SAN-GAT-RL: A Multimodal Drone and AI-Driven Framework for Enhanced Wheat Pest and Disease Detection in Dynamic Agricultural Environments

Aiming at the problems of low efficiency and insufficient spatial coverage of traditional methods in the detection and prediction of wheat pests and diseases, this study proposes an intelligent detection and prediction system that combines multimodal data fusion of unmanned aerial vehicles (UAVs) and artificial intelligence. This system preprocesses multimodal data through dynamic geometric correction and adaptive sparse filtering, designs hierarchical fusion strategies (pixel-level weighted fusion, feature level graph neural network, decision level Markov random field), and innovatively integrates sparse adaptive network (SAN), graph attention network (GAT), and reinforcement learning (RL) modules. SAN dynamically screens key features through the sparse attention mechanism, GAT models the on-site spatial dependency relationship, and RL optimizes the dynamic strategy of unmanned aerial vehicles through methods such as sensor switching and path planning. The experiment is based on the AgriVision dataset and the PlantVillage dataset to verify the robustness of the system in complex scenarios. The results show that the recall rates of the SAN-GAT-RL framework for the two adopted datasets have reached 92.3% and 89.7% respectively, and the F1 scores have both exceeded 90 %. The reduction of the missed detection rate under complex conditions is achieved through dynamic strategies. And the verification of the effectiveness comparison of the methods was carried out through ablation experiments. This research provides high-precision and low-energy consumption solutions for precision agriculture and sustainable development, and has significant practical application value.