Date of Award
Spring 2024
Document Type
Thesis (Master's)
Department or Program
Biochemistry and Cell Biology
First Advisor
Robert Hawley
Abstract
High Mountain Asia (HMA) is home to the largest mass of glaciers and ice outside the north and south polar regions. HMA glaciers are projected to experience accelerated mass loss from higher greenhouse gas emissions through the end of the century. Many studies of glacier mass balance and mass loss in HMA obtain glacier area from the Randolph Glacier Inventory (RGI). However, the RGI is designed to show glacier area across the world that is accurate to the year 2000 and, as a result, is not an accurate representation of the current state of glacier area in HMA. Additionally, glacier outlines in HMA are delineated using various manual and semi-automated mapping methods which are labor- and time-intensive making them difficult to repeat on a large spatial or temporal scale.
To address this issue, I developed an automated method to classify glacier area, leveraging machine learning via a random forest classifier model. I used remotely sensed data including Landsat-7 Enhanced Thematic Mapper Plus multispectral imagery, a digital elevation model, and glacier velocity data to train the model. I tested two different geographic regions, the Karakoram Mountains and the Pamir Alay Mountains, for model training and accuracy. Within each region, the model was trained and tested on different-sized tiles. Overall, the model trained in the Karakoram Mountains more accurately identified glacier cover than the model trained in the Pamir Alay Mountains. Test 05a, in the Karakoram Mountains with a tile size of 0.75° by 0.5°, was the best at classifying glacier cover based on its ice-covered classification results compared to the other tiles and had an accuracy score of 0.9615. The most important feature for identifying glacier cover in test 05a was glacier velocity.
I then applied the trained model to classify glacier area across the Tian Shan Mountains using satellite imagery from two time periods, 1999-2010 (Epoch 1) and 2015-2023 (Epoch 2) to assess change in glacier area since around the year 2000. I calculated a 10.1% decrease in glacier area from 10,183 ± 601 km2 in Epoch 1 to 9,156 ± 540 km2 in Epoch 2. Despite some persistent areas that the model misclassified as glacier-covered that were not glaciated, the model results overall appeared to agree well with the RGI polygons.
Recommended Citation
Halvorson, Victoria Elizabeth, "Automated Glacier Classification in High Mountain Asia using Machine Learning and a Random Forest Classifier" (2024). Dartmouth College Master’s Theses. 154.
https://digitalcommons.dartmouth.edu/masters_theses/154