Research On Polarimetric SAR Image Classification Based On Deep Convolutional Neural Network Control

Polarized Synthetic Aperture Radar(PolSAR)is an advanced microwave detection system plays an irreplaceable role in civil and military fields.Among them,polarization SAR image scene classification is one of the important tasks of PolSAR image interpretation,and has received more and more attention.The traditional method of scene classification usually restricted by prior knowledge and noise,satisfactory accuracy can not be obtained.The deep learning method can automatically extract the essential features from the data and use high-level spatial information and polarization information to reduce noise interference,and achieve high-precision classification of PolSAR images.This thesis explored the PolSAR image classification method based on deep convolutional neural network.The main contents are as follows:1.An ensemble-transfer framework based on deep convolutional neural networks is proposed.The method makes full use of the polarization information and spatial information in the image to achieve high-precision classification on small samples.Firstly,the effects of three advanced polarization decomposition methods on the classification of PolSAR images are studied.Then three classical models are transfer to polarization SAR image classification.Finally,the appropriate model is used to integrate the advanced polarization parameters.2.Visualize the working mode of convolutional neural networks in image classification,explore the processing methods of different modules for different data and the reasons for some phenomena,and verify the applicability and reason of transfer optical models to SAR image classification.3.A multi-scale and multi-polarization semantic segmentation model based on full convolutional network is proposed.The model combines high-order polarization information and multi-scale spatial information to achieve smooth semantic segmentation on small samples.Firstly,the classification network with fully connected layers is improved into a semantic segmentation network capable of retaining image space information and the PolSAR image is semantically segmented using pre-training parameters.Then the polarization decomposition parameters are merged with the original data to realize multi-polarization multi-scale semantic segmentation.