Machine Learning based Seasonal Streamflow Forecasting in Cold-Region Catchments: Insights from LamaH-Ice dataset

Machine learning (ML) remains one of the best approaches for long-term seasonal streamflow forecasting in cold regions owing to its capacity to capture nonlinearity between inputs and outputs, as well as its scalability across hydroclimatic regimes. ML’s main advantage lies in the generalizability of these models when applied to heavily glacierized catchments. In this data-driven study, we mainly utilize the Extreme Gradient Boosting (XGBoost) regression to train and test seasonal streamflow predictions using the LArge-SaMple DAta for Hydrology and Environmental Sciences for Iceland (LamaH-Ice). This new dataset for Iceland, published in 2024 consists of topographic, hydroclimatic, land cover, vegetation, soil, geological, and glaciological attributes that are essential for understanding cryosphere–hydrology processes in cold regions. For more than 100 basins, time series information on meteorological forcings and variables relevant to cold-region hydrology, such as MODIS (Moderate Resolution Imaging Spectroradiometer) snow cover, glacier albedo are also available. The majority of gauged rivers in LamaH-Ice are reported to have minimal human disturbances, making the dataset particularly unique. The XGBoost model demonstrates strong predictive skill across the study basins, as indicated by Kling-Gupta Efficiency (KGE) and Nash-Sutcliffe Efficiency (NSE) metrics exceeding 0.98. Ultimately high-precision streamflow forecasting is needed to track hydrometeorological hazards and to aid our ability to manage water resources in cold regions, which are a source for irrigation and hydropower.

References

Helgason, Hordur Bragi, and Bart Nijssen. “LamaH-Ice: LArge-SaMple DAta for Hydrology and Environmental Sciences for Iceland.” Earth System Science Data, vol. 16, no. 6, 13 June 2024, pp. 2741–2771, doi:10.5194/essd-16-2741-2024. 

Strahlendorff, Mikko, et al. “Forestry Climate Adaptation with HarvesterSeasons Service—a Gradient Boosting Model to Forecast Soil Water Index SWI from a Comprehensive Set of Predictors in Destination Earth.” Frontiers in Remote Sensing, vol. 5, 20 Dec. 2024, doi:10.3389/frsen.2024.1360572.