Autonomous Repeat Image Feature Tracking (autoRIFT) and Its Application for Tracking Ice Displacement

Abstract

In this paper, we build on past efforts with regard to the implementation of an efficient feature tracking algorithm for the mass processing of satellite images. This generic open-source feature tracking routine can be applied to any type of imagery to measure sub-pixel displacements between images. The routine consists of a feature tracking module (autoRIFT) that enhances computational efficiency and a geocoding module (Geogrid) that mitigates problems found in existing geocoding algorithms. When applied to satellite imagery, autoRIFT can run on a grid in the native image coordinates (such as radar or map) and, when used in conjunction with the Geogrid module, on a user-defined grid in geographic Cartesian coordinates such as Universal Transverse Mercator or Polar Stereographic. To validate the efficiency and accuracy of this approach, we demonstrate its use for tracking ice motion by using ESA’s Sentinel-1A/B radar data (seven pairs) and NASA’s Landsat-8 optical data (seven pairs) collected over Greenland’s Jakobshavn Isbræ glacier in 2017. Feature-tracked velocity errors are characterized over stable surfaces, where the best Sentinel-1A/B pair with a 6 day separation has errors in X/Y of 12 m/year or 39 m/year, compared to 22 m/year or 31 m/year for Landsat-8 with a 16-day separation. Different error sources for radar and optical image pairs are investigated, where the seasonal variation and the error dependence on the temporal baseline are analyzed. Estimated velocities were compared with reference velocities derived from DLR’s TanDEM-X SAR/InSAR data over the fast-moving glacier outlet, where Sentinel-1 results agree within 4% compared to 3–7% for Landsat-8. A comprehensive apples-to-apples comparison is made with regard to runtime and accuracy between multiple implementations of the proposed routine and the widely-used “dense ampcor" program from NASA/JPL’s ISCE software. autoRIFT is shown to provide two orders of magnitude of runtime improvement with a 20% improvement in accuracy.