Air pollution is hardest to understand in street canyons. Roads have buildings on both sides and emit right next to where people are standing. In these relatively narrow areas, air flow can get funnelled, held, slowed, reversed, or trapped.
Consequently, air pollution patterns within street canyons can change significantly in a very short distance. Peaks in concentration can occur on one side of the road or the other, at a junction, by the edge of a building, or around a bus stop. A measure taken at one location may not represent what exists only metres away.
AI is well-placed to help explain some of these patterns, particularly when a network of sensors is available alongside traffic data, meteorological data, and model outputs. However, machine learning must not be used to replace physics in street canyon air pollution analysis. AI can help identify patterns, but those patterns must be explained.
The Complex Nature of Street Canyons
A street canyon is more than just a street bounded by buildings. It is a small-scale flow environment, where street width, building height, wind direction, flow on the roof, and traffic pollution interact.
When wind crosses a street canyon, the air flow within the canyon may recirculate, rather than just pass through. On one side of the street, concentrations may be high, and on the other they may be low. On other occasions, the situation is reversed, and depending on the wind direction and wind speed, pollutants may remain close to the source for a longer period of time.
This is one reason why short pollution events at street level matter for urban policy. A short period of bad ventilation within a street canyon can cause conditions that are not captured in long-term averages.
The challenge is not just measuring them; it is understanding them.
How AI Can Be Used to Interpret
Machine learning can be helpful where you have a large number of interacting variables, such as in street canyons, where pollution may depend upon wind speed, wind direction, traffic flow, queueing vehicles, temperature, building geometry, and background pollution.
AI may be able to help understand when you see high concentrations. It may, for instance, identify a particular side of a street where the concentration is higher when the wind is coming from one direction. Or, it might help reveal short periods when low wind speed is associated with traffic congestion.
It can also help to discriminate between different types of events. An anomaly on a single sensor might appear similar to a high concentration all across the street or a plume moving down the street. With enough data, and in some cases at high frequency, AI can help distinguish between these.
This is a more practical application of machine learning in urban air pollution dispersion, when it helps to interpret the evidence rather than to replace it.
Street Canyons and Networks of Sensors
To measure pollution at street level is already difficult. To understand it within a street canyon is even harder. There might be different conditions on the windward and leeward side of the street, or higher up on the street, or in the middle of the canyon versus on the edges.
A network of sensors can help provide a better picture. With sensors at multiple points, you might see whether a high concentration repeats at one place or if it moves through the street canyon, and whether the variation in concentration depends on the wind or traffic.
This is where AI can add value, by picking up signals from the network and highlighting things that are unusual. The AI could point to groups of sensors acting in unison, short periods with an increase in values or places where the concentrations are behaving differently from their surroundings on the street.
Yet as long as the questions asked of an urban pollution monitoring network are ill-matched to the sensor placements, the network will not be of much use. There is no point in having data, no matter how big or small the monitoring network, unless the placement is well suited to the questions asked.
Placing sensors without thought for airflow or source location or exposure will generate lots of data but will not give much information on why the canyon is behaving in certain ways.
AI is great at handling the data, but it cannot fix a poorly designed monitoring network.
Traffic, Timing and Local Peaks
Traffic plays a key role in the pollution of street canyons, but its effects are not the same everywhere. Traffic stops, queues, turns and accelerates. Emissions might rise significantly at signals, junctions or points of congestion. It is precisely at these spots that people are more likely to be exposed to the pollutants.
AI can help bridge the connection between traffic flow and changes in concentrations. It might highlight that there are recurrent short periods of elevated values and link them to traffic timing, congestion or specific vehicle flows. The AI might also help explain what is happening in one canyon compared with another. It might find out if there are notable differences between rush hour and quiet times.
This is significant because short-term peak prediction is just as dependent on the timing as it is on the location of the peaks. There might not always be an elevated concentration at one place throughout the day, but there might be an important increase in concentration for several short periods throughout the week.
That is a useful distinction to make for policymaking, as it could enable better-informed decisions around monitoring, traffic management, street design and exposure reduction.
