{"id":"4ab1b5b5-4fe8-45a7-8d9a-efb9ea5849f7","arxiv_id":"2509.09606","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A multi-scale UNet with atrous spatial pyramid pooling predicts UAV mmWave pathloss maps from distance, LOS, and building mask inputs, reporting state-of-the-art accuracy on two datasets.","lead":"This paper trains a UNet-based neural network that predicts radio signal loss maps for drone-to-ground millimeter-wave communication using building maps, line-of-sight masks, and distances. It reports lower errors than several baselines on both a new ray-traced dataset and the public RadioMapSeer benchmark, and shows the model stays accurate across cities, drone altitudes, and noisy inputs.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Comparison with RadioUNet is confounded by unequal input channels; the claimed performance gain may reflect extra input information (log-distance map and LOS mask) rather than architectural superiority.","rationale":"The reader's weakest_assumption directly identifies the input-channel confound in the comparison with RadioUNet. This is the most load-bearing concern because the paper's headline result—'consistently outperforming all baselines'—depends on the comparison being an apples-to-apples test of architecture. Giving the proposed model log-distance and LOS mask inputs while RadioUNet lacks them violates that condition. The paper even notes the different input configurations, so it is a known design choice rather than an accidental omission. However, the paper may still contribute via the dataset and the vectorized LOS algorithm; the performance claim is what becomes conditional. The identical values in Table 9 for distance-noise experiments add a secondary red flag about evaluation rigor, but they do not directly attack the comparison. Therefore, the reader's CONDITIONAL verdict remains appropriate: the work is potentially valuable but the core claim needs a controlled comparison. No change to the verdict is warranted.","tokens_in":15597,"tokens_out":2255,"duration_ms":27211,"concrete_test":"Retrain the RadioUNet baseline with the same three input channels as the proposed model (log-distance map, LOS mask, building mask) on the same train/test split and data preprocessing, keeping its architecture fixed. If the RMSE gap versus the proposed model shrinks to within ~0.2 dB or reverses on either dataset, the central claim of architectural superiority is unsupported; if the gap persists, the input-channel confound is less severe.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim that the proposed multi-scale UNet 'outperforms' RadioUNet rests on a comparison that does not hold input information constant. In Table 6, the proposed model receives three input channels: a log-distance map, a binary LOS mask, and a building occupancy mask. The RadioUNet baseline receives either a transmitter-location mask plus building mask (2-channel) or these plus sparse pathloss measurements at 300 randomly chosen positions (3-channel). Log-distance and LOS mask are highly informative: log-distance encodes free-space attenuation, and the LOS mask directly marks shadow regions. RadioUNet must infer these implicitly from the building geometry and transmitter location, while the proposed model has them as explicit inputs. Thus the RMSE difference (3.15 vs 4.59 dB on the in-house dataset; 3.97 vs 4.23 dB on RadioMapSeer) could be driven primarily by the additional information, not by the multi-scale encoder or ASPP bottleneck. The paper acknowledges the input difference but does not control for it. A fair comparison would give RadioUNet (or any baseline) the same three input channels. Without this control, the central claim of architectural superiority is not established. Additionally, Table 9's distance-noise rows show identical RMSE/MAE/NMSE for 1%, 5%, and 10% noise (e.g., 3.18, 2.39, 0.000499 for all near-field rows), which suggests either a rounding artifact or that the distance-channel noise is not actually affecting the inputs as described; this further casts doubt on the robustness analysis and the reliability of the evaluation pipeline, though it is secondary to the main confound.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a UNet-based architecture for UAV-to-ground millimeter-wave pathloss prediction. The model consumes three 128×128 input channels (log-distance map, binary LOS mask, building occupancy mask) and uses multi-branch encoder blocks with parallel convolutions, an ASPP bottleneck, and transposed-convolution decoding. The authors generate an in-house 28 GHz ray-tracing dataset over five urban sites at three UAV altitudes and four transmitter positions each, using a CI-pathloss fallback for NLOS receivers, and they also evaluate on the RadioMapSeer benchmark at 5.9 GHz. They report lower RMSE/MAE/NMSE than classical ML, empirical, and RadioUNet baselines, along with cross-city generalization, altitude robustness, and noise-sensitivity results. A vectorized LOS estimation algorithm and