Answers are grounded in this paper and cite their sources.
Accurate estimation of Snow Water Equivalent (SWE) is crucial for water resource management and hydropower operations in snow-dominated regions. This study presents a surrogate model based on a ConvLSTM architecture to predict daily SWE at a high spatial resolution (1 km×1 km) in the Drammen River Basin, Southern Norway. Designed to emulate the SeNorge snow model, the proposed framework uses meteorological inputs, such as precipitation and temperature from the SeNorge observational dataset, along with topographic features. To mitigate error propagation in the prediction process, the model was trained using an autoregressive scheme with scheduled sampling, allowing the model to incrementally shift from using true SWE target values to its own prior estimates as subsequent inputs. The surrogate model was also driven and evaluated using input data from ERA5-Land dataset, to measure its potential for generalisation and transfer, as well as to identify overfitting issues. Results demonstrate that the surrogate model effectively captures spatial and seasonal SWE dynamics in the study area, showing strong agreement between the target SeNorge SWE dataset and the simulated ConvLSTM driven by SeNorge and ERA5 datasets, with median KGE values greater than 0.7 and 0.5, respectively. In addition, a comparison with in-situ observations demonstrates that the ConvLSTM model driven by both datasets achieves KGE values greater than 0.7. This work highlights the potential of data-driven deep learning approaches for developing robust tools to support hydrological prediction.
No takes yet. Share an insight, caveat, or question.
Ávila et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: