Application of the Combined Feature Tracking and Maximum Cross-Correlation Algorithm to the Extraction of Sea Ice Motion Data From GF-3 Imagery

Mingci Li, Chunxia Zhou, Bing Li, Xiaoli Chen, Jianqiang Liu, Tao ZengIEEE JSTARS 2022 · DOI

This paper, published in the IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, details a significant step forward in polar remote sensing: the first successful application of a hybrid tracking algorithm to extract sea ice motion (SIM) from China’s Gaofen-3 (GF-3) satellite imagery.

TL;DR

Researchers have successfully adapted a hybrid "Feature Tracking and Maximum Cross-Correlation" (FT-MCC) algorithm to Gaofen-3 satellite data, filling a critical gap in global sea ice monitoring. By introducing a new "locally consistent" filtering method and identifying optimal imaging parameters (80m resolution in VH polarization), the team achieved sea ice tracking accuracy comparable to established international standards.

What Problem Does This Paper Address?

Monitoring how Arctic sea ice moves is vital for understanding climate change and ensuring safe maritime navigation. While we have several satellites capable of this, China’s Gaofen-3—a high-resolution C-band Synthetic Aperture Radar (SAR)—had not yet been utilized for automated sea ice motion extraction.

The technical challenge lies in the algorithms themselves. Traditional Maximum Cross-Correlation (MCC) is robust but computationally expensive and struggles with ice rotation. Feature Tracking (FT) is fast and handles rotation well but produces unevenly distributed data. Furthermore, standard methods for cleaning up "noisy" data usually rely on a fixed correlation threshold, which often fails because sea ice appearance changes drastically between summer and winter.

Methods

The authors employed a multi-stage FT-MCC pipeline:

  1. Initial Tracking: They used FT algorithms (specifically comparing A-KAZE and ORB) to find distinct "features" on the ice and track them between two images.
  2. Interpolation: These sparse, scattered vectors were interpolated onto a regular grid.
  3. Refinement: The MCC algorithm was then used to "fine-tune" these grid vectors, combining the speed of FT with the precision of MCC.
  4. Novel Filtering: The core innovation is the Locally Consistent (LC) flow field filtering. Instead of discarding vectors based on a simple quality score (which varies by season), the LC filter removes vectors that deviate too far from the average velocity of their neighbors. This "sanity check" is much more effective at preserving valid data in difficult conditions.

Datasets

  • Primary Data: Gaofen-3 (GF-3) Level-1A images in Fine Stripe II (FSII) mode.
  • Validation Data: Sentinel-1 SAR images (the current industry standard).
  • Reference Products: CMEMS (Copernicus Marine Environment Monitoring Service) SIM data.
  • Ground Truth: Manually extracted SIM vectors to serve as a baseline for "perfect" tracking.

Key Insights and Ideas

  • The "Goldilocks" Resolution: Counter-intuitively, the highest resolution isn't always best. At 20m resolution, "speckle noise" (the grainy salt-and-pepper look of radar images) overwhelmed the features. The researchers found that 80m resolution provided the best balance, offering enough detail for tracking while suppressing noise.
  • VH Polarization Wins: The study compared VH (vertical transmit, horizontal receive) and VV (vertical transmit, vertical receive) polarizations. VH images consistently provided more reliable vectors. The authors used Information Entropy to prove why: VH images contain "richer" grayscale information, essentially giving the algorithm more unique textures to grab onto.
  • Algorithm Efficiency: While both A-KAZE and ORB versions of the algorithm produced similar accuracy, A-KAZE was significantly faster, making it the preferred choice for processing large volumes of satellite data.

Key Results

The FT-MCC algorithm proved highly accurate when applied to GF-3 data:

  • Velocity Uncertainty: 0.119 cm/s to 0.287 cm/s (roughly 100 to 250 meters per day).
  • Directional Uncertainty: 4.119° to 5.930°.
  • Consistency: The results were highly consistent with those derived from Sentinel-1, proving that GF-3 can serve as a reliable independent source for global sea ice monitoring.

Limitations and Future Work

The main limitation is the geographic and temporal scope. The study focused on specific regions of the Arctic during selected winter and summer windows. Because sea ice characteristics vary wildly across the entire Arctic basin, the authors note that their conclusions might not yet be universal.

Future Work will involve:

  • Expanding the testing to cover the entire Arctic.
  • Utilizing the growing constellation of GF-3 satellites to improve the "temporal resolution" (how often we get a fresh look at the ice).
  • Refining the algorithm to handle even more complex ice deformation zones where the "locally consistent" assumption might be challenged.

Summary updated

Abstract

In this study, an algorithm combining feature tracking and maximum cross-correlation (FT-MCC) for the extraction of sea ice motion (SIM) vectors were applied to Gaofen-3 (GF-3) imagery, filling the gap of SIM extraction using GF-3 imagery. The locally consistent flow field filtering method is proposed to replace the filtering method based on the correlation coefficient threshold in FT-MCC to improve filtering effectiveness of SIM results extracted by FT-MCC. A comparison of the probability density distributions (PDDs) of the correlation coefficients of SIM vectors extracted by FT-MCC from images with different resolutions revealed high reliability for SIM vectors extracted for images with an 80 m spatial resolution. A comparison of the PDDs of the correlation coefficients of SIM vectors obtained from images with different polarization modes showed more reliable SIM vectors were extracted from vertical transmit horizontal receive (VH) polarization images than from corresponding vertical transmit vertical receive (VV) polarization images. The SIM vectors extracted from GF-3 images by two methods (FT(A-KAZE)-MCC and FT(ORB)-MCC) derived from the FT-MCC algorithm were highly consistent in terms of accuracy and reliability. SIM vectors extracted manually and from Sentinel-1 images were used as reference data to verify the SIM results extracted from GF-3 images, for which the uncertainties in the magnitude and direction of the extracted SIM vectors were found to be 0.119 cm/s–0.287 cm/s (103 m/d–248 m/d) and 4.119°–5.930°, respectively.
BibTeX
@article{li2022application,
  author = {Mingci Li and Chunxia Zhou and Bing Li and Xiaoli Chen and Jianqiang Liu and Tao Zeng},
  title = {{Application of the Combined Feature Tracking and Maximum Cross-Correlation Algorithm to the Extraction of Sea Ice Motion Data From GF-3 Imagery}},
  journal = {IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing},
  year = {2022},
  doi = {10.1109/jstars.2022.3166897},
}
Figure