Global Quantification of Black Carbon in Seasonal Snow: A Physically and Observationally Constrained Machine-Learning Framework

Black carbon in seasonal snow (BCS) critically influences the Earth system by reducing surface albedo (snow darkening), perturbing radiative balance, and accelerating snowmelt. However, its climatic and hydrological impacts remain poorly quantified because high-quality data sets are scarce. This study introduces a novel Dual Random Forest (DRF) framework that synergistically constrains BCS concentration estimates through both physical mechanisms and observational fidelity. By pretraining with six spatiotemporally complete Earth System Model (ESM) simulations, driven by multisource reanalysis data, and fine-tuning against global field observations, the DRF generates a 44-year (1981–2024) global monthly BCS data set at 0.5° × 0.625° resolution. This data set achieves higher accuracy in BCS patterns and concentrations compared with observations (spatial correlation R = 0.92; normalized mean error NME = 31%), outperforming both single-constrained machine learning approaches (ESM- or observation-only) and traditional methods (CMIP5, CMIP6, and MODIS; R = 0.07–0.82, NME = 51%–1951%). It also excels in capturing seasonal variations and long-term annual trends, better reproducing increasing trends driven by anthropogenic forcing and decadal variability modulated by natural climate variability. This data set is hence valuable for robust quantification of BCS-induced radiative forcing and attribution of snowmelt acceleration to snow darkening versus climate warming on a global scale.