Synergistic spectral-spatial fusion in hyperspectral Imaging: Dual attention-based rice seed varieties identification

Abstract

Accurate classification of rice seeds is important for improving crop yields, optimizing the breeding process and achieving sustainable agricultural production. Hyperspectral imaging technology has demonstrated a wide range of potential applications in seed classification tasks due to its ability to capture both spectral and spatial information. However, most of the existing hyperspectral classification methods focus only on spectral information and ignore the role of spatial features, resulting in limited classification performance. To address this problem, this study proposes a Principal Component Analysis-based Vision Transformer Non-Local Dual Attention Network (PCA-VNDANet). PCA is used for dimensionality reduction to eliminate redundant information, while the model leverages a Vision Transformer (ViT) module to extract spatial features, overcoming the limitations of traditional convolutional neural networks in modeling global dependencies. Additionally, a non-local spectral attention module is introduced to construct a spectral-spatial collaborative attention mechanism, further enhancing classification performance. In addition, the decision-making process of the model is feature visualized using Gradient-weighted Class Activation Mapping (Grad-CAM++) to enhance the interpretability of the model. The experimental results show that PCA-VNDANet achieves a classification accuracy of 94.87 %, which is at least 0.6 % better than the existing comparison methods. Meanwhile, the parameters, model size and FLOPs are 2,969,016, 11.34 MB and 91.89 M respectively. This study provides an efficient and accurate technical means for seed classification based on hyperspectral imaging and shows a broad application prospect in complex agricultural tasks.