{"id":"98dffe69-14ed-4470-ab3c-82fdf2dbb59b","arxiv_id":"2501.02175","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"A 2.8 GHz CSI measurement campaign under artificial rain shows rain-correlated attenuation and delay-spread changes, and a ResNet-based classifier distinguishes no rain, moderate rain, and heavy rain from 20-second CSI windows.","lead":"This paper reports the first channel-state-information (CSI) measurements of rainfall attenuation at 2.8 GHz, using an artificial rain setup, and trains a neural network to classify rain intensity from 20-second CSI windows. If the result holds outside the lab, existing sub-6 GHz cellular links could serve as opportunistic rain sensors for ISAC networks.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Wet-antenna confound is the load-bearing weakness: exposed antennas in artificial rain make RSS/PDP changes attributable to water films, not atmospheric rainfall.","rationale":"The reader's weakest-assumption analysis identifies the wet-antenna confound as the load-bearing vulnerability, and I agree. The central claim requires that water on the antenna hardware contributes negligibly to the measured RSS/PDP changes; the paper provides no evidence for this, and its own introduction acknowledges wet radomes as a known source of attenuation overestimation. My independent reading of the measurement sections confirms that no wet/dry control, radome treatment, or sensitivity analysis is present. This is not an internal inconsistency—the signal processing and classification pipeline are coherent—but it is an external validity threat that directly affects both the quantitative attenuation values and the trained classifier's features. The classification results, while internally consistent and based on a temporal train/test split, cannot rescue the physical interpretation if the input features are dominated by antenna wetting. I therefore keep the reader's CONDITIONAL verdict unchanged: the paper is a plausible proof-of-concept but requires a wet-antenna control or a natural-rain validation before the attenuation and classification claims can be accepted as rainfall sensing. I did not find a separate, more load-bearing concern; the lack of released code/data and the same-session evaluation are secondary to the wetting issue.","tokens_in":14732,"tokens_out":2821,"duration_ms":31390,"concrete_test":"Repeat the experiment with the antennas sheltered by transparent covers so rain falls along the 7 m path but not on the hardware. Compare RSS and PDP across no-rain, moderate, and heavy rain with sheltered antennas against the unsheltered baseline. If the attenuation and PDP trends mostly disappear, the reported rain signature is antenna wetting; if they persist, the atmospheric-rain interpretation is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim—measurable rain-induced attenuation at 2.8 GHz and CSI features that classify rainfall—rests on attributing observed RSS drops (1.86 dB moderate, 3.28 dB heavy over a 7 m path), PDP decay changes (nPDP 1.52→1.49→1.41), and RMS delay spread changes (6.41→11.77→10.70 ns) to rain in the propagation channel. The experiment exposes both the horn antenna and the patch antennas directly to artificial rainfall, yet no control isolates antenna wetting. The authors themselves cite [33]–[35] in Section I, noting that water films on radomes cause additional attenuation and can overestimate rainfall intensity, but Sections II and III report no wet-antenna baseline, no radome treatment, no drying procedure after each run beyond 'the system is reset,' and no sensitivity analysis. Given the short 7 m path, the free-space specific attenuation at 2.8 GHz is orders of magnitude below the observed few-dB excess; water films on the antennas or near-field spray is a quantitatively plausible alternative explanation for both the RSS and PDP changes. Since the same PDP differences feed RainGaugeNet's input features, the classifier may be learning the wetting state of the testbed rather than rainfall in the air. This confound undermines the attenuation measurements, the fitted power-law trend (Table II), and the interpretation of classification accuracy as rain sensing.