{"id":"9a6348bb-be5c-47bf-962b-d895428cbe4d","arxiv_id":"2501.00056","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A graph neural network trained on CCTV traffic flows, path signatures, and weather predicts hourly NO2 concentrations across London, but validation is limited and no baseline is reported.","lead":"Traffic counts pulled from London's CCTV cameras, together with weather and location data, are used by a graph neural network to estimate hourly NO2 pollution levels across the city. The idea is to give cities a low-cost pollution monitoring option using cameras they already have, but the evidence in this paper is not yet strong enough to support it.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central claim unverified: no ablation shows CCTV traffic features contribute to NO2 predictions beyond spatial/environmental covariates.","rationale":"The reader's rationale explicitly mentions the missing ablation and held-out test, but the stated weakest_assumption focuses on the graph adjacency structure. In my read, the graph structure is secondary: even a perfect adjacency would not rescue the claim if the predictions do not depend on camera data. The missing no-CCTV baseline is the single most load-bearing gap, because it directly tests whether CCTV features are necessary. The paper has real strengths: a large processed dataset, a plausible pipeline from video to signatures, and multiple architectural baselines. However, none of those establish that the camera signal matters. The proposed check is feasible with the existing model and data: train a small set of variants and compare on a proper holdout. If the ablation shows degradation, the central claim would gain support; if not, the 'pseudo-sensor' framing is misleading. Given the current manuscript, I agree with rejection, though a conditional acceptance could be justified if a supplementary ablation were added.","tokens_in":37250,"tokens_out":2798,"duration_ms":33112,"concrete_test":"Retrain the best-performing model (Model 19 in Table S5) under three conditions with identical graph, hyperparameters, and environmental/spatial features: (a) full camera-derived features; (b) camera-derived features removed; (c) camera-derived features randomly permuted across camera nodes per hour to break the traffic-NO2 link while preserving marginal distributions. Evaluate on a held-out set of hours not used in training, and ideally at held-out sensor locations. If MSLE/MAE/R2 in conditions (b) or (c) are not substantially worse than condition (a), the CCTV-derived features contribute nothing and the central claim fails.","verdict_should_be":"REJECT","load_bearing_attack":"The paper's central claim is that CCTV-derived traffic flows act as pseudo-NO2 sensors, generating reliable NO2 estimates at locations without sensors. For this to hold, the camera features must add predictive signal beyond what can be obtained from environmental and spatial covariates alone. The manuscript never demonstrates this. Table S5 reports 20 models, all of which include traffic-flow features (with or without signatures); there is no baseline that removes or permutes the CCTV-derived inputs while keeping the same graph, weather, and locational features. Because NO2 is strongly spatially autocorrelated, and because Section 2.6 deliberately uses a fully connected sensor graph and the model also receives weather and land-use data, the reported validation metrics (MSLE 0.0285-0.0526, R2 0.21-0.82) could be achieved by spatial interpolation of NO2 plus diurnal/seasonal weather patterns, with no actual contribution from the 133 million frames. The absence of a held-out temporal or spatial split further weakens the inference: the validation set in Table S5 draws from the same sensor locations and hours used for training, so the model may memorize sensor-specific NO2 fields rather than learn a camera-to-NO2 mapping. Without an ablation, the central claim is unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a pipeline that extracts traffic modal flows from London CCTV footage, summarizes hourly streams with path signatures, and uses graph neural networks to predict hourly NO2 concentrations at 144 sensor locations. The authors claim that CCTV cameras act as pseudo-NO2 sensors, enabling city-wide pollutant maps and policy insights such as temporal lags of traffic effects of up to 6 hours. The empirical section reports validation metrics for 20 model configurations, along with spatial regression and Granger causality analyses.","tokens_in":37500,"tokens_out":9507,"duration_ms":92930,"significance":"The problem is important, and assembling a large CCTV dataset with per-mode traffic counts is non-trivial. The combination of path signatures and graph-to-graph learning is a reasonable technical approach. However, the central claim is not established by the reported experiments. No ablation removes the CCTV-derived features, no independent test set or temporal/spatial holdout is used, and the reported validation metrics are therefore not sufficient to separate the contribution of camera data from spatial interpolation and environmental covariates. The paper also contains internal inconsistencies in data counts and in the Granger lag results. If the missing experiments were performed and the claims held, the work would be a useful contribution to urban air-quality monitoring.","major_comments":[{"comment":"The paper never tests a model that excludes the CCTV traffic inputs: all 20 models in Table S5 include traffic flow features, with or without signatures. Because NO2 is strongly spatially autocorrelated and the model also receives weather, land use, and a fully connected NO2 sensor