Estimates of forest degradation: An algorithm based on active learning, maximum likelihood and PCA for change detection
Abstract
Land use management and control of deforestation in mega cities became an important problem due to pressure of urban sprawl. In addition, green areas and fountain-heads must be preserved in these cities, for the sake of living quality. For the purpose of monitoring deforestation and non-legal use of lands, this paper presents an approach to detect changes on the remaining areas of Atlantic Forest in São Paulo. We describe an algorithm for image classification and change detection based on the active learning, which is improved by a pre-clustering task, then a change detection step based on histogram analysis of the second PCA (Principal Component Analysis) component. We applied the algorithm over 4 multispectral Landsat 5/TM images representing the years 1986, 1996, 2003, 2011 of an eastern area of São Paulo city. With the proposed active learning technique we were able to categorize the pixels used in the change detection step. We estimated that the remaining 31% of forest in 1986 reaches a minimum of 25% in 2003, but afterwards it recovered to 27% of the total area in 2011.