A Neuromorphic Dual-Stage Swin Transformer for Wetland Classification on Edge Platforms

Abstract

Accurate coastal wetland classification using hyperspectral remote sensing is crucial for conservation, restoration, and sustainable management of these ecologically vital regions. With the proliferation of edge computing, there is a growing demand for real-time hyperspectral analysis on sensing platforms such as satellites and unmanned aerial vehicles. However, the spectral complexity of wetland scenes, the high computational cost of modern CNNs and Transformers, and frequent memory accesses impose substantial overhead on resource-constrained edge platforms. To address these challenges, we propose the Dual-Stage Spiking Swin Transformer (D2S-SwinT), a neuromorphic architecture that integrates the feature representation capability of Swin Transformers with the event-driven efficiency of brain-inspired computation. The model employs a compact dual-stage design to reduce architectural redundancy and leverages the Expectation Compensation and Multi-Threshold (ECMT) mechanism to transform dense arithmetic operations into sparse, event-driven processes. Extensive experiments on four wetland datasets show that D2S-SwinT reaches 99.62% OA on Huanghekou at T = 4. Additional experiments on IP and SA provide evidence that the model generalizes beyond wetland-specific scenes. At T = 1, its normalized operation-level energy is reduced to 8% of the ANN-proxy reference. At T = 3, a memristor-based architecture-level mapping yields a projected OA of 96.69% and an energy per sample equal to 0.66% of the software reference. These results indicate the potential of D2S-SwinT for accurate and energy-efficient hyperspectral wetland monitoring on resource-constrained edge platforms.