graph, the reported validation performance may be achievable without any camera data. The authors must add a baseline that removes or permutes the CCTV-derived features while keeping everything else identical, and report the resulting MSLE and R2. This is necessary to support the paper's core claim that CCTV cameras act as pseudo-NO2 sensors.","section":"Section 2.6, Table S5"},{"comment":"All metrics are reported on the validation set, and the models are selected by early stopping on validation loss. There is no description of how the validation split is formed (random, temporal, or spatial), no independent test set, and no held-out period or held-out sensor locations. A random split that mixes hours from the same sensors in training and validation can yield optimistic R2 values due to spatial and temporal autocorrelation. The authors should report results on a held-out temporal period (for example, a set of days) and on a held-out spatial subset of NO2 sensors to support city-wide generalization claims.","section":"Section 4, Table S5"},{"comment":"Data counts are contradictory. Section 2.1 states 133,132,866 frames from 907 locations over 112 hours; the Methods (Materials) state 65,493,858 frames from 892 locations over 56 hours, and Section 2.2 says n=906. The abstract repeats the '133 million frames' figure. The paper must reconcile these numbers, as the dataset description is the basis for the headline claim and for reproducing the experiments.","section":"Section 2.1 vs Methods"},{"comment":"The abstract and text claim Granger-causal lags 'of up to 6 hours', but Table 3 reports only lags 1 through 5, with no lag-6 column. Many reported lags are significant in some tests but not others (for example, Car count at Hackney 1.01301002, Lag 5: alpha p=0.1267 and delta p=0.1267, but beta p<0.0001). The analysis is restricted to a handful of camera locations, yet the policy implications in Section 3 are city-wide. Either report the 6-hour lag results or correct the claim, and discuss the inconsistency of significance across the four tests.","section":"Section 2.4, Table 3"},{"comment":"The choice of a fully connected NO2 graph and a KNN (K=10) traffic graph is asserted with a physical rationale ('air can diffuse freely') but not empirically validated. In particular, the fully connected target graph may enable the model to interpolate NO2 directly from neighboring sensors, making the traffic input redundant. The authors should test alternative graph constructions (for example, distance-based or thresholded graphs for the NO2 sensors) and compare performance, ideally as part of the ablation described in the first major comment.","section":"Section 2.6"}],"minor_comments":[{"comment":"The camera count is given as n=906 in Section 2.2, but Section 2.1 and the abstract refer to 907 cameras; please correct the inconsistency.","section":"Section 2.2"},{"comment":"Panels B and D have identical captions ('scatter plot for the actual and predicted data for all sensor locations and all dates'); clarify which panel corresponds to training and which to validation, or indicate the actual distinction.","section":"Figure 5"},{"comment":"Several p-values are reported as 0.0000 (for example, 'car count' and 'rainfall'); these should be reported as p<0.0001 or with a more precise notation.","section":"Table 1"},{"comment":"The definition of the truncated signature is garbled; the notation 'NM' and 'hat S' is not standard, and the expression for S^n is not an iterated integral. Please provide a standard reference or a correct formulation.","section":"Equation (16)"},{"comment":"The labels 'Task 1' and 'Task 2' in Table S5 are reversed relative to the descriptions in Section 2.6 (Methods), where Task 1 is the NO2 surface and Task 2 is a single location. Please align the terminology.","section":"Table S5 vs Section 2.6"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses an important problem and the dataset assembly is substantial, but the current evidence does not support the abstract's claims. The missing ablation and the lack of a proper holdout validation are load-bearing and should be addressed before further consideration. The inconsistent frame counts and Granger lag statements also need correction. I recommend requesting a major revision rather than outright rejection because the missing experiments are well within the scope of a revision, and the authors may be able to demonstrate the value of the CCTV features with additional baselines."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper has a good idea and a load-bearing gap. The pitch — repurpose existing CCTV cameras as pseudo-NO2 sensors by converting traffic video into modal flows, summarizing them with path signatures, and feeding a graph network that outputs NO2 at unmonitored locations — is genuinely new as an applied pipeline, and the scale is real: hundreds of cameras and tens of millions of frames. The policy framing is sensible, and the limitations section is honest about camera placement bias and the large compute cost. The descriptive statistics, such as the hourly spatial regressions and the NO2 clock, are useful on their own as a summary of London's traffic-NO2 relationships. Those parts deserve credit.\n\nThe problem is that the central claim is unverified. Every model in Table S5 includes the CCTV traffic features, so there is no ablation showing what the camera data contributes. Given how strongly NO2 is spatially autocorrelated, a fully connected sensor graph, weather variables, and location features could plausibly produce the reported validation R2 (up to 0.82) by spatial interpolation plus diurnal and seasonal patterns, with no real signal from the video. The stress-test concern lands.