Radio U-Net: Deep Learning for Radio Maps

• Radio U-Net is a family of U-Net based architectures that map structured radio inputs like urban layouts and sparse measurements to dense radio maps.
• It leverages encoder-decoder networks with skip connections and domain adaptation techniques for accurate and low-latency radio propagation predictions.
• Applications include urban pathloss estimation, indoor millimeter-wave coverage, and radio astronomy segmentation with performance metrics such as ~1 dB RMSE.

Searching arXiv for "RadioUNet" and "Radio U-Net" to ground the article in the relevant literature. Searching for papers with exact phrases and close variants. The label Radio U-Net has been used for several U-Net-derived convolutional networks in radio propagation and radio astronomy. In one major usage, RadioUNet denotes a deep learning method for estimating large-scale pathloss, or a “radio map,” from urban geometry, transmitter location, and optionally sparse measurements (Levie et al., 2019). In later work, the same naming appears in indoor wave-propagation prediction with a Double U-Net and transfer learning (Fu et al., 2024), and in segmentation of faint diffuse radio sources in interferometric surveys (Stuardi et al., 2024). Across these usages, the common formulation is an encoder-decoder with skip connections that maps structured radio-domain inputs to dense spatial outputs.

1. Terminology, scope, and problem formulations

In the radio-map literature, the central object is the large-scale pathloss function $G(x,y)$, defined in dB as

$G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$

where $P_{Rx}$ and $P_{Tx}$ are received and transmitted powers (Levie et al., 2019). Classical empirical models are written as

$L(d) = L_0 + 10 \alpha \log_{10}(d),$

with $d=\|x-y\|_2$, but such radially symmetric laws fail to capture shadowing by buildings, street canyons, and related effects (Levie et al., 2019).

RadioUNet replaces that radial approximation with a learned mapping from a rasterized scene description to a dense pathloss field. One summary writes this as

$$\hat f = \mathrm{RadioUNet}(M,s,T) \approx G(\cdot,\cdot),$$

where $M$ encodes the $2D$ geometry, $s$ encodes sparse in-situ measurements, and $G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$0 encodes the transmitter location (Levie et al., 2019). A closely related formulation used in localization-oriented work is

$G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$1

with $G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$2 a binary city map and $G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$3 a one-hot base-station map (Yapar et al., 2022).

The same architectural template has also been applied to other radio-domain segmentation tasks. In radio astronomy, a U-Net was used to classify clean signal and radio-frequency interference signatures in $G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$4 time-ordered data (Akeret et al., 2016). A later astronomical paper explicitly titled “Radio U-Net” used a fully convolutional network to detect radio halos, relics, and cosmic web filaments in radio surveys (Stuardi et al., 2024). This suggests that the term is best understood as a family name rather than a single immutable architecture.

2. Canonical RadioUNet for urban radio-map estimation

The original RadioUNet formulation is a U-Net encoder-decoder with skip connections and no fully connected layers (Levie et al., 2019). One implementation summary describes a network with 9 resolution levels down to $G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$5, then 9 up levels back to $G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$6, with channel counts increasing through $G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$7 and decoding symmetrically to a one-channel output (Levie et al., 2019). A later operational description used in real-time localization presents a five-scale encoder and decoder with channels $G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$8, a bottleneck with two $G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$9 convolutions of $P_{Rx}$0 channels, and skip-concatenations at matched resolutions (Yapar et al., 2020). Both descriptions preserve the essential structure: repeated convolutional blocks, downsampling, symmetric upsampling, and long skip connections.

The canonical inputs are raster images defined on a common $P_{Rx}$1 grid. The city map $P_{Rx}$2 is a binary image with value $P_{Rx}$3 inside building footprints and $P_{Rx}$4 outside; the transmitter location is encoded as a one-hot channel $P_{Rx}$5 or as a base-station map $P_{Rx}$6 (Levie et al., 2019). Optional channels include cars or other obstacles, and in the sample-aided version an additional channel $P_{Rx}$7 holds measured pathloss values at a few pixel locations, with zero elsewhere (Levie et al., 2019). Heights are uniform in the baseline dataset, with buildings at $P_{Rx}$8 and Tx/Rx at $P_{Rx}$9 (Levie et al., 2019).

The output is a single-channel dense radio map. In one formulation, $P_{Tx}$0 approximates the true pathloss $P_{Tx}$1 in dB from base station $P_{Tx}$2 to grid cell $P_{Tx}$3 (Yapar et al., 2020). In another, the target maps are pathloss images truncated at $P_{Tx}$4 and linearly scaled to $P_{Tx}$5 gray levels (Levie et al., 2019). This normalization allows the pathloss-prediction problem to be treated as an image-to-image regression task.

