Machine Learning-based Virtual Densification of Sparse Environmental Variables (Solar Radiation and PM10) from Co-located Basic Weather Observations across South Korea

Among the weather networks in South Korea, although basic variables are recorded at 600 to 744 stations nationwide, only approximately 67 and 31 are equipped with sensors for detecting solar radiation (GSR) and particulate matter (PM10), respectively. Exploiting this asymmetry, we propose a machine-learning-based virtual densification to facilitate estimations of sparse variables from co-located basic-weather variables, evaluated based on leave-station-out (LSO) cross-validation, excluding entire stations to emulate sensor-free locations. Solar radiation was found to be effectively densified (LSO R2 = 0.76 ± 0.03), far exceeding the Hargreaves-Samani baseline (R2 = 0.54) and was stable without coordinates, indicating a physical relationship. Comparatively, PM10 reaches only R2 = 0.28 from weather alone, a control for emission-driven variables. Although machine learning (RandomForest, XGBoost) achieves significantly higher predictive accuracy than the empirical and linear baselines (p < 0.001), the two models are mutually indistinguishable. Given that solar radiation is a required wet-bulb globe temperature input, the virtual solar field can directly support heat-disaster risk assessment in locations in which sensors are absent.