BMFormer: azimuth-conditioned radio map prediction for ground-projected low-altitude task coverage planning
Abstract
Radio maps describe the power of received signals in space and support task-driven wireless resource control. For ground-projected low-altitude task areas such as temporary venues and corridor footprints, the coverage also depends on the azimuth of the directional antenna. This paper proposes BMFormer, an azimuth-conditioned predictor that learns a signed directional residual relative to an omnidirectional prior with a FiLM-conditioned U-Net and an optional Transformer bottleneck. In 30 ray-tracing deployment scenarios from one building layout, BMFormer achieves 3.15 dB MAE on held-out directional maps; direct prediction of RSRP, a 14-channel residual U-Net, and removing FiLM increase MAE to 4.72, 4.56, and 4.27 dB, respectively. Sectors are selected from predicted maps but scored on ray-tracing ground truth. At −95 dBm, the selection of the predicted-map sector raises the mean crowd-weighted GT coverage from 23.1% to 47.6%; a 40-venue placement check gives 51.9% predicted-selected GT coverage versus 32.1% omnidirectional and 56.3% sampled-sector oracle coverage. The evidence remains limited to sampled sectors in a single-layout 2D benchmark, with height-dependent air-ground channels, unseen layouts, and live measurements left for future work.