A breath of fresh (AI)r

A breath of fresh (AI)r

Cleaning data can help solve one of the globe’s ‘dirtiest’ health issues

Despite fewer headlines, air pollution kills far more people each year than Covid-19. Now, researchers have shown that machine learning can help clean data from low-cost sensors, empowering communities to better monitor air quality.

“If I’m exposed to any triggers, I have a strong reaction. My throat tightens and becomes sore and itchy. My lungs feel swollen, and my breath gets heavy,” explains Robert Johnstad, who has struggled with asthma since childhood. A chronic lung disease severely worsened by air pollution.

“On days with poor air quality, just being outside can be challenging. My quality of life heavily depends on accessibility to clean and unpolluted surroundings,” he says.

Unfortunately for Johnstad, a breath of fresh air is not always guaranteed, neither in his hometown of Trondheim, Norway, nor in most other inhabited areas on the globe.

In fact, 91 per cent of the world’s population lives in places where air pollution levels exceed the World Health Organisation’s (WHO) guideline limits. This exposure increases the risk of respiratory and heart disease in the population, and a staggering 4.2 million deaths every year occur due to exposure to outdoor air pollution. Furthermore, a recent health briefing from the European Environment Agency shows that air pollution continued to be a significant cause of premature death and illness in Europe in 2019. By improving the air quality to the levels recommended by the WHO, more than half of the premature deaths caused by air pollution could have been prevented.

But despite the extent of the problem, the key to better safeguarding citizens from hazardous air might be found in the results from a small AI project in Trondheim.

Trondheim’s air quality is satisfactory most of the year, but especially the use of studded tires on ice-free roads during wintertime forms harmful particulates, which cause severe quality deterioration, particularly dangerous for those with respiratory and cardiovascular diseases. “The air quality situation is often worst downtown, an area I travel to daily due to work,” says Johnstad.

Global issue, local responses

Since 2018, Telenor Research, together with the Norwegian University of Science and Technology (NTNU) and the Information Technologies Institute, Centre for Research and Technology Hellas (ITI-CERTH), has collaborated with Trondheim municipality to improve the city’s air quality services through a pilot funded by the European Artificial Intelligence project AI4EU.

“In short, we have explored how Artificial Intelligence, and more specifically Machine Learning, can be used to improve air quality data coming from low-cost IoT sensors and air quality services further by combining pollution data with other information,” explains Sigmund Akselsen, Senior Research Scientist in Telenor Research.

To obtain air quality data, municipalities depend heavily on sensors. However, the challenge for most cities is that air quality is a hyperlocal phenomenon, varying from street to street, and thus, detailed data is required to achieve adequate monitoring.

Microsensor: Local measurements of air quality using high-end sensors are, for most municipalities, not feasible due to high costs. An industrial sensor can cost as much as 50 times more than a low-cost microsensor (to the left), so better calibration of data from the latter might be the future solution for many cities.

“Today, policy makers are forced to balance between coverage and deployment costs. A bulky, industrial sensor type provides accurate data but is expensive, both in deployment and maintenance. Lower-cost sensors reduce the expenses significantly, allowing for a denser coverage of regions, but they aren’t capable of providing high-quality data,” says Akselsen.

“If we could improve the data quality coming from low-cost sensors so that it meets the minimum requirements for a public evaluation of air quality levels, we would help solve a severe global challenge. And the results from our air quality pilot in Trondheim show that this is possible.”

Co-location: Two co-located micro- and industrial sensors are placed at Elgeseter and Torget in Trondheim city. While the sensors at Elgeseter are placed at a location experiencing heavy traffic (local air pollution), the sensors at Torget are situated on top of a 15-meter-tall building to measure background pollution (air pollution not produced locally). Here, Bjørn Borud (from the left) and Hans Jørgen Grimstad, partners at Lab5e, and Bjørn Villa, Head of IT at Trondheim Municipality, inspects the low-cost microsensor at Torget.

