Machine Learning Calibration of Low-Cost Sensor PM2.5 data

Irfan Yaqoob, Vijay Kumar, Shafique Ahmad Chaudhry

202411 citationsDOI

Abstract

Recent Advancements in low-cost air quality sensors have significantly expanded their potential applications in local air quality and environment monitoring. In contrast to regulatory PM<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2.5</inf> monitors, these low-cost sensors are more affordable, portable, and easier to maintain. This enables ambient air quality monitoring with higher spatio-temporal resolution. However, the accuracy and reliability of low-cost sensors are generally inferior to those of regulatory monitors, and their performance varies across different environments and sensor technologies. Therefore, calibrating data from low-cost sensors is crucial to mitigate systematic biases and errors, thereby making it usable for informed policy-making decisions. This study builds and evaluates the performance of various machine-learning models to calibrate PM<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2.5</inf> data from Purple Air(PA) sensors. The calibration models are developed using the PM<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2.5</inf> concentration, temperature (T), and relative humidity (RH) of the PA, with PM<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2.5</inf> concentration measurements from Federal Monitors. We have evaluated the performance of various models including Linear Regression (LR), Random Forest (RF), Gradient Boosting (GB), k-nearest Neighbors (KNN), and Neural Networks (NN) using several metrics, including the coefficient of determination (R2), root mean square error (RMSE), and mean absolute error (MAE). Spatial validation of the model is conducted using leave-one-location-out (LOLO) and Leave-Multiple-Locations-Out (LMLO) methods. Furthermore, we utilized k-fold cross-validation methods to ensure robust evaluation of the model performance across different locations and times. The Neural Network outperforms all the remaining models with an R2 of 0.88, a root mean squared error (RMSE) of 1.75 (µg/m3), and a mean absolute error (MAE) of 3.13 (µg/m3). In comparison, the Random Forest model achieved an R2 of 0.87, an RMSE of 1.84 (µg/m3), and a mean squared error (MSE) of 3.39 (µg/m3). Gradient Boosting showed an R2 of 0.86, an RMSE of 1.89 (µg/m3), and an MSE of 3.60 (µg/m3). The K-Nearest Neighbors (KNN) model resulted in an R2 of 0.83, an RMSE of 2.12 (µg/m3), and an MSE of 4.49 (µg/m3). Lastly, the Linear Regression model achieved an R2 of 0.83, an RMSE of 2.13 (µg/m3), and an MSE of 4.57 (µg/m3).

Topics & Concepts

CalibrationComputer scienceArtificial intelligenceRemote sensingGeologyStatisticsMathematicsAir Quality Monitoring and Forecasting

Machine Learning Calibration of Low-Cost Sensor PM<sub>2.5</sub> data

Abstract

Topics & Concepts