Abstract—A Feature Selective Linguistic Classifier (FeSLiC) is proposed in this paper, for land cover classification from hyperspectral images. FeSLiC is a Genetic Fuzzy Rule-Based Classification System (GFRBCS), designed under the Iterative Rule Learning (IRL) approach. A local feature selection scheme is employed, designed to guide the genetic evolution, through the evaluation of deterministic information about the relevancy of each feature with respect to its classification ability. A simplification post-processing stage significantly enhances the interpretability of the derived model, by reducing its structure size. The performance of the classifier is finally optimized through a genetic tuning stage. Comparative results using an Earth Observing-1 (EO-1) Hyperion satellite image indicate the effectiveness of the proposed methodology in handling high- dimensional feature spaces. I. INTRODUCTION VER the past two decades, the development of satellite hyperspectral sensor technologies has provided new means to improve the accuracy of land cover classification from satellite images. Hyperspectral sensors collect several (typically 200 or more) narrow spectral bands, from the visible to the short-wave infrared portions of the electromagnetic spectrum, providing an almost continuous spectral reflectance signature. It has been reported that hyperspectral data are capable of producing both genus- and species-level classifications, whereas multispectral data are well-suited for genus-level classifications [1]. Particularly, in land cover classification of forests, where typically different species of the same genus coexist, it has been shown that hyperspectral satellite imagery can significantly increase the classification accuracy [2]. Fuzzy Rule-Based Classification System (FRBCSs) could accurately describe the overlapping signatures of closely related vegetation species, in a way easily interpretable by humans. However, constructing FRBCSs for such high- dimensional feature spaces is far from straightforward, since the number of possible fuzzy rules increases exponentially with the increase of the feature space. Therefore, some sort of feature reduction mechanism must be employed, Manuscript received February 7, 2010. D. G. Stavrakoudis and J. B. Theocharis are with the Department of Electrical and Computer Engineering, Division of Electronics and Computer Engineering, Aristotle University of Thessaloniki, 54124 Thessaloniki, Greece (e-mail: jstavrak@auth.gr; theochar@eng.auth.gr). G. N. Galidaki and I. Z. Gitas are with the Laboratory of Forest Management and Remote Sensing, School of Forestry and Natural Environment, Aristotle University of Thessaloniki, 54124 Thessaloniki, Greece (e-mail: galidaki@for.auth.gr; igitas@for.auth.gr). combined with a powerful search algorithm. Since the beginning of the previous decade, the enhanced search capabilities of Genetic Algorithms (GAs) [3] have been extensively used in the derivation of Fuzzy Rule-Based Systems (FRBSs), giving rise to the field of Genetic FRBSs (GFRBSs) [4]. Feature selection mechanisms can be easily employed in GFRBSs, leading to compact fuzzy rule bases, thus increasing the inherent interpretability properties of FRBSs. In constructing GFRBSs, four basic genetic learning approaches can be identified [5]: the Pittsburg, the Michigan, the Genetic Cooperative-Competitive Learning (GCCL), and the Iterative Rule Learning (IRL). In the first three approaches, despite their different formulation, the population (or a part of it) represents the complete fuzzy rule base. Conversely, in the IRL approach, each chromosome represents a single rule and a rule generation algorithm in repeatedly invoked, iteratively adding fuzzy rules to the rule base, one at a time. Thus, the dimensionality of the search space is reduced, making the derivation of the fuzzy rule base easier and faster. Prominent examples of GFRBSs following the IRL approach are the MOGUL [6] and the SLAVE [7]–[8] systems. In this paper, we propose a novel GFRBCS, targeted at effectively handling high-dimensional feature spaces. The model is constructed through a three-stage process. The first stage constructs a preliminary rule base, under the principles of the IRL approach. During this stage, a novel feature selection scheme is integrated in the genetic rule generation algorithm, based on the notion of the Feature Partition Vector (FPV). The latter provides deterministic information about the relevancy of each feature with respect to its classification ability, guiding the GA in effectively selecting each rule’s relevant features faster. Subsequently, a simplification post-processing stage follows, aiming at eliminating any possible redundancy introduced in the first stage, reducing the complexity of the fuzzy rule base. Finally, the performance of the classifier is optimized through a genetic tuning stage. The rest of the paper is organized as follows. In Section II, we describe the proposed system’s fuzzy inference mechanism implementation. In Section III, the three-stage learning algorithm is detailed. The inclusion of a novel feature selection scheme is described in Section IV. Experimental results are reported in Section V, using a Hyperion hyperspectral satellite image. We conclude the paper in Section VI, with a summary of the proposed system. Enhancing the Interpretability of Genetic Fuzzy Classifiers in Land Cover Classification from Hyperspectral Satellite Imagery Dimitris G. Stavrakoudis, Georgia N. Galidaki, Ioannis Z. Gitas, and John B. Theocharis O WCCI 2010 IEEE World Congress on Computational Intelligence July, 18-23, 2010 - CCIB, Barcelona, Spain FUZZ-IEEE 978-1-4244-8126-2/10/$26.00 c 2010 IEEE 1277