Removing redundant rules from a neuro-fuzzy classification system for remotely sensed imagery.

Recent advancements made in the capture of remotely sensed imagery (RSI) have begun to produce images with increased levels of detail and information. The traditional statistical techniques used for RSI classification are proving inadequate and incapable of handling higher resolution multispectral data. Subsequently, neural network classifiers are gaining in popularity due to their inherent suitability to image recognition. However, neural network classifiers often provide incomprehensible feedback on decisions taken to solve image classification problems. When coupled with fuzzy logic, a neuro-fuzzy network provides comprehensible rules defined during the image classification process. These defined rules, however, often contain considerable overlap and redundancy. The author is investigating this specific advancement by attempting to develop a neuro-fuzzy classification system that is able to remove its own redundant classification rules. As the rule reduction component is the core of the new system, algorithms used in existing classifiers are considered as the basis for implementing the new system. The neuro-fuzzy classification system is developed using the IDL programming language and imported as a module into the ENVI image analysis software package. The effectiveness of the new application will be investigated through repetitive testing comparisons with existing neural-network classifiers, against known test imagery, to ensure that suitable results are being produced.

* Wood C.D. Computer Science Department, Southwest Missouri State University.