Glacial lake mapping and monitoring using SAR and optical satellite data

: development and applications of algorithms

Abstract

Glaciers around the world are retreating and thinning due to atmospheric warming resulting in the formation and expansion of glacial lakes. Glacial lakes can affect glacier mass balance and glacier dynamics. They are also a source of catastrophic glacial lake outburst floods (GLOFs). Remote sensing-based mapping and monitoring of glacial lakes are crucial for people living in the nearby glacierized catchments for the early detection of GLOFs with minimal costs. Although many remote sensing techniques already exist for mapping glacial lakes, the generic challenges of approaches include cloud cover, frozen glacial lake surface, turbidity, and shadows from mountains and clouds. Moreover, the remote sensing-based monitoring routines are of low frequency and mainly rely on optical remote sensing data. In this thesis, I have used multi-source remote sensing data from sensors such as Landsat, Sentinel-1 Synthetic Aperture Radar (S-1 SAR), Sentinel-2 Multispectral Instrument (S-2 MSI), and digital elevation models (DEMs) to develop, test, and validate glacial lake mapping and monitoring algorithms that address aforementioned challenges. When necessary, I have also used field photographs and high-resolution remote sensing data to validate results. I mainly focused on the Bhutanese Himalayas, and the alpine regions were also considered. The main approaches I used include threshold-based image segmentation, rule-based image segmentation, the Normalised Difference Water Indices (NDWIs), Persistent Scatterer Interferometry (PSI), and machine learning techniques (e.g., random forest).

Using S-1 SAR data, a semi-automated method for lake mapping based on a radar backscatter intensity threshold was presented for the Bhutanese Himalayas. The lake mapping results were accurate. However, the overall accuracy was low due to false positives requiring multiple post-processing steps to improve the accuracy. The main advantage of using S-1 SAR for lake mapping was the high temporal frequency of data acquisition and no cloud cover allowing repeated mapping of glacial lakes. Another glacial lake mapping method was proposed using S-1 SAR data, S-2 MSI data, rule-based image segmentation, and random forest classifier algorithm. The method was tested and validated in eight test sites across the alpine regions. It integrated the strengths of multi-source data and multiple techniques to map glacial lakes. The overall accuracy of the method was greater than 95% in all the test sites.

The time series S-1 SAR data, Google Earth Engine (GEE), and PSI technique were used to develop a method for monitoring glacial lakes at high frequency (every six days) in the Bhutanese Himalayas. The method was validated against the field photographs taken in 2016 and DEM differencing results generated using SPOT-7 and Pleiades data. The integrated use of GEE and time series S-1 SAR data improved hazard assessment techniques by quantifying glacier and basin melt areas, glacier and lake freeze-thaw cycle, and lake area and open water area changes. The PSI results revealed information about the stability of surrounding slopes and moraines of glacial lakes. Using the proposed approaches, an objective and systemic near-real-time mapping and monitoring of glacial lakes across alpine regions should be possible with certainty where field investigation is challenging due to rugged topography, geopolitics, and limited financial resources. Integrating knowledge, methods, tools, and techniques into the early warning systems by decision-makers and scientists is recommended.

Date of Award	30 Nov 2021
Original language	English
Awarding Institution	University of St Andrews
Supervisor	Tobias Bolch (Supervisor) & Doug I Benn (Supervisor)