Improving Antarctic Topography: Utilizing Pre-Trained Models for Rock Outcrop Updates

The Antarctic Digital Database (ADD) is a comprehensive compilation of the best available topographic data for Antarctica, serving as an essential tool for researchers navigating and understanding the continent’s ever-changing landscape. Key layers, such as the Antarctic coastline, are regularly updated through visual control and manual editing. In contrast, other data layers, like Rock Outcrop, are updated only upon request, typically during the creation of maps for specific regions.

Antarctic topographic mapping primarily relies on remote sensing data rather than ground surveys, which is unlike most populated areas of the world. Monitoring changes in polar regions is crucial for understanding global climate change. Therefore, the increasing use of pre-trained foundation models based on remote sensing data is expected to be beneficial for Antarctic mapping.

This study aimed to utilize models pre-trained on large satellite datasets to update the Rock Outcrop layers using a small amount of training data. The current Rock Outcrop layer, generated in 2016 based on Landsat 8 imagery, has become outdated.

For this work, a Sentinel-2 mosaic of averaged, mostly cloud-free images from the 2023/2024 Antarctic summer season was generated using Google Earth Engine for the British Antarctic Territory (BAT). The mosaic was exported as a tiled raster image, consisting of 4,990 chips. Rock outcrop labels were created for 20 tiles, evenly spread across the BAT.

The ResNet18 model, pre-trained on the SSL4EO-S12 dataset with Sentinel-2 RGB MOCO weights, was trained, resulting in an F1 score of 0.91. To achieve proper cartographic representation, the generated predictions underwent a generalization process. The results will be published as part of the next ADD update.

The updated layer shows a 28% increase in the area of exposed rock compared to the 2016 layer. This significant change could be attributed to both the data and methods used, as well as actual changes in snow coverage in the study area. Further updates, which now require minimal effort to implement, may help explain the observed dynamics.