Improving Accuracy in PM2.5 Interpolation Using AI and ML

Abstract

Air quality prediction is increasingly critical, especially considering recent events like the Canadian Wildfires releasing hazardous particulate matter over the US. The growing awareness of AI and machine learning have been increasingly used to facilitate applications in scientific studies, notably PM2.5 retrieval. PM2.5 retrieval uses machine learning (ML) techniques to estimate accurate PM2.5 values by considering predictors such as meteorological variables. One of the challenges in PM2.5 retrieval is dealing with different predictors with varying spatial resolutions. Prior to retrieval, predictor variables needed to be interpolated into a uniform grid, which currently lacks a standardized and validated model. Different interpolation methods (Inverse Distance Weighting (IDW), Kriging, and Natural Neighbor) offer techniques for estimating values in PM2.5 retrieval. Understanding the results involves assessing the spatial behavior of the data and validating the interpolation methods using metrics like Root Mean Squared Error (RMSE) and R-squared (R2). In this study, different interpolation models in ArcGIS Pro were used to interpolate meteorological variables. The models employed estimated values at a uniform grid with a spatial resolution of 1 km x 1 km. Interpolation methods were carefully evaluated using validation metrics to assess their effectiveness and accuracy in capturing spatial patterns and variations at this resolution.