsen et al. (2018) concluded that the lake had lost approximately half of its surface area in the given time frame. While each index mentioned does extract water data from a target area, they are not all the same. Choosing which index or indices to use depends on a few factors, including, but not limited to, what data is available, the topography of the target area, and the objective of the study. Each index requires different bands in the EM spectrum. If that data is not available for the specific target area, then a different index must be used. The AWEI is useful in areas where steep topographic changes can create shadows in the images but may not be needed in flat areas. The MNDWI is useful for eliminating noise in an image. Many studies either used the indices in tandem to create the best image or used multiple indices separately to compare the results of each. 2.7 Remote Sensing Aquaculture Applications 2.7.1 SAR Applications Prasad et al. (2019) used SAR data from Sentinel-1 satellite to assess the coastal aquaculture in India from September 2014 to June 2017. The SAR instrument on Sentinel-1 operates at a C-band frequency of 5.5 GHz. The study used all available VH polarized SAR data in the interferometric wide-swath (IW) mode and the ground range detected high-resolution 25 (GRDH) format. In the IW mode, three sub-swaths were captured using terrain observation with progressive scans SAR (TOPSAR). TOPSAR is a form of ScanSAR imaging where the antenna beam is switched cyclically between multiple adjacent sub-swaths to obtain data (ESA, 2020c). This Sentinel SAR method resulted in a swath width of 250 km and 5 m by 20 m spatial resolution. Prasad et al. (2019) methods can be explained in five steps: preprocessing, calculation of a temporal median layer, topographic masking, segmentation, and spatial analysis. All data was obtained from the free and open-source Sentinel Application Platform (SNAP). Preprocessing involved removing thermal noise, converting intensity values, and correcting terrain distortions. The temporal median layer was calculated in the VH polarization at the pixel level. VH polarization was used because it is slightly better for differentiating land and water areas because of its more pronounced bimodal distribution of backscatter values. The temporal median layer was averaged over time, greatly diminishing speckle noise. Aquaculture and rice fields can be differentiated despite often having the same features because the aquaculture ponds will appear much darker in the temporal median layer because of their year-round water presence. The topographic masking excluded mountain areas and rough terrain since aquaculture typically occurs in flat environments. The target area was then broken into 11 approximately equal sub-regions for segmentation, then the Euclidian intensity pixel distance criterion was applied to see if two pixels belonged to the same segment. The final step involved using object-based image filtering with aquaculture pond characteristics inputs to identify ponds. After validating the results with reference datasets, the methods led to an accuracy of 97% for identifying aquaculture ponds. Ottinger et al. (2017) used similar methods when using high-resolution Sentinel-1 SAR data to identify and map aquaculture ponds in China and Vietnam from September 2014 to September 2016. The target areas, more specifically, were the Mekong Delta, Red River Delta, 26 Pearl River Delta, and Yellow River Delta. High spatial resolution is needed for detecting certain pond features like embankments, levees, or dikes. The study also used Sentinel-1, but instead of data from both Sentinel-1A and -1B, Ottinger used Sentinel-1A dual-polarized (VV and VH) data in IW and Ground Range Detected High Resolution (GRDH). Sentinel-1B was not used because it finished its commissioning phase in September 2016. After in situ observations, more than 3000 aquaculture ponds were identified in the target areas. These would be used to calculate accuracy of the computer identified ponds. Only ponds within 20 km of the shoreline were selected for sampling. The data for the Ottinger et al. (2017) study was also obtained from SNAP. The preprocessing included removal of thermal processing. After this, a radiometric calibration was performed, and then a terrain correction applied. Since many aquaculture ponds are enclosed, they have a very low backscatter value and distinct characteristics. To reduce speckling, the pixel-wise median was calculated, a process of comparing pixels to their surrounding pixels to remove outliers. Similarly to the previously discussed study, Ottinger et al. (2017) also calculated a temporal median image, terrain masking, and segmentation. However, unlike Prasad et al. (2019) Ottinger et al. (2017) performed an edge sharpening, finding that bilateral denoising and non-local means filter as the two best methods. Both methods successfully removed noise, but the non-local means filter did a better job of preserving detail. The average overall accuracy of their methods was 83%. These two studies used similar methods and obtained high accurate results. The Prasad et al. (2019) study followed the Ottinger et al. (2017) study and had better results. Both studies looked at aquaculture in low-lying coastal areas, similar to the landscape in Bangladesh, which is the focus of our study. 27 2.7.2 Multispectral and Panchromatic Imaging Applications An alternative to radar imaging is optical imaging, such as multispectral and/or panchromatic imaging. Virdis (2014) used high and very-high resolution panchromatic imaging, SPOT5 and Worldview-1 (both optical sensor satellites), and machine learning to detect shrimp farms in the Tam Giang-Cau Hai Lagoon in central Vietnam (70 km in length). To classify aquaculture shrimp farms, the study looked at panchromatic images from SPOT5 Level11A (5 m pixel size) and Worldview-1 Ortho-Ready Standard OR2A panchromatic images (0.5 m pixel size). The data collected first went through a geometric correction and spatial accuracy assessment. The images then were cropped to fit the two areas of interest (AOIs), had contrast- enhanced, and edge detection filtering. The study used a non-commercial software called SPRING to segment the image through clustering. The clusters were then compared to a previously created aquaculture reference database of the area to find accuracy. The SPOT5 images led to an accuracy of 84.7% and 93.2% for AOI1 and AOI2, respectively. The Worldview-1 images led to an accuracy of 90.6% band 95.7% for AOI1 and AOI2, respectively. The algorithm implemented through the SPRING software worked well with both SPOT5 and Worldview-1, especially in AOI2, where there was a higher contrast between ponds and embankments. Zeng et al. (2019) used medium resolution multispectral images to extract aquaculture ponds from water surfaces around inland Liangzi Lake in China. The study took medium- resolution multispectral image data from Landsat-5 (Thematic Mapper, TM) and Landsat-8 (Operational Land Imager, OLI), as well as high-resolution observation data from GaoFen-1 (World Field of View, WFV). The data was obtained via the United States Geological Survey (USGS) and the China Center for Resources Satellite Data and Application (CRESDA). All images 28 chosen had to be cloud-free and from December to January for minimal vegetation coverage. The water pixels were extracted using the MNDWI and the NDWI. These indices were chosen after calculating numerous indices and their accuracy for each data source. The MNDWI and NDWI had the best accuracy. The MNDWI was applied to the OLI and TM data, while the NDWI was applied to GaoFen-1 data. The water pixels were then separated into segments and geometrical features were calculated. The feature vectors formed were used as inputs to Support Vector Machine (SVM) classifier, an algorithm used for classification, regression, and outlier detection. The SVM classifier detected whether the water pixels were natural water surfaces or aquaculture. The results were validated from previously digitized and labeled water types in the area. The user accuracy ranged from 91.8% to 99.2% depending on the area of the lake (either eastern or western Liangzi Lake) and which image sensor. The lowest accuracy, 91.8% was the TM sensor in eastern Liangzi Lake. The highest accuracy was the OLI sensor in western Liangzi Lake. The methods used in this study can be applied to identify inland aquaculture ponds around the world. Landsat satellites have been collecting data since 1972, meaning the archives have great potential to monitor historic changes in inland lake environments. 2.7.3 Table of Relevant Studies on SAR and Optical Remote Sensing Table 8 lists relevant studies that utilized radar (SAR) and/or optical satellite imagery. 29 Table 8. List of relevant studies for fishpond detection using radar and optical satellite imagery. Resolution Product References Country Algorithm Prediction accuracy Software Spatial Temporal Synthetic aperture radar (SAR) Two images at ERS-2 (C band VV different dates polarization), 12.5m (ERS-2) Source: ESA Travaglia et al., 2004 Philippines (ERS-2) Visual interpretation 95% Radarsat-1 (C band HH 6.25m (Radarsat-1) Preprocessing: ERDAS, ArcGIS A single image polarization) (Radarsat-1) Object-based (multi-temporal Preprocessing: Radar Analysis Radarsat-1 (C band HH 6m Less than a month segmentation). 83.1% (Liu et al., 2010) China Package - PCI; ERDAS polarization) (raw 4.6x5.1m) (three images total) Parameters: scale, color, and kappa 0.81 Processing: Definiens shape. 10m (after 5x1 Source: Sentinel Scientific Data multi-looking) Object-based mapping Overall 84% (ranging Hub/Google Earth Engine Less than two weeks Sentinel-1A/ B (C band (Ottinger et al., 2017, (raw 5x20m) (temporal filtering, topographic from 80% to 88%) Preprocessing: Sentinel China, Vietnam (66 to 192 scenes in VV + VH polarization) 2018) Ground Range masking, connected component kappa 0.68 (ranging Application Plattform (SNAP) a two-year period) Detected High segmentation) from 0.59 to 0.77) Processing: GDAL tools, Orfeo Resolution (GRDH) Toolbox (OTB) Source: Alaska Satellite Facility's Sentinel-1A/B Single (Ballester-Berman et al., One image per study Unsupervised Wishart data portal Spain, Norway 2.3x14m Not reported Look Complex (SLC) 2018) site was used classification Processing: Sentinel Application Platform (SNAP) Images are tiled into patches. A primitive feature vector is extracted from each patch. The TerraSAR-X, Images for a single patches are then clustered and (Dumitru et al., 2018) Albania, Greece Not reported From 80 to 95% Not reported Sentinel-1B date were used classified using a Support Vector Machine classifier and predefined labels from optical imagery. 30 Table 8 (cont’d) Resolution Product References Country Algorithm Prediction accuracy Software Spatial Temporal Source: Sentinel Scientific Data Object-based mapping Hub/Google Earth Engine Imagery collected Sentinel-1A/B (C band (temporal filtering, topographic Preprocessing: Sentinel (Prasad et al., 2019) India 10m for a two-year Overall 90% VH polarization) masking, connected component Application Plattform (SNAP) period segmentation) Processing: GDAL tools, Orfeo Toolbox (OTB) Optical Two images were 43 (satellite) out of Landsat 4 and 5 employed (dry (Kapetsky, 1987) Zimbabwe 1:250,000 scale Visual interpretation 77 (aerial photos) Not reported imagery season, and end of (56%) rainy season) IRS 1D (panchromatic 65 to 75% of ponds images) 6m (IRS 1D) Two images at (De Graaf et al., 2000) Bangladesh Visual interpretation larger than 1,000 m2 Not reported SPOT (multi-spectral 20m (SPOT) different dates are detected. images) CORONA (panchromatic), CORONA (2m) SPOT-3 (multispectral), SPOT-3 (1:50,000) Up to two images of ERS-1 (SAR), ERS-1 (12.5m) each platform along SIR-C (SAR), (Huda et al., 2010) Bangladesh SIR-C (30m) multiple years Visual interpretation Not reported Not reported X SAR, X SAR (25m) between 1972 and Landsat 5 (multi- Landsat 5 (30m) 2003 spectral), IRS (6m, 23m) IRS (panchromatic, multi-spectral) Image segmentation using basin-detection and SPOT5 Level1A, 5m (SPOT5) One image for each region-growing techniques. (Virdis, 2014) Vietnam > 95% SPRING Worldview-1 0.5m (Worldview) product Unsupervised clustering classification ISOSEG Algorithm incorporating Optical imagery (Z. Yu, 2019) Bangladesh Not reported Not reported spectral and spatial filtering on Not reported Google Earth Engine multi-temporal images. IKONOS, 0.5 - 1m Quickbird, (panchromatic) Images from 2000 to (Gusmawati et al., 2017) Indonesia Visual interpretation > 80% ArcGIS Worldview-2, 1.5 - 4m 2015 Worldview-3 (multispectral) 31 Table 8 (cont’d) Resolution Product References Country Algorithm Prediction accuracy Software Spatial Temporal For Sentinel-1, same as Ottinger et al. (2018) For Landsat: 1) Cloud and cloud shadow masking 2) NDWI calculation 3) Difference of 10th and 90th 66 to 174 scenes in a percentiles (to discard rice two-year period for paddy fields) Sentinel-1 4) Set threshold for land-water Sentinel-1A, 10m (Sentinel-1) separation (trial and error, Overall 89% Landsat 5/7/8 Surface (Stiller et al., 2019) China Google Earth Engine 30m (Landsat) Yearly cloud-free visual interpretation) kappa 0.78 Reflectance (Level 2) image composites 5) Comparison with Sentinel-1 between 1984-2016 detected ponds, water using Landsat proportion over time is obtained 6) For each year, Sentinel-1 fishponds that had at least 1/3 of their surface are retained. 7) False positives are detected using the Global Surface Water (GSW) raster dataset. Water surface extraction is performed using NDWI, MNDWI, AWEI, NDVI, WRI, NDMI. Thresholds are determined using the Otsu method (Otsu, 1979). The most accurate water and non-water binary image is selected. Up to three images Source: USGS, China Center for Landsat 5/8, Gaofen-1 from the end of Water segments are obtained, Resources Satellite Data and 30m (Landsat) Overall > 94% Wide Field of View (Zeng et al., 2019) China December to and for each one, geometric Application 16m (Gaofen) kappa > 0.8 (WFV) January for each features (area, regularity, Preprocessing: FLAASH tool in product perimeter) are determined based ENVI 5.3 on boundary tracing and contour-based regularity and curvature computations. A SVM classifier incorporating the geometric features is used to separate aquaculture from natural water surfaces. 32 Table 8 (cont’d) Resolution Product References Country Algorithm Prediction accuracy Software Spatial Temporal Updating approach using visual interpretation and automatic Source: USGS Images for 5 years classification based on water Preprocessing: FLAASH tool in Overall > 87% Landsat 5/7/8 (Ren et al., 2019) China 30m between 1984 and surface extraction with NDWI ENVI 5.0 kappa > 0.80 2016 and object-based classification Processing: eCognition Developer (parameters: scale, shape, and 8.64 compactness). Sentinel-2: Images for two time periods: February Visual interpretation (2.5m 10m (Sentinel-2) and May (selection Sentinel-2A/B, data) Source: Copernicus open access 2.5m (Cartosat 1 based on agricultural Cartosat 1 PAN, (Anand et al., 2020) India Not reported hub PAN + IRS LISS operations). IRS LISS IV Water surface extraction using Processing: ArcGIS IV) NDWI (Sentinel-2) Cartosat-1 + IRS: A single merged image. Decision-tree classifier: 1) Non-water land elimination using a water extraction index 2) Intensive aquaculture ponds Revisit time: 16 extraction using the MNDWI days and the 8-neighborhood Landsat 5/8 Surface (Duan, Li, Zhang, Liu, et Seven-time slices Laplacian operator. Overall >91% China 30m Google Earth Engine Reflectance Tier 1 al., 2020) from 1988 to 2018, 53) Extensive aquaculture ponds kappa > 0.79 year each, between identification using two shape April to October indexes representing the regular shape of these ponds. 4) Merging different pond types) Landsat: 86% (pond count) 75% (pond spatial Landsat 5/7/8, Sentinel- 30m (Landsat) Multiple images for allocation) Source: USGS Earth Explorer (Al Sayah et al., 2020) France NDWI 2A 10m (Sentinel-2A) October Sentinel: (Landsat) 93% (pound count) 84% (pond spatial allocation) 33 2.8 Basic Machine Learning and its Applications in Detecting Aquaculture Farms Machine learning is computer algorithms for translating human ways of learning into machines (Faul, 2019). Machine learning is founded on the need for a computer to convert data examples into knowledge (Kubat, 2017). It is a field of computer science that studies techniques for obtaining results to complex issues that are difficult to solve using conventional programming methods. Machine learning algorithms have provided solutions to very complex problems, from internet searching to speech recognition (Rebala et al., 2019). Benefits to machine learning include a reduction in programming time, ability to customize and scale products, and complete “unprogrammable” tasks (Google, 2020). Machine learning generally uses two different techniques: supervised learning and unsupervised learning. Supervised learning trains a model on known input and output data to predict future outputs by using classification and/or regression. Classification is a technique that predicts discrete responses, meaning that it categorizes input data (e.g., is a pond is a fishpond or not). Different classification algorithms include SVM, Discriminant Analysis, Naïve Bayes, and Nearest Neighbor. Regression is a supervised learning technique that predicts continuous responses (e.g., temperature changes). Examples of different regressions are Linear Regressions, Support Vector Regression, Ensemble Methods, Decision Trees, and Neural Networks. Unsupervised learning is a model that finds hidden patterns or structures in input dataset. This method uses clustering, a method of grouping data together based on certain similarities or patterns. Different types of clustering methods include k-Means, k-Medoids, Hierarchical, Gaussian Mixture, Neural Networks, and Hidden Markov Model (MathWorks, 2016d). Table 9 and Table 10 describe some common supervised and unsupervised algorithms and when they are best used. 34 Table 9. Common supervised algorithms and their applications (MathWorks, 2016a). Algorithm Name Process Applications Logistic Creates a model that predicts the  Data is clearly separated by a Regression probability of a binary event single, linear boundary  Can be a baseline for more complex classifications k Nearest Categorizes based on characteristics  Establishes benchmark learning Neighbor (k-NN) on nearest neighbors in a dataset rules  Memory usage is not important  Prediction speed is less of a concern Support Vector Classifies through discovering linear  Data only has two classes Machine (SVM) decision boundary that separates all  High-dimensional, nonlinearly data points from one group to separable data another  Easy to interpret Neural Network Highly connected networks relate  Modeling non-linear systems inputs to outputs  Data is available in increments  If there is a possibility of unexpected change in input data  Model does not need to be easily interpreted Naïve Bayes Assumes presence of a certain  Small datasets with many feature is unrelated to the presence features of another feature  Easy to interpret  Model will encounter circumstances not covered in training data 35 Table 10. Common unsupervised algorithms and their applications (MathWorks, 2016b). Algorithm Name Process Applications k-Means Divides data into k number of  k is known mutually exclusive clusters  Fast clustering of large data sets k-Medoids Divides data into k number of  k is known mutually exclusive clusters, but  Fast clustering of categorical cluster center must be a data point data  Scales to large datasets Hierarchical Creates nested sets of clusters into a  Number of clusters unknown binary, hierarchical tree  Provides visualization for selection Self-Organizing Neural network that changes a  Provides data visualized in 2D or Map dataset into a multidimensional map 3D  Deduces dimensionality of data by preserving the shape Using machine learning occurs in steps. The first step is to access the proper data, then preprocess data (e.g., check for outliers or missing data points), derive certain features (i.e., turning raw data into information), train models using the features, iterate to find the best model, and then integrate the best model into a productive system (MathWorks, 2016c). Machine learning techniques such as SVM has been used extensively for remote sensing classification as a supervised method and can classify aquaculture ponds from natural water surfaces. For example, Zeng et al. (2019) used the SVM classifier algorithm to identify aquaculture farms in an inland lake in China. Geometric features, like perimeter, area, and contour-based regularity, are examples of input features into an SVM. The Zeng et al. (2019) study used geometric features from satellite images as training datasets. These datasets enabled the algorithm to pick out aquaculture ponds from natural surface water with an accuracy of at least 91.8%. In 36 addition to supervised classification, machine learning can be unsupervised, meaning without being presented a dataset with a known outcome. In SNAP there is the Wishart classifier, which has been used to identify differences in aquaculture structures within SAR imagery (Ballester- Berman et al., 2018). WorldView-1 and SPOT5 imagery have been also used in partnership for fishpond classification using the unsupervised Isogeg classifier in the Spring software (Virdis, 2014). 