E-mambaSR: a multiexpert selection scanning mechanism for remote sensing image super-resolution

Remote sensing image super-resolution reconstruction plays a crucial role in improving image resolution, enhancing the accuracy of ground feature recognition, and optimizing resource management. Existing models, mainly based on CNNs and transformers, still have limitations in global receptive fields and computational efficiency. The recently proposed Mamba model uses a linear complexity mechanism to capture long-range dependencies while reducing computational costs. However, the existing Mamba method relies on a single scanning approach, which limits the model’s ability to capture complex contextual relationships in remote sensing images. Therefore, we propose the E-mambaSR model based on an expert selection scanning mechanism. First, the routing instruction network extracts semantic information from the image. Then, the expert routing network selects specific scanning strategies based on the extracted image information, enabling multiexpert scanning beyond a single approach. In addition, we introduce a contrastive learning strategy, leveraging effective data augmentation to provide rich positive and negative samples, improving high-frequency information recovery, and avoiding overly smooth images. This method enhances feature sequence modeling and improves the Mamba model’s ability to reconstruct textural details. Extensive experiments on AID, DIOR, and WHU-RS19 datasets show that E-mambaSR improves peak signal-to-noise ratio and structural similarity index measure by 0.35 dB and 0.0089, respectively, compared with the VmambaIR model.