DEVELOPING ACTIVITY DATA FROM REMOTE SENSING FOR REDD+ MONITORING IN TROPICAL MIOMBO WOODLANDS By Daniel J. Zelenak A THESIS Submitted to Michigan State University in partial fulfillment of the requirements for the degree of Forestry Master of Science 20 18 ABSTRACT DEVELOPING ACTIVITY DATA FROM REMOTE SENSING FOR REDD+ MONITORING IN TROPICAL MIOMBO WOODLANDS By Daniel J. Zelenak The country of Malawi is experiencing some of the highest rates of deforestation in Africa. Of a unique type of dry woodland that is made up from a mosaic of upper and lower canopy species. Miombo woo dlands are threatened by expanding agriculture and demand for fuel wood, which is the primary energy source for the majority of Efforts to map areas of Miombo degradation can be assisted by remote sensing technologies and data sets, w hich can be much more cost efficient than in - situ measurements. An effort was made to map the locations of deforestation and degradation of - series Landsat data. NDVI and fractional cover were used to quantify t he presence of vegetation throughout time, and a change algorithm is used to identify areas of vegetation loss. This information can be used to develop activity data which guides REDD+ efforts in the country of Malawi. iii This thesis is dedicated to my family. Thank you for your unwavering support. iv ACKNOWLEDGEMENTS The work of th is thesis was made possible by funding from the Malawi PERFORM project and also by support from the Global Observatory for Ecosystem Services Lab at Michigan State University Department of Forestry. v TABLE OF CONTENTS LIST OF TABLES ............................................ .............................................................................. vi LIST OF FIGURES ................................ ...................................................................................... viii KEY TO ABBREVIATIONS .......................................................................................................... x Chapter 1 : I ntroduction ........................................................................................ . ........................... 1 I. Remote Sensing of Forests and W oodlands ..... .. ......................................... . . ............................ 1 II. Climate Change, REDD+, and Malawi ................................................ ......... . .......................... 4 Chapter 2: Methods .................. .. ................................................................................................... . 11 I: Using Landsat Medium Resolution Data . ................................... . .................. . ........................ 11 II: Data Acquisition for Malawi........ ......................................................................................... 13 III: Workflow . .. .......................................................................................................................... 19 Chapter 3: Results ............. ......... .................................................................... . ............................... 26 I. National Multispectral Mosaics ............................................................................................. 26 II. National NDVI Mosaics ..... ... ......... ....................................................................................... 30 III. National f C Datasets ..... ........... ................................................................ . ................... . ........ 35 IV. Accuracy Assessment Using High - Resolution Imagery . .... . ..................... . ................. . ........ 42 V. Accuracy Assessment of the Landsat - based f C Product . .............................................. . ........ 49 VI. Change Detection Algorithm ......................................................................... . ...................... 54 VII: Change Detection Algorithm Results .... .. ................................................... . ........................ 56 VIII: Validation of Change Detection Algorithm Results .................. . .............. . ........................ 59 IX: Discussion and Conclusion ................................ . ................................................................. 61 BIBLIOGRAPHY .......................................................................................................................... 65 vi LIST OF TABLES Table 1: Descriptive information for the Landsat satellites used for this study ........................... 11 Table 2: Landsat scenes used to create the 1998 - 2002 mosaic. ................................ .................... 15 Table 3: Landsat scenes used to create the 2008 - 2010 mosaic. ................................ .................... 16 Table 4: Landsat scenes used to create the 2014 - 2015 mosaic. ................................ .................... 17 Table 5: High resolution satellite images used for the purpose of accuracy assessment of the fractional cover product ................................ ................................ ................................ ................ 42 Table 6: The confusion matrix generated for the Perekezi GeoEye - 1 high resolution vegetation/non - vegetation thematic layer. ................................ ................................ .................... 44 Table 7: Accuracy totals for the Perekezi GeoEye - 1 high resolution vegetation/non - vegetation thematic layer. ................................ ................................ ................................ ............................... 44 Table 8: Conditional kappa statistics for the individual classes, as well as the overall kappa statistic for the Perekezi GeoEye - 1 classification. ................................ ................................ ........ 45 Table 9 : The confusion matrix generated for the Liwonde Worldview - 3 high resolution vegetation/non - vegetation thematic layer. ................................ ................................ .................... 45 Table 10: Accuracy totals for the Liwonde Worldview - 3 high resolution vegetation/non - vegetation thematic layer. ................................ ................................ ................................ ............. 45 Table 11: Conditional kappa statistics for the individual classes, as well as the overall kappa statistic for the Liwonde Worldview - 3 classification. ................................ ................................ .. 46 Table 12: The confusion matrix generated for the Liwonde Worldview - 2 high resolution vegetation/non - vegetation thematic layer. ................................ ................................ .................... 46 Table 13: Accuracy totals for the Liwonde Worldview - 2 high resolution vegetation/non - vegetation thematic layer. ................................ ................................ ................................ ............. 46 Table 14: Conditional kappa statistics for the individual classes, as well as the overall kappa statistic for the Liwonde Worldview - 2 classification. ................................ ................................ .. 47 vii Table 15: Perekezi SLR results. Simple linear regression results from the comparison of the Landsat derived fC (the independent variable) and the high resolution GeoEye - 1 fC (the dependent variable). ................................ ................................ ................................ ...................... 51 Table 16: Liwonde (West) SLR results. Simple linear regression results from the comparison of the Landsat derived fC (the independent variable) and the high resolution Worldview - 3 fC (the dependent variable). ................................ ................................ ................................ ...................... 52 Table 17: Liwonde (East) SLR results. Simple linear regression results from the comparison of the Landsat derived fC (the independent variable) and the high - resolution Worldview - 2 fC (the dependent va riable). ................................ ................................ ................................ ...................... 53 Table 18: 2002 to 2010 change results. Area in hectares of each change category for the time period from 2002 to 2010. The percent of total forest area for each change category is also shown. ................................ ................................ ................................ ................................ ........... 57 Table 19: 2010 to 2015 change results. Area in hectares of each change category for the time period from 2010 to 2015. The percent of total forest area for each change category is also shown. ................................ ................................ ................................ ................................ ........... 57 Table 20: Confusion matrix for change detection. The confusion matrix that was created as part of the accuracy assessment of the change algorithm. The rows represent the classified change algorithm product, and the columns represent the ground - truth reference totals. The diagonal of the table represents correctly labeled change classes. ................................ ................................ .. 60 Table 21: Accuracy totals for change detection. The accuracy totals are shown for the producer's, user's, and overall accuracies. ................................ ................................ ..................... 60 Table 22: Kappa (K^) statistics for change detection. Conditional kappa values calculated for each individual class as well as the kappa value for the overall classification. ............................ 61 viii LIST OF FIGURES Figure 1: Malawi Path/Row Scenes. Malawi is covered by 11 nominal WRS - 2 path/row scenes which are shown here overlaid onto the Malawi international border for reference. ................... 18 Figure 2: Flowchart showing the workflow used to generate the national - scale datasets and change results ................................ ................................ ................................ ................................ 25 Figure 3: The 1998 - 2002 multispectral mosaic. The image is comprised of TM and ETM+ images stitched together using an averag ing function for overlap areas and path to path histogram matching to reduce seamlines. ................................ ................................ ................................ ...... 27 Figure 4: The 2008 - 2011 multispectral mosaic is composed almost entirely of TM imagery due to the SLC - off error on the ETM+ sensor for this time period. One ETM+ SLC - off image was used for gap - filling path/row 167/72. An averaging function was applied to overlap areas and no histogram match ing was required. ................................ ................................ ................................ 28 Figure 5: The 2014 - 2015 multispectral mosaic is composed entirely of OLI imagery. An averaging function wa s used for overlap areas and no histogram matching was required. .......... 29 Figure 6: The national NDVI mosaic for the 1998 - 2002 time period. ................................ ......... 32 Figure 7: The national NDVI mosaic for the 2008 - 2010 time period. ................................ ......... 33 Figure 8: The national NDVI mosaic for the 2014 - 2015 time period. ................................ ......... 34 Figure 9: The national fC dataset for the time period of 1998 - 2002. ................................ ........... 39 Figure 10: The national fC dataset for the time period of 2008 - 2010. ................................ ......... 40 Figure 11: The national fC dataset for the time period 2014 - 2015. ................................ .............. 41 Figure 12: Perekezi GeoEye - 1 Imagery. The multispectral image is on the left with the random sample points overlaid. The thematic vegetation/non - vegetation image is on the right. ............. 47 Figure 13: Liwonde Worldview - 3 Imagery. The multispectral image is on the left with the random sample points overlaid. The thematic vegetation/non - vegetation image is on the right. 48 Figure 14: Liwonde Worldview - 2 Imagery. The multispectral image is on the le ft with the random sample points overlaid. The thematic vegetation/non - vegetation image is on the right. 48 ix Figure 15: Comparison of the same spatial area shown in both the 30m fC product derived from Landsat data (left image) and the 0.5m thematic classification of vegetation and non - vegetation (right image). ................................ ................................ ................................ ................................ . 50 Figure 16: Perekezi regression of Landsat f C and HiRes f C. Simple linear regression plot between the Landsat f C (independent variable) and the hi - resolution f C (dependent variable) using t he Perekezi Geoeye - 1 imagery. ................................ ................................ .......................... 51 Figure 17: Liwonde (West) regression of Landsat f C and HiRes f C. Simple linear regression plot between the Landsat f C (independent variable) and the hi - resolution f C (dependent variable) using the Liwonde Worldview - 3 imagery. ................................ ................................ ................... 