Abstract
In recent years, remote sensing technologies have played a crucial role in the detection and management of natural disasters. In this context, deep learning models are of great importance for the early detection of natural disasters such as landslides. Landslide segmentation is a fundamental tool for the development of geographic information systems, natural disaster management and risk mitigation strategies. In this study, we propose a new semantic segmentation model called LandslideSegNet to improve early intervention capabilities for potential landslide scenarios. LandslideSegNet incorporates an encoder-decoder architecture that integrates local and contextual information, advanced encoder-decoder residual blocks and Efficient Hybrid Attentional Atrous Convolution. Thanks to this structure, the model is able to extract high-resolution feature maps from remote sensing imagery, accurately delineate the landslide areas and minimize the loss of contextual information. The developed LandslideSegNet model has shown significantly higher accuracy rates with fewer parameters compared to existing image segmentation models. The model was trained and tested using the Landslide4Sense dataset specially prepared for landslide detection. LandslideSegNet achieved an accuracy of 97.60% and 73.65% mean Intersection over Union of 73.65 on this dataset, demonstrating its efficiency. These results indicate the potential usability of the model in landslide detection and related disaster management applications.