Towards Developing an Automated Science Analysis System for Mars Surface Exploration.

We are continuing development of algorithms that will facilitate development of automated systems to assist robotic or human explorers in identifying rocks and minerals in the field. Over the past year, we have focused on algorithms with the ability to identify igneous rocks from images and spectra and on building our database of rocks and minerals. Our collection currently contains over 700 igneous, sedimentary, metamorphic rocks and mineral samples that we have identified, analyzed, and imaged. Images are taken under controlled lighting and at fixed distances. We are in the process of obtaining Raman and visible, near- and mid-infrared spectra of the entire collection to help identify the minerals that comprise the samples. Analysis of both the physical properties and the relative mineral abundances of a sample form the basis of rock identification and classification. This extensive dataset allows us to optimize and test the algorithms under a variety of conditions. We will report on the current ability of our algorithms to identify and discriminate rock types with a variety of input data. When considering color only, using the weighted k-nearest neighbors approach, the algorithm correctly identified greater than 70% of the felsic rocks, at least 70% of the intermediate rocks, and greater than 80% of the mafic rocks. Using a similar approach for texture, the algorithm correctly identified 85% of the plutonic rocks and 76% of the volcanic rocks. We have used both Bayesian and Decision Tree automated reasoning approaches to combine the results of the color and texture algorithms. Based on our tests to date, the Decision Tree method has given the best results, correctly identifying at least 80% of granites and granodiorites and greater than 70% of andesites and basalts using color and texture algorithms combined. In addition to the generally better performance, the Decision Tree method has the advantage of allowing one to trace back the algorithm's line of reasoning in reaching a final identification. Since a hierarchical method more closely follows the line of reasoning used by practicing geologists, we feel it may meet with greater acceptance in spacecraft and field applications. We are continuing to improve our algorithms and will report on their current state of development.