Land-cover mapping in the Brazilian Amazon using SPOT-4 vegetation data and machine learning classification methods

The main objective of this study is to evaluate the feasibility of deriving a land-cover map of the state of Mato Grosso, Brazil, for the year 2000, using data from the 1 km SPOT-4 VEGETATION (vGT) sensor. For this purpose we used a VGT temporal series of 12 monthly composite images, which were further transformed to physical-meaningful fraction images of vegetation, soil, and shade. Classification of fraction images was implemented using several recent machine learning developments, namely, filtering input training data and probability bagging in a classification tree approach. A 10-fold cross validation accuracy assessment indicates that filtering and probability bagging are effective at increasing overall and class-specific accuracy. Overall accuracy and mean probability of class membership were 0.88 and 0.80, respectively. The map of probability of class membership indicates that the larger errors are associated with cerrado savanna and semi-deciduous forest.