PaperlayerView exact revision on arXiv

Source-linked AI summary

RadioUNet: Fast Radio Map Estimation with Convolutional Neural Networks

Ron Levie, Çağkan Yapar, Gitta Kutyniok, Giuseppe Caire

arXiv:1911.09002v3eess.SPcs.ITcs.LGstat.ML

TL;DR

Wireless applications need accurate pathloss maps despite complex urban propagation that simple distance-based models do not capture. The paper proposes RadioUNet, a UNet-based deep-learning estimator trained on simulated radio maps and adapted toward real-life settings, and reports higher accuracy with much faster computation than prior methods.

  • Problem

    Accurate pathloss knowledge is needed across transmitter–receiver locations, while radial distance-based models can be misleading in environments with buildings, street canyons, and obstacles.

  • Method

    RadioUNet uses UNet-based deep learning to estimate radio maps from urban geometry and transmitter information, optionally incorporating pathloss measurements and car locations.

  • Results

    RadioUNet significantly outperforms previously proposed methods in accuracy and runtime, with roughly three orders of magnitude faster execution than RBF interpolation.

  • Takeaways & Limitations

    The method provides an efficient radio-map estimator for settings requiring pathloss knowledge, including dynamic environments where measurements can be supplied in real time.

  • Takeaways & Limitations

    The demonstrated fingerprint-localization method works well when estimated maps are accurate and the number of reported signal strengths is large enough but not too large.

Abstract

from arXiv · show

In this paper we propose a highly efficient and very accurate deep learning method for estimating the propagation pathloss from a point $x$ (transmitter location) to any point $y$ on a planar domain. For applications such as user-cell site association and device-to-device link scheduling, an accurate knowledge of the pathloss function for all pairs of transmitter-receiver locations is very important. Commonly used statistical models approximate the pathloss as a decaying function of the distance between transmitter and receiver. However, in realistic propagation environments characterized by the presence of buildings, street canyons, and objects at different heights, such radial-symmetric functions yield very misleading results. In this paper we show that properly designed and trained deep neural networks are able to learn how to estimate the pathloss function, given an urban environment, in a very accurate and computationally efficient manner. Our proposed method, termed RadioUNet, learns from a physical simulation dataset, and generates pathloss estimations that are very close to the simulations, but are much faster to compute for real-time applications. Moreover, we propose methods for transferring what was learned from simulations to real-life. Numerical results show that our method significantly outperforms previously proposed methods.

1 Introduction

Radio maps describe signal-strength loss between transmitter–receiver locations and are important for wireless applications. Existing estimation approaches range from measurement interpolation and model-based fitting to physics-based prediction.

  • Radio maps: Pathloss measures large-scale signal attenuation between transmitter and receiver locations.Buildings, reflections, diffraction, street-canyon waveguides, and line-of-sight obstacles can affect it.
  • Paper objective: The paper develops RadioUNet, a deep-learning method for estimating radio maps.The method is introduced as an approach for estimating the pathloss function over urban environments.
  • Applications: Accurate pathloss knowledge supports device-to-device link scheduling, user-cell site association, localization, security, power control, and other wireless tasks.In link scheduling, pathloss information helps manage mutual interference among simultaneously active links.
  • Prior approaches: Radio-map estimation methods are grouped into data-driven interpolation, model-based data fitting, and model-based prediction.Examples include Kriging, tomography, ray tracing, dominant-path models, and empirical models.

1.3 Radio Map Prediction Using Deep Learning

The paper frames RadioUNet as an image-to-image estimator that maps urban geometry and transmitter information to radio maps. It also develops datasets and transfer methods intended to support refined or real-life deployment settings.

  • Deep-learning prediction: RadioUNet produces a radio map from a city map and transmitter source, unlike earlier networks tied to one fixed map.Earlier approaches fit a 4D pathloss function for a specific map and require retraining for different city maps.
  • Training data: RadioMapSeer contains 56,000 simulated radio maps with city maps, transmitter locations, and car information.The dataset includes simulations generated with different coarse-simulation variants.
  • Training data: IRT4 provides 1,400 high-accuracy simulations, sparsely sampled to imitate real-life measurement scenarios.The paper uses IRT4 as a surrogate for real-life measured radio maps.
  • Transferability: Transfer methods refine coarse-simulation knowledge using smaller sets of high-accuracy sparse samples.The stated goal is transferring learned representations from simulated data toward real-life deployments.
  • Applications: RadioUNet is demonstrated in toy applications where accurate pathloss knowledge is useful, including coverage classification.The passage introduces two applications but does not provide their complete results.