2.9 Limitations of Remote Sensing Techniques in Detection of Small Waterbodies Despite previous research, identifying small waterbodies or aquaculture ponds on a large scale using remote sensing data can be a very challenging task mostly due to:  There is no universal index to use for extracting water from an image (Zeng et al., 2019)  There are multiple machine learning algorithms could be used for a fishpond detection, but there is no grantee that any or all of them work for an area of interest (Virdis, 2014; Dumitru et al., 2018; David Ballester-Berman et al., 2018; Zeng et al., 2019)  Inland aquaculture comes in many different forms (Jahan et al., 2015). Because of this, there are no apparent geometric features to provide for an algorithm. Some aquaculture ponds like homestead ponds and beels form naturally with the shape of the land. Ghers and rice-fish plots are aquaculture coinciding with agriculture, meaning the shape of the pond will take the shape of the rice field it is built into.  There is no freely available high-resolution (< 5 m pixel size) imagery to improve current fishpond detection (Z. Yu et al., 2020). 37 2.10 Goals There is currently a lack of understanding of the extent of aquaculture in Bangladesh. The goals of this thesis are to find appropriate remote sensing data, applying different manual and machine learning techniques to distinguish various aquaculture, and verify the manual and machine learning algorithms against ground-truthing. The result will be used to improve official statistics and to enhance the capacity for aquaculture production and regulations within Bangladesh. 38 3 HARNESSING MACHINE LEARNING TECHNIQUES FOR LARGE-SCALE MAPPING OF INLAND AQUACULTURE WATERBODIES IN BANGLADESH 3.1 Introduction Aquaculture (farming fish and other aquatic animals) is growing at an extremely rapid pace, increasing around 6.5 times, from 13 million tonnes to 85 million tonnes over the past thirty years (FAO, 2020c), and is forecast to continue growing rapidly in coming decades (Kobayashi et al., 2015) Most of this growth has occurred in Asia, which accounts of 89% global aquaculture production (FAO, 2020c). Aquaculture farm expansion in Asia has been concentrated mainly in water abundant deltaic regions (Bernzen et al., 2021). The growth of aquaculture since the 1990s has been particularly rapid in Bangladesh, induced by demand caused by rising incomes, urbanization, a growing population, and declining supplies of fish from capture fisheries. This development has occurred in a largely unplanned and spontaneous way as farmers have converted rice paddies into fishponds that typically offer much higher returns than paddy cultivation (Belton et al., 2018). As a result, Bangladesh is now the 5th largest aquaculture producing country in the world (FAO, 2020c). Aquaculture in Bangladesh has shifted dramatically from subsistence to commercial farming in the last three decades (Hernandez et al., 2018). During this time, the farmed fish market grew from 124,000 tons to 2.4 million tons, and the number of people involved in the aquaculture value chain tripled (Hernandez et al., 2018; Department of Fisheries, 2018). This dynamic growth has created a huge new industry that makes substantial contributions to the rural economy in many parts of the country, but the rapid emergence of the sector means that it performance is not well understood by policymakers and researchers (Jahan et al., 2015; Hernandez et al., 2018, Belton & Azad, 2012). Some observers contend that unplanned expansion of aquaculture is creating 39 competition for arable land necessary to produce rice, Bangladesh's primary food crop, creating pressures on limited space that need to be addressed (Yu et al., 2018; Hashem et al., 2014). Moreover, studies have suggested that national aquaculture statistics may not have kept pace with this rapid growth, meaning that production volumes might be underestimated (Belton & Azad, 2012). Accurately determining the physical extent of aquaculture in Bangladesh can be approached in different ways. One is to conduct physical surveys. These would provide accurate information about the use of waterbodies throughout the region, but because of the large areas and populations involved, this approach would be highly resource-intensive (Rhodes et al., 2015). Using remote sensing to identify where aquaculture is occurring is an alternative approach with potential to provide data for aquaculture management, including site selection for potential farmers, mapping (Anand et al., 2020), environmental monitoring, and aquaculture inventory (Ottinger et al., 2016). Remote sensing is helpful for analyzing large areas and obtaining information quickly, but it is difficult to determine use from an identified body of water. Generally, three types of remote sensing products are used for aquaculture structure detection: synthetic-aperture radar (SAR) imagery, medium-resolution multispectral imagery, and high- and very-high-resolution multispectral imagery (Ottinger et al., 2016). SAR is a technique for creating fine-resolution images from a radar system (NASA, 2020a). The term “synthetic” refers to the processing method of backscattered waves to improve the azimuthal resolution, which allows for smaller spatial resolutions (Woodhouse, 2017). Multispectral imaging refers to sensors that can simultaneously capture many wavelength bands across the electromagnetic spectrum (Coffey, 2012). Medium spatial resolutions refer to any images with pixel sizes of 10 to 30 m, whereas high and very high spatial resolutions refer to any images with pixel sizes of 30 cm to 5 m per pixel 40 (Earth Observing System, 2019). Ultimately, selecting a particular product mostly depends on trade-offs between cost, spatial resolution, temporal availability, and surface area coverage (Bello & Aina, 2014, Zoran et al., 2010). Some current methods for utilizing SAR imagery to identify aquaculture structures include object-based mapping in China, Vietnam (Ottinger et al., 2018), and India (Prasad et al., 2019) and Wishart classifiers applied to coastal regions of Spain and Norway (Ballester-Berman et al., 2018). However, filtering to remove speckle noise in SAR imagery negatively impacts its spatial resolution (Z. Yu et al., 2020). For medium-resolution optical imagery, the current methods include object-based mapping and classification along China’s coasts (Ren et al., 2019), utilizing the Normalized Difference Water Index (NDWI) to determine aquaculture value thresholds in Southern China (Stiller et al., 2019), and a combination of pixel selection and image segmentation to find aquaculture in northern Bangladesh (Z. Yu et al., 2020). Meanwhile, methods utilizing Landsat medium resolution multi-spectral imagery (Ren et al., 2019) do not work for identifying small (approximately 0.002 km2) fishponds because of its 30 m spatial resolution. Finally, high- resolution optical imagery can be resource-intensive and expensive; however, it has been used for visual aquaculture interpretation in India (Anand et al., 2020) and Indonesia (Gusmawati et al., 2017). This study aims to identify bodies of water throughout the study region and determine if their use is for aquaculture. This can lead to the creation of an aquaculture pond identification method that does not rely on tedious surveys for information and can locate drastically different shapes and sizes of fishponds. However, what makes Bangladesh so different from previous studies is that there are many types of fishponds, all with different characteristics. ‘Commercial ponds’ (purpose- built ponds producing fish exclusively for sale) and ‘homestead ponds’ (small multi-use backyard 41 ponds), usually contain water year-round but can vary considerably in size. Ghers (modified diked rice paddies), only contain water during certain months of the year and can vary widely in terms of size and extent of integration with field crops. Ghers can be as small as 0.002 km 2 to as large as 0.4 km2, and tend to not have a standard shape, instead following the topography. In the remainder of the paper for simplicity, we refer to homestead ponds, commercial ponds, and ghers using the catchall term ‘fishponds’, except when emphasizing salient differences between these waterbody types. Among the studied methods, the method outlined in Yu et al. (2020) was deemed to be the best fit for this study because 1) similar to this study, their method was also developed and tested in Bangladesh, although their research was conducted in the Natore district in northern Bangladesh, 2) both studied regions fall in the same agroecological zone (Department of Agricultural Extension, 2021), 3) the logic behind their method is supported by open access code and can be easily implemented to other regions, and 4) inputs required for their method are free while using other methods rely on high- and very-high-resolution multispectral imagery can cost thousands of dollars. Therefore, in this study, we hypothesized that the Yu et al. (2020) method could be employed to achieve the research goal. Building on the Yu et al. (2020) method, first, we examined its performance within the study area and then proposes several improvement strategies to 1) enhance water detection specifically for small water bodies, 2) advance the edge detection for fishponds, and 3) achieve better fishpond classification. 42 3.2 Materials and Methods 3.2.1 Study Area The study area is comprised of seven districts in southwest and south-central Bangladesh, with a total area of 17,385 km2 (Figure 5). These are Bagerhat, Barisal, Bhola, Gopalganj, Jessore, Khulna, and Satkhira. These districts were chosen because they are home to some of the highest concentrations of aquaculture farms in the country, including nearly all of Bangladesh’s shrimp and prawn farms (Department of Fisheries, 2018). The districts are also in the USAID Feed the Future (FtF) Zone of Influence meaning they are being targeted to ensure long-term economic sustainability in farming (Feed the Future, 2021). For this research, we focus on aquaculture in enclosed waterbodies (ponds and ghers), and exclude farming taking place in stocked natural waterbodies, pens, and cages (Department of Fisheries, 2018). The study region excludes the Sundarbans National Park in the southern parts of Satkhira, Khulna, and Bagerhat. The land use for the study region was calculated from data from GlobeLand30 (China Ministry of Natural Resources, 2020). The study region is made up of 62% cultivated land (9966 km2), 23% artificial surfaces (3661 km2), 13% water bodies (2136 km2), and 2% wetlands (296 km2). Forest, shrubland, grassland, and bare land each encompass less than 1% of the total land use (Figure A1). 43 Figure 5. Study area is broken into seven individual districts. The seven districts are primarily made up of cultivated land, artificial surfaces, water bodies, and wetlands (Table A1). Gopalganj, Bagerhat, and Jessore have the highest percentages of cultivated land (all over 65%). Satkhira has the highest percentage of water bodies at 39% - even higher than their cultivated land. Bhola is one of the few districts with wetlands as a land cover and has the highest percentage of wetlands with 13%. The study area falls into three different climate subregions: the south-eastern zone, south- western zone, and south-central zone (Momtaz & Shameem, 2016). The south-eastern zone includes Bhola and the southern part of Barisal. This region sees high rainfall from May to September. Northern Barisal, Gopalganj, southern Khulna, and southern Satkhira fall within the south-central zone. This area has less rain on average than the south-eastern zone. The south- 44 western zone is comprised of northern Khulna, northern Satkhira, and Jessore and has the lowest levels of rainfall on average out of the three zones (Bangladesh Meteorological Department, 2021). The study area has four seasons throughout the year that are described as winter (December to February), pre-monsoon (March to May), monsoon (June to September), and post-monsoon (October and November) (Momtaz & Shameem, 2016). The heavy rainfall of the monsoon season makes it challenging to obtain could-free satellite images during that season. The districts within the study region see anywhere from 1200 mm to 2800 mm of rainfall during the monsoon season (Bangladesh Water Development Board, 2019). The south-eastern zone has a slightly longer rainy season with higher precipitation from May to September (Bangladesh Meteorological Department, 2021). 3.2.2 Overview of the Base Method As described in the Introduction section, the Yu et al. (2020) method was selected as a Base Method for detecting fishponds for the study area. In general, only a small number of changes (e.g., temporal and spatial extend of satellite imageries) to the original code were needed to fit the method for the study area. The Base Method comprises six main steps: data acquisition, water index calculation, initial mask, threshold optimization, majority vote, and fishpond identification (Figure 6). 45 Figure 6. Flowchart outlining the major steps of the Yu et al. (2020) for fishpond identification. 46 Step 1. Data Acquisition: The process starts with combining data from Sentinel-2 Level-1C (Top of Atmosphere Reflectance) from 1/1/2020 to 12/31/2020. The images are then filtered for cloud coverage, keeping only those images with less than 10% cloud pixels. We obtained 311 images for our study region during 2020 that include any image that covers any part of our study region. This is one of the major differences with the original Yu et al. (2020) study as they only obtained four images for their region of interest during the study period of 2016. In addition, the size of our study region is just under 33 times larger than the Yu et al. (2020) study area. During the 2020 study period, no images were available between mid-June and the end of September due to the monsoon occurrence. Step 2. Water Index Calculation: To identify where water is located within the study area, three different water indexes were used. These indexes analyze different wavelengths of reflected light to determine the presence of water. Using the images left after the cloud filter, we calculated the three water indexes: the Normalized Difference Water Index (NDWI) (S. K. McFeeters, 1995), the Modified NDWI (MNDWI) (Xu, 2006), and the Automated Water Extraction Index with no shadow (AWEI) (Feyisa et al., 2014). Step 3. Initial Mask: Previous studies showed that MNDWI performed the best among the three water indices (Ji et al., 2009, Zhou et al., 2017, Jiang et al., 2014) and also recommended by Yu et al. (2020) method; the mask is created from this index. The MNDWI is filtered by the ‘greater than’ function. If an MNDWI pixel value is greater than 0, the function returns a value of 1. The filtered MNDWI images are reduced into one image using the allNonZero reducer command in the Google Earth Engine (GEE). The reducer assigns the pixel values of 0 or 1 based on all values obtained from every image in the collection at a specific location. If every value at that pixel location was a non-zero, then the final image has a pixel value of 1 at that location. If there is even 47 one 0 value in any image at that location, then the final image has a pixel value of 0. This creates an initial binary mask that is converted to polygons. The polygons are enlarged with a buffer of 5 pixels to capture values of pixels that surround bodies of water. The polygons are then rasterized, creating a new image called MNDWI Mask (MM). The MM is then applied to the NDWI, original MNDWI, and AWEI image collections. Step 4. Threshold Optimization: This initiates a process called Otsu Segmentation which optimizes the threshold between foreground and background in images (water and not-water) (Otsu, 1979). The segmentation process creates a histogram of values from the locations the MM provided. The histogram should be bimodal with a buffer of 5 pixels, the two modes being water and non-water. The threshold between the two modes is found using the sum of squares. The Otsu Segmentation process occurs within each index’s image collection separately. Step 5. Majority Vote: The optimized NDWI, MNDWI, and AWEI image collections after Otsu Segmentation are then reduced into three images (one for each index) using the allNonZero reducer. This creates three separate images that show where each water index identified water in every image during the time period. The three reduced index images are then reduced to one image through the Mode/Majority Vote method. The mode method sees what value is present at that pixel location most often and assigns the final image that value. The final combined image depicts where water is in the majority of indexes and also where water is located in every image. This image is then converted into polygons and these polygons are given object-based features (OBFs). The OBFs include the Iso-Perimetric Quotient, Solidity, Patch Fractal Dimensions, Convexity, and Square Pixel Metric. The Convexity, Solidity, and Path Fractal Dimensions describe basic geometric characteristics of the pond and how complex it may be (Jiao et al., 2012). The Iso- Perimetric Quotient measures how similar an object is to compact shapes like circles (Q. Yu et al., 48 2006). Lastly, the Square Pixel Metric analysis for the shape convexity and is similar to the Iso- Perimetric Quotient (Frohn, 2007). Step 6. Fishponds Detection: Two machine learning techniques are utilized following the Yu et al. (2020) recommendations: Classification and Regression Trees (CART) and Logistic Regression (LR). These two algorithms were run in R Studio utilizing the caret package (Kuhn, 2008). The algorithms are trained by analyzing the calculated OBFs of the ground-truthing data. The base method determined algorithm performance by using five-fold cross-validation. The Base Method used images from the GEE platform to determine what defined a fishpond visually. The fishponds were then manually digitized for their study area to train their algorithms. In addition, non-fishpond water bodies were identified by removing digitized lakes from the Tibetan Plateau. 