Learning to Recognize Volcanoes on Venus

  • Authors:
  • Michael C. Burl;Lars Asker;Padhraic Smyth;Usama Fayyad;Pietro Perona;Larry Crumpler;Jayne Aubele

  • Affiliations:
  • Jet Propulsion Laboratory, MS 525-3660, 4800 Oak Grove Drive, Pasadena, CA 91109, USA and California Institute of Technology. E-mail: burl@aig.jpl.nasa.gov;Jet Propulsion Laboratory, MS 525-3660, 4800 Oak Grove Drive, Pasadena, CA 91109, USA and Stockholm University;Jet Propulsion Laboratory, MS 525-3660, 4800 Oak Grove Drive, Pasadena, CA 91109, USA and University of California, Irvine;Jet Propulsion Laboratory, MS 525-3660, 4800 Oak Grove Drive, Pasadena, CA 91109, USA and Microsoft Research;California Institute of Technology and Università di Padova;Brown University and New Mexico Museum of Natural History & Science;Brown University and New Mexico Museum of Natural History & Science

  • Venue:
  • Machine Learning - Special issue on applications of machine learning and the knowledge discovery process
  • Year:
  • 1998

Quantified Score

Hi-index 0.00

Visualization

Abstract

Dramatic improvements in sensor and image acquisition technologyhave created a demand for automated tools that can aid in the analysis oflarge image databases. We describe the development of JARtool, a trainablesoftware system that learns to recognize volcanoes in a large data set ofVenusian imagery. A machine learning approach is used because it is mucheasier for geologists to identify examples of volcanoes in the imagery thanit is to specify domain knowledge as a set of pixel-level constraints. Thisapproach can also provide portability to other domains without the need forexplicit reprogramming; the user simply supplies the system with a new setof training examples. We show how the development of such a system requiresa completely different set of skills than are required for applying machinelearning to “toy world” domains. This paper discusses importantaspects of the application process not commonly encountered in the“toy world,” including obtaining labeled training data, thedifficulties of working with pixel data, and the automatic extraction ofhigher-level features.