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents RainGaugeNet, a CSI-based rainfall classification system operating at 2.8 GHz. The authors build a controlled artificial-rain testbed with a USRP transmitter and receiver, collect RSS and CSI under no rain, moderate rain (5 mm), and heavy rain (20 mm) in LoS and NLoS configurations at low and high wind speeds, and report three physical observations: RSS mean and variance shift with rainfall; the PDP follows a power-law decay whose exponent decreases with rainfall; and RMS delay spread increases with rainfall onset. They then train a ResNet-based neural network on 40x20 PDP matrices (400 ns x 20 s) to classify rainfall intensity, achieving average accuracies above 90% in LoS and above 85% in NLoS. The central claims are that this is the first CSI-based demonstration of measurable rainfall attenuation at sub-6 GHz and that CSI features support accurate rainfall classification.","tokens_in":14870,"tokens_out":5656,"duration_ms":51056,"significance":"If the central claims survive experimental scrutiny, the paper would be a valuable data point for ISAC-based weather sensing. Its strengths are the explicit hardware description, the use of a calibrated rain gauge, the temporal separation of training and test data, and the comparison against RSS-based and PDP-based baselines, which make the classification comparison meaningful. The paper also honestly reports challenges and does not oversell per-class performance. However, the physical interpretation of the attenuation measurements is currently undermined by an unaddressed wet-antenna confound; if that is resolved (or bounded), the CSI-based classification result would be the most transferable contribution.","major_comments":[{"comment":"The measurement setup exposes both the horn transmit antenna and the two patch receive antennas directly to artificial rainfall, yet no control experiment isolates the wet-antenna contribution. The introduction itself cites [33]-[35] to note that water films on radomes introduce additional attenuation and can cause overestimation of rainfall intensity. Over the 7 m path, the reported RSS drops (1.86 dB moderate, 3.28 dB heavy) are far larger than the free-space rain attenuation predicted by the ITU-R model at 2.8 GHz, so antenna wetting or near-field spray is a quantitatively plausible alternative explanation for the attenuation, the PDP decay changes, and the RMS delay spread increase. I request either a dry-versus-wet antenna control (e.g., wetting the antennas while keeping the propagation path dry), a protective radome with known wet-antenna loss, or a sensitivity analysis that bounds the wet-antenna contribution; without one of these, the central claim that the observed attenuation is due to rain in the propagation channel is not established.","section":"Sections I, II-B"},{"comment":"The power-law decay factors in Table II (nPDP = 1.52, 1.49, 1.41) are presented as evidence that the decay factor decreases with rainfall intensity, but no confidence intervals or statistical test are reported. The RMSE values in the same table range up to 4.99 dB for the no-rain condition, which is of the same order as the differences among the fitted η0 values and among the mean RSS attenuation levels; the differences in nPDP are small relative to the fitting uncertainty. Please report per-realization distributions or bootstrap confidence intervals for nPDP and perform a significance test before asserting O3.","section":"Section II-D, Table II, Observation O3"},{"comment":"The headline claim of 'over 90% accuracy in LoS and over 85% in NLoS' is supported only as an average over wind conditions. Several per-condition results fall well below these thresholds: 80.14% for LoS low-wind heavy rain, 65.07% for NLoS low-wind no rain, and 58.10% for NLoS high-wind moderate rain. Because the robustness claim is central to the paper's ISAC positioning, the authors should report per-condition confidence intervals, discuss the failure modes, and qualify the abstract claim by stating that these are average accuracies.","section":"Section III-D, Table V"}],"minor_comments":[{"comment":"Please specify what 'the system is reset' entails in Section II-B; currently it is too vague to determine whether antenna surfaces were dried between runs.","section":"Section II-B"},{"comment":"The units of RSS and of the attenuation values in Fig. 4 and the text should be stated consistently (dBm for RSS, dB for attenuation).","section":"Section II-C"},{"comment":"The notation XPDP is used both as the random variable and as the normal-distributed model error; distinguish the distribution parameter σ from the RMSE notation σRMSE for clarity.","section":"Section II-D, Eq. (5)"},{"comment":"In the description of RainGaugeNet-single, the sentence stating that it 'shares the same architecture as RainGaugeNet but uses data only from the first time snapshot' should clarify whether the remaining 19 time slots are zero-filled, as suggested by Fig. 11(a), since this affects the interpretation of its degenerate behavior.","section":"Section III-C"},{"comment":"Reference [21] contains a typo in its title ('arbrelation' should be 'a-b relation'); elsewhere, 'poisson' should be capitalized in Section I.