Landslide detection from satellite imagery using a deep learning technique

Taiwan is subject to severe natural hazards like earthquakes and typhoon, which often cause landslides in mountainous area, claiming crops, property safety and even lives. Monitoring the occurrence of landslides using remote sensing is an annual task for government institutions. However, the task had been extremely labor-intensive and time consuming. In order to solve the problem, this study proposed a deep learning technique for automatic landslide classification from satellite imagery in order to get a more accurate and robust classification results. The classification model is based on the U-Net convolutional neural network, implemented with the deep learning toolkit, CNTK. The model takes pairs of satellite imagery and ground truth label as the input and produces predicted classified labels as the output. The model is trained on pairs of FORMOSAT-2 imagery and ground truth labels. The ground truth is classified into 5 classes: vegetation, riverbed, landslides, water and miscellaneous. To best separate landslides from other impervious land cover like riverbed and farmlands, slope degree is added to satellite imagery to provide distinguish information for classification. Imagery from Kompsa-3 is used to test the sensor independency of the model. The study produces a result of a robust classification model that is able to distinguish landslides from the satellite imagery. The landslides classification results are accurate for FORMOSAT-2 imagery. The model is also able to produce similar results on Kompsat-3 imagery when scaled to the image depth of 8-bit despite of the difference in spatial resolution. The model is reusable, and the process is fully automated. We expect the model will be useful on landslides monitoring and inventory mapping, which are elementary task for hazard mitigation and susceptibility mapping.