Geometry-Forcing Masked Completion for Radio Maps in Unseen Cities

A radio map gives the received signal strength at every point of a city for a given transmitter. Neural networks can predict these maps from building layouts about four orders of magnitude faster than ray-tracing simulators. They have two weaknesses. They need thousands of densely simulated maps to train, and their accuracy drops sharply on a city whose architecture they have not seen. Every entrant in the first pathloss prediction challenge got worse on Istanbul after training on six European cities (The First Pathloss Radio Map Prediction Challenge). In a real deployment, the cheap and plentiful data is a building map plus a few hundred scattered signal measurements, not a dense map. We propose to train and adapt radio map networks with a self-supervised task built from those scattered measurements. The network hides some measurements and predicts them from the rest plus the building geometry.

The central design choice is the shape and size of what gets hidden. Radio maps are spatially smooth. If small random points are hidden, the network can fill them in by local interpolation, which is essentially classical Kriging, and it learns nothing about how buildings cast shadows. We will instead hide whole street segments and blocks larger than a typical shadow. Neighbouring values are then useless, and the network has to reason from the layout. A network with attention at several spatial scales (local windows, plus sparse grid attention spanning the whole map) is the right model to use that signal. The same objective can also be run at test time on a new city, using only that city's own scattered measurements, with no dense ground truth required.

Core Claim

The value of masked self-supervision for radio maps depends on the size of the masks relative to the field's correlation length. Masks shaped like streets and blocks, larger than shadow regions, combined with an encoder that attends at multiple scales, will cut error in non-line-of-sight areas far from any measurement. This holds both when only a few dense maps are labeled and when the model is adapted per city at test time on unseen urban geometry. Random-point masking or a single-scale encoder will not deliver these gains.

Relation to Prior Work

The closest work is FM-RME. It pretrains an attention autoencoder with masked reconstruction across space, time and frequency, adds a graph network that builds in translation and rotation symmetry, and reports zero-shot transfer to an unseen dataset and robustness at 10% sampling density. It motivates our proposal because it shows masked pretraining helps radio maps. It differs from ours in three ways: it uses single-scale attention, it does not study how mask geometry relates to propagation structure, and it does not adapt per city at test time. It is also evaluated only on its own simulated spectrum grids.

RMTransformer supplies our encoder. It is a multi-axis vision transformer, combining local window and dilated grid attention across resolutions from 128×128 down to 8×8, and it reports a 62% RMSE gain over RadioUNet. It is fully supervised on 19,016 paired maps and has no answer for the case where dense labels are scarce. SymNet shows that attention gates and building-distance input channels help reconstruction from as few as 20 sampled points. It also needs dense supervision.

RadioUNet established that feeding sparse measurements as an input channel lets a convolutional network correct for missing building data. We keep that input format but also use the measurements as a training target. SpectrumFM motivates the direction from the signal side: masked pretraining reaches target accuracy with 10% of labels on 1D IQ signals. The authors explicitly leave spatial and geographic context open.

One finding constrains our design. RF-MatID reports that the large vision model DINOv3 did not consistently beat ResNet-50 or MLPs under geometric shift. Scale and generic pretraining alone therefore do not buy robustness in RF. Our bet is that the pretext task must encode propagation structure, which is why mask geometry, not model size, is the variable we study.

Research Program

The model takes four input channels: a building height map, a transmitter location map, a sparse measurement map, and a binary mask marking where measurements exist. The encoder is RMTransformer-style, with multi-axis attention blocks at four resolutions and a convolutional decoder with skip connections. Let $G$ be the geometry, $S$ the set of measured pixels with values $m_i$, and $W \subset S$ a withheld subset. The self-supervised loss is

$$ \mathcal{L}_{\text{ssl}} = \mathbb{E}_{W \sim \pi_s}\; \frac{1}{|W|}\sum_{i \in W} \big(f_\theta(G, m_{S\setminus W})_i - m_i\big)^2 , $$

where $f_\theta$ is the network's predicted map and $\pi_s$ is a distribution over withheld sets at spatial scale $s$. Loss is computed only at withheld measured pixels, so no dense map is ever needed.

We will compare three mask families. The first withholds random points. The second withholds street-aligned strips, traced along the free-space skeleton of the building map. The third withholds square blocks whose side ranges from 8 pixels to a quarter of the map. The main mixture samples all three.

Training has two stages. Pretraining runs on many cities that have only sparse measurements, then fine-tunes on a small number of cities with dense maps, using full-map loss plus $\mathcal{L}_{\text{ssl}}$. Test-time adaptation then runs a few hundred steps of $\mathcal{L}_{\text{ssl}}$ on each new city's own measurements before prediction.

