formulations are possible (Schaepman-Strub et al.
2006), but we will use reflectance here in the ordinary Lambertian sense. This assumption should generally hold for the geologic materials of interest. Note
that the light-colored sediment area in spectra I-III is
associated with a higher average reflectance, as well
as unique spectral features such as the dip near 2
micrometers. These spectra were smoothed using
local linear regression, but some lingering noise
spikes at longer wavelengths evidence the lower signal level in these spectral regions.
Science Autonomy Methods
Zoë’s science autonomy system includes two basic
capabilities that operate on mesoscale and
macroscale features respectively. Smart targeting can
identify science features in rover navigation imagery
and use this information to point the Vis-NIR spectrometer. Adaptive path planning navigates on scales
of tens or hundreds of meters, using satellite images
to select waypoints with distinctive or novel spectra.
We describe each of these techniques in turn.
Zoë began each autonomous target selection process
by acquiring a navigation camera image. On-board
image processing then analyzed the scene to find
large contiguous regions of a desired terrain class.
Typically these classes were rough surface features
like rock outcrop or bright sediment patches with
distinctive spectral signatures. Upon finding a feasible target, the rover recalibrated its Vis-NIR spectrometer, pointed at the feature, and collected a
small 3 x 3 raster of spectra centered on the target of
interest. For context, it also acquired a high-resolution color image of the scene.
The image analysis used a random forest pixel classification system described in previous work (Foil et
al. 2013; Wagstaff et al. 2013) and adapted to the
Atacama environment. This supervised classification
method learns a mapping from local pixel intensities
to the surface class of that pixel. The model is instantiated as an ensemble of decision trees trained in
advance. At run time, the rover tested each pixel in
the new image and averaged the classification of
each tree in the ensemble. The end result was a classification map of the entire image, along with asso-
00.5 1 1. 5 2 2. 5
Figure 6. Panoramic Camera Subframe.
Dive spectrometer fields of view (PP24), and associated reflectance spectra.