Comparative analysis of classification algorithms and multiple sensor data for land use/land cover classification in the Brazilian Amazon

Abstract

A comparative analysis of land use/land cover (LULC) classification results in the Brazilian Amazon based on four classification algorithms and four remote sensing datasets was conducted in order to better understand the selection of a classification algorithm suitable for a specific remote sensing data. It is shown that maximum likelihood classifier (MLC) provided reasonably good classification accuracy when Landsat Thematic Mapper (TM) or the TM and Advanced Land Observing Satellite Phased Array type L-band Synthetic Aperture Radar (ALOS PALSAR) data-fusion images were used, but nonparametric algorithms such as classification tree analysis for TM multispectral bands and K-nearest neighbor for the combination of TM and PALSAR data provided better classification than MLC. Individual PALSAR dataset is not suitable for detailed LULC classification and has much poorer classification accuracy (47.6% to 59.4%) than Landsat TM image (79.7% to 84.9%). However, integration of TM and PALSAR data through the wavelet-merging technique improved classification accuracy. It is implied that the importance of selecting a suitable classification algorithm for a specific dataset by considering such factors as overall classification accuracy and time and labor involved in a classification procedure. Important information for guiding the selection of remote sensing dataset and associated classification algorithms for LULC classification in the moist tropical regions is also provided.