Improved learning strategies for multimodular fuzzy neural network systems: a case study on image classification

Abstract

This paper explores two different methods for improved learning in multimodular fuzzy neural network systems for classification. It demonstrates these methods on a case study of satellite image classification using 3 spectral inputs and 10 coastal vegetation covertype outputs. The classification system is a multimodular one; it has one fuzzy neural network per output. All the fuzzy neural networks are trained in parallel for a small number of iterations. Then, the system performance is tested on new data to determine the types of interclass confusion. Two strategies are developed to improve classification performance. First, the individual modules are additionally trained for a very small number of iterations on a subset of the data to decrease the false positive and the false negative errors. The second strategy is to create new units, ‘experts’, which are individually trained to discriminate only the ambiguous classes. So, if the main system classifies a new input into one of the ambiguous classes, then the new input is passed to the ‘experts’ for final classification. Two learning techniques are presented and applied to both classification performance enhancement strategies; the first one reduces omission, or false negative, error; the second reduces comission, or false positive, error. Considerable improvement is achieved by using these learning techniques and thus, making it feasible to incorporate them into a real adaptive system that improves during operation.