Date of Award
Doctor of Philosophy
Large-scale study of glaciers improves our understanding of global glacier change and is imperative for monitoring the ecological environment, preventing disasters, and studying the effects of global climate change. In recent years, remote sensing imagery has been preferred over riskier and resource-intensive field visits for tracking landscape level changes like glaciers. However, periodic manual labeling of glaciers over a large area is not feasible due to the considerable amount of time it requires while automatic segmentation of glaciers has its own set of challenges. Our work aims to study the challenges associated with segmentation of glaciers from remote sensing imagery using machine learning, improve on the performance of existing methods using deep learning techniques, and interpret the working mechanism of these deep learning models.
In this dissertation, several machine learning and deep learning techniques were used to delineate glaciers from Landsat-7 imagery and we observed that the U-Net based model outperformed the other methods. While the methods used in this research are generalizable across all alpine glaciers, we evaluate our performance on the glaciers in the Hindu Kush Himalayas(HKH) as the HKH is one of the world's most sensitive region for climate change. Alpine glaciers, as the ones seen in HKH, have clean ice/snow surface where they form. As these Clean Glacial Ice (CIG) move down the valleys, they sometimes gather a significant covering of dirt, rocks, and boulders on their surface and are known as Debris Glacial Ice (DGI). As expected, our experimental results verify that segmenting DGI is significantly harder than segmenting CIG. To improve the performance on DGI, we introduce a novel Self-Learning Boundary-Aware loss (L_SLBA). L_SLBA combines masked dice loss and boundary loss to simultaneously learn multiple objectives during the training process for improved performance. Experimentally, we show that L_SLBA outperforms the commonly used dice loss for DGI mapping. We also propose feature-wise saliency scores to quantify the contributions of each channel in the input image towards the final label. This can help identify which features are most important in the context of glacier mapping and has the potential to change the way people view glaciers and the features associated with them, leading to a better understanding in monitoring them.
A limitation of U-Net based model is the need for densely labeled training data where a label is assigned to every pixel within the image. Due to the time and effort associated with creating dense labels, having access to training data remains a limiting factor in many applications of landcover mapping. To solve this problem, we introduced a technique to train the U-Net based model using sparsely labelled data where only some pixels for a given image are labelled. Unlike dense labels, sparsely labelled samples can be collected in large numbers in a relatively fast and cheap manner and often without the need for an expert. We used the technique to segment water bodies in the Arctic National Wildlife Refuge (ANWR) using the 4 Band Orthorectified NOAA Airborne Imagery and sparse training labels. This also shows that the methods used in this research can be used to segment geomorphological features on the Earth's surface other than glaciers and across different types of satellite and overhead imagery for mapping.
In recent years, large volumes of public freely-available large-scale satellite images have been made available. However, there exists a knowledge gap on accurate pixel level understanding of what is going on in these images. In this dissertation, we present multiple methods for segmentation of geomorphological landscape features from overhead imagery. These methods can have major implications in understanding global challenges such as climate change and anthropogenic impacts to ecosystems (i.e. deforestation, urbanization, land use change).
Received from ProQuest
Aryal, Bibek, "Glacier Segmentation From Remote Sensing Imagery Using Deep Learning" (2022). Open Access Theses & Dissertations. 3645.
Computer Sciences Commons, Environmental Sciences Commons, Geology Commons, Remote Sensing Commons