Underwater acoustic target recognition using SVM ensemble via weighted sample and feature selection

Abstract

The accuracy of underwater acoustic target recognition (UATR) system can be improved by ensemble of support vector machine (SVM) classifiers. However, the ensembles are often large, leading to extra high computational and storage cost. To solve this problem, we propose a novel AdaBoost method based on weighted sample and feature selection (WSFSelect-SVME). The AdaBoost method constructs an ensemble of classifiers iteratively focusing each new individual SVM classifier on the most difficult samples. Weighted immune clonal sample selection algorithm and mutual information sequential forward feature selection algorithm are utilized to keep the performance of each new individual SVM classifier while reducing the number of samples and features in the training set. The classification performance of the proposed method is examined on the UCI Sonar dataset and a real-world underwater acoustic target dataset. Experiment results on two datasets show that, compared to AdaBoost SVM ensemble (SVME) algorithm, the WSFSelect-SVME algorithm obtains better classification accuracy with the number of samples decreasing respectively to 45% and 50%, and the number of features decreasing to 33% and 51%. The experimental results revealed that the proposed algorithm can reduce the space complexity of the ensemble while improving the accuracy compared to the AdaBoost SVME algorithm.