a public code release are also presented.","tokens_in":16020,"tokens_out":10369,"duration_ms":111795,"significance":"If the results hold, the paper would make several useful contributions to radio-map prediction for UAV communications: a public codebase, a vectorized LOS mask computation that substantially accelerates preprocessing, a large synthetic ray-tracing dataset, and a systematic evaluation across cities, altitudes, and input perturbations. The multi-scale encoder plus ASPP design is plausible, and the release of reproducible training/evaluation scripts is a clear strength. However, the headline claim of outperforming RadioUNet is not yet established because the comparison does not hold input information constant, and the altitude and noise experiments have methodological issues. The underlying framework is sound enough that targeted additional experiments could validate the central claim.","major_comments":[{"comment":"The comparison with RadioUNet is confounded by input channels. The proposed model receives a log-distance map and an LOS mask, which directly encode free-space attenuation and shadow regions; RadioUNet receives only transmitter/building masks (2-channel) or those plus 300 sparse pathloss measurements (3-channel). The RMSE gap (3.15 vs 4.59 dB on the in-house set; 3.97 vs 4.23 dB on RadioMapSeer) may therefore reflect an information advantage rather than the proposed multi-scale/ASPP architecture. To support the 'outperforms all baselines' claim, the authors should train RadioUNet with the same three input channels, or ablate the proposed model to the RadioUNet input representation, so that the comparison isolates architecture.","section":"6.1.3, Table 6"},{"comment":"The altitude-robustness experiment does not test altitude invariance. The text states that 'we train three separate instances of our model for each altitude,' so each model only ever sees one transmitter altitude. The tight RMSE range (3.17–3.28 dB) shows that altitude-specific models achieve similar accuracy, but it does not show that a single model is insensitive to altitude. Please train one model on all altitudes and evaluate per altitude, and/or perform cross-altitude transfer (e.g., train at 25 m and test at 45 m), before claiming the architecture is altitude-invariant.","section":"6.3, Table 8"},{"comment":"The distance-noise rows are identical across perturbation levels: Dist. Noise (Near) gives RMSE=3.18, MAE=2.39, NMSE=0.000499 for 1%, 5%, and 10%; Dist. Noise (Far) gives the same RMSE=3.17, MAE=2.38, NMSE=0.000498 for all three levels. This is implausible if the described Gaussian perturbation is actually applied, and it suggests either that the noise is not reaching the model, that the same affected receivers are reused, or that the metrics are rounded beyond meaningful precision. The authors must re-run this experiment and report the actual variation across noise levels; otherwise the noise-robustness conclusion is not supported.","section":"6.4, Table 9"},{"comment":"The in-house test set uses only one held-out transmitter per city (15 test scenarios). A single transmitter location per environment is a thin basis for the main accuracy claim and for the comparison with RadioUNet, since results could depend on the particular choice of that transmitter. The cross-city protocol in Table 7 is stronger, but the main benchmark should either use multiple random held-out transmitters or repeated splits to quantify variance and demonstrate that the reported gains are stable.","section":"5.1, Table 6"}],"minor_comments":[{"comment":"For RadioMapSeer, LOS masks are generated using the proposed vectorized algorithm, but the paper does not explain how wall heights for these environments are obtained. Since the algorithm requires 3D walls, this is a reproducibility gap; please specify the data source and any assumptions.","section":"5.2"},{"comment":"The throughput units are inconsistent: 'one test sample' is described as one row of input features, but the computation uses 15×6×128×128 spatial samples. Please report throughput in patches/s or in pixels/s consistently, and align the MLP comparison with the CNN input format.","section":"6.5, Table 10"},{"comment":"Reference [15] appears garbled: 'M. A. I. F. Takawira, F. Tariq and A. Imran' for DeepREM does not match the actual author list. Please correct.","section":"References"},{"comment":"The term 'altitude-invariant' is too strong for the presented evidence; 'consistent performance across altitude-specific models' would be more accurate. This is related to the design issue in the second major comment.","section":"6.3"},{"comment":"The CI fallback uses n=3.0 and σ=6.8 dB from prior literature; this is a reasonable choice, but it means the labels are partially synthetic. A brief sentence acknowledging that this fallback may increase the apparent learnability of the maps would strengthen the discussion.","section":"2.2"}],"recommendation":"major_revision","confidential_remarks":"The most serious issue is the input-channel confound in the RadioUNet comparison. If a same-input control cannot close the gap, the contribution reduces to an incremental architecture evaluated on a synthetic dataset. I would also ask for raw per-seed results for Table 9; identical distance-noise rows across 1%, 5%, and 10% are a red flag for a coding or reporting error. The dataset release and the vectorized LOS algorithm are valuable, and I would be willing to review a revised version that addresses the confound and the altitude/noise methodology."