Friday, 12 June 2015

Revisiting my dissertation work on Mars

In my MSc dissertation at Aberystwyth University, I wrote about mid-latitude glaciers on Mars. I linked the 1309 candidate mid-latitude Martian glaciers published by Colin Souness with topographic data from Mars Express, and made derived topographic layers in GDAL and LandSerf. I made a layerstack of Mars Express images, topography and derived layers and segmented it in RSGISLib, and did some statistical processing to create a kind of naive Bayes classifier to indicate which areas have similarity in topography to the Souness glaciers.

Due to time and space constraints, I did not actually present full results on all of the 179 HRSC tiles within the dissertation. I had intended to do some more afterwards, but at one point my external hard drive failed and I had to re-download some of the data.
A figure from my dissertation with colourized MOLA elevation, overlaid on a Mars Orbiter Camera wide-angle mosaic images by transparency, showing the general area. Unfortunately I lost the MOC images and spent some time trying to reproject them again correctly since, and I kept getting an offset between it and the other data. Fortunately there is a version available from the USGS that has already been put into the same projection as MOLA.

In the dissertation, I calculated the distribution of the various variables, both for the whole of the HRSC data tiles I used, and for the areas coinciding with the locations of the Souness glaciers. By comparing the two, a Bayesian classifier is derived.

The head, extent and context areas are explained in the dissertation, the head simply being a small circle around the location of the glacier head as recorded in Souness, the extent being calculated from the width, head, centre and foot recorded, and the context is the extent polygon expanded by 3 times.

I present results from two tiles which were written about in the dissertation, h0248 and h2279, but will be expanded on a little below.

update (13/06/15)

I have corrected an error in the classification script that gave spuriously small values of the 'Extents' classifier function, and revised it to take the nadir image data out of the classifier, since it is scaled separately for each tile, it doesn't make a lot of sense to consider it in the context of global statistics.


larger version (25m/pixel)
Tile h0248
The region near the bottom of the image shows a number of Souness glaciers, which are in an area which is indicated by the classifier, however there are several regions that are indicated by the classifier which do not have any Souness glaciers. There are other glacial features such as crater fills that were beyond the scope of Souness's catalogue, which focused on forms analogous to valley glaciers on Earth.


larger version (25m/pixel)
Tile h2279
Again, a number of Souness glaciers are shown, which have a high classifier value but regions with high values exist that do not have any Souness catalogued objects.
The classifier functions are clearly not specific to the Souness catalogue type glaciers, but may indicate general areas where glacial forms occur, or possibly areas where glacial forms existed in the past.


  1. David, that's brilliant.

    A hard disk failure - oh no !!! That'd upset most of us. Well recovered.

    I have re-read the bit on page 43 of your dissertation where you describe constructing the fuzzy classifier for this purpose. Very interesting. Did you spot any cases (on Mars or Earth) where the classifier didn't identify features which nevertheless looked (to a human) like glaciers ? (If so, what meant that ?)

  2. I'll have to look through a bit more to be sure. I found an error in my classification script that made the values based on the extent areas much smaller than they should have been, and I also decided it made little sense to use the nadir image data in the classifier, because it is scaled separately for each tile from 0-255, which will be clear in the next post.