Trends in Antarctic Ice Speed 2014-2023 From Big Data Processing of Satellite Observations

The speed at which the Antarctic Ice Sheet (AIS) flows from the continental interior to the ocean is a key indicator of its stability, and satellite-derived ice velocity measurements play a major role in our assessment of changes in ice dynamics. The Sentinel-1 constellation of synthetic aperture radar (SAR) satellites, part of the European Commission’s Copernicus program, has acquired repeat images of the AIS margins at a combination of 6 and 12-day intervals since 2014, leading to a dramatic improvement in the spatio-temporal resolution and coverage of key velocity measurements at the edge of the AIS.

Such modern Earth observation satellites provide ever increasing volumes of data which can be used to study changes over time; however, this growing archive poses two key issues when performing timeseries analysis at continental scale. Firstly, generation of dense, pixelwise time series from thousands of successive observations, each stored in separate files corresponding to observation date, can require extremely large numbers of file reads, limiting computation speeds and increasing the memory required for data handling. Secondly, with ever increasing volumes of data, analysis must be able to scale effectively and easily handle out-of-core computation where datasets are larger than the available memory.

In this study, we present results from an analysis pipeline built on the Xarray and Dask python packages, and deployed on a HPC service, which allows both large scale interactive analysis in Jupyter notebooks as well as traditional batch processing. We first use an ice velocity processing chain to generate Antarctic-wide mosaics of ice speed on a 100 m grid for each combination of 6 and 12-day Sentinel-1 repeat observation dates, using the GAMMA Remote Sensing software to derive ice displacements from offset tracking of the SAR image pairs. The resulting stack of 2-dimensional mosaics is then restructured into a 3-dimensional data cube with dimensions x, y, time to facilitate time series analysis, overcoming the issue of excessive file reads and memory requirements by storing chunks of time series data together using the Zarr storage format.

Using this pipeline we investigate trends in ice speed across the AIS, performing time series outlier removal on 11 billion time series and subsequently calculating linear rates of change across both grounded and floating ice during the study period. We present the resulting map of ice speed trends and highlight time series of notable individual outlet glaciers and ice shelves.