Multitemporal spectral mixture analysis for Amazonian land-cover change detection

The complex landscape and environmental conditions in the moist tropical region often result in poor land-cover change detection accuracy using traditional change detection methods. This paper explores linear spectral mixture analysis (LSMA) of multitemporal thematic mapper (TM) images to detect land-cover change in Rondonia, Brazilian Amazon basin. Three image endmembers (shade, green vegetation, and soil) were developed based on a combination of field data and image scatterplots. An unconstrained least-squares solution was used to unmix the multitemporal TM images into three fractions. Then, fraction image differencing results were used to analyze land-cover change/non-change detection. The detailed "from-to" change detection was implemented using a pixel-by-pixel comparison of classified images, which were developed using a decision tree classifier on the multitemporal fraction images. This study indicates that LSMA is a powerful image processing tool for land-cover classification and change detection. The multitemporal fraction images can be effectively used for land-cover change detection. The stable and reliable multitemporal fraction images developed using LSMA make the change detection possible without the use of training sample datasets for historical remotely sensed data. This characteristic is particularly valuable for the land-cover change detection in the Amazon basin.