Meteorite and Rock Classification Results

Liam Pedersen

The preliminary results of the classifier tests are encouraging re the ability of the system to distinguish among meteorites and non-meteorites. Ability to recognize individual rock types is not as good, but still encouraging. A rigorous analysis of the results is needed. Nonetheless, the next few examples illustrate the average performance.

Kirkpatrick Basalt ("Meteorwrong")

This sample was collected in Antarctica up by Dr Ralph Harvey of the US ANSMET program, because of its similarity in appearance to a meteorite. Terrestrial rocks such as this one are sometimes referred to as "meteorwrongs" because of their resemblance to meteorites. As such, they particularly prone to being mis-classified as meteorites by both humans and robots (it should be noted that the opposite has been known to occur as well: researchers have occasionally assumed actual meteorites to be meteorwrongs).

This meteorwrong is a Kirkpatrick basalt, the extrusive equivalent of a Ferrar dolerite. This is an igneous rock formed by slow cooling lava.

Note the green tint of the rock image from the robot mounted high-resolution camera (Figure 1B) compared to an image of the same rock on blue ice, obtained with a handheld digital camera. The blue channel of the camera is clearly faulty. Because of this the current iteration of the classifier is unable to segment the rock from the snow, so the image is unusable for classification purposes.

Fortunately, rocks can also be identified from their spectrum, and the classifier is equipped to do this, with or without a suitable image available.

Figure 3 is the Bayes network classifier output given the spectra (Figure 2) of the rock (Figures 1a, 1b). The rock is correctly identified as an igneous, not an extraterrestrial (read meteorite) rock with a network certainty of 92.3%. However, the rock is misidentified as a granite, not a basalt. This most likely occurs because the network has not previously been exposed to many basalt samples, so it has a poorly defined model of them.

Note the other rock and meteorite types that the classifier is (theoretically) capable of recognizing.

Stony Meteorite (Chondrite)

This is a fragment of a stony meteorite from Caraweena in Australia, lent to us by Dr Bill Cassidy of the University of Pittsburgh. The exterior, covered by a dark fusion crust, is spectrally distinct from the lighter interior. It is a type of meteorite known as a chondrite, characterized by the presence of small silicate spherules called chondrules. These are the most common kinds of meteorites, accounting for approximately 80% of all meteorte finds.

The two spectra in Figure 5 clearly indicate the difference in reflectance spectra of the interior and the crust of the meteorite.

Based on the crustal spectra, the classifier is 84% certain it is a meteorite, and identifies it as a chondrite with 95% probability if it were indeed a meteorite.

Looking at the interior spectra only the classifier believes it to be extraterrestrial or a terrestrial ignous rock with 51% and 40% probabilities respectively. If a meteorite, the classifier gives a 99.8% chance of it being a chondrite. If it is assumed to be a rock, granite (69%) and basalt (7%), both igneous rocks, are the most likely candidates.

Quartz veined Marble

This rock is a large sample of marble, veined with quartz or calcite (it was not possible to be certain in the field). The classifier (see Figure 8) correctly identifies it as a non-meteorite, and gives a reasonable likelihood that it is a marble. However, there is also some confusion with granite, possibly because of it looks similar to the region (pinkish section) of the rock that was sampled. Furthermore, the classifier has been previously trained with marble spectra taken with a different spectrometer, and granite spectra taken with the instrument currently in use, so there is a bias towards granites.

Return to Latest Updates