\n\nThe validation is also weak in a standard way that matters here: metrics are reported on the validation set with early stopping, and validation draws from the same sensor locations and hours as training. There is no held-out temporal or spatial split, so sensor-specific memorization is not ruled out.\n\nTwo concrete red flags. The abstract says 133,132,866 frames from 112 hours at 907 locations; the methods say 65,493,858 frames from 56 hours at 892 locations. That factor-of-two discrepancy needs reconciling. And the abstract and discussion claim Granger lags up to 6 hours, but the supplementary table only tests lags 1–5, with significance mixed across the four tests. The headline lag claim is not backed by the table as presented. No code is released, so these things are hard to check independently.\n\nI agree with the reader's REJECT for the current version — the evidence does not support the headline. But the flaws are fixable, not fatal to the idea. A proper ablation that removes the CCTV features, a held-out time period, and reconciled data counts would go a long way; if the camera features survive that test, this is a solid applied contribution to urban sensing. My recommendation: send it to peer review, not desk reject it, and insist on exactly those additions.","headline":"Good idea, unverified claim: no ablation shows the CCTV features beat weather-plus-interpolation, validation is in-sample, and the data counts disagree — but the pipeline deserves a serious referee.","tokens_in":37995,"tokens_out":6412,"would_cite":true,"duration_ms":63032,"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":"CCTV cameras can act as citywide NO2 sensors","keywords":["air quality monitoring","NO2 prediction","CCTV traffic analysis","graph neural networks","path signature","city-scale sensing","urban policy","London"],"falsifier":"Hold out a random subset of NO2 sensors from training and run portable reference monitors at those locations and at several sensor-free camera sites; if the graph model's predictions at those points are no better than a spatial interpolation baseline, the CCTV signal is not carrying the prediction. A simpler lab check is to retrain the model with the traffic-flow channels replaced by shuffled noise and compare validation R2: if it barely drops, the camera data contributes little.","tokens_in":37026,"feed_emoji":"🚦","tokens_out":6611,"duration_ms":69123,"temperature":0.7,"pith_summary":"The paper tries to establish that existing CCTV cameras, originally installed for traffic monitoring, can be repurposed as pseudo-NO2 sensors. It uses 133 million frames from 907 London cameras to extract road-user flows, summarizes each hourly stream with a path signature, and feeds the result through a graph neural network to predict hourly NO2 concentrations across the city, including at locations without any pollution sensor. If this works, cities with dense camera networks but sparse air-quality monitors could get street-level pollution maps without buying new hardware. The paper also claims that different traffic modes affect NO2 on different time lags, up to six hours, and argues that this undercuts policies that simply push truck traffic into the night.","feed_headline":"Traffic-camera video predicts citywide NO2 pollution","feed_subtitle":"A graph model trained on 133 million London camera frames fills gaps where NO2 sensors are scarce.","key_machinery":"The load-bearing object is the path signature, a summary of a multidimensional stream that is invariant to reparameterization; here it is computed to depth N=3 from a 13-channel tensor holding modal flows and stationary statuses across 11 file increments per hour. The signature compresses the order of road-user events in a camera scene into a fixed-size feature while retaining the information needed for prediction. That feature is combined with environmental and spatial variables and passed through a graph-to-graph neural network with two graphs: a K-nearest-neighbour traffic graph (K=10) for the cameras, chosen to mimic the road network, and a fully connected graph for the NO2 sensors, chosen because air can diffuse freely. The model learns to generate NO2 values on the sensor graph from signals on the camera graph, which is what makes prediction at unsensed locations possible.","core_discovery":"The central claim is that ground-level NO2 can be inferred from camera-derived traffic dynamics through a graph-to-graph mapping: input features at 907 camera nodes (per-mode traffic counts, stationary statuses, congestion, plus weather and locational factors) are transformed into an NO2 field at 144 sensor locations, with no requirement that camera and sensor positions coincide. The strongest reported city-wide model, a Conv1D multi-branch graph network with signature inputs of depth three, reaches an MSLE of 0.0285, an MAE of 0.5197, and an R2 of 0.82 on validation data. The authors interpret the learned temporal lags, such as stationary trucks showing a one-hour Granger-causal effect and moving trucks a five-hour effect, as evidence that NO2 dynamics are tied to the chemical lifetime of NO2 and to the timing of traffic. The practical conclusion is that a hardware-free, CCTV-based sensing layer is feasible for urban NO2 monitoring and can inform adaptive policy.","pith_inferences":["The validation is performed only at the 144 sensor locations, so a decisive field test would place portable NO2 monitors in camera-rich, sensor-poor blocks and compare model output against them; if accuracy collapses, the model is mostly interpolating sensor readings.","Because the NO2 graph is fully connected, the network