3. Data generation, losses, and transfer mechanisms

The standard training corpus is RadioMapSeer, built from 700 distinct $P_{Tx}$6 urban scenes extracted from OpenStreetMap and simulated at $P_{Tx}$7 resolution, $P_{Tx}$8, and $P_{Tx}$9 bandwidth (Levie et al., 2019). For each scene, 80 Tx locations were simulated using the Dominant Path Model and $L(d) = L_0 + 10 \alpha \log_{10}(d),$0-bounce Intelligent Ray Tracing, both with and without cars, yielding 56,000 radio-map images (Levie et al., 2019). A higher-accuracy IRT4 set with 1,400 maps was generated for transfer experiments (Levie et al., 2019).

The basic training objective is pixel-wise mean-squared error,

$L(d) = L_0 + 10 \alpha \log_{10}(d),$1

or, equivalently,

$L(d) = L_0 + 10 \alpha \log_{10}(d),$2

over all pixels and training examples (Levie et al., 2019). Optimization uses Adam with $L(d) = L_0 + 10 \alpha \log_{10}(d),$3, batch size $L(d) = L_0 + 10 \alpha \log_{10}(d),$4, and $L(d) = L_0 + 10 \alpha \log_{10}(d),$5 epochs in one summary; a localization-oriented account describes pre-training on $L(d) = L_0 + 10 \alpha \log_{10}(d),$6 coarse maps, optional fine-tuning on the $L(d) = L_0 + 10 \alpha \log_{10}(d),$7 IRT4 maps, and on-the-fly random flips and $L(d) = L_0 + 10 \alpha \log_{10}(d),$8 rotations (Levie et al., 2019).

A distinctive aspect of RadioUNet is domain adaptation from coarse simulations to higher-fidelity or partially observed regimes. One approach trains on random convex combinations of Dominant Path Model and IRT2 outputs so that the network learns only the common coarse patterns and then generalizes zero-shot to IRT4 or real measurements (Levie et al., 2019). A second approach freezes the first U-Net and trains a small U-Net “adaptor” on sparse IRT4 samples using a weighted MSE on measured pixels,

$L(d) = L_0 + 10 \alpha \log_{10}(d),$9

A third approach, RadioUNet$d=\|x-y\|_2$0, incorporates sparse measurements directly as an input channel so that the network learns to blend geometry-based prior with data-driven interpolation (Levie et al., 2019).

4. Quantitative performance and localization use

On held-out data, RadioUNet reports test RMSE on gray levels of $d=\|x-y\|_2$1–$d=\|x-y\|_2$2, corresponding to $d=\|x-y\|_2$3 RMSE on an $d=\|x-y\|_2$4 range, with runtime of $d=\|x-y\|_2$5–$d=\|x-y\|_2$6 per $d=\|x-y\|_2$7 map on GPU (Levie et al., 2019). The no-sample variant, RadioUNet$d=\|x-y\|_2$8, is reported at approximately $d=\|x-y\|_2$9 RMSE, while RadioUNet$$\hat f = \mathrm{RadioUNet}(M,s,T) \approx G(\cdot,\cdot),$$0 with 50 samples achieves $$\hat f = \mathrm{RadioUNet}(M,s,T) \approx G(\cdot,\cdot),$$1; adaptation on IRT4 yields best RMSE of approximately $$\hat f = \mathrm{RadioUNet}(M,s,T) \approx G(\cdot,\cdot),$$2 gray, or $$\hat f = \mathrm{RadioUNet}(M,s,T) \approx G(\cdot,\cdot),$$3 (Levie et al., 2019). Compared with model-based simulation and interpolation baselines, the reported advantages are simultaneously lower error and orders-of-magnitude lower latency (Levie et al., 2019).

In a later localization pipeline, RadioUNet serves as the radio-map engine inside LocUNet and related RSS-based positioning methods (Yapar et al., 2022). The user reports RSS from known base stations, which is converted to pathloss values $$\hat f = \mathrm{RadioUNet}(M,s,T) \approx G(\cdot,\cdot),$$4, and the predicted maps $$\hat f = \mathrm{RadioUNet}(M,s,T) \approx G(\cdot,\cdot),$$5 are queried through level sets

$$\hat f = \mathrm{RadioUNet}(M,s,T) \approx G(\cdot,\cdot),$$6

Votes are accumulated as $$\hat f = \mathrm{RadioUNet}(M,s,T) \approx G(\cdot,\cdot),$$7, the maximal-vote region $$\hat f = \mathrm{RadioUNet}(M,s,T) \approx G(\cdot,\cdot),$$8 is extracted, and the location estimate is the center of mass

$$\hat f = \mathrm{RadioUNet}(M,s,T) \approx G(\cdot,\cdot),$$9

(Yapar et al., 2020). The end-to-end urban localization error is reported as $M$0–$M$1 on average, with worst-case $M$2, while a single forward pass for one BS map estimation takes approximately $M$3–$M$4 on an NVIDIA Quadro GP100 or approximately $M$5–$M$6 on a modern $M$7-core CPU (Yapar et al., 2020).

