Radio Map Prediction from Noisy Environment Information and Sparse Observations

Many works have investigated radio map and path loss prediction in wireless networks using deep learning, in particular using convolutional neural networks. However, most assume perfect environment information, which is unrealistic in practice due to sensor limitations, mapping errors, and temporal changes. We demonstrate that convolutional neural networks trained with task-specific perturbations of geometry, materials, and Tx positions can implicitly compensate for prediction errors caused by inaccurate environment inputs. When tested with noisy inputs on synthetic indoor scenarios, models trained with perturbed environment data reduce error by up to 25% compared to models trained on clean data. We verify our approach on real-world measurements, achieving 2.1 dB RMSE with noisy input data and 1.3 dB with complete information, compared to 2.3-3.1 dB for classical methods such as ray-tracing and radial basis function interpolation. Additionally, we compare different ways of encoding environment information at varying levels of detail and we find that, in the considered single-room indoor scenarios, binary occupancy encoding performs at least as well as detailed material property information, simplifying practical deployment.

💡 Research Summary

This paper addresses the practical problem of indoor radio‑map (path‑loss) prediction when the available environment information is incomplete, outdated, or noisy, and only a limited number of ground‑truth signal‑strength measurements are at hand. The authors formulate the task as “uncertain environment prior + sparse reliable observations” and propose two key contributions.

First, they introduce a systematic data‑augmentation strategy called Simulated Noise as Data Augmentation (SNDA). For each synthetic base environment they generate ten perturbed copies by randomly shifting object and transmitter positions (average offset 0.5 m) and by varying material electromagnetic parameters (permittivity, conductivity, thickness) within realistic ranges. Ray‑tracing simulations (Wireless InSite, 5.92 GHz, up to three reflections, transmissions and one diffraction) produce the ground‑truth radio maps for all perturbed versions. By training convolutional neural networks (CNNs) on this enlarged set of noisy input–label pairs, the model learns to be robust to the kinds of structured errors that occur in real indoor settings.

Second, they compare three levels of environment encoding: (1) binary occupancy (presence/absence of any object at several height slices), (2) semantic class encoding (material categories such as wood, metal, glass, concrete, free space), and (3) detailed material‑property encoding (continuous values of permittivity, conductivity, thickness). All encodings are rasterized into multi‑channel tensors; transmitter location is added via a one‑hot map and a Euclidean distance map, while sparse RSSI observations are inserted into a separate observation map.

The authors generate a comprehensive synthetic dataset: 3 601 distinct room layouts with realistic furniture arrangements, five transmitter positions per layout, yielding 18 005 ray‑traced radio maps. Each map is represented on a 32 × 32 grid (30 cm spacing) and normalized to