Open447
I'm quite dubious about crowdsourcing in the pure sense, because the problem is too much like screening bags - people aren't very good at detecting significant rare events in a large number of insignificant ones. Further, as a lot of the seabed is going to be quite different oceanographically, most of the variation in it will be rocks and the reviewers will lock on to the rocks because they're what's available (anchoring; target fixation).
But there might be some role for it as part of a semiautomatic process - for example, if some kind of statistical model has thrown up x number of probables, there could be value in having the crowd work through them, as the targets probably aren't very amenable to automatic detection. If we were looking for the fin, for example, that would have some nicely defined corners, but we're looking for "stuff that looks engineered, but has been smashed" and that's quite subjective.
There are some interesting methodologies for this sort of thing; one would be pairwise comparison (a.k.a kittenwar). Rather than "is there an aeroplane here?" or worse "here are 100 very similar sonar images, can you spot a piece of wreckage?", this presents the user with 2 randomly selected probables and asks them to pick one - eventually, you should end up with the images sorted by the totality of the users' preferences, which should be a valid Bayesian search. There's a more formal version called the Analytical Hierarchy Process, where you make the pairwise comparison on the basis of parameters you choose in advance.
As well as kittenwar and Am I Hot Or Not, this has been used for quite a few hard problems - like examining British politicians' expenses claims, and making technology decisions in big companies.
Alternatively, if you were trying to build a statistical method of identifying targets on the seabed, you might use something similar to classify the seabed into different topographies, or to confirm that a zone was clear of wreckage so that you could use it to calibrate the model.
Depending how self-similar the seabed topography is at the relevant scale, it might be possible to use a Fourier analysis to compare similar areas of seabed - the answer to my assumption looks to be "not very", but it might be possible to classify it into relatively self-similar zones.