Radio Map Prediction from Noisy Environment Information and Sparse Observations

TL;DR

This work tackles indoor radio map prediction under realistic imperfect environment information by training CNNs with structured perturbations of geometry, materials, and transmitter positions (SNDA). It compares different environment encodings and demonstrates that SNDA-trained models achieve superior robustness to input errors, even with sparse observations, with binary occupancy encoding often performing best. On real measurements, the SNDA-trained CNN attains around 2.1 dB RMSE, improving to 1.3 dB when a blocking object is included in the input, and substantially outperforming ray-tracing and classical interpolation baselines. The findings suggest a practical path for fast, robust radio-map estimation in indoor settings, with potential extensions to multi-room and millimeter-wave scenarios and other frequencies.

Abstract

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.

Figures (6)

• Figure 1: Example of a room from our dataset. Plot of 2D geometry (top) and 3D model after import into ray-tracer (bottom).
• Figure 2: CNN trained with a random number of observations from $[0\%, 20\%]$ and different inputs (Material Properties, Classes, Binary encoding, No Environment) compared with baseline methods on the test set generated with RT.
• Figure 3: Top row: environment slices showing material classes at 0.3/0.6/0.9/1.2m height for the environment in Fig. \ref{['fig:project148']}/ground truth radio map. Second row: slices for noisy version of the environment. Third row: 1% Input observations and predictions for the model trained without SNDA, receiving clean environment information/ the model trained without SNDA, receiving noisy environment information/ the model trained with SNDA, receiving clean environment information/ the model trained with SNDA, receiving noisy environment information. RMSE in dB below. Fourth and Fifth row: Same for 5% and 10% observation given.
• Figure 4: Test results for Perturbations of only Tx locations/only object locations (Env)/Both/None, for varying degrees of noise and model trained with/without SNDA.
• Figure 5: Measurement Scenario - photo and representation in ray-tracer with annotations: Tx in green, blockage in blue, measurement areas in black with red or yellow frame depending on whether they lie in LoS of the Tx or not, respectively.