Early detection of rice blast using UAV hyperspectral imagery and multi-scale integrator selection attention transformer network (MS-STNet)

Abstract

Rice blast is one of the most destructive diseases of rice leaves, seriously affecting rice production and quality. An accurate and rapid large-scale disease detection method is essential for rice production management. This study employed unmanned aerial vehicle (UAV) hyperspectral remote sensing technology for continuous observation of rice blast in the field. Advanced deep-learning techniques were utilized and combined with UAV data to detect rice blast. Firstly, the sensitivity and importance of canopy reflectance and texture features in disease monitoring were assessed. Considering the limitations of single texture features, the rice blast texture indices (RBTIs) were constructed by multiple texture features. Secondly, based on characteristic wavelengths, RBTIs, and their combinations, an effective rice blast detection framework based on the transformer network, multi-scale integrator selection attention transformer network (MS-STNet) model, was proposed. By incorporating multi-scale integrator and adopting a multi-scale and multi-pooling strategy that considered the interactions between different layers, the ability of the model to capture fine-grained information was enhanced. The top-k selection mechanism was introduced to generate corresponding attention masks, preserving the most contributive feature combinations while maintaining the global structural information of the input. The results demonstrated that the MS-STNet model could adequately learn significant features at different scales, demonstrating excellent accuracy and strong spatial adaptability in both field experiments. Compared with single texture features, the model using RBTIs as inputs demonstrated superior classification performance, with a maximum increase in overall accuracy (OA) of 4.27%. Furthermore, the model constructed by combining spectral features and RBTIs outperformed models built using only spectral features or RBTIs, with a maximum OA of 96.98% and Kappa of 96.22%. Overall, the feature-based combination method can improve the early phases of rice blast classification accuracy. The study results can provide valuable reference for accurately monitoring rice blast using UAV hyperspectral imagery.