Bias correction for near-real-time estimation of snow water equivalent using machine learning algorithms: A case study in the Tuolumne River basin, California

Accurately estimating snow water equivalent (SWE) in near-real-time is important for water resources management and water supply forecasting in snow-dominant regions. However, conventional SWE estimation approaches have large uncertainties in mountainous regions due to complex terrain, snow-vegetation interactions, and other challenging factors. This study develops a SWE bias correction framework (SWE-BCF) that utilizes the Airborne Snow Observatories (ASO) SWE data and machine learning (ML) algorithms to correct biases in a near-real-time SWE estimation linear regression model (LRM). The spatial distribution of LRM SWE residuals, which are estimated using the ASO SWE, is explicitly modeled using multiple ML algorithms and evaluated using a leave-one-out cross-validation (LOOCV) workflow. A wide range of commonly used ML algorithms is examined to model LRM SWE residuals, including Gaussian Process Regression (GPR), Support Vector Machine (SVM), Bayesian Regularized Neural Networks (BRNN), Random Forest (RF), and Gradient Boosting Machine (GBM). The results show that all ML algorithms substantially improve LRM SWE estimation accuracy. While the Kruskal-Wallis test indicates no significant difference ( p -value >0.05) among the bias correction models, the RF model outperforms others, with the highest median R 2 (0.89), the lowest median RMSE (69 mm), MAE (41 mm), and NRMSE (37.4 %), as well as the second-best median PBIAS (−6.6 %) in the LOOCV for correcting SWE bias. Four performance metrics (R 2 , MAE, RMSE, NRMSE) show significant improvements ( p -value <0.05) over the original LRM model, highlighting the effectiveness of SWE-BCF in correcting the spatial patterns of SWE. However, the correction in the basin-wide average SWE, as indicated by the PBIAS values, exhibits high variance and does not show significant improvement ( p -value >0.05). Among the three land cover types in the Upper Tuolumne River Basin, the alpine area showed the most substantial SWE improvements with the SWE-BCF. The structural adaptability of the SWE-BCF enables its transferability to various geographic locations and SWE datasets, allowing for an extension of coverage and frequency of more accurate SWE estimates. This potential advancement may improve water management decisions which rely on accurate water supply forecasts. • We design a machine-learning framework to bias correct spatial distribution of SWE. • The framework is informed by airborne remote sensing observations. • The RF model shows the best performance in SWE bias correction. • The RF-based SWE-BCF improves SWE accuracy by 9 % in median R 2 and 43 % in median RMSE. • The framework can extend the coverage and frequency of more accurate SWE estimates.