【摘要】 为了解决现有模型在小样本高光谱小麦数据集上存在的特征捕捉不充分、全局依赖性建模薄弱等性能不足的问题，提出了一种基于光谱-空间协同特征融合网络(Spectral-spatial cooperative feature fusion network, SSCFF-Net)的深度学习模型。该模型采用多阶段架构，充分利用局部和全局特征的互补优势：一方面，通过结合二维卷积神经网络(Two-dimensional convolutional neural network, 2D-CNN)、三维卷积神经网络(Three-dimensional convolutional neural network, 3D-CNN)与多种注意力机制，高效捕捉高光谱数据的局部光谱-空间细节；另一方面，借助Transformer注意力机制和1D-CNN结构，有效解决了其全局依赖性问题。该模型通过深度挖掘小麦高光谱图像的光谱维与空间维信息，构建光谱-空间协同特征融合机制，针对性解决现有模型的性能缺陷。此外，模型中采用的数据增强和残差设计已被证明能显著降低对标注数据的依赖，同时，提升模型对小样本数据集的适应性，特征压缩与模块化设计的结合则可进一步提升计算效率。实验结果表明，SSCFF-Net模型具备准确区分不同小麦品种的能力，在HyperLeaf 2024数据集上分类准确率达到99.4%,有效解决了深度学习在小麦小样本高光谱数据集时准确率较低的问题。更多还原

【Abstract】 To address the performance shortcomings of existing models on small-sample hyperspectral wheat datasets, such as inadequate feature capture and weak modelling of global dependencies, a deep learning model is proposed, based on the spectral-spatial cooperative feature fusion network(SSCFF-Net). This model adopts a multi-stage architecture that fully leverages the complementary strengths of local and global features: on one hand, it efficiently captures local spectral-spatial details in hyperspectral data by integrating two-dimensional convolutional neural network(2D-CNN), three-dimensional convolutional neural network(3D-CNN), and multiple attention mechanisms. On the other hand, it effectively addresses global dependency issues through the Transformer attention mechanism and 1D-CNN structure. By deeply mining spectral and spatial information from wheat hyperspectral imagery, the model constructs a spectral-spatial cooperative feature fusion mechanism that specifically addresses performance limitations of existing models. Furthermore, the incorporated data augmentation and residual design have been demonstrated to significantly reduce reliance on labelled data while enhancing the model’s adaptability to small-sample datasets; the combination of feature compression and modular design further improves computational efficiency. Experimental results demonstrate that the SSCFF-Net model possesses the capability to accurately distinguish between different wheat varieties, achieving a classification accuracy of 99.4% on the HyperLeaf 2024 dataset. This effectively resolves the issue of low accuracy encountered when applying deep learning to small-sample hyperspectral datasets of wheat.更多还原