Attention driven deep convolutional network with optimized learning for accurate landslide detection and monitoring

Abstract

Abstract Effective landslide monitoring is essential for mitigating risks to infrastructure and communities, particularly in geologically unstable regions. Traditional monitoring methods, such as ground surveys and visual inspections, are time-intensive and lack early detection capabilities. To address these limitations, this study employs feature fusion and enhanced Deep Convolutional Neural Networks (DCNNs) for landslide detection. The model is built upon a fine-tuned, pre-trained VGG16 architecture, adapted to a new landslide dataset. Key modifications include the integration of a spatial attention mechanism, optimized learning rate schedules, attention-based Global Average Pooling (GAP), and the Lookahead Adam optimizer, all aimed at improving feature extraction, model convergence, and generalization. Experimental results demonstrate that the proposed approach achieves high accuracy, with performance ranging from 90% to 96% across different datasets and training iterations. Using the Kaggle Landslide Dataset, the model attained a training accuracy of 93%, with validation and testing accuracies of 95.2% and 95.8%, respectively. Comparable results were observed with the NASA Landslide Inventory, confirming the robustness of the method. The findings highlight the potential of DCNN-based models, augmented with attention mechanisms, as a reliable and efficient tool for landslide monitoring, significantly outperforming conventional assessment methods.