Comparative sensitivity analysis of kernel parameters for support vector machines on land use classification: A case study of rapideye and spot

The development level of countries and living of societies in a high welfare level are based on efficient use and sustainable management of their social, cultural and economical resources. Based on CLC 2006 data, the land in Turkey consists of 42.35% agricultural areas and 54.04% forest and semi natural vegetation.Therefore, reliable and detailed information regarding LULC are critical important for environmental decision makers and local authorities on strategic planning and sustainable resource management of country's land. Remote sensing technology has an important role in efficient and proper planning and management of natural resources. Multispectral high resolution images are preferred and commonly used since providing the spectral, spatial and temporal information while producing thematic maps. Image classification is the most common way to gather thematic information, among the all collateral methods. The quality and accuracy of the thematic information are based on algorithm used as well as satellite image and its band combination. One of the machine learning algorithms SVMs which are commonly used in recently since its high classification capability and one of the conventional classification algorithms MLC whose validity has been accepted in literature are preferred in this study. Multispectral high resolution RapidEye and SPOT images are used in this study since they have been effectively used for vegetation, forest and wetland mapping and gave successful results. These images has been preferred since their band features obtained besides their high resolution The study area is located Aydin province and covered by agricultural lands. Study area has 12 land use classes as they are corn (I.crop, II.crop, III.crop), cotton (well-developed, moderate-developed, weak-developed), soil ( wet, moist, dry), pasture and brush land mixed area,settlement area and water body. In-situ data was collected at the acquisition date of satellite images by experts from Department of Soil Science and Plant Nutrition in Ege University, İzmir. Four different types of kernel and 139 different models have been used for SVM classification.The effects on choice of kernel and its internal parameters have been investigated and optimum parameters having highest classification accuracy have been determined.Moreover, the success and efficient use of satellite images and classification algorithms on agricultural lands have been examined. SVM classification results based on selected various model have superior performance than MLC in most cases, however MLC is better than SVM rarely in a few cases. For both type of satellite images, the polynomial kernel had the highest classification accuracies which are 85.6% and 73.3% for RapidEye and SPOT, respectively. Keywords: Support vector machines, kernel functions, comparative sensitivity analysis, RapidEye, SPOT