Scene Structure Based Neural Radio-Frequency Radiance Fields for Channel Knowledge Map Construction

Channel knowledge map (CKM), as an emerging technology for environment-aware communication, corresponds transmitter/receiver locations to wireless channel states and can be used for the acquisition of channel state information (CSI) in wireless communications. However, accurately and efficiently constructing this location-to-channel mapping remains a key challenge. Inspired by the success of neural radio-frequency radiance fields (NeRF2) in predicting the spatial spectrum, we propose NeRF2 with scene structure (NeRF2-SS), which uses the scene structure as strong prior information in the neural network to predict the spatial spectrum. Specifically, NeRF2-SS introduces two new modules: 1) a scene structure based sampling method concentrating samples near the surfaces of the scatterers, and 2) a lightweight network architecture that distinctly processes near-surface and free-space samples. Simulation results show that NeRF2-SS achieves high-precision spatial spectrum prediction with an 85-90% reduction in inference latency compared to NeRF2.