Classification of Remote Sensing Images based on Ant Colony Optimization
基于蚁群智能的遥感影像分类新方法

This paper presents a bottom-up approach to mi prove the classification performance for remote sensing applications. Top-down approaches, such as statistical classifiers, have inherited lmi itations in dealingwith complicated relationships in classification. Forexample, data correlation between bandsofremote sensing mi ageryhas caused problems in generating satisfactory classification with statisticalmethods. In this paper, ant colony optmi ization (ACO) based on swarm intelligence is used to mi prove classification performance. Actually, ACO is a complex multi-agent system, in which agentswith smi ple intelligence can complete complex tasks through cooperation such as classification problems. Ants guide their route selection based on pheromone, which is accumulated from the collectivemovements of individual ants. In thisway, an ant learns from the pastexperience ofothers. Ant-Miner is different from decision tree approaches. The entropymeasure is a local heuristicmeasure, which considers only one attribute at a tmi e, and so it is sensitive to attribute correlation problems. Whereas inAnt-Miner, pheromone updating tends to cope betterwith attribute correlation, since pheromone updating is directly based on the performance of the rule as awhole. Thus, Ant-Miner should have great potential in mi proving remote sensing classification. In this study, an Ant-Miner program for discovering classification rules is developed for the classification of remote sensing mi ages. In theAnt-Miner program, the route search by an ant colony is to find the best links between attribute nodes and class nodes. An attribute node corresponds to a band value of remote sensing mi ages. An attribute node can only be selected once andmustbe associated with a class node. Each route corresponds to a classification rule, and discovering a classification rule can be regarded as searching for an optmi al route. To enableACO to effectively classify remote sensing mi agery of very large data sets, original band values are sliced into a number of intervals by using a discretization technique. The ACO method is more explicit and comprehensible than mathematical equations. Our study in Guangzhou city indicates that the ant colony-based classifier yields better accuracy than conventionalmaxmi um likelihood classifiers and decision tree classifiers. The overall accuracy of theACO method is 88.6%, with aKappa coefficientof0.861. The decision treemethod has an accuracy of85.4% and aKappa coefficientof 0.822. Themaxmi um likelihoodmethod has an accuracy of83. 3% and aKappa coefficient of0. 796. The results clearly support the conclusion that themethod explored in thispapercan bemore effective than conventional classificationmethods.