A Variable-Receptive-Field Model for Joint Radio Map Construction and Radiation Source Localization in Urban Environments

With the growing demand for smart-city electromagnetic monitoring, radiation source localization based on received signal strength indication (RSSI) remains challenging in complex urban environments because RSSI measurements are affected by non-line-of-sight propagation, multipath fading, and shadow fading. Motivated by this challenge, we design a lightweight autoencoder named LocAutoencoder to jointly perform radio map construction and radiation source localization in urban environments. In the encoder, a variable-receptive-field module uses multiple dilation rates to extract multi-scale wireless propagation features directly from both RSSI measurements and urban map information. The sequential dilated convolutions expand the receptive field without substantially increasing the parameter count. In the decoder, a dual-task decoder is designed with two output branches: one branch reconstructs the radio map, whereas the other estimates the source coordinate through probability heatmap generation and centroid calculation. The lightweight autoencoder model computes the probability-weighted centroid within the network to directly output the estimated coordinates, enabling accurate estimation of the radiation source location without requiring a separate localization subnetwork. Finally, simulations and comparative experiments demonstrate the effectiveness of the proposed network in terms of localization accuracy, robustness, and computational efficiency.