轻量级的油菜籽品种高光谱特征识别方法OA

Accurate identification of rapeseed varieties is of paramount importance for ensuring seed purity and grain-oil security.However,existing hyperspectral detection models suffer from parameter redundancy and massive computational overhead,severely restricting their deployment on resource-constrained platforms,such as portable devices and unmanned aerial vehicles(UAVs).To address this issue,this paper proposes a lightweight hyperspectral recognition model termed SSL-SwinT(Spectral-Adaptive-Selector Spectral-Spatial-Hybrid-Encoder Light-Swin-Transformer).Specifically,a lightweight Swin-Transformer(SwinT)backbone is first constructed to achieve structural compression through configuration optimization and input dimensionality reduction.Subsequently,a Spectral Adaptive Selector(SAS)is designed.Based on global average pooling(GAP)and a multi-layer perceptron(MLP)attention mechanism,the SAS dynamically filters effective spectral bands,substantially eliminating the interference of redundant information.Finally,a Spectral-Spatial Hybrid Encoder(SSHE)is constructed,which employs depth wise separable convolutions to realize the decoupling and synergistic extraction of spatial-spectral features.Experimental validation on a hyperspectral dataset comprising 11 rapeseed varieties(with 128 spectral bands)demonstrates that SSL-SwinT achieves an 82％compression rate in both parameter count(reduced from 27.72M to 4.99M)and model volume(reduced from 105.81MB to 17.44MB),while maintaining an outstanding F1 score of 99.57％.Ablation studies further verify the synergistic enhancement effect among the proposed modules.This research provides a reliable solution for the lightweight edge deployment of agricultural hyperspectral technologies.