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper would benefit from a reviewer with experience in wet-antenna attenuation in microwave backhaul links. The 'first CSI-based' claim appears plausible given the references, but the authors should double-check against more recent work. The main risk is that the classification results, while impressive, may not transfer to field deployments where antenna wetting is uncontrolled; the editor may want to emphasize this in the decision letter."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, the real news: this is the first time anyone has used CSI rather than RSS for rainfall classification at sub-6 GHz, and the measurement campaign at 2.8 GHz under controlled artificial rain is a legitimate new empirical dataset. The paper is clearly written, the temporal train/test split is a good practice, and the comparisons against RSS-Net, CNN, and single-snapshot baselines are fair. If the goal is to show that a CSI-based classifier can distinguish no-rain/moderate/heavy in a testbed, RainGaugeNet does that and the numbers are internally consistent.\n\nThe soft spot is exactly where the reader's report puts it. The antennas are exposed directly to the artificial rain, and the paper itself cites [33]–[35] on wet radome attenuation but reports no wet-antenna control, no radome treatment, and no drying procedure beyond 'the system is reset.' Over a 7 m path at 2.8 GHz, the measured 1.86 dB and 3.28 dB excess are far above ITU-R physical predictions, and water films on the horn and patch antennas are a quantitatively plausible alternative explanation. This confound hits more than just the RSS numbers: the PDP and delay-spread changes feed the classifier, so the model may be learning the wetting state of the testbed rather than rain in the air. The authors should either correct for wet antenna or re-frame the claim from 'rainfall attenuation' to 'rainfall-induced end-to-end effects.'\n\nThe power-law decay trend (O3) is also shakier than the text suggests. nPDP moves from 1.52 to 1.49 to 1.41 with RMSEs as high as 4.99 dB, and no confidence intervals are given. That is a fragile fit to hang a physical law on. And no code or data are released, so the dataset is not independently usable yet.\n\nThat said, the central idea is not bad. For a subfield that is actively looking for opportunistic rainfall sensing in 5G networks, a CSI-based proof-of-concept is worth having, even if the current version overclaims the attenuation physics. I'd send it to peer review with a request for major revision: quantify the wet-antenna contribution, add a second site or natural rain validation, and tone down O3. As it stands, it belongs in the 'promising but not yet trustworthy' pile.","headline":"A genuinely new CSI-based rainfall sensing dataset and classifier, but the unaddressed wet-antenna confound makes the headline attenuation numbers unreliable.","tokens_in":15566,"tokens_out":2488,"would_cite":false,"duration_ms":23408,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Sub-6 GHz radio channels carry a measurable rainfall signature at 2.8 GHz.","keywords":["rainfall attenuation","sub-6 GHz","channel state information","power delay profile","rainfall classification","integrated sensing and communications","2.8 GHz channel measurement","ResNet1D"],"falsifier":"Re-run the 2.8 GHz, 7 m measurement with the horn and patch antennas kept dry, for example heated or shielded, while artificial rain falls only between them; if the 1.86 dB and 3.28 dB attenuation values and the fitted decay-factor shifts largely disappear, the rain-in-the-air attribution is disproved.","tokens_in":14373,"feed_emoji":"🌧️","tokens_out":8192,"duration_ms":73069,"temperature":0.7,"pith_summary":"This paper tries to establish that rain is not invisible to sub-6 GHz radio links: using a 2.8 GHz channel-sounding setup, it reports measurable attenuation, a power-law decay in the power delay profile, and a rainfall-dependent delay spread, and then shows a neural network can classify no-rain, moderate-rain, and heavy-rain conditions from 20 seconds of channel state information (CSI). The authors' claim is that CSI, which is already collected in modern cellular links, carries enough rainfall information to act as a distributed rain gauge for integrated sensing and communications. If true, dense sub-6 GHz networks could monitor rainfall without new spectrum, dedicated weather hardware, or waiting for millimeter-wave infrastructure. The paper positions RainGaugeNet as the first CSI-based rainfall classifier, with