Image analysis, ice floes and other such wonderful stuff

In 2005 I wrote an honours thesis at the Institute of Southern Ocean and Antarctic Studies.

Luckily it went pretty well, and the results were presented at the 2005 IGS International Symposium on Sea Ice, in Dunedin, New Zealand. The results were also presented at the SCAR open science conference in 2006.

Now, I'm working on another project in the same vein - attempting to build a tool that determines the amount of rough and smooth ice present in an image. Yep - machine vision... arrr! Hopefully, we'll be able to plug the output into our floe size algorithm, overcoming one of the major issues with identifying floes!

As if by magic...

Long days and long nights have led to a partial implementation of a split-and-merge image segmentation algorithm using a Local Binary Pattern operator [LBP] and grey level [contrast] variance [VARc] to do the dirty work of determining rough and smooth ice areas.

It is a partial implementation for many reasons, some of which are my inexperience as a programmer, and others which are more like 'we have found a shortcut that works OK for us, for now'. In essence, using information derived from neighbourhoods around each image pixel, we do the first part of the job - splitting the image into roughly homogeneous blocks - and assign a loose classification to each block based on some knowledge about the distribution of grey levels and texture in each.

While this doesn't give us discrete objects, and given the coarse resolution [minimum 4x4 pixel blocks], isn't so hot for using as an input for floe size distribution processing, we do get a pretty good idea, with maybe 10% error, about the contribution of rough and smooth snow to the total area of snow-covered ice in an image.