Pruning State Space Models for Lightweight Remote Sensing Image Classification

With the rapid advancement of satellite remote sensing technology, real-time on-board processing of optical imagery has become a critical demand for applications such as land resource management and disaster emergency response. However, spaceborne platforms face extreme resource constraints: existing satellites lack the necessary computational capacity, and the quadratic computational complexity of traditional Transformer architectures make them difficult to adapt. Although State Space Models (SSMs) offers significant improvements in efficiency with their linear computational complexity, a major gap remains for on-board deployment, and current research on lightweighting SSM architectures is scarce. This study proposes a lightweight remote sensing image classification algorithm via pruning of State Space Models, constructing a dual-track pruning methodology: Structured pruning progressively removes blocks based on a perturbation-based evaluation, which measures the importance of individual Mamba Blocks within the baseline model, thereby uncovering the critical role of deep modules. Unstructured pruning uses magnitude pruning and a momentum reset mechanism to address the issue of gradient coupling interference. This process demonstrates that sparse, high-performance sub-networks can be uncovered within SSMs. Experiments on three public remote sensing image datasets show that the proposed method reduces the number of model parameters by 50% to 70%, with negligible loss in classification accuracy.