Differentiable Heatmap Regression for Source Localization Based on RSS and AoA Measurements

Signal source localization in urban environments is challenged by severe multipath effects caused by dense building obstructions, resulting in significant signal attenuation and distortion. Existing deep learning-based methods often rely on reconstructing the signal distribution from passively collected data to infer the source location. However, these approaches typically require auxiliary equipment and high-resolution datasets, leading to high deployment and computation costs that limit real-world applicability. To address these limitations, we propose LocHeatmap, a differentiable heatmap regression framework for non-cooperative source localization based on RSS and AoA measurements. Unmanned aerial vehicles (UAVs) collect sparse and noisy received signal strength (RSS) and angle of arrival (AoA) measurements, which are fused with environmental topography and reconstructed into complete radio maps using a dual-path autoencoder. These reconstructed maps are then processed by a U-Net model to generate a physically grounded heatmap, supervised by labels derived from a segmented propagation model that captures signal decay characteristics. Finally, Soft-argmax enables differentiable coordinate extraction and backpropagation for end-to-end optimization. Experimental results under varying sampling rates and noise levels show that LocHeatmap significantly improves localization performance over existing methods, achieving 74% RMSE reduction on average in complex urban scenarios.