52 Figure 18: Liwonde (east) regression of Landsat f C and HiRes f C. Simple linear regression plot between the Landsat f C (independent variable) and the hi - resolution f C (dependent variable) using the Liwonde Worldview - 2 imagery. ................................ ................................ ................... 53 Figure 19: ERDAS graphical model used to carry out the change detection between two fC datasets. ................................ ................................ ................................ ................................ ......... 55 Figure 20: The change algorithm results are shown here for the two lengths of time 2002 - 2010 and 2010 - 2015. The types of change that occurred in each period are color coded. Inset A shows the detail area around Perekezi Forest Reserve which includes a western portion of the South Viphy a Forest Reserve. Inset B shows the detailed area around Ntchisi Forest Reserve. Inset C shows the detailed area around Liwonde Forest Reserve, which also includes the southern portion of Liwonde National Park and the northern portion of the Zomba - Malosa Forest. All protected forests are indicated by the green areas on the main map and insets. ........ 58 x KEY TO ABBREVIATIONS 6S: Second Simulation of the Satellite Signal in the Solar Spectrum AFOLU: Agriculture, Forestry and Other Land Use AOT: Aerosol Optical Thickness AVHRR: Advanced Very High Resolution Radiometer DN: Digital N umber DOS: Dark Object Subtraction DOY: Day of Y ear EROS: Earth Resources Observation and Science Center ETM+: Enhanced Thematic Mapper Plus FOLU: Forestry and Other Land Use f C: Fractional Cover GHG: Greenhouse Gas GLOVIS: Global Visualization Tool GtCO 2 eq/yr: Gigatons of Carbon Dioxide Equivalents per Year Ha: Hectare IPCC: Intergovernmental Panel on Climate Change LUC: Land - use Change LULCC: Land - use and Land Cover Change LiDAR: Light D etection and R anging MODIS: Moderate Resolution Imaging Spectroradiometer MSAVI - 2: Modified Soil Adjusted Vegetation Index 2 NDVI: Normalized Difference Vegetation Index xi NIR: Near I nfrared OLI: Operational Land Imager PERFO RM: Protecting Ecosystems and Restoring Forests in Malawi SWIR: Short - wave I nfrared TM: Thematic Mapper TOA: Top of A tmosphere UNFAO: Food and Agriculture Organization of the United Nations USGS: United States Geological Survey WRS - 2: World Reference System 2 1 Ch apter 1: Introduction I. Remote Sensing of Forests and Woodlands Remote sensing is the science of collecting observational data on a phenomena using an instrument which is not in direct contact with the phenomena being observed. Since the early emphasis has been focused on monitoring the land surface. Such instru ments have been deployed in aircraft or on earth orbiting satellites and have use d active and passive sources of energy for detection and measurement of surface properties. Some forms of active remote sensing include RADAR and SONAR, in which the instrume nt provides the source of energy that is being emitted to interact with a target surface and measured as a reflected signal. Passive remote sensing refers to instruments that use a natural source of energy to collect information about a target. For examp le, a satellite that measures the radiant energy from solar output reflected from the earth surface is considered a passive instrument ; such instrument s measure the interactions of the energy with the surface of the earth . The radiant energy, or electromagnetic energy , that strikes the earth is absorbed or reflected by differen t surface objects to varying degrees according a specific spectral response of the surface material , allowing measurements of the surface properties across different parts o f the electromagnetic spectrum. This allows materials to be identified based on their unique spectral signatures (Lillesand, Kiefer, and Chipman 2008) . One of the most important earth observations is that of land used and land cover change (LULCC) monito ring using optical sensors on polar orbiting satellites, such as the Landsat series which has been continuously operating through a series of platforms since 1973. 2 The use of remote sensing to identify different types of land use and land - cover as well as monitor change has been well established as a body of theory and application in the remote sensing technical communities. One of the great advantage s of routine repeat observations from orbital systems is the ability to consistently monitor change by comparing imagery throughout time for the same areas of the earth. As such s atellite remote sensing has also been demonstrated as an effective means for determining the extent of deforestation and forest degradation , which are forms of LULCC (Goetz and Dubaya 2011 ; Skole and Tucker 1993 ) . Forests are of particular interest in the field of LULCC because they play a significant role in many important environmental processes and problems, such as the global carbon cycle (Bonan 2008), surface - c limate interactions , quantification of richness of biodiversity, and information on human systems including livelihood s and food security ( UN FAO 2010) . Much work has been done using satellite data for measuring forest cover and forest cover loss in the tropics (Broich, et al. 2011; Broich, et al. 2011; Hansen, et al. 2008; Hansen, Stehman, et al. 2008; Hansen et al. 2003; Hansen , et al. 2010; Hansen et al. 2009; Navratil 2013; Skole and Tucker 1993; Matricardi et al. 2007; Matricardi et al. 2010) . The abundanc e of research focused on tropical deforestation shows how important the tropical forest ecosystems are to the rest of the biosphere , and it also highlights how serious the problem of forest loss is in these regions . Although far less studied than closed canopy tropical forests, LULCC in tropical woodlands and s a vannas around the world is also an important agent of the global environment . Th ere is particular concern for the woodlands of Africa, especially the Miombo woodlands and the rich biodiversity that they contain . These ecosystems cover approximately 2.7 million km 2 and constitute the largest area of tropical seasonal woodlands globally (Frost 1996) . However, as elsewhere (e.g. South America ) , m uch of the remote sensing monitoring work in Africa has 3 focused on tropical rainforests (e.g. Matthew C. Hansen, Roy, et al. 2008) . Miombo woodlands have not been studied as extensively with remote sensing, but some country - specific work has been done to examine the potential conversion land supporting Miombo . T he woodlands referred to as Miombo are a mosaic of over 8,500 different species . The most prevalent canopy species are from the family Fabaceae, subfamily Caesalpinioideae, and include the genera Brachystegia , Julbern ardia , and Isoberlinia . While Miombo can occur as a closed - canopy woodland, most places in which it occurs are open canopy systems with an abundance of under - canopy flora that thrive due to the ability of sunlight to pass through the upper canopy layer . The geographic distribution of Miombo in Africa includes the region north of the E quator extending south of the Tropic of Capricorn. This includes Tanzania, the Democratic Republic of the Congo, Angola, Zambia, Malawi, Zimbabwe, and Mozambique. A distinc tion can be made between wet and dry Miombo, which is based upon the amount of rainfall that a particular region receives. However, all Miombo is characterized by a warm seasonal wet season (Frost 1996; Ribeiro et al. 2012; Desanker et al. 1997) . Miombo experiences seasonal variation which makes certain times of the year more ideal for remote sensing observations. There are 5 distinguishable phenology periods: a warm and dry pre - rain season in September and October, an early rainy seas on in November and December, a mid - rainy season in January and February, a late rainy season in March and April, and a cool earl y dry season from May to August . The majority of Miombo species lose their leaves in the dry season from August to October. New leaves begin for ming in the warm and dry period before the rainy season begins (Chidumayo and Frost 1996) . An ideal time to use remote sensing for monitoring Miombo is during the early dry season before significant leaf - loss has begun (Hudak and Wessman 2000). Furthermo re, the minimum canopy cover for a land area to 4 be defined as a forest is 10% and must cover a minimum of 0.5 hectares according to a definition used by the United Nations Food and Agriculture Organization (UNFAO 2015) . U sing leaf - on imagery makes it poss ible to detect forest cover at the Landsat resolution of 30 - meters. II. Climate Change, REDD+, and Malawi Human activities , such as the burn ing of fossil fuels , cause the release of carbon dioxide (CO 2 ) to the atmosphere. CO 2 is one of the gaseous form s that carbon takes in the global carbon cycle . The global carbon cycle includes process es that cy c le carbon between biological and geochemical forms: carbon enters the biosphere via primary production (i.e. photosynthesis) and is transferred to the hydrosphere and atmosphere via chemical reactions or physical processes . Respiration returns carbon to the atmosphere. Removals from the atmosphere to the oceans occur through both biological production and physical - chemical diffusion. The oceans return ca rbon to the atmosphere through biological and physical - chemical process. CO 2 is a GHG because its molecular structure readily absorbs radiation and r adiates heat energy would otherwise escape into space . This is the greenhouse effect, which as a naturally occurring process is important because this trapped heat allow s the earth to be warm enough to support life . CO 2 is just one of several radiatively - active gasses (e.g. chlorofluorocarbons, methane (CH 4 ) and nitrous oxide (N 2 O) ) that are released by human activities. CO 2 is an important GHG when discussing the human role in climate change because it is capable of remaining in the atmosphere for long periods of time , which allow s for a net accumulation if total emissions of CO 2 are greater than sinks over time. Studies have shown that the burning of fossil fuels has caused a steady and quantifiable increase in the concentration of atmospheric CO 2 since the industrial revolution (Keeling et al. 1976; IPCC 1990); f rom 1750 to 2011 there w as approximately 2,040 5 GtCO 2 in emissions, of which 40% still remains in the atmosphere (IPCC 2014a). GHGs are known to be responsible for the majority of the observed increase in global average surface temperatures by 0.5 °C from 1951 to 2010. (IPCC 2014a; IPCC 2014b) . Deforestation and forest degradation are significant contributors to the total amount of anthropogenic GHG emissions. These forms of LULCC represent emissions of CO 2 because they diminish or entirely remove the capacity of the la nd to assimilate carbon from the atmosphere via primary production , and they reduce the overall stock of carbon stored on land . When stocks of carbon in ecosystems are reduced by deforestation and degradation, carbon in transferred to the atmosphere throu gh biological decomposition or rapid oxidation (e.g. fires). Regeneration of ecosystems provide removals of carbon from the atmosphere when stocks of carbon increase. Coupled with increasing amounts of CO 2 emissions from burning fossil fuels, these forms o f LULCC represent a net flux of CO 2 to the atmosphere. Globally the Agriculture, Forestry, and Other Land Use (AFOLU) sector contribute d approximately 10 - 12 GtCO 2 eq/yr (Gigatons of CO 2 equivalents per year) between 1970 and 2010 representing about a quarter of all anthropogenic GHG emissions (Smith et al. 2014) . Of th ese emissions , a significant portion originated from the sub - sector of Forestry and Other Land Use (FOLU) which includes Land Use Change (LUC , also referred to as land use and l and - cover change using the abbreviation LULCC ) . It is estimated that deforestation and degradation of forests contribute d about 12% of all anthropogenic GHG emissions between 1980 and 2010 (van der Werf et al. 2009). It is widely accepted in the scientific community that avoid ing the most catastrophic repercussions of climate change will require keeping the increase in global surface temperature GHG emiss ions), the increase in temperature will likely exceed 4°C by the year 2100. 