Expert Classification

Producer Field Guide

Producer Field Guide

Expert classification can be performed using IMAGINE Expert Classifier. The expert classification software provides a rules-based approach to multispectral image classification, post-classification refinement, and GIS modeling. In essence, an expert classification system is a hierarchy of rules, or a decision tree, that describes the conditions under which a set of low level constituent information gets abstracted into a set of high level informational classes. The constituent information consists of user-defined variables and includes raster imagery, vector coverages, spatial models, external programs, and simple scalars.

A rule is a conditional statement, or list of conditional statements, about the variable’s data values and/or attributes that determine an informational component or hypotheses. Multiple rules and hypotheses can be linked together into a hierarchy that ultimately describes a final set of target informational classes or terminal hypotheses. Confidence values associated with each condition are also combined to provide a confidence image corresponding to the final output classified image.

IMAGINE Expert Classifier is composed of two parts: Knowledge Engineer and Knowledge Classifier. Knowledge Engineer provides the interface for an expert with first-hand knowledge of the data and the application to identify the variables, rules, and output classes of interest and create the hierarchical decision tree. Knowledge Classifier provides an interface for a nonexpert to apply the knowledge base and create the output classification.

Knowledge Engineer

Using Knowledge Engineer, you can open knowledge bases, which are presented as decision trees in editing windows.

Knowledge Engineer Editing Window


In the dialog above, the upper left corner of the editing window is an overview of the entire decision tree with a green box indicating the position within the knowledge base of the currently displayed portion of the decision tree. Drag the green box to change the view of the decision tree graphic shown in the right side display window. The branch containing the currently selected hypotheses, rule, or condition is highlighted in the overview.

The decision tree grows in depth when the hypothesis of one rule is referred to by a condition of another rule. The terminal hypotheses of the decision tree represent the final classes of interest. Intermediate hypotheses may also be flagged as being a class of interest. This may occur when there is an association between classes.

The figure below represents a single branch of a decision tree depicting a hypothesis, its rule, and conditions.

Example of a Decision Tree Branch


In this example, the rule, Gentle Southern Slope, determines the hypothesis, Good Location. The rule has four conditions depicted on the right side, all of which must be satisfied for the rule to be true.

However, the rule may be split if either Southern or Gentle slope defines the Good Location hypothesis. While both conditions must still be true to fire a rule, only one rule must be true to satisfy the hypothesis.

Split Rule Decision Tree Branch


Variable Editor

Knowledge Engineer also uses a Variable Editor when classifying images. The Variable editor provides for the definition of the variable objects to be used in the rules conditions.

The two types of variables are raster and scalar.

Raster variables may be defined by imagery, feature layers (including vector layers), graphic spatial models, or by running other programs.

Scalar variables my be defined with an explicit value, or defined as the output from a model or external program.

Evaluating the Output of Knowledge Engineer

The task of creating a useful, well-constructed knowledge base requires numerous iterations of trial, evaluation, and refinement. To facilitate this process, two options are provided. First, you can use Test Classification to produce a test classification using the current knowledge base. Second, you can use Classification Pathway Cursor to evaluate the results. Use this tool to move a crosshair over the image in a Viewer to establish a confidence level for areas in the image.

Knowledge Classifier

Knowledge Classifier is composed of two parts: an application with a user interface, and a command line executable. Enter a limited set of parameters to control the use of the knowledge base. The user interface is a wizard to lead you through entering parameters.

After selecting a knowledge base, you select classes. The following is an example classes dialog:

Knowledge Classifier Classes of Interest


After you select the input data for classification, the classification output options, output files, output area, output cell size, and output map projection, the Knowledge Classifier process can begin. An inference engine then evaluates all hypotheses at each location (calculating variable values, if required), and assigns the hypothesis with the highest confidence. The output of Knowledge Classifier is a thematic image, and optionally, a confidence image.