Land use mapping is essential for the understanding of global change processes, especially in regions which are experiencing great pressure for development such as the Amazon. Traditionally, these mappings have been done using visual interpretation techniques of satellite imagery, that provide satisfactory results but are time-consuming and highly cost. In this paper, a technique of image segmentation based on region growing algorithm, followed by a per-field non-supervised classification, is proposed. Thus, the thematic classification is based on a set of image elements (pixels), benefiting from contextinformation, therefore minimizing the limitations of the digital processing techniques based on single pixels (per-pixel classification). This approach was evaluated in a typical test site of the Amazon region located to the north of Manaus, AM, using both original Landsat Thematic Mapper images and their decomposition into endmembers such as green vegetation, wood material, shade and soil, named mixture image in this paper. The results were validated by a reference map obtained from proved visual interpretation techniques of satellite imagery and by field check and indicated that automatic classification is feasible to map land use in Amazonia. Statistics tests indicated that there was significant agreement between the automated digital classifications and the reference map (at 95% confidence level).
Automated thematic mapping; image segmentation; per-field non-supervised classification; land use change; remote sensing