3.2.3 Proposed Improvements We identified Yu et al. (2020) method as having the most promising and applicable methods for our region of interest in southern Bangladesh, but with some modifications. From initial surveys, we found that many of the assumptions in the Yu et al. (2020) paper do not apply to the south-west and south-central regions of Bangladesh (Table 11). 49 Table 11. Base Method assumptions, shortcomings, and alternative approaches Base Method Shortcoming of the Base Method and Improvement assumptions alternative approach number Fishponds are filled with Some fishpond types in Bangladesh may dry 1, 3 water year-round out for a portion of the year (e.g., Homestead ponds). Other fishponds may be planted with rice for part of the year (e.g., Gher). Therefore, assuming all fishponds are filled with water all year is not correct. Instead of focusing on areas that have water for an entire year, we should focus on time periods when each type of fishpond holds water. Fishponds are Bangladesh's landscape is very diverse, and in 2, 4 surrounded by non- many areas, different types of land use can be water found around fishponds such as trees, rice paddies, buildings. In some regions, fishponds are very close with very narrow boundaries (< 10 m apart, which is the highest resolution of imageries used here). This assumption impacts the buffer size for the MM. Fishponds are easy to There is a high probability that small fishponds 5 detect visually from cannot be detected correctly through visual images. Non-fishponds observation of satellite imageries. The water bodies from preferred method is to use ground-truthing another region were fishpond collections from the study region that selected for algorithm are diverse in size, shape, type, and training. surrounding areas. CART and LR are the Support Vector Machine and Random Forest 6 preferred machine are identified as promising methods (Maxwell learning techniques for et al., 2018) for differentiating between this application fishponds and non-fishponds classes. In order to address these shortcomings, the following changes were proposed to improve the Base Method (1) identify a period of time that all types of fishponds are filled with water, (2) modify the buffer size to improve differentiating fishponds from their surroundings, (3) change the image reducer and water index combination to enhance waterbody detection, (4) introduce a convolution filter for edge detection, (5) utilize Support Vector Machine (SVM) and Random 50 Forest (RF) as two additional machine learning techniques, and (6) utilize ground-truthing data for better algorithm training. Improvement 1: As described earlier, the shortcoming of the Base Method in assuming year-round fishponds impoundment is not valid in our study region as many fishponds are dried for a certain part of the year. In addition, using many images for the entire year for a large study area, such as the one here, can significantly increase the computational time or prohibit the implementation of the Base Method for a large study area. In order to address this shortcoming, we identified a period of time in which all types of fishponds are filled with water. For this period, two sets of analyses were performed. First using all images with less than 10% cloud cover over the span of one month. Secondly, in order to further reduce the computational needs for data processing in our large study region, a single image for each district during the month of interest was selected for the second set of analyses. Using images for a single month, relevant satellite images were obtained for the period of October 11th to November 13th of 2020 instead of the full year. Therefore, the total number of images was reduced from 311 for the whole year observation to 43 for one month observation. For the second scenario, one single day during the month was selected. Since our study area is very large compared to the Base Method, each district has a different day that falls between October 11th and November 13th. The images chosen were based on visual inspection and clarity, such as cloud coverage and contrast between water and non-water. The days chosen for each district are October 13, 2020 for Barisal, October 28, 2020 for Bagerhat and Gopalganj, November 5, 2020 for Jessore and Satkhira, and November 7, 2020 for Bhola and Khulna. Improvement 2: The Base Method assumes that fishponds are surrounded by non-water pixels, but this is not the case in some of the districts in our study region. The Satkhira and Khulna districts, 51 for example, have fishponds and agricultural fields that are very close together with dikes and buffers that are smaller than 10 m. We tested buffer sizes of 0, 1, 3, and 5 pixels to identify the best size as the buffer in the Base Method is a very important step in creating the MM. The MM is a binary raster that is one of the inputs into the Otsu Segmentation function. Where the MM has a value of 1, the Otsu Segmentation function creates a histogram of index values at those locations. The buffer selects an additional region around the areas where the MNDWI has identified water that is assumed to be land. The histogram created by Otsu should be bimodal, where one peak represents water and the other represents non-water. Having a bimodal histogram makes the threshold values from Otsu more accurate. The potential issue with a buffer of 5 pixels is that it assumes that the water bodies identified by MNDWI during the MM process are surrounded by 50 m of land, trees, or buildings. This is not the case in many of the districts in our study region. In areas where aquaculture ponds are being used, they tend to be very close together, with only a dike separating them. Dikes are much smaller than 50 m; some are even smaller than the 10 m pixel that we can see. Fishponds surrounded by other water bodies could change the histogram from bimodal to unimodal since most pixels identified by the MM are water. We wanted to see if the buffer size impacts the histogram and threshold value from the Otsu Segmentation. If that threshold impacts how much water we identify and whether they have clear boundaries for a better classification. Improvement 3: The Base Method functions under the assumption that water is present in fishponds throughout the year and in every image. To implement this assumption, they utilize an all-or-nothing reducer called allNonZero and combine all three water index images. We know that some fishponds are drained for maintenance or may not be visible due to an imaging error (e.g., sensor issues, platform issues, or angle issues). With the all-or-nothing style of reducer, the Base 52 Method removes too many potential fishponds and under detects. Therefore, we tested an image reducer that would take the value of what is present in the majority of images instead of all. The allNonZero reducer in the Base Method labels final image pixels with a value of 0 (or non-water) if even one image does not have water at that location. This is a very strict reducer that eliminates too much water during the pre-processing and establishes a lower bound estimate. It does not take into account potential cloud interference or the periods of time when ponds may be dried or have rice within them. The allNonZero reducer is used two times in the Base Method. The first instance is in creating the MM when combining all the MNDWI images. The second instance is when combining the calculated index images into one summary image for each index (Step 5: Majority Vote). The first instance of the allNonZero reducer remains the same since we want very strict water locations to create the MM. As an improvement, we replaced the second instance of the allNonZero reducer with the Mode reducer. The Mode reducer, instead of requiring water to be present for every image requires that water be present for most of the images. The Mode reducer allows for more water pixels to be identified in the final classification. The Base Method combines images from all three water-identifying indexes together to create one final summary image. We assumed this muddles the shape and location of identified water bodies, decreasing the overall amount of detected water. To improve upon this, we tested four different index combinations to evaluate what performed best in each district. The combinations are the Base Method, AWEI individually, NDWI individually, and MNDWI individually. This test will show whether combining the three indexes negatively impacts the fishpond identification results or filters out likely non-fishpond water bodies. Improvement 4: Increasing the number of water pixels, whether from reducing the time frame, changing the buffer, or using the Mode reducer, will cause many water bodies to merge, forming 53 much larger polygons. This problem is exacerbated by the issue of dikes and other dividers commonly being less than 10 m wide, much smaller than the best resolution images that are used in this study. These large polygons have a much higher chance of being labeled as a non-fishpond by the classifier because of their size and odd shapes. To try to reduce the size of these polygons, we introduced a Laplacian 5×5 convolution filter to smooth the images (Gao et al., 2018). The convolution filter was applied to the best index or index combination image from Improvement 3 of each district and enhances the differences between water and non-water. The convolution filter is used in ArcGIS Pro 2.4. We chose Laplacian 5×5 as the filter type since it visually performed better than the other filter types (e.g., Smoothing 3×3, Smoothing 5×5, Laplacian 3×3). After the convolution filter is used, we put the smoothed index image through Otsu Segmentation with the MM mask. The process follows the same as the previously improved Base Method. This improvement aims to provide the Otsu Segmentation process with a clearer index image than the one provided by the Base Method. Improvement 5: The Base Method assumes that fishponds are easy to detect visually from standard RGB images, but that may depend more on when the images were captured and what surrounds the fishpond. Many fishponds in Bangladesh are hard to detect visually because they are surrounded by trees or many rice paddies and they have different fishpond layouts. In the study region, four types of fishponds were identified that include gher with and without rice, commercial ponds, and homestead ponds. All four types of fishponds vary in average size, typical shape, and period with water, all highlighted in Table 12. Gher without rice is, on average, much larger than all the other fishpond types. Homestead ponds and gher without rice can have round and irregular shapes, whereas gher with rice and commercial ponds typically are rectangular. The period in which each of these fishpond types is filled with water varies as well. 54 Commercial and homestead ponds are typically permanent ponds that have water year-round. Gher can have water or not because of their design to cultivate both fish and crops throughout the year. All of these differences between fishpond types make it very difficult to locate a diverse range of fishponds visually from satellite images. Table 12. Average size, shape, and period with water for gher with and without rice, commercial ponds, and homestead ponds. Characteristic Gher with rice Gher without Commercial Homestead rice pond pond Average size (m2) 1620-2020 4047 1620 810 Typical Shape Rectangular Rectangular, but Rectangular Round could have curved edges Period with water May - February - Year-round Year-round December December Predominant Freshwater Shrimp and fish Fish Fish farming system prawn, fish, rice, and vegetables The Base Method utilizes historical GEE images to visually find and trace fishponds. The traced fishponds are not directly used for training, but instead, they used the shape of the polygon of water identified at the traced locations through the Base Method for training. For their non- fishpond data, they manually trace water bodies from a Tibetan plateau for algorithm training using the Joint Research Centre (JRC) Yearly Water Classification dataset that has 30-m spatial resolution (Pekel et al., 2016). Tibet is north of Bangladesh and has a very different elevation and climate. In comparison, for this study, we used 1,728 ground-truthing polygons provided by surveyors in Bangladesh to validate the Base Method's performance and the proposed 55 improvements. The ground-truthing data consists of 996 fishponds and 741 non-fishpond waterbodies (Table A2 and Figure A2). Using ground-truthing data additionally makes validation results reliable instead of trusting that the waterbody chosen visually was a fishpond or not. The ground-truthing data was created by manually drawing borders around the known fishponds and non-fishponds in GEE. This results in the training dataset having smoother edges than what the water identification process will be. This is because the water identification process takes the shape of pixel groupings that appear blocky with hard edges. For this simplification of shape to not impact the results, we used the ‘Simplify Polygon’ function in ArcGIS to simplify the polygons identified through Steps 1-4. Improvement 6: The two machine learning algorithms used in the Base Method are CART and LR due to their wide use and applications in addition of being transcribed into GEE. In addition to these algorithms, SVM and RF were identified as promising methods and also recommended by Maxwell et al. (2018), considering many factors such as training data, requirements, and computational cost. SVM was selected because it is useful for finding an optimal boundary between two classes (Maxwell et al., 2018) and performs reliably when trained with a smaller dataset (Ramezan et al., 2019). RF is a larger combination of many decision trees and is easy to optimize, so it is reasonable to compare its results with the CART method (Shi & Yang, 2016). 3.2.4 Evaluation Criteria for Comparing Fishpond Detection Algorithms To compare the performance of the different fishpond detection algorithms, we used several criteria, including: 1) The number of ground-truthing fishponds that are correctly identified, 2) The percentage of the ground-truthing area that is correctly identified by the classifier 3) The number of fishponds classified, 56 4) The recall: Recall is the ratio of true positive fishponds to the total of all known fishponds, whether true positive or false negative (Equation (8)). 5) The precision: Precision is the ratio of true positive fishponds to all identified fishponds, whether true positive or false positive (Equation (9)). 6) The F1 score: The F1 score (Equation (10)) is the harmonic mean of the recall and precision and is used to provide a more insightful characteristic of performance than the arithmetic mean (Sasaki, 2007). The range for recall, precision, and F1 score is 0.0 to 1.0, with 1.0 being the highest score for all of them. TP (8) Recall = TP + FN TP (9) Precision = TP + FP Recall × Precision (10) F1 score = 2 × Recall + Precision To determine whether the proposed improvements significantly impacted the results, we compared the areas that are correctly identified between the Base Method and the proposed improvement method against the ground-truthing fishpond areas. The comparison was made using a confidence interval for the mean relative error between the identified area and the known ground- truthing area (Abramowitz & Stegun, 1972). 57 3.3 Results and Discussion 3.3.1 Base Method Performance Evaluation within the Study Area in Detecting Waterbodies and Fishponds The performance of the Base Method was evaluated based on 1) how well it identified water bodies and 2) how well it classified those water bodies as fishponds. Figure 3 shows the percentage of the ground-truthing area that are correctly identified at pre-classifier (waterbody identification) and post-classifier (fishpond identification) stages. First, concerning detecting water areas within the known fishponds, the highest detection was in Bagerhat and had 9.4% of ground-truthing fishpond area overlap. The second highest was Gopalganj with 5.9%, but the rest of the districts saw less than 1% ground-truth fishpond area detection. Second, regarding the applicability of classifiers to differentiate between fishponds and non-fishponds, the LR method was identified more ground-truth fishponds correctly than CART. The performance post-classifier will always be the same or less than the waterbody identification since the classifiers are only applied to the areas identified as water. Bagerhat had the highest correct classification out of the seven districts. Most of the other districts had correct classifications under 1%, except Gopalganj had around 5% classification after the LR classifier (Table A3). The CART classification true positives were lower than the LR, with the highest being Bagerhat’s 5% ground-truth area identification. In Bagerhat, the Base Method using the LR classifier correctly identified 16 ground- truthing fishponds out of 235 (9% of the total ground-truthing fishpond area) (Figure 7). Having said that, still the performance of the Base Method, even in the Bagerhat, is low. 58 10% 9% Precentage of the correctly idefnitifed 8% 7% 6% 5% waterbodies/fishponds 4% 3% 2% 1% 0% Bagerhat Barisal Bhola Gopalganj Jessore Khulna Satkhira District Pre-Classifier Logistic Regression Classification and Regression Trees Figure 7. Percentage of the ground-truthing area that were correctly identified as waterbodies, classified as fishpond using logistic regression method, and classified using the classification and regression trees method for different districts within the study area. Meanwhile, the Bagerhat district is the only one of the seven in which the results are significant enough to visualize. Therefore, in Figure 8, we visually compared the performance of the classifiers that were used in the Base Method. While CART correctly identifies fewer ground- truthing fishponds (Figure 8C), LR over-classifies water bodies as fishponds (Figure 8B). Also notable in Figure 8 are the large polygons that are identified as water and how those polygons are classified by LR and CART. LR tends to classify the large polygons as water, whereas CART is more conservative and classifies them as non-fishponds. The polygons are much larger than the ground-truthing fishponds in the area, and in some cases, overlap the ground-truthing fishponds. These large polygons make individual fishpond detection difficult since they typically encompass multiple fishponds or several agricultural lands (e.g., rice paddies). 