A Benchmarking Dataset for Arctic Ice Monitoring using Radar Satellite Images
The rapid monitoring of icebergs potentially crossing arctic shipping routes as well as long-term climate research issues have already led to several attempts to exploit the content of Synthetic Aperture Radar (SAR) satellite images. The European Sentinel-1 mission with its twin satellites allows fr...
| Main Authors: | , , |
|---|---|
| Format: | Conference Object |
| Language: | unknown |
| Published: |
2021
|
| Subjects: | |
| Online Access: | https://elib.dlr.de/144805/ https://phiweek.esa.int/detailed-programme |
| Summary: | The rapid monitoring of icebergs potentially crossing arctic shipping routes as well as long-term climate research issues have already led to several attempts to exploit the content of Synthetic Aperture Radar (SAR) satellite images. The European Sentinel-1 mission with its twin satellites allows free and rapid access to its carefully processed image data. As a consequence, the EU-funded research project ExtremeEarth has designed and implemented several highly automated and innovative explainable machine learning and deep neural network algorithms allowing us to train and operate an automated sea-ice monitoring system. In this respect, we analyse its most important innovative aspects and compare it with other approaches that have been developed by various nations. But first of all, in order to verify and implement efficient and innovative methods, we need sufficient datasets with high quality. Here in this contribution, we will present an efficient and very accurate routine based on active learning being able to generate sea-ice datasets annotated by several human experts providing their detailed knowledge with respect to local semantic image classification. Our selected target area is located near the North Pole, in the East part of Greenland. This part of Greenland is called Belgica Bank and was affected by a global warming effect between 2018 and 2019, when a high volume of ice was melted or transformed into water. For demonstration, we selected 36 images (one image per month acquired in 2018, 2019, and 2020) out of 183 available images for this area acquired by Sentinel-1A and Sentinel-1B. The Sentinel-1A/B images were tiled into patches of 256×256 pixels, and for each patch a semantic label was attached. In addition, a semantic classification map was generated for each analyzed Sentinel-1 image. The benchmarking dataset is presented as patches of 256×256 pixels grouped into 8 semantic classes. These classes are: Black image edge, Glaciers, Icebergs, Mountains, Old ice, First-Year ice, Young ice, and Water ... |
|---|