Reseach On Classificaiton Algorithms Of Polarimetric Synthetic Aperture Radar Image

Polarimetric synthetic aperture radar, short for PolSAR, has become one of the world’s most advanced technologies in international remote sensing field for earth observation. And the classification processing of PolSAR images has become an important role of the application of remote sensing technology. Imagery classification results have been widely used among the field of agriculture, land use, forestry, meteorology, geology, oceanography, sea ice detection and so on. Polarimetric coherence matrix and three polarimtric parameters, namely polarimetric scattering entrophy, average scattering angle and anistrophy, are taken into account as image features,then target decomposition theme, fuzzy clustering algorithm and BP neural network algorithm are applied in research on claasification of PolSAR image in this dissert.First, Research on classification of PolSAR image based on target decomposition is conducted.Unsupervised classification schemes consisting of three polarimetric parameters H, A, alpha formed from Cloude-Pottier decomposition, which have characterristics of rotation invariance independent of coordinate system, are pointed out in summary. Then the distribution of Covariance matrix of multi-look PolSAR data which obeys complex Wishart distribution is taken into consideration, so two algorithms, one of which is a Wishart classifier based on H/alpha and the other is a Wishart classifier based on H/A/alpha, are listed separately. Besides point out that Wishart classifier based on H/A/alpha is the best in using the three polarimetric parameters. Finally, a supervised classification algorithm based on Wishart distribution is introduced and a quantitative evaluation for the algorithm performance is measured by producer accuracy, user accuracy, overall accuracy and Kappa coefficient based on confusion martrix.Second, Research on classification of PolSAR image based on fuzzy theory is conducted.Introduce fuzzy theory into H/alpha plane, a comparative analysis result between unsupervised classification algorithm based on H/alpha and unsupervised classification algorithm based on H/alpha using fuzzy clustering algorithm is made. The result indicates the new algorithm has better performance. Then apply the FCM algorithm in classification of dual polarimetric and single polarimetric SAR image. The result show the more images feature, the better accuracy. Considering that the FCM algorithm is sensitive to noise and initial center, an algorithm combined Otsu with KFCM is proposed. Experiment results show the new algorithm has tolerance with noise and outliers to some extent and has better performance. And the parameters in algorithm can be optimized further for better accuracy.Third, Research on classification of PolSAR image based on neural network is conducted.Using H, A, alpha and coherence matrix as image features, a comparative analysis between a supervised classification algorithm based on BP neural network and a supervised classification algorithm based on Wishart classifier is conducted. Experiment results show that BP algorithm outperforms traditional Wishart maximum likelihood classifer. BP neural network is sensitive to initial weights and thresholds, falls into local minima easily and converges slowly, so gene algorithm is used to optimize the weights and thresholds of BP neural network. Experiment results show that the performance of BP neural network optimized is better than BP neural network algorithm, and noted that parameters in the algorithm can be researched further.