Conference,
Towards the global Quantification of Permafrost Loss along Arctic Coasts
(May 2022)
DOI: http://dx.doi.org/10.13140/RG.2.2.36303.36002/1
Abstract
Roughly one quarter of exposed land on the Northern Hemisphere and more than 65% of terrestrial area above 60°N are underlain by permafrost. Permafrost, defined as ground material which remains frozen for at least two consecutive years, stores around 1460-1600 billion tonnes of organic carbon, roughly twice the amount of carbon in the atmosphere. The climate change driven degradation of permafrost has major implications for the environment and could potentially turn frozen ground from a carbon sink to a carbon source that consequently could cause trillions of dollars in global economic damage. A widespread process and linked to the degradation of permafrost is Arctic coastal erosion. Roughly 30-34% of coastlines on the Earth are affected by permafrost. Erosion processes of permafrost-coasts cause a release of carbon to the oceans, alter fish and wildlife habitats, force changes in the Arctic ecosystem and endanger human settlements and infrastructure. In order to fully assess the implications of eroding permafrost-coastlines, a continuous quantification of the erosion rates, the associated loss of permafrost and the stored carbon is required at high spatial resolution for the entire Arctic. In this context, spaceborne earth-observation is a suited tool to map and monitor changes of the Arctic coast remotely with high confidence; however, frequent cloud cover causes data gaps in optical imagery and thus limits the usability, repeatability, and transferability. In contrast, Synthetic Aperture RADAR (SAR) imagery is capable of overcoming such limitations thanks to its ability to gather surface information independent of sun illumination or weather condition. Here we present a novel approach for the automated mapping of annual Arctic coastal erosion rates based on Sentinel-1 SAR scenes. The method employs deep learning and Change Vector Analyses (CVA). Results are presented for ten test sites across the Arctic covering about 1830 km of Arctic coastline. In a first step, a reference coastline for the year 2020 was computed via a deep learning workflow. Nine different U-Net architectures were combined to generate a high-quality coastline product which acted as a reference for the erosion-rate quantification. The U-Net architectures were pre-trained based on the ImageNet database (14. Mio images) and further trained with about 185.000 Sentinel-1 Pseudo-RGB scenes, which were generated from annual median and annual standard deviation VV/VH backscatter. Consequently, CVA was executed on these spatiotemporal metrics to estimate the erosion rates. On average, the standard deviation proved to be higher for water, whereas median backscatter was higher on land; thus, a change from high median backscatter to high standard deviation backscatter can be interpreted as a change from land to water. Furthermore, working on annual composites instead of single scenes significantly reduced the speckle effect and the geolocation uncertainty. Our analysis revealed erosion rates of up to 90 m/year for some areas across the investigated Arctic coastline. The approach can be applied for very large scales, prospectively even for the entire Arctic. Eventually, the generated products may act as a means to quantify the loss of frozen ground, the release of stored carbon stocks, and might be useful for further analyses related to changes in Arctic coastal environments.
