How can channel knowledge maps model FR3 UAV base station attenuation?

A 16,180-sample FR3 map set with real building heights

The dataset holds 16,180 CKM samples drawn from 3,236 urban locations across 66 cities, all simulated at 7.125 GHz with UxNB transmitter heights from 10 m to 500 m. Each sample is a 513×513 grid at 1 m horizontal resolution, which the authors note corresponds to 24 wavelengths at this frequency.

Every sample stores six aligned layers: two environmental layers (building height profile and a binary building occupancy mask) and four channel layers (binary LOS/NLOS visibility, attenuation in dB, multipath RMS delay spread in ns, and multipath RMS angular spread in degrees). In-building elements are set to zero and are meant to be masked out during training and evaluation.

The environmental layers come from global-scale 3D building datasets rather than OpenStreetMap alone. The paper argues this matters because OSM height tags are available in less than 10% of cases, while UxNB modeling depends heavily on deployment height. The authors also note that the two global building sources have geographic gaps, so they searched the more recent one first and used the other as backup.

FR3 and 500 m heights break terrestrial assumptions

FR3 sits between 7 and 24 GHz, and the paper frames it as a balance between the reliable propagation of sub-6 GHz FR1 and the large bandwidth of millimeter-wave FR2. The 7.125 GHz carrier was chosen because WRC-23 identified it as the only FR3 test band available in all ITU regions.

The authors argue that existing CKM datasets are unsuitable for this setting for three stated reasons: they cover only FR1 or FR2, they do not use UxNB-specific geometries such as 300–500 m coverage diameters and elevated transmitters, and they rely on less accurate 3D city models. The only dataset with any 3D channel content, UrbanRadio3D, uses receiver heights up to 20 m, which the paper calls insufficient.

The geometry itself is the reason. With UxNBs at up to 500 m, the resulting channel differs substantially from terrestrial channels, and air-to-ground modeling needs accurate building height information. This is the gap the dataset is built to fill.

The LOS branch adds a two-ray term and a fitted lookup table

The model separates the map into LOS and NLOS regions using the dataset visibility mask, then combines a LOS attenuation estimate and an NLOS estimate element-wise. Reported accuracy assumes this supplied mask; generating it via ray tracing or shadow projection is excluded from the timing but is described as being in the same order of magnitude as the proposed solution.

For LOS, the attenuation is free-space path loss plus a coherent ground-reflection two-ray term, plus a height-dependent scalar bias and a residual term capturing horizontal distance and transmitter height effects. The two-ray term keeps the phase difference between the direct and reflected paths, with an effective reflected-field amplitude and a fitted phase offset.

Calibration happens in two steps. In each 5 m height bin, candidate reflection amplitudes from 0 to 1.5 in steps of 0.05 and 48 uniformly spaced phases in [−π, π) are evaluated, and the combination minimizing bin RMSE is kept; the remaining radial term is the mean residual error. In implementation this becomes a lookup table over height and distance, with linear interpolation at inference.

The NLOS branch recalibrates COST 231-Hata with local geometry

For NLOS receivers, the model starts from a COST 231-Hata baseline and augments it with 13 contextual features derived from the building height grid and building presence map. These include building density, normalized building heights, shadowed ground fraction at two window sizes (15 and 41 pixels, corresponding to ±7 m and 20 m scales), a normalized path elevation angle, a deterministic shadowing proxy, and interaction cross-terms.

The paper is explicit that COST 231-Hata was designed for sub-3 GHz terrestrial links, so here it is used as a regression feature with coefficients fitted on training data rather than as a physically valid model at 7.125 GHz. The prediction is a dot product between the 14-element feature vector and a calibrated weight vector.

Weights are fitted per regime. Scenes are partitioned into nine regimes from three ITU topologies (suburban, urban, dense urban) crossed with three altitude bins: low (h_tx ≤ 60 m), medium (60 m < h_tx ≤ 100 m), and high (h_tx > 100 m). For each regime, up to 1,024 NLOS channel realizations per training sample feed a ridge regression with an L2 penalty that omits the intercept.

Unseen-city error is 1.93 dB overall but 3.54 dB in NLOS

The split is strictly by city: 10,840 samples from 37 training cities, 2,750 from 15 validation cities, and 2,590 from 14 final test cities, with no city moved between partitions. After excluding building interiors and channels lacking valid ray tracing outputs, the test set evaluates over 4.23×10^8 ground receivers.

