Seabed sediments classification based on side-scan sonar images with adaptive feature fusion and metric learning

Machine learning has become a typical method in seabed sediment classification. Focusing on integrating the advantages of various features and improving the classification performance of K-Nearest Neighbor (KNN) in seabed sediment classification, we propose an adaptive weight feature fusion method and apply metric learning to seabed sediment classification for the first time. First, extract features with the histogram of oriented gradient (HOG), local binary pattern (LBP), and generalized search tree (GIST). Then fuse these three features with the proposed adaptive weight feature fusion method. Finally, the gradient boosting large margin nearest neighbor (GB-LMNN) is used for classification. Experimental results show that the overall accuracy reached 93.5% with the GB-LMNN based on adaptive weight feature fusion. In comparison with random forest (RF), original KNN, and support vector machine (SVM), the GB-LMNN achieved higher accuracy. Besides, after applying the proposed adaptive weight feature fusion method, the accuracy of RF, KNN, SVM, and GB- LMNN were improved. The metric learning classifier gives a new approach to seabed sediment classification.