Adaptive multiple kernel fusion model using spatial-statistical information for high resolution SAR image classification

The current high-resolution (HR) synthetic aperture radar (SAR) image classification is confronted with the challenges of the complex spatial patterns and highly variable backscattering of objects. Data-based methods, such as convolutional neural networks (CNNs), can well extract spatial features but ignore valuable statistical knowledge of SAR data. Model-based methods can utilize the non-stationary and non-Gaussian statistical properties of SAR images, while they fail to encode the local spatial patterns. In this paper, an adaptive multiple kernel fusion model with superpixel regularization (AMKFM-SPR) is proposed for HR SAR image classification, which combines the advantages of both deep spatial features and multiscale statistical properties to improve classification accuracy. For the deep spatial features, mul-tilayer features are extracted by using the pre-trained CNN model. Then, a hybrid pooling strategy is designed to encode the feature maps into more distinguishable features from the perspective of spatial structure and local statistics. For the scattering statistical features, the Gabor magnitudes of HR SAR data are modeled by the log-normal distribution and projected into the cumulative distribution function space. Further, the covariance matrix is calculated for mapped multivariate data to build global statistical features. For the feature fusion, multiple kernels are constructed to describe the spatial and statistical information. An improved centered kernel alignment (ICKA) scheme is utilized to adaptively determine the kernel weights and effectively exploit the complementarity among different features. Next, the fusion kernel is fed into a support vector machine to produce the initial classification map. Finally, to compen-sate for the limitations of the patch-based classification way, superpixel regularization (SPR) is performed on the class probability map to enhance label consistency and refine the spatial detail. Experiments on three real HR SAR images show the superiority of the proposed method over other related algorithms. (c) 2022 Published by Elsevier B.V.