Advancing cold-region hydrology with large-sample data and deep learning: insights from Icelandic catchments

Abstract

This thesis advances understanding of cold-region hydrology by combining large-sample data development, hydrological trend analysis, and deep learning. Iceland serves as the study domain because of its extensive hydrological observations and its unique landscape shaped by snow, glaciers, permeable volcanic terrain, and minimal human influence on rivers. These conditions allow natural hydrological processes to be examined in a rapidly warming Arctic environment.The first part introduces LamaH-Ice, a dataset that compiles streamflow records, detailed weather information, glacier mass-balance measurements, snow cover, and catchment characteristics for more than 100 Icelandic basins. It fills an important gap by providing consistent multi-basin information from a high-latitude region where observations are typically scarce. The second part examines how Icelandic streamflow has changed over the past few decades. The analysis shows higher cool-season streamflow and lower summer flows in many rivers, with baseflow acting as a buffer that dampens summer flow declines and moderates winter and spring increases. Glacier-fed systems have shifted from increasing melt-season discharge over the past fifty years to weaker or negative tendencies in the past thirty years. Streamflow variability has also declined, and baseflow now makes up a larger share of total runoff. The third part evaluates a regional Long Short-Term Memory model trained across Icelandic catchments. The model predicts streamflow with strong skill, and an examination of its internal states shows that it learned hydrologically plausible snow and ice dynamics even when trained only on streamflow. Static physiographic catchment characteristics contribute little to raw predictive accuracy but play an important role in interpretability. Multi-task experiments show that adding a cryospheric output improves the model’s internal representation of snow and ice processes, but does not improve streamflow predictions under the tested architecture. Together, these contributions provide new data resources, clarify how Icelandic hydrology is changing, and demonstrate the potential of deep learning to capture processes in snow- and glacier-influenced environments.