High-resolution downscaling of source resolved PM2.5 predictions using machine learning models

Accurate predictions of source resolved atmospheric PM2.5 concentrations at high resolutions using chemical transport models (CTMs) require expensive CTM simulations and development of high-resolution emissions inventories. We use multiple machine learning (ML) approaches to downscale coarse-resolution (36 x 36 km2) CTM predictions to 1 x 1 km2 spatial resolution. ML predictions include concentrations of the major chemical components of PM2.5 and the contributions of its major emissions sources. Inputs for the ML models include 36 x 36 km2 source resolved CTM predicted concentrations of all PM2.5 components, meteorological data, and several land-use (LU) variables. The output of our ML models is the 1 x 1 km2 source-resolved concentrations of all major PM2.5 components in southwestern Pennsylvania (5184 km2 domain) during February and July 2017. Models were trained and validated using 1 x 1 km2 resolution source-and species-resolved CTM predictions of PM2.5 from recent complementary studies. The best overall performance was found using a random forest (RF) model, where species and source resolved PM2.5 concentrations were reproduced with low normalized mean bias (|NMB| < 0.01). The downscaling model captures the spatial distribution of PM2.5 both by component and source, with some discrepancies when predicting the plumes of large point sources that have long-range impacts. In a test of generalizability to unknown domains, the model differentiates well between areas that are primarily urban, rural, or industrial but faces challenges with the reproduction of the effects of large point sources of PM2.5 when entire quadrants are removed from the training data. The results represent a proof of concept for downscaling low-resolution CTM predictions using native high-resolution CTM predictions in training.

Automated summary

Machine learning approaches were used to downscale coarse-resolution CTM predictions to high spatial resolutions, successfully predicting PM2.5 concentrations and source contributions in southwestern Pennsylvania in February and July 2017. The best performing model was a random forest model, which accurately reproduced PM2.5 concentrations for all components and sources.