6 Therefore, any effective climate change mitigation policies and measures must include mechanisms to actively reduc e emissions and increas e sinks in the AFOLU sector, which must fo cus on reducing deforestation and forest degradation. Remote sensing presents an efficient and accurate means to detect deforestation and forest degradation over large areas. In many cases this allows for monitoring of forests at a scale that would be impractical with ground - based observations and measurements. The UNFAO definition of forest requires a minimum 10% canopy cover and a minimum area of 0.5 Ha , with trees reaching a height of 5 meters or greater . Deforestation is defined as the conversion of forest to some other form of land use, which includes a permanent reduction in canopy cover to below 10%. Forest degradation is more difficult to define succinctly ; it involves long term reduction of canopy cover due to human or natural ac tivities but keeps the requirements for land cover classification of forest (i.e. the canopy cover remains above 10%). Forest degradation also implies a long - term reduction in the capacity of a forest to sustain biodiversity, provide resources, and seques ter carbon from the atmosphere (UNFAO 2015) . The need to accurately measure activity data as related to forest clearing and degradation is essential for the implementation of REDD+ (Reduced Emissions from Deforestation and and enhancement of carbon stocks) (Corbera and Schroeder 2011). The IPCC defines activity data as a magnitude of the emissions or removals due to some human activity in a particular sector over a certain period of time. In the forestry sector, activity data may refer to measurements of area or area change of forested land. Ov erall emissions are calculated by multiplying an emission factor coefficient by the activity data. An emission factor coefficient 7 therefore is some value which relates the activity data to a physical measurement of emissions (Houghton et al. 1997). Miombo woodlands are the dominant forest type in Malawi. They are an ecosystem of particular concern due to recent increased rates of deforestation, which in turn drives GHG emissions as well as biodiversity loss. In the country of Malawi, Miombo woodlands are essential to the local population as a source of domestic energy (Abbot and Homewood 1999). Th e demand for fuelwood has driven encroachment of wood harvesting forests as other sources have become depleted. A s local sources become depleted, those responsible for collecting fuel wood for their households must travel increasingly farther abroad into the wilderness to find good quality wood (Brouwer, Hoorweg, and van Liere 1997). In addition to mitigating greenho use gases, forests and woodlands in Malawi play an essential role in water quality and soil retention. Expanding agriculture contributes to increased run - off, which in turn increases the amount of suspended solids and nutrients entering water bodies . Stu dies have been done using remote sensing products to quantify the relationship between conversion of Miombo to negative impacts on water supplies. To provide improved information for land management practices in the Upper Shire River Catchment Palamuleni e t al used Landsat data to map land cover change. Imagery from 1989 and 2002 was processed using a maximum likelihood classifier to identify land cover types spatially. The authors used these data to show that cultivated land increased while savanna, shru bs, and woodlands showed a decrease in cover area. These changes in land cover were found to directly contribute to negative water impacts in the Upper Shire river catchment, such as increased run - off and degradation of ground - water supplies during the dr y season (Palamuleni et al. 2011). 8 In Malawi, d during this time period. A study using AVHRR and MODIS NDVI to determine which land cover types were b eing converted to agriculture in the time period of (Chavula, Brezonik, and Bauer 2011). The study found that savanna and woodland experienced significant loss due to agricultural expansion from 1982 - 1995, but then a slight increase in this land cover type was observed from 1995 to 2005. They did concede the caveat of comparing AVHRR and MODIS due to the difference in resolutions (8km resolution for AVHRR compared to 500m resoluti on for MODIS). These large differences in resolution make it difficult to make meaningful comparison in vegetation cover at the smaller MODIS scale. Other remote sensing work has been done to map land cover for the entire country of Malawi. One such study compared the efforts of 4 separate methodologies carried out by different teams to map land cover and land use change in Malawi to assist with greenhouse gas reporting (Haack, Mahabir, and Kerkering 2015). This review showed that e ach of the individ ual assessments produced different results in land cover change, due to differences in methods. This highlights the challenge of developing significant and reproducible remote sensing analysis of forest cover in Malawi. Hudak and Wessman used NDVI derived from 4 Landsat MSS imagery and field measurements to estimate changes in woody canopy cover from the satellite data. A mean NDVI value from one of the dates was used to normalize NDVI for the other imagery to enable meaningful comparison between the diffe rent images . Overall, they found a 1.8% reduction in woody canopy between 1981 and 1992 (Hudak and Wessman 2000). 9 There has been very l ittle work on measuring degradation of open canopy woodlands and specifically Miombo woodlands in Malawi at the national scale. Most of the remote sensing work to date has focused on closed canopy forests ; this work has been shown to be both accurate and e ffective for detecting change in these types of forests . Open canopy woodlan ds are more difficult to observe change in due to the presence of background soil signal in remote sensing data and other factors such as seasonality (e.g. differences in canopy during wet vs. dry seasons) . Detecting change in open canopy woodlands is vit al to REDD+ efforts which require activity data for these regions, especially in countries such as Malawi whose forest lands are dominated by this type of vegetation. Furthermore, detecting change within open canopy woodlands in which the land cover itsel f does not change is also important for developing activity data. This change is referred to as forest degradation and in countries such as Malawi represents a decrease why it is vital to national REDD+ efforts . The aim of the research of this thesis has been to use Landsat data to demonstrate the feasibility of mapping of degradation and deforestation in sparse nation REDD+ Program. Furthermore, this research seeks to demonstrate a technique for detecting forest degradation at the national level, which could be deployed by nations using freely available data and technologies . RED D + efforts in Malawi will be be tter informed by an accurate method to provide reliable woodland degradation monitoring and mapping . Th ese methods could then inform efforts to produce national estimates of emissions and removals from a ctivity d ata, as one component of a National Forest Monitoring System . The goals of this thesis are therefore to provide the technical groundwork for conducting analysis of degradation of Miombo woodlands, demonstrate the usefulness of using freely available Landsat imagery as the source of analysis 10 data, and to provide spatial locations and area estimates of degraded or deforested Miombo in Malawi . The techniques and data produced by this study will provide actionable scientific data and methodologies for the Malawi PERFORM project, which is a mechanism f or assisting Malawi with preparing for REDD+. The areas of interest for this study are specifically those gazetted as protected forest lands in the country of Malawi. 11 Chapter 2 : Methods I: Using Landsat Medium Resolution Data The Landsat data archive provides an invaluable source of earth observations which are ideal for studying land cover change for various reasons. Landsat satellites have acquired i magery for most of the earth on a regular basis from 1972 up to the present day. Imagery is collected and made available to the public free of charge via the USGS Earth Resources Observation and Science Center (EROS) (Wulder et al. 2012; Woodcock et al. 2008) . All Landsat data used in this study were acquired from the USGS EROS data archive using the Global Visualization Viewer (GLOVIS) Tool (http://glovis.usgs.gov). Landsat 5 Landsat 7 Landsat 8 Sensor Thematic Mapper Enhanced Thematic Mapper Plus Operational Land Imager and Thermal Infrared Sensor Beginning Date of Operation March 1984 April 1999 April 2013 End Date of Operation November 2011 Still in Operation 1 Still in Operation Repeat Time 16 days 16 days 2 16 days 2 Table 1 : Descript ive information for the Landsat satellites used for this study . 1 Landsat 7 ETM + imagery collected since May 2013 has gaps of missing data due to the failure of an on - board mechanism responsible for correcting the forward motion of the spacecraft . 2 The repeat times for Landsat 8 and Landsat 7 are 16 days, but they are in a staggered orbit allowing for a scene to be visited every 8 days The Thematic Mapper (TM) senso r was deployed on Landsat 5 in March 1984, and was operational until November 2011. The Enhanced Thematic Mapper Plus ( ETM+ ) sensor was deployed on Landsat 7 (Landsat 6 was launched in October 1993 but failed to reach orbit) in April 1999. On May 31, 2003 the scan - line corrector (SLC) on Landsat 7 , which adjusted the 12 instrument for the f orward motion of the spacecraft, experienced a failure that caus es all imagery collected since to have missing strips of data . Although the data collected after the SLC failure is still radiometrically useful, it is difficult to use in change studies due to the signi ficant areas of imagery missing (Wulder et al. 2008; Williams, Goward, and Arvidson 2006) . For this reason - avoided when possible . Most recently , the Operational Land Imager ( OLI ) deployed on - board the Landsat 8 platfo rm was launched in February 2013 and is fully operational (Irons, Dwyer, and Barsi 2012) . The spatial resolutions for the TM sensor , ETM+ sensor , and OLI sensor are all 30 meters. This medium - scale resolution provides a consistent spatial measurement over time that is adequate for landscape - scale studies (Woodcock and Strahler 1987; Skole and Tucker 1993; Matthew C. Hansen and Loveland 2012) . Global coverage is p rovided by Landsat instruments on a 16 - day period, so any given area of the earth is imaged once every 16 days by a single sensor , with the Landsat 7 ETM+ and Landsat 8 OLI sensors being the two that are currently active . Furthermore, Landsat 7 and Landsa t 8 orbits are staggered by 8 days meaning that any area of the earth is imaged once every 8 days by these two platforms . Therefore, the temporal resolution of 16 days and in some cases 8 days when the missing scan - line corrector data are acceptable makes the Landsat data archive incredibly useful for earth monitoring . The radiometric resolution of the sensors refers to the bit - depth which in turn gives an idea of how detailed the imagery is , or how many unique gray - levels are present in a single ban d of the image . TM and ETM+ s ensors collect data at 8 bits meaning that each band can have a maximum of 256 gray levels , while OLI collects data at 12 - bits giving it 4096 maximum gray levels per band . The 12 - bit imagery is ideal due to its greater signal - to - noise ratio which is a 13 result of the greater bit depth , but data from the OLI is only available since 2013, which is only part of the historical record of this study. T he spectral resolution of the TM, ETM+, and OLI sensors provide coverage in the vi sible, near - infrared (NIR), and short - wave infrared (SWIR) wavelengths . Each Landsat platform contains a thermal infrared imaging sensor, but these products are not used in this study . The visible, NIR, and SWIR portions of the electromagnetic spectrum a re extremely useful for detecting different types of land cover (cf. vegetation, water, solid earth, snow and ice) . T his study focus es on vegetation cover , and the visible and NIR portions of the spectrum provide the most useful information and the coverage provided by all Landsat instruments is ideal for this purpose. Furthermore, radiometric calibrations have been well documented and this allow s for the comparative analyses of imagery acquired from different Landsat platforms. The consistency of d ata acquisition and the ability to perform radiometric calibration allows for comparisons of imagery to accurately detect changes over space and time . We use Landsat data for a specific application that involves measuring a parameter referred to as veget ation fractional cover the percentage of canopy vegetation to bare substrate within a given pixel ( Matricardi et al. 2007; Matricardi et al. 2010; Matricardi et al. 2013 ) . The readily - available medium resolution satellite data from Landsat sensors represents a well - known and widely used tool for measuring forest cover change and also forest carbon stocks. II: Data Acquisition for Malawi The country of Malawi is covered by 11 nominal WRS - 2 Landsat scenes from three paths and 6 rows of the Landsat World Reference System (Figure 1 ) . Landsat imagery was acquired for each of the path/row combinations covering Malawi, resulting in full national c overage at three different general time periods 2000 ( + 2), 200 9 (+ 1 ) and 2015 ( + 1) (Table s 2, 3, & 4) ) . In 14 order to perform a change analysis of Miombo woodlands, three epochs were selected for which to compare the fraction of vegetation cover over time. The use of epochal datasets comprising imagery that spans + 2 - 3 years rather than single date acquisitions was necessary. Th ere are a few reasons for this: cloud cover limits the availability of usable imagery at any given time, the coverage provided by Landsat sensors is periodic and clouds, haze, or smoke may sev erely limit the number of usable acquisitions, and in some instances scenes may not be available in the archive if they were not delivered to EROS from their original ground receiving stations. Three mosaic datasets were created covering the early dry sea son phenological period to study degradation and deforestation on the national scale for Malawi. This early phenological period of the dry season lasts from approximately May through the end of August . However, leaf senescence in some locations begins as early as July and reaches a maximum in August, so we target acquisition s between May and June. I magery w ere selected with the aim of using the most temporally similar day - of - year (DOY) sets. All Landsat data w ere projected to the WGS 84 datum and UTM Zo ne 36 South projection, and clipped using a national boundary shapefile. This boundary shapefile was obtained from the Malawi Spatial Data Platform website (MASDAP 2013) . Images w ere organized in terms of individual paths, and the country of Malawi is cov ered by three WRS - 2 paths: 167, 168 and 169. Paths correspond to the North - South direction of the satellite in orbit and therefore represent data that were collected in a single flyby. This means that scenes within an individual path can be collected on the same day of year with north and south adjacent scenes being acquired sequentially. This is advantageous because th ese data can be more seamlessly mosaicked and processed as a single path dataset due to the uniformity of atmospheric conditions and sun - earth - sensor geometry. 