2 Background and Preliminaries

The paper expresses pathloss in a logarithmic scale motivated by wireless-interference theory, then reviews CNN, UNet, supervised-learning, optimization, and transfer-learning concepts used by RadioUNet.

  • Wireless communication: The relevant pathloss representation is logarithmic, with negative exponents treated through their positive part and large negative values truncated.This follows the paper’s discussion of GDoF exponents and the TIN regime.
  • Wireless communication: Pathloss in dB is defined as PL = (PRx)dB − (PTx)dB.The paper states that truncation and rescaling make this function suitable for deep-learning estimation.
  • CNNs: A CNN applies convolution, activation, pooling, up-sampling, and optionally fully connected layers to feature maps.A feature map is defined on a 2D grid with feature channels.
  • UNets: UNets use convolution, pooling, up-sampling, and activation layers without fully connected layers.Their encoder contracts spatial resolution while increasing feature channels, and the decoder reverses this process.
  • Training: Supervised learning tunes UNet parameters so outputs approximate corresponding training targets, typically using gradient-based optimization.The stated objective is evaluated over example input–output pairs, with SGD operating by batches.
  • Optimization: SGD explores parameter configurations along a highly non-exhaustive descent path, so network expressiveness alone does not ensure high-quality training.The paper motivates more deliberate control of parameter exploration.
  • Transfer learning: Transfer learning trains on one data distribution and evaluates on another when target-domain data is difficult or costly to obtain.The paper calls performance in new domains transferability.

3 The RadioMapSeer Dataset

RadioMapSeer provides simulated urban radio maps across cities, transmitter locations, simulation fidelities, and uncertainty conditions. The section also defines pathloss representation and error evaluation through truncation and gray-level conversion.

  • RadioMapSeer contains 700 city maps with 80 transmitter locations per map and corresponding coarse radio-map simulations.Coarse simulations use DPM and IRT2; fine IRT4 simulations are provided for the first two transmitters per map.
  • The dataset covers Ankara, Berlin, Glasgow, Ljubljana, London, and Tel Aviv using 256 × 256m2 urban map regions.Buildings and roads are represented as polygons and converted into 256 × 256 binary morphological images.
  • Coarse radio maps store pixelwise transmitter-to-location pathloss in dB, while higher-accuracy IRT4 simulations provide a refined evaluation surrogate.IRT4 uses four ray-geometry interactions and includes two transmitter locations for each map.
  • Dataset uncertainty is modeled with cars and missing buildings, including four perturbed datasets with m = 1, . . . , 4 missing buildings.Car-inclusive simulations perturb simulations based on maps without cars.
  • Pathloss values are truncated below the analytic noise floor and converted to gray levels, with zero denoting values below PL,trnc and one the maximal transmitter gain.The chosen thresholds are PL,thr = −127dB and PL,trnc = −147dB.
  • Gray-level conversion prevents weak-signal regions from dominating normalized error, aligning evaluation with the communication relevance of stronger signals.The paper notes that very negative pathloss values are effectively irrelevant and should not dominate overall error.

4 Estimating Radio Maps via RadioUNets

The paper introduces RadioUNet methods for learning radio-map estimation in multiple scenarios and evaluates them against state-of-the-art methods.

  • RadioUNet comprises deep-learning methods that estimate radio maps under different input, learning, and simulation conditions.
  • The methods are evaluated for accuracy and compared with state-of-the-art approaches.

4.1 Motivation for RadioUNet

RadioUNet uses UNets because radio-map estimation maps city-and-transmitter images to radio-map images and benefits from translation-invariant image processing. Its encoder extracts urban-geometry relationships and its decoder synthesizes the map.

  • UNets are a natural choice because the task maps an image of city geometry and transmitter location to an image representing the radio map.
  • RadioUNet incorporates translation invariance, matching the symmetry of the underlying physical phenomenon without requiring training.
  • The encoder extracts global relationships among urban features and the transmitter, while the decoder uses them to synthesize the radio map.The paper contrasts this global-information interpretation with the view that the network directly mimics ray tracing or Maxwell’s equations.

