Utilizing deep neural networks for landslide detection and segmentation in remote sensing imagery

Understanding the occurrences of historic landslide events is crucial for supporting strategies aimed at reducing disaster risks. Drawing from insights obtained in the 2022 Landslide4Sense competition, we present a methodological framework reliant on a deep neural network design for the detection and segmentation of landslides using input from various remote sensing sources. Our approach involves using a U-Net architecture, initially trained with cross entropy loss, as a baseline. We then enhance this architecture by employing diverse deep learning techniques. Specifically, we engage in feature engineering by creating new band data derived from the original bands, thereby improving the quality of the remote sensing image input. Concerning the network architecture, we substitute the conventional convolutional layers in the U-Net baseline with a residual-convolutional layer. Additionally, we introduce an attention layer that capitalizes on a multi-head attention scheme. Furthermore, we generate multiple output masks at three distinct resolutions, forming an ensemble of three outputs during the inference process to augment performance. Lastly, we propose a composite loss function that integrates focal loss and IoU loss to train the network effectively. Our experiments on the Landslide4Sense challenge's development set yield an F1-score of 84.07 and an mIoU score of 76.07. Our optimized model surpasses both the challenge baseline and the proposed U-Net baseline, improving the F1-score by 6.8/7.4 and the mIoU score by 10.5/8.8, respectively.