Research On PolSAR Images Classification Via Automatic Search-based Multi-scale CNN

As one of the important issues in high-resolution polarimetric synthetic aperture radar(PolSAR)image interpretation,land cover classification has become a hot topic in the field of remote sensing recently.At present,remarkable deep learning algorithms based on convolutional neural networks(CNNs)have shown excellent performance in the intelligent interpretation of PolSAR images.However,standard CNNs are limited to single-scale image slices and convolution kernels,which leads to information loss in the process of feature extraction for some targets with different scales.In addition,to tune the hyperparameters of CNNs manually relies on a large amount of professional knowledge and rich computing resources,the manually designed architecture might be suboptimal.Therefore,how to efficiently design a PolSAR-tailored neural network is an important topic.Combining the deep learning theory and the properties of PolSAR system,this paper conducts in-depth research on complex-valued multi-scale CNN with automatic designed hyperparameters.The main work involves the following three aspects:First,beginning from the matrix description methods of the target scattering characteristics of PolSAR system,the basic working units,model structures,optimization methods and working characteristics of the PolSAR-tailored CNNs are discussed.Considering the complex-valued form of PolSAR dataset,this paper mainly focuses on how each module of CNNs works in the complex-valued domain,and analysis the characteristics of the multi-scale PolSAR image slices.Then,different presentation forms of each target in the image slices are summarized totally.Second,considering that the information loss may be generated when the singlescale CNNs are employed to process multi-scale targets in PolSAR images,and the coherent speckle noise of the PolSAR images will affect the quality of feature extraction.In this paper,the wavelet transform operation is integrated into the CNNs to fully maintain the low-frequency components in the image features and remove the high-frequency components like coherent speckle noise.In order to fully highlight the importance of the central local area of the image slices,1 × 1 complex-valued convolution is used to extract the abstract features of the channel dimension.In addition,a complex-valued transformer network is constructed to extract the correlation characteristics between different local regions to make up for the lack of the local connection pattern in the convolutional operations.Finally,a complex-valued multi-scale CNN model with multiple branches is constructed to achieve PolSAR image classification.Third,according to the topology of the complex-valued multi-scale CNN,referring to the problem of high calculation and human participation for designing hyperpatameters of CNNs manually and combining the basic theory of neural architecture search(NAS),a search space based on some key hyperparameters such as the height,width and the number of channels of the complex-valued convolution kernels,the size of image slices,and the number of neurons in the fully connected layers are constructed.In order to improve the efficiency of the search algorithm,the search process is improved and perfected from three perspectives,such as the differentiable search strategy,the nonlinear activation method of the architecture parameters and the construction of the fused convolutional modules in the parent network,so as to search a set of optimal hyperparameter combination.After completing the search process of the hyperparameters of the complex-valued multi-scale CNN,the model parameters are retrained to achieve PolSAR image classification.