Analysing feature-tracking methods to derive flow rates at John Evans Glacier, Canadian High Arctic

 

 

William D. Harcourt

 

Introduction

 

Recent climatic change in the Arctic has been amplified due to the reduction in sea ice extent, which accelerates the positive ice-albedo feedback, and enhances warming in the northern hemisphere high latitudes. In the Queen Elizabeth Islands (QEI), which contain 14% of the ice covered area outside of Greenland and Antarctica, accelerated mass loss has been attributed to a persistently negative mass balance. While this is mainly attributed to the dynamic acceleration of marine-terminating glaciers (van Wychen et al., 2016; Harcourt et al., in review), while significant surface melt in response to elevated summer temperatures has been observed, affecting the flow rates of land-terminating glaciers. However, the mechanisms responsible for the changes observed across the region remain unclear, and there is a pressing need for more robust methods to derive ice flow estimates. Quantifying the velocity of glaciers is a first order constraint on understanding changes in ice discharge, as well as long- and short-term changes in ice dynamics. Estimating glacier velocity has been undertaken using both optical and Radar satellites, so that velocity results can be obtained on both inter-annual and intra-annual timescales. Specifically, Synthetic Aperture Radar (SAR) imagery can be used to track the movement of Radar backscatter, while optical satellites can be used to measure the displacement of features on the ice surface using image matching techniques.

 

Key Objectives

 

Given the need for further systematic comparison of available feature-tracking techniques, this study had three key objectives:

1)      Analyse whether a standardised method image pre-processing can be implemented to derive ice flow on a slow-moving glacier.

2)      Compare and contrast the effect of automated pre-processing on algorithm choice.

3)      Investigate long-term changes in ice dynamics at a land-terminating glaciers in the Canadian High Arctic

 

Methods and Study Site

 

John Evans Glacier (JEG) covers 165 km2 of ice, 0.15% of the total ice covered area across the QEI (Fig. 1). This site was chosen because of the availability of field velocity from 2000 (Bingham et al., 2003), which is used as a validation dataset for the feature-tracking analysis.

 

Text Box: Fig. 1 Location map of John Evans Glacier (JEG).Deriving an estimate of surface displacement is undertaken by measuring the degree of similarity between groups of pixels (patches) across two images of a defined temporal baseline (𝐼1 and 𝐼2), and the local peak in the correlation matrix (derived from the cross correlation algorithm) is used to estimate displacement (Fig. 2).

 

Three algorithms were chosen for the comparison based on the use in the literature (Table 1). Each of these implement one or more of the following image matching methods; Oriented Correlation (OC), Normalised Cross-Correlation (NCC), and Phase Correlation (PC).

 

Image pre-processing followed the automated workflow of Dehecq et al. (2015), which utilise all available Landsat images of JEG to derive an ice flow estimate. Because of the slow flow rates of JEG (22 m yr-1), only the 15 m panchromatic band was used, as this was smaller than the assumed maximum displacement over 1 year.

 

Following on from the automated processing method of Dehecq et al. (2015), image pairs are constructed from the same path and row Text Box: Fig. 2 Feature-tracking overview.combinations, and the feature-tracking is then computed on each image pair in order to produce a velocity stack. A median filter is then computed on the velocity stack in space and time in order to merge the individual velocity results, which generates a conservative estimate of glacier flow which is suitable given the small strain rates experienced at JEG.

Text Box: Table 1 Algorithms used in the study.

Algorithm

Description

Image Georectification and Feature Tracking Toolbox (ImGRAFT)

Matlab script which is fully automated in the processing chain of Dehecq et al. (2015), but untested within the literature.

Co-registration of Optically Sensed Images and Correlation (COSI-corr)

Plug-in for ENVI, and has been shown to perform the best compared to other algorithms.

Image Cross-Correlation Software (CIAS)

User-friendly interface, computing velocity results using vector points, thus requiring more post-processing.

 

 

Results and Discussion

The success of each algorithm was assessed based on three criteria; the RMSE with respect with the field results (Bingham et al., 2003), the movement over assumed stable bedrock, and the coherence of the velocity results (Table 2).

Text Box: Table 2 Success rate of each algorithm.

Algorithm

RMSE

Average Deviation over Bedrock

Coherence

COSI-corr

39.05

1.61

0.19

ImGRAFT NCC

47.34

6.92

0.26

ImGRAFT OC

40.06

4.5

0.24

ImGRAFT PC

26.26

3.03

0.28

CIAS NCC

59.5

5.03

N/A

CIAS OC

24.86

4.12

0.36

 

It is shown that the ImGRAFT PC algorithm (automated within Dehecq et al., 2015), works best, however the CIAS OC method achieves the best RMSE with respect to the field results. However, COSI-corr produces less deviation over bedrock, suggesting that the algorithm has little background noise. Regression analysis (not shown here)  of the field and feature-tracking results show that each merged velocity result underestimates true ice surface displacement, which may be an artefact of a median fusion filter on a small set of image pairs. The ImGRAFT PC method was thus used to derive an estimate of 2013/14 velocity (Fig. 3), showing a decrease in velocity by 1.73 m yr-1 over the 14 year period. This is concurrent with a reduction in ice thickness as shown by airborne geophysical surveys, suggesting that ice flow is reducing as a result of reduced driving stress, as well as being affected by surface melt due to the spatially coherent flow regime (Fig. 3).

 


Text Box: Fig. 3 Velocity change between 2000 and 2013/14.

 

Conclusion

 

The results drew the following conclusions:

·         An automated pre-processing chain is sensitive to algorithm choice, however individual steps are not sensitive to the parameters used.

·         Algorithms using Fourier-based methods (COSI-corr and ImGRAFT PC) achieve better results at JEG most likely due to the small changes observed.

·         JEG shows a slowing down of ice displacement, both in response to surface melt and reduce ice discharge.

 

Future studies should consider the following:

·         Understand the local environmental conditions, and apply algorithms based upon pre-determined displacement measurements

·         Understand the changing nature of land-terminating glaciers, and their sensitivity to climate-induced changes.