15 Table 2 : Landsat scenes used to create the 1998 - 2002 mosaic. d Landsat 5, and onboard Landsat 8. The next 6 digits represent the path and row in the format PPPRRR. The next 4 digits represent the year, and the next 3 dig its represent the Julian day of year. The final characters in the image ID represent the receiving station and archive version number. Path/Row Image ID Acquisition Date 169/67 LE71690672002128SGS00 5/8/2002 169/68 LE71690682002128SGS00 5/8/2002 169/69 LE71690692002128SGS00 5/8/2002 169/70 LE71690702002128SGS00 5/8/2002 169/70 LE71690702001125SGS00 5/5/2001 168/68 LT51680681998134JSA00 5/14/1998 168/68 LE71680682002153SGS00 6/2/2002 168/68 LT51680682001126JSA00 6/6/2001 168/69 LT51680691998134JSA00 5/14/1998 168/70 LT51680701998134JSA00 5/14/1998 168/71 LT51680711998134JSA00 5/14/1998 167/70 LE71670702002146SGS00 5/26/2002 167/71 LE71670712002146SGS00 5/26/2002 167/72 LE71670722002146SGS00 5/26/2002 16 Path/Row Image ID Acquisition Date 169/67 LT51690672009155JSA02 6/4/2009 169/68 LT51690682009155JSA02 6/4/2009 169/69 LT51690692009155JSA02 6/4/2009 169/70 LT51690702009155JSA02 6/4/2009 168/68 LT51680682009164MLK00 6/13/2009 168/68 LT51680682008146JSA00 5/25/2008 168/68 LT51680682009148JSA02 5/28/2009 168/69 LT51680692009164MLK00 6/13/2009 168/69 LT51680692008178JSA00 6/26/2008 168/69 LT51680692009148JSA02 5/28/2009 168/69 LT51680692009180JSA02 6/29/2009 168/70 LT51680702009164MLK00 6/13/2009 168/70 LT51680702009148JSA02 5/28/2009 168/70 LT51680702009180JSA02 6/29/2009 168/71 LT51680712009148JSA02 5/28/2009 168/71 LT51680712008130JSA00 5/9/2008 167/70 LT51670702008155JSA01 6/3/2008 167/70 LT51670702009173JSA03 6/22/2009 167/71 LT51670712009157JSA02 6/6/2009 167/72 LT51670722008139MLK00 5/18/2008 167/72 LT51670722010128JSA00 5/8/2010 167/72 LT51670722008123JSA00 5/2/2008 167/72 LE71670722009149ASN00 5/29/2009 Table 3 : Landsat scenes used to create the 2008 - 2010 mosaic. 17 Table 4 : Landsat scenes used to create the 2014 - 2015 mosaic. Paths adjacent to the east and west correspond to different flyby dates. Therefore the imagery of adjacent paths across a given East - West row will be slightly different in terms of reflectance due to differences in atmospheric conditions, climate conditio ns such as temperature and precipitation, and sun - earth - sensor geometry. Normalization of these differences was necessary to reduce patch - work results in the mosaic datasets . Normalization of an image is accomplished by using a simple linear regression m odel to relate one image to another image . Path/Row Image ID (gap - fill order) Acquisition Date 169/67 LC81690672015156LGN00 6/5/2015 169/67 LC81690672014137LGN00 5/17/2014 169/68 LC81690682015156LGN00 6/5/2015 169/68 LC81690682014137LGN00 5/17/2014 169/69 LC81690692015156LGN00 6/5/2015 169/70 LC81690702015156LGN00 6/5/2015 168/68 LC81680682014178LGN00 6/27/2014 168/68 LC81680682014130LGN00 5/10/2014 168/69 LC81680692015165LGN00 6/14/2015 168/69 LC81680692015133LGN00 5/13/2015 168/70 LC81680702015165LGN00 6/14/2015 168/70 LC81680702015133LGN00 5/13/2015 168/71 LC81680712015165LGN00 6/14/2015 168/71 LC81680712015133LGN00 5/13/2015 168/72 LC81680722015165LGN00 6/14/2015 167/70 LC81670702014155LGN00 6/14/2014 167/70 LC81670702015158LGN00 6/7/2015 167/71 LC81670712014155LGN00 6/14/2014 167/71 LC81670712015158LGN00 6/7/2015 167/72 LC81670722014155LGN00 6/14/2014 167/72 LC81670722015158LGN00 6/7/2015 18 Figure 1 : Malawi Path/Row Scenes. Malawi is covered by 11 nominal WRS - 2 path/row scenes which are shown here overlaid onto the Malawi international border for reference. 19 Ideally these two images will overlap and cover some similar ground features that remain static between the two acquisitions. Reference points selected over these static areas can be used to generate the linear regression model that is used to estimate the normalization between the two input images ( Elvidge et al. 1995; Matricardi et al. 2013) . Epoch 1 for the early dry season mosaic consists of imagery acquired in May of 1998 and 2002, epoch 2 was comprised of imagery collected in May and June of 2008 and 2010, and epoch 3 is comprised of imagery acquired in May 2014 and 2015. I II: Workflow Raw Landsat data were acquired using the USGS GLOVIS online tool ( http://glovis.usgs.gov ) . Individual path/row scene multispectral datasets were first converted to at - sensor - radiance from the raw digital numbers (DNs), and then converted to top o f atmosphere (TOA) reflectance by using the scene specific gain/offsets, scene specific solar elev ation angles and sun - earth distance along with sensor - specific exoatmospheric spectral irradiances using the following equations (Chander, Markham, and Helder 2009) : Conve rt raw DNs to spectral radiance: Convert spectral radiance to TOA reflectan ce: Clouds , cloud shadows, and water were detected and masked for each scene by using the fmask algorithm . Fmask, which stands for function of mask, is a completely self - contained algorithm that uses an input Landsat scene to generate a thematic product which identifies each pixel as clear land, water, cloud, cloud shadow, or snow . This is advantageous sinc e no other 20 ancillary data or scene - specific parameterization is required to generate the mask, allowing for a more streamlined and efficient method of cloud and cloud - shadow detection. Fmask works by taking advantage of known physical and spectral propert ies of clouds to identify potential cloud pixels. Cloud shadows are identified using spectral characteristics with the near infrared band and the satellite viewing angle to accurately estimate shadow placement (Zhu and Woodcock 2012). Each mask was appli ed to its parent Landsat scene to recode as 0 any pixels that were flagged as cloud, cloud shadow, or water . Furthermore, a buffer of 10 pixels surrounding any A serious cavea t of using 10 pixels as a buffer is that there is a very likely increase in commission error causing usable clear - land pixels to be ignored. This was mostly relevant for detected cloud and cloud shadow pixels, since clouds can have hazy pixels immediately surrounding them that are not detected by the algorithm , which themselves can cast a shadow that lowers the reflectance in the vicinity of the corresponding cloud - shadow . The 10 - pixel buffer helps to minimize the inclusion of contaminated pixels in the final masked products. The TOA reflectance conversion corrects for differences in sensor offsets and gains, solar zenith angles, sun - earth distance, and solar irradiances at the different times of acquisition. However, the TOA reflectance data s till suffer from atmospheric interference that causes scattering , absorption, and refraction of light between the earth surface and the satellite sensor. There are numerous atmospheric correction methods which exist in varyin g degrees of complexity . A ra diative transfer model such as the 6S (Second Simulation of the Satellite Signal in the Solar Spectrum) attempts to account for this atmospheric interference to obtain a true surface reflectance (Vermote et al. 1997). Models such as these are incredibly c omplex and oftentimes require image - specific ancillary data such as the aerosol optical thickness (AOT) to 21 produce accurate results . AOT is measured by ground - based weather stations and is not available at the necessary spatial or temporal distribution required to cover a single Landsat image due to the highly v ariable nature of the atmosphere both over space and time . A highly detailed and complex AOT correction was therefore not used for this study. On the other hand, other methods exist that approx imate an atmospheric correction. Such methods are not as accurate but are much simpler to apply to such a large dataset as was used in this study. A dark object subtraction (DOS) technique was therefore employed to account for the additive nature of atmo spheric interference (Chavez Jr. 1988). Using the information within each individual image, it is possible to identify this effect and minimize it by calculating band - specific correction factors. A starting band value in the visible spectrum was selected first, and then the accompanying band correction factors were calculated from this starting value based on a relative scattering model (Chavez 1996). The DOS technique can be easily applied to the imagery and was therefore chosen to produc e the surface reflectance data for this study. Surface reflectance datasets for each epoch w ere then mosaicked by path, and gap - fill ing of masked pixels was accomplished where multip le seasonally homogenous images were available. All of the images used for this process are listed in Tables 2, 3, and 4. The 3 path datasets for each epoch were combined into a national mosaic using an averaging function for overlap areas to minimize contrast between adjacent paths. The averaging function was a method of normaliza tion by which a simple linear regression is applied to the overlap region of two adjacent scenes. These mosaic d atasets were then clipped to the Malawi national boundary to produce the final multispectral national mosaic for each epoch . A Normalized Dif ference Vegetation Index (see Chapter 3 Section II for description of this calculation) was calculated from the multispectral path mosaics rather than from the national 22 mosaics due to differences between adjacent paths that become apparent when combining them into a single dataset. A s mentioned above, a djacent paths do not mosaic seamlessly due to slightly different levels of greenness resulting from differences in acquisition dates. This can have deleterious effects for the national mosaic change algori thm results by causing an unnatural patch - working effect. To reduce this variability a scaling process was applied to the vegetation index path datasets to minimize false change detection owing to variations in pixel values for adjacent paths. The scalin g process mentioned here is the normalization by means of regression modeling using sample points of static surface targets that are present in the two input scenes or mosaics. A built - in random point generator for Arc Map ( ArcGIS for Desktop (version 10.2. 