Classification of remote sensing data has long been a fundamental technique for studying vegetation and land cover. Furthermore, land use and land cover maps are a basic need for environmental science. These maps are important for crop system monitoring and are also valuable resources for decision makers. Therefore, an up-to-date and highly accurate land cover map with detailed and timely information is required for the global environmental change research community to support natural resource management, environmental protection, and policy making. However, there appears to be a number of limitations associated with data utilization such as weather conditions, data availability, cost, and the time needed for acquiring and processing large numbers of images. Additionally, improving the classification accuracy and reducing the classification time have long been the goals of remote sensing research and they still require the further study.To manage these challenges, the primary goal of this research is to improve classification algorithms that utilize MODIS-EVI time-series images. A supervised self-organizing map (SSOM) and a soft supervised self-organizing map (3SOM) are modified and improved to increase classification efficiency and accuracy. To accomplish the main goal, the performance of the proposed methods is investigated using synthetic and real landscape data derived from MODIS-EVI time-series images. Two study areas are selected based on a difference of land cover characteristics: one in Thailand and one in the Midwestern U.S.The results indicate that time-series imagery is a potentially useful input dataset for land cover classification. Moreover, the SSOM with time-series data significantly outperforms the conventional classification techniques of the Gaussian maximum likelihood classifier (GMLC) and backpropagation neural network (BPNN). In addition, the 3SOM employed as a soft classifier delivers a more accurate classification than the SSOM applied as a hard classifier. Furthermore, the 3SOM-F, which applies both pure and mixed pixels during the training process, accomplishes more accurate and realistic classification results than the 3SOM-P, which applies only pure pixels in the training process. Therefore, these results suggest that the 3SOM-F could be considered the most appropriate method for land cover classification using time-series imagery. However, the results also demonstrate that there is uncertainty in the classification accuracy associated with network design architecture and internal parameter settings. As a result, the suitable neural network configuration should be investigated for the best performance of the classifier. Additionally, two study areas, Thailand and the Midwestern U.S., are selected to investigate the performance of the 3SOM-F. All results confirmed that the classification performance of the 3SOM-F is effective even when it is applied to real landscape data in both study areas.The proposed techniques will benefit detailed land cover classification at the regional scale. The spatial pattern of land cover classes can be valuable information for managing and understanding the environment as well as monitoring land cover change. Furthermore, the advantages of this research will contribute to various disciplines such as map updating, agricultural area estimation, cartography, and urban planning.
Read