59 Figure 8. Example of the Base Method results in Bagerhat. Orange polygons represent ground- truthing fishponds. Blue polygons are classified as fishponds. Red polygons are classified as non-fishponds. (A) Ground-truthing fishponds. (B) Logistic Regression classification results. (C) Classification and Regression Trees classification results. (Centroid of the pictures above: 22° 35' 59.4096" N, 89° 34' 57.414" E) The threshold optimization (Step 4) was created to pinpoint the water bodies most likely to be fishponds. For the small study area that the Base Method was applied, this technique works very well, especially since the layout of their study area has clear divided water bodies. However, even broken into the seven districts in our study area, the pixel selection technique does not perform to the same standard. The underlying assumptions of the Base Method are not valid for our study area and must be adjusted to make accurate fishpond identification possible. Overall, the Base Method is under-identifying water bodies before the classifier is even applied. In addition, the Base Method's performance in detecting fishponds is very low, the highest of all districts being 9%, and can be enhanced significantly. In the following sections, we will evaluate the effectiveness of proposed improvements on waterbody and fishpond detections in a very large and diverse study area. 60 3.3.2 Identifying the Best Period of Image Collection for Detecting Fishponds (Improvement 1) We tested one year, one month, and one day time periods for image collection to evaluate how the number of images impacts the threshold optimization results. The one year period combined information for 311 images while one month lowered the number of images to 43. These totals are for the study area as a whole and not for each specific district. Table 13 compares the waterbody identification results from each time period. Rows 1, 3, and 5 highlight the ground- truth fishpond area percentage identified and how it improves with the shorter time for every district. The district of Bhola performed the worst out of all districts with 0% water identification for both year and month periods and 3% for the day. The poor performance in Bhola could be explained by the low numbers of ground-truthing data (27 locations) we have for the district. Generally, the fewer ground-truthing fishponds we have, the harder it is to identify any of them specifically. Bagerhat had the best water identification during the one-month period with 52% of the ground-truthing fishpond area identified, but Jessore had the best performance in a single day with 91% area identification. The districts of Barisal, Gopalganj, Khulna, and Satkhira all saw significant improvement over the Base Method, but performed worse than the Bagerhat and Jessore values (Table A4-Table A10). All seven districts improved water identification when we changed the time period from a year to either a month or a day. The single-day images had the highest results out of the three periods. 61 Table 13. The percentage of ground-truthing fishponds area and total ground-truthing fishponds correctly identified in all districts using different lengths for image processing (i.e., the entire year, one month, one day). District Period of Performance image criteria collection Bagerhat Barisal Bhola Gopalganj Jessore Khulna Satkhira One year Percentage of 9.4% 0.2% 0% 6% 1% 0.01% 0% GT* fishpond area identified pre- classifier GT* 25 of 235 1 of 168 0 of 7 of 77 13 of 1 of 163 0 of 208 fishponds 27 113 identified One Percentage of 52% 4.3% 0% 22% 44% 36% 66% month GT* fishpond area identified pre- classifier GT* 139 of 8 of 168 0 of 44 of 77 70 of 71 of 163 75 of fishponds 235 27 113 208 identified One Day Percentage of 67% 16.3% 3% 36% 91% 67% 85% GT* fishpond area identified pre- classifier GT* 182 of 28 of 168 1 of 66 of 77 99 of 112 of 115 of fishponds 235 27 113 163 208 identified *Ground truthing Changing the period from the full 2020 year to just October 11 to November 13, 2020 proved to be effective in identifying more water than the Base Method. The Base Method utilizes the allNonZero reducer in GEE to identify water that is present year-round. A longer time period means more images to compare and more chances for water to not be present. All districts saw an improvement in water detection from the Base Method and from the month period when using an image for one day. Figure 9 shows the drastic difference in water detection between the three periods, but also how small boundaries between ponds and agriculture can blend together to form large polygons. Figure 9 also compares a location in which we had good water detection, Jessore, to an area with very poor water detection, Bhola. The method performs much better when using a 62 shorter time frame with fewer images. The best period proved to be the single day. We do not expect that the day chosen will drastically impact results, but the image must have low cloud coverage (e.g., less than 10%) and it must be during a time period when we are fully confident fishponds contain water. Because of these guidelines and Bangladesh’s climate, we are only able to obtain a few images that fit these standards. To ensure that the best time period is a day, we used both single day and single month for Improvements #2 and #3. 63 Figure 9. Visual comparison of the three-time periods in southern Jessore and southern Bhola. (A) Jessore ground-truthing fishponds (orange). (B) Jessore ground-truthing fishponds and full 64 Figure 9 (cont’d) year water (red). (C) Jessore ground-truthing fishponds and one-month water (purple). (D) Jessore ground-truthing fishponds and single-day water (blue). (Centroid: 22° 57' 49.9248" N, 89° 16' 46.3188" E). (E) Bhola ground-truthing fishponds (orange). (F) Bhola ground-truthing Fishponds and full year water (red). (G) Bhola ground-truthing Fishponds and one-month water (purple). (G) Bhola ground-truthing Fishponds and single-day water (blue). (Centroid: 22° 10' 7.0464" N, 90° 41' 27.546" E) 3.3.3 Testing the Buffer Size for Threshold Optimization (Improvement 2) The buffer size impacts the amount of water identified by the Threshold Optimization process. The larger the buffer, the more pixel values are included in determining the threshold value between water and non-water. Having the best possible threshold will significantly improve upon water and fishpond identification. We tested buffer sizes of 5, 3, 1, and 0 pixels within the MM for both single day and one month time periods. The water identification pre-classifier for Bagerhat with a buffer size of 5 was 67% for one day, but that dropped to 58% with a buffer size of 0 (Table A11). This was the same trend across all districts for both time frames except for Bhola (Table A11-Table A17). None of the attempted methods have resulted in any success in Bhola higher than the 3% of ground-truthing area identified pre-classifier achieved from the 5-pixel buffer for a single day. The results showed that the buffer of five pixels performed better than any other size buffer for both time periods. Figure 10 shows the histogram of NDWI values at the MM location for Khulna. The final results for each district are from using the Base Method, in which all three indices are combined. Khulna was chosen for the example since it had a noticeable change from the buffer size and NDWI is one of the most used water indices. Khulna saw a large decrease in ground-truth water identification from 67% to 59% when changing the 5-pixel buffer to 0-pixel for a single day. Khulna saw a similar decrease from 36% to 27% for 5-pixel to 0-pixel buffer for the month period. The histograms for 0- and 1-pixel buffers have large peaks between -0.2 and -0.1 that represent non-water pixels. As the buffer size increases, the second peak around 0.3 to 0.4 increases 65 considerably. With NDWI, it is commonly interpreted that values above 0.3 are water bodies and values between 0 and 0.3 may be water bodies (Stuart K. McFeeters, 2013). These histograms show that with the increased buffer size, water pixels become more prominent. The 5-pixel buffer may work best because it includes more pixels in total. The more pixels included in the histogram, the more reflective the histogram is to the land use in the district. Additionally, the 5-pixel buffer works better in districts with small fishpond boundaries because it provides more opportunities for endmember pixels, or pure water pixels, to be accounted for. Figure 10. NDWI reflectance values for MM in Khulna for a single-day image. Top left to bottom right: 0-pixel buffer, 1-pixel buffer, 3-pixel buffer, 5-pixel buffer 66 3.3.4 Determining the Combination of Image Reducer and Water-Identifying Index to Improve Waterbody and Fishpond Detection (Improvement 3) Both the image reducer and water-identifying indices impact the amount of water identified for classifying. The Base Method resulted in too few of the ground-truthing fishponds being identified. Therefore, we hypothesized that determining the best combination of image reducer and indices will significantly improve the overall detections. To increase the amount of water identified, we tested two image reducers (i.e., Modeand allNonZero) and four different water- identifying index combinations (i.e., AWEI, MNDWI, NDWI, and combination of AWEI; MNDWI; and NDWI). The Mode reducer increased the area percentage of ground-fishponds identified for all districts (Table A18-Table A24) compared to the allNonZero reducer (Table A25- Table A31). The exception to this is Bhola, in which most of the results were 0% except for the NDWI index in which the Mode reducer identified 7%. Figure 11 shows the comparison of the combined indexes, AWEI, MNDWI, and NDWI in correctly identifying the ground-truth fishpond areas for each district. The combination of all three indexes that represents the Base Method did not perform the best for any district. Figure 11 also shows that the MNDWI by itself also does not perform the best for any district (Table A32-Table A38). 67 100% GT Fishpond Area Identified Pre-Classifier 90% 80% 70% 60% Combined 50% AWEI 40% MNDWI 30% NDWI 20% 10% 0% Bagerhat Barisal Bhola Gopalganj Jessore Khulna Satkhira District Figure 11. Ground-truthing (GT) fishpond area identified pre-classifier per district for the combined index, only AWEI, only MNDWI, and only NDWI results using the single day images Testing the image reducer type proved again that the single day time period was one of the best improvements. Testing the Mode reducer on the month-long period resulted in less water identification than the single day (Table A18-Table A24), but more than the same period using the allNonZero reducer (Table A25-Table A31). Meanwhile, the Mode reducer did improve water identification, but since single-day images only have utilized one image, the reducer has no impact on its performance. 3.3.5 Implementing Edge Detection with a Convolution Filter to Improve Fishpond Boundary Detection (Improvement 4) Considering the results from Improvement 1 through 3, the best combination of time of image collection, buffer size for the threshold optimization, and water index for each district was identified. Two major combinations are as follows: (1) single day, 5-pixel buffer, with AWEI index for Bagerhat, Gopalganj, Jessore, and Khulna and (2) single day, 5-pixel buffer, with NDWI index for Barisal, Bhola, and Satkhira (Table A39). Since the best combinations for every district entail 68 using the single day image, the reducer type is no longer relevant. A potential reason why MNDWI does not perform to the same standards as NDWI and AWEI is that MNDWI was created to mainly differentiate water from developed land (Wicaksono & Wicaksono, 2019), which is not the case within our study area. In fact, the majority of the representative fishponds (ground-truth) are not located near significant developed land. Our preliminary literature review also supported this assumption that is most likely the case with the majority of fishponds in the country. In the next step, we apply a Laplacian 5×5 convolution filter for edge detection to identify the best image combination for each district. The purpose of introducing edge detection is to begin with an enhanced image that highlights sharp differences in pixel values. The difference in pixel values shows where boundaries are located around different plots of land. The convolution filter emphasized the differences in the select index images to identify the boundaries around water. The spectrum of index values widens after the filter has been applied, allowing for differences to become more apparent numerically. For example, the original range of NDWI values in Barisal is -0.50 to 0.65. After using the filter, the range lengthens from -14.83 to 8.95. Visually, the filter highlights boundaries but does make many of the pixels become a very similar grey color (Figure 12B). 69 Figure 12. NDWI image in western Barisal (A) before and (B) after the Laplacian 5×5 convolution filter was applied (Centroid: 22° 51' 38.37" N, 90° 5' 53.28" E) The results of applying the Laplacian 5×5 convolution filter were mixed; however, every district saw a large increase in the overall number of fishponds identified, with some now seeing over 100,000 fishponds (Table A40). Bagerhat, Bhola, Gopalganj, Jessore, and Khulna had all their ground-truthing fishponds identified before classification. Jessore performed the best with edge detection seeing all 113 ground-truthing fishponds and 91.5% of the area correctly identified before any classifier was used. Despite Jessore’s high performance, the water body identification pre-classifier decreased with adding edge-detection from 95% to 91.5%. Satkhira experienced a similar result from adding edge-detection with the ground-truth area identified decreasing from 91% to 82%. Oppositely, Gopalganj, Barisal, and Bhola all increased their ground-truthing fishpond area identification at the pre-classifier stage from 54% to 68%, 26% to 58%, and 20% to 57%, respectively. The same district divide was seen with the post-classifier results as well. Barisal and Bhola saw an increase in their post-classifier area identification for both LR and CART. The ground-truthing fishpond area percentage for Jessore decreased from 94.3% to 91.4% with LR and from 3% to 0% with CART using edge-detection. Satkhira saw a slight increase from 0.3% to 5% for the CART classifier, but its LR dropped from 93% to 56% ground-truth area identification. 70 3.3.6 Evaluating the Impact of Ground-Truthing Data on Machine Learning Training (Improvement 5) To train machine learning classifiers, the Base Method manually traced suspected fishponds from historical satellite images. The water bodies they traced may have only been representative of one type of fishpond because homestead ponds and gher with rice are very difficult to identify as fishponds. To improve the machine learning classifier training, we used detailed ground-truthing data for all four types of ponds described previously. In addition to this, we had ground-truthing data for non-fishpond water bodies that would provide juxtaposing training data. Table 14 compares the performance of the machine learning classifiers when they are trained with ground-truthing data versus historical imagery that was used in the Yu et al. (2020). The recall significantly increased from 0.538 to 0.989 for LR and from 0.495 to 0.827 for CART. This shows that the ground-truthing data is decreasing the number of false negatives. The precision score decreased from 0.788 for LR and 0.773 for CART to 0.594 and 0.738, respectively, when using the ground-truthing data for machine learning training. A decrease in precision implies an increase in the number of false positives. The increases in recall and F1 score show that utilizing ground-truthing data improves the performance of both the LR and CART classifiers, but the decrease in precision increases the number of false positives compared to the Base Method. 71 Table 14. Comparing statistics from Yu et al. (2020) machine learning training with historical imagery to our ground-truthing data training. Training method Performance Logistic Regression Classification and Criteria Regression Trees Yu et al. (2020) Precision 0.788 0.773 Historical imagery Recall 0.538 0.495 F1 Score 0.640 0.604 Ground-truthing data Precision 0.594 0.738 Recall 0.898 0.827 F1 Score 0.715 0.780 3.3.7 Adding Random Forest and Support Vector Machine Classifiers to Determine the Best Classifier for the Data (Improvement 6) In addition to LR and CART, RF and SVM are widely recognized classifiers and can be used for many different research applications that require the differentiation of two or more classes (Agmalaro et al., 2021), Yan et al., 2021, Iordache et al., 2020). To improve the classification results, we added RF and SVM since they are identified as promising classifiers (Maxwell et al., 2018). Overall, the LR and SVM classifiers identified more ground-truthing fishponds both in terms of area percentage and total number than the CART and RF (Table A41 and Table A42). Figure 13 shows how the results of the RF and CART are similar to each other as well as how similar the results from LR and SVM are. The district with the best performance is Jessore, with both LR and SVM classifying 91% of the ground-truthing fishpond area and all 113 ground- truthing fishponds correctly. Jessore is also where there is the most drastic difference in performance by classifiers was observed. This could be due to the predominant presence of large water polygons in Jessore. LR and SVM tend to classify the large polygons are fishponds, whereas RF and CART do not. The RF and CART classifiers in Jessore performed the worst out of any district with 3% ground-truthing fishpond area identification and 3 of 113 fishponds correctly identified. Introducing two additional classifier types yielded similar results to the Base Method. 72 100% Ground-truthing Fishpond Area Identified Post- 90% 80% 70% 60% 50% Classifier 40% 30% 20% 10% 0% Bagerhat Barisal Bhola Gopalganj Jessore Khulna Satkhira Districts Logistic Regression Classification and Regression Trees Random Forest Support Vector Machine Figure 13. Ground-truthing fishpond area percentage identified by each classifier type. Table 15 compares the precision, recall, and F1 scores for each of the four classifier types we used. The F1 scores are all very similar to each other, with the range from lowest value (0.706) to highest (0.780) being only 0.074. The LR recall and precision results are interesting because it has the worst precision, but the best recall meaning the LR is very good at preventing false negatives, but the worst at preventing false positives. RF has the exact opposite results with the highest precision and the lowest recall. Both LR and RF have lower F1 scores. CART and SVM fall between LR and RF for precision and recall, but because of their consistency, they have the two higher F1 scores. From these scores, we can conclude that LR is over-classifying fishponds and RF is under-classifying fishponds. When looking at the results in Figure 13, we can conclude that since SVM performed very similarly to LR, it must also be over-classifying. In addition, since RF typically identifies less than CART and RF under-classifies, CART can be the best final option. 