Automatic Landslide Detection with Satellite Imagery using Attention-boosted CNN Model

A Landslide is a dangerous geological event that causes an intimidating impact on human life and demolition of infrastructure anywhere in the world. Landslides occur very frequently in the rainy season when large masses of rock, debris, and soil move downward a slope due to some natural phenomenon and some man-made activities. In a catastrophic emergency, reliable and efficient landslide detection can provide logical information to save life. To meet the requirement of relief operations with accuracy and in time, this research proposes an automatic landslide detection with satellite images instead of the site visit process in the traditional approach. Convolutional neural networks (CNN) are much more efficient than traditional methods in detecting landslides by reducing the time required to identify relevant features. The implementation of the attention module in CNN models in remote sensing image processing can improve the global context modeling and feature detection of the model. This work uses deep learning CNN with attenuation mechanisms and optimization to extract landslides from satellite images to automatically identify landslides. The proposed method is divided into three steps: augmentation of labeled datasets; introduce the attention module in the decoder to suppress press feature map noise with three different CNN backbone networks (ResNet50, ResNet101, Google Net) for training and performance evaluation of the proposed algorithm on accuracy, precision and F1 score of three parameters. The result indicates that the proposed CNN attention module model with ResNet 101 as a backbone network has high precision, precision, and the F1 score can be used as an effective landslide detection method to help in emergency rescue.