Long-term calibration models to estimate ozone concentrations with a metal oxide sensor

Area of Science:

• Environmental Science
• Sensor Technology
• Atmospheric Chemistry

Background:

• Ozone (O3) is a harmful air pollutant. Low-cost metal oxide (MO) sensors offer potential for enhanced ozone monitoring but face data quality challenges.
• The University of Utah's AirU network uses MO sensors, demonstrating good lab response to ozone despite intra-sensor variability.
• Regulatory ozone measurements often lack the spatiotemporal resolution needed for comprehensive air quality assessment.

Purpose of the Study:

• To evaluate the field performance of low-cost MO sensors for ozone (O3) detection.
• To develop and compare calibration models (MLR and ANN) for predicting O3 concentrations using MO sensor data.
• To assess the generalizability of these calibration models for predicting ozone levels from similar sensors.

Main Methods:

• Eight AirU sensor packages were co-located with regulatory ozone monitors for one year.
• Multiple linear regression (MLR) and artificial neural network (ANN) models were developed using sensor data and meteorological variables.
• Variable selection was performed using LASSO, MLR, and ANN to identify key predictors for ozone concentration.

Main Results:

• The ANN calibration model showed superior performance (R² = 0.767) compared to MLR (R² = 0.491) on a holdout set of MO sensors.
• Both models demonstrated moderate predictive power for ozone concentrations in the subsequent five months (ANN R² = 0.567, MLR R² = 0.427).
• Key predictors identified included MO sensor readings, temperature, and solar radiation.

Conclusions:

• Low-cost MO sensors, when calibrated with an ANN model, can effectively supplement reference methods for ozone monitoring.
• This approach enhances the understanding of ozone's spatial and temporal variations.
• The developed ANN model shows promise for improving the accuracy and applicability of low-cost sensor networks in air quality studies.