Attention U-Mamba: A Simple and Efficient Method for Landslide Segmentation

Abstract

Landslides cause significant casualties and property damage worldwide. Integrating optical remote sensing with deep learning is crucial for effective landslide segmentation. This study introduces attention U-Mamba (AUM), a novel approach combining state-space models (SSMs) with a U-shaped network. AUM leverages CNNs for local feature extraction and Mamba for global context, benefiting from Mamba’s linear complexity to reduce parameters while enhancing performance. Evaluated on a public landslide dataset against seven state-of-the-art methods, the AUM achieves state-of-the-art performance with only 15.89 M parameters—60% fewer than DeepLabV3 (39.63 M)—while attaining an $F1$ score of 87.81%, mIOU of 79.82%, and precision of 84.84%, demonstrating superior efficiency and accuracy.