Accurate detection of rice blast using UAV hyperspectral red-edge bands and deep learning method based on...

Rice blast is a fungal disease that threatens rice yield and quality, and timely detection methods are essential to control its spread. Unmanned aerial vehicle (UAV)-based optical Remote Sensing offers a promising approach for monitoring disease in field environments. However, the complex symptoms of rice blast make it challenging to capture the sensitivity of hyperspectral features to disease progression. This study first proposed a collaborative gray wolf optimizer feature selection method with mutual information ranking (MI-CGWO) for characterizing the spectral response of rice blast. Subsequently, the impact of different combinations of canopy red-edge reflectance and fraction vegetation coverage (FVC) on rice blast was evaluated. On this basis, a channel fusion dense cross-attention transformer (CFXFormer) disease monitoring model was developed. The model incorporated a channel interaction fusion module (CIFM) to mitigate modal discrepancies. Additionally, a cross-attention mechanism using Gaussian and self-attention was introduced to capture relationships between fine-grained and coarse-grained data. The results demonstrated that MI-CGWO achieved better classification accuracy with fewer features in a shorter time. Furthermore, vegetation indices (VIs) covering the red-edge and FVC significantly improved model performance compared to models based on conventional VIs, with average OA and Kappa reaching 95.07 % and 93.80 %. Meanwhile, the ideal red edge band may be located at about 701-711 nm and 739 nm. Compared to other classification methods, CFXFormer achieved the best classification performance, with a maximum OA and Kappa of 96.59 % and 95.71 %. Overall, this study offers scientific support for precision agriculture and provides a reference for detecting other crop diseases.