Rigorous testing of the classifier is incomplete. Results using the data used to train the classifier are excellent. However, there is a significant danger that it has been "over trained", that is simply learned the correct classifications of all the samples in the training and test set. If this is so, it may be of little use in identifying previously unseen examples.
The agenda for the December 1998 Antarctic field season is to test the classifier under true field conditions: that is using data acquired from new rocks by the robot mounted sensors, and with limited human input in the data collection process. This will establish conclusively whether the current system (sensors and classifier) is indeed practical for field robotic deployment, and its suitability for robotic meteorite search.
Over a period of several weeks, weather permitting, the Nomad robot will be driven through the areas of Independence Moraine from where the data and rock samples were acquired last season. Images and spectral data from new samples will be collected, and fed to the classifier for assessment. Samples will be taken from each rock so examined, and the classifier output compared to that from a field geologist who will be with us.
The Patriot and Independence Hills regions of Antarctica are not believed to harbor meteorites. Therefore, to test the classifier utility for meteorite search an area will have to be seeded with meteorites. A geologist looking at high-resolution images returned by the rover will designate possible candidates, that the robot classifier will then tackle. Success will be the robot correctly classifying 90% of the meteorites, and rejecting 95% of all rocks.
In addition, in the robot will be used to gather more data for use in further perfecting the classifier. An expedition further inland from Patriot Hills to the polar plateau to search for meteorites will be an opportunity to acquire more data from different geographical areas, particularly ones where meteorites may be found. A hand carried version of the spectrometer on Nomad and a digital camera will be used for this.
A suitable classifier must therefore be able to handle data from multiple sensors, not all working simultaneously, and subject to the noise and inaccuracies the robotic deployment usually implies. Furthermore, because of the cost of acquiring data, rationally scheduling sensor deployments to identify objects of interest in a cost optimal manner is highly desirable.
To properly create a Bayes network classifier, it is necessary to deduce the statistical relationships between different sensor readings and states of the world. This is easiest done given large case files recording the identity and properties of samples to be identified, along with all sensor readings. However, obtaining substantial amounts of data is difficult, especially as the number of sensors increases. Nevertheless, there is a large corpus related data available, documenting the performance of different individual sensors on similar or related objects. The challenge is to use this at present incompatible data to boost the performance of a Bayes network classifier and allow it to be used in situations dissimilar to the originally mentioned full sample set. In the case of rock identification, this means identifying rocks not found in the original survey area.
Specifically, the issues I wish to address in my thesis are:
However, this work has wider applications than meteorite search. A particularly useful, yet similar task, would be the robotic location and detection of anti-personnel landmines. Should the Bayes network classification approach prove fruitful for meteorite search and geological surveying, adapting and demonstrating it for landmine detection would be a significant contribution to robotic engineering.
Find out about the robotic meteorite search demonstration in the current Antarctic expedition.