Evolutionary Neural Network Architecture Search Methods For Polarimetric SAR Image Classification

With the increasing application of polarimetric SAR systems in civil and military fields,the requirements for polarimetric SAR image interpretation are getting higher and higher,so polarimetric SAR image classification has attracted the attention of many researchers.In recent years,with the development of deep learning,researchers have proposed many deep learning-based polarimetric SAR image classification models.Most of these models are obtained by manual design,which requires the designer to have rich experience in model design,and manual design is easy to introduce human errors,resulting in the problem of low classification accuracy.In addition,designing a deep neural network model for each polarimetric SAR image will bring a lot of labor costs.In response to the above problems,this paper proposes a series of deep neural network architecture search methods based on evolutionary computing around the polarization SAR image classification problem,which are summarized as follows:1.An evolutionary convolutional neural network architecture search method for polarimetric SAR image classification is proposed.In this method,the complex-valued convolution is used as the complex-valued feature extraction operator of polarimetric SAR images,the complex-valued convolutional neural network is encoded as an individual through a fixedlength encoding method,and the classification accuracy is used as the fitness of the individual.The population is optimized,and finally a complex-valued convolutional neural network with high classification accuracy is obtained.In order to obtain a suitable complexvalued convolutional neural network architecture,the proposed method searches in four dimensions: input resolution,network depth,convolution kernel size,and the number of output channels of convolutional layers.The super network is designed by adopting the design idea of network block,and the individual crossover is performed in the unit of genome in the evolution process to generate offspring individuals,thereby improving the search efficiency.The learning rate is adjusted through the learning rate decay strategy,so that the complex-valued convolutional neural network is fully trained.The experimental results show that the complex-valued convolutional neural network obtained by this method has the advantage of high classification accuracy.2.An evolutionary multi-object graph neural network architecture search method for polarimetric SAR image classification is proposed.The method uses complex-valued graph operators to model the non-local features of polarimetric SAR images,takes both classification accuracy and network complexity as optimization goals,and obtains a graph neural network with high classification accuracy and low complexity.In this method,a variable-length individual encoding method is designed,and the depth of the graph neural network is represented by the length of the individual,and the depth of the graph neural network is searched.In addition,a variable-length individual crossover method is also designed.By randomly selecting the crossover points in the two parent individuals,the generated offspring individuals have different lengths,thereby generating graph neural networks with different depths.During the mutation process,the genes are randomly initialized by probability to increase the diversity of individuals.Finally,it is proved by experiments that the graph neural network obtained by this method has high classification accuracy and low complexity.3.A federated evolutionary multi-objective convolutional neural network architecture search method for polarimetric SAR image classification is proposed.This method combines private data sources for automatic design of deep convolutional neural networks through horizontal federated learning.By replacing conventional convolutions with depthwise separable convolutions,the amount of parameters of the basic network is reduced,thereby reducing the resource consumption during model transmission.By taking the parameter quantity of the model as one of the optimization goals,the classification accuracy and the parameter quantity of the model are simultaneously optimized in the evolution process,which further reduces the model transmission cost in the federated learning scenario while improving the classification accuracy.Through the average aggregation method and the weighted aggregation method,the deep convolutional neural network quickly converges in the two-stage training process.Finally,through experiments on multiple polarimetric SAR images,it is proved that this method can reduce the transmission cost of the model to a certain extent while ensuring high classification accuracy under the premise of protecting data privacy.