246 IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 1, NO. 4, OCTOBER 2004 Knowledge Discovery From Multispectral Satellite Images Arun Kulkarni and Sara McCaslin Abstract—A new approach to extract knowledge from multi- spectral images is suggested. We describe a method to extract and optimize classification rules using fuzzy neural networks (FNNs). The FNNs consist of two stages. The first stage represents a fuzzi- fier block, and the second stage represents the inference engine. After training, classification rules are extracted by backtracking along the weighted paths through the FNN. The extracted rules are then optimized by use of a fuzzy associate memory bank. We use the algorithm to extract classification rules from a multispectral image obtained with a Landsat Thematic Mapper sensor. The scene represents the Mississippi River bottomland area. In order to verify the rule extraction method, measures such as the overall accuracy, producer’s accuracy, user’s accuracy, kappa coefficient, and fidelity are used. Index Terms—Fuzzy neural network (FNN), rule generation, su- pervised learning, fuzzy associative memory (FAM), multispectral images. I. INTRODUCTION O NE OF THE methodologies used for extracting useful information from raw data involves fuzzy neural net- works (FNNs). Many FNN models have been proposed in the literature [1]–[6]. Until recently, FNNs have been viewed as “black boxes,” which successfully classify data samples, but without anything for the user to see that explains how the network reached the decisions. Recently, there is a growing interest in the research community not only to understand how the neural network or the FNN arrived at a decision but how to decode information stored in the form of connection strengths in the network. One of the major directions taken in this endeavor involves the extraction of fuzzy if-then rules from FNNs. Andrews et al. [7] in their survey article provide the most important features of the published techniques for extracting rules from trained artificial neural networks. They also describe techniques for extracting fuzzy rules from neurofuzzy systems. Survey articles by Mitra et al. [8] and Mitra and Hayashi [9] describe various methodologies for generating fuzzy rules from FNNs. Mitra and Pal [10], [11] have proposed two methods for rule generation. In the first method, they have treated the network as a black box and using the training set input and the network output to generate a set of if-then rules. The second method is based on the backtracking algorithm. In this method, a path is traced from active nodes in the output layer back to the feature input layer based on the maximal paths through the network. The path is calculated as the sum of the input to a node in the network multiplied by the weight of its associated link. At least one rule can be obtained for each data point using this methodology. Any input feature to which a path may be found is considered in producing the if portion of the rules. The Manuscript received April 7, 2004; revised July 5, 2004. The authors are with the Computer Science Department, The University of Texas at Tyler, Tyler, TX 75799 USA (e-mail: Arun_Kulkarni@UTTyler.edu). Digital Object Identifier 10.1109/LGRS.2004.834593 if-part is called the antecedent part and the then-part is called the consequent part. A confidence factor looking only at the inputs and the outputs is calculated for each rule. Wang and Mendel [12] developed a five-step algorithm for directly extracting the rules from a training dataset. First, the input and output spaces are divided into fuzzy regions followed by determining the degrees of the inputs and outputs in different regions. Next, the inputs are assigned to a region with a max- imum degree, and a rule is extracted from the input–output data pair. A degree is then assigned to each rule using the degree of membership of each part of the antecedent multiplied with the degree of membership of the consequent. After the rules are extracted, a combined fuzzy rule base, or a fuzzy associative memory (FAM) bank, is created to store the rules such that each box of possible input combinations has in it only the rule with the highest degree. After all samples from the dataset have been processed, inputs can then be applied to determine a mapping based on the FAM using a centroid defuzzification formula. In order to compare various rule extraction methods, we need some quantitative measures for rule evaluation. Taha and Ghosh [13] and Mitra et al. [8] have proposed measures such as overall ac- curacy, producer’s accuracy, user’s accuracy, kappa coefficient, and fidelity. There are many application areas of rule extrac- tion including medical diagnosis [13] and remote sensing image analysis [14]. In this letter, we present a method for rule extraction that com- bines the backtracking algorithm and a FAM bank for rule op- timization. We train the network with training samples from a satellite image and generate classification rules. The rest of this letter is organized as follows. Section II describes FNN models. Section III deals with rule generation and optimization. Sec- tion IV deals with implementation and results, and Section V provides conclusions. II. FUZZY NEURAL NETWORK MODELS In this section, we describe two FNN models. The first model consists of three layers. The second model consists of four layers, and we have used a backpropagation-learning algorithm to train the network. The rule extraction algorithm is the same for both models. A. Fuzzy Neural Network Models A three-layer fuzzy perceptron model is shown in Fig. 1. The first layer is an input layer. The second layer is used for fuzzifi- cation wherein input feature values are mapped to membership values, and the last layer implements the fuzzy inference en- gine. Layers and represent a two-layer feedforward net- work. The connection strengths connecting these layers encode fuzzy rules used in decision-making. In order to encode decision rules, we have used a gradient descent search technique. The al- gorithm minimizes the mean squared error between the desired output and the actual output. Layers in the model are described below. 1545-598X/04$20.00 © 2004 IEEE