Deep reinforcement learning-based adaptive fusion method for low-altitude multi-source heterogeneous sensing data

Low-altitude intelligent systems are proliferating faster than the fusion algorithms meant to support them, and the gap widens precisely where sensing conditions refuse to stay still. This paper proposes a deep reinforcement learning framework for adaptive fusion of multi-source heterogeneous low-altitude sensing data. The architecture is hierarchical, comprising data perception, feature extraction, fusion decision and output application layers, so that each layer can be optimized on its own and new sensors integrated without redesigning the pipeline. A unified representation maps heterogeneous sensor outputs into a shared feature space through modality-specific encoders and a cross-modal alignment mechanism. Our central contribution recasts fusion-weight determination as a sequential decision problem and solves it with a Dueling Deep Q-Network whose action space is finely discretized: every weight component is quantized onto 21 uniform levels rather than emitted as a continuous value. A composite reward balancing accuracy, timeliness and smoothness steers policy optimization. Experiments on a purpose-built multi-modal dataset yield 1.52 m fusion RMSE, 94.7% detection accuracy and 91.2% tracking precision at 12.3 ms average latency, comfortably within real-time budgets. Stress tests confirm graceful rather than catastrophic degradation under sensor failure and noise injection, with recovery achieved through autonomous weight redistribution. Taken together, these results advance adaptive fusion methodology for low-altitude autonomous platforms operating in complex, changeable environments.