Object-based Classification of High Spatial Resolution Remote Sensing Images in Ethiopia Using Machine Learning Approaches

GIS;remote sensing;School for Environment and Sustainability

Lu, ChuyingJain, Meha ;

University of Michigan

关键词: GIS;    remote sensing;    School for Environment and Sustainability;   

Others  :  https://deepblue.lib.umich.edu/bitstream/handle/2027.42/148830/Chuying_Lu_Thesis.pdf?sequence=1&isAllowed=y

瑞士|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

Remote sensing image classification is the important process of extracting land use and land cover (LULC) information and has been widely used in a range of fields. With the availability of high spatial resolution images, object-based image analysis together with machine learning classification algorithms has received increasing attention and use.The main goal of this research is to conduct supervised object-based classification experiments based on Random Forest (RF) and Support Vector Machine (SVM) on high spatial resolution images in Benishangul (BG), Gambella (GM), Oromia (OR), Ethiopia. Performance of the classifiers were compared through analyzing the classification results. Multi-variate linear regression models were built to explore the relationships between factors and classification performance. Two questions were addressed: Are SVM or RF appropriate to be applied to mapping LULC in Ethiopia? and What factors influence classification results? Another objective was to explore the possibility to improve classification performance in terms of accuracy of features extracted. Temporal features were included and the effectiveness of which was examined. When trained the data without temporal features, the mean overall accuracy is 0.72 for SVM, 0.74 for RF. The effectiveness of the two classification approaches differed by site. They were significantly difference in OR and GM, where SVM overperformed RF. Because the dataset was unbalanced, SVM had an advantage. The results of the linear regression analysis suggested that the area of class and sample counts had notable impacts on classification performance. Inclusion of temporal features improved results when using SVM, but had little influence on RF.