Polarimetric SAR Data Compressive Representation And Its Applications

As a passive remote sensing technique,Synthetic Aperture Radar(SAR)transmits and receives microwave pulses to obtain the terrain information all-time and all-weather.Polarmetric Synthetic Aperture Radar(PolSAR)has two transmit and receive mode,so the PolSAR data contain more information including the geometric chracteristics,scattering characteristics,texture characteristics and dielectric constant of the terrain objects.More and more PolSAR systems have come into service,and the techniques of PolSAR is developing,which makes PolSAR an indispensable and importance option in remote sensing observation.In the big data era,we are facing with a new chanllenge in the interpretation of the PolSAR image.To deal with the high-resolution and streaming PolSAR data,the scalability of the PolSAR data is taken into consideration.A good solution is compressively express the PolSAR data to obtain the balance of information preservation/task requirement and data storage and computation.In this thesis,the PolSAR polarimetric features compressive expression methods are exploited.The main work and contributions of this thesis are listed as follows:(1)The PolSAR data dimensionality reduction based on feature extraction is exploited.The manifold learning based and other dimensionality reduction methods are applied to the polarimetric feature data.The Principle Component Analysis(PCA),Isomap,Laplacian Eigenmaps(LE),t-SNE and undercomplete autoencoder are used to reduce the features dimensionality and visualize the features.(2)The compressive representation of PolSAR data based on the construction of coresets is exploited.Based on the markov chain model,the Bayesian coresets are obtained by using the weighted likelihood function to fit the original likelihood function.The Greedy Iterative Geodesic Ascent(GIGA)strategy is used to build the PolSAR features coresets by finding the optimal geodesic.(3)The compressive representation of PolSAR data based on sketch learning is exploited.Based on the mixture model estimated by sketch learning,a mixture-modeled PolSAR feature terrain classification framework is proposed.In this framework,the sketch learning is utilized to estimate the mixture model parameters of the homogeneous areas in PolSAR image to compressively represent the PolSAR data.The total Square Loss(tSL)divergence is used to measure the similarity of different mixture model in this framework.The framework is incremental which also available for streaming data and time-sequence PolSAR images.The compressive expression of the real PolSAR data are used in classification experiments.The experimental results demonstrate the effectiveness of the compressive representation strategies exploited in this thesis.