Abstract
This work was supported by NASA Grant No. NAGW-925 “Earth Observation Research - Using Multistage EOS-Iike Data” (Principal lnvestigators: David A. Landgrebe and Chris Johannsen). The Anderson River SAR/MSS data set was acquired, preprocessed, and loaned to us by the Canada Centre for Remote Sensing, Department of Energy Mines, and Resources, of the Government of Canada. The importance of utilizing multisource data in ground-cover^ classification lies in the fact that improvements in classification accuracy can be achieved at the expense of additional independent features provided by separate sensors. However, it should be recognized that information and knowledge from most available data sources in the real world are neither certain nor complete. We refer to such a body of uncertain, incomplete, and sometimes inconsistent information as “evidential information.” The objective of this research is to develop a mathematical framework within which various applications can be made with multisource data in remote sensing and geographic information systems. The methodology described in this report has evolved from “evidential reasoning,” where each data source is considered as providing a body of evidence with a certain degree of belief. The degrees of belief based on the body of evidence are represented by “interval-valued (IV) probabilities” rather than by conventional point-valued probabilities so that uncertainty can be embedded in the measures. There are three fundamental problems in the muItisource data analysis based on IV probabilities: (1) how to represent bodies of evidence by IV probabilities, (2) how to combine IV probabilities to give an overall assessment of the combined body of evidence, and (3) how to make a decision when the statistical evidence is given by IV probabilities. This report first introduces an axiomatic approach to IV probabilities, where the IV probability is defined by a pair of set-theoretic functions which satisfy some pre-specified axioms. On the basis of this approach the report focuses on representation of statistical evidence by IV probabilities and combination of multiple bodies of evidence. Although IV probabilities provide an innovative means for the representation and combination of evidential information, they make the decision process rather complicated. It entails more intelligent strategies for making decisions. This report also focuses on the development of decision rules over IV probabilities from the viewpoint of statistical pattern recognition The proposed method, so called “evidential reasoning” method, is applied to the ground-cover classification of a multisource data set consisting of Multispectral Scanner (MSS) data* Synthetic Aperture Radar (SAR) data, and digital terrain data such as elevation, slope, and aspect. By treating the data sources separately, the method is able to capture both parametric and nonparametric information and to combine them. Then the method is applied to two separate cases of classifying multiband data obtained by a single sensor, in each case, a set of multiple sources is obtained by dividing the dimensionally huge data into smaller and more manageable pieces based on the global statistical correlation information. By a Divide-and-Combine process, the method is able to utilize more features than the conventional Maximum Likelihood method.
Date of this Version
7-1-1990
Comments
This work was supported by NASA Grant No. NAGW-925 “Earth Observation Research - Using Multistage EOS-Iike Data” (Principal lnvestigators: David A. Landgrebe and Chris Johannsen). The Anderson River SAR/MSS data set was acquired, preprocessed, and loaned to us by the Canada Centre for Remote Sensing, Department of Energy Mines, and Resources, of the Government of Canada.