Calibrating into meaningful input

The air quality data used in the pilot is generated by a network of 25 low-cost sensors compiled by local IoT platform company Lab5e and deployed all over Trondheim by the municipality. Two of the low-cost sensors play a particularly crucial role as they’re put next to an industrial sensor each.

“Data from the more accurate industrial sensors has been used as the reference point in this pilot. By comparing the incoming data from the two different sensor types, we have found that the low-cost sensors tend to underestimate the pollution levels,” explains Tiago Santos Veiga, Postdoctoral fellow at the Department of Computer Science at NTNU, who has been working with the pilot’s data exploration. He adds:

We have also observed that low-cost sensors measurements are affected by external factors, such as humidity and possibly other meteorological factors as temperature and air pressure to name a few. This, together with a low correlation between low-cost and reference sensors, led us into investigating automatic calibration models.

Calibration: “Sensor calibration can be defined as transforming the measurements from sensors in a way that the calibrated measurements follow closely those from reference instruments,” explains Kerstin Bach, associate professor in computer science at NTNU. Here, she discusses results from the air quality pilot with Tiago Santos Veiga in the Norwegian Open AI Lab in Trondheim

(Photo: Kai T. Dragland/NTNU)

Based on the data analysis, the researchers applied Machine Learning techniques to develop an air quality prediction model that takes input measurements from the microsensor and external factors and targets the measurements from the co-located reference sensor at Elgeseter.

“The result was a much higher correlation between the data from the microsensor and the reference sensor. High correlation was also the case when we applied the prediction model trained on data from Elgeseter on the measurements from the microsensor at Torget,” explains Kerstin Bach, associate professor at the Norwegian Open AI Lab and NTNU’s main project contact.

“Our results show that the calibrating process manages to transform the data from low-cost sensors into meaningful input. This can enable municipalities to obtain the minimal information required to do a proper assessment of the air quality.”

Correlation: Visible are plots of datapoints from micro- versus reference sensor, before and after calibration. "The points have become more centered along a straight line. A straight line implies a perfect correlation, so the intuitive interpretation is that the calibration model works as the correlation has increased," says Sigmund Akselsen, Senior Research Scientist in Telenor Research. He also states that one of Telenor's aims with the project has been to build competence within Machine Learning as the technology will become important for the company's ability to capture new growth opportunities in the future.

Accelerating AI, empowering societies

Within the wide range of possibilities for services that use air quality data, the researchers have discussed two applications that benefit from the calibration procedure.

“We addressed the desire for real-time services from the municipality through exploration of a warning system for prediction of pollution peaks in the next 24 hours, and visualisation techniques with spatial and temporal patterns of the air quality measurements,” says Akselsen.

For Trondheim municipality, this work has provided insights for building improved air quality services that could lead to better decision-making support and the realisation of new digital health and safety solutions.

“We believe many groups will benefit from accessing better air quality data. For example, individuals can better assess if there’s a risk connected to current or future air quality, while decision-makers can more adequately consider when and where to perform cleaning actions,” says Thea Berg Lauvsnes, department head of environmental health and pollution in Trondheim Municipality.

In 2013, Trondheim municipality was ordered by the Norwegian Environmental Agency to improve the air quality situation after Norway had been taken to the EFTA court over high pollution levels in a number of Norwegian cities. “Since then, we have taken actions in terms of traffic regulations and introducing better cleaning and dust binding measures. Still, we want to further extend our efforts, and that includes testing and evaluation of intelligent decision support systems,” says Lauvsnes.

Improved advance warning to cleaning services operators and more detailed information in the form of the visualised measured and forecasted air quality levels are just a few examples of what improved measurements can enable, according to the department head.

“We could also develop monitoring and forecasting services delivered through mobile apps for citizens to plan their short-term activities and incentivise them to choose green transport options. That would surely be a game changer. So, in sum, the use of IoT sensor data and Machine Learning methods in air quality monitoring enables better decisions. Thereby, the municipality of Trondheim, and its citizens, are empowered to better safeguard themselves and their environment,” concludes Lauvsnes.