Evolving Automated Feature Extraction Algorithms for Planetary Science

Introduction: Planetary exploration missions have returned a wealth of imagery data over the last 40 years. The problem is how to make best use of it all. Thoroughly analyzing such large datasets manually is impractical, but developing handwritten feature extraction software is difficult and expensive. The current project explores the use of machine learning techniques to automate the development of feature extraction algorithms for the Mars Orbiter Camera (MOC) narrow angle dataset using Los Alamos National Labo-ratory's GENIE machine learning software. GENIE uses a genetic algorithm to assemble feature extraction algorithms from low-level spatial and spectral image processing steps. Each algorithm is evaluated against user-provided training data, and the most accurate ones are allowed to "reproduce" to build new solutions. The result is automated feature extraction algorithms cus-tomized to the dataset at hand and the current feature of interest. A graphical user interface is used to provide training data, allowing map-makers without programming experience the ability to generate new feature extraction algorithms. Mars Global Surveyor (MGS) [1] has been studying Mars since 1997. The narrow angle dataset produced by the MOC [2] provides imagery with a spatial resolution of approximately 3 meters/pixel in a broad visible/near-infrared spectral range (0.50-0.90 mm). Since its arrival, MOC has taken over 112,000 images, which have been used to study various planetary processes. Craters were selected as our feature of interest because they are a easily recognizable feature that can be used to derive important information about a surface [3-4]. GENIE: GENIE [5-8] uses techniques from genetic algorithms (GA) [9-11] and genetic programming (GP) [12] to construct spatio-spectral feature extraction algorithms for multi-spectral remotely sensed imagery. Both the algorithm structure and the parameters of the individual image processing steps are learned by the system. GENIE has been described at length elsewhere [5-8], so we will only present a brief description here. In particular, the present work explores applying GENIE to panchromatic imagery [13-14]. GENIE begins by randomly generating a population of candidate image-processing algorithms from a collection of spectral and textural image processing operators, including local neighborhood statistics, texture measures, spectral band-math operations (e.g. ratios of bands), and gray-scale morphological filters with various shapes of structuring elements. Each candidate algorithm consists of a number of these image-processing operators, which together generate a vector of processed images in an intermediate, non-linear