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Plainly: this is a solid, clearly written engineering paper. What's new: a vectorized LOS computation algorithm that's genuinely useful for pre-processing, a new ray-traced UAV mmWave dataset at 28 GHz with three altitudes across five urban environments, and systematic evaluations of altitude sensitivity and input noise. The architecture itself—a UNet with multi-scale branches and an ASPP bottleneck—is a sensible combination of known components rather than a conceptual leap. The code and dataset are released, which is real value.\n\nThe main soft spot is the headline comparison to RadioUNet. The proposed model receives three input channels: log-distance, LOS mask, and building mask. RadioUNet gets either a transmitter mask plus building mask, or those plus sparse pathloss measurements. Log-distance and LOS mask directly encode free-space attenuation and shadowing, so the reported RMSE gap (3.15 vs 4.59 dB on the in-house set) could be driven by the extra information, not by the multi-scale encoder. The paper acknowledges the input difference but doesn't control for it. A fair test would give RadioUNet the same three channels, or remove the extra channels from the proposed model. The RadioMapSeer result narrows the gap (3.97 vs 4.23 dB), but the same confound applies.\n\nTwo other concerns: Table 9 shows exactly identical RMSE/MAE/NMSE values for 1%, 5%, and 10% distance noise. That looks like a bug or rounding artifact; it needs a fix or a clear explanation. And the in-house test set is small (one held-out transmitter per city), with no error bars or repeated runs, so the reported differences may not be stable. Recent baselines like PEFNet and RadioFormer are cited but not evaluated.\n\nNone of this sinks the paper. The dataset, the vectorized LOS algorithm, and the altitude/noise analyses are useful contributions on their own. But the load-bearing 'outperforms state-of-the-art' claim needs a controlled comparison before it can be trusted.\n\nWho's it for: researchers in radio map prediction and UAV channel modeling. It deserves serious peer review, with the confound and Table 9 addressed. I'd accept it for review with major revisions; I'd be hesitant to cite the head-to-head numbers until the input sets are matched, though I'd consider citing the dataset and method.","headline":"Useful engineering paper with a new dataset and fast LOS algorithm, but the RadioUNet comparison is confounded by unequal input channels.","tokens_in":16423,"tokens_out":3208,"would_cite":true,"duration_ms":34380,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A multi-scale UNet with an ASPP bottleneck predicts UAV-assisted mmWave pathloss maps more accurately than existing baselines, with RMSE of 3.15 dB on an in-house dataset and 3.97 dB on a public benchmark, while staying robust across cities","keywords":["UAV communications","mmWave propagation","pathloss prediction","deep learning","UNet","ASPP","LOS estimation","ray tracing"],"falsifier":"Retrain the strongest baseline (the 3-channel UNet) with the same three input channels as the proposed model—log-distance map, LOS mask, and building mask—on the identical train/test split. If its RMSE drops to 3.15 dB or below on the in-house dataset, or close to 3.97 dB on RadioMapSeer, then the architecture's claimed advantage is not isolated from input engineering.","tokens_in":15500,"feed_emoji":"📡","tokens_out":3750,"duration_ms":41315,"temperature":0.7,"pith_summary":"The paper proposes a deep learning architecture for predicting pathloss maps in UAV-assisted millimeter-wave networks. The model, built on UNet, combines multi-scale convolutional feature extraction with an atrous spatial pyramid pooling bottleneck to capture both local and global propagation patterns. Using three input channels—log-distance, line-of-sight mask, and building mask—the model is shown to outperform classical machine learning, empirical models, and a standard UNet baseline on both a new ray-traced dataset (28 GHz) and the public RadioMapSeer benchmark (5.9 GHz). The paper also claims strong cross-city generalization, altitude invariance, and resilience to distance noise, while introducing a vectorized LOS mask algorithm that speeds up data preprocessing.","feed_headline":"UNet predicts UAV mmWave pathloss to 3.15 dB RMSE","feed_subtitle":"New architecture beats baselines on two datasets, generalizes across cities and UAV altitudes.","key_machinery":"The central mechanism is