2) 2015) was used to generate 10 0 0 random sample points with a minimum spacing of 120 meters . The location of all points was constrained using a polygon boundary that covered the spatial extent of each path - overlap area. This polygon was manually digitized using ArcMap. The vegetation index pixel values were extracted at each point location from the path datasets and added to the attribute table of the random points vector file. A column of pixel values was created for e ach path, and the values in each row represent the pixel values for an identical location shared by each path dataset allowing for the values at each location to be analyzed. The columns were copied from the ArcMap attribute table into an Excel spreadshee t to perform regression analysis. Path - to - path scaling was accomplished by identifying areas within the overlap of adjacent paths for which no the NDVI change was less than 20 percent. This was done to filter out areas where the change in NDVI was more l ikely to be the result of an actual land - cover change, rather than the result from atmospheric effects. It can be assumed that the vegetation index values for these areas should be nearly identical and that any difference is the 23 result of differing condit ions at the time of acquisition rather than differences in vegetation cover (Hudak and Wessman 2000) . Mosaicking the scaled path dataset with the reference path dataset allowed for a visual inspection of the success of the scaling effort. The 3 paths wer e then mosaicked together and clipped to the Malawi national boundary to create the national NDVI datasets for each time period. The NDVI was used in the calculation of vegetation fractional cover ( f C) . The fractional cover un - mixes the NDVI pixels between two pure components: vegetation and bare soil. Therefore, the final result is a continuous fields raster dataset in which each pixel represents the proportion of pure vegetation on a scale from 0 - 1. This is then re scaled to 0 - 100 to represent percentage vegetation cover by multiplying the f C by 100. NDVI provide a relative abundance of vegetation within each pixel as compared to other signal contributors (e.g. soil) within the pixel. Vegetation indices such as NDVI theref ore provide a physical measurement of vegetation, but rather are representations of the ratios between particular bands of remote sensing data ( Carlson and Ripley 1997 ). Spectral mixture analysis (SMA) has been shown to provide a more accurate representation of vegetation on the ground than simply using NDVI (Elmore, et al. 2000). A continuous fields product that quantifies a gradient of vegetation abundance across space and is comparable over time is the proposed product for the detection of miombo degradation in this thesis. To achieve this, a spectral un - mixing method similar to SMA was employed to determine the percent abundanc e of vegetation in each pixel. This impr oves upon the information provided by NDVI by separating out the signal components within a pixel and provides continuous fields data that estimate the physical abundance of vegetation as a percentage within each pixel. The usefulness of these data for det ecting degradation is that percentage quantities can be compared over time to determine if 24 the relative abun d ance of vegetation at a given pixel has decreased. This is an improvement over simply classifying the spectral data into thematic cover types and comparing classes for a given pixel over time (i.e. a methodology useful for the detection of closed canopy deforestation) . As s capacity for primary production while still maintaining the same land cover type. This type of change would not be apparent when comparing thematic class types across time ; rather it would be measurable when comparing the percentage abundance of vegetation over time. Therefore, NDVI was used as an input to a two - endmember SMA to generate continuous fields products that provide the abundance of vegetation as a percentage within each pixel. The two - endmember approach uses NDVI to separate the signal components of the NDVI value into a vegetation and non - vegetation percent age. This process is described in greater detail in Chapter 3 Section III. Visual analysis was used to generate endmember sample areas for each path that represented pure Miombo and pure bare substrate pixels that remained consisten t across all time periods and seasons. Areas of interest (AOIs) were drawn in ERDAS Imagine ( ERDAS Imagine 2014) for these endmember sample areas. The pixel values were then collected from the AOIs using the Spectral Signature tool in ERDAS Imagine . The values sampled for each endmember were combined to calculate the average national dataset endmember values. These values were t hen entered into the fractional cover model to calculate f C at the national scale for each time period. 25 Figure 2 : Flowchart showing the workflow used to generate the national - scale datasets and change results 26 Chapter 3: Results I. National Multispectral Mosaics A national multispectral mosaic was generated for each of the three time periods. The purpose of these is more for reference purposes, but they were also used for some of the change detection assessment. The mosaic s were stitched together using an averaging function, wherein the mean value of overlapping pixels in adjacent paths of ima ges was used in order to reduce the patchwork effect between adjacent paths. Histogram matching was also employed to further normalize the multispectral imagery by using path 169 as the reference dataset. Also, the surface reflectance datasets calculated with the DOS technique w ere used as the source of the mosaic components. 27 Figure 3 : The 1998 - 2002 multispectral mosaic. The image is comprised of TM and ETM+ images stitched together using an averaging function for overlap areas and path to path histogram matching to reduce seamlines. 28 Figure 4 : The 2008 - 2011 multispectral mosaic is composed almost entirely of TM imagery due to the SLC - off error on the ETM+ sensor for this time period. One ETM+ SLC - off image was used for gap - filling path/row 167/72. An averaging function was applied to overlap areas and no histogram matching was required. 29 Figure 5 : The 2014 - 2015 multispectral mosaic is composed entirely of OLI imagery. An averaging function was used for overlap areas and no histogram matching was required. 30 II. National NDVI Mosaics Plant physiology has unique properties which allow s remote sensing of plant characteristics. Photosynthesis causes plant leaves to be highly absorptive of the visible red portion of the electromagnetic spectrum. Also, the physical structure of plant leaves makes them highly reflective of the near infrare d portion of the electromagnetic spectrum. A healthy tree canopy will therefore absorb most of the red visible light it receives for use in primary production while simultaneously reflecting much of the near - infrared light that also strikes it (Tucker 1979) . These spectral characteristics are nearly unique to plants and photosynthesis and can be used to easily detect the presence of vegetation, measure the extent of plant coverage , or even estimate primary production via remote sensing. The r ed and near - infrared (NIR) reflectances of a surface can be plotted on an orthogonal axis to create what is referred to as a feature space image. This plot visualizes the similarity and variation between two bands . Since the red visible and near - infrared bands are sensitive to vegetation, portions of the plot between these bands correlate to either the presence of vegetation or bare soil. This helps explain the basis for vegetation indices and how they utilize the unique spectral characteristics of veget ation in the red - visible and near - infrared portions of the electromagnetic spectrum (Lillesand, Kiefer, and Chipman 2008). A widely used vegetation index is the normalized difference vegetation index (NDVI). NDVI has historically been used as the index of choice for the study of vegetation dynamics in the field of remote sensing for decades. It is easily calculated from the red and near - infrared bands of satellite imagery, such as that acquired by Landsat, using the following equation: 31 NIR and RED here refer to the reflectance values of the near - infrared and red bands respectively. NDVI has been shown to be directly correlated to biophysical characteristics of vegetation via a linear relationship with leaf area index (LAI) up to a particular value generally between an LAI of 3 and 4 (Carlson and Ripley 1997). NDVI becomes insensitive to increases in LAI beyond a certain value, and therefore cannot measure differences in very dense canopies once LAI saturation has been reached ( Tucker 1979 ). Therefore, NDVI may not be the best vegetation index for regions consisting of dense tropical canopies. Miombo, however, has characteristically open canopy making NDVI suitable for detecting the presence and vitality of Miombo woo dlands (Frost 1996). The NDVI mosaic datasets were calculated directly from the multispectral mosaic datasets using a graphical model in ERDAS Imagine. The goal of these was to represent a high quality record in time of the abundance of vegetation at the national scale. The values of NDVI range from - displayed below in grayscale with darker regions corresponding to lower NDVI values and therefore lower abundances of vegetat ion. These datasets were used as direct inputs into the model used to calculate fC at the national scale using a simple two - endmember linear un - mixing algorithm. 32 Figure 6 : The national NDVI mosaic for the 1998 - 2002 time period. 33 Figure 7 : The national NDVI mosaic for the 2008 - 2010 time period. 34 Figure 8 : The national NDVI mosaic for the 2014 - 2015 time period. 35 III. National fC D atasets The spatial resolution of Landsat is 30 meters, which provides a default minimum mapping unit that is ideal f or regional studies. However, the area within a Landsat pixel, which is about 900 m 2 , is relatively large for representing the measurement of a heterogenous land cover. Ea ch pixel is potentially a mixture of land covers (e.g. canopy or soil) or land cover components (e.g. soil, green vegetation, shadow, stems and branches), all contributing to the overall reflectance signal . Thus, the reflectance signal for a given pixel i s the sum of its land - cover s or cover components . These sub - pixel land - cover elements can therefore be quantified at the pixel - scale using a measure of fractional cover, i.e. the percent for which each land cover e signal for a particular wavelength, or range of wavelengths when speaking in terms of individual Landsat bands. Q uantif i cation of these land cover types across a landscape consistently is integral to the use of satellite imagery to measure change across time (Adams et al. 1995). There are a number of methods available for calculating fractional cover. Spectral mixture analysis (SMA) is the process by which the fraction components within a pixel are calculated for a given band with given endmember values that are unique to each land cover type for each band. An endmember is simply the reflectance value for a particular homogenous land cover type at a particular wavelength or Landsat band. Therefore, SMA requires knowledge of not only the land cover type s present within a pixel, but also each land analysis (LSMA), a typical scenario is separating vegetation such as forest from non - photosynthetic vegetation (NPV) and non - vegetation or bare soil (Cochrane and Souza 1998). An equation can then be used to describe the fraction component of each land cover type within 36 the pixel and solved by combining a least - square estimator with the constraining factor that all fra ctions must sum to 1 for each pixel (Lillesand, Kiefer, and Chipman 2008). Multiple endmember spectral mixture analysis (MESMA) builds upon SMA by charging that endmember values are variable throughout a satellite image (Roberts et al. 1998). MESMA ther efore accounts for this extra variation in an attempt to generate a more accurate fractional cover, but requires more field data measurements to model the variation of endmember types across the landscape. MESMA is therefore a much more labor and computat ionally intensive algorithm best suited for very specialized scenarios for which extensive knowledge of the component reflectance characteristics exists. The se two methods SMA and MESMA -- were employed to calculate the fraction of land cover types for multispectral data. This requires a great deal of computing and also requires determining several endmember values for each of the different land cover types. The number of bands in the image ty pically determines the maximum number of endmembers required (Lillesand, Kiefer, and Chipman 2008). The SMA approach used with multispectral imagery can be simplified by applying a linear un - mixing model to a vegetation index (Qi et al. 2000; Xiao and Moo dy 2005; Matricardi et al. 2010). As described previously, vegetation indices take advantage of the unique spectral characteristics of vegetation generally in the visible red and near infrared portions of the electromagnetic spectrum to describe the level present in an image. A linear un - mixing model uses a pure vegetation endmember and a pure bare soil endmember to calculate the fraction of vegetation present in each pixel with the following formula: 37 Where VI is the value of the input vegetation index, VI bare is the endmember vegetation index for bare substrate, and VI veg is the endmember vegetation index for vegetation. The result is a raster file of continuous - . Selecting the best repre sentative endmembers for pure land - cover types is essential for correctly un - mixing the materials of interest within each pixel. There may be many different vegetation types within a Landsat scene. However, the location of Miombo in the study area is wel l documented. This allowed for accurately selecting pure pixels for determining endmembers associated with Miombo vegetation. To ensure consistency between fractional cover datasets, endmember samples were taken from identical areas of interest (AOIs) fo r both time periods. This was done to minimize the effects of atmospheric scattering due to haze and effects due to phenology as a result of meteorological or atmospheric differences across the two dates. This was because while both time periods are cons istent in terms of seasonality and day of year (DOY) they may still exhibit slight differences due to these aforementioned factors. Similarly, for bare substrate endmember s , samples were taken from the same AOIs across both time periods for areas that wer e consistently bare, such as roads and large barren agriculture fields. Endmembers were sampled from each of the paths of Landsat imagery in or around the three principle study areas identified by the Malawi PERFORM project. These are the Perekezi, Ntchisi, and Liwonde forest reserves . The three national fC datasets were produced directly from the national NDVI mosaic datasets. Endmembers were selected from each NDVI mosaic which represented pure vegetation and pure bare soil land cover types. These were found by selecting areas of the image that represented these pure land covers by drawing polygons around the areas of interest. The signature editor tool in ERDAS was then used t o collect the NDVI values within each polygon. 