Classification Of PolSAR Image Based On Fuzzy PSO With Target Decomposition

Polarimetric Synthetic Aperture Radar(Pol SAR) has been meeting the requirements in the aspect of acquiring imagines, which are free of light, weather and other reasons, for a full 24 hours, so polarimetric SAR is widely applied in military and civilian life. Nowadays there are increasingly many polarimetric SAR images, and its abundant information, polarimetric SAR image processing and interpretation have played more and more important role in national defense and economic development. In this thesis, In this thesis, to the question of polarimetric SAR classification, three new features on classification are proposed.Firstly, one modified classification method based on fuzzy particle swarm optimization with scattering decomposition is proposed. Based Cloude Decomposition theorem, we use FCM clustering algorithm instead of the traditional classifier with complex Wishart is adopted to improve the clustering center iterative update method. In this way, we can reduce the computational complexity and make the program running time short. And aiming at defects of the FCM clustering algorithm, we use the iterative optimization process of Particle Swarm Optimization(PSO) instead of gradient descent of Fuzzy C-means(FCM), which can avoid the original cluttering centers trapping in local optimum.. Ultimately, making clustering centers achieve the global optimal and achieve better classification results. The experiments show that the improved method for the classification is more accurate.Secondly, we proposed a method with combination of fuzzy particle swarm optimization(FPSO) and the polarization target decomposition, continuing to replace the conventional iteration of cluttering center with the complex Wishart classifier by the method mentioned above.At the same time, for the roll-invariant of the Cloude decomposition, the polarization characteristics of the targets are extracted. However, neither can the Cloude decomposition give us the details of targets without using the amplitude information for imaging nor it preserves polarization scattering characteristics. So we extract scattering characteristics by the Freeman decomposition for this problem. Combined the characteristics introduced by two incoherent decomposition, classification is more neutral closed to what targets are.Finally, aiming at improving the classification accuracy, we introduced the quantum particle swarm optimization(QPSO), and presented a novel terrain classification method based on Fuzzy Quantum Particle Swarm Optimization and the target decomposition theorem.Employing advanced PSO, QPSO, to substitute gradient descent, it is an improvement of the two methods. Combining the theory of quantum-motion with PSO, and replacing the PSO with QPSO, not only does the algorithm have the capacity of avoiding the degradation of completely random optimization, but it can deal with the problem of premature convergence.Through the comparison for these two methods above, we found the new algorithm is able to achieve higher classification accuracy.