Monitoring land use in Amazonia based on image segmentation and neural networks

The objective is to show the methodological steps of the combined use of segmentation and thematic image classification by neural systems. The area under study is the region around the Tucurui electric power plant (SE of Para State). The image segmentation is based on the spectral characteristics, using the region growing algorithm. Each segment is labeled with thematic classes: 1. Basic categories: forest, regrowth (initial and advanced), crop, pasture, water; 2. Interfering categories: shadow and clouds. During the labelling of these segments, partial decision factors are assigned, among the Boolean concepts of “false” and “true”, within the fuzzy logics approach. For the architecture of this classification system, the following descriptors are used: spectral, geometric, textural and contextual. The segments labelled as training areas are used to feed and feedback this neural system. Results of the identification of deforestation in Amazonia, presented an overall performance above 92%. It is well-known that the landuse classes in Amazonia are quite complex. Thus, one can expect that the thematic classification by neural networks, would allow the definition of transition phenomena such as different types of pasture, different stages of regrowth, among others. The basic concept of this integrated image analysis is the use of segments of the scene, as the main information source, instead of pixels. This procedure allows a high degree of trust for the digital classification of images from Amazonia