38 The signature editor tool allowed for the calculation of the overall mean for each endmember value by combining the area of interest samples associated with each land cover type. The tool also allowed for th e verification of quality samples by the use of histograms to ensure samples had a normal distribution. Once the endmembers were determined, they were input into an ERDAS graphical model which calculated the f C directly from the NDVI mosaic dataset. This produced a national f C dataset for each time period. 39 Figure 9 : The national f C dataset for the time period of 1998 - 2002. 40 Figure 10 : The national f C dataset for the time period of 2008 - 2010. 41 Figure 11 : The national f C dataset for the time period 2014 - 2015. 42 IV. Accuracy Assessment Using High - Resolution Imagery High resolution satellite imagery (spatial resolution of 1 m or less) was purchased from Apollo Mapping to support an accuracy assessment and validation of the f C results . A search was performed for imagery with minimal cloud cover and ideal nadir conditions for coverage of the 3 study areas. After searching, t hree images were selected for purchase that cover ed the forest reserves Perekezi and Liwonde . These images were also selected based on their close temporal proximity to the Landsat imagery used in the early dry season analysis. The Perekezi scene was a GeoEye - 1 50 cm pan - sharpened orthorectified 4 - band image acquired on May 31, 2014. Parts of Liwonde were covered by two images covering the east and west portions of the protected forest area. The west scene was a WorldView - 3 50 cm pan - sharpened orthorectified 4 - band image acquired on May 7, 2015 , and t he east scene was a WorldView - 2 50 cm pan - sharpened orthorectified 4 - band image acquired on April 28, 201 5 . Table 5 : High resolution satellite images used for the purpose of accuracy assessment of the fractional cover product High resolution imagery was obtained to create a classification of forest cover as a GIS layer which could be used as a reference layer surrogate to ground validation (i.e. ground truth) . The reasoning behind this is that the high - resolution data represents an unmixed land cover image depicting spatial objects, usually individual trees, without mi xed spectral responses. Thus, high resolution data directly portray what the models using medium - resolution Landsat imagery Image Target Sensor Acquisition Date Perekezi GeoEye - 1 5/31/2014 Liwonde (west) WorldView - 3 5/7/2015 Liwonde (east) WorldView - 2 4/28/2015 43 predict . However, classification of high - resolution imagery using traditional statistical classifiers is rather difficult because t he sensor detects a much higher degree of spatial and spectral variation within each land cover type. Therefore , it is necessary to introduce other decision factors or to altogether reduce the spectral variation to avoid over - abundant noise in the classified product . In high - resolution imagery, the spatial distribution of pixels can be combined with the spectral characteristics of pixels to aid in developing contiguous groups of land cover classes in what is known as object - based classifica tion . Alternatively, unsupervised classification can be used to first partition the multispectral image into a pre - determined number of classes based on the spectral characteristics of the image. Then, these classes can be manually combined into the desi red final classified product. Accuracy assessment was performed on the final classified product to determine the validity for further use. Classification and accuracy assessment of high - resolution imagery was performed using ERDAS Imagine. For this wor kflow, the Advanced RGB clustering tool was used to first convert the DNs into a thematic layer of several hundred classes based on their spectral values. These classes were then grouped together manually by using the Grouping Tool to recode their class v alues to values corresponding to either vegetation or non - vegetation. The Neighborhood M ajority F unction with a 3x3 moving window size was applied to the original classified products to remove noise. The Accuracy A ssessment tool was used to generat e 100 random ly placed reference points. T he original high - resolution images with the false - color near - infrared band combination were then used as reference layer s to code each sample point as vegetation , non - vegetation, or background. These reference clas ses corresponded to the classes present in the high - resolution classified image . The final overall accuracy was 79% , 86%, and 89% for the Perekezi GeoEye - 1 image , the Liwonde Worldview - 3 image , and Liwonde Wordlview - 3 image 44 respectively. Based on the Kappa Statistic, the level of agreement for the classification is considered moderate for the Perekezi scene and substantial for the two Liwonde scenes (Landis and Koch 1977). Confusion Matrix Perekezi GeoEye - 1 Image Reference Data Classified Data Other Bare Soil Vegetation Row Total Other 2 0 0 2 Bare Soil 0 33 11 44 Vegetation 0 10 44 54 Column Total 2 43 55 100 Table 6 : The confusion matrix generated for the Perekezi GeoEye - 1 high resolution vegetation/non - vegetation thematic layer. Accuracy Totals Perekezi GeoEye - 1 Image Class Name Reference Totals Classified Totals Number Correct Producers Accuracy Users Accuracy Other 2 2 2 --- --- Bare Soil 43 44 33 76.74% 75.00% Vegetation 55 54 44 80.00% 81.48% Totals 100 100 79 Overall Classification Accuracy = 79.00% Table 7 : Accuracy totals for the Perekezi GeoEye - 1 high resolution vegetation/non - vegetation thematic layer. 45 Kappa (K^) Statistics Perekezi GeoEye - 1 Image Overall Kappa Statistics = 0.5910 Conditional Kappa for each Category Class Name Kappa Other 1 Bare Soil 0.5614 Vegetation 0.5885 Table 8 : Conditional kappa statistics for the individual classes, as well as the overall kappa statistic for the Perekezi GeoEye - 1 classification. Confusion Matrix Liwonde Worldview - 3 Image Reference Data Classified Data Other Bare Soil Vegetation Row Total Other 3 0 0 3 Bare Soil 0 38 1 39 Vegetation 0 13 45 58 Column Total 3 51 46 100 Table 9 : The confusion matrix generated for the Liwonde Worldview - 3 high resolution vegetation/non - vegetation thematic layer. Accuracy Totals Liwonde Worldview - 3 Image Class Name Reference Totals Classified Totals Number Correct Producers Accuracy Users Accuracy Other 3 3 3 --- --- Bare Soil 51 39 38 74.51% 97.44% Vegetation 46 58 45 97.83% 77.59% Totals 100 100 86 Overall Classification Accuracy = 86.00% Table 10 : Accuracy totals for the Liwonde Worldview - 3 high resolution vegetation/non - vegetation thematic layer. 46 Kappa (K^) Statistics Liwonde Worldview - 3 Image Overall Kappa Statistics = 0.7375 Conditional Kappa for each Category Class Name Kappa Other 1 Bare Soil 0.9477 Vegetation 0.5849 Table 11 : Conditional kappa statistics for the individual classes, as well as the overall kappa statistic for the Liwonde Worldview - 3 classification. Confusion Matrix - Liwonde Worldview - 2 Image Reference Data Classified Data Other Bare Soil Vegetation Row Total Other 1 0 0 1 Bare Soil 0 50 11 61 Vegetation 0 0 38 38 Column Total 1 50 49 100 Table 12 : The confusion matrix generated for the Liwonde Worldview - 2 high resolution vegetation/non - vegetation thematic layer. Accuracy Totals - Liwonde Worldview - 2 Image Class Name Reference Totals Classified Totals Number Correct Producers Accuracy Users Accuracy Other 1 1 1 --- --- Bare Soil 50 61 50 100.00% 81.97% Vegetation 49 38 38 77.55% 100.00% Totals 100 100 89 Overall Classification Accuracy = 89.00% Table 13 : Accuracy totals for the Liwonde Worldview - 2 high resolution vegetation/non - vegetation thematic layer. 47 Kappa (K^) Statistics Liwonde Worldview - 2 Image Overall Kappa Statistics = 0.7838 Conditional Kappa for each Category Class Name Kappa Other 1 .00 Non - Vegetation 0.6393 Vegetation 1 .00 Table 14 : Conditional kappa statistics for the individual classes, as well as the overall kappa statistic for the Liwonde Worldview - 2 classification. Figure 12 : Perekezi GeoEye - 1 Imagery. The multispectral image is on the left with the random sample points overlaid. The thematic vegetation/non - vegetation image is on the right. 48 Figure 13 : Liwonde Worldview - 3 Imagery. The multispectral image is on the left with the random sample points overlaid. The thematic vegetation/non - vegetation image is on the right. Figure 14 : Liwonde Worldview - 2 Image ry . The multispectral image is on the left with the random sample points overlaid . The thematic vegetation/non - vegetation image is on the right. 49 V . Accuracy Assessment of the Landsat - based f C Product All GIS - related operations for this purpose were executed using ArcMap with the Spatial Analysis extension enabled. Determining the accuracy of the fractional cover product was carried out using the previously created high - resolution vegetation/non - vegetation classification map. This accuracy assessment was therefore done for the study areas of Perekezi and Liwonde due to the high - resolution imagery purchased for those sites. First, subsets of the study areas were generated from the national fractional cover dataset early dry season time period 2014 - 2015, which matched the high - resolution imagery acquisition dates and phenology most closely. To relate 30 meter Landsat - derived fractional cover to the high - resolution 0.5 meter vegetation/non - vegetation classification, it was necessary to generate a standardized sampling grid. It was first decided to sample the fractional cover value for a rando m ly generated selection of 100 Landsat pixels . The extent of the high - resolution image area was digitized, and a grid of 30x30 meter squares was generated in this extent, along with the center point for each grid cell. The generated grid was edited so th at it aligned with the extent of the Landsat pixels. A random selection of 100 center points was made. The center points were utilized to extract the value of f C at each Landsat pixel. The intersecting square polygons from the sampling grid were extract ed for each of the 1 00 points as well and saved as a new polygon shapefile resulting in 1 00 square polygons . The result was a quantity of 1 00 center points and 1 00 square cells aligned with the Landsa t fractional cover grid cells. The utility of these two separate shapefiles is as follows: The point shapefile was used to extract the Landsat fractional cover value for each pixel that was intersected by a point feature, and then those values were entered into the attribute table of the point shapefile so th at each row now had a corresponding fractional cover value. The point features were then related to the 50 square cell polygon features using a spatial join, so that each row of the point attribute table had a corresponding square cell ID. Therefore, the po ints could then be grouped together within the attribute table based on the ID from each of the 1 00 corresponding square cells. The square cell polygon shapefile was used to calculate the high - resolution fractional cover . This was performed by tabulating pixel counts by area . For each square cell , the number of pixels from each class of the high - resolution thematic image w as counted and sorted based on class value. The fractional cover was then calculated for each cell as the sum of vegetation class pixels divided by the sum of all pixels per cell and multiplied by 100. The results of the tabulation were then joined to the attribute table of the point shapefile based on the square cell polygon ID field. These results were copied into a comma separated values (.csv) file and opened in R for statistical analysis (R Development Core Team 2008) . The Landsat fractional cover could then be compared to the high - resolution fractional cover based on the ID of the square cell polygon. Figure 15 : Comparison of the same spatial area shown in both the 30m fC product derived from Landsat data (left image) and the 0.5m thematic classification of vegetation and non - vegetation (right image). 51 Figure 16 : Perekezi regression of Landsat f C and HiRes f C. Simple linear regression plot between the Landsat f C (independent variable) and the hi - resolution f C (dependent variable) using the Perekezi Geoeye - 1 imagery. Perekezi SLR Results Dependent variable: Hi - Res f C Landsat fC 0. 7734 *** ( 0.7734 ) Constant 14.585 *** ( 4.050 ) Observations 220 R 2 0.477 Adjusted R 2 0.475 Residual Std. Error 5.927 (df = 21 9 ) F Statistic 1 99 . 137 *** (df=1; 219 ) Note: *p<0.1; **p<0.05; ***p<0.01 Table 15 : Perekezi SLR results. Simple linear regression results from the comparison of the Landsat derived f C (the independent variable) and the high resolution GeoEye - 1 f C (the dependent variable). 