4.2 Different Setting in Radio Map Estimation

RadioUNet supports map-only and measurement-assisted inputs, accurate or perturbed maps, and coarse or randomized simulations. Training on coarse simulations can transfer to refined or real-world targets through zero-shot generalization, optionally aided by measurements.

  • City map and transmitter location: In the map-only setting, the UNet receives morphological images of the city map and transmitter location and estimates the radio map.
  • City map, transmitter location, and measurements: Measurement-assisted inputs add pathloss measurements to map and transmitter channels when the nominal map may not fully represent reality.The network combines map-based estimation with interpolation of accurate measurements.
  • Large and dense simulation dataset: Training uses dense 2D gray-level radio-map images generated from DPM, IRT2, or randomized combinations, with or without cars.Randomized simulations encourage reliance on broad radio-map behavior rather than exact coarse-simulation details.
  • Transfer to refined or real-life maps: Transfer to IRT4 or real-life maps is treated as zero-shot generalization because the network is trained only on coarse simulations, optionally supplemented by measurements.

4.3 RadioUNet Architectures

RadioUNet uses one or two UNets to estimate radio maps from urban-map inputs, with a second UNet enabling retrospective correction, adaptation, or thresholding.

  • RadioUNet Architectures: The basic RadioUNet uses two, three, or four input channels and produces a one-channel estimated radio map.The channels depend on whether measurements and cars are included.
  • RadioUNet Architectures: A WNet composes a second UNet whose inputs include the first UNet’s output alongside the original inputs.The first and second UNets together form the W-shaped architecture.
  • Retrospective Improvement: The second UNet inspects the first estimate with the city map and other inputs, then synthesizes an improved radio map.This retrospective stage is intended to correct visible inconsistencies with the map and physical phenomenon.
  • Retrospective Improvement: Retrospective improvement performs especially well when the first UNet is small, reducing network size without degrading performance.The paper reports this comparison in Fig. 5(a).
  • Training: The WNet is trained in two phases: first the initial UNet learns ground-truth maps, then the frozen first UNet supports training of the second.Both phases use MSE loss.
  • Transfer Adaptation: For transfer to sparse IRT4 data, the first UNet learns coarse simulations before the second UNet is trained on the small sparse dataset.The adaptation UNet is trained first on coarse data and then further trained on sparse IRT4 data.

4.4 Training

Training uses a fixed supervised split of RadioMapSeer and MSE-based optimization, with validation selection and testing on coarse or IRT4 simulations.

  • Dataset Splits: The 700 RadioMapSeer maps are split into 500 training, 100 validation, and 100 test maps.A second 400/100/200 split is also evaluated to examine the original test-set size.
  • Dataset Splits: The 400/100/200 split reuses the original 100 test examples within its 200-example test set.The reported comparison is intended to assess whether the original test set is too small.
  • Optimization: Models are trained with supervised MSE loss using Adam at learning rate 10^-4, for 50 epochs with batch size 15 and no regularization.The model with the smallest validation error is selected for testing.
  • Evaluation: Testing uses either coarse simulations or IRT4 simulations on test maps, depending on the evaluation setting.Validation error determines which epoch is retained.

4.5 RadioUNet Performance

RadioUNet performance is evaluated across simulation, sparse-measurement, transfer, and missing-building scenarios using NMSE and RMSE.

  • Transferability: Both sparse-IRT4 adaptation and randomized coarse-map training promote transferability across evaluation settings.RadioUNetC and RadioUNetS distinguish models without and with input measurements.
  • Transferability: RadioUNetS is trained and tested with a random number of input measurements between 1 and 300.Zero-shot IRT4 evaluates coarse-trained methods directly, while adaptation trains a second UNet using sparse IRT4 measurements.
  • Metrics: Performance is reported using both NMSE and RMSE, with PL,thr = -127dB for the pathloss threshold.The best IRT4 result in each category is marked in bold in Fig. 3.
  • Architecture and Transfer: Retrospective improvement is effective when the first UNet is small, while random coarse simulations plus an adaptor promote transferability with missing buildings.These comparisons appear in Fig. 5(a) and Fig. 5(b).
  • Dataset Robustness: The 400/100/200 and original 500/100/100 splits have comparable performance, indicating that the original 100-example test set is not too small.The two test sets share the last 100 examples of the larger test set.

