AutoPolCNN: A neural architecture search method of convolutional neural network for PolSAR image classification

Abstract

Convolutional neural networks (CNNs), as a kind of typical classification model known for good performance, have been utilized to cope with polarimetric synthetic aperture radar (PolSAR) image classification. Nevertheless, the performances of CNNs highly rely on well-designed network architectures and there is no theoretical guarantee on how to design them. As a result, the architectures of CNNs can be only designed by human experts or by trial and error, which makes the architecture design is annoying and time-consuming. So, a neural architecture search (NAS) method of CNN called AutoPolCNN, which can determine the architecture automatically, is proposed in this paper. Specifically, we firstly design the search space which covers the main components of CNNs like convolution and pooling operators. Secondly, considering the fact that the number of layers can also influence the performance of CNN, we propose a super normal module (SNM), which can dynamically adjust the number of network layers according to different datasets in the search stage. Finally, we develop the loss function and the search method for the designed search space. Via AutoPolCNN, preparing the data and waiting for the classification results are enough. Experiments carried out on three PolSAR datasets prove that the architecture can be automatically determined by AutoPolCNN within an hour (at least 10 times faster than existing NAS methods) and has higher overall accuracy (OA) than state-of-the-art (SOTA) PolSAR image classification CNN models.