52 Figure 17 : Liwonde (West) regression of Landsat f C and HiRes f C. Simple linear regression plot between the Landsat f C (independent variable) and the hi - resolution f C (dependent variable) using the Liwonde Worldview - 3 imagery. Table 16 : Liwonde (West) SLR results. Simple linear regression results from the comparison of the Landsat derived f C (the independent variable) and the high resolution Worldview - 3 f C (the dependent variable). Liwonde ( W est) SLR Results Dependent variable: Hi - Res f C Landsat fC 0.8555 *** ( 0.0353 ) Constant 10.696 ***( 2.870 ) Observations 290 R 2 0. 670 Adjusted R 2 0. 669 Residual Std. Error 6.273 (df = 28 9 ) F Statistic 588 . 049 *** (df=1; 28 9 ) Note: *p<0.1; **p<0.05; ***p<0.01 53 Figure 18 : Liwonde (east) regression of Landsat f C and HiRes f C. Simple linear regression plot between the Landsat f C (independent variable) and the hi - resolution f C (dependent variable) using the Liwonde Worldview - 2 imagery. Liwonde ( E ast) SLR Results Dependent variable: Hi - Res f C Landsat fC 0. 8906 *** (0.0 303 ) Constant 5.2009 ***( 2.3503 ) Observations 165 R^2 0. 8405 Adjusted R^2 0. 8340 Residual Std. Error 5.572 (df = 164 ) F Statistic 864.485 *** (df=1; 164 ) Note: *p<0.1; **p<0.05; ***p<0.01 Table 17 : Liwonde ( E ast) SLR results. Simple linear regression results from the comparison of the Landsat derived fC (the independent variable) and the high - resolution Worldview - 2 fC (the dependent variable). 54 V I . Change Detection Algorithm The usefulness of time series satellite imagery is that it provides measurements which directly record how a given landscape has been altered over time both by anthropogenic and natural forces. Multispectral imagery such as Landsat contains a large amount of information that is difficult to quantify in n - dimensions where n is the number of bands present. Hence, classification techniques can be used to quantify targeted land cover types, or as is used in this study, vegetation indices and fractional cover a nalysis can be used to directly measure biophysical characteristics of vegetation. This can then be used to measure changes in vegetation such as Miombo woodlands over time . A graphical model was built in ERDAS Ima gine to process change detection between two spatially related periods of time using the national f C datasets as inputs. The change model had 3 main functions. A m inimum change detection threshold of 1 0 was used as a filter because it was assumed that an y reduction in f C of less than this was both negligible and likely temporary . The first function in the model calculated the diffe rence between the earlier date f C and the later date f C ( f C 2 f C 1 ) so that a decrease in f C across time yielded a negative value. A conditional statement was included in this difference function to ignore any results that were greater than or equal to 0, which means any positive or null changes in f C were ignored. In other words, only decreas es in vegetation are being examined in this study. Two simultaneous functions were also applied to the individual f C inputs. These functions binned the f C datasets into 6 categories based on the actual f C values. The binned f C datasets were used as inputs along with the results of the differencing function into the last function of the model. This function compared the binned classes of f C 1 and f C 2 to determine if a pixel was either degradation or 55 deforestation. The differenci ng results were used to only accept change that was at least a magnitude of 1 0 or greater, meaning that minimum decrease of 1 0 in f C was enforced here. Deforestation and degradation were categorized based on how a pixel changed between f C classes, and als o whether or not that pixel met the minimum change requirement. This minimum requirement was enforced for both degradation and deforestation. A land cover dataset was used to mask out non - forest areas. This meant that changes in f C would only be examined in areas designated broadly as protected forest areas. Figure 19 : ERDAS graphical model used to carry out the change detection between two f C datasets . 56 VII: Change Detection Algorithm Results Between the epochs of 200 0 and 20 09 a total of 248,007 Ha of forest land was degraded , but remained forest , while 16,202 Ha of forest land was deforested based on the comparison of fC values for the two datasets corresponding to these dates. Between the y ears of 2010 and 2015 a total of 254,25 7 Ha of forest land was degraded and 9,037 Ha of forest land was deforested based on the comparison of fC values for the two datasets corresponding to these dates. For the time period of 2010 to 2015, of the 254,25 7 Ha of forest that was degraded, 36,159 Ha was in forest that were further degraded from forest degraded in 2000 - 2009 was further degraded. Also, between 2010 and 2015, of the 9,037 Ha of deforested land, 5,547 Ha of that occurred on forested land that was previously degraded between 2002 and 2010. Likewise, between 2010 and 2015 218,097 Ha of the 254,25 7 Ha of degradation occurred on forest land that had not been previously degraded. Similarly, 3,490 Ha of the total 9,037 Ha that were deforest ed between 2010 and 2015 occurred on land that was not previously degraded between 2002 and 2010. The total area of protected forests in Malawi, as determined from the available land - cover data set , is 2,399,381 Ha. Between 2002 and 2010, 10.34% of protected forests were degraded and 0.68% of protected forests were deforested. Between 2010 and 2015, 10.59% of protected forests were degraded and 0.38% of protected forests were deforested. Figure 20 shows the mapped results of the change detection algorithm. Three insets are provided in addition to the national scale map; each inset shows a protected forest area. Two of these protected forest areas were also used for the validation of the f C product via the use of high - resolution imagery. Inset A sho w s Perek ezi forest reserve, inset B shows N t chisi forest reserve, and inset C shows Liwonde forest reserve. Areas of degradation for the 2 change 57 periods are clearly visible in each reserve. A number of previously identified drivers is likely responsible, howeve r there may be a few caveats to these results. Primarily, orchards were not considered in the change algorithm. All land within the protected forest reserves was considered to be forest . However, orchards are present for example in southeastern Perekezi , and it is possible that these could contribute to false positives in the change results. 2 002 to 2010 Change Results Change Type Area (ha) % of Total Forest Area Degradation 2 48 ,0 07 10.34 % Deforestation 16,202 0.68 % Total Forest Area 2, 399 , 381 11.01 % Table 18 : 2002 to 2010 change results. Area in hectares of each change category for the time period from 2002 to 2010. The percent of total forest area for each change category is also shown . 2010 to 2015 Change Results Change Type Area (ha) % of Total Forest Area Degradation 2 54 , 257 10.59 % Deforestation 9,037 0.38 % Total Forest Area 2, 399 , 381 10.97 % Table 19 : 2010 to 2015 change results. Area in hectares of each change category for the time period from 2010 to 2015. The percent of total forest area for each change category is also shown. 58 Figure 20 : The change algorithm results are shown here for the two lengths of time 2002 - 2010 and 2010 - 2015. The types of change that occurred in each period are color coded. Inset A shows the detail area around Perekezi Forest Reserve which includes a western portion of the South Viphya Forest Reserve. Inset B shows the detailed area around Ntchisi Forest Reserve. Ins et C shows the detailed area around Liwonde Forest Reserve, which also includes the southern portion of Liwonde National Park and the northern portion of the Zomba - Malosa Forest. All protected forests are indicated by the green areas on the main map and i nsets. 59 VIII: V alidation of Change Detection Algorithm Results The a ccuracy a ssessment for the results of the change algorithm was carried out by calculating a confusion matrix from a set of stratified random sample points . The GeoEye - 1 image acquired for the area covering the Perekezi forest reserve was used as a reference layer to confirm the presence of the classified change types. For this scenario, 100 stratified random points were generated and the parameter for a minimum number of points per class was set to 10 to ensure deforestation was sampled due to its relatively smaller area compared to the other two change classes. Google Earth imagery was also used as a reference for earlier time period high - resolution imagery by spatially - linking an ERDAS viewer to the Google Earth user interface. However, earlier time period high - resolution imagery was not available for most sample points so inferences had to be made as to the nature and time period for areas of forest that were cleared or degraded in the purchased high resolution imagery due to a lack of reference. If a sample point fell within an open agricultural field in the more recent high resolution image, then Google Earth was referenced to confirm whether or not that cl earing was made within the time period measured by the Landsat imagery. If the clearing was made prior to the Landsat imagery acquisition, then it should be considered an area of no change. The Landsat multispectral mosaic images from the two time period s were used as a reference as well to observe the source of any potentially detected change directly. The purchased high resolution imagery was however used as the primary deciding factor in assigning reference classes to the sample points. If no earlier reference point was available, then a reas that were obviously cleared for agriculture or that had sparse canopy cover and proximity to agriculture or settlements were assigned degradation and deforestation values depending on the openness of the land wher e the sample points were located. This was suitable for testing the accuracy of the Landsat - derived change detection 60 because it was shown earlier that the high - resolution classified percent vegetation cover correlated fairly well to the Landsat fractional cover. Confusion Matrix for Change Detection Reference Classified No Change Degradation Deforestation Row Total No Change 46 7 1 5 4 Degradation 4 21 3 28 Deforestation 1 1 1 6 1 8 Column Total 51 2 9 20 100 Table 20 : Confusion matrix for change detection. The confusion matrix that was created as part of the accuracy assessment of the change algorithm. The rows represent the classified change algorithm product, and the columns represent the ground - truth reference totals. The diagonal of the table represents correctly labeled change classes. Accur acy Totals for Change Detection Class Name Reference Totals Classified Totals Number Correct Producers Accuracy Users Accuracy No Change 51 5 4 46 90.2 % 85.2 % Degradation 29 28 21 7 2.41 % 75 % Deforestation 20 1 8 1 6 80.0 % 88.89 % Totals 100 100 8 3 Overall Classification Accuracy = 8 3 % Table 21 : Accuracy totals for change detection. The accuracy totals are shown for the producer's, user's, and overall accuracies. 61 Kappa (K^) Statistics for Change Detection Overall Kappa Statistics = 0. 7201 Conditional Kappa for each Category Class Name Kappa No Change 0. 6977 Degradation 0. 6479 Deforestation 0. 8611 Table 22 : Kappa (K^) statistics for change detection. Conditional k appa values calculated for each individual class as well as the kappa value for the overall classification. Tables 20 - 22 show the results of the accuracy assessment. The confusion matrix was built first to determine the number of correctly or incorrectly assigned change categories (table 20 21 ). Finally, these accuracies were used to calculate the overall kappa statistic (table 22 ). The overall kappa statistic equaled 0.720 1, which is a substantial result indicating the success of the change algorithm for detecting degradation and deforesta tion (Landis and Koch 1977). IX: Discussion and Conclusion T hree Malawi national - scale Landsat multispectral mosaics representing unique epochs of time were generated. A continuous fields fractional cover of vegetation was calculated from each of these Landsat mosaics . A technique was successfully demonstrated which uses the f C product to detect changes in vegetation cover over time. Negative change in vegetation cover within protected forest lands represents forest degradation from one epoch to the next . This type of change detection is particular ly important for the forestry sector due to the proliferation of both high and low impact forest use. It is necessary for REDD+ efforts in Malawi to be able to accurately quantify current levels of forest degradation to generate reference emission levels. The f C products allowed for a robust and accurate method of directly locating areas where degradation and deforestation h ave occurred. Therefore, it is a recommendation that the 62 technique demonstrated here be implemented as part of a measurement, reporting, and verification (MRV) tool, which is a requirement of UNFCCC and IPCCC for REDD+ in Malawi and other countries. The a ctual quantification of biomass removal may be estimated based on the f C change results, but it should be verified and recalibrated with extensive on - the - ground measurements. Other forms of remote sensing such as LiDAR and high - resolution satellite imager y may provide a means of further calibrating the Landsat - derived fractional cover and allow for a more accurate biomass measurement. Mapping the locations of degradation and deforestation with Landsat imagery as shown in this study is an effective means t o direct REDD+ mitigation strategies and These change products can also be used to develop remote sensing - based activity data to assist with calculating reference emission le vels for REDD+. Both degradation and deforestation were observed across the entire national extent of Malawi. Based on the analysis of Landsat data, the extent of these areas was quantified. Similarly, the use of remote sensing data allowed for locating the hotspots of deforestation and forest degradation across the country of Malawi. It is not surprising that these areas coincide with locations of protected forest areas. This is likely due to the fact that much of the country has already been extensive ly deforested. The remaining areas of Miombo woodland are therefore primarily located within these protected forests. The protected forest status may prevent land cover change from forest to agriculture which prevents deforestation. However, forest degr adation may still occur because these remaining woodlands represent a source of fuel wood and other wood - based resources, which is driven by the extensive poverty and lack of alternative energy resources for the majority of Malawians. 