A Method on Land Cover Classification by Combining Unsupervised Algorithm and Training Data

Abstract In this paper, a method on land cover identifying by combining unsupervised algorithm and training data (CUT) was developed. The procedures of land cover classification by using the CUT method are: (a) to carry out remotely sensed image classification by using an unsupervised algorithm (e.g. ISODATA unsupervised classification) to make a land cover classification map, MAP1, with n classes, where n is much greater than the proposed number of land cover classes, m, in the study area; (b) to collect training data for each of the proposed m classes; (c) to make a mask by using training data sets and statistically compute MAP1; (d) to assign the class h in MAP1 to class c in the final classification map, MAP2, if and only if the number of pixels in class c is with the maximum ratio at the statistic. The CUT method was also used to produce a land cover classification map in a test area, Ansan City of Korea, with Thematic Mapper (TM) data acquired by Landsat‐5. The accuracy analysis on the classificatio...