Lightweight deep learning model for automatic landslide prediction and localization

There has been a lot of interest in utilizing remote sensing images to anticipate landslides. We propose a novel framework for automatic landslide detection and landslide region localization from the input remote sensing image. The framework consists of pre-processing, dynamic segmentation, automatic feature extraction, classification, and localization. The pre-processing is the integrated step that performs atmospheric corrections, geometric corrections, and unnecessary region removal with denoising using 2D median filtering. The pre-processed image is then segmented using the dynamic segmentation approach to extract the Region of Interest (ROI). We propose lightweight Convolutional Neural Network (CNN) layers for automatic feature extraction and scaling using the ResNet50 model. The CNN layers are designed systematically for automatic feature extraction to improve accuracy and reduce computational requirements. The Long-Term Short Memory (LSTM), Artificial Neural Network (ANN), and Support Vector Machine (SVM) classifiers are designed to perform the landslide prediction. If landslides are forecast, the post-processing stages are intended to identify potential landslide locations. The experimental results show that the proposed CNN-LSTM model outperformed the existing solutions in terms of accuracy, F1 score, precision, and recall rates. The experimental outcomes reveal that the proposed model improves the overall prediction accuracy by 2% and reduces the computational complexity by 35% compared to state-of-the-art methods.