D2Mamba: A mamba-based method for floodway obstructions segmentation from multispectral satellite imagery

Abstract

Floodway obstructions are critical anthropogenic factors that exacerbate flood risks, playing a key role in flood control and ecological protection. Current obstruction extraction methods face two main limitations. First, high-resolution visible-light imagery is highly sensitive to lighting and terrain, which restricts its applicability in complex hydrological environments. Second, most methods cannot capture precise obstruction boundaries and fail to meet the fine-grained requirements of floodway management. To address these challenges, we propose D2Mamba, an automated obstruction extraction framework. The method first uses a dual-domain Mamba block (DDMB) to extract key obstruction features. It then employs an adaptive feature fusion module (AFFM) that uses attention mechanisms to fuse encoder features at the same level, enhancing representation capabilities. To balance dual-domain feature learning, we designed a dual-domain loss function (DLoss) to constrain joint optimization of spatial and frequency domain features. To evaluate the effectiveness of the proposed D2Mamba, we constructed a multispectral floodway obstructions dataset (MFOD), including 8301 images and corresponding masks. Comprehensive experiments showed that D2Mamba outperformed 15 state-of-the-art (SOTA) methods on MFOD, with F1, IoU, and OA exceeding other methods by 0.14–14.25%, 3.05–24.47%, and 0.05–14.78%, respectively. Furthermore, the proposed method also demonstrated robust adaptability in large-scale obstruction extraction within the Datengxia area.