Explainable VoI-aware energy-efficient cooperative multi-UAV data collection

Timely retrieval of high-priority sensor data is essential for decision-making in wireless sensor networks (WSNs), yet most multi-unmanned aerial vehicle (UAV) trajectory planners treat all readings as equally valuable, wasting flight energy on low-priority data while time-critical information decays. Energy-efficient UAV-assisted data collection, therefore, depends on the ability to distinguish sensor importance in real time, while understanding the resulting policy decisions calls for explainable mechanisms that earn operator trust. In our work, we present Explainable Value of Information (VoI) aware Cooperative Multi-Agent Deep Q-Network (VoI-CMADQN) as a cooperative double deep Q-network that navigates a UAV fleet, maximizing VoI retrieval at minimum energy cost while maintaining fleet-wide spatial coverage. Each UAV receives its own reward, computed from the VoI it personally collects, the sensors it covers, and the energy it spends; a spatial repulsion term discourages clustering, VoI-urgency observations navigate UAVs toward high-priority sensors, and potential-based shaping speeds up early learning. We then apply Permutation Feature Importance (PFI) to provide explainability and identify which observation components influence the policy at each scale. We evaluate the performance of our proposed approach against baseline and state-of-the-art approaches under different network scenarios, demonstrating. Ablation study by stripping out one mechanism at a time pins per-UAV reward decomposition as the single biggest contributor. PFI analysis reveals that the policy progressively shifts its decision priorities from spatial coordination at smaller scales to resource budgeting at larger ones, adapting its internal strategy to match the dominant operational bottleneck.