Towards high-resolution air pollutants sensing through dense low-cost sensor networks – a case study in Munich

Being Germany’s 3^rd largest city, Munich has the nation’s highest number of daily commuters leading to high traffic volumes and adding up to the urban air pollution through emissions from combustion and tire and brake wear. Although urban air quality is generally improving over the past years, all measurement stations by the Bavarian State Office for Environment (LfU) in Munich still exceed the critical NO₂ level of 10 µg/m³ as suggested in the updated WHO guidelines 2021.

For quantifying air quality at high spatiotemporal resolution in the city, we developed a self-sufficient low-cost sensor system equipped with electrochemical cells (ECs) for measuring NO₂, NO, CO and O₃ as well as an optical particle counter.

In summer 2023, we started setting up a dense sensor network of 25 sites in the inner city and since then, we have gathered 22 million data points. Prior to installation at their final locations, we mounted each sensor system for several weeks at an automated measurement station operated by the LfU in order to acquire high-accuracy reference data for calibrating our system. During operation of the sensor network, several nodes are occasionally returned to the reference site and three sensor nodes have been mounted there continuously ever since. In total, 8 million data points have been gathered during these co-location measurements. The combination of frequent rotation of sensor nodes between network locations and reference site as well as long-term co-location nodes yields a unique dataset for a novel calibration approach. Here, we analyze the calibration performance of NO₂; other pollutants will follow.

Firstly, we analyzed the correlation of the EC’s raw hourly signal to the reference station by assessing their coefficient of determination (R²). Remarkably, highest R² values occurred in fall and winter time with temperatures in the range of -5 to +15 °C. Lowest and even negative R² values occurred during summer and during long-term co-locations facing seasonal changes.

Secondly, for assessing a real-world scenario, we analyzed the performance of one node with long-term co-location at the reference station. The raw EC data yields an R² of 0.35 over 27 weeks. By applying a Random Forest Regressor using the first 30 % of the data points for training and including temperature, humidity and NO as features, R² could be increased to 0.7. Currently, we are developing a novel calibration strategy that leverages this extensive co-location data set with advanced machine learning methods to increase the calibration performance and to map air quality within our sensor network at high resolution and accuracy.