Research On Polarimetric SAR Imagery Superpixel Segmentation And Object-Oriented Classification

Synthetic Aperture Radar(SAR)technology,which can work all-weather and day-and-night,is an important branch of remote sensing technology.Compared with the single polarized SAR,polarimetric SAR(PolSAR)can record the complete polarization scattering information of featureobjects.Due to the sensitivity of polarized electromagnetic wave to target materials,physical structure,and geometric shapes,the classification accuracyof feature targets can beimproved greatly by using PolSAR measurement.However,because of the special imaging mechanism,the SAR image processing is difficult,and interpretation accuracy is low,limiting the wide application of SAR technology.This paper aims to improve processing speed and interpretation precision of PolSAR image.Based on the idea of object-oriented analysis,we made a deep research onsuperpixelsegmentation and object-oriented classificationof PolSAR image.The principal research contents and conclusions are as follows:1)Systematically research the basic theory of radar polarization measurement,including vectorization and matrix description of polarized electromagnetic wave,the statistic characteristics of PolSAR data,polarization synthesis,etc.Further study on the basic scattering mechanism modelsand characteristics of the typical targets.On this basis,further summarize and research the polarimetric target decomposition theory and the typical methods.2)Systematically researchexisting PolSAR image segmentation algorithms.To tackle the problem of slow speed and poor segmentation results of the existing algorithms,the simple linear iterative clustering(SLIC)algorithm for the optical image is introduced into the PolSAR image processing field.Then some processing steps of SLIC such as distance measurement,initializationof the clustering center,post-processing,etc was improvedaccording to the characteristics of PolSAR image.Based on these improvements,a superpixel segmentation algorithm named PolSLIC was proposed.Finally,two groups of airborne L-band PolSAR data are used for experiments,proving that the proposed algorithm has obvious advantagesin terms of processing speed and segmentation effect comparing with the original SLIC algorithm and the commonly used segmentation algorithms in thePolSAR field.3)Systematically researchexisting PolSAR image classification algorithms.The Restricted Boltzmann Machines(RBM)and the Adaptive Boosting(AdaBoost)framework are combined.On the one hand,the object-oriented ideas are used to overcome the influence of speckle noise in PolSAR image and speed up the processing speed.On the other hand,amultiple classifiers integration framework based on deep learning algorithm is used to overcomethe problem thatit is difficult to achieve highclassification precision by a single classifier due to the complexpolarimetric scattering mechanism oflfeaturetargets in PolSAR image.The finally proposedRBM-AdaBoost algorithm avoids the demand for a mass of data,thus more suitable for object-oriented approach.Experiments by using airborne L-band PolSAR data proved that the proposed method is better than the stacked RBM model and other commonly used PolSAR classification methods in terms of classification accuracy.Althoughsome research resultsin terms ofPolSAR image superpixel segmentation and object-oriented classification has been made,but there are still someproblems remain to be further studied and discussed.Such as in aspect of superpixel segmentation,how much the pre-processing step such asfiltering affects the segmentation results,how much the post-processing step affectsthe segmentation results,and how tofurther introducethe texture properties and the polarimetric scattering properties of PolSAR data into the proposed PolSLIC algorithm.In terms of object-oriented classification,the proposed classification method does not consider the statistical properties and the polarimetricscattering characteristicsof PolSAR data,and the execution timeof the algorithm is too long,the efficiency is not satisfactory.