The MH370 Tomnod Tiles: How the experts are doing it

I've seen a number of comments on the MH370 subreddit calling the TomNod crowdsourcing pointless. Most seem to focus on the cell resolution (the "grain") and coverage area but don't touch upon how experts would actually sift through the data, so I thought I'd give a brief introduction. This is not meant to disparage TomNod or its users; just to outline what the experts are doing right now in their search efforts.

We don't really need to dwell upon the tiny area:

The data has been collected by WorldView-2. Launched in 2009, this satellite is spinning around the world 770 km above the ground. The best "colour" pixellation (grain size) available to anyone but the US government is 0.5 metres by 0.5 metres... in contrast, LANDSAT-8 has a grain size of 30 metres. Anyone can buy imagery from this satellite but it's (understandably) extraordinarily expensive.

In the science, "colour" imagery is called "multispectral". The satellite's sensor is measuring the amount of a number of different wavelengths reflecting from the Earth. In the online TomNod tiles, we are seeing a composite of a few different bands located in the visible spectrum to make a "true colour image". But WorldView-2 in fact measures outside the visible spectrum (~0.4 to ~0.7 micrometres), too: (Image source)

So there are eight multispectral bands, two of which are between 0.75 and 1.050 micrometres... the "Near Infrared" range. Each pixel has a value for the amount of reflectance from each wavelength.

Different materials have different spectral signatures... they reflect different amounts of each wavelength. For example within the visible wavelength, trees reflect more green (~0.55 micrometres) than they do blue (~0.47) or red (~0.70). Here's a graph comparing the "spectral signature" of a few materials; the visible spectrum is located on the left: (Image source)

The soil here would be a reddish-brown.

Now, you can run through pixels looking for high levels of green and call those vegetation, or you can amplify the difference between vegetation and bare soil by comparing two or more different bands. The most famous algorithm for this is probably the Normalized Difference Vegetation Index, which compares the level of infrared and red reflectance within each cell through the algorithm (Near Infrared - Red)/(Near Infrared + Red) to make healthy vegetation "jump out". If you've seen a vegetation map like this you've seen the NDVI in action -- every cell has a value derived from this simple calculation, and a colour gradient has been assigned with higher values being greener: (Image source)

Every material has its own "spectral signature" and there are thousands of algorithms to make each material "stand out" from its surroundings. You can be assured that developed militaries know the alloys used in their rivals' forces. Note that while we're just looking at a satellite with eight multispectral bands, there exists "hyperspectral" imagery with hundreds of bands that can differentiate materials with great precision -- this is most likely what the US government has. (Atmospheric scatting makes improvements in spatial resolution quite difficult so their "top secret" improvements more likely relate to temporal and band resolution). They would also know the spectral signature of Boeing's Alloy 7055 and are using an algorithm to distinguish it from its surroundings. This would be made more difficult by the possibility that if it's floating, it is covered by a thin layer of water.

So they're running through the available satellite imagery, using a matrix with the parameters "minimum class size" (individual pixels might just pick up rubbish... they're looking for a grouping of cells with the same signature) and "sample interval" (they might be skipping over cells and only checking every fifth one... the amount of data here is immense and time is an issue). THEN, when Boeing's alloy is spotted, they're probably using simple human (highly experienced) eyes to look over the pixels, just like we're doing with the TomNod photos, though probably running other algorithms as well. This was most likely the process involved for Australia's "sighting".