An adaptive Window Backtracking Snow Assimilation Algorithm Improved Snowmelt Runoff Forecasting in a Data-Scarce Mountain Basin

Accurately estimating mountain snow is essential for predicting spring snowmelt runoff. Integrating snow observations with model simulations through data assimilation techniques represents a highly promising approach for runoff forecasting in global data-scarce basins. However, existing snow assimilation algorithms often compromise model physical consistency by failing to trace model biases back to meteorological forcing. This study developed a new snow assimilation scheme using MODIS remote sensing snow cover data with VIC-CAS model. We defined a one-day adaptive backtracking window that operated at the subgrid scale from August 31 of the prior year to December 31 of the current year, tracing snow simulation errors to both the initial snow accumulation and precipitation within a day. The scheme was applied to a data scarce basin on the northern slope of the Tianshan Mountains, China, and better model performance was achieved. The root-mean-square error (RMSE) of the simulated snow cover fraction decreased from 30% to 15%, and the mean bias declined from 19% to 6%. The Nash–Sutcliffe efficiency (NSE) for daily runoff simulation increased from 0.58 to 0.82. The improved backtracking algorithm effectively suppressed runoff overestimation by reducing snow during the accumulation period, and suppressing new snow accumulation by reducing precipitation during ablation, decreasing initial snow water equivalent (SWE), or both, thus avoiding runoff overestimation. Furthermore, error tracing combined with solid precipitation feedback improved total precipitation estimates, enhancing mixed rain-snow peak flow predictions. This scheme provides a practical solution for operational snowmelt runoff forecasting in global data-scarce mountain basins.