A Physically and Observationally Constrained Dual Machine Learning Framework for Satellite Retrieval of Equivalent Black Carbon in Snow

This dataset provides daily, 500 m multi-model ensemble mean estimates of equivalent black carbon in snow (eBCsnow) over northern Xinjiang, Northwest China. The dataset was generated using DML-Sat, a physically and observationally constrained dual machine-learning framework. Here, eBCsnow is defined as the black carbon concentration required to reproduce the optical snow-darkening effect caused by light-absorbing particles in snow. DML-Sat consists of two components. DML-Sat_A learns the physically consistent relationships among snow spectral reflectance, eBCsnow concentration, snow grain size, and snow cover fraction from SNICAR radiative-transfer simulations. Four snow-grain morphologies and two black-carbon mixing states are considered, producing eight physical configurations. The trained models are driven by daily MODIS/Terra MOD09GA seven-band surface reflectance and solar zenith angle to generate preliminary eBCsnow estimates. DML-Sat_B subsequently uses field observations from Northwest China to correct systematic biases in the preliminary estimates. Uncertainties from field measurements and MODIS inputs are propagated through a Monte Carlo calibration procedure. Four machine-learning algorithms—Random Forest, XGBoost, LightGBM, and CatBoost—with two configurations for each algorithm are used to represent model structural uncertainty. For each valid MODIS pixel, eight observation-corrected estimates are generated, and their ensemble mean is used as the final eBCsnow concentration. The dataset can support studies of snow-darkening variability, snow-albedo reduction, radiative forcing, snowmelt acceleration, and the influences of anthropogenic emissions and natural dust sources.