Machine Learning Calibration to Air Sensor Data During Wildfire Smoke Events

Wildfires significantly contribute to air pollution events, often driving fine particulate matter <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$(\text{PM}_{2.5})$</tex> concentrations above 200 μg/m<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sup>, and causing severe air quality degradation. Low-cost air quality sensors like PurpleAir (PA) are widely deployed for air quality monitoring, but their accuracy diminishes under extreme weather conditions. Linear correction methods have been developed to calibrate and correct bias in PA sensor data; however, these methods may fail to capture nonlinear relationships in the data, particularly at high PM2.5 concentrations. This study uses machine learning algorithms, including Neural Networks (NN), Random Forests (RF), and Linear Regression (LR) to develop calibration models for PA sensors' PM2.5 measurements with U.S. Environmental Protection Agency(EPA) reference monitors' readings as ground truth. The models are evaluated using the coefficient of determination <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$(R^{2})$</tex>, the mean square error (RMSE), and the normalized RMSE (NRMSE) across the breakpoints of the Air Quality Index (AQI). The neural network and Random Forest exhibit superior results and achieved <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$(R^{2})$</tex> of 0.98 and (RMSE) of 5.39 μg/m<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sup> and 5.59 μg/m<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sup>, respectively. The linear regression model resulted in an <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$R^{2}$</tex> of 0.93, and an RMSE of 11.92 μg/m<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sup>. Unlike prior correction methods that rely on multiple regression equations, machine learning models provide a single unified solution capable of handling diverse weather conditions, including extremes, ensuring consistent performance across varying scenarios. AQI results demonstrate that NN and RF significantly outperform existing correction methods, achieving higher accuracy and robustness across all AQI categories, including wildfire smoke conditions. These findings highlight the potential of machine learning to improve low-cost sensor calibration, ensuring reliable air quality monitoring during wildfire events.