Comparison of Land-cover Classification Methods in the Brazilian Amazon Basin

Numerous classifiers have been developed and different classifiers have their own characteristics. Controversial results often occurred depending on the landscape complexity of the study area and the data used. Therefore, this paper aims to find a suitable classifier for the tropical land cover classification. Five classifiers – minimum distance classifier (MDC), maximum likelihood classifier (MLC), fisher linear discriminant (FLD), extraction and classification of homogeneous objects (ECHO), and linear spectral mixture analysis (LSMA) – were tested using Landsat Thematic Mapper (TM) data in the Amazon basin using the same training sample data sets. Seven land cover classes – mature forest, advanced succession forest, initial secondary succession forest, pasture, agricultural lands, bare lands, and water – were classified. Overall classification accuracy and kappa analysis were calculated. The results indicate that LSMA and ECHO classifiers provided better classification accuracies than the MDC, MLC, and FLD in the moist tropical region. The overall accuracy of LSMA approach reaches 86% associated with 0.82 kappa coefficient.