Multisource Data Fusion for Snowmelt Runoff Onset Date Estimation in Forested Areas

Abstract

Accurate estimation of the runoff onset date (ROD) is critical for water resource management and the assessment of snowmelt-related hazards. Currently, remote sensing retrieval of ROD primarily relies on the backscatter of synthetic aperture radar (SAR). However, when using SAR data for ROD estimation in forested areas, microwave signals are severely attenuated due to scattering and reflection from branches and tree trunks, making it difficult to capture the state changes of snow beneath the forest canopy. Considering the strong correlation between ROD and vegetation phenology, this study proposes a machine learning-driven framework that leverages extreme gradient boosting (XGBoost) to enhance the accuracy of ROD estimation in forested areas through the synergistic fusion of multisource heterogeneous data. The model integrates 20 spatiotemporal features, including SAR backscatter, vegetation index, air temperature, and geospatial data. Focusing on California, USA, as the study area, we extracted eight SAR features from 3925 Sentinel-1 images, four vegetation phenological features from 207 MODIS images, and three meteorological features from ERA5-Land hourly temperature data for the period February to July 2017–2025. In addition, five geospatial features were derived from Copernicus DEM (GLO-30) and ESA WorldCover datasets. The model was trained and validated against ground-truth ROD data from 125 snow water equivalent (SWE) stations over nine years, yielding a total of 850 samples, using station-grouped five-fold cross-validation. The results show that the XGBoost ensemble model incorporating spatiotemporal features achieves higher accuracy compared to RF, SVM, LightGBM, MLP, and MLR, with R, mean absolute error (MAE), and RMSE reaching 0.63, 10.60 d, and 15.08 d, respectively. When analyzed separately by land cover type, XGBoost yielded the best accuracy among forest stations (R = 0.63, MAE = 10.81 d, RMSE = 15.26 d), whereas RF achieved the highest performance at nonforest stations (R = 0.67, MAE = 9.16 d, RMSE = 13.70 d). Stability analysis indicates no significant variance in MAE across different latitudes, elevations, aspects, or years (p>0.05), confirming the model's robustness. Designed for the historical monitoring of snowmelt, this model enables accurate large-scale estimation of ROD in forested areas under complex mountainous terrain for the first time.