Remote Sensing and Geoinformation EARSeL, 2011 not only for Scientific Cooperation The Use of Artificial Intelligence Optimization Algorithms in Unsupervised Classification Ümit Haluk Atasever 1 , Coşkun Özkan 1 , Filiz Sunar 2 1 Erciyes University Engineering Faculty, Geomatics Engineering Department, Remote Sensing Division, Kayseri, Turkey; {uhatasever, cozkan@erciyes.edu.tr} 2 Istanbul Technical University Civil Engineering Faculty, Geomatics Engineering Department, Maslak, Istanbul, Turkey; fsunar@itu.edu.tr Abstract. One of the most important digital image processing steps in the remote sensing is undoubtedly classification. As a branch of classification, unsupervised classification is a general concept which defines natural structures and groups in data without training. The K-means (KM) and Fuzzy C-means (FCM) are the most common unsupervised classification methods. For over the last two decades, Artificial Intelligence (AI) optimization algorithms (heuristic algorithms) such as Genetic Algorithm (GA), Particle Swarm Optimization (PSO), Differential Evolution (DE) have been using for different nonlinear problems in many disciplines successfully. As an AI optimization algorithm, Artificial Bee Colony (ABC) was recently proposed. PSO is inspired from the social behaviors of bird and fish shoals, GA is similar to the evolutionary process in the nature, DE is a GA-based intuitional algorithm and ABC which is inspired from honey bees. These AI optimization algorithms are especially preferred when the classical deterministic methods are inadequate because of too many parameters and data sets are not homogenous. These AI tools are effectively used in remote sensing, as well. The main objective of this study is to examine the effectiveness of AI optimization algorithms for unsupervised classification problem. Especially the potential of the ABC among others is the most overrating point of this study. The basic approach for unsupervised classification is to determine the centers of clusters of user specified number by using pattern similarities based on the Euclid norm. As dataset, beside to Landsat TM imagery, some benchmark data named Wine, Breast Tissue and Image Segmentation obtained from UCI da- tabase are used. The performances of AI tools are compared with conventional KM and FCM. Keywords. Particle Swarm Optimization Algorithm, Genetic Algorithm, Differential Evolution Algorithm, Artificial Bee Colony, Landsat TM Image, Clustering 1. Introduction Land cover mapping is one of the most important application areas of remote sensing discipline. Classification is preferential step for producing thematic spatial information from satellite image data. Classification can be defined as grouping of similar pixels by aid of mathematical equations which describes neighborhood relations of pixels. There are basically two approaches for image classification. In supervised classification, training areas are selected by user and these data are used for training of classification procedure. In unsupervised classification, there is no training process. So, instead of training process, clustering is done according to spectral brightness value. Most frequently used methods in unsupervised classification of satellite imagery are FCM, K-Means. Besides heuristic optimization methods such as GA, DEA and PSO are also being used for solution of problems. ABC is relatively new algorithm according to other heuristic methods. But due to be- ing effective, adaptive and rapid method, many researchers have begun to use ABC in their study areas. In this study, ABC algorithm have been used for calculation of class centroids and compared with GA, DEA, PSO Fuzzy C-Means (FCM) and K-Means to evaluate its efficiency.