Multi-source field sensor data fusion based on cross modal attention mechanism and reinforcement learning driven pesticide application

The intensification of global agriculture demands precise and sustainable pest management strategies, as indiscriminate pesticide application continues to cause environmental degradation and reduce crop resilience. Existing approaches often rely on unimodal sensing or static rule-based spraying, which fail to capture the heterogeneous and dynamic nature of crop-pest-environment interactions. To address this limitation, we propose a multi-source field sensor data fusion framework that combines a cross-modal attention mechanism with a reinforcement learning-driven model for optimizing pesticide applications. The method integrates Unmanned Aerial Vehicle (UAV) hyperspectral imagery, soil and weather sensors, and pest monitoring signals through adaptive attention, encodes temporal dynamics with recurrent structures, and optimizes spraying actions via a PPO-based policy network. Experiments across rice, maize, and soybean datasets demonstrate superior performance, achieving the lowest RMSE (0.162), highest spray precision (88.3%), and notable pesticide reduction (18.3%) compared with state-of-the-art baselines. These findings highlight the potential of cross-modal AI and adaptive control to advance sustainable crop protection, providing a scalable paradigm for intelligent agriculture.