Research on landslide detection method based on improved VMamba-UNet+

As remote sensing technology improves, the segmentation of landslide targets has become more important in disaster prevention, control, and urban construction, and it plays a significant role in disaster loss evaluation and post-disaster rescue. Therefore, this paper presents an improved UNet-based landslide segmentation algorithm. China has a large proportion of land in landslide-prone areas, and remote sensing technologies are becoming a preferred method for investigating and monitoring landslides. Rapid and precise recognition of landslide-prone regions through high-resolution remote sensing images is essential for disaster reduction. Although Convolutional Neural Networks have made progress in automatic detection, they have small receptive fields which can't capture long-range spatial dependencies. Conversely, Transformer-based methods provide global modeling capability but suffer from quadratic computational complexity, making it difficult to implement real-time processing. In order to solve such problems, this paper puts forward VMamba-UNet+, a new kind of light-weighted landslide detection network based on Visual State Space Model (VMamba). It uses CSM to get global features with linear computation cost. Additionally, it includes Dynamic Snake Convolution (DSC) to improve the extraction of features from irregular landslide boundaries, as well as the Spatial Group-wise Enhance (SGE) module to efficiently suppress complex background noise. On the landslide dataset, VMamba-UNet+ gets an mIoU of 0.8716 and an F1-score of 0.9278. These numbers show that the suggested model keeps low computing costs and does a great job at separating things, giving a clear and not using much power way to find landslides fast.