Publications

Abstract : High resolution satellite images serve as an eye in the sky and are helpful in generating landslide inventories. Satellite images of landslides can be used as an input for deep learning models that are specifically designed for image processing and assess the risk posed by landslides on communities. Fully satellite image-based landslide risk evaluation is still a new venture that needs a lot of improvements before it can be deployed in real-time. In this work, we assess landslide hazard to communities near debris scars of slopes that have experienced landslides in the past, using three pre-trained deep learning algorithms. We followed a 70:30 scheme for training and validation with over 2000 images of two classes: Urban and Debris. Finally, the trained neural network is tested on images that contain human settlements near the debris scars. Based on class probabilities predicted by the algorithm, the sites were ranked for eminent risks in the future. It was observed that the validation accuracy of Alexnet, Resnet and NASNet-Large were 92%, 96% and 98%, respectively. The risk classification for four test images indicates: Alexnet and NASNet-Large extracted features same as observed in the satellite images however, Resnet50 is very sensitive to the features and could not predict the same as observed in satellite images. Given the accuracy of predictions, such algorithms can be further modified and deployed to create landslide hazard risk maps for various communities around the world.