Determining the Suitability and Accuracy of Various Statistical Algorithms for Satellite Data Classification

Abstract

Land use and land cover (LULC) data is very important for determining the nature and mechanism of different land surface, hydrological processes. The production of land use land cover map, using an image classification is one of the most common applications of remote sensing. However, image classification is a complex process that may be affected by many factors including spatial resolution, classifier used, training sets, etc. This paper briefly reviews the suitability of different methods of classification that are commonly used and their impact on classification accuracy. Three different supervised classification techniques (Maximum likelihood, Mahalanobis Distance, and Minimum Distance) were applied in Kashmir valley for the classification of the IRS LISS-III (2008) image in thirteen different LULC classes; agriculture, aquatic vegetation, barren land, built-up, exposed rock, forest, horticulture, pastures, plantation, riverbed, scrub land, snow, and water. The classified maps where then visually compared with each other and the accuracy of classified map was assessed using the reference data sets which consisted of a large number of ground samples collected in each land cover category. The overall accuracy for Maximum likelihood classifier was 89%, for Mahalanobis distance was 54% and for Minimum distance was 48%. It was observed that the Maximum likelihood method gave the best results and good agreement between classes extracted from the classified maps and field observations. Mahalanobis distance method has overestimated agriculture land, plantation and built-up. Minimum distance method overestimated overestimated water, built-up and horticulture and underestimated agriculture. The selection of a suitable classification method is significant for improving classification accuracy.