A transfer learning approach for simulation of Greenland Ice Sheet meltwater runoff

Surface melt and runoff are responsible for much of the Greenland Ice Sheet’s mass loss over the last few decades. However, complex hydrological processes such as refreezing, percolation, and lateral flow make it challenging to accurately simulate meltwater runoff with physically-based models. In this study, we present a new approach for predicting daily meltwater runoff on the Greenland Ice Sheet using the principles of transfer learning. We first develop deep learning models to emulate the regional climate model Modèle Atmosphérique Régional (MAR). We then fine-tune these emulators with in-situ observations of meltwater runoff from three catchments on the Greenland Ice Sheet. We find that the inclusion of observational data during the fine-tuning stage yields a 10–59% reduction in meltwater runoff error rate relative to emulators that are only trained using climate model data. Across all three catchments, the average error rate of the MAR emulators was ±19%, while the average error rate of the TL models was ±12%. We also find that models developed using transfer learning can generalize to unseen catchments, provided that they are trained using in-situ observations from the same region. Finally, we find that the transfer learning models can generalize to unseen years and fill gaps in proglacial discharge measurements, although this works best when they are trained on many (i.e. >2) years of observational data. Overall, our study demonstrates the advantages of conducting transfer learning by fine-tuning deep learning emulators with in-situ observations to yield more accurate predictions of Greenland Ice Sheet meltwater runoff.