Data Reduction Using Statistical and Regression Approaches for Ice Velocity Derived by Landsat-8, Sentinel-1 and Sentinel-2

Abstract

During the last decade, the number of available satellite observations has increased significantly, allowing for far more frequent measurements of the glacier speed. Appropriate methods of post-processing need to be developed to efficiently deal with the large volumes of data generated and relatively large intrinsic errors associated with the measurements. Here, we process and combine together measurements of ice velocity of Russell Gletscher in Greenland from three satellites—Sentinel-1, Sentinel-2, and Landsat-8, creating a multi-year velocity database with high temporal and spatial resolution. We then investigate post-processing methodologies with the aim of generating corrected, ordered, and simplified time series. We tested rolling mean and median, cubic spline regression, and linear non-parametric local regression (LOWESS) smoothing algorithms to reduce data noise, evaluated the results against ground-based GPS in one location, and compared the results between two locations with different characteristics. We found that LOWESS provides the best solution for noisy measurements that are unevenly distributed in time. Using this methodology with these sensors, we can robustly derive time series with temporal resolution of 2–3 weeks and improve the accuracy on the ice velocity to about 10 m/yr, or a factor of three compared to the initial measurements. The presented methodology could be applied to the entire Greenland ice sheet with an aim of reconstructing comprehensive sub-seasonal ice flow dynamics and mass balance.

Keywords

• ice velocity
• time series
• post-processing
• data reduction
• non-parametric regression
• multi-sensor data