Next-Day PM10 Prediction Maps for South Korea Using Machine Learning with Meteorological and Transboundary Air Quality Data
TL;DR
An observation-driven framework for generating nationwide next-day PM10 prediction maps over South Korea using machine learning extends point-based machine learning to nationwide spatial PM10 mapping and provides an observation-driven complement to operational air quality forecasting and environmental management.
Abstract
PM10 poses significant risks to public health, and reliable short-term prediction is essential for air quality management. In this study, we developed an observation-driven framework for generating nationwide next-day (D + 1) PM10 prediction maps over South Korea using machine learning. The framework integrates ground-based PM10 observations from the AirKorea monitoring network, meteorological variables derived from the Local Data Assimilation and Prediction System (LDAPS), and a regional transport proxy based on PM10 observations from upwind locations in China. Station-based predictors were converted into spatially continuous grids using kriging and subsequently used as inputs for two tree-based ensemble models: Random Forest (RF) and Extreme Gradient Boosting (XGBoost). Because random cross-validation can yield optimistic estimates for spatiotemporally autocorrelated data, we report complementary validation results using a region-based geographic holdout (Correlation Coefficient (CC) 0.859–0.924; Root Mean Square Error (RMSE) 10.12–13.40 μg/m3 for RF) and leave-one-year-out (LOYO) evaluation (CC ≈ 0.63–0.77), with random cross-validation reported only as an upper-bound reference. Under chronological expanding window evaluation, RF and XGBoost achieved RMSE values of 18.75 and 19.08 μg/m3, respectively, compared with 23.19 μg/m3 for persistence. Nevertheless, the performance of the framework degraded in years with strong episodic variability, particularly in 2021. Ablation experiments further demonstrate that meteorological predictors and the China PM10 proxy substantially improve predictive accuracy, while systematic underestimation of extreme concentrations remains a key limitation. The framework produces next-calendar-day daily mean predictions using only inputs available after completion of the current-day observation window, thereby avoiding future information. It extends point-based machine learning to nationwide spatial PM10 mapping and provides an observation-driven complement to operational air quality forecasting and environmental management.