ETGC2-net: an enhanced transformer and graph convolution combined network for landslide detection

Landslide detection is one of the crucial tasks in geological hazard prevention and control. Accurate detection and prediction of landslide areas contribute to taking appropriate preventive and mitigation measures in advance. However, it still faces challenges in terms of timeliness and accuracy. Our design motivation is to seek a network that can demonstrate excellent performance and relatively low computational cost in practical landslide area detection tasks. The proposed method in this paper can simultaneously capture local information and global dependencies in the images while achieving outstanding performance with fewer parameters. Moreover, considering the spatial correlation of landslides in remote sensing images, we introduce graph convolution to model the internal topological relationship within areas, with superpixels serving as guidance. Extensive experiments on two publicly available landslide datasets demonstrate that our network outperforms traditional and existing state-of-the-art methods while achieving the optimal balance between efficiency and performance, thus exhibiting broad potential for practical applications.