Research On Land Cover Classification And Road Networks Extraction Of SAR Imagery Based On Deep Learning

Synthetic Aperture Radar(SAR)images interpretation which was applied in military surveillance,intelligent control,oceanographic observation,disaster monitoring,geology surveying,resource exploration,precision agriculture and environmental protection,has been a hot research area of computer vision and image processing for many years.Recent several decays,lots of scholars engage in the reach of SAR images interpretation method.The features extraction always take a lot of priori knowledge in the detection algorithms in the past.Deeply relying on the selecting the appropriate algorithm and setting the suitable parameters,the detection or classification results is quite discrepant when setting the different parameter even using the same algorithm.Recent years,as data explosion,algorithms based on deep learning,are widely used in automatic analysis of SAR images.Deep learning models establish on tremendous data,make the machine to learn as human.The output of model is used in generating the loss which descend step by step when optimizing the model.Convolutional Neural Networks(CNN),a kind of Neural Network,directly receive the images as inputs,are constructed of several convolution layers used in features extraction and some full connected layers end up with a softmax layer.As avoiding the complicate images pre-processing,CNN simplify SAR images interpretation.In this dissertation,two algorithms based on deep learning are proposed to apply in the SAR land cover classification and road networks extraction.The innovation and research results are as follows:In order to solve the problem that the past land cover classification algorit hms cannot reach the state of the art results in small sample sets.A Semi-Sup ervised Learning(SSL)method combines the Support Vector Machine(SVM)and Convolutional Neural Networks(CNN)is proposed in the condition of lacking data,to reach the similar precision as the supervised convolution neural networ k.The features of images are composed of gray-scale feature,radar cross secti on,Garbor Wavelet Transform,Gray-level co-occurrence matrix.Principal Com ponent Analysis(PCA)is proposed to reduce the dimension of features before t raining the SVM model,which is used in labeling the raw data set for generat ing the training set of CNN model according to the confidence coefficient.The loss is computed in the form of cross entropy after the forward propagation a nd optimized through backward propagation using Stochastic Gradient Descent(SGD)algorithm.It turns out that the SSL only utilizing the small sample set can reach the similar accuracy as the supervised CNN model.The other part of this dissertation is road network detection of SAR images using Generative Adversarial Network(GAN)to solve the problem of inevitable Speckle Noise interference in past algorithm.Meanwhile the parameters are also hard to set.Training set is composed of several pairs,SAR images and their counterpart road images depicted formerly.The generator maps the SAR images to road images which try to confuse the discriminator.The mission of discriminator is distinguish the inputs are generated by the generator or the road image from training set.The loss is optimized in backward propagation and the performance of either generator and discriminator is improved during the adversarial system.The SAR road network is generated using sliding window in SAR base map.Fuzzy C Means algorithm is acquired to suppress the noise generated by GAN.It turns out that GAN can reconstruct more road networks than past algorithms.