Pol-NAS: A Neural Architecture Search Method With Feature Selection for PolSAR Image Classification

With the development of deep learning, more and more neural networks have been used in polarimetric synthetic aperture radar (PolSAR) image classification and obtain good results. As we all know, the performances of neural networks highly depend on well-designed neural architectures. Besides, the features input to neural networks also have a huge impact on the classification results. Both architecture design and feature selection are time consuming and require human expertise. Therefore, in this article, we propose a neural architecture search method with feature selection (Pol-NAS) for PolSAR image classification. It can automatically search and obtain a good architecture, including intracell and intercell structure and the number of layers in the search stage. Meanwhile, all the features commonly used in PolSAR data interpretation, rather than part of them, are input to the model in order to avoid selecting the size of an optimal feature subset, which is a hyperparameter and usually different for different models. Then, we propose the feature attention block (FA block) and redesign the stem layers by combining the FA block and the original stem layers. Thus, Pol-NAS can adaptively find the importance of each feature in the training stage by using the redesigned stem layers. With the help of Pol-NAS, we only need to prepare the data and wait for the classification results. Experimental results on three real PolSAR datasets show that the performance of Pol-NAS is better than that of state-of-the-art PolSAR image classification models.