73 Table 15. Comparing the training performance of the four classifiers on the validation ground- truthing data. Performance Logistic Classification and Random Support Vector Criteria Regression Regression Trees Forest Machine Precision 0.594 0.738 0.784 0.720 Recall 0.898 0.827 0.645 0.822 F1 Score 0.715 0.780 0.706 0.768 The total number of classified fishponds for each district also supports the conclusion that CART is the best middle ground (Table A41 and Table A42). Figure 14 shows the classifier results from LR, CART, RF, and SVM in eastern Gopalganj as compared to the ground-truth fishponds in that location. Since the data shows that LR is over-classifying, we can conclude SVM must also be since SVM classifies 3 to 5 times as many fishponds as LR does. This conclusion is also backed by the fact that LR and SVM are very similar classifiers, so it is realistic that they would behave similarly (Zhang et al., 2003). The CART classifier numbers always fall below LR and above RF levels. This could also support our previous finding that CART provides the best classification out of the four. 74 Figure 14. Comparison example of classified fishponds in eastern Gopalganj using (A) Logistic Regression, (B) Classification and Regression Trees, (C) Random Forest, and (D) Support Vector Machine. Ground truthing fishponds are in orange and classified fishponds are in blue. (Centroid: 23° 0' 17.9136'' N, 90° 3' 46.8936'' E). 3.3.8 The Overall Enhancement in Fishpond Detection as Results of the Proposed Improvements based on Medium-Resolution and High-Resolution Imagery We calculated the mean relative error for four different scenarios to determine whether the six proposed improvements considerably enhanced the capability of the Base Method for detecting fishponds. The scenarios that are considered here are the Base Method before classification, the Enhanced Method (considering all improvements) before classification, the Base Method using LR, and the Enhanced Method using CART. Table 16 outlines these statistics. The mean relative difference closest to 0 is the Enhanced Method without classifier (lower and upper limits of -34.4% 75 and -31.3%). This means that the Enhanced Method is regularly 31-34% under detecting the ground-truthing area. The other three scenarios had mean relative differences very close to -100%. The Enhanced Method CART mean relative difference was better than for the Base Method pre- and post-LR. This shows that the Enhanced Method without classification considerably improves upon the Base Method water identification. Table 16. The lower and upper limits for the mean relative difference of the Based and Enhanced Methods performances before and after a machine learning classifier. Mean relative difference Scenarios Lower limit Upper limit Base Method before -99.4% -98.5% classifier Enhanced Method before -34.4% -31.3% classifier Base Method with Logistic -99.7% -97.7% Regression Enhanced Method with -91.5% -88.6% Classification and Regression Trees Having access to high-resolution imagery can improve the results of the Enhanced Method significantly. A WorldView-2 image was obtained for an area of 45 km 2 in Jamalnagar, Satkhira to analyze the performance of the Enhanced Method. The high-resolution image was used to identify the boundaries around all land plots and intersected those boundaries with the final Enhanced Method for Satkhira. Figure 15 compares the boundary identification from the WorldView-2 image, the Base Method water identification, the Enhanced Method water identification, and the intersect of the WorldView-2 boundaries and the Enhanced Method. We applied the CART classification to the intersect of WorldView-2 and the Enhanced Method to compare it to the Enhanced Method fishpond identification alone. There are 47 ground-truth 76 fishponds in this region. The Enhanced Method correctly identified 9 of the 47 whereas the intersect correctly identified 29 of the 47 fishponds. In addition, the area of fishponds correctly identified was doubled with the introduction of the high-resolution boundaries. The improvement in fishpond and area identification in one small region shows that having access to high-resolution imagery would refine the results of the Enhanced Method even further. In summary, using high- resolution imagery would significantly improve boundary detection around fishponds and provide more accurate water locations. ¯ Figure 15. Impacts of high-resolution imagery in Jamalnagar, Satkhira. (A) High-resolution boundary identification, (B) Base Method water identification, (C) Enhanced Method water 77 Figure 15 (cont’d) identification, and (D) Intersect of Enhanced Method and high-resolution boundaries. (Centroid: 22° 34' 57.6624'' N, 89° 12' 12.1536'' E) 3.3.9 Land Use and Ground Truth Fishpond Characteristics to Explain Trends in Results All seven districts within the study region are very different from each other. This makes creating one over-arching method very difficult. We looked at the land use surrounding the ground- truth fishponds to see if there was a trend in districts that performed better/worse than others. To do this, we tested buffer sizes of 10 m, 50 m, and 100 m around ground-truth fishponds in each district separately (Table A43-Table A45). The predominant land use types in all districts were cultivated land, water bodies, and artificial surfaces. Bhola and Barisal had the lowest percentage of ground-truthing fishpond area identified before the machine learning classifier. These two districts had the highest percentage of artificial surfaces, with Bhola seeing 73% and Barisal having just under 50% with the 100 m buffer. All other districts had artificial surfaces as 26% or less of their land use at 100 m. Water and artificial surfaces may have very similar reflectance values (Worden & de Beurs, 2020). In this instance, since there was a very high percentage of artificial surfaces and a very low percentage of water, the indexes may be identifying artificial surfaces instead of water during Otsu Segmentation. The median size ground-truthing fishpond also varied significantly by district (Table A46). The district with the smallest median size ground-truthing fishpond was Bhola (713 m 2), followed by Barisal (833 m2), then Satkhira (1252 m2). All other districts had median ground-truthing fishponds larger than 2,200 m2. The combination of having the least and smallest ground-truthing fishponds may be why the improved methods performed the worst in Bhola. Barisal had more 78 ground truth fishponds with 168, but they were additionally much smaller than other districts. This could explain why the Enhanced Method performed worse in Barisal than other districts. 3.4 Conclusion Aquaculture is becoming increasingly more important in regions that relies heavily on fish for food, but capture fisheries are producing less and less. Identifying where aquaculture is occurring is critical in understanding the industry, how it operates, and the areas it can improve in. With this information, we can estimate total production, inform policy, or analyze trends in aquaculture locations. This study showed that the approach taken for identifying fishponds must be tailored to the characteristics of the fishponds, the fishponds’ surrounding areas, and the satellite data being used. The best approach for this is to test the time period for image capture, the buffer size for image optimization with Otsu Segmentation, the combination of water-identifying indexes, the image reducer type, and the machine learning classifier type. Utilizing edge-detection techniques and high-resolution imageries can improve the overall waterbody detection. For south-west and south- central Bangladesh, limiting the number of images for a specific time period results in an increased number of waterbodies detected. One reason for this is that additional errors can be introduced when combining multiple images together if an image does not capture the period with water. The buffer size for threshold optimization of water and non-water had an enormous impact on the results. The 5-pixel buffer performed the best for our study region. This proves that including more pixels in image thresholding results in better thresholds altogether, especially when the shape and size of the fishponds change throughout the year. Of the three water indices, NDWI and AWEI performed the best for the region, but this varies on location and land use/cover characteristics. The Mode reducer did allow for more waterbodies to be identified than the allNonZero image 79 reducer when working with more than one image for the study region. The image reducer type, in the end, did not matter for the Enhanced Method since one image is usually selected for fishpond detection. The training dataset for machine learning plays an important role in identifying the best classifier. Having diverse and representative ground-truthing fishponds will improve the machine learning classifiers. Some of the challenges that were observed in this study include 1) size of the study area: due to the size of the study area and the limitation of the GEE platform computational power, many of the tasks proposed in this study could not be executed for the entire region. One way to address this issue was to divide the study area into seven districts and apply various tasks separately for each district; 2) threshold optimization: in the original Base Method, one threshold value for water detection was obtained for the study region; however, our results showed that this approach does not work in a large region. Therefore, in this study, we optimized the threshold values based on the water index values at the district level. One alternative is to train the system based on the region's physiographical and climatological characteristics instead of its political boundaries. Using agroclimotological zone is one alternative but is not a good representation of the fishpond distribution and attributes; and 3) high-resolution imagery: Our study showed that using high- resolution imagery can significantly improve the overall performance of the Enhanced Method in detecting fishponds. Having high-resolution imagery would be very beneficial in identifying the boundaries between fishponds and breaking up any large polygons. As the cost for high-resolution imagery decreases, these six strategies will become increasingly more useful and 4) rainy season: The rainy and monsoon season in Bangladesh is from June through mid-October. This limits the availability of high-quality images with low cloud coverage. In fact, our study showed that the viable images for this study are mainly available from January to May and October to December. 80 One solution can be the combined usage of SAR and optical imaging in identifying small waterbodies. Ultimately, verification of the Enhanced Method in other regions is crucial to examine the reliability of the proposed method. 81 4 OVERALL CONCLUSION Aquaculture in Bangladesh is growing very rapidly and plays an important role in addressing agriculture and food security. Therefore, mapping and classifying fishponds is essential for economic purposes and understanding how land use is changing. Better statistics on aquaculture come from two different methods: surveys or remote sensing. Surveys are time- consuming and costly to be a viable option for the large-scale study. Meanwhile, utilizing remote sensing data can be a good alternative as it can cover a large area and can be implemented in a platform such as GEE that is available for free. However, the products from remote sensing do not have the same accuracy as the survey and should be calibrated based on ground-truthing data. The research found that the process used to identify fishponds must be specific to the region where it is applied, as the complexity of the region does not allow for the development of the universal fishpond detection algorithm. Therefore, we developed six strategies for identifying fishponds that can build off each other and can be applied broadly. The findings from each of these strategies are presented below: (1) When determining the best time-period for image identification, it was concluded that the fewer images analyzed, the higher the fishpond identification was. This relies on knowing when the typical fishpond has water and the climate of the region to determine the best single image to use. (2) To optimize the threshold between water and non-water, using a larger buffer size around waterbodies yielded higher fishpond identification. (3) Regardless of the region of study, several water-identifying indexes should be tested to determine whether a single index or combination of indices should be selected to detect existing waterbodies. 82 (4) In general, detecting fishpond shapes in Bangladesh can be challenging as the boundary widths around fishponds are generally smaller than the resolution of the optical imagery. Therefore, utilizing edge-detection techniques can increase the number of waterbodies identified and can be used to differentiate between land plots. (5) Performing a limited but targeted ground-truthing survey can help the overall performance of machine learning techniques. (6) Due to spatial and temporal variabilities of fishpond types and shapes in Bangladesh, not a single machine learning technique could be identified as the best. Therefore, it is recommended to test different machine learning classifiers to ensure the selection of the most robust technique for different regions (e.g., district). 83 5 FUTURE RESEARCH RECOMMENDATIONS This research provided six strategies to improve fishpond identification in the south-west and south-central Bangladesh. However, due to the limitations of using medium-resolution imagery, the results cannot be further improved as the boundary of the fishponds in the region is smaller, in many cases, than the resolution of the satellite imagery. Therefore, additional research should be mainly focused on fishpond boundary detection. Below, more specific recommendations are provided for future studies:  Apply the six improvement strategies in different Southeast Asian countries with similar fishpond prevalence as this study to evaluate the overall robustness of the proposed strategies.  Adjust the proposed improvement strategies with high-resolution imagery to better detect fishpond shapes and characteristics.  Explore the applicability of other satellite imagery and datasets as inputs to the proposed fishpond detection strategies. Among the existing imageries, SAR seems to be promising as it has been used for detecting waterbodies. In addition, the problem with cloudiness is not impacting the quality of the imagery as radar sensors can see through clouds.  Improve the threshold optimization through better identification of fishpond zones (e.g., land use/land cover) within a study area. As the utilized boundary for this study (i.e., district) does not capture the spatial variabilities of fishpond settings. 84 APPENDIX 85 Figure A1. Land use map of the study region (adapted from region (China Ministry of Natural Resources, 2020)) 86 Figure A2. Distribution of Ground Truthing Data locations by type. 87 Table A1. Land cover as percentage of area for each district within the study region (China Ministry of Natural Resources, 2020) Percentage of area within each District Land Cover Bagerhat Barisal Bhola Gopalganj Jessore Khulna Satkhira Cultivated Land 69.9 60.7 52.5 84.1 67.4 56.0 38.2 Forest 0 0.1 2.0 0 0 0 0 Grass Land 0 0.1 1.0 0 0 0 0 Shrub Land 0 0 0.3 0 0 0 0 Wetland 0.4 0.7 13.3 0.1 0.2 0.4 0.1 Water Body 6.3 9.2 7.0 2.2 3.0 22.6 39.0 Artificial 23.4 29.1 23.8 13.6 29.4 21.0 22.7 Surfaces Bareland 0 0.1 0.1 0 0 0 0 88 Table A2. Numbers of ground-truthing data fishponds and non-fishponds by district. District Number of fishponds Number of non-fishponds Bagerhat 235 144 Barisal 168 32 Bhola 27 2 Gopalganj 77 131 Jessore 113 111 Khulna 163 146 Satkhira 208 175 89 Table A3. Base method comparison for all seven districts. District Bagerhat Barisal Bhola Gopalganj Jessore Khulna Satkhira Classifier Type LR* CART** LR* CART** LR* CART** LR* CART** LR* CART** LR* CART** LR* CART** Percentage of GT*** fishpond area identified 9% 9% 0.2% 0.2% 0% 0% 6% 6% 0.8% 0.8% 0% 0% 0% 0% pre-classifier GT*** fishponds identified pre- 25 of 13 of 25 of 235 1 of 168 1 of 168 0 of 27 0 of 27 7 of 77 7 of 77 13 of 113 1 of 163 1 of 163 0 of 208 0 of 208 classification 235 113 Percentage of GT*** fishpond area identified post-classifier 9% 5% 0% 0% 0% 0% 6% 5% 0.5% 0.2% 0% 0% 0% 0% GT*** fishponds identified post- 16 of 235 5 of 235 0 of 168 0 of 168 0 of 27 0 of 27 3 of 77 2 of 77 4 of 113 2 of 113 0 of 163 0 of 163 0 of 208 0 of 208 classification * Logistic Regression ** Classification and Regression Trees *** Ground truthing 90 Table A4. Time period comparison for Bagerhat. Time Frame Year Month Day Classifier Type LR * CART ** LR * CART ** LR * CART** Buffer Size 5 5 5 5 5 5 Percentage of GT*** fishpond area identified 9% 9% 52% 52% 67% 67% pre-classifier GT*** fishponds identified pre- 25 of 235 25 of 235 139 of 235 139 of 235 182 of 235 182 of 235 classification Percentage of GT*** fishpond area identified post-classifier 9% 5% 51% 19% 66% 16% GT*** fishponds identified post- 16 of 235 5 of 235 117 of 235 49 of 235 159 of 235 51 of 235 classification * Logistic Regression ** Classification and Regression Trees *** Ground truthing 91 Table A5. Time period comparison for Barisal. Time Frame Year Month Day Classifier Type LR * CART ** LR * CART ** LR * CART** Buffer Size 5 5 5 5 5 5 Percentage of GT*** fishpond area identified 0.2% 0.2% 4% 4% 16% 16% pre-classifier GT*** fishponds identified pre- 1 of 168 1 of 168 8 of 168 8 of 168 28 of 168 28 of 168 classification Percentage of GT*** fishpond area identified 0% 0% 4% 4% 13% 12% post-classifier GT*** fishponds identified post- 0 of 168 0 of 168 3 of 168 4 of 168 11 of 168 10 of 168 classification * Logistic Regression ** Classification and Regression Trees *** Ground truthing 92 Table A6. Time period comparison for Bhola. Time Frame Year Month Day Classifier Type LR * CART ** LR * CART ** LR * CART** Buffer Size 5 5 5 5 5 5 Percentage of GT*** fishpond area identified 0% 0% 0% 0% 3% 3% pre-classifier GT*** fishponds identified pre- classification 0 of 27 0 of 27 0 of 27 0 of 27 1 of 27 1 of 27 Percentage of GT*** fishpond area identified post-classifier 0% 0% 0% 0% 3% 3% GT*** fishponds identified post- 0 of 27 0 of 27 0 of 27 0 of 27 1 of 27 1 of 27 classification * Logistic Regression ** Classification and Regression Trees *** Ground truthing 93 Table A7. Time period comparison for Gopaglanj. Time Frame Year Month Day Classifier Type LR * CART ** LR * CART ** LR * CART** Buffer Size 5 5 5 5 5 5 Percentage of GT*** fishpond area identified 6% 6% 22% 22% 36% 36% pre-classifier GT*** fishponds identified pre- classification 7 of 77 7 of 77 44 of 77 44 of 77 66 of 77 66 of 77 Percentage of GT*** fishpond area identified post-classifier 6% 5% 19% 11% 30% 23% GT*** fishponds identified post- 3 of 77 2 of 77 19 of 77 6 of 77 36 of 77 22 of 77 classification * Logistic Regression ** Classification and Regression Trees *** Ground truthing 94 Table A8. Time period comparison for Jessore. Time Frame Year Month Day Classifier Type LR * CART ** LR * CART ** LR * CART** Buffer Size 5 5 5 5 5 5 Percentage of GT*** fishpond area identified 1% 1% 44% 44% 91% 91% pre-classifier GT*** fishponds identified pre- 13 of 113 13 of 113 70 of 113 70 of 113 99 of 113 99 of 113 classification Percentage of GT*** fishpond area identified post-classifier 0.5% 0.2% 43% 6% 91% 3% GT*** fishponds identified post- 4 of 113 2 of 113 59 of 113 9 of 113 90 of 113 7 of 113 classification * Logistic Regression ** Classification and Regression Trees *** Ground truthing 95 Table A9. Time period comparison for Khulna. Time Frame Year Month Day Classifier Type LR * CART ** LR * CART ** LR * CART** Buffer Size 5 5 5 5 5 5 Percentage of GT*** fishpond area identified pre-classifier 0% 0% 36% 36% 67% 67% GT*** fishponds identified pre- 1 of 163 1 of 163 71 