We also need a stand-in for the gap between simulation and reality. RadioMapSeer offers several simulator fidelities over the same 700 city layouts (RadioUNet). We treat the coarse dominant-path model as "cheap simulation". We treat high-fidelity ray tracing with cars as "reality", and draw sparse measurements from it with added noise. This is only a proxy. The file contains no dense real-world radio map dataset, so over-the-air validation is out of scope and stated as a limitation.

Baselines

The naive reference points are ordinary Kriging on the sparse measurements and RadioUNet_S trained supervised on the small labeled set. Kriging is the interpolation floor our mask-scale argument says random masking collapses toward. FM-RME reports beating Kriging at high sparsity but quotes no numbers; no reported result found on Lacuna for Kriging RMSE on RadioMapSeer.

The strongest supervised competitor is the multi-axis transformer of RMTransformer, which reports RMSE 0.007148 on the USC dataset with full supervision, together with PMNet. PMNet is the challenge winner, with RMSE 0.0383 on the challenge leaderboard reported in The First Pathloss Radio Map Prediction Challenge. For both, we will train our reimplementation with sparse-measurement input channels at each label budget, using the same parameter count, GPU-hours and number of hyperparameter trials as our method. With full labels, we will first check that our reimplementations come close to those reported numbers.

The strongest self-supervised competitor is FM-RME-style pretraining: random masking with a single-scale transformer encoder. FM-RME reports no numerical RMSE in the file, so we reimplement it under the same compute budget.

Two ablations isolate the key ingredient. The first crosses mask family (random points versus street and block) with encoder type (flat transformer versus multi-axis). The second toggles test-time adaptation on and off.

Our method counts as beating the field if it reduces RMSE on the Istanbul blind set relative to all baselines at matched compute, at label budgets of 1%, 5% and 10% of training cities. The gain must hold across three seeds. It must be concentrated in non-line-of-sight pixels more than 20 pixels from the nearest measurement.

Why not just…?

Why not just pretrain on cheap low-fidelity simulation and fine-tune on a few high-fidelity maps? This is the most serious threat, because simulator outputs are dense and cost about 100 s per map, not field labor (RadioUNet). Low-fidelity pretraining, however, inherits the simulator's systematic errors, such as missing cars and uniform materials, exactly where reality departs from it. Our objective trains on samples drawn from "reality", even though they are sparse. We will run dominant-path pretraining followed by fine-tuning as a baseline and also combine it with our objective. If simulator pretraining alone matches our method on the ray-tracing-with-cars target, the contribution shrinks to test-time adaptation, and we will report it that way.

Why not just fine-tune on the new city's measurements with a pointwise loss? That is our objective with random-point masks, so it is precisely the mask-scale ablation. Our claim is falsified if random-point adaptation matches block adaptation in the far, non-line-of-sight stratum.

Why not just use Kriging or a supervised U-Net with measurements as input? Kriging cannot predict shadows inside a fully unmeasured block. Supervised RadioUNet_S has no mechanism to adapt to an unseen city. The far-from-measurement error stratification will show whether either closes the gap.

Why not just use an off-the-shelf pretrained vision backbone? RF-MatID found DINOv3 offered no consistent advantage under geometric shift. We include an ImageNet-initialized multi-axis encoder, fine-tuned identically, as a control.

Proposed Experiments

The primary benchmark is RadioMap3DSeer. We train on six cities and test on the 84 held-out Istanbul maps, following the challenge protocol of masking building footprints (The First Pathloss Radio Map Prediction Challenge). Label budgets range from 1% to 100% of training cities, and measurement density from 0.1% to 1% of pixels. The fidelity-gap proxy uses RadioMapSeer's dominant-path and ray-tracing-with-cars variants. The USC dataset serves as a full-supervision sanity check against RMTransformer's reported number. CKMImageNet's 28 GHz channel gain maps, available at three resolutions, test whether the mask-scale effect carries to a different frequency and to rural and factory environments; no reported result found on Lacuna for CKMImageNet.

Metrics are RMSE on free-space pixels, broken down by line-of-sight status and by distance to the nearest measurement, plus coverage error at a fixed received-power threshold.

The decisive check is a curve: test RMSE against mask block size, plotted beside the empirical correlation length of the field (from the measurements' variogram). The hypothesis predicts that gains appear only once masks exceed that length. It also predicts that the multi-axis encoder benefits more than the flat encoder in that regime. A flat curve, or an equal benefit for both encoders, refutes the core claim.