Transforming CCTV cameras into NO$_2$ sensors at city scale for adaptive policymaking

TL;DR

This work addresses the lack of real-time city-scale NO$_2$ monitoring by turning ubiquitous CCTV footage into pseudo-sensors of NO$_2$ concentrations. It introduces a graph-to-graph neural framework that fuses high-frequency traffic signals extracted from 907 London CCTV cameras with environmental factors to predict ground-level NO$_2$ across the city, leveraging a multi-level spatiotemporal representation and path signatures of traffic flows. Key findings show significant, sometimes lagged, relationships between traffic modes and NO$_2$ levels (lags up to 6 hours), and demonstrate that the Graph-to-Graph model can reliably reconstruct city-wide NO$_2$ surfaces with MSLE in the ~0.037–0.052 range and $R^2$ up to 0.94. The approach offers a cost-effective, scalable method for adaptive policymaking in cities lacking dense sensor networks, while highlighting policy implications (e.g., nocturnal heavy-vehicle patterns) and limitations related to data quality and computation.

Abstract

Air pollution in cities, especially NO\textsubscript{2}, is linked to numerous health problems, ranging from mortality to mental health challenges and attention deficits in children. While cities globally have initiated policies to curtail emissions, real-time monitoring remains challenging due to limited environmental sensors and their inconsistent distribution. This gap hinders the creation of adaptive urban policies that respond to the sequence of events and daily activities affecting pollution in cities. Here, we demonstrate how city CCTV cameras can act as a pseudo-NO\textsubscript{2} sensors. Using a predictive graph deep model, we utilised traffic flow from London's cameras in addition to environmental and spatial factors, generating NO\textsubscript{2} predictions from over 133 million frames. Our analysis of London's mobility patterns unveiled critical spatiotemporal connections, showing how specific traffic patterns affect NO\textsubscript{2} levels, sometimes with temporal lags of up to 6 hours. For instance, if trucks only drive at night, their effects on NO\textsubscript{2} levels are most likely to be seen in the morning when people commute. These findings cast doubt on the efficacy of some of the urban policies currently being implemented to reduce pollution. By leveraging existing camera infrastructure and our introduced methods, city planners and policymakers could cost-effectively monitor and mitigate the impact of NO\textsubscript{2} and other pollutants.

Figures (7)

• Figure 1: Multi-level representation of all data sources. (A) The six layers of factors presented in this research. (B) The spatial representation as graph of knowledge of the camera (red nodes) and NO2 (blue nodes) inputs. The locations of the cameras and NO2 sensors do not need to align.(C) The multi-level representation of the studied data modalities shows several spatial and temporal resolutions in which different data modalities are aligned to conduct this research.
• Figure 2: Capturing the order of events from micro-level to city-scale (A) Shows 1) sequential frames of a given video file at Piccadilly circus in London as an input of one given CCTV camera, 2) vector representation of road users in an estimated bird’s eye view map with Google Maps to validate the geographical localisations of road users, and 3) temporal representation of road users within a given file based on the tracked system. (B) The relationship between hourly observed traffic flow data and the unseen temporal intervals among various file increments representing the stream of paths $X \in \mathbb{R}^{(nXtXc)}$, given that n is the number of cameras $(n=907)$, t is the number of file increments that make an hour of traffic modes $(t=11)$ and $c$ is the number of channels for traffic modal flows and their stationary status $(c=13)$(C) The tensor representation of the generated paths with all its channel and their unique computed signatures ($Sig^N,N=3$) that summarise the paths of varied traffic modal flows and their actions in a given scene.
• Figure 3: The Spatial patterns of NO2 and traffic at city-scale. (A) The association between the studied variables relies on Pearson’s correlation. (B) Hot spot analysis using significant Moran’s I z-value (P < 0.05) to highlight the outliers of NO2 across different hours of the day (the rest of the 24 hours are presented in supplementary). (C) Hot spot analysis using significant Moran’s I z-value to highlight the outliers of total flow across different hours of the day. (D) Statistically significant results ($p<0.05$,$r^2=0.4$, $spatial r^2=0.23$,and $df=88020$) of the spatial two-stage least-square model, variables are shown based on the sign and weight of their $\beta$ value.
• Figure 4: NO$_2$ Clock. It shows the hourly average levels of NO$_2$ and the factors influencing these levels, based on a spatial regression model for each hour. It features three concentric circles: the innermost represents the average NO$_2$ concentration per hour, the middle circle shows factors negatively correlated with NO$_2$, and the outermost highlights positively correlated factors. Each factor’s influence is quantified by a $\beta$ value, indicating its effect size relative to the hourly covariates, factors, and overall impact on NO$_2$. All $\beta$ values are standardised across all hours. For simplicity and clarity, the figure displays only four variables, although the full model considers a more extensive range of variables detailed in Table S1.
• Figure 5: Graph-to-Graph model to predict NO2 surface at a given time from camera feeds. (A) The overall method for the developed a signature-based graph neural network to generate a surface of NO2 ground-level from camera inputs. The arrows represent the flow of information from the CCTV footage to the prediction of the NO2 ground-level. (B) A scatter plot for the actual and predicted data for all sensor locations and all dates. (C) The results of training and validation loss and evaluation metrics for training and validation sets. (D) A scatter plot for the actual and predicted data for all sensor locations and all dates. (E) NO2 prediction for different hours of the day, aggregated at a borough level. This figure is created by the first author using python programming.