of 163 71 of 163 112 of 163 112 of 163 classification Percentage of GT*** fishpond area identified post-classifier 0% 0% 33% 9% 65% 5% GT*** fishponds identified post- 0 of 163 0 of 163 60 of 163 19 of 163 95 of 163 15 of 163 classification * Logistic Regression ** Classification and Regression Trees *** Ground truthing 96 Table A10. Time period comparison for Satkhira. Time Frame Year Month Day Classifier Type LR * CART ** LR * CART ** LR * CART** Buffer Size 5 5 5 5 5 5 Percentage of GT*** fishpond area identified 0% 0% 66% 66% 85% 85% pre-classifier GT*** fishponds identified pre- classification 0 of 208 0 of 208 75 of 208 75 of 208 115 of 208 115 of 208 Percentage of GT*** fishpond area identified post-classifier 0% 0% 66% 1% 84% 1% GT*** fishponds identified post- 0 of 208 0 of 208 72 of 208 9 of 208 103 of 208 14 of 208 classification * Logistic Regression ** Classification and Regression Trees *** Ground truthing 97 Table A11. Buffer size comparison for Bagerhat for both month and single day. Time frame Month Month Month Month Day Day Day Day Classifier CART* CART* CART type LR* * LR* * LR* CART** LR* ** LR* CART** LR* CART** LR* CART** LR* CART** Buffer Size 5 5 3 3 1 1 0 0 5 5 3 3 1 1 0 0 Percentage of GT*** fishpond area 52% 52% 50% 50% 45% 45% 42% 42% 67% 67% 66% 66% 61% 61% 58% 58% identified pre- classifier GT*** fishponds identified 139 of 139 of 136 of 136 of 111 of 111 of 95 of 95 of 182 of 182 of 180 of 180 of 157 of 157 of 141 of 141 of pre- 235 235 235 235 235 235 235 235 235 235 235 235 235 235 235 235 classificatio n Percentage of GT*** fishpond area identified 51% 19% 49% 17% 44% 0% 41% 13% 66% 16% 65% 15% 60% 13% 57% 15% post- classifier GT*** fishponds identified 117 of 49 of 107 of 48 of 87 of 80 of 39 of 159 of 51 of 156 of 50 of 138 of 46 of 131 of 44 of 0 of 235 post- 235 235 235 235 235 235 235 235 235 235 235 235 235 235 235 classificatio n * Logistic Regression ** Classification and Regression Trees *** Ground truthing 98 Table A12. Buffer size comparison for Barisal for both month and single day. Time frame Month Month Month Month Day Day Day Day Classifier CART * CART * CART type LR* * LR* * LR* CART** LR* ** LR* CART** LR* CART** LR* CART** LR* CART** Buffer Size 5 5 3 3 1 1 0 0 5 5 3 3 1 1 0 0 Percentage of GT*** fishpond area identified 4.3% 4.3% 3% 3% 1% 1% 1% 1% 16% 16% 13% 13% 10% 10% 6% 6% pre- classifier GT*** fishponds identified pre- 8 of 8 of 7 of 7 of 2 of 2 of 28 of 28 of 22 of 22 of 12 of 12 of 8 of 8 of 4 of 168 4 of 168 classificati 168 168 168 168 168 168 168 168 168 168 168 168 168 168 on Percentage of GT*** fishpond area identified 4% 4% 2% 3% 1% 0% 0% 1% 13% 12% 12% 8% 9% 6% 5% 4% post- classifier GT*** fishponds identified 3 of 4 of 3 of 4 of 0 of 1 of 11 of 10 of 11 of 6 of 9 of 3 of 4 of 3 of 1 of 168 0 of 168 post- 168 168 168 168 168 168 168 168 168 168 168 168 168 168 classificati on * Logistic Regression ** Classification and Regression Trees *** Ground truthing 99 Table A13. Buffer size comparison for Bhola for both month and single day. Time frame Month Month Month Month Day Day Day Day Classifier CART* CART* CART type LR* * LR* * LR* CART** LR* ** LR* CART** LR* CART** LR* CART** LR* CART** Buffer Size 5 5 3 3 1 1 0 0 5 5 3 3 1 1 0 0 Percentage of GT*** fishpond area identified 0% 0% 0% 0% 0% 0% 0% 0% 3% 3% 1% 1% 0% 0% 0% 0% pre- classifier GT*** fishponds identified pre- classificatio 0 of 27 0 of 27 0 of 27 0 of 27 0 of 27 0 of 27 0 of 27 0 of 27 1 of 27 1 of 27 1 of 27 1 of 27 0 of 27 0 of 27 0 of 27 0 of 27 n Percentage of GT*** fishpond area identified 0% 0% 0% 0% 0% 0% 0% 0% 3% 3% 0% 0% 0% 0% 0% 0% post- classifier GT*** fishponds identified post- 0 of 27 0 of 27 0 of 27 0 of 27 0 of 27 0 of 27 0 of 27 0 of 27 1 of 27 1 of 27 0 of 27 0 of 27 0 of 27 0 of 27 0 of 27 0 of 27 classificatio n * Logistic Regression ** Classification and Regression Trees *** Ground truthing 100 Table A14. Buffer size comparison for Gopalganj for both month and single day. Time frame Month Month Month Month Day Day Day Day Classifier CART type LR* ** LR* CART** LR* CART** LR* CART** LR* CART** LR* CART** LR* CART** LR* CART** Buffer Size 5 5 3 3 1 1 0 0 5 5 3 3 1 1 0 0 Percentage of GT*** fishpond area identified 22% 22% 20% 20% 13% 13% 9% 9% 36% 36% 35% 35% 32% 32% 30% 30% pre-classifier GT*** fishponds identified pre- 44 of 44 of 42 of 42 of 24 of 24 of 13 of 13 of 66 of 66 of 65 of 65 of 62 of 62 of 55 of 55 of classification 77 77 77 77 77 77 77 77 77 77 77 77 77 77 77 77 Percentage of GT*** fishpond area identified 19% 11% 18% 13% 8% 0% 9% 6% 30% 23% 30% 23% 28% 21% 28% 18% post- classifier GT*** fishponds identified 19 of 6 of 19 of 36 of 22 of 38 of 21 of 36 of 20 of 37 of 15 of post- 9 of 77 8 of 77 0 of 77 7 of 77 3 of 77 77 77 77 77 77 77 77 77 77 77 77 classification * Logistic Regression ** Classification and Regression Trees *** Ground truthing 101 Table A15. Buffer size comparison for Jessore for both month and single day. Time Month Month Month Month Day Day Day Day frame Classifier CART * CART * CART * CART * CART * CART * CART * CART* type LR* * LR* * LR* * LR* * LR* * LR* * LR* * LR* * Buffer 5 5 3 3 1 1 0 0 5 5 3 3 1 1 0 0 Size Percentag e of GT*** fishpond area identified pre- 44% 44% 38% 38% 22% 22% 18% 18% 91% 91% 91% 91% 89% 89% 86% 86% classifier GT*** fishponds identified pre- 70 of 70 of 62 of 62 of 43 of 43 of 35 of 35 of 99 of 99 of 97 of 97 of 91 of 91 of 90 of 90 of classificat 113 113 113 113 113 113 113 113 113 113 113 113 113 113 113 113 ion Percentag e of GT*** fishpond area identified post- classifier 43% 6% 37% 6% 21% 0% 17% 6% 91% 3% 90% 1% 88% 1% 86% 1% GT*** fishponds identified 59 of 9 of 55 of 13 of 33 of 0 of 26 of 10 of 90 of 7 of 90 of 5 of 88 of 6 of 85 of 5 of post- 113 113 113 113 113 113 113 113 113 113 113 113 113 113 113 113 classificat ion * Logistic Regression ** Classification and Regression Trees *** Ground truthing 102 Table A16. Buffer size comparison for Khulna for both month and single day. Time frame Month Month Month Month Day Day Day Day Classifier type CART* CART* CART* CART* CART* LR* * LR* * LR* * LR* * LR* CART** LR* CART** LR* CART** LR* * Buffer Size 5 5 3 3 1 1 0 0 5 5 3 3 1 1 0 0 Percentag e of GT*** fishpond area identified 36% 36% 32% 32% 29% 29% 27% 27% 67% 67% 66% 66% 62% 62% 59% 59% pre- classifier GT*** fishponds identified pre- 71 of 71 of 66 of 66 of 60 of 60 of 58 of 58 of 112 of 112 of 110 of 110 of 103 of 103 of 98 of 98 of classificat 163 163 163 163 163 163 163 163 163 163 163 163 163 163 163 163 ion Percentag e of GT*** fishpond area identified 33% 9% 30% 8% 27% 0% 0% 7% 65% 5% 65% 6% 61% 5% 56% 9% post- classifier GT*** fishponds identified 60 of 19 of 56 of 19 of 48 of 0 of 0 of 15 of 95 of 15 of 93 of 16 of 90 of 17 of 84 of 21 of post- 163 163 163 163 163 163 163 163 163 163 163 163 163 163 163 163 classificat ion * Logistic Regression ** Classification and Regression Trees *** Ground truthing 103 Table A17. Buffer size comparison for Satkhira for both month and single day. Time frame Month Month Month Month Day Day Day Day Classifier CART* CART* CART CART* type LR* * LR* * LR* CART** LR* ** LR* CART** LR* CART** LR* CART** LR* * Buffer Size 5 5 3 3 1 1 0 0 5 5 3 3 1 1 0 0 Percentage of GT*** fishpond area identified pre- 66% 66% 64% 64% 61% 61% 60% 60% 85% 85% 84% 84% 84% 84% 83% 83% classifier GT*** fishponds identified pre- 75 of 75 of 73 of 73 of 69 of 69 of 67 of 67 of 115 of 115 of 108 of 108 of 101 of 101 of 94 of 94 of classificati 208 208 208 208 208 208 208 208 208 208 208 208 208 208 208 208 on Percentage of GT*** fishpond area identified post- 66% 1% 64% 2% 61% 1% 60% 5% 84% 1% 84% 1% 83% 1% 83% 1% classifier GT*** fishponds identified 72 of 9 of 69 of 9 of 65 of 62 of 10 of 103 of 14 of 101 of 13 of 91 of 6 of 91 of 6 of 7 of 208 post- 208 208 208 208 208 208 208 208 208 208 208 208 208 208 208 classificati on * Logistic Regression ** Classification and Regression Trees *** Ground truthing 104 Table A18. Mode reducer applied for one month period in Bagerhat comparing water-identifying index combinations. Index Type Combined AWEId MNDWIe NDWIf Classifier Type LR a CART b LR a CART b LR a CART b LR a CART b Buffer Size 5 5 5 5 5 5 5 5 Percentage of GTc fishpond 64% 64% 74% 74% 57% 57% 62% 62% area identified pre-classifier GTc fishponds identified pre- 171 of 235 171 of 235 183 of 235 183 of 235 145 of 235 145 of 235 182 of 235 182 of 235 classification Percentage of GTc fishpond 63% 19% 73% 15% 56% 22% 60% 19% area identified post-classifier GTc fishponds identified post- 141 of 235 59 of 235 174 of 235 37 of 235 129 of 235 57 of 235 142 of 235 51 of 235 classification a Logistic Regression b Classification and Regression Trees c Ground truthing d Automated Water Extraction Index e Modified Normalized Difference Water Index f Normalized Difference Water Index 105 Table A19. Mode reducer applied for one month period in Barisal comparing water-identifying index combinations. Index Type Combined AWEId MNDWIe NDWIf Classifier Type LR a CART b LR a CART b LR a CART b LR a CART b Buffer Size 5 5 5 5 5 5 5 5 Percentage of GTc fishpond 12% 12% 14% 14% 8% 8% 17% 17% area identified pre-classifier GTc fishponds identified pre- 17 of 168 17 of 168 20 of 168 20 of 168 11 of 168 11 of 168 40 of 168 40 of 168 classification Percentage of GTc fishpond 11% 9% 13% 11% 8% 7% 13% 13% area identified post-classifier GTc fishponds identified post- 12 of 168 6 of 168 15 of 168 10 of 168 7 of 168 5 of 168 14 of 168 13 of 168 classification a Logistic Regression b Classification and Regression Trees c Ground truthing d Automated Water Extraction Index e Modified Normalized Difference Water Index f Normalized Difference Water Index 106 Table A20. Mode reducer applied for one month period in Bhola comparing water-identifying index combinations. Index Type Combined AWEId MNDWIe NDWIf Classifier Type LR a CART b LR a CART b LR a CART b LR a CART b Buffer Size 5 5 5 5 5 5 5 5 Percentage of GTc fishpond 0% 0% 0% 0% 0% 0% 7% 7% area identified pre-classifier GTc fishponds identified pre- 0 of 27 0 of 27 0 of 27 0 of 27 0 of 27 0 of 27 5 of 27 5 of 27 classification Percentage of GTc fishpond 0% 0% 0% 0% 0% 0% 0% 0% area identified post-classifier GTc fishponds identified post- 0 of 27 0 of 27 0 of 27 0 of 27 0 of 27 0 of 27 0 of 27 0 of 27 classification a Logistic Regression b Classification and Regression Trees c Ground truthing d Automated Water Extraction Index e Modified Normalized Difference Water Index f Normalized Difference Water Index 107 Table A21. Mode reducer applied for one month period in Gopalganj comparing water-identifying index combinations. Index Type Combined AWEId MNDWIe NDWIf Classifier Type LR a CART b LR a CART b LR a CART b LR a CART b Buffer Size 5 5 5 5 5 5 5 5 Percentage of GTc fishpond 30% 30% 48% 48% 14% 14% 33% 33% area identified pre-classifier GTc fishponds identified 58 of 77 58 of 77 62 of 77 62 of 77 22 of 77 22 of 77 61 of 77 61 of 77 pre- classification Percentage of GTc fishpond 24% 19% 46% 30% 10% 9% 26% 20% area identified post- classifier GTc fishponds identified 25 of 77 14 of 77 49 of 77 26 of 77 10 of 77 7of 77 26 of 77 16 of 77 post- classification a Logistic Regression b Classification and Regression Trees c Ground truthing d Automated Water Extraction Index e Modified Normalized Difference Water Index f Normalized Difference Water Index 108 Table A22. Mode reducer applied for one month period in Jessore comparing water-identifying index combinations. Index Type Combined AWEId MNDWIe NDWIf Classifier Type LR a CART b LR a CART b LR a CART b LR a CART b Buffer Size 5 5 5 5 5 5 5 5 Percentage of GTc 88% 88% 94% 94% 85% 85% 85% 85% fishpond area identified pre-classifier GTc fishponds identified 95 of 113 95 of 113 98 of 113 98 of 113 87 of 113 87 of 113 98 of 113 98 of 113 pre- classification Percentage of GTc 87% 1% 94% 3% 84% 0% 84% 3% fishpond area identified post-classifier GTc fishponds identified 88 of 113 5 of 113 94 of 113 6 of 113 84 of 113 2 of 113 87 of 113 10 of 113 post- classification a Logistic Regression b Classification and Regression Trees c Ground truthing d Automated Water Extraction Index e Modified Normalized Difference Water Index f Normalized Difference Water Index 109 Table A23. Mode reducer applied for one month period in Khulna comparing water-identifying index combinations. Index Type Combined AWEId MNDWIe NDWIf Classifier Type LR a CART b LR a CART b LR a CART b LR a CART b Buffer Size 5 5 5 5 5 5 5 5 Percentage of GTc 65% 65% 69% 69% 61% 61% 64% 64% fishpond area identified pre-classifier GTc fishponds identified 104 of 163 104 of 163 107 of 163 107 of 163 93 of 163 93 of 163 115 of 163 115 of 163 pre- classification Percentage of GTc 64% 7% 67% 6% 60% 9% 62% 6% fishpond area identified post-classifier GTc fishponds identified 93 of 163 18 of 163 96 of 163 17 of 163 88 of 163 21 of 163 94 of 163 14 of 163 post- classification a Logistic Regression b Classification and Regression Trees c Ground truthing d Automated Water Extraction Index e Modified Normalized Difference Water Index f Normalized Difference Water Index 110 Table A24. Mode reducer applied for one month period in Satkhira comparing water-identifying index combinations. Index Type Combined AWEId MNDWIe NDWIf Classifier Type LR a CART b LR a CART b LR a CART b LR a CART b Buffer Size 5 5 5 5 5 5 5 5 Percentage of GTc fishpond 84% 84% 84% 84% 82% 82% 86% 86% area identified pre-classifier GTc fishponds identified 100 of 208 100 of 208 102 of 208 102 of 208 93 of 208 93 of 208 126 of 208 126 of 208 pre- classification Percentage of GTc fishpond 83% 1% 83% 1% 82% 1% 86% 1% area identified post- classifier GTc fishponds identified 92 of 208 6 of 208 95 of 208 9 of 208 88 of 208 6 of 208 112 of 208 6 of 208 post- classification a Logistic Regression b Classification and Regression Trees c Ground truthing d Automated Water Extraction Index e Modified Normalized Difference Water Index f Normalized Difference Water Index 111 Table A25. allNonZero reducer applied for one month period in Bagerhat comparing water-identifying index combinations. Time frame Month Month Month Month Index Type Combined Combined AWEId AWEId MNDWIe MNDWIe NDWIf NDWIf Classifier type LRa CARTb LRa CARTb LRa CARTb LRa CARTb Buffer Size 5 5 5 5 5 5 5 5 Percentage of GTc fishpond area identified pre-classifier 52% 52% 60% 60% 48% 48% 52% 52% GTc fishponds identified pre- 157 of classification 139 of 235 139 of 235 156 of 235 156 of 235 122 of 235 122 of 235 157 of 235 235 Percentage of GT c fishpond area identified post-classifier 51% 19% 60% 14% 47% 18% 50% 14% GT c fishponds identified post- 117 of 235 49 of 235 134 of 235 46 of 235 108 of 235 50 of 235 115 of 235 49 of 235 classification a Logistic Regression b Classification and Regression Trees c Ground truthing d Automated Water Extraction Index e Modified Normalized Difference Water Index f Normalized Difference Water Index 112 Table A26. allNonZero reducer applied for one month period in Barisal comparing water-identifying index combinations. Time frame Month Month Month Month Index Type Combined Combined AWEId AWEId MNDWIe MNDWIe NDWIf NDWIf Classifier type LRa CARTb LRa CARTb LRa CARTb LRa CARTb Buffer Size 5 5 5 5 5 5 5 5 Percentage of GTc fishpond area identified 4.3% 4.3% 6% 6% 3% 3% 6% 6% pre-classifier GTc fishponds identified pre- classification 8 of 168 8 of 168 10 of 168 10 of 168 6 of 168 6 of 168 18 of 168 18 of 168 Percentage of GT c fishpond area identified 4% 4% 6% 5% 2% 2% 4% 5% post-classifier GT c fishponds identified post- 3 of 168 4 of 168 5 of 168 4 of 168 3 of 168 3 of 168 5 of 168 6 of 168 classification a Logistic Regression b Classification and Regression Trees c Ground truthing d Automated Water Extraction Index e Modified Normalized Difference Water Index f Normalized Difference Water Index 113 Table A27. allNonZero reducer applied for one month period in Bhola comparing water-identifying index combinations. Time frame Month Month Month Month Index Type Combined Combined AWEI d AWEI d MNDWI e MNDWI e NDWI f NDWIf Classifier type LRa CARTb LRa CARTb LRa CARTb LRa CARTb Buffer Size 5 5 5 5 5 5 5 5 Percentage of GTc fishpond area identified 0% 0% 0% 0% 0% 0% 0% 0% pre-classifier GTc fishponds identified pre- classification 0 of 27 0 of 27 0 of 27 0 of 27 0 of 27 0 of 27 0 of 27 0 of 27 Percentage of GT c fishpond area identified 0% 0% 0% 0% 0% 0% 0% 0% post-classifier GT c fishponds identified post- 0 of 27 0 of 27 0 of 27 0 of 27 0 of 27 0 of 27 0 of 27 0 of 27 classification a Logistic Regression b Classification and Regression Trees c Ground truthing d Automated Water Extraction Index e Modified Normalized Difference Water Index f Normalized Difference Water Index 114 Table A28. allNonZero reducer applied for one month period in Gopalganj comparing water-identifying index combinations. Time frame Month Month Month Month Index Type Combined Combined AWEId AWEId MNDWIe MNDWIe NDWIf NDWIf Classifier type LRa CARTb LRa CARTb LRa CARTb LRa CARTb Buffer Size 5 5 5 5 5 5 5 5 Percentage of GTc fishpond area identified 22% 22% 38% 38% 10% 10% 25% 25% pre-classifier GTc fishponds identified pre- classification 44 of 77 44 of 77 56 of 77 56 of 77 14 of 77 14 of 77 53 of 77 53 of 77 Percentage of GTc fishpond area identified 19% 11% 35% 26% 7% 3% 21% 16% post-classifier GT c fishponds identified post- 19 of 77 6 of 77 38 of 77 24 of 77 8 of 77 3 of 77 22 of 77 12 of 77 classification a Logistic Regression b Classification and Regression Trees c Ground truthing d Automated Water Extraction Index e Modified Normalized Difference Water Index f Normalized Difference Water Index 115 Table A29. allNonZero reducer applied for one month period in Jessore comparing water-identifying index combinations. Time frame Month Month Month Month Index Type Combined Combined AWEId AWEId MNDWIe MNDWIe NDWIf NDWIf Classifier type LRa CARTb LRa CARTb LRa CARTb LRa CARTb Buffer Size 5 5 5 5 5 5 5 5 Percentage of GTc fishpond area identified 44% 44% 53% 53% 24% 24% 66% 66% pre-classifier GTc fishponds identified pre- classification 70 of 113 70 of 113 74 of 113 74 of 113 41 of 113 41 of 113 86 of 113 86 of 113 Percentage of GT c fishpond area identified 42.8% 6% 52.3% 10% 23.0% 4% 65.1% 10% post-classifier GT c fishponds identified post- 59 of 113 9 of 113 71 of 113 9 of 113 38 of 113 10 of 113 74 of 113 12 of 113 classification a Logistic Regression b Classification and Regression Trees c Ground truthing d Automated Water Extraction Index e Modified Normalized Difference Water Index f Normalized Difference Water Index 116 Table A30. allNonZero reducer applied for one month period in Khulna comparing water-identifying index combinations. Time frame Month Month Month Month Index Type Combined Combined AWEId AWEId MNDWIe MNDWIe NDWIf NDWIf Classifier type LRa CARTb LRa CARTb LRa CARTb LRa CARTb Buffer Size 5 5 5 5 5 5 5 5 Percentage of GTc fishpond area identified 36% 36% 43% 43% 33% 33% 31% 31% pre-classifier GTc fishponds identified pre- classification 71 of 163 71 of 163 79 of 163 79 of 163 67 of 163 67 of 163 64 of 163 64 of 163 Percentage of GT c fishpond area identified 33% 9% 42% 43% 32% 8% 28% 10% post-classifier GT c fishponds identified post- 60 of 163 19 of 163 69 of 163 79 of 163 62 of 163 21 of 163 47 of 163 17 of 163 classification a Logistic Regression b Classification and Regression Trees c Ground truthing d Automated Water Extraction Index e Modified Normalized Difference Water Index f Normalized Difference Water Index 117 Table A31. allNonZero reducer applied for one month period in Satkhira comparing water-identifying index combinations. Time frame