Mass movements’ scars classification using data mining techniques

Mass movements are destructive natural phenomena that can lead to serious problems such as economic loss, damage to natural resources and even injuries and deaths. Efforts have been made to semi automate the interpretation of remote sensing data in order to improve efficiency and support specialists in recognizing mass movements’ scars. However, this approach is still incipient in Brazil. This study presents results of semiautomatic classification of mass movements’ scars that occurred in Nova Friburgo (Rio de Janeiro state, Brazil) in 2011 by using segmentation and applying data mining techniques. Two classifications were compared, from C4.5 and CART decision tree algorithms. Data mining techniques confirmed that mass movements have different spectral characteristics from other classes, allowing its detection from remote sensing images. The overall accuracy of C4.5 algorithm was 62.6%, while CART was 66.4%. The errors occurred mainly in urban areas and in unpaved roads located at higher altitudes. Spectral digital elevation model (DEM) average, blue band and NDVI were the more appropriate attributes to distinguish mass movements patterns. This methodology offered an alternative, that still needs improvements, to produce data about statistics and spatial distribution of mass movements, providing information to be used, for instance, as parameters in susceptibility maps and models, assisting public policies focused on natural disasters.