Day 4

Detailed paper information

Back to list

Paper title Towards a circumpolar coastline characterisation based on Sentinel-1/2 and machine learning
  1. Aleksandra Efimova b.geos Speaker
  2. Annett Bartsch b.geos GmbH, AT
  3. Barbara Widhalm b.geos
  4. Goncalo Vieira IGOT - Universidade de Lisboa
  5. Sofia Bilbao
Form of presentation Poster
  • C1. AI and Data Analytics
    • C1.04 AI4EO applications for Land and Water
Abstract text Arctic regions are one of the most rapidly changing environments on earth. Especially Arctic coastlines are very sensitive to climate change. Coastal damages can affect communities and wildlife in those areas and increased erosion leads to higher engineering and relocation costs for coastal villages. In addition, erosion releases significant amounts of carbon, which can cause a feedback loop that accelerates climate change and coastal erosion even further. As such, a detailed examination of coastal ecosystems, including shoreline types and backshore land cover, is necessary.
High spatial resolution datasets are required in order to represent the various types of coastlines and to provide a baseline dataset of the coastline for future coastal erosion studies. Sentinel-2 data offer good spatial and temporal resolution and may enable the monitoring of large areas of the Arctic. However, some relevant classes have similar spectral characteristics. A combination with Sentinel-1 (C-band SAR) may improve the characterization of some flat coastal types where typical radar issues such as layover or shadow do not occur.
This study is comparing a Sentinel-1/2 based tundra landcover classification scheme, which is developed for full pan-Arctic application, with another landcover classification, created specifically for mapping Arctic coastal areas (Sentinel-2 only). Both approaches are based on machine learning using a Gradient Boosting Machine. The Arctic Coastal Classification is based on Sentinel-2 data and considers 12 bands with 5 target classes while the Sentinel-1/2 based tundra landcover classification scheme is based on five Sentinel-2 bands (temporally averaged) and Sentinel-1 data acquired at VV polarization and results in more than 20 classes.
Results show that even the best classification algorithms show limitations at specific coastal settings and sea water conditions. The analysis demonstrates (1) the need for a coastal specific classification in this context, (2) the need for date specific mapping, but consideration of several acquisitions to capture general coastal dynamics, (3) the potential of detailed Arctic landcover mapping schemes to derive subcategories and (4) the need to separate settlements and further infrastructure.