The occurrence of pollution hotspots is seldom haphazard; instead, it arises when a confluence of factors — such as emissions intensity, street canyon geometry, wind characteristics, and exposure levels — combine in such a manner that they generate a recurring or momentary concentration. A busy intersection, a narrow alleyway, a bus shelter in a stationary queue, or a blocked cul-de-sac can all serve as a hotspot.
Although we may utilise machine learning to identify such hotspots, the method must be used appropriately. The usefulness of a machine-learning approach does not lie in the ability of algorithms to uncover pollution problems; rather, it lies in the capability to integrate multifaceted urban data and extract potential problems that might merit more scrutiny.
This might be helpful when dealing with complex street networks, but hotspot detection still requires ground truth evidence, physical interpretation, and model validation.
Why Are Hotspots Difficult to Identify?
A pollution hotspot is not necessarily a place with a high quantity of emissions. Instead, it might be where air pollutants are concentrated, where individuals remain in that location for a greater duration, or where the local wind patterns blow a plume.
This forms one part of why street-level air pollution is so difficult to predict. A small alteration in building elevation, wind direction, traffic flow, or the position of the source of an emission can result in a significant change in where the build-up of concentrations will happen.
Consequently, a monitoring system might miss some of this variation. One fixed monitor might deliver a good measure for that location, but does not represent all nearby streets or intersections. A whole city model may present the more general pattern but may not necessarily account for the local differences between one junction and the next.
A machine learning approach might be helpful here because we can apply it across several variables simultaneously, comparing against monitoring data with traffic conditions and weather, road configuration, land use patterns and the outputs of our models to identify those places where an elevated concentration can be expected more frequently.
Where Might a Machine Learning Approach Be Helpful?
A machine learning method is well adapted for detecting patterns. If a street network is complex, the patterns of pollution might be determined by the interaction of multiple factors simultaneously.
For example, a place might only emerge as a hotspot when the wind is calm with traffic levels rising. Another might display peaks when the wind comes from a specific direction. A third might be influenced by a plume passing from a nearby road, rather than being formed by local emissions.
A machine learning method might prove valuable in spotting this connection in a large dataset. It might highlight the repetitive behaviour of hotspots that are otherwise less obvious when viewed as daily averages. It might also help us to identify whether a particular place might actually form a persistent hotspot, an isolated event, or an issue with a sensor.
This can be helpful, for example, when we are making use of urban pollution monitoring networks to obtain high-frequency data. At that point, the task is no longer only gathering the information but also making sense of that information.
Hotspots Are About Exposure, Not Just Concentration
A hotspot really matters in relation to exposure.
A high concentration on a deserted service road may be less of a concern for exposure than a smaller, but frequently repeated peak at a bus stop, a school entrance, a crossing or on a cycle route. For policy making, it is not just a matter of where pollution is high. It is a matter of where people are going to breathe it.
AI could assist in combining pollution data with data relating to movement and activity. Pedestrian routes, bus stops, queueing traffic and the layout of streets, for example, may all assist in identifying where exposure is likely to be significant.
This is where hotspot detection meets the idea of short-term pollution events in urban policy. A place may not have the highest average concentration, but it could have frequent short-term peaks when people are actually there.
The distinction is crucial. Averages can conceal where short-term exposure actually happens.
Using AI with Sensor Networks
Sensor networks can provide the data required to identify local variation. Sensors positioned at close intervals could indicate that levels vary markedly along a main road, along a side road and at a crossroad nearby.
AI may assist in interpreting this, identifying similar locations, identifying anomalous readings, identifying repeated peaks and predicting where hotspots are likely to be between sensors.
But the result depends on the quality of the network. Sensors that are badly calibrated, have poor rationale for their locations or exist predominantly in obvious places will lead to a weak AI model.
Machine learning, for example, that has been trained principally on main road data will be less trustworthy for courtyards, small side streets or pedestrian spaces. It may tell you what sensor locations you have got rather than where exposure is likely to be important.
Hotspot detection thus has to start with good network design. AI can help to interpret a network, but it cannot make up entirely for a network that is asking the wrong question.
Linking Hotspots to Urban Form
Complex networks of streets affect airflow. Buildings can block it, funnel it or divert it. Intersections may enhance mixing. Street canyons can trap air. Openings between blocks may form unanticipated airways.
AI can help identify statistical relationships between urban structure and air concentration. It may suggest that specific street profiles may be prone to repeat hotspots at certain times of weather.
