Primary forest and land cover contextual classification using JERS-1 data in Amazonia, Brazil

The authors present a methodology for land cover and primary forest mapping in Amazonia using textural features derived from JERS-1 data and classified with a multilayer perceptron based contextual method. Land cover classification is an important step towards the use of radar data as a tool for land use change studies in Amazonia. Also, primary forest classification is an important issue in ecosystem studies and economical assessment of sustainable timber exploitation. The use of radar data, particularly L-band data, is justifiable as large Amazonian area is permanently cloud covered. Considering a set of primary forest and land use classes of interest in the Tapajos National Forest and adjacent regions, Para State, Brazil, it was investigated which classes could be distinguished using textural features derived by co-occurrence and matched filtering techniques. Nondiscriminating classes were grouped together to form new classes resulting in two classes of primary forest, three classes of land use, water and aquatic vegetation. The feature set with higher overall accuracy was used to classify a small mosaic of the region, using a contextual neural network based classifier with 87% overall accuracy.