Spatial-strip Mamba for multisource remote sensing image fusion

Multisource remote sensing image classification aims to fuse complementary information from heterogeneous sensors such as hyperspectral and LiDAR to enhance the fine-grainedness and robustness of land cover recognition. Convolutional neural networks are constrained by local receptive fields, making it difficult to capture large-scale spatial context. Although Vision Transformers can capture long-range dependencies, their quadratic computational complexity leads to bottlenecks in memory usage and inference efficiency when processing high-resolution remote sensing imagery. The recently proposed Mamba architecture, leveraging state-space models with low computational complexity, offers a new paradigm for efficient visual understanding. However, its direct application to remote sensing fusion faces two challenges: first, flattening two-dimensional images into one-dimensional sequences disrupts the local adjacency relationships among pixels, thereby weakening the discriminative capability for textures and edges; second, the scales of land cover objects in remote sensing scenes vary dramatically, while existing Mamba variants lack a flexible multiscale modeling mechanism. To address the above issues, this paper proposes an end-to-end classification network named spatial-strip Mamba. The network adopts a U-shaped encoder-decoder architecture, embedding the spatial locality-aware Mamba (SLAM) module and the decomposed strip multiscale attention (DSMA) module at multiple hierarchical levels. SLAM introduces a channel-wise depthwise convolutional residual path inside the Mamba block, re-injecting two-dimensional local spatial inductive biases into the sequential modeling process with nearly low parameter increase, thereby recovering fine-structured information. DSMA decomposes a large square convolutional kernel into a combination of asymmetric one-dimensional strip convolutions, constructing a multiscale receptive field pyramid ranging from fine-grained to large-scale contexts with extremely low parameter cost, and adaptively recalibrates the weights of each scale via channel attention. Experiments conducted on three public datasets Muufl, Trento, and Augsburg demonstrate that spatial-strip Mamba achieves higher performance than current state-of-the-art methods in terms of classification accuracy. Particularly notable improvements are observed in challenging categories such as road connectivity, vegetation discrimination, and railway extraction, confirming the effectiveness and superiority of the SLAM and DSMA designs.