Automated Detection of Small and Shallow Landslides after the 2010 Madeira Island Flash-Floods in VHR Imagery
Heleno, Sandra; Lousada, Maura; Pereira, Maria João; Pina, Pedro
CERENA, Instituto Superior Técnico, PORTUGAL
In recent years, semi-automated methods have been developed to map landslides from remote-sensed imagery. The number and variety of approaches is increasing, but the performances reported on the literature, highly dependent on the spectral and spatial features of the images and also on landscape characteristics, are very uneven. This still is an open problem, where novel methodological contributions to automatize the detections with a high degree of confidence are welcome, namely in regions where the landslides are frequent and their number particularly high. This is a common situation in Madeira Island (Portugal) where recently thousands of small and shallow landslides were triggered by heavy rainfall on February 20 2010, resulting in fast transport of solid material with tragic consequences. Assessing its hazard through landslide inventory maps is therefore essential for risk management purposes. These numerous, small and shallow types of landslides present a challenge to image classifiers and this study addresses such difficulties, reporting with detail the successful and critical situations. We use very high spatial resolution images (GeoEye-1, 0.5m/pixel and orthophotos with 0.4m/pixel) acquired before and after the 2010 triggering event in Madeira Island, to test several automated detection methods, seeking for a robust approach able to produce a detailed and complete identification of landslides. We perform an accurate segmentation (delineation of contours) of all the structures (objects) resolved in the remotely sensed images, aiming at landslides scars as small as about 4-6 m2 in area. The classification of segmented objects into landslides/non-landslides is obtained through the use of adaptive methods (such as Support Vector Machine) that are able to learn with examples provided during the training phases. We present our preliminary findings, supported and validated by in-situ measurements in different locations of Madeira Island.