Extraction of land use map under arid and semi-arid conditions using frequency-based classifier

摘要

In this study a cover frequency-based classifier approach was tested in an attempt under arid and semiarid conditions to extract a land use information map from TM multispectral images. This classification procedure involves two stages. In the first stage, the TM data were first classified into 14 land cover types using a supervised maximum-likelihood classifier (MLC). In the second stage, the pixel window size 7*7 was moved all over the land cover map obtained from the first stage to extract the cover frequency tables for each of the 7 land use classes. These frequency tables were then employed in the classification of 7 land use classes using a supervised minimum city-block distance classifier. Also the study area was classified using the conventional MLC method for comparison. For each land use map, a confusion matrix was obtained by comparing the classification results for the test samples with the ground truth map. A Kappa Index of Agreement (KIA) was compared for two classification error matrices. Also the accuracy of the individual category was measured using a coefficient of condition Kappa. The overall accuracy measured by Kappa Index of Agreement (KIA) was 0.86 for MLC method. It was significantly improved to 0.91 with the cover frequency method.