Mapping a wetland complex in the Brazilian savannah using an Ikonos image: assessing the potential of a new region-based classifier

The Pandeiros is a unique wetland in northern Minas Gerais, Brazil. Even though it is an official wildlife sanctuary, it suffers from increasing degradation because of livestock grazing and other forms of human activity. The absence of specific laws defining and protecting wetlands in Brazil is partly responsible for this situation. Having no official designation, most environments of the Pandeiros have not yet been identified or characterized as wetlands. This article presents a comparison between an unsupervised region-based classification approach and traditional supervised and unsupervised classification techniques for mapping different environments within this complex wetland system. All three classification methods were tested and their results compared with field data and with an interpreted reference map. The results suggest that the traditional supervised classification approach and the new region-based approach have good potential for identifying the main vegetal physiognomies, but the latter yielded superior results, with an estimated accuracy of 95% compared with 88% for the former. The region-based approach had the advantage of a much superior spatial consistency and allowed easier visual interpretation of the classified image. In a second accuracy assessment, we compared the results with a visually interpreted reference map and obtained total accuracies of 70%, 41%, and 39% for the region-based, maximum likelihood, and unsupervised ISODATA algorithms, respectively.