- Research Article
- 10.1080/01431161.2025.2557591
DBENet: dual-branch encoder network based on the visual state space model for semantic segmentation of remote sensing image
- Sep 28, 2025
- International Journal of Remote Sensing
- Hengyou Wang + 4 more +4
Semantic segmentation is a fundamental issue for remote sensing images. However, modelling global and local information is a challenging task. Compared to the insufficient receptive field of traditional CNNs and the high complexity of Transformers, Mamba, built on the state space model (SSM), has emerged as an alternative for establishing long-distance dependency relationships while maintaining linear computational complexity. Thus, to better capture global and local information, we propose a remote sensing semantic segmentation method based on the Visual State Space Model, simply referred to as the dual-branch encoder network (DBENet). It employs the cross-scanning mechanism to calculate attention. Specifically, we integrate ResNet and visual state space models as dual branches into the encoder to capture global and local features. Additionally, we design the Global-local Transformer Block (GLTB) and the Feature Enhancement Block (FEB) to enhance global and local features. In the GLTB, we put the dual input into respective enhancement components. One input is used to enhance the global feature extracted by VSS (Visual State Space), while the other is utilized to enhance the local feature extracted by ResNet. The FEB employs a 1 × 1 convolution to enhance the features along the channel dimension. Extensive experiments on three public datasets, ISPRS Vaihingen, ISPRS Potsdam and LoveDA (Urban), demonstrate the effectiveness and potential of the proposed DBENet. The proposed method achieves better performance compared to state-of-the-art methods. The source code will be released after publication at https://github.com/MathMaths-liuh/MathMaths-liuh.
Read more