63 Between 2002 and 2 010 248,007 Ha of Miombo woodlands were degraded, and between 2010 and 2015 254,257 Ha of Miombo woodlands were degraded based on the analysis presented here. Deforestation of Miombo woodlands between 2002 and 2010 totaled 16,202 Ha and between 2010 and 2 015 totaled 9,037 Ha. This supports the conclusion that much of Malawi is already deforested, and that the remaining protected forests are still susceptible to exploitation that results in forest degradation. One caveat to the national f C datasets is the use of static endmembers. Because the f C algorithm works by taking in a single endmember value for each of the two cover - types, it may not be the most accurate depiction of vegetation cover across a large area (Johnson, Tateishi, and Kobayashi 2012) . This is due to the variation in land cover types, pred ominant forms of vegetation, and distribution of micro - climates across a large region such as Malawi. Despite atmospheric corrections and image normalization, differences may still exist between adjacent and gap - fill Landsat scenes which could necessitate individual scene - specific endmembers. This could mean that scene - specific endmembers might be more effective for measuring f C at the scale of individual fore st reserves. Spatial interpolation of endmembers is one theoretical solution to this problem (Johnson, Tateishi, and Kobayashi 2012). More work should be done on which type of spatial interpolation is more effective, and how accurately the endmembers of target land cover types can be modeled in such a way as to generate a more meaningful f C over a large area. However, despite this caveat, the technique demonstrated of using a two - endmember SMA was sufficient for accurately measuring the presence of veget ation across the landscape at a 30 - meter resolution. There is potential for more remote sensing work to be done in Malawi. By adding more dates to the analysis, a higher temporal resolution could be obtained. This would allow for a 64 clearer picture to eme rge which indicates when and where Miombo woodlands are being degraded . This can certainly help to inform mitigation and outreach efforts with local communities. Furthermore, the ability to quantify biomass directly from Landsat satellite observations would allow for a robust and efficient method to develop activity data necessary for developing REDD+ in Malawi. A technique for using f C to map change in open canopy Miombo woodlands has been demonstrated. Specific locations of degradation and deforestation within Malawi protected forest areas have been produced. The usefulness of the freely available Landsat data archive has also been demonstrated as a reliable and scientifically robust data source for generating v egetation cover maps. Activity data related to forest loss and degradation can be acquired by c omparing f C across time to calculate delta f C values . These delta values have been shown in this study to be a promising method for estimating forest degradati on in Malawi, and these methods have potential for reproducibility to develop ac tivity data for other countries. These efforts have also assisted with the Malawi PERFORM project by providing actionable scientific data describing locations and magnitudes of Miombo degradation in Malawi . PERFORM is a capacity building project to assist Malawi with developing a framework to prepare for REDD+ and increase forest conservation. PERFO RM was designed as a primary mechanism between the United States Government and Government of Malawi to promote low ering emissions via conser vation and sustainable forestry practices. 65 B IBLIOGRAPHY 66 BIBLIOGRAPHY Journal of Applied Ecology 36 (3): 422 33. Adams, John B., Donald E. Sabol, Valerie Kapos, Raimundo Almeida Filho, D ar A. Roberts, Based on Fractions of Endmembers: Application to Land - Cover Change in the Brazilian Remote Sensing of Environment 52 (2): 137 54. ArcGIS for Desk top (version 10.2.2). 2015. Redlands, CA: ESRI Inc. Science 320 (5882): 1444 49. Broich, Mark, Matthew C. Hansen, Peter Potapov, Bernard Adusei , Erik Lindquist, and Stephen - Series Analysis of Multi - Resolution Optical Imagery for International Journal of Applied Earth Observation and Geoinformation 13 (2): 277 91 . Broich, Mark, Matthew Hansen, Fred Stolle, Peter Potapov, Belinda Arunarwati Margono, and Temporal Variation across Sumatera and Kalimantan, Indonesia 2000 Env ironmental Research Letters 6 (1): 014010. Fuel: Responses of Rural Households to Decreasing Fuelwood Availability, Ntcheu World Development 25 (2): 255 66. Remote Sensing of Environment 62: 241 52. y of Current Radiometric Calibration Coefficients for Landsat MSS, TM, ETM+, and EO - 1 ALI Remote Sensing of Environment 113: 893 903. - Object Subtraction Technique for Atmospheric Scattering Correction o Remote Sensing of Environment 24 (3): 459 79. - Based Atmospheric Corrections Photogrammetric Engineering & Remote Sensing 62 (9): 1025 36. 67 Chavula, Geoffrey, Patrick Brezonik, an Change (LULC) in the Lake Malawi Drainage Basin, 1982 - International Journal of Geosciences 02 (02): 172. The Miombo in Transition: Woodlands and Welfare in Africa , edited by Bruce Campbell. Bogor, Indonesia: CIFOR. International Journal of Remote Sensi ng 19 (17): 3433 40. Environmental Science & Policy , Governing and Implementing REDD+, 14 (2): 89 99. Network: Framework for a Terrestrial Transect Study of Land - Use and Land - Cover Change in the International Geosphere - Biosphere Program. Elmore, Andrew J. , Mustard, John F., Mann Quantifying Vegetation Change in Semiarid Environments: Precision and Accuracy of Spectral Mixture Analysis and the Normalized Difference Vegetation Index Remote Sensing of Environment 73 ( 1 ): 87 - 102 Elvidge, Christopher and Yuan, Ding & Weerackoon, R.D. and Lunetta, R.S.. 1995 . Relative Radiometric Normalization of Landsat Multispectral Scanner (MSS) Data Using an Automatic Scattergram - Controlled Regression. Photogrammetric Engineering & Remote Sens ing . 61. 1255 - 1260. ERDAS Imagine . 2014. Madison, AL: Intergraph Corporation. The Miombo in Transition: Woodlands and Welfare in Africa , edited by Bruce Campbell. Bogor, Indonesia: CIFOR. Implications for Measuring and Monitoring Forest Carbon Stocks and Carbon Management 2 (3): 231 44. nsing - Derived National Geocarto International 30 (3): 270 92. Landsat Legac y Special Issue 122 (0): 66 74. 68 Hansen, Matthew C., David P. Roy, Erik Lindquist, Bernard Adusei, Christopher O. Justice, and Monitoring of Forest Cover and Change in t Earth Observations for Terrestrial Biodiversity and Ecosystems Special Issue 112 (5): 2495 2513. Proc Natl Acad Sci U S A 10 7 (19): 8650 55. Hansen, Matthew C., Stephen V. Stehman, Peter V. Potapov, Belinda Arunarwati, Fred Stolle, Envir onmental Research Letters 4: 12. Hansen, Matthew C., Stephen V. Stehman, Peter V. Potapov, Thomas R. Loveland, John R. G. Clearing from 2000 to 2005 Quantified by Using Multi temporal and Multiresolution Proceedings of the National Academy of Sciences 105 (27): 9439 44. Hansen, M. C., R. S. DeFries, J. R. G. Townshend, M. Carroll, C. Dimiceli, and R. A. Sohlberg. Earth Interactions 7 (10): 1 15. Guidelines for National Greenhouse Gas Inventories. v. 1: Greenhouse Gas Inventory Reporting Instructions. - v. 2: Greenhouse Gas Inventory Workbook. - v. 3: G reenhouse Applied Geography 20 (2): 155 75. - Remote Sensing of Environment 25 (3): 295 309 . Report. Cambridge, Great Britain, New York, NY, USA and Melbourne Australia: IPCC. IPCC . 2014a. Climate Change 2014: Synthesis Report. Contribution of Working Groups I, II, and III to the Fifth Assessment Report of the Intergovernmental Panel on Climate Change . Edited by Core Writing Team, R.K. Pachauri, and L.A. Meyer. Geneva, Switzerland: IPCC. 69 IPCC Climate Change 2014: Impacts, Adaptation, and Vulnerability. Part A: Global and Sectoral Aspects. Contribution of Working Group II to the Fifth Assessment Report of the Intergovernmental Panel on Climate Ch ange , edited by C. B. Field, V. R. Barros, D. J. Dokken, K. J. Mach, M. D. Mastrandrea, T. E. Bilir, M. Chatterjee, et al., 1 32. Cambridge, United Kingdom, and New York, NY, USA: Cambridge University Press. Irons, James R., John L. Dwyer, and Julia A. Ba Remote Sensing of Environment , Landsat Legacy Special Issue, 122 (July): 11 21. Fractional Green Vegetation Cover Using Spatially - Remote Sensing 4: 2619 34. Keeling, Charles D., Robert B. Bacastow, Arnold E. Bainbridge, Carl A. Ekdahl, Peter R. Tellus 28 (6): 538 51. or Biometrics 33 (1): 159 74 . Lillesand, Thomas, Ralph W Kiefer, and Jonathan Chipman. 2008. Remote Sensing and Image Interpretation . 6th ed. John Wiley & Sons. http://www.masdap.mw/layers/geonode%3Amwi_adm0. Matricardi, E. A. T., D. L. Skole, M. A. Cochrane, M. Pedlowski, and W. Chomentowski. 2007. International Journal of Remote Sensing 28 (1): 63 82. Matricardi, Eraldo A. T., David L. Skole, Marcos A. Pedlowski, Walter Chomentowski, and Luis Remote Sensing of Environment 114 (5): 1117 29. Matricardi, Eraldo A. T., David L. Skole, Marcos A. Pedlowski & Walter Chomentowski (2013) Assessment of forest disturbances by selective logging and forest fires in the Brazilian Amazon using Landsat data, Inter national Journal of Remote Sensing, 34:4, 1057 - 1086 . - Use Change in the Districts of West Kalimantan and East Kalimantan, Indonesia: Assessment of District and Forest Management Unit Wide Historical Emi Internationale Zusammenarbeit (GIZ) GmbH FORCLIME Forests and Climate Change Programme. 70 Remote Se nsing of Environment 48 (2): 119 26. Qi, J, R. C Marsett, M. S Moran, D. C Goodrich, P Heilman, Y. H Kerr, G Dedieu, A Agricultural and Fore st Meteorology 105 (1 3): 55 68. R Development Core Team. 2008. R: A Language and Environment for Statistical Computing. Vienna, Austria: R Foundation for Statistical Computing. http://www.R - project.org. Ribeiro, Natasha, Aniceto Chaúque, Faruk Mamugy, a nd Micas Cumbana. 2012. Remote Sensing of Biomass in the Miombo Woodlands of Southern Africa: Opportunities and Limitations for Research . INTECH Open Access Publisher. ping Chaparral in the Santa Monica Mountains Using Multiple Endmember Spectral Mixture Remote Sensing of Environment 65 (3): 267 79. Science 260 (5116): 1905 10. Smith, Pete, Mercedes Bustamante, Helal Ahammad, Harry Clark, Hongmin Dong, Elnour Ab Climate Change 2014: Mitigation of Climate Change. Contribution of Working Group III to the Fifth Assessment Report of the Intergovernmental Panel on Climate Change , edited by O. Edenhofer, R. Pichs - Madruga, Y. Sokona, E. Farahani, S. Kadner, K. Seyboth, A. Adler, et al. Cambridge, United Kingdom and New York, NY, USA.: Cambridge University Press. UNFAO. 2010. Global Forest Resources Assessment 2010: Main Rep ort . Food and Agriculture Organization of the United Nations. Food and Agriculture Organization of the United Nations. Tucker, C. J. 1979. Red and photographic infrared l inear combinations for monitoring vegetation. Remote Sensing of Environment, 8, 127 150. Vermote, E.F., D. Tanre, J.L. Deuze, M. Herman, and J. - IEEE Transactions on Geoscience and Remote Sensing 35 (3): 675 86 . Earth Interactions 9 (22): 1 18 . 71 Werf, G. R. van der, D. C. Morton, R. S. DeFries, J. G. J. Olivier, P. S. Kasibhatla, R. B. Nature Geoscience 2: 737 38. Williams, Darrel L., Samuel Goward, and Terry A Photogrammetric Engineering & Remote Sensing 72 (10): 1171 78. Woodcock, Curtis E., Richard Allen, Martha Anderson, Alan Belward, Robert Bindschadler, Science 320 (5879): 1011. Remote Sensing of Environment 21 (3): 311 32. Wulder, Michael A., Jeffrey G. Masek, Warren B. Cohen, Thomas R. Loveland, and Curtis E. Landsat Legacy Special Issue 122 (0): 2 10. Wulder, M ichael A, Stephanie M Ortlepp, Joanne C White, and Susan Maxwell. 2008. - 7 SLC - Canadian Journal of Remote Sensing 34 (2): 93 99. Methods for Estimating Fractional Green Vegetation Cover within a Desert - to - Upland Transition Zone in Central New Remote Sensing of Environment 98 (2 3): 237 50. - Based Cloud and Cloud Shadow D etection in Remote Sensing of Environment 118 (0): 83 94.