could in principle learn a purely spatial smooth of the sensor targets; shuffling or zeroing the traffic features while retraining would reveal how much of the prediction actually comes from the CCTV signal.","The same architecture might transfer to PM2.5, but its different atmospheric lifetime means the lag structure and graph assumptions would need to be re-estimated rather than reused.","The policy criticism based on Granger causality is predictive rather than causal; an intervention study, such as a real change in lorry routing times, would be needed to confirm that nighttime truck bans worsen morning NO2."],"forward_implications":["Cities with extensive CCTV coverage but few air-quality monitors could generate hourly NO2 maps with no new sensing hardware.","The reported Granger lags imply that moving heavy truck traffic to night hours will not necessarily reduce morning peak exposure; policy should weigh total emissions across all hours.","Because camera and sensor locations do not have to align, the graph-to-graph approach can in principle be transferred to other cities and to other pollutants.","With inference under 0.1 seconds per hour, the prediction side is fast enough for near-real-time monitoring if the video-processing front end can be scaled."],"supporting_citations":[{"why":"Supplies the Transport for London CCTV feeds that provide the 133 million frames from 907 camera locations used as the input data.","marker":"[33]"},{"why":"Provides the rough-path theory and path signature used to summarise the hourly traffic streams in a reparameterization-invariant form.","marker":"[35, 36]"},{"why":"Supplies the graph convolutional network formalism that the predictive graph-to-graph models extend and compare against.","marker":"[45]"},{"why":"Provides the graph attention mechanism used in one of the model architectures for focusing on relevant neighbouring nodes.","marker":"[46]"},{"why":"The YOLOv5 detector used to identify road users in the camera frames.","marker":"[54]"},{"why":"The DeepSort tracker used to assign unique identifiers to road users across sequential frames.","marker":"[55]"},{"why":"The TopView method that projects detected road users into a bird's eye view without calibrated camera parameters, enabling geo-referenced tracking.","marker":"[58]"},{"why":"Anchors the interpretation of temporal lags by documenting the variable lifetime of NOx in urban air.","marker":"[20]"}],"fun_headline_variants":["City CCTV traffic data predicts NO2 pollution","133M camera frames transform into NO2 maps","Traffic cams forecast NO2 levels for policy","London CCTV video turns into NO2 sensor grid","Graph model links traffic flow to NO2 spikes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the K-nearest-neighbour traffic graph and the fully connected sensor graph capture the real physics of NO2 dispersion, and that traffic features at 907 cameras plus weather data are enough to determine NO2 at 144 sensor locations and everywhere in between.","fun_headline_variants_meta":{"raw":{"variants":["City CCTV traffic data predicts NO2 pollution","133M camera frames transform into NO2 maps","Traffic cams forecast NO2 levels for policy","London CCTV video turns into NO2 sensor grid","Graph model links traffic flow to NO2 spikes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000211,"raw_usage":{"total_tokens":1435,"prompt_tokens":990,"completion_tokens":445,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":606,"completion_tokens_details":{"reasoning_tokens":374}},"tokens_in":606,"tokens_out":445,"duration_ms":5409,"temperature":1.0,"reasoning_tokens":374,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T23:30:33.408039+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Hold out a random subset of NO2 sensors from training and run portable reference monitors at those locations and at several sensor-free camera sites; if the graph model's predictions at those points are no better than a spatial interpolation baseline, the CCTV signal is not carrying the prediction. A simpler lab check is to retrain the model with the traffic-flow channels replaced by shuffled noise and compare validation R2: if it barely drops, the camera data contributes little.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Transport for London CCTV feeds that provide the 133 million frames from 907 camera locations used as the input data."},{"cited_title":"Comput Soc Netw 6, 11 (2019)","cited_arxiv_id":null,"evidence_quote":"Supplies the graph convolutional network formalism that the predictive graph-to-graph models extend and compare against."},{"cited_title":"Accessed: 2023-02-12 (2018)","cited_arxiv_id":null,"evidence_quote":"Provides the graph attention mechanism used in one of the model architectures for focusing on relevant neighbouring nodes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The YOLOv5 detector used to identify road users in the camera frames."},{"cited_title":"In: 2016 IEEE International Conference on Image Processing (ICIP), pp","cited_arxiv_id":null,"evidence_quote":"The DeepSort tracker used to assign unique identifiers to road users across sequential frames."},{"cited_title":"TopView: Vectorising road users in a bird's eye view from uncalibrated street-level imagery with deep learning","cited_arxiv_id":"2412.16229","evidence_quote":"The TopView method that projects detected road users into a bird's eye view without calibrated camera parameters, enabling geo-referenced tracking."},{"cited_title":"Science 366, 723–727 (2019)","cited_arxiv_id":null,"evidence_quote":"Anchors the interpretation of temporal lags by documenting the variable lifetime of NOx in urban air."}],"review_version":1}