PolSAR Terrain Classification Based On Deep Learning And Architecture Search

Polarimetric Synthetic Aperture Radar(Pol SAR)is a kind of active microwave remote sensing sensor.Compared with passive imaging methods such as visible light imaging,Pol SAR has many advantages such as working in all-day and all-weather conditions.So,it is a very important means of data acquisition,and it is also a technological high point of competition among countries.Besides,Pol SAR terrain classification is an important task of Pol SAR image interpretation.In recent years,with the development and wide application of deep learning,many experts and scholars have proposed to use deep learning models to deal with Pol SAR terrain classification and achieved good results.However,the classification accuracy of the deep learning model is very much related to the hyperparameter setting/network architecture design.If they are not properly set,the model can hardly achieve the ideal classification accuracy.Unfortunately,there is no theoretical guarantee for the hyperparameter setting/network architecture design of the model.It is usually determined by human experts’ experience or by trial-and-error.In order to solve the above problems,some scholars have proposed hyper-parameter optimization and neural architecture search methods in the field of visible light image interpretation to automatically set hyperparameters of the model and design architectures of the network,which greatly save manpower.However,there are only a few works in the field of Pol SAR image interpretation,and they can only automatically set part of the hyperparameters in convolutional neural networks.So,the automatic hyperparameter setting and network architecture design are not realized in the real sense.Therefore,this dissertation explores and designs the corresponding hyperparameter optimization and neural architecture search methods for deep learning models related to the task of Pol SAR terrain classification,aiming to realize the automatic hyperparameter setting and network architecture design.At the same time,for the problems of the high time and space cost of the architecture search method and the difficulty of designing a network architecture with a good tradeoff of classification accuracy and inference latency,an improved scheme is proposed to address these problems,which further improves the practicality of the neural architecture search methods.The specific research results are as follows:(1)For the problem of insufficient feature extraction capability of the autoencoder model and the difficulty of setting hyperparameters of the model properly in Pol SAR terrain classification,this dissertation proposes a Pol SAR terrain classification method based on stacked autoencoder and multi-objective optimization to solve the problem.First,in the pre-training stage,the method uses reconstruction loss and cross-entropy loss to train the autoencoder layer by layer,which drives the model to learn features that are beneficial to improving the classification accuracy and improves the feature extraction capability of the model.Then,for the difficulty in setting the hyperparameters(the number of layers,the number of nodes in each layer,the type of activation function in each layer,the balancing parameter in the loss function,and the number of training epochs)of the autoencoder properly,a hyperparameter optimization method with the help of a multi-objective optimization algorithm for the autoencoder is designed to achieve automatic setting of the hyperparameters of the autoencoder.At the same time,the proposed method also optimizes the initial value of the weight of each layer,reducing the risk of falling into local optimum during training.(2)For the difficult problem of the architecture design of convolutional neural networks in Pol SAR terrain classification,this dissertation proposes a Pol SAR terrain classification method based on convolutional neural networks and differentiable architecture search,which realizes the automatic architecture design of convolutional neural networks(the type of operations in each layer and the number of layers).The method proposes a search space with a variable number of layers and constructs a supernet based on it.Then a gradient-based architecture search framework is used.The method using architecture parameters to represent different sub-networks,designing the corresponding loss function,taking the architecture parameters as independent variables of the loss function,and optimizing the weights and architecture parameters of the network through the gradient descent method to realize the automatic architecture design of the convolutional neural network.(3)For the difficult issue of architecture design of fully convolutional networks in Pol SAR terrain classification,this dissertation proposes a Pol SAR terrain classification method based on fully convolutional networks and differentiable architecture search,which realizes the automatic architecture design of fully convolutional networks(the type of operations in each layer and the number of layers).The method proposes a three-level search space,including intra-cell,inter-cell,and the layer-num search space.Meanwhile,a supernet is constructed using the three-level search space.Different sub-network architectures are represented by the architecture parameters corresponding to the intra-cell,inter-cell,and the layer-num search space.The architecture parameters are regarded as independent variables of the training loss function.Then a gradient-based architecture search framework is used.The weights and architecture parameters of the network are optimized by a gradient descent method to achieve the automatic architecture design of the fully convolutional network.Meanwhile,in order to avoid selecting the input features and determining the number of input features manually,and further improve the model classification accuracy,this dissertation also proposes a feature attention module.The network should take all features as the input in the training stage,and the stem layers can automatically learn the weight of each feature and realize the automatic selection of input features by using this module in the first few layers of the network(the stem layers).(4)The architecture search method of fully convolutional networks is the one having the highest classification accuracy among all the architecture search methods.For its bottleneck problem of high time and space cost,this dissertation proposes a Pol SAR terrain classification architecture search method based on the improved search space and feature selection method.We find that the intra-cell search space is more important than the inter-cell search space and the inter-cell connection has a little effect on the model classification accuracy.Therefore,the simplification of the search space can be realized by deleting the inter-cell search space,which significantly reduces the time and space cost of the supernet as well as the search method.Meanwhile,for the architecture search method of fully convolutional networks with feature selection,this dissertation finds that taking all features as the inputs of the network results in a problem that there is a large amount of input data needing to be loaded,which reduces the efficiency of the network in the training stage and that of predicting the full image.Therefore,this dissertation proposes an unsupervised feature selection method to filter the input features,which significantly reduces the amount of input data and improves the efficiency of the network in the training stage and that of predicting the full image with only a small loss of classification accuracy(less than 1%).(5)In practical application scenarios,the model is not only required to have high classification accuracy but also low inference latency.Although the above methods can automatically design architectures and achieve the state-of-the-art classification accuracy,they cannot search for network models with a good tradeoff between classification accuracy and inference latency.Therefore,this dissertation proposes a Pol SAR terrain classification method based on joint search of classification accuracy and inference latency to automatically design network models with a good tradeoff.First,the method improves the intra-cell search space so that the search method can search architectures within a search space with operations having a good tradeoff.In addition,the method proposes a loss function associated with the network inference latency,which explicitly guides the search method to retain sub-networks with a good tradeoff in the search stage.Finally,the method can realize the automatic design of network architectures with a good tradeoff.