PoISAR Terrain Classification Based On Bayesian Deep Learning

In remote sensing detection,Polarimetric Synthetic Aperture Radar(Pol SAR)has allweather,all-weather working characteristics.The use of Pol SAR images to classify features and terrain is widely used in urban planning,natural Disasters,prediction aspects of agricultural production.In recent years,with the rapid development of deep learning methods,the development of Pol SAR terrain classification has become more accurate and fast.However,the labeled samples of Pol SAR data are very rare,and the training process of deep learning methods usually requires a large number of labeled samples,which leads to the extremely limited application of deep learning methods in Pol SAR terrain classification.Aiming at the small sample problem in Pol SAR terrain classification,this paper applies fully convolutional network and Bayesian deep learning to Pol SAR terrain classification and makes a study of a series of classification methods based on the this model.The main work is described as following:First,A Pol SAR terrain classification method based on Bayes FCN model is proposed.This method is based on FCN,which is convenient to realize the pixel level classification task of Pol SAR data.It introduces the idea of Bayesian deep learning,initializes the weights and biases of all convolutional layers in the form of a Gaussian distribution,and realizes the backpropagation of bayesian by means of variational inference and local reparametrisation of convolutions.By combining the KL divergence of Bayesian learning with the cross entropy loss,the parameters of the network are learned together to improve the classification accuracy of the model.The training of the model only needs a small number of labeled samples,which effectively solves the problem of small samples.Second,A Pol SAR terrain classification method based on Bayes FCN-Mix Match model is proposed.Although Bayesian deep learning is used to learn the uncertainty of data,it is not used for unlabeled samples.In order to further improve the classification accuracy of the model,the Mix Match semi-supervised learning idea is introduced to enhance the data of the samples,and then calculate the average class probability of the unlabeled samples,and then obtain the minimized "guess" label of the unlabeled samples under the Sharpen algorithm.Unlabeled samples and labeled samples with "guess" label use Mix Up algorithm to mix the samples to get the training set of the model.The loss value is composed of supervised items,unsupervised items and KL divergence,so that the model can use a large number of unlabeled samples to assist training under a small number of labeled samples,and further improve the classification effect of the model in this way.Third,A Pol SAR terrain classification method based on Sparse Bayes FCN-Mix Match model is proposed.One obvious disadvantage of the Bayes FCN model is that the weights and biases of all convolutional layers follow the Gaussian distribution,and its parameters are changed from single values to mean and variance parameters.In the same network structure,the parameters of Bayes FCN model are twice as high as those of FCN model,and too high parameters can easily lead to model over fitting and affect the classification accuracy of the model.In this paper,on the basis of Bayesian deep learning and combined with sparse Bayesian method,the model has sparse characteristics.At the same time,in order to ensure the stability of the model,the mean teacher method is introduced to make the model parameters more smooth,so that the network improves the classification accuracy and more robust.