MambaOutRS: A hybrid CNN-Fourier architecture for remote sensing image classification

While State Space Models (SSMs) scale linearly with sequence length, adapting them to 2D vision often increases architectural complexity without commensurate benefit for non-autoregressive tasks. We introduce MambaOutRS, a hybrid convolutional architecture that combines Gated CNN blocks for local spatial modeling with a Fourier Filter Gate (FFG) for efficient global context extraction in the frequency domain. Evaluated on UC Merced, AID, EuroSAT, and NWPU-RESISC45, MambaOutRS achieves state-of-the-art or competitive performance while maintaining a favorable parameter budget. Ablation studies confirm that the FFG is the primary driver of performance improvement, enabling the model to capture global spatial patterns that are inaccessible to purely local convolutional operations. This work demonstrates that combining gated convolutions with frequency-domain filtering provides a highly efficient and practical alternative to recurrent SSM-based designs for remote sensing image classification.