Polarimetric SAR Image Classification Based On Machine Learning

Polarimetric Synthetic Aperture Radar(PolSAR)is a kind of high resolution coherent imaging system which can transmit electromagnetic waves actively and receive its echoes.The imaging system has the capability with all-time and all-weather.Due to the long wavelength of electromagnetic waves,the SAR system is rid of the influence of the meteorological conditions such as clouds and haze like the optimal imaging system.The PolSAR images contain rich features scattering information and texture characteristics and dielectric constant information which can greatly enhance the recognition ability of the object.We studied the problem of PolSAR image classification based on machine learning in this thesis with the aim of improving the level of PolSAR data automatic interpretation.There is of important significance in environmental monitoring,resource exploration and military reconnaissance.In this thesis,we studied the PolSAR image classification technology under the framework of supervision and non-supervision.Our main work and contributions are summarized as follows:(1)We proposed a supervised PolSAR image classification algorithm which is based on ensemble learning.Taken both the theoretical background of ensemble learning and the characteristics of PolSAR image imaging into account,an ensemble learning framework is designed.The proposed algorithm utilized SVM as the basic learner and Random Forest as a meta-learning device to combine the results.The experimental results show the effectiveness of the ensemble framework in classification accuracy.(2)We proposed an unsupervised PolSAR image classification algorithm by combining feature selection and large scale spectral clustering.The clustering Forest feature selection algorithm is used to reduce feature dimension and redundancy.The image clustering is realized by the large-scale spectral clustering algorithm,which integrates the spatial information by super-pixel segmentation.(3)We proposed a supervised PolSAR image classification algorithm based on deep learning(CNN).A convolution neural network structure applied to PolSAR image is proposed,which is a combination of the LeNet and NIN network.The input of the network is data of two channel consisting of the real and imaginary parts to reserve data information maximally.The full connection layer with lots of parameters is replaced by the global mean pooling layer.From the experimental results,we learnt that the classification performance of the algorithm is better than the competitors,which show the reasonability and feasibility of the network design(4)We proposed an unsupervised PolSAR image classification algorithm based on deep learning(SAEs).The density peak clustering algorithm is used to select the representative points of each super-pixel.Then the low-dimensional manifold characteristics are obtained through spectral clustering.The clustering of PolSAR image is accomplished by using the deep embedded clustering structure.The experimental results show the necessity of selecting the representative point with DPC algorithm and the introduction of DEC can further improve the classification accurancy.