Overall test RMSE is 1.93 dB, split into 1.74 dB for LOS and 3.54 dB for NLOS. The paper notes that only 7.4% of those receivers are in NLOS states because of the elevated UxNB geometry, so the element-weighted overall score is LOS-dominated; 424 of the 2,590 test maps nonetheless contain more than 25% NLOS receivers, which is why regime-specific results are reported separately.

By ITU topology, suburban shows the lowest overall error at 1.70 dB, urban reaches 2.30 dB, and dense urban is worst at 2.57 dB overall and 4.13 dB in NLOS. Per-city RMSE ranges from 1.61 dB in Halifax to 2.59 dB in Osaka, with an unweighted 14-city mean of 1.98 dB. The authors state that fitted coefficient vectors change substantially across topologies, so a single global set of slopes would mix distinct relationships.

Higher UxNBs are easier, and the neural baseline splits the result

Breaking the test error down by altitude, overall RMSE falls monotonically from 2.14 dB in the low regime to 1.97 dB in the medium regime and 1.71 dB in the high regime. The paper aligns this with the observation that higher-altitude channels are usually more stable and predictable.

Against a height-conditioned GMM U-Net trained on the same split and evaluated on the same 2,590 maps and mask, the proposed model is better overall (1.93 dB versus 4.36 dB) and better in LOS (1.74 dB versus 4.43 dB). The neural baseline is better in NLOS (3.34 dB versus 3.54 dB), which the authors read as evidence that neural output layers suit multipath NLOS channels while the calibrated LOS model is more stable.

The paper cautions against ranking these numbers against published CKM and radio map models, since those were designed under different frequencies, heights, and generated outputs. It offers scale context only: RadioUNet reports 1.60 dB on fixed-height 256×256 RadioMapSeer grids, TSRDiff 3.62 dB under sparse target sampling, and RO-HMTGP 4.19–6.00 dB on empirical RSRP data.

Each calibrated term earns its place, and runtime is 0.06 s per map

Ablation on the 2,750-sample validation set shows the coherent ground-reflection term cuts LOS RMSE from 3.79 dB to 1.74 dB. For NLOS, the raw uncalibrated COST 231-Hata model yields an RMSE of 25 dB at 7.125 GHz, while the version with fitted regression coefficients and topology- and height-specific calibration reaches 3.20 dB on validation.

The authors also report that a compact model retaining eight of the 14 features, using only the larger window and no interaction cross-terms, preserves over 90% of the full model's improvement over COST 231-Hata.

On an AMD Ryzen 5 5600X with GPU disabled, the proposed model has a median CPU runtime of 0.06 s per map with supplied masks and no file I/O. The paper reports 66.8k stored scalars plus 126 NLOS coefficients and 3.684M counted MACs per map, against 12.73G for RadioUNet and 52.71G for PMNet on 256×256 grids, and 102.289 s for MATLAB ray tracing on 513×513. The paper labels these timings orientative and not directly comparable.

Trust the LOS number first, then check the NLOS share

The most decision-relevant split in this paper is not the headline 1.93 dB but the gap between 1.74 dB in LOS and 3.54 dB in NLOS, combined with the fact that only 7.4% of evaluated receivers are NLOS. A reader comparing this model against alternatives should first ask what fraction of their own deployment's receivers fall in NLOS, because the element-weighted overall figure is dominated by the easier LOS case.

The second split is by altitude regime. Error falls monotonically from 2.14 dB in the low regime to 1.71 dB in the high regime, so the model's accuracy claim is strongest for high-altitude UxNBs and weakest near the 10 m end of the tested range. The third is topology: dense urban NLOS at 4.13 dB is more than twice the suburban LOS figure of 1.60 dB.

Two boundaries qualify these numbers. Accuracy assumes the visibility mask is supplied, and generating that mask is excluded from the reported runtime. The dataset is ray-traced simulation, not measurement, so the 1.93 dB describes agreement with ray tracing rather than with field data. Within those bounds, the paper's own priority is clear: the calibrated LOS branch is the stable part of the result, and the NLOS branch is where a neural approach already beats it.

Sources used in this answer