Automatic landslide detection from remote sensing images using supervised classification methods

The creation of a landslide inventory map by manual interpretation of remote sensing images is very time-consuming. This study aims at developing an automated procedure for the detection of landslides from multi-spectral remote sensing images. According to the type of landslide, the parameters for detecting the slope instabilities will differ. In a first step, predefined input parameters derived from the images are incorporated in a supervised pixel classification algorithm. In this study, we use a maximum likelihood classification method, which shows positive preliminary results. In order to evaluate the accuracy and applicability of the method, the results are compared with ANN classification. Segmentation of the output image (containing likelihood values to be a landslide) into landslide and nonlandslide areas is conducted by using the double threshold technique in combination with a histogram-based thresholding. Additional filtering of the detected blobs based on shape and geomorphologic properties allows to eliminate spurious areas. Validation of the results is done by comparison with manually defined landslides.