MULTIPLE CLASSIFIER COMBINATION FOR TARGET IDENTIFICATION FROM HIGH RESOLUTION REMOTE SENSING IMAGE

摘要

Target identification from high resolution remote sensing image is a common task for many applications. In order to improve the performance of target identification, multiple classifier combination is used to QuickBird high resolution image, and some key techniques including selection and design of member classifiers, classifier combination algorithm and target identification methods are investigated. A classifier ensemble is constructed at first, consisting of seven member classifiers: Decision Tree Classifier (DTC) and NaiveBayes classifier, J4.8 decision tree classifier, simple classifier OneR, IBK classifier, feed-forward Neural Network (NN) and Support Vector Machine (SVM). Weighted Count of Errors and Correct results (WCEC) measure is used to select five classifiers for further combination. DTC, J4.8, NN, SVM and IBK are selected and their independence and diversity are evaluated. Some standard MCS methods, such as Boosting, Bagging, linear combination and non-linear combination are experimented to extract road from QuickBird image. The results show that multiple classifier combination can improve the performance of image classification and target identification.