Linking Canyon Patterns to Hotspots
Street canyons often contain high concentrations but it might not be clear why. An area with high concentrations might be due to local emissions, poor ventilation, recirculation, pollution travelling down from a nearby road or something else.
AI can help identify likely locations of hotspots within complex street networks. It may reveal that certain parts of a canyon repeatedly have a high value or that high concentrations only occur at a street under particular wind conditions and traffic flows.
That kind of insight can help boost hotspot identification with AI, especially when the goal is to better understand exposure rather than just show areas with high concentrations.
However, identifying a hotspot is not enough. We also need to know why a hotspot is there. Without this, the results will be difficult to use for policy.
The Role of Physical Evidence
Street canyon dispersion is a physical process. AI is good at finding patterns in the data, but it relies on physical evidence to explain those patterns.
Wind tunnel experiments are very important in this respect as they offer a chance to investigate street geometry, source location and wind direction in a controlled way. They can reveal whether pollutants are likely to stay trapped in a canyon, whether a pollution plume will be blown over to one side of a street or how pollution varies near a junction.
It is for that reason that wind tunnel modelling with controlled evidence remains important for interpreting AI results. If machine learning detects that a certain concentration pattern repeatedly occurs, controlled experiments can help determine whether there is a credible physical explanation for what is going on.
Without the evidence, the AI might detect a correlation without being able to explain it in terms of the process behind the correlation.
Physical and Digital Methods in Concert
AI can work alongside numerical models. A machine learning method could be used to compare model outputs and observations, identify situations where a model underestimates canyon concentrations, or find situations where a model works poorly.
Street canyon models typically use simplified assumptions about geometry, turbulence and boundary conditions, and a method like machine learning could be useful to identify when these assumptions are important.
Thus, a combined approach to physical and digital modelling is crucial, since physical experiments are used to explain the processes and models are used to explore scenarios, and a method like machine learning is used to interpret large amounts of data and compare the results.
Ultimately, the best evidence is evidence from using these approaches jointly, rather than placing them in competition.
Why Validation Is Key
Model results can look very accurate in street canyons when maps, classifications or modelled concentration curves are presented. This appearance of accuracy can be misleading if the model has not been properly validated.
A machine learning model may be trained on one canyon but perform poorly on another; may be trained mostly on typical wind conditions but poorly on extreme wind directions; or trained on data at a single street and have problems predicting the concentration at other streets.
This is why model validation for policy is so important for machine learning. If the model is going to be used to underpin decisions about exposure, monitoring and traffic management then the performance must be assessed at that level.
In street canyons this means assessing whether the machine learning model captures the variation in concentrations from one side to the other, the short-term peaks and the impact of wind direction and local plumes, and not just overall performance.
Interpreting Patterns for Street Canyon Evidence
Street canyon evidence is useful for policy decisions when it can be used to answer questions, for example:
- Where in the street canyon do people receive highest exposure?
- Are the concentrations higher for some traffic conditions?
- Is the pollution the same or very different at the different sides?
- Is there a value to adding monitoring data in certain places?
- Is a hotspot associated with emissions, poor air movement or plume movement?
- Should the model be used for a street environment like this?
Machine learning can be used to support analysis to answer these questions, identifying candidate patterns, unusual patterns and comparing models. But the evidence is useful because of interpretation.
A pattern in a street canyon cannot currently be used to support an action even if it looks like something significant. If a pattern in the street canyon can be explained using measurements from sensors or a model validated in the environment, then it is a pattern to watch out for.
The Appropriate Role for AI in Street Canyons
There is a useful role for AI in street canyons for interpreting patterns, for example, analysing high-frequency monitoring data, detecting recurring peaks in concentrations, comparing behaviour of different sensors, identifying hotspots or comparing models.
But the role should be bounded. Street canyon pollution is not just a data challenge but a physical and policy challenge, and involves understanding emissions, air movement and buildings, exposure and uncertainty.
Thus, used in a careful way, AI will help make the evidence easier to interpret, but in an uncritical way it is going to create more complex maps and model outputs that cannot be compared to the physical evidence in a street canyon.
A useful future is not where AI replaces urban dispersion research, but where it helps researchers and policy-users make sense of the street canyon evidence we currently need to understand pollution in urban streets.
