Uncertainty-Aware Informative Path Planning for UAV: An Interpretable Bayesian-DRL Approach
Abstract
Efficient unmanned aerial vehicle (UAV)-based crop monitoring requires disease-relevant observations within a limited flight budget. This paper presents a task-oriented informative path planning (IPP) framework for motion-efficient disease source localization in partially observable agricultural environments. A voxelwise Bayesian belief represents the latent disease field, while a twin delayed deep deterministic policy gradient (TD3) policy generates continuous sensing actions from the UAV pose, local crop descriptors, row-alignment information, and spectral disease cues. The reward combines posterior covariance-trace reduction, local disease evidence, and translational motion cost, and the accumulated belief supports source estimation. The method was compared with coverage path planning, three adaptive IPP methods, and privileged oracle References under a common protocol. Across three fixed-hotspot layouts and one random-hotspot condition, each evaluated over 20 trials, it achieved 90-100% localization success and mean localization errors of 3.80-4.73 m. Motion cost was reduced by 41.9-65.5% relative to coverage planning and by 20.9-65.8% relative to the adaptive baseline with the highest average localization success. Sensitivity and ablation studies examined voxel resolution, prior initialization, and framework components. Field-derived LiDAR and spectral maps further assessed trajectory-level applicability to measured agricultural data. Note to Practitioners—This work addresses UAV-based crop monitoring tasks in which disease-related evidence must be acquired under a limited flight budget. Conventional coverage planning systematically surveys the field but may require unnecessary traversal when the operational objective is to identify a disease-relevant region rather than to reconstruct the entire field. The proposed method combines a Bayesian spatial belief with a continuous TD3 policy that selects sensing motions according to domain-aware agricultural features and translational motion cost. In simulation, the method maintained high disease source localization success while providing a favorable overall tradeoff between localization reliability and traversal cost relative to CPP and the evaluated adaptive IPP methods. The field-data evaluation applies the planners to maps constructed from actual LiDAR and RGN measurements and is intended to assess trajectory-level applicability rather than biologically verified disease diagnosis. Physical UAV deployment will additionally require sensor calibration, onboard computation, localization under canopy-induced disturbances, and independently confirmed disease ground truth.