Remote sensing image super-resolution via hybrid mamba and convolutional architectures

Lightweight remote sensing image super-resolution (RSISR) aims to generate high-resolution remote sensing images (RSIs) with improved texture and structural details while keeping computational cost low. Most lightweight RSISR methods use convolutional neural networks (CNNs), which are excellent at capturing local spatial features but struggle to model long-range dependencies because their receptive fields are naturally limited. Even though Transformer-based methods address this problem by using global attention mechanisms, they aren't very practical for lightweight, large-scale remote sensing applications because they are complex and require many parameters. Mamba, a state-space-based sequence modeling framework, has recently demonstrated that it can model long-range data effectively with linear computational complexity. This makes it a suitable choice for efficient global context modeling. These observations have led us to propose a hybrid convolutional network–Mamba architecture for RSISR. The proposed model includes a ConvNet branch for capturing local high-frequency details and a Spectral Mamba module for efficiently modeling long-range dependencies across spatial and spectral dimensions. We present ConvMambaB as the fundamental computational unit, integrating convolutional feature extraction, channel-mixing MLP enhancement, and Mamba-based state-space modeling within a progressive residual refinement framework. The ConvMambaGroup backbone consists of several ConvMambaB blocks stacked on top of each other. This enables effective feature integration at different levels. Extensive tests on three benchmark remote sensing datasets show that the proposed method achieves reconstruction performance that is either competitive with or better than existing lightweight models, while maintaining a low number of parameters and computational complexity.