Quantifying Global Glacier Change using Foundation-Surrogate AI Methods

Keywords: machine learning, glaciology, spatial data science, cryosphere, climate change
Abstract Type: Poster Abstract

Abstract

This study aims to assess the validity of using a deep learning model as a foundation-surrogate model to model a known physical glacier surface mass balance model. The model that the "AI surrogate model" will attempt to replicate is sourced from Pygem, with the inputs and outputs being the same as applicable(Rounce, 2020). This model will also serve as a foundation model as it is trained from all regions that the Pygem model was trained on and has generalizability. This foundation model is also a surrogate model as it learns the nuanced non-linearities of Pygem via deep machine learning. Users can take the output from the foundation-surrogate model and train to specific sub-regions to glean better accuracy. When tested, the foundation-surrogate AI model has a cross-validated RMSE of 0.77 and MAPE of 2.39%. When tested on the Greenland and the Central/Eastern Himalayas, which the model was not trained on, the model has a RMSE and MAPE of [1.88,1.40] and [5.86,4.79]% respectively which emphasizes the model's generalizability to other regions. The model was also tested on the Alps and Karakoram, which the model was trained partially on along with data from Alaska, Arctic North, and Patagonia. The RMSE and MAPE for these regions are [2.32,2.46] and [9.52,9.68]% respectively. Transfer learning can be applied by re-training the finer layers which yields a 21% improvement. This surrogate-foundation model has computational benefits and flexibility and is fully open sourced to be used by anyone interested.