Learning class descriptions from a data base of spectral reflectance of soil samples

Abstract

A program was developed to learn class descriptions from positive and negative training examples of spectral reflectance data of bare soils. The program combined a form of learning referred to as “learning by example” with the generate-and-test paradigm to provide a robust learning environment that could handle error prone data. The program was tested by having it learn class descriptions of various categories of organic carbon content, iron oxide content, and particle size distribution in soils. These class descriptions were then used to classify an array of targets. The class descriptions in this program were comprised of a series of different relationships (greater-than, first-maximum, second-maximum, first-minimum, second-minimum, convex, and concave) between combinations of spectral bands. The program found the sequence of relationships between bands that contained the most important information to distinguish the classes. The concept being learned was a sequence of relationships that optimized the discrimination of a class. Training examples of spectral reflectance data of bare soils were taken from data of Stoner and Baumgardner (1981). Classes of any soil characteristic that is in the data base can be explored. The characteristics include organic carbon content, iron oxide content, particle size distribution, soil order, mineralogy, cation exchange capacity, drainage class, and Munsell color. Combinations of the first four characteristics were used to make class descriptions that were used to classify the soil samples into the five classes observed by Stoner and Baumgardner (1981). Other classes of organic carbon, iron oxide, drainage, and clay content were explored using this method. The program was tested by learning class descriptions of these classes and subsequently classifying an array of unknown soil data. The results showed a high classification accuracy of 77% for separating classes of high and low organic carbon content. Physical explanations for the class descriptions obtained were presented.