Similarity-Based Multiple Kernel Learning Algorithms for Classification of Remotely Sensed Images

Abstract

Multiple kernel learning (MKL) algorithms are proposed to address the problems associated with kernel selection of the kernel-based classification algorithms. Using a group of kernels rather than one single kernel, the MKL algorithms aim to provide better classification efficiency. This paper presents new similarity-based MKL algorithms to classify remote-sensing images. These algorithms find the optimal combination of kernels by maximizing the similarity between a combination of kernels and an ideal kernel. In this framework, we initially introduced three similarity measures to be used: kernel alignment, norm of kernel difference, and Hilbert–Schmidt independence criterion. Then, we proposed to solve the optimization problems of the MKL algorithm associated with each similarity measure adopting heuristic and convex optimization methods. The performances of the proposed algorithms were compared with a single kernel support vector machines as well as other MKL algorithms for classifying the features extracted from the high-resolution and hyperspectral images. The results demonstrated that the similarity-based MKL algorithms performed better than other algorithms, especially when their optimization problems were solved using the convex optimization methods or when few training samples were available. Moreover, when the optimization problems of these algorithms were solved using the heuristic optimization methods, they were able to yield acceptable performances and were faster than other MKL algorithms.