Polarimetric SAR Image Classification Based On Convolutional Neural Network

Polarimetric synthetic aperture radar(POLSAR)image refers to high-resolution imaging obtained by radar based on synthetic aperture principle and receiving electromagnetic waves of multiple polarization combinations.It has more abundant polarization scattering information.Polarimetric SAR image classification is of great significance for the interpretation of polarized synthetic aperture radar images.It has been widely used in natural disaster detection,observation of crop growth,and ship tracking and identification.In recent years,the deep learning method has developed rapidly in the field of computer vision.The deep learning network acquires image structure information more comprehensively through multi-layer training of input samples.The deep convolutional neural network plays an irreplaceable role in the classification and recognition of images with the small amount of local connection weight sharing.The thesis analyzes the application of convolutional neural networks in polarimetric SAR images.The main work are as follows:A supervised classification method for polarimetric SAR images based on convolutional neural networks is presented.According to the structure of convolutional neural network and the training characteristics of multi-class data sets,the multi-channel feature combination of polarimetric SAR data is constructed.Firstly,the polarization coherence matrix is subjected to Cloude decomposition to obtain the polarization feature set containing scattering information,and combined with color features.And a training sample set that obtains multi-dimensional features of the texture features representing the spatial structure information of the polarimetric SAR image,and learns the multi-layer feature information through the convolutional neural network training,and then tests the polarized SAR image by using the trained network model to obtain the final classification result.In order to reduce the redundancy between features and further shorten the training time,a more reasonable and effective feature combination is obtained.Based on multi-feature classification,a supervised classification algorithm for polarimetric SAR images with feature selection and convolutional neural network is implemented.The original feature parameter set is obtained based on the polarization SAR image data and target decomposition.Then the random forest method is used to evaluate the importance of the feature parameter set,and the optimal polarization feature is selected according to the feature importance ranking.The optimal polarization characteristics are taken as input,and finally the image classification is realized by the convolutional neural network classifier.The effectiveness of the test is verified by two sets of measured data collected by the US AIRSAR airborne system,and compared with the existing classical supervised classification algorithm.The results show that the convolutional neural network method can improve the classification accuracy and after feature selection.The training time is shortened and the amount of calculation is smaller.