Abstract
The ability of synthetic aperture radar (SAR) to capture maritime phenomena is widely acknowledged. However, ocean SAR scene automatic classification remains challenging due to speckle noise interference, the nonlinearities and poor distinguishability of different geophysical phenomena. Kernel entropy component analysis (KECA) was recently proposed as a feature extraction approach. It is capable of handling nonlinear data and revealing different structures of interest. However, KECA suffers from high computational complexity, meaning it cannot penetrate deep for finer feature extraction. To address this issue, this paper proposes an efficient multilayer convolutional kernel network (denoted as KECANet) equipped with KECA for ocean SAR scene classification. The pivoted Cholesky decomposition is employed to accelerate KECA filtering in the network. KECA was trained on hand-labeled but limited samples describing ten oceanic or atmospheric phenomena. Several conventional and state-of-the-art deep learning methods were also included for comparison. According to the classification experiments, KECANet can greatly improve the classification precision of geophysical phenomena, considering that the precision, recall and F-score values increased by 13.3%, 2.3% and 12.2% in average. Overall, the results suggest that KECA is a promising approach for various applications in remote sensing image recognition.