Assessment of supervised methods for mapping rainfall induced landslides in VHR images

In this work we develop and compare three different supervised approaches for semi-automatic mapping of landslides, including the separation of landslide source and transport areas, using a single GeoEye-1 image acquired after a rainfall-induced landslide event in Madeira Island. The methodologies cover object-based classification using support vector machine (SVM) algorithms; pixel-based classification using textons; and object-based classification with a rule-set framework. The assessment was made by comparison of the results obtained in the validation areas with the ground-truth landslide mapping. In what concerns landslide recognition, the results of the object-based and pixel-based machine-learning approaches have higher accuracy when compared with the rule-set method. The object-based SVM approach achieves false positive rate FPR=20% and false negative rate FNR=18% for landslide area detection, while the pixel-based texton method displays even higher accuracy (FPR=19% and FNR=9%) although at higher computational cost and slower execution. In what concerns internal mapping of landslide source areas, the three methods show lower but still reasonably good performance, in particular in the sunnier east-facing slopes.