Month Month Month Month Index Type Combined Combined AWEI d AWEI d MNDWI e MNDWI e NDWI f NDWIf Classifier type LRa CARTb LRa CARTb LRa CARTb LRa CARTb Buffer Size 5 5 5 5 5 5 5 5 Percentage of GTc fishpond area identified 66% 66% 69% 69% 63% 63% 62% 62% pre-classifier GTc fishponds identified pre- classification 75 of 208 75 of 208 75 of 208 75 of 208 73 of 208 73 of 208 80 of 208 80 of 208 Percentage of GT c fishpond area identified 66% 1% 69% 3% 63% 1% 61% 2% post-classifier GT c fishponds identified post- 72 of 208 9 of 208 72 of 208 6 of 208 70 of 208 8 of 208 72 of 208 9 of 208 classification a Logistic Regression b Classification and Regression Trees c Ground truthing d Automated Water Extraction Index e Modified Normalized Difference Water Index f Normalized Difference Water Index 118 Table A32. Water-identifying index combination comparison for Bagerhat using single day time period. Time frame Day Day Day Day Index Type Combined Combined AWEI d AWEI d MNDWI e MNDWI e NDWI f NDWIf Classifier type LR a CARTb LRa CARTb LRa CARTb LRa CARTb Buffer Size 5 5 5 5 5 5 5 5 Percentage of GTc fishpond area identified 67% 67% 76% 76% 62% 62% 67% 67% pre-classifier GTc fishponds identified pre- classification 182 of 235 182 of 235 190 of 235 190 of 235 155 of 235 155 of 235 197 of 235 197 of 235 Percentage of GT c fishpond area identified 66% 16% 76% 13% 61% 16% 65% 16% post-classifier GT c fishponds identified 159 of 235 51 of 235 181 of 235 30 of 235 138 of 235 50 of 235 164 of 235 54 of 235 post- classification a Logistic Regression b Classification and Regression Trees c Ground truthing d Automated Water Extraction Index e Modified Normalized Difference Water Index f Normalized Difference Water Index 119 Table A33. Water-identifying index combination comparison for Barisal using single day time period. Time frame Day Day Day Day Index Type Combined Combined AWEI d AWEI d MNDWI e MNDWI e NDWI f NDWIf Classifier type LRa CARTb LRa CARTb LRa CARTb LRa CARTb Buffer Size 5 5 5 5 5 5 5 5 Percentage of GTc fishpond area identified 16.3% 16.3% 19% 19% 11% 11% 26% 26% pre-classifier GTc fishponds identified pre- classification 28 of 168 28 of 168 28 of 168 28 of 168 12 of 168 12 of 168 61 of 168 61 of 168 Percentage of GT c fishpond area identified 13% 12% 15% 15% 10% 6% 20% 17% post-classifier GT c fishponds identified 11 of 168 10 of 168 13 of 168 12 of 168 8 of 168 3 of 168 21 of 168 16 of 168 post- classification a Logistic Regression b Classification and Regression Trees c Ground truthing d Automated Water Extraction Index e Modified Normalized Difference Water Index f Normalized Difference Water Index 120 Table A34. Water-identifying index combination comparison for Bhola using single day time period. Time frame Day Day Day Day Index Type Combined Combined AWEI d AWEI d MNDWI e MNDWI e NDWI f NDWIf Classifier type LRa CARTb LRa CARTb LRa CARTb LRa CARTb Buffer Size 5 5 5 5 5 5 5 5 Percentage of GTc fishpond area identified 3% 3% 3% 3% 0% 0% 20% 20% pre-classifier GTc fishponds identified pre- classification 1 of 27 1 of 27 2 of 27 2 of 27 0 of 27 0 of 27 10 of 27 10 of 27 Percentage of GT c fishpond area identified 3% 3% 3% 0% 0% 0% 12% 9% post-classifier GT c fishponds identified post- 1 of 27 1 of 27 1 of 27 0 of 27 0 of 27 0 of 27 4 of 27 3 of 27 classification a Logistic Regression b Classification and Regression Trees c Ground truthing d Automated Water Extraction Index e Modified Normalized Difference Water Index f Normalized Difference Water Index 121 Table A35. Water-identifying index combination comparison for Gopalganj using single day time period. Time frame Day Day Day Day Index Type Combined Combined AWEI d AWEI d MNDWI e MNDWI e NDWI f NDWIf Classifier type LRa CARTb LRa CARTb LRa CARTb LRa CARTb Buffer Size 5 5 5 5 5 5 5 5 Percentage of GTc fishpond area identified 36% 36% 54% 54% 21% 21% 38% 38% pre-classifier GTc fishponds identified pre- classification 66 of 77 66 of 77 67 of 77 67 of 77 38 of 77 38 of 77 69 of 77 69 of 77 Percentage of GT c fishpond area identified 30% 23% 50% 28% 15% 15% 31% 23% post-classifier GT c fishponds identified 36 of 77 22 of 77 56 of 77 26 of 77 20 of 77 13 of 77 37 of 77 22 of 77 post- classification a Logistic Regression b Classification and Regression Trees c Ground truthing d Automated Water Extraction Index e Modified Normalized Difference Water Index f Normalized Difference Water Index 122 Table A36. Water-identifying index combination comparison for Jessore using single day time period. Time frame Day Day Day Day Index Type Combined Combined AWEI d AWEI d MNDWI e MNDWI e NDWI f NDWIf Classifier type LRa CARTb LRa CARTb LRa CARTb LRa CARTb Buffer Size 5 5 5 5 5 5 5 5 Percentage of GTc fishpond area identified 91.2% 91% 95% 95% 90% 90% 88% 88% pre-classifier GTc fishponds identified pre- classification 99 of 113 99 of 113 102 of 113 102 of 113 91 of 113 91 of 113 103 of 113 103 of 113 Percentage of GT c fishpond area identified 90.7% 3% 94.3% 3% 89.8% 2% 87.6% 4% post-classifier GT c fishponds identified post- 90 of 113 7 of 113 95 of 113 7 of 113 89 of 113 5 of 113 92 of 113 7 of 113 classification a Logistic Regression b Classification and Regression Trees c Ground truthing d Automated Water Extraction Index e Modified Normalized Difference Water Index f Normalized Difference Water Index 123 Table A37. Water-identifying index combination comparison for Khulna using single day time period. Time frame Day Day Day Day Index Type Combined Combined AWEI d AWEI d MNDWI e MNDWI e NDWI f NDWIf Classifier type LRa CARTb LRa CARTb LRa CARTb LRa CARTb Buffer Size 5 5 5 5 5 5 5 5 Percentage of GTc fishpond area identified 67% 67% 74% 74% 63% 63% 67% 67% pre-classifier GTc fishponds identified pre- classification 112 of 163 112 of 163 117 of 163 117 of 163 103 of 163 103 of 163 122 of 163 122 of 163 Percentage of GT c fishpond area identified 65% 5% 73% 6% 62% 5% 56% 7% post-classifier GT c fishponds identified post- 95 of 163 15 of 163 109 of 163 19 of 163 94 of 163 15 of 163 88 of 163 18 of 163 classification a Logistic Regression b Classification and Regression Trees c Ground truthing d Automated Water Extraction Index e Modified Normalized Difference Water Index f Normalized Difference Water Index 124 Table A38. Water-identifying index combination comparison for Satkhira using single day time period. Time frame Day Day Day Day Index Type Combin Combined AWEId AWEId MNDWIe MNDWIe NDWIf NDWIf ed Classifier type LRa CARTb LRa CARTb LRa CARTb LRa CARTb Buffer Size 5 5 5 5 5 5 5 5 Percentage of GTc fishpond area identified 85% 85% 84% 84% 84% 84% 91% 91% pre-classifier GTc fishponds identified pre- 115 of classification 115 of 208 114 of 208 114 of 208 100 of 208 100 of 208 159 of 208 159 of 208 208 Percentage of GT c fishpond area identified 84% 1% 84% 2% 84% 1% 91% 0% post-classifier GT c fishponds 103 of identified 14 of 208 103 of 208 14 of 208 93 of 208 6 of 208 133 of 208 6 of 208 208 post- classification a Logistic Regression b Classification and Regression Trees c Ground truthing d Automated Water Extraction Index e Modified Normalized Difference Water Index f Normalized Difference Water Index 125 Table A39. Best results of the index tests for each of the seven districts. District Bagerhat Barisal Bhola Gopalganj Jessore Khulna Satkhira Water Index AWEI d NDWI e NDWI e AWEI d AWEI d AWEI d NDWIe Image date 10/28/2020 10/13/2020 11/7/2020 10/28/2020 11/5/2020 11/7/2020 11/5/2020 Classifier LRa CARTb LRa CARTb LRa CARTb LRa LRa CARTb LRa CARTb LRa CARTb LRa Type Buffer Size 5 5 0 0 5 5 5 5 5 5 5 5 5 5 Percentage of GTc Fishpond area identified 76% 76% 26% 26% 20% 20% 54% 54% 95% 95% 74% 74% 91% 91% pre-classifier GT c Fishponds 190 of 190 of 61 of 61 of 10 of 10 of 67 of 67 of 102 of 102 of 117 of 117 of 159 of 159 of Identified 235 235 168 168 27 27 77 77 113 113 163 163 208 208 Percentage of GTc Fishpond area identified 76% 13% 20% 17% 12% 9% 50% 28% 94.3% 3% 73% 6% 91% 0.3% post-classifier GTc Fishponds Identified 181 of 21 of 16 of 56 of 26 of 109 of 133 of 30 of 235 4 of 27 3 of 27 95 of 113 7 of 113 19 of 163 6 of 208 After 235 168 168 77 77 163 208 Classification a Logistic Regression b Classification and Regression Trees c Ground truthing d Automated Water Extraction Index e Normalized Difference Water Index 126 Table A40. Results from adding the Laplacian 5×5 convolution filter for each district. District Bagerhat Barisal Bhola Gopalganj Jessore Khulna Satkhira Water Index AWEId NDWIe NDWIe AWEId AWEId AWEId NDWIe Image date 10/28/2020 10/13/2020 11/7/2020 10/28/2020 11/5/2020 11/7/2020 11/5/2020 Classifier Type LRa CARTb LRa CARTb LRa CARTb LRa CARTb LRa CARTb LRa CARTb LRa CARTb Buffer Size 5 5 5 5 5 5 5 5 5 5 5 5 5 5 Percentage of GTc 76% 76% 58% 58% 57% 57% 68% 68% 91.5% 91% 78% 78% 82% 82% Fishpond area identified pre- classifier GT c Fishponds 235 of 235 of 160 of 160 of 27 of 27 of 77 of 77 of 113 of 113 of 163 of 163 of 206 of 206 of Identified 235 235 168 168 27 27 77 77 113 113 163 163 208 208 Percentage of GTc 25% 2% 44% 37% 43% 33% 19% 3% 91.4% 0% 56% 1% 56% 5% Fishpond area identified post- classifier GTc Fishponds 131 of 40 of 235 90 of 168 83 of 168 16 of 14 of 30 of 10 of 113 of 3 of 113 134 of 239 of 138 of 56 of Identified After 235 27 27 77 77 113 163 163 208 208 Classification # of Hits 131 40 90 83 16 14 30 10 113 3 134 23 138 56 # of Classified 94794 66043 219955 185758 115902 98263 47046 33129 123786 83825 97045 66392 155887 126110 Fishponds # of GTc Fishponds 235 235 168 168 27 27 77 77 113 113 163 163 208 208 a Logistic Regression b Classification and Regression Trees c Ground truthing d Automated Water Extraction Index e Normalized Difference Water Index 127 Table A41. Comparison of LR, CART, RF, and SVM results when applied to ground-truthing fishpond data for Bagerhat, Barisal, Bhola, and Gopalganj during one day with that district’s best performing water-identifying index. District Bagerhat Barisal Bhola Gopalganj Water Index AWEIf NDWIg NDWIg AWEIf Image Date 10/28/2020 10/13/2020 11/7/2020 10/28/2020 Classifier Type LRa CARTb RFc SVMd LRa CARTb RFc SVMd LRa CARTb RFc SVM LRa CART RFc SVMd d b Buffer Size 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 e Percentage of GT 76% 76% 76% 76% 58% 58% 58% 58% 57% 57% 57% 57% 68% 68% 68% 68% Fishpond area identified pre- classifier GTe Fishponds 235 of 235 of 235 of 235 of 160 of 160 of 160 of 160 of 27 of 27 of 27 of 27 of 77 of 77 of 77 of 77 of Identified 235 235 235 235 168 168 168 168 27 27 27 27 77 77 77 77 Percentage of GTe 25% 2% 2% 26% 44% 37% 35% 42% 43% 33% 33% 41% 19% 3% 3% 20% Fishpond area identified post- classifier GTe Fishponds 131 of 40 of 37 of 153 of 90 of 83 of 76 of 103 of 16 of 14 of 14 of 17 of 30 of 10 of 10 of 38 of Identified Post- 235 235 235 235 168 168 168 168 27 27 27 27 77 77 77 77 Classifier # of Hits 131 40 37 153 90 83 76 103 16 14 14 17 30 10 10 38 # of Classified 94794 66043 58466 426858 219955 185758 154127 666932 115902 98263 80723 446448 47046 33129 28794 24344 Fishponds 1 # of GTe Fishponds 235 235 235 235 168 168 168 168 27 27 27 27 77 77 77 77 a Logistic Regression b Classification and Regression Trees c Random Forest d Support Vector Machine e Ground truthing f Automated Water Extraction Index g Normalized Difference Water Index 128 Table A42. Comparison of LR, CART, RF, and SVM results when applied to ground-truthing fishpond data for Jessore, Khulna, and Satkhira during one day with that district’s best performing water-identifying index. District Jessore Khulna Satkhira Water AWEIf AWEIf NDWIg Index Image 11/5/2020 11/7/2020 11/5/2020 Date Classifier LRa CARTb RFc SVMd LRa CARTb RFc SVMd LRa CARTb RFc SVMd Type Buffer 5 5 5 5 5 5 5 5 5 5 5 5 Size Percentage 91.5% 91% 91.5% 91% 78% 78% 78% 78% 82% 82% 82% 82% of GTe Fishpond area identified pre- classifier GTe 113 of 113 113 of 113 113 of 113 113 of 113 163 of 163 163 of 163 163 of 163 163 of 163 206 of 208 206 of 208 206 of 208 206 of 208 Fishponds Identified Percentage 91.4% 0% 0.3% 91% 56% 1% 1% 56% 56% 5% 4% 56% of GTe Fishpond area identified post- classifier GTe 113 of 113 3 of 113 3 of 113 113 of 113 134 of 163 23 of 163 23 of 163 136 of 163 138 of 208 56 of 208 51 of 208 147 of 208 Fishponds Identified Post- Classifier # of Hits 113 3 3 113 134 23 23 136 138 56 51 147 129 Table A42 (cont’d) # of 123786 83825 74365 667161 97045 66392 58987 477444 155887 126110 104324 389366 Classified Fishponds # of GTe 113 113 113 113 163 163 163 163 208 208 208 208 Fishponds a Logistic Regression b Classification and Regression Trees c Random Forest d Support Vector Machine e Ground truthing f Automated Water Extraction Index g Normalized Difference Water Index 130 Table A43. Land use as percent of total area for 10-meter buffer around ground truth fishpond locations for each district. District Land Use Bagerhat Barisal Bhola Gopalganj Jessore Khulna Satkhira Cultivated Land 79 46 21 67 90 63 6 Forest 0 0 0 0 0 0 0 Grass Land 0 0 0 0 0 0 0 Shrub Land 0 0 0 0 0 0 0 Wetland 0 0 0 0 0 0 0 Water Body 8 0 0 0 1 19 82 Artificial Surfaces 13 54 79 33 9 18 12 Bareland 0 0 0 0 0 0 0 Unknown 0 0 0 0 0 0 0 131 Table A44. Land use as percent of total area for 50-meter buffer around ground truth fishpond locations for each district. District Land Use Bagerhat Barisal Bhola Gopalganj Jessore Khulna Satkhira Cultivated Land 80 46 26 68 86 61 10 Forest 0 0 0 0 0 0 0 Grass Land 0 0 0 0 0 0 0 Shrub Land 0 0 0 0 0 0 0 Wetland 0 0 0 0 0 0 0 Water Body 7 0 0 0 1 16 69 Artificial Surfaces 13 54 74 32 13 23 21 Bareland 0 0 0 0 0 0 0 Unknown 0 0 0 0 0 0 0 132 Table A45. Land use as percent of total area for 100-meter buffer around ground truth fishpond locations for each district. District Land Use Bagerhat Barisal Bhola Gopalganj Jessore Khulna Satkhira Cultivated Land 78 50.4 27 71 82 59 12 Forest 0 0 0 0 0 0 0 Grass Land 0 0 0 0 0 0 0 Shrub Land 0 0 0 0 0 0 0 Wetland 0 0 0 0 0 0 0 Water Body 7 0.1 0 0 2 15 63 Artificial Surfaces 15 49.5 73 29 16 26 25 Bareland 0 0 0 0 0 0 0 Unknown 0 0 0 0 0 0 0 133 Table A46. Median ground truth fishpond size for each district. District Statistic Bagerhat Barisal Bhola Gopalganj Jessore Khulna Satkhira Median Ground truth fishpond size 2503 833 713 2237 3169 2972 1252 (m2) 134 BIBLIOGRAPHY 135 BIBLIOGRAPHY Abramowitz, M., & Stegun, C. A. (1972). Handbook of Mathematical Functions with Formulas, Graphs, and Mathematical Tables (9th ed., p. 14). Dover. Agmalaro, M. A., Sitanggang, I. S., & Waskito, M. L. (2021). Sentinel 1 Classification for Garlic Land Identification using Support Vector Machine. 9th International Conference on Information and Communication Technology, 440–444. Ahmed, N., Thompson, S., & Glaser, M. (2019). Global Aquaculture Productivity, Environmental Sustainability, and Climate Change Adaptability. Environmental Management, 63(2), 159–172. https://doi.org/10.1007/s00267-018-1117-3 Al Sayah, M. J., Nedjai, R., Abdallah, C., & Khouri, M. (2020). On the use of remote sensing to map the proliferation of aquaculture ponds and to investigate their effect on local climate, perspectives from the Claise watershed, France. Springer Nature Switzerland. https://doi.org/10.1007/s10661-020-08250-0 Anand, A., Krishnan, P., Kantharajan, G., Suryavanshi, A., Kawishwar, P., Raj, U., Rao, C. S., Choudhury, S. B., Manjulatha, C., & Babu, D. E. (2020). Assessing the water spread area available for fish culture and fish production potential in inland lentic waterbodies using remote sensing: A case study from Chhattisgarh State, India. Remote Sensing Applications: Society and Environment, 17. https://doi.org/10.1016/j.rsase.2019.100273 Ballester-Berman, D. J., Sanchez-Jerez, P., & Marino, A. (2018). Detection of aquaculture structures using Sentinel-1 data. Proceedings of the European Conference on Synthetic Aperture Radar, EUSAR, 2018-June, 888–891. https://servicio.pesca.mapama.es/acuivisor/ Bangladesh Meteorological Department. (2021). Maproom: Seasonal Climate Analysis. Maproom. http://datalibrary.bmd.gov.bd/maproom/Climatology/Climate_Analysis/seasonal.html?resol ution=0.05&YearStart=2020&YearEnd=2020&seasonStart=Jan&seasonEnd=Dec&var=Rai n&yearlyStat=Mean Bangladesh Water Development Board. (2019). Summary of Rainfall in Bangladesh for the Year 2017 & 2018. Bello, O. M., & Aina, Y. A. (2014). Satellite Remote Sensing as a Tool in Disaster Management and Sustainable Development: Towards a Synergistic Approach. Procedia - Social and Behavioral Sciences, 120, 365–373. https://doi.org/10.1016/j.sbspro.2014.02.114 Belton, B., & Azad, A. (2012). The characteristics and status of pond aquaculture in Bangladesh. Aquaculture, 358–359, 196–204. https://doi.org/10.1016/j.aquaculture.2012.07.002 Belton, B., Bush, S. R., & Little, D. C. (2018). Not just for the wealthy: Rethinking farmed fish consumption in the Global South. Global Food Security, 16, 85–92. 136 https://doi.org/10.1016/J.GFS.2017.10.005 Bernzen, A., Pritchard, B., Braun, B., Belton, B., & Rigg, J. (2021). Guest Editorial: Geographies of engagement, livelihoods and possibility in South and Southeast Asian deltas. Singapore Journal of Tropical Geography, 42(2), 197–202. https://doi.org/10.1111/SJTG.12373 Biggs, J., von Fumetti, S., & Kelly-Quinn, M. (2017). The importance of small waterbodies for biodiversity and ecosystem services: implications for policy makers. In Hydrobiologia (Vol. 793, Issue 1, pp. 3–39). Springer International Publishing. https://doi.org/10.1007/s10750- 016-3007-0 Charles, H., Godfray, J., Beddington, J. R., Crute, I. R., Haddad, L., Lawrence, D., Muir, J. F., Pretty, J., Robinson, S., Thomas, S. M., & Toulmin, C. (2010). Food Security: The Challenge of Feeding 9 Billion People. In New Series (Vol. 12, Issue 5967). China Ministry of Natural Resources. (2020). GlobeLand30. http://www.globeland30.org/defaults_en.html?type=data&src=/Scripts/map/defaults/En/bro wse_en.html&head=browse Chu, D. (2020). Remote Sensing of Land Use and Land Cover in Mountain Region. In Remote Sensing of Land Use and Land Cover in Mountain Region. https://doi.org/10.1007/978-981- 13-7580-4 Coffey, V. C. (2012, April 1). Multispectral Imaging Moves into the Mainstream. Optics & Photonics. https://www.osa- opn.org/home/articles/volume_23/issue_4/features/multispectral_imaging_moves_into_the_ mainstream/ CRISP, & National University of Singapore. (2001). Infrared Remote Sensing. https://crisp.nus.edu.sg/~research/tutorial/infrared.htm De Graaf, G., Khan, G., Omar, M., Lubna, F., Abdullah, Y., & Mamun, A. (2000). Environment and GIS Support Project for Water Sector Planning EGIS-II FISH-GIS An Introduction to the Use of Geographical Information Systems and Remote Sensing in Fisheries Monitoring, Analysis and Management. www.Nefisco.org Department of Agricultural Extension. (2021). Agro-Ecological Zones -AEZs Maps. https://www.bamis.gov.bd/en/page/aezs-maps/ Department of Fisheries. (2018). Yearbook of Fisheries Statistics of Bangladesh 2017-18. www.fisheries.gov.bd Duan, Y., Li, X., Zhang, L., Chen, D., Liu, S., & Ji, H. (2020). Mapping national-scale aquaculture ponds based on the Google Earth Engine in the Chinese coastal zone. Aquaculture, 520. https://doi.org/10.1016/j.aquaculture.2019.734666 Duan, Y., Li, X., Zhang, L., Liu, W., Liu, S., Chen, D., & Ji, H. (2020). Detecting spatiotemporal 137 changes of large-scale aquaculture ponds regions over 1988–2018 in Jiangsu Province, China using Google Earth Engine. Ocean and Coastal Management, 188, 105144. https://doi.org/10.1016/j.ocecoaman.2020.105144 Dumitru, C. O., Schwarz, G., & Datcu, M. (2018). Earth Observation Semantics and Data Analytics for Coastal Environmental Areas. International Journal on Advances in Software, 11(3), 390–399. http://www.iariajournals.org/software/2018, Earth Observing System. (2019, December 4). Satellite Data: What Spatial Resolution Is Enough For You? Earth Observing System. https://eos.com/blog/satellite-data-what-spatial- resolution-is-enough-for-you/ Ellowitz, H. I. (1992). An introduction to microwave remote sensing. Microwave Journal, 35(9), 69. https://go-gale- com.proxy1.cl.msu.edu/ps/i.do?p=ITOF&u=msu_main&id=GALE%7CA13412901&v=2.1 &it=r&sid=summon ESA. (2020a). CBERS - Second Generation . https://earth.esa.int/web/eoportal/satellite- missions/c-missions/cbers-3-4 ESA. (2020b). Sentinel-1 - Missions . https://sentinel.esa.int/web/sentinel/missions/sentinel-1 ESA. (2020c). TOPSAR Processing - Sentinel-1 SAR Technical Guide - Sentinel Online. https://sentinel.esa.int/web/sentinel/technical-guides/sentinel-1-sar/products- algorithms/level-1-algorithms/topsar-processing ESA. (2020d). WorldView-1. https://directory.eoportal.org/web/eoportal/satellite-missions/v-w- x-y-z/worldview-1 FAO. (1989). 