5 Comparison of RadioUNet to state-of-the-art

RadioUNet is compared with interpolation, tomography, ray-tracing-related, and prior deep-learning methods on accuracy and computational cost. It is reported as both more accurate and substantially faster in the tested settings.

  • Accuracy and Practicality: RadioUNet outperforms data-driven interpolation, tomography, and the previously proposed deep-learning approach in the reported comparison.The alternatives require separate training or optimization for each map, whereas RadioUNet is trained offline once.
  • Environmental Variation: RadioUNet can use pathloss measurements as input to handle cars when trained on simulations containing cars.This contrasts with methods requiring recomputation when environmental features such as moving cars change.
  • Model-Based Simulation: RadioUNet estimates radio maps roughly two to three orders of magnitude faster than the efficient dominant pathloss method.Reported runtimes are 10^-3sec to 10^-2sec for RadioUNet, roughly 1sec for WinProp, 10sec for IRT2, and 10^2sec for IRT4.
  • Data-Driven Interpolation: RadioUNet is roughly three orders of magnitude faster than RBF interpolation and five orders faster than tensor completion interpolation.The comparison plots average NMSE over 80 transmitters as a function of the number of samples.
  • Tomography: Tomography models penetration attenuation but omits diffraction and reflections, making it less realistic than DPM and IRT for high-frequency urban signals.Its reported runtime is on the order of 10^2sec.
  • Complexity Comparison: RadioUNet’s trained-network runtime is measured separately from training, which is performed offline once for all.This runtime definition applies to the computational comparison.

6 Applications

The applications use RadioUNet radio-map estimates for coverage classification and fingerprint localization. Coverage prediction benefits from a staged architecture, while localization depends on accurate maps and a suitable number of reported signal strengths.

  • 6.1 Coverage Classification: Coverage classification determines whether a transmitter’s pathloss exceeds a threshold at a receiver location.It supports received-signal-strength and interference decisions.
  • 6.1 Coverage Classification: Simple UNets fail to predict coverage maps reliably because gradient descent does not exhaustively search parameter configurations.The proposed approach therefore models coverage through the underlying radio-map phenomenon.
  • 6.1 Coverage Classification: The WNet architecture combines RadioUNet radio-map prediction with a second thresholding UNet that computes the coverage map.The RadioUNet receives the city and transmitter inputs; the thresholding UNet also receives the predicted radio map.
  • 6.1 Coverage Classification: Curriculum learning trains the thresholding UNet from a soft coverage map with α = 1 to a sharp transition with α = 128.The soft map is interpreted as the probability that a location lies in the coverage area.
  • 6.1 Coverage Classification: Figure 7 presents SWNet service-map accuracy across thresholds and an example service map.The caption identifies the middle panel as a coverage map at threshold 0.5 and the right panel as RMSE across thresholds.
  • 6.2 Pathloss Based Fingerprint Localization: Fingerprint localization intersects ε-level sets from multiple estimated radio maps to identify a receiver location.A localized intersection indicates that the receiver has been located with high probability.
  • 6.2 Pathloss Based Fingerprint Localization: Localization accuracy depends on accurate radio maps and a number of reported signal strengths that is large enough but not too large.Prediction-error outliers can exclude the true location, while too few signal strengths leave multiple candidate locations.
  • 6.2 Pathloss Based Fingerprint Localization: Random subsets of transmitters and random ε values mitigate outliers by generating multiple candidate localization sets.The method selects the non-empty outcome with the smallest variance; an example uses K = 10, J = 5, R = 5, and ε = 0.03.

7 Conclusion

The paper introduces RadioUNet and the RadioMapSeer dataset, develops simulation-to-real transfer approaches, and reports superior runtime and accuracy against the state of the art.

  • 7 Conclusion: RadioUNet estimates radio maps from city geometry and transmitter location, optionally incorporating pathloss measurements and car locations.The paper presents it as a deep learning method for simulating radio maps.
  • 7 Conclusion: RadioMapSeer provides 56,000 simulated radio maps across different city and transmitter locations, with accompanying city, transmitter, and car maps.Its simulations include coarse variants and, in one type, cars that perturb received signal strength.
  • 7 Conclusion: Simulation-to-real transfer approaches are developed, and the methods demonstrate superior runtime and accuracy relative to the state of the art.The conclusion reports these outcomes without specifying a numerical comparison in the supplied passage.
Loading 1911.09002v3…