a multi-scale feature extraction and fusion block: each encoder stage runs parallel convolutions with different kernel sizes (1x1, 3x3, 5x5, and 7x7 in the first stage), concatenates the outputs, and fuses them with a 1x1 convolution. The bottleneck uses an ASPP module with dilated convolutions (rates 1, 2, 4) plus a global average pooling branch to aggregate context across multiple receptive fields. A second component is a fully vectorized LOS mask computation algorithm that filters wall-facing geometry and performs 3D intersection tests using tensor broadcasting, enabling fast generation of training labels.","core_discovery":"The paper claims that its proposed UNet variant, which uses multi-branch convolutions (1x1, 3x3, 5x5, and a 7x7 in the first stage) fused by 1x1 convolutions and an ASPP bottleneck, achieves the best pathloss prediction accuracy among all compared methods. On the in-house ray-traced dataset of five urban environments, it reports RMSE of 3.15 dB, MAE of 2.37 dB, and NMSE of 0.00049; on RadioMapSeer it reports the lowest RMSE of 3.97 dB, MAE of 2.03 dB, and NMSE of 0.0011. The model also maintains RMSE between 3.16 and 3.53 dB when trained on four cities and tested on a held-out city, and stays between 3.17 and 3.28 dB across UAV altitudes of 25, 35, and 45 m. These results are attributed to t","pith_inferences":["The reported gains may partly stem from input information rather than architecture alone: the proposed model receives a log-distance map and a line-of-sight mask, while the primary baseline receives only transmitter/building masks or sparse measurements. Giving the baseline the same inputs could narrow the performance gap.","Altitude invariance may be a byproduct of using 3D distance and LOS masks as inputs, since these already encode transmitter height through geometry; the architecture may not need to learn altitude as a separate factor.","A testable extension is to apply the same architecture and inputs to other frequencies (e.g., sub-6 GHz) or to predict additional channel quantities like delay spread or angle of arrival, which would test whether the learned spatial features generalize beyond pathloss.","The vectorized LOS algorithm could be reused in other visibility or ray-tracing preprocessing pipelines, independent of the neural network, as a standalone tool."],"forward_implications":["If the model's claimed accuracy holds, UAV-assisted mmWave network planning could switch from computationally heavy ray tracing to sub-second deep-learning predictions, making real-time deployment feasible.","Cross-city generalization of roughly 3.2 to 3.5 dB RMSE suggests a single trained model might be reused across urban environments without per-city retraining, reducing data collection costs.","The tight RMSE range across UAV altitudes (3.17 to 3.28 dB) implies the model could support altitude-varying missions without needing altitude-specific models.","The vectorized LOS mask algorithm could accelerate generation of large synthetic datasets for other radio map learning tasks, lowering the barrier to training data creation."],"fun_headline_variants":["Multi-scale UNet predicts UAV pathloss with 3.15 dB RMSE","New UNet beats baselines in mmWave pathloss accuracy","Generalizable UNet for UAV pathloss across cities and altitudes","Faster LOS makes large-scale pathloss datasets possible","ASPP fusion sharpens UNet pathloss predictions to 3.15 dB"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The reported superiority over the UNet baseline assumes that the comparison isolates architecture quality, but the proposed model receives additional high-information inputs (log-distance map and LOS mask) that the baseline does not get; if the baseline were given the same channels, the accuracy gap might shrink or disappear.","fun_headline_variants_meta":{"raw":{"variants":["Multi-scale UNet predicts UAV pathloss with 3.15 dB RMSE","New UNet beats baselines in mmWave pathloss accuracy","Generalizable UNet for UAV pathloss across cities and altitudes","Faster LOS makes large-scale pathloss datasets possible","ASPP fusion sharpens UNet pathloss predictions to 3.15 dB"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000834,"raw_usage":{"total_tokens":3493,"prompt_tokens":775,"completion_tokens":2718,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":519,"completion_tokens_details":{"reasoning_tokens":2626}},"tokens_in":519,"tokens_out":2718,"duration_ms":24869,"temperature":1.0,"reasoning_tokens":2626,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T18:44:26.520494+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the strongest baseline (the 3-channel UNet) with the same three input channels as the proposed model—log-distance map, LOS mask, and building mask—on the identical train/test split. If its RMSE drops to 3.15 dB or below on the in-house dataset, or close to 3.97 dB on RadioMapSeer, then the architecture's claimed advantage is not isolated from input engineering.","supporting_citations":[],"review_version":1}