IRGS (Iterative Region Growing With Semantics) is a computer vision algorithm designed for region-based scene segmentation and classification. MAGIC (MAp-Guided Ice Classification) is the software system where IRGS and other algorithms are implemented.

MAGIC was originally created to read and classify SAR (synthetic aperture radar) images provided by the Canadian Ice Service (CIS) along with associated ice maps. IRGS provides a pixel-level classification of each polygon (i.e. subregion) knowing only the number of ice types in that polygon. For each polygon, the CIS analyst assigns the ice type concentrations, classes, and floe sizes and stores this information in an “Egg Code”. So, although a polygon may be assigned 40% multi-year ice, the map does not pinpoint the location of the multi-year ice in the polygon. For the image on the right, the polygon of interest is outlined in red; its associated egg code shows that it contains four different classes of ice. IRGS takes this information to generate a pixel-level classification where each pixel can be assigned a particular ice type.

MAGIC uses one of many provided classification algorithms (e.g., IRGS) and overlays the results on the image. A sample classification is shown below – note that this image is a portion of a much larger original image and had to be shrunk by a factor of 4 for display purposes. IRGS is a Markov Random Field (MRF) based approach which has been shown to perform significantly better classifications of SAR sea ice imagery compared to other algorithms. Multiple classification results from different algorithms can be saved for comparison.

How does IRGS work? IRGS begins by performing an oversegmentation of the polygon or image using a watershed algorithm.  Using a Markov model based on Gaussian statistics and inter-region edge strength, the resulting regions are then iteratively (a) classified knowing the number of ice types in the scene and then (b) merged in a greedy fashion based on energy minimization.  A more detailed description of IRGS and how it has been extended to multi-band and polarimetric data can be found in the journal papers listed at the bottom of the page.  An animation showing regions being merged by IRGS is found below.

The image on the right below shows the results of classifying the polygon above using IRGS. Within MAGIC the overlay can be turned on and off and the transparency adjusted to compare the classification result with the underlying image.

MAGIC v2.0 and the associated classification algorithms are still under active development.
The images have been significantly down-scaled for use in this article. Do not use them for research purposes without permission.

RADARSAT–2 Data and Products © MacDONALD, DETTWILER AND ASSOCIATES LTD. (2004–2008) – All Rights Reserved “RADARSAT” is an official mark of the Canadian Space Agency.SAR Images in this article are provided by the Canadian Ice Service.

Related people

Directors

Students

Alumni

Related research areas

Multiresolution Techniques

Remote Sensing

Related publications

Theses

Leigh, S., “Automated Ice-Water Classification using Dual Polarization SAR Imagery”, Department of Systems Design Engineering, Waterloo, ON, Canada, University of Waterloo, pp. 110, 2013.

Bandekar, N.., “Illumination and Noise-Based Scene Classification – Application to SAR Sea Ice Imagery“, Department of Systems Engineering, Waterloo, ON, Canada, University of Waterloo, pp. 77, 2012. Get it here.

Clausi, D. A., “Texture Segmentation of SAR Sea Ice Imagery”, Department of Systems Design Engineering, Waterloo, Ontario, Canada, University of Waterloo, pp. 176, 1996.