SRAD-CNN for adaptive synthetic aperture radar image classification

Abstract

ABSTRACTThe performance of synthetic aperture radar (SAR) image classification based on a conventional convolutional neural network (CNN) is limited by a trade-off between immunity to speckle noise and the ability to locate boundaries accurately. Difficulties regarding the accurate location of boundaries are a result of the smoothing effect of the pooling layer. To address this issue, we propose a novel framework called SRAD-CNN for SAR image classification. In this framework, we apply a filtering layer constructed according to prior knowledge of the speckle reducing anisotropic diffusion (SRAD) filter. The filtering layer can not only reduce speckle but also enhance the boundaries. The main parameter that controls the degree of filtering can be optimized adaptively by a backpropagation algorithm. Image patches adaptively filtered by the filtering layer are then put into the CNN layers to assign a label. Due to the effect of the filtering layer, for our proposed SRAD-CNN, both the speckle noise immunity and the sensitivity to boundaries are superior to those of conventional CNN.To confirm the performance of the proposed SRAD-CNN, we conducted experiments using both simulated and real SAR images. The experimental results demonstrated that the parameter of the filtering layer could be optimized adaptively for different scenes, different noise levels, and different image resolutions. The SRAD-CNN outperformed the conventional CNN in both overall classification accuracy and maintenance of boundary accuracy on images with different resolutions and noise levels with limited training samples.