Data Prospecting with CORPRAL: Pre-attentive Vision Model at Work

I. A. Galkin, B. W. Reinisch, G.M. Khmyrov, A.V. Kozlov, J. Grinstein, 
and S. F. Fung

The Cognitive Online Rpi Plasmagram Rating Algorithm (CORPRAL) is the automated
data prospecting tool developed to find interesting examples of signal
propagation in the 1.2 million plasmagram image archive from the Radio Plasma
Imager (RPI) onboard the IMAGE spacecraft. The RPI instrument is a radar whose
transmitted radio waves may reflect from important magnetospheric structures,
such as the plasmapause and the magnetopause, and return to the spacecraft
location to be detected. The CORPRAL prospector draws attention of the human
analysts to the RPI plasmagrams that contain traces of remote reflections (~18%
of all images), thus helping to relieve the search efforts in the otherwise
overwhelming RPI data repository. To find the echo traces in plasmagrams,
CORPRAL employs a pre-attentive vision model that replicates human ability to
identify important objects in the field of view without willful concentration of
attention. The importance of such objects is determined by their ÒsaliencyÓ, the
ability to stand out against the background. The saliency evaluation is done
subconsciously, without a priori concepts of the object shape. The RPI imagery
dataset providesd an excellent testbed for studies of the model performance on
low saliency objects immersed in irregular background noise. The paper discusses
results of statistical evaluation of the CORPRAL performance on a collection of
~25,000 plasmagram images interpreted manually. Overall accuracy of plasmagram
interpretation varies depending on the applied measuring program but remains in
the mid-90% range. As the prevalence of plasmagrams that contain traces goes
below 10%, we do observe lower positive predicted values (PPV) of the CORPRAL
analysis (~50%), indicating that automatically selected plasmagrams are
interesting only with 50% chance. We will discuss future development efforts and
provide examples and scenarios where RPI data prospecting is instrumental in
knowledge discovery.

_______________
Fall Meeting, American Geophysical Union, San Francisco, U.S.A., 2005