over 90% accuracy in line-of-sight and over 85% in non-line-of-sight conditions.","feed_headline":"Rain leaves a measurable signature in 2.8 GHz radio data","feed_subtitle":"Twenty seconds of channel-state readings let a network tell no rain from moderate and heavy rain in over 90% of line-of-sight tests.","key_machinery":"The load-bearing object is the power delay profile (PDP), a 40-tap, 400-ns snapshot of how received power is distributed across propagation delays, obtained by inverse discrete Fourier transform of the channel frequency response. RainGaugeNet feeds a 40 x 20 matrix of PDP taps across 20 seconds into a dual-path ResNet1D network: one path extracts spatial multipath structure per snapshot, the other extracts temporal correlations across snapshots, and the final fully connected layers output the three-class rainfall label.","core_discovery":"At 2.8 GHz over a 7 m link, rainfall attenuation is real but small and nonlinear: average received signal strength drops by 1.86 dB under moderate rain and 3.28 dB under heavy rain, whereas the ITU-R high-frequency power-law model predicts 0.0047 dB and 1.33 dB. The measured power delay profile follows a power-law decay model, with the fitted decay factor falling from 1.52 with no rain to 1.49 with moderate rain and 1.41 with heavy rain, and the RMS delay spread rises from 6.41 ns to 11.77 ns in moderate rain before easing to 10.70 ns in heavy rain. From these CSI-derived features, RainGaugeNet classifies three rainfall intensities with average accuracy above 90% in line-of-sight settings and above 85% in non-line-of-sight settings, using 20 consecutive 1-second CSI snapshots and outperforming RSS-only, single-snapshot, and plain CNN baselines.","pith_inferences":["A controlled experiment with dry antennas or wet-antenna compensation would separate airborne rain attenuation from water films on radomes; until that control exists, the reported 1.86 dB and 3.28 dB values should be read as upper bounds on true atmospheric attenuation.","The same CSI features could be extended from three-class classification to continuous rain-rate regression, using a rain gauge as ground truth over longer outdoor links.","The non-monotonic delay spread, peaking at moderate rain, suggests rain-onset detection may be more reliable than intensity estimation, since onset and heavy rain produce different multipath signatures.","Natural rainfall differs from artificial rainfall in drop-size distribution and wind, so validating on outdoor natural rain events is a direct next test of the classifier's generality."],"forward_implications":["If the effect is real, every sub-6 GHz base station that already records CSI could in principle act as a rain gauge without additional transmit power or dedicated hardware.","Rainfall onset increases RMS delay spread, meaning rain is not only a loss effect; it changes the multipath structure that channel equalizers and beamformers must track.","The reported decay-factor trend, with smaller fitted decay factors under heavier rain, gives a physical model-based feature for rainfall intensity beyond raw RSS.","Twenty seconds of CSI is short enough for near-real-time network-level rainfall mapping in integrated sensing and communications systems."],"supporting_citations":[{"why":"Supplies the ITU-R baseline rain attenuation predictions (0.0047 dB and 1.33 dB) against which the measured values are compared.","marker":"[16]"},{"why":"Prior evidence at 1.8 GHz that mobile phones detect RSS variations caused by rain, motivating the sub-6 GHz feasibility claim.","marker":"[22]"},{"why":"The RSS-based LTE/4G rainfall classifier that serves as the accuracy baseline and informs the observed RSS variance trend.","marker":"[23]"},{"why":"S-band rain attenuation measurements in Greece that confirm significant rain-induced attenuation at sub-6 GHz frequencies.","marker":"[3]"},{"why":"Statistical modeling of climatic influence on a 5 GHz microwave link, supporting the existence of measurable sub-6 GHz rain effects.","marker":"[26]"},{"why":"Refined power-law relationships for sub-6 GHz rain attenuation, which the paper's PDP decay model builds on.","marker":"[24]"},{"why":"Provides the stochastic tapped-delay-line channel model that frames the PDP tap analysis.","marker":"[40]"},{"why":"Supplies the power-law decay model for PDP fitting used to extract the decay factor observations.","marker":"[41]"},{"why":"The multipath identification algorithm used to extract multipath components and compute RMS delay spread.","marker":"[45]"},{"why":"Identifies the wet-antenna attenuation hazard that the experiment does not control, marking the main threat to the