End-to-End UAV-Enabled Adaptive 3-D Radio Mapping via Joint Optimization of Sparse Sampling and Reconstruction

Accurate radio environment map (REM) construction proves critical for efficient wireless spectrum management. Although conventional 2-D REMs have demonstrated effectiveness in wireless network optimization, they inherently overlook vertical signal strength variations, which are vital for uncrewed aerial vehicle (UAV) operations, particularly in urban landscapes with skyscrapers or diverse terrain features. Current estimation approaches, including ground-based crowdsourcing, random sampling, and predetermined trajectory measurements, show limited capability in generating high-fidelity 3-D REMs. This study proposes a joint optimization framework for UAV-enabled adaptive 3-D radio mapping, integrating 3-D REM construction with adaptive aerial sampling. At the heart of the construction module, a dual-branch encoder-decoder architecture fuses multiscale features from sparse aerial measurements with building structural data, explicitly modeling obstruction effects through offline pretraining and online refinement to enhance generalization. For adaptive sampling, a diffusion-based trajectory planner dynamically optimizes UAV measurement paths by integrating environmental priors (e.g., building layouts), effectively overcoming the sparse-reward limitations inherent in reinforcement learning (RL) methods. Experimental validation demonstrates significant performance improvements across all evaluation metrics. Compared to 2-D per-layer estimation methods, our 3-D estimator achieves 49% superior structural similarity (SSIM) in construction accuracy, while the feature fusion module yields a 37% reduction in mean squared error (mse). The diffusion-based planner outperforms RL approaches by achieving 45% lower mse and 18% higher SSIM in resultant map quality after 5000 step iterations.