A separate empirical study compared RadioUNET with several U-Net variants for radio-frequency power prediction and found that RadioUNET achieved average MAE $M$8 on the publicly available RadioMapSeer $M$9 dataset, while deeper customized variants achieved lower MAE on that benchmark (Ozyegen et al., 2020). That study attributes RadioUNET’s under-performance in that setting to a relatively shallow bottleneck and fixed filter sizes (Ozyegen et al., 2020). This is a methodological contrast rather than a contradiction: the cited works use different datasets, simulation pipelines, and evaluation protocols.

5. Indoor and three-dimensional extensions

A later indoor-propagation model again uses the name Radio U-Net, but in a distinct architectural form (Fu et al., 2024). The task is to predict indoor millimeter-wave coverage maps from $2D$0 floor plans and furniture layouts rather than from urban outdoor footprints. Inputs are encoded as a $2D$1 tensor: a relative permittivity map $2D$2, a conductivity map $2D$3, and Free Space Path Loss,

$2D$4

while the output is a $2D$5 received-power map in dBm (Fu et al., 2024).

This indoor Radio U-Net is a two-stage (“Double”) U-Net. Each stage uses EfficientNet-B5 pretrained on ImageNet as the encoder backbone, attention gates on selected skip connections, and decoder blocks with two $2D$6 convolutions, batch normalization, ReLU, and Dropout(0.5) (Fu et al., 2024). The first stage predicts a coarse map; the second refines it by taking $2D$7 as input (Fu et al., 2024). Training uses ridge-regularized MSE,

$2D$8

with $2D$9, Adam, linear decay from $s$0 to $s$1 over $s$2 epochs, batch size $s$3, early stopping after three validation-loss plateaus, and augmentation by $s$4 rotations and vertical flips (Fu et al., 2024). Across four scenarios, the final-stage predictions report RMSE $s$5–$s$6 and MAE $s$7–$s$8, while the stage-one coarse predictions are $s$9–$G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$00 RMSE; the paper summarizes this as $G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$01–$G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$02 RMSE reduction and $G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$03 inference on a GPU (Fu et al., 2024).

Another U-Net-based REM predictor, REM-U-net, is trained on the large-scale RadioMap3DSeer dataset of 56,080 radio environment maps and evaluated in the context of the 2023 IEEE ICASSP Signal Processing Grand Challenge (Sallouha et al., 2023). Its classical U-shape uses four encoder and four decoder levels with $G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$04 convolutions, inputs such as building height $G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$05, transmitter height $G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$06, and a fractional line-of-sight map $G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$07, and predicts a $G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$08 path-gain map in dB (Sallouha et al., 2023). The best reported hold-out result is nRMSE $G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$09 with average inference time $G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$10; exact per-pixel LoS input lowers nRMSE to $G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$11 but increases runtime to $G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$12 (Sallouha et al., 2023). A plausible implication is that later “radio U-Net” systems increasingly combine geometry encodings with explicit propagation side information such as LoS, FSPL, or material parameters.

6. Radio-astronomy segmentation and RFI mitigation

Outside wireless pathloss prediction, U-Net-based radio models have been used for two-dimensional time-frequency and image-domain segmentation. An early example is “Radio frequency interference mitigation using deep convolutional neural networks”, which applies a U-Net to $G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$13 time-ordered spectrograms from radio telescopes (Akeret et al., 2016). The published architecture is a symmetric U-shaped encoder-decoder with three down/up-sampling levels, 64 base filters, unpadded $G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$14 convolutions with ReLU, dropout with keep probability $G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$15, $G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$16 weight decay $G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$17, and a final $G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$18 convolution to $G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$19 classes followed by a pixel-wise softmax,

$G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$20

Inputs are single-channel $G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$21 time-frequency images and outputs are $G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$22 one-hot masks (Akeret et al., 2016). On simulated data, the U-Net achieved ROC AUC $G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$23, PR AUC $G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$24, and max $G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$25, compared with PR AUC $G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$26 and max $G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$27 for SEEK SumThreshold; throughput during prediction is reported as $G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$28 per GPU (Akeret et al., 2016).

