SCMamba: A Spatial–Channel Multiscale State-Space Network for High-Resolution Image Scene Classification

High-resolution remote sensing images play a crucial role in urban planning and environmental monitoring; however, their multi-scale and diverse backgrounds pose significant challenges for classification tasks. To address this issue, we propose an enhanced framework that integrates multi-scale features and spatial–channel attention on top of a state space model (SSM) backbone. Specifically, we draw upon the concept of a bidirectional feature pyramid (BiFPN) for multi-scale feature fusion, incorporating skip connections and residual improvements to strengthen information flow across different feature levels. Meanwhile, leveraging large-kernel convolutions and channel attention, we design a spatial–channel fusion module tailored to the demands of remote sensing images, thereby enabling deeply separable reweighting for irregular ground objects and multi-channel spectral responses. Experimental results show that on the GID dataset, our method achieves a Precision of 94.61%, surpassing CNN-based networks (VGG16, ResNet-18), Transformer-based architectures (ViT, Swin-T), and other Mamba-based models (Vim-tiny, VMamba-tiny, Samba) by approximately 13%, 6%, and 8%, respectively. Meanwhile, on the EuroSat and AID datasets, our method achieves Precision values of 98.19% and 94.65%, respectively, thereby providing a feasible solution for the efficient intelligent analysis of urban remote sensing imagery. The code will be available at https://github.com/qxin-g/SCMamba.