Linear Spectral Mixture Analysis of Tm Data for Land-use and Land-cover Classification in Rondônia, Brazilian Amazon

The mixed pixels in remotely sensed data are one of the main error sources resulting in poor classification accuracy using traditional classification methods. In order to improve classification accuracy, linear spectral mixture analysis (LSMA) has been used to handle the mixed pixel problems. This paper aims to achieve an appropriate processing routine of LSMA through the comparison of classification results derived from different processing methods, i.e., constrained and unconstrained least-square solutions, different numbers of endmembers, different image bands used, and minimum noise fraction (MNF) transformation. A Landsat Thematic Mapper (TM) image of June 18, 1998, was used, and field data were collected in Rondônia, Brazilian Amazon. Seven classes are defined: mature forest, intermediate secondary succession (SS2), initial secondary succession (SS1), pasture, agriculture, water, and bare land (including urban areas, roads, and bare soil for cultivation). This study indicates that using constrained or unconstrained least-square solutions, atmospherically corrected or raw TM images in LSMA do not produce significant difference in the overall classification accuracy. However, reducing correction between image bands used in LSMA is useful in improving fraction quality and classification accuracy. Selection of four endmembers (green vegetation, shade, bright soil, and dark soil) and bands TM 3, 4, 5, and 7 provided the best classification accuracy. The overall classification accuracy reached 86%. This study shows that selecting appropriate endmembers and image bands is crucial for developing high quality fraction images using LSMA.