SD-Mamba: A lightweight synthetic-decompression network for cross-modal flood change detection
Year: 2025
Type: article
Abstract: Cross-modal flood change detection using optical and SAR images has become one of the most commonly used techniques for monitoring the progression of flooding events. Existing methods fail to adequately capture the interrelationship between semantics and changes, which limits the potential for effective flood detection. To address this issue, we propose a lightweight Synthetic-decompression network. The synthetic component is divided into four stages, each of which employs a Multi-branch Asymmetric Part-convolution block (MAPC) and a Temporal Semantic Interaction module (TSIM) to extract semantic features from dual-temporal images. Subsequently, these features are fed into the Temporal-mamba (T-Mamba), which uses 4D Selective Scanning (SS4D) to traverse temporal change information in four directions. The decompression component employs a three-stage Asymmetric Coordinate-convolution block (ACoord-Conv) to project the change results onto the source images, thereby indirectly supervising the model’s detection performance. Compared to the 22 state-of-the-art (SOTA) lightweight methods, SD-Mamba achieves an optimal balance between computational efficiency and detection accuracy. Under the same computational conditions, SD-Mamba demonstrated superior performance to other Mamba-based models, with an improvement of 1.01% in mIoU, while maintaining a lightweight structure with only 5.32M parameters and 12.24G floating-point operations (FLops). The code is available at https://github.com/yaoshuang-yaobo/SD-Mamba . • A lightweight SD-Mamba network for cross-model flood change detection. • A twin backbone composed of MAPC and TSIM extracts semantic features. • A T-Mamba traverses the changing sequence using a 4D selective scanning strategy. • A D-Net to decompress the change maps to the source images.
Authors: Yu Shen, Shuang Yao, Zhenkai Qiang, Guanxiang Pei
Institution: Lanzhou Jiaotong University
Language: English
Subfield: Media Technology
Field: Engineering
Domain: Physical Sciences
SDG: Climate action
Open Access status: gold
Estimated APC paid: $3,550
OpenAlex ID: W4407481955