A Landslide Detection Network Integrating Mamba and Deformable Sliding Window Attention

Abstract

Landslide detection is essential for geological disaster mitigation, yet existing deep learning methods still struggle with the high cost of global feature modeling, limited receptive fields in window-based attention, and insufficient fusion of local and global information. To address these challenges, we propose MTTNet, an efficient framework for multisource landslide extraction that integrates Mamba state space modeling with a deformable sliding window attention (DSWA) mechanism. MTTNet employs a visual state-space (VSS) block to capture long-range dependencies and a DSWA-based encoder that assigns different convolutional kernel sizes to attention heads, enabling the extraction of high-resolution and semantically rich features. To reduce global computational overhead, we design an MTC module that generates informative tokens using both SSM and Transformer structures, avoiding dense pairwise interactions. A VSS-based decoder further enhances generalization across varying image resolutions. Experiments on multiple public datasets demonstrate that MTTNet consistently outperforms state-of-the-art methods.