rain-attenuation attribution.","marker":"[33]"}],"fun_headline_variants":["Sub-6 GHz rain attenuation measured: CSI classifies intensity in 20s","RainGaugeNet: 20-second CSI data identifies rain intensity at 2.8 GHz","Rain alters 2.8 GHz signals: neural net classifies rain from CSI data","First CSI-based rain gauge at 2.8 GHz achieves 90%+ accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper attributes the measured received-power drops and power-delay-profile changes to attenuation by raindrops in the 7 m propagation path, and the load-bearing premise is that water films on the antennas and radomes do not contribute most of that signal; the paper identifies wet-antenna attenuation as a known hazard but reports no wet-antenna control.","fun_headline_variants_meta":{"raw":{"variants":["Sub-6 GHz rain attenuation measured: CSI classifies intensity in 20s","RainGaugeNet: 20-second CSI data identifies rain intensity at 2.8 GHz","Rain alters 2.8 GHz signals: neural net classifies rain from CSI data","First CSI-based rain gauge at 2.8 GHz achieves 90%+ accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000558,"raw_usage":{"total_tokens":2664,"prompt_tokens":964,"completion_tokens":1700,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":580,"completion_tokens_details":{"reasoning_tokens":1608}},"tokens_in":580,"tokens_out":1700,"duration_ms":13486,"temperature":1.0,"reasoning_tokens":1608,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:13:59.085933+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the 2.8 GHz, 7 m measurement with the horn and patch antennas kept dry, for example heated or shielded, while artificial rain falls only between them; if the 1.86 dB and 3.28 dB attenuation values and the fitted decay-factor shifts largely disappear, the rain-in-the-air attribution is disproved.","supporting_citations":[{"cited_title":"Propagation data and prediction methods required for the design of terrestrial line-of-sight systems,","cited_arxiv_id":null,"evidence_quote":"Supplies the ITU-R baseline rain attenuation predictions (0.0047 dB and 1.33 dB) against which the measured values are compared."},{"cited_title":"The impact of weather condition on radio-based distance estimation: A case study in gsm networks with mobile measurements,","cited_arxiv_id":null,"evidence_quote":"Prior evidence at 1.8 GHz that mobile phones detect RSS variations caused by rain, motivating the sub-6 GHz feasibility claim."},{"cited_title":"Rainfall estimation based on the intensity of the received signal in a lte/4g mobile terminal by using a probabilistic neural network,","cited_arxiv_id":null,"evidence_quote":"The RSS-based LTE/4G rainfall classifier that serves as the accuracy baseline and informs the observed RSS variance trend."},{"cited_title":"A rain estimation model based on microwave signal attenuation measurements in the city of ioannina, greece,","cited_arxiv_id":null,"evidence_quote":"S-band rain attenuation measurements in Greece that confirm significant rain-induced attenuation at sub-6 GHz frequencies."},{"cited_title":"Statistical modeling of the climatic influence on a 5 ghz microwave link: A tropical weather case empirical study,","cited_arxiv_id":null,"evidence_quote":"Statistical modeling of climatic influence on a 5 GHz microwave link, supporting the existence of measurable sub-6 GHz rain effects."},{"cited_title":"Harnessing the radio frequency power level of cellular terminals for weather parameter sensing,","cited_arxiv_id":null,"evidence_quote":"Refined power-law relationships for sub-6 GHz rain attenuation, which the paper's PDP decay model builds on."},{"cited_title":"Impulse response modeling of indoor radio propagation channels,","cited_arxiv_id":null,"evidence_quote":"Provides the stochastic tapped-delay-line channel model that frames the PDP tap analysis."},{"cited_title":"The ultra-wide bandwidth indoor channel: from statistical model to simulations,","cited_arxiv_id":null,"evidence_quote":"Supplies the power-law decay model for PDP fitting used to extract the decay factor observations."},{"cited_title":"Time-domain channel measurements and small-scale fad- ing characterization for ris-assisted wireless communication systems,","cited_arxiv_id":null,"evidence_quote":"The multipath identification algorithm used to extract multipath components and compute RMS delay spread."},{"cited_title":"Modeling of wet antenna attenuation for precipitation estimation from microwave links,","cited_arxiv_id":null,"evidence_quote":"Identifies the wet-antenna attenuation hazard that the experiment does not control, marking the main threat to the rain-attenuation attribution."}],"review_version":1}