Airvoice has initiated research on developing a neural network capable of classifying air pollution episodes

2023-11-02 17:04

When an#nbsp;air pollution episode starts in#nbsp;a#nbsp;city, it’s important to#nbsp;understand the expected scale of#nbsp;the problem as#nbsp;early as#nbsp;possible to#nbsp;respond most effectively. As#nbsp;part of#nbsp;our internal research projects, our mathematical modeling team is#nbsp;training neural networks to#nbsp;classify air pollution episodes by#nbsp;their potential size and possible cause. This is#nbsp;based on#nbsp;an#nbsp;array of#nbsp;data from about 500 events that occurred in#nbsp;the same city over a#nbsp;year.

We#nbsp;think of#nbsp;a#nbsp;city as#nbsp;a#nbsp;space with a#nbsp;large pool of#nbsp;air#nbsp;— a#nbsp;part of#nbsp;the atmosphere where impurities are collected and mixed, which all citizens breathe.

However, certain areas of#nbsp;this pool may have local pollution sources, resulting in#nbsp;air quality that can be#nbsp;considerably worse than the average. Pollution episodes may also arise from the transfer of#nbsp;pollutants across city/community borders (typically from wildfires) or#nbsp;from the buildup of#nbsp;harmful substances in#nbsp;the air (often due to#nbsp;weather conditions like cold and still weather).

The sooner an#nbsp;episode is#nbsp;identified, the quicker and more effective actions can be#nbsp;taken. For example, if#nbsp;the pollution came from outside, city authorities may not be#nbsp;able to#nbsp;address the source directly but can minimize community’s exposure by#nbsp;warning people and organizations. These might include advice to#nbsp;stay indoors, close windows, turn on#nbsp;ventilation systems, and requests for enterprises to#nbsp;reduce emissions (as#nbsp;the air is#nbsp;already dirty).

Our model analyzes measurements from an#nbsp;entire network of#nbsp;monitoring stations and anticipates the scale and dynamics of#nbsp;pollution dispersion. Currently, it#nbsp;classifies an#nbsp;episode with 90% accuracy within 24 hours of#nbsp;its onset. The classification accuracy is#nbsp;nearly 100% for common events and lower for rare ones, which is#nbsp;attributed to#nbsp;limited data for neural network training.

We#nbsp;continue our research, as#nbsp;we#nbsp;still have a#nbsp;lot of#nbsp;questions, such as "how to#nbsp;train a#nbsp;model to#nbsp;predict the pollution episode type and scale prior to#nbsp;its start."

We#nbsp;will keep you updated on#nbsp;the results.

*Photo taken on#nbsp;November 2, 2023 in#nbsp;Gurugram (India)