But statistical associations are no substitute for physical understanding. If the AI identifies a hotspot near a corner of a block, the next step is to ask why. Is it due to local recirculation? Is it the position of the source? Inadequate air exchange? The direction of the plume on another street?
Studies with controlled settings such as wind tunnel experiments remain useful, as they enable us to test the physical mechanisms causing the patterns.
Absent the understanding of how it all works, we know where with AI but not why.
Field Campaigns Are Still Important
The detection of a hotspot via AI is limited by the quality of data it can learn from.
Field campaigns offer real-world data such as traffic, meteorology, buildings, background levels and people movement. They capture the variability that may be absent from a purely digital approach.
The development of policy based on field campaign evidence involves the matching of observations to a suitable interpretation. AI can aid this matching by parsing through the complexity, however the observations need to be valid, appropriate and robust in the first instance.
Hotspots that rely on tenuous evidence have limited value. Hotspots supported by field measurements, sensor data, modelling results and physical interpretation are of much greater utility.
AI and Model-Derived Hotspot Mapping
Models are commonly used to provide estimates for pollution in areas without direct measurements. AI can add to this by combining observational and model data to generate more accurate hotspot maps.
AI may help in correcting model bias, estimating concentrations in gaps between monitoring sites, or pinpointing areas in which a model is uncertain.
The resultant maps have their place in planning, but there are dangers. AI-generated hotspot maps could appear credible even where the underlying uncertainty is significant. Maps created without the model being tested at the appropriate scale could be misleading.
Therefore, model validation needs to be considered a policy issue. The detection of a hotspot could be used to prioritise monitoring or street management, for example, or the interpretation of health outcomes. We need to be sure that the detection methodology can be verified.
The Role of Tracer Evidence in Understanding Hotspots
Some hotspots are created by pollutant pathways, not just direct sources nearby. The pollutant may move down one street, past a junction and over the top of roof height back down to street level in another.
AI might identify this pattern, but tracer evidence will support it. Data from urban tracer experiments will identify how material moves through the city streetscape starting from a known release.
That is important to hotspot analysis as it avoids the simplistic assumption that the nearest source is responsible. In an urban dispersion scenario, the pathway is equally important as the point of origin.
The Applications of AI Hotspot Detection in Policy
AI-driven hotspot identification can also serve a range of policy functions.
It can assist in prioritising the locations for monitoring efforts, particularly where past observations point to recurrent brief spikes. It can contribute to the design of traffic management strategies to identify those areas where queuing or acceleration may be adding to local exposure. It can inform the planning of public space to identify those places where people wait, walk or cycle in proximity to the likely concentration peaks.
It can also support emergency preparedness. Within dense cities, planning for the release of airborne hazardous materials relies on an understanding of the areas where pollutants are likely to move rapidly and which areas should be given priority for monitoring.
In all of these examples, AI should be viewed as a support for decision-making processes. It can help to pinpoint those places that warrant attention, but should not be used in isolation to trigger any response.
Limits and Risks
There are several risks in applying AI to hotspot identification.
- Poor training data. If training data are incomplete, biased or not representative, the model may uncover patterns which do not transfer to other conditions.
- Over-confidence. AI models produce outputs that can appear precise, especially when plotted in a visual format such as a map. A predicted hotspot should not be assumed certain without verification.
- Lack of transferability. A model trained within one street network may not transfer to a different city with different buildings, different traffic movements or different meteorology.
- Limited transparency. The users of a hotspot map may need to know more than just where the hotspot is. They may need to know why a hotspot is there and what is possible in reducing exposure.
None of these risks render AI useless. They merely specify the circumstances in which it can or should be used.
A Practical Use for AI
AI can assist in identifying pollution hotspots, but only when used in conjunction with field measurements, deployed sensor networks, tracer evidence, physical modelling and validated dispersion models.
Its role is to help manage and structure this diverse collection of evidence. It can pick out patterns, anomalies and candidate hotspots to support the design of monitoring and testing programmes. It can help researchers evaluate the combined impact of multiple possible exposures.
However, AI should not be considered an authority in its own right. A pollution hotspot is defined not simply as a set of data points but as a physical phenomenon with important implications for policy.
The strongest evidence-based approach is one that is led by evidence. AI is useful to reveal patterns, field data used to ground these in reality, physical modelling to explain how they arise and validation to determine whether the results are precise enough to support decision-making.
In complex urban street networks, this combination is vital. Pollution hotspots emerge from a complex mix of flows, built features, meteorology, emissions and exposure. AI can help illuminate the interrelationships between these factors but it cannot replace the need to understand and account for them.
