Extraction of Coral Features from MBES data using GRASS

In 2006 I was faced with a problem. Where I worked at the time we were acquiring large volumes of seabed information in the form of discrete multi-beam echo sounder (MBES) data for bathymetry and side scanning sonar to identify seabed features. Our vector database of classified coral outlines had not been updated based on these datasets for a few years and it was a big task to re-interpret the lot.

We were not recording backsctter intensity from the MBES and SSS (side scanning sonar) has some inherently funky geometry which makes it hard to work with and integrtate with MBES for automated feature extraction purposes (although I did some interesting work with this in 1999/2000 that I may post if I find the results). I had some previous experience with terrain analysis and thought I could lend a hand.

I had done quite a bit of work in the past using 2nd differentials of terrain (more commonly known as rugosity). If the first differential of a terrain dataset is slope - or rate of change of the vertical component over a horizontal surface (this is essentially the vertical variability or undulation of the terrain), then the second differential is "slope slope" or the rate of change of the rate of change of the vertical over the horizontal. This is essentially a measure of how rugged a surface is. The theory is that coral tends to exhibit higher rugosity than the surrounding seabed.

Another technique to find the edges of the coral is to perform "inflection" analysis. In this case you take a smoothed lower resolution version of the terrain (spline fitting is often well suited to this) and intersect it with a higher resolution model. The point at which the absolute differences between the two models is zero is often a pretty good approximation of the edge of the Coral. This method sometimes seems to work better than using the 1st differential (high slope) alone.

A combination of slope, rugosity, inflection analysis and various "grow" and "thin" filters eventually results in a fairly accurate extraction of the coral areas - although a bit of tuning is required depending on the the resolution and general properties of the input terrain model .

I had worked on similar problems in the past using custom filters and kernels in ER Mapper and also using ESRI technology in Spatial Analyst. As I was mucking around with GRASS under Linux at the time I decided to use this environment instead. I was extremely impressed! I found it extremely powerful that I could write bash scripts to automate and tie together many processes.

After successful extraction of coral features in GRASS we ran vectorisation of the results in Spatial Analyst, performed some sinuosity filters to remove jagged effects and then updated the coral outlines data in GIS.

The screen capture below shows the results against a dataset over varying quality. It should be noted that thus was taken part way through the extraction process and does not represent the final classification. Unfortunately I could not find a screen cap of what was finally extracted nor could I find the bash scripts we used in GRASS.... I'll keep looking.