Regional Scale Generalizations of Firn Thickness and Snow Accumulation in Southeast Alaska: an Applied Deep Learning Framework
High elevation mountain glaciers are referred to as the “water towers of the world” due to their ability to store water internally and release it later through melt and runoff. Much of the work to understand these complex environments has focused on cold based polar glaciers in high latitude regions...
| Main Author: | |
|---|---|
| Format: | Text |
| Language: | unknown |
| Published: |
DigitalCommons@UMaine
2022
|
| Subjects: | |
| Online Access: | https://digitalcommons.library.umaine.edu/etd/3741 https://digitalcommons.library.umaine.edu/cgi/viewcontent.cgi?article=4806&context=etd |
| Summary: | High elevation mountain glaciers are referred to as the “water towers of the world” due to their ability to store water internally and release it later through melt and runoff. Much of the work to understand these complex environments has focused on cold based polar glaciers in high latitude regions, neglecting temperate glaciers and associated snowpack which are more responsive than polar systems to minor climatic changes. Advancements in the availability of open source data, deep learning theory, and computational efficiency has opened up new methods of “data hungry” modeling that may be well suited to the type of complex terrain-climate interactions common in these environments. Through the use of regression based deep neural networks (DNNs), this study explores the applicability of this method on a 400 MHz ground penetrating radar dataset collected on the Juneau Icefield between 2012 and 2021. The primary task of these networks is to find the statistical relationship between englacial layer thickness and a collection of topographic and climatological data with the goal of upscaling measurements from radar transects to the entire region. These models are separated into two primary tasks, the prediction of firn thickness, and the prediction of annual accumulation thickness. These tasks are approached using two different data structures, one that uses only a single year of data, and one that uses all available years of data. Models are validated using the standard “leave on glacier out (LOGO) and “leave one year out” (LOYO) holdout methods depending on the data structure. Results indicate that the available data for the Juneau Icefield is better suited for the prediction of annual accumulation thickness rather than firn thickness, with much better regional scale models produced by the multiyear data structure. Through this method, realistic simulations of icefield scale firn and snowpack distributions are generated and can be paired with existing models to output estimations of key environmental characteristics ... |
|---|