Date of Award

12-2024

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Mathematics and Systems Engineering

First Advisor

Nezamoddin Nezamoddini-Kachouie

Second Advisor

Steven Lazarus

Third Advisor

Ryan White

Fourth Advisor

Joo Young Park

Abstract

Glaciers around the world have experienced a trend of recession within the past century. Quantification of glacier variations using satellite imagery is of great interest due to the importance of glaciers as freshwater resources and as indicators of climate change. The potential methods to quantify glacier variations with increasing complexity include detecting the terminus location, quantifying the glacier surface area, and measuring glacier volume. Although there are methods in literature designed purposefully for glacier area segmentation that have achieved acceptable results, they are often localized to the region where their training data were acquired and further rely on training sets that were often curated manually to highlight glacial features. It is practically impossible to perform a worldwide study of glacier dynamics using manual methods. Therefore, an automated or semi-automated method is highly desirable. The objectives of this dissertation are to build upon our previous works by improving upon the 2D glacier representation for glacier area segmentation, investigate the impact of patch size on deep learning glacier area segmentation, model glacier area with respect to relevant climate factors, and discuss the different annotated datasets of glacier area. We found that conventional image segmentation methods, edge detection and region growing, do not perform consistently for glacier area segmentation due to low contrast and signal-to-noise ratio. A deep learning method using the U-net was proposed to segment glacier area in the Aoraki massif in New Zealand. This model has demonstrated promising performance while trained on a relatively small dataset. The sensitivity of the model to patch size was assessed, as well as the difference between binary and multiclass classification. Models trained on larger patch sizes perform better than on smaller patch sizes. A post-processing denoising method using filtering methods is proposed to alleviate the segmentation inconsistencies along the stitching lines. In order to assess what model to use for modeling glacier area, the generalized additive model was used on glacier area data obtained from the methodology used in previous work and on climate data. Several labeled datasets are investigated and compared in order to further understand the annotated data scarcity for this problem.

Download

Included in

Mathematics Commons

Share

COinS