Limitations of Environmental Extrapolation in Low-Cost Carbon Monoxide and Fine Particulate Matter Sensors

Low-cost air quality sensors have grown in popularity. Calibrating these sensors in environments that perfectly mirror the environments they will be used in is difficult and can be costly, so researchers often need to extrapolate calibration models outside of the environmental and pollutant data ranges they are trained on. As such, understanding the impact that this extrapolation has on the resulting measurements is important when using this data. In this study we conducted a multi-season colocation between two low-cost sensors (Plantower PMS5003 for PM2.5 and Alphasense CO-B4 for CO) and Federal Equivalence Method monitors. We then trained multiple calibration models on different selective ranges of relative humidity, temperature, and pollutant concentration data to test how those models performed when extrapolated into data ranges. There were three main takeaways in this study. First, there was minimal bias error in calibration model performance when a model is applied to data that falls within its training data range. Second, extrapolating models into data that was outside of their training data ranges led to increases in bias error. Third, extrapolating models into higher pollutant concentrations than what the models were trained on increases random error.