4. AQUACULTURE METHODS AND PRACTICES: A SELECTED REVIEW. http://www.fao.org/3/t8598e/t8598e05.htm FAO. (2014). A-2 World aquaculture production by culture environment. http://www.fao.org/fishery/docs/STAT/summary/a-2.pdf FAO. (2018). WORLD FISHERIES AND AQUACULTURE. www.fao.org/publications FAO. (2020a). FIGIS - Fisheries Statistics - Aquaculture. http://www.fao.org/figis/servlet/TabLandArea?tb_ds=Aquaculture&tb_mode=TABLE&tb_ act=SELECT&tb_grp=COUNTRY FAO. (2020b). National Aquaculture Sector Overview - Bangladesh. http://www.fao.org/fishery/countrysector/naso_bangladesh/en FAO. (2020c). The State of World Fisheries and Aquaculture. In The State of World Fisheries and Aquaculture 2020. FAO. https://doi.org/10.4060/ca9229en Faul, A. C. (2019). A Concise Introduction to Machine Learning. In A Concise Introduction to 138 Machine Learning. Chapman and Hall/CRC. https://doi.org/10.1201/9781351204750 Feed the Future. (2021, January). Bangladesh. https://www.feedthefuture.gov/country/bangladesh/ Feyisa, G. L., Meilby, H., Fensholt, R., & Proud, S. R. (2014). Automated Water Extraction Index: A new technique for surface water mapping using Landsat imagery. https://doi.org/10.1016/j.rse.2013.08.029 Frohn, R. C. (2007). The use of landscape pattern metrics in remote sensing image classification. Https://Doi.Org/10.1080/01431160500212229, 27(10), 2025–2032. https://doi.org/10.1080/01431160500212229 Funge-Smith, Simon; Phillips, M. J. (2001). Aquaculture Systems and Species. http://www.fao.org/3/AB412E/ab412e07.htm Gade, M., & Stoffelen, A. (2019). An Introduction to Microwave Remote Sensing of the Asian Seas. In V. Barale & M. Gade (Eds.), Remote Sensing of the Asian Seas (pp. 81–101). Springer, Cham. https://doi.org/10.1007/978-3-319-94067-0_4 Gao, J., Du, Z., Shi, Z., Xu, Z., Cao, Q., & Tang, R. (2018). Switching impulse noise filter based on Laplacian convolution and pixels grouping for color images. Signal, Image and Video Processing, 12(8), 1523–1529. https://doi.org/10.1007/S11760-018-1308-7 Gautam, V. K., Gaurav, P. K., Murugan, P., & Annadurai, M. (2015). Assessment of Surface Water Dynamicsin Bangalore Using WRI, NDWI, MNDWI, Supervised Classification and K-T Transformation. Aquatic Procedia, 4, 739–746. https://doi.org/10.1016/J.AQPRO.2015.02.095 Google. (2020). Introduction to Machine Learning | Machine Learning Crash Course. Google. https://developers.google.com/machine-learning/crash-course/ml-intro Gusmawati, N., Soulard, B., Selmaoui-Folcher, N., Proisy, C., Mustafa, A., Le Gendre, R., Laugier, T., & Lemonnier, H. (2017). Surveying shrimp aquaculture pond activity using multitemporal VHSR satellite images-case study from the Perancak estuary, Bali, Indonesia. https://doi.org/10.1016/j.marpolbul.2017.03.059 Hashem S., Akter T., Salam M. A., H. M. T. (2014). Aquaculture planning through Remote Sensing Image Analysis and GIS tools in Northeast region, Bangladesh. International Journal of Fisheries and Aquatic Studies, 1(5), 127–136. www.fisheriesjournal.com Heck, J. E., Nieves, J. W., Chen, Y., Parvez, F., Brandt-Rauf, P. W., Howe, G. R., & Ahsan, H. (2010). Protein and amino acid intakes in a rural area of Bangladesh. Food and Nutrition Bulletin, 31(2), 206. https://doi.org/10.1177/156482651003100203 Hernandez, R., Belton, B., Reardon, T., Hu, C., Zhang, X., & Ahmed, A. (2018). The “quiet revolution” in the aquaculture value chain in Bangladesh. Aquaculture, 493(June 2017), 456–468. https://doi.org/10.1016/j.aquaculture.2017.06.006 139 Huda, K. M. S., Atkins, P. J., Donoghue, D. N. M., & Cox, N. J. (2010). Small water bodies in Bangladesh. Area, 42(2), 217–227. https://doi.org/10.1111/j.1475-4762.2009.00909.x Iordache, M.-D., Mantas, V., Baltazar, E., Lewyckyj, N., & Souverijns, N. (2020). Application of Random Forest Classification to Detect the Pine Wilt Disease from High Resolution Spectral Images. International Geoscience and Remote Sensing Symposium, 776026, 4489– 4492. Islam, M. M., Barman, A., Kundu, G. K., Kabir, M. A., & Paul, B. (2019). Vulnerability of inland and coastal aquaculture to climate change: Evidence from a developing country. Aquaculture and Fisheries, 4(5), 183–189. https://doi.org/10.1016/j.aaf.2019.02.007 Jahan, K., Belton, B., Ali, H., Dhar, G., & Ara, I. (2015). Aquaculture technologies in Bangladesh: An assessment of technical and economic performance and producer behavior. In WorldFish. JAXA. (2020). ALOS-2 Project / PALSAR-2. Ji, L., Zhang, L., & Wylie, B. K. (2009). Analysis of dynamic thresholds for the normalized difference water index. Photogrammetric Engineering and Remote Sensing, 75(11), 1307– 1317. https://doi.org/10.14358/PERS.75.11.1307 JianCheng, S., Yang, D., JinYang, D., LingMei, J., LinNa, C., KeBiao, M., Peng, X., WenJian, N., Chuan, X., Qiang, L., ChenZhou, L., Peng, G., Qian, C., YunQing, L., Jing, C., AnQi, W., HeJia, L., & YinHui, W. (2012). Progresses on microwave remote sensing of land surface parameters. Sci China Earth Sci, 55(7), 1052–1078. https://doi.org/10.1007/s11430- 012-4444-x Jiang, H., Feng, M., Zhu, Y., Lu, N., Huang, J., & Xiao, T. (2014). An Automated Method for Extracting Rivers and Lakes from Landsat Imagery. Remote Sensing 2014, Vol. 6, Pages 5067-5089, 6(6), 5067–5089. https://doi.org/10.3390/RS6065067 Jiao, L., Liu, Y., & Li, H. (2012). Characterizing land-use classes in remote sensing imagery by shape metrics. ISPRS Journal of Photogrammetry and Remote Sensing, 72, 46–55. https://doi.org/10.1016/J.ISPRSJPRS.2012.05.012 Kapetsky, J. M. (1987). Satellite remote sensing to locate and inventory small water bodies for fisheries management and aquaculture development in Zimbabwe. CIFA Occas. Pap. Karim, M., & Little, D. C. (2018). The impacts of integrated homestead pond-dike systems in relation to production, consumption and seasonality in central north Bangladesh. Aquaculture Research, 49(1), 313–334. https://doi.org/10.1111/are.13462 Khorram, S., Koch, F. H., van der Wiele, C. F., & Nelson, S. A. C. (2012). Introduction. In Remote Sensing (pp. 1–15). https://doi.org/10.1007/978-1-4614-3103-9_1 Khorram, S., van der Wiele, C., Koch, F., Nelson, S., & Potts, M. (2016). Principles of Applied Remote Sensing. https://doi.org/10.1007/978-3-319-22560-9 140 Kobayashi, M., Msangi, S., Batka, M., Vannuccini, S., Dey, M. M., & Anderson, J. L. (2015). Fish to 2030: The Role and Opportunity for Aquaculture. Aquaculture Economics and Management, 19(3), 282–300. https://doi.org/10.1080/13657305.2015.994240 Kubat, M. (2017). An Introduction to Machine Learning. In An Introduction to Machine Learning. Springer International Publishing. https://doi.org/10.1007/978-3-319-63913-0 Kuhn, M. (2008). Building predictive models in R using the caret package. Journal of Statistical Software, 28(5), 1–26. https://doi.org/10.18637/jss.v028.i05 Lam, M. E. (2016). The Ethics and Sustainability of Capture Fisheries and Aquaculture. Journal of Agricultural and Environmental Ethics, 29(1), 35–65. https://doi.org/10.1007/s10806- 015-9587-2 Liu, K., Ai, B., & Wang, S. (2010). Fishpond change detection based on short term time series of RADARSAT images and object-oriented method Reconstruction of green open space in perspective of social equity: assessment, mechanism and optimization View project Environment Remote sensing, View project Fish-pond Change Detection Based on Short Term Time Series of RADARSAT Images and Object-oriented Method. https://doi.org/10.1109/CISP.2010.5647643 MathWorks. (2016a). Applying Supervised Learning. MathWorks. (2016b). Applying Unsupervised Learning. MathWorks. (2016c). Getting Started with Machine Learning. MathWorks. (2016d). Introducing Machine Learning. Maxwell, A. E., Warner, T. A., & Fang, F. (2018). Implementation of machine-learning classification in remote sensing: An applied review. International Journal of Remote Sensing, 39(9), 2784–2817. https://doi.org/10.1080/01431161.2018.1433343 McFeeters, S. K. (1995). The use of the Normalized Difference Water Index (NDWI) in the delineation of open water features. Https://Doi.Org/10.1080/01431169608948714, 17(7), 1425–1432. https://doi.org/10.1080/01431169608948714 McFeeters, Stuart K. (2013). Using the Normalized Difference Water Index (NDWI) within a Geographic Information System to Detect Swimming Pools for Mosquito Abatement: A Practical Approach. Remote Sensing 2013, Vol. 5, Pages 3544-3561, 5(7), 3544–3561. https://doi.org/10.3390/RS5073544 Mohsen, A., Elshemy, M., & Zeidan, B. A. (2018). Change detection for Lake Burullus, Egypt using remote sensing and GIS approaches. Environmental Science and Pollution Research, 25(31), 30763–30771. https://doi.org/10.1007/s11356-016-8167-y Momtaz, S., & Shameem, M. I. M. (2016). The Research Setting. Experiencing Climate Change in Bangladesh, 41–68. https://doi.org/10.1016/b978-0-12-803404-0.00004-1 141 Mukherjee, N. R., & Samuel, C. (2016). Assessment of the Temporal Variations of Surface Water Bodies in and around Chennai using Landsat Imagery. Indian Journal of Science and Technology, 9(18), 1–7. https://indjst.org/articles/assessment-of-the-temporal-variations-of- surface-water-bodies-in-and-around-chennai-using-landsat-imagery NASA. (2020a). Get to Know SAR – NASA-ISRO SAR Mission (NISAR). https://nisar.jpl.nasa.gov/mission/get-to-know-sar/overview/ NASA. (2020b). SAR | NASA Earth Science Disasters Program. https://disasters.nasa.gov/instruments/sar NASA. (2020c, July 29). About Terra | Terra. https://terra.nasa.gov/about NASA EarthData. (2020). Remote Sensors | Earthdata. https://earthdata.nasa.gov/learn/remote- sensors Natural Resources Canada. (2015). Fundamentals of Remote Sensing. https://www.nrcan.gc.ca/sites/www.nrcan.gc.ca/files/earthsciences/pdf/resource/tutor/funda m/pdf/fundamentals_e.pdf NSIDC. (2020, April 3). Remote Sensing. https://nsidc.org/cryosphere/seaice/study/infrared_remote_sensing.html Otsu, N. (1979). A Threshold Selection Method from Gray-Level Histograms. IEEE Trans Syst Man Cybern, SMC-9(1), 62–66. https://doi.org/10.1109/TSMC.1979.4310076 Ottinger, M., Clauss, K., & Kuenzer, C. (2016). Aquaculture: Relevance, distribution, impacts and spatial assessments - A review. Ocean and Coastal Management, 119, 244–266. https://doi.org/10.1016/j.ocecoaman.2015.10.015 Ottinger, M., Clauss, K., & Kuenzer, C. (2017). Large-scale assessment of coastal aquaculture ponds with Sentinel-1 time series data. Remote Sensing, 9(5). https://doi.org/10.3390/rs9050440 Ottinger, M., Clauss, K., & Kuenzer, C. (2018). Opportunities and challenges for the estimation of aquaculture production based on earth observation data. Remote Sensing, 10(7). https://doi.org/10.3390/rs10071076 PBS. (2007, October). NOVA | Sputnik Declassified | What Satellites See (non-Flash) | PBS. https://www.pbs.org/wgbh/nova/sputnik/sate-nf.html Pekel, J.-F., Cottam, A., Gorelick, N., & Belward, A. S. (2016). High-resolution mapping of global surface water and its long-term changes. Nature 2016 540:7633, 540(7633), 418– 422. https://doi.org/10.1038/nature20584 Prasad, K. A., Ottinger, M., Wei, C., & Leinenkugel, P. (2019). Assessment of coastal aquaculture for India from Sentinel-1 SAR time series. Remote Sensing, 11(3). https://doi.org/10.3390/rs11030357 142 Ramezan, C. A., Warner, T. A., & Maxwell, A. E. (2019). Evaluation of sampling and cross- validation tuning strategies for regional-scale machine learning classification. Remote Sensing, 11(2). https://doi.org/10.3390/rs11020185 Rashid, S., Minot, N., & Lemma, S. (2019). Does a “Blue Revolution” help the poor? Evidence from Bangladesh. Agricultural Economics, 50(2), 139–150. https://doi.org/10.1111/agec.12472 Rebala, G., Ravi, A., & Churiwala, S. (2019). An Introduction to Machine Learning. Springer International Publishing. https://doi.org/10.1007/978-3-030-15729-6 Ren, C., Wang, Z., Zhang, Y., Zhang, B., Chen, L., Xi, Y., Xiao, X., Doughty, R. B., Liu, M., Jia, M., Mao, D., & Song, K. (2019). Rapid expansion of coastal aquaculture ponds in China from Landsat observations during 1984-2016. https://doi.org/10.1016/j.jag.2019.101902 Rhodes, C. J., Henrys, P., Siriwardena, G. M., Whittingham, M. J., & Norton, L. R. (2015). The relative value of field survey and remote sensing for biodiversity assessment. Methods in Ecology and Evolution, 6(7), 772–781. https://doi.org/10.1111/2041-210X.12385 Ross, L. G., Telfer, T. C., Falconer, L., Soto, D., & Aguilar-Manjarrez, J. (2013). Site selection and carrying capacities for inland and coastal aquaculture. In FAO Fisheries and Aquaculture Proceedings No. 21 (Issue December). http://web.a.ebscohost.com.proxy1.cl.msu.edu/ehost/ebookviewer/ebook?sid=6361a65c- 22b2-4e69-afa9-22ca9df1e735%40sessionmgr4007&vid=0&format=EB Sasaki, Y. (2007). The truth of the F-measure (Issue January 2007). Sattar, M. N. (2019). The Freshwater Aquaculture Revolution in Bangladesh: Impacts on Land, Water, and Livelihoods. Shah, M. S. (2003). Human resource development activities infisheries sector. Fish Fortnight Compendium 2003, 57–59. Shamsuzzaman, M. M., Islam, M. M., Tania, N. J., Al-Mamun, M. A., Barman, P. P., & Xu, X. (2017). Fisheries resources of Bangladesh: Present status and future direction. Aquaculture and Fisheries. https://reader.elsevier.com/reader/sd/pii/S2468550X16300260?token=D4D1183AFAF673C C35884FDC579258AC3B92C6F79612827C172DF1BC8C94C66545D953E385C931B859 370D634C2EF949&originRegion=us-east-1&originCreation=20210603182845 Shi, D., & Yang, X. (2016). An assessment of algorithmic parameters affecting image classification accuracy by random forests. Photogrammetric Engineering and Remote Sensing, 82(6), 407–417. https://doi.org/10.14358/PERS.82.6.407 Stiller, D., Ottinger, M., & Leinenkugel, P. (2019). Spatio-Temporal Patterns of Coastal Aquaculture Derived from Sentinel-1 Time Series Data and the Full Landsat Archive. https://doi.org/10.3390/rs11141707 143 Subasinghe, R. (2017). FAO Fisheries and Aquaculture Circular FIAA / C1140 ( En ) WORLD AQUACULTURE 2015 : A BRIEF OVERVIEW Fisheries and WORLD AQUACULTURE 2015 : A BRIE F OVERVIEW (Vol. 1140, Issue January). Thilsted, S. H., Thorne-Lyman, A., Webb, P., Bogard, J. R., Subasinghe, R., Phillips, M. J., & Allison, E. H. (2016). Sustaining healthy diets: The role of capture fisheries and aquaculture for improving nutrition in the post-2015 era. Food Policy, 61, 126–131. https://doi.org/10.1016/j.foodpol.2016.02.005 Travaglia, C., Profeti, G., Aguilar Manjarrez, J., & Lopez, N. A. (2004). Mapping coastal aquaculture and fisheries structures by satellite imaging radar. FAO. Fisheries Technical Paper, 459, 58. http://www.fao.org/documents/card/en/c/d8c9fd81-4655-55f9-b359- 8107c6629622/ UCSB. (2020). GeoEye-1 Gives Google Highest Resolution Imagery. https://geog.ucsb.edu/geoeye-1-gives-google-highest-resolution-imagery/ United Nations et al. (2003). Integrated Environmental and Economic Accounting 2003. US Department of Commerce, N. O. and A. A. (n.d.). What is aquaculture? USGS. (2020). Landsat 8. https://www.usgs.gov/land-resources/nli/landsat/landsat-8?qt- science_support_page_related_con=0#qt-science_support_page_related_con Virdis, S. G. P. (2014). An object-based image analysis approach for aquaculture ponds precise mapping and monitoring: A case study of Tam Giang-Cau Hai Lagoon, Vietnam. Environmental Monitoring and Assessment, 186(1), 117–133. https://doi.org/10.1007/s10661-013-3360-7 Wicaksono, A., & Wicaksono, P. (2019). Geometric accuracy assessment for shoreline derived from NDWI, MNDWI, and AWEI transformation on various coastal physical typology in Jepara Regency using Landsat 8 OLI imagery in 2018. Geoplanning: Journal of Geomatics and Planning, 6(1), 55–72. https://doi.org/10.14710/geoplanning.6.1.55-72 Woodhouse, I. H. (2017). Introduction to Microwave Remote Sensing. In Introduction to Microwave Remote Sensing. Taylor & Francis Group. https://doi.org/10.1201/9781315272573 Worden, J., & de Beurs, K. M. (2020). Surface water detection in the Caucasus. International Journal of Applied Earth Observation and Geoinformation, 91, 102159. https://doi.org/10.1016/J.JAG.2020.102159 Xu, H. (2006). Modification of normalised difference water index (NDWI) to enhance open water features in remotely sensed imagery. International Journal of Remote Sensing, 27(14), 3025–3033. https://doi.org/10.1080/01431160600589179 Yan, W., Wang, R., Shengyi, Y., & Jianping, L. (2021). Tight Reservoirs Classification using Random Forest. 3rd International Conference on Intelligent Control, Measurement and 144 Signal Processing and Intelligent Oil Field, Icmsp, 1–5. https://doi.org/101109 Yu, Q., Gong, P., Clinton, N., Biging, G., Kelly, M., & Schirokauer, D. (2006). Object-based detailed vegetation classification with airborne high spatial resolution remote sensing imagery. Photogrammetric Engineering and Remote Sensing, 72(7), 799–811. https://doi.org/10.14358/PERS.72.7.799 Yu, Z. (2019, April 4). Fish Ponds Detection in Bangladesh on Google Earth Engine: A case study of Singra Upazila. Yu, Z., Di, L., Rahman, M. S., & Tang, J. (2020). Fishpond mapping by spectral and spatial- based filtering on google earth engine: A case study in singra upazila of Bangladesh. Remote Sensing, 12(17). https://doi.org/10.3390/RS12172692 Yu, Z., Di, L., Tang, J., Zhang, C., Lin, L., Yu, E. G., Rahman, M. S., Gaigalas, J., & Sun, Z. (2018, September 27). Land use and land cover classification for Bangladesh 2005 on google earth engine. International Conference on Agro-Geoinformatics. https://doi.org/10.1109/Agro-Geoinformatics.2018.8475976 Zeng, Z., Wang, D., Tan, W., & Huang, J. (2019). Extracting aquaculture ponds from natural water surfaces around inland lakes on medium resolution multispectral images. International Journal of Applied Earth Observation and Geoinformation, 80, 13–25. https://doi.org/10.1016/j.jag.2019.03.019 Zhang, J., Jin, R., Yang, Y., & Hauptmann, A. G. (2003). Modified Logistic Regression: An Approximation to SVM and Its Applications in Large-Scale Text Categorization. Twentieth International Conference on Machine Learning. Zhou, Y., Dong, J., Xiao, X., Xiao, T., Yang, Z., Zhao, G., Zou, Z., & Qin, Y. (2017). Open Surface Water Mapping Algorithms: A Comparison of Water-Related Spectral Indices and Sensors. Water 2017, Vol. 9, Page 256, 9(4), 256. https://doi.org/10.3390/W9040256 Zoran, L. F. V., Ionescu Golovanov, C., & Zoran, M. A. (2010). Spectral characteristics and feature selection of satellite remote sensing data for land use/cover changes assessment in the Romanian northwestern Black Sea coastal area. Earth Resources and Environmental Remote Sensing/GIS Applications, 7831, 78311P. https://doi.org/10.1117/12.864793 145