Unsupervised Iceberg Detection in SAR Satellite Images Using Recurrent Principal Component Analysis
Abstract
The shape and distribution of icebergs and their trajectory through oceans from source to where they break up and melt can be diagnostic of ice-sheet, ocean and atmospheric dynamics. Existing methods of iceberg tracking are optimised for open ocean and work less well for icebergs within sea ice. The study (funded by the Alan Turing Institute) aims to reveal pattern and changes in iceberg populations and their trajectories through the coastal seas.
We have developed a novel unsupervised greedy algorithm to detect icebergs in Synthetic Aperture Radar (SAR) satellite images with horizontal-horizontal (HH) and horizontal-vertical (HV) polarisation in near coastal regions as well as in open water. The algorithm employs solely pixel intensity values. Novelty lies in the use of recurrent Principal Component Analysis (PCA) combining HH and HV to identify the parametric direction of highest variability in the data. Along this direction, Dirichlet Process (DP) clustering identifies clusters of similar data values, typically resulting in three clusters: open water, sea ice and icebergs. Cluster assignments with high probabilities are accepted, and PCA applied again to the remaining data giving a new direction of highest variability; this approach can be compared to looking at the data from a different 'parametric angle'. The algorithm proceeds in an iterative manner until almost all the data are clustered, and we have a heat map of pixels identified as icebergs. The advantage of this approach is that there are no fixed thresholds determining cluster membership. It is entirely determined from the data. Thus the computational complexity is very low, making it applicable to the scales and rates at which data are generated. Moreover, the generated heat maps can be used as input to supervised learning algorithms. Some post-processing is necessary to split adjacent icebergs or use convolution to resolve shadows caused by ridges or crevasses. We present our initial results and manual validation. The next steps are generating labelled training data from this unsupervised approach, and tracking icebergs between successive images to recreate their trajectories.- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2020
- Bibcode:
- 2020AGUFMC004.0005F
- Keywords:
-
- 0555 Neural networks;
- fuzzy logic;
- machine learning;
- COMPUTATIONAL GEOPHYSICS;
- 0758 Remote sensing;
- CRYOSPHERE;
- 1910 Data assimilation;
- integration and fusion;
- INFORMATICS;
- 1942 Machine learning;
- INFORMATICS