A more recent astronomical use adopts the explicit title “Radio U-Net: a convolutional neural network to detect diffuse radio sources in galaxy clusters and beyond” (Stuardi et al., 2024). Here the task is semantic segmentation of faint, extended radio emission in survey images rather than pathloss regression. The network is a U-Net-style fully convolutional autoencoder with four down-sampling stages, a bottleneck, and five up-sampling stages; each encoder stage repeats $G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$29 convolution, batch normalization, and ReLU twice, then applies $G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$30 max-pooling and dropout $G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$31 (Stuardi et al., 2024). In the decoder, skip connections are implemented by summation, not concatenation, and the final $G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$32 convolution with softmax produces background and source-probability maps (Stuardi et al., 2024). The model was trained on synthetic radio observations derived from cosmological MHD simulations and evaluated on LoTSS cluster cutouts. Reported results include mean IoU $G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$33 on the synthetic test set, compared with IoU $G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$34 for a $G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$35 threshold baseline, and a cluster-level operating point on 246 clusters with accuracy $G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$36, precision $G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$37, and recall $G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$38 at $G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$39 (Stuardi et al., 2024). The paper also states that 83\% of clusters exhibiting diffuse radio emission were accurately identified (Stuardi et al., 2024).

Later RFI work continues the same architectural line with added multi-scale and attention modules. EMSCA-UNet for RFI detection uses a depth-$G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$40 U-Net with Efficient Channel Attention and a Multi-Scale Convolutional Attention block, is trained on 1,384 samples from the $G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$41-m Yunnan radio telescope, and reports Precision $G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$42, Recall $G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$43, F1 $G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$44, and IoU $G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$45, outperforming baseline U-Net by approximately $G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$46 in F1 and IoU (Gu et al., 2024). Although this model is not itself titled Radio U-Net, it illustrates how radio-domain U-Net systems have evolved toward explicit multi-scale and attention mechanisms.

7. Limitations, ambiguities, and research directions

The primary limitations reported for RadioUNet-style radio-map estimators are tied to geometry fidelity, dimensionality, and transfer. The original urban model requires accurate building footprints and heights; missing or erroneous obstacles degrade performance (Levie et al., 2019). It is also based on a $G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$47 assumption, with no explicit vertical multi-floor variation or antenna-tilt modeling, and it targets single-Tx estimation rather than simultaneous multi-Tx interference maps (Levie et al., 2019). Transfer to real measured data remains an open issue: the IRT4-based experiments are described as promising, but the same source notes that ultimate proof requires real-world measurement campaigns (Levie et al., 2019).

Indoor and survey-imaging variants exhibit different failure modes. The indoor Double U-Net highlights the computational trade-off between full ray tracing and learned surrogates, but its training set is still produced by a deterministic simulator and its reported robustness is bounded by the diversity of generated furniture configurations and propagation settings (Fu et al., 2024). In diffuse-source detection, false negatives are associated with small, high-$G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$48 halos under-represented in training, while false positives are often extended radio galaxies not present in the synthetic training distribution; fine-tuning on the available real LoTSS data does not improve peak accuracy beyond $G(x,y) = P_{Rx}(dB) - P_{Tx}(dB),$49 (Stuardi et al., 2024).

A recurrent misconception is that Radio U-Net denotes a single fixed architecture. The literature does not support that reading. The name refers variously to outdoor urban pathloss estimation (Levie et al., 2019), indoor millimeter-wave coverage prediction with a Double U-Net and EfficientNet-B5 (Fu et al., 2024), and radio-astronomy image segmentation (Stuardi et al., 2024). What remains stable across these uses is the U-shaped encoder-decoder, dense spatial prediction, and reliance on skip connections to preserve fine-scale structure. A plausible implication is that “Radio U-Net” functions less as a proprietary model name than as a domain-specific pattern: a U-Net whose inputs encode radio-relevant scene structure and whose outputs are dense spatial fields, masks, or maps.

Current directions in the cited literature include explicit adaptation from coarse to higher-fidelity simulations (Levie et al., 2019), incorporation of sparse measurements into the input representation (Levie et al., 2019), transfer learning with pretrained encoders and attention gates (Fu et al., 2024), preprocessing that injects line-of-sight structure (Sallouha et al., 2023), and multi-scale attention for RFI segmentation (Gu et al., 2024). Taken together, these developments indicate a broad methodological trajectory from plain U-Net surrogates toward hybrid systems that combine geometric priors, learned attention, domain adaptation, and physics- or simulation-informed supervision.