Different methods proposed so far for accurate classification of land cover types in polarimetric synthetic aperture radar (SAR) image are data specific and no general method is available. A novel hybrid framework for this classification was developed in this work. A set of effective features derived from the coherence matrix of polarimetric SAR data was proposed. Constituents of the feature set are wavelet, texture, and nonlinear features. The proposed feature set has a strong discrimination power. A neural network was used as the classification engine in a unique way. By exploiting the speed of the conjugate gradient method and the convergence rate of the Levenberg-Marquardt method (near the optimal point), an overall speed up of the classification procedure was achieved. Principal component analysis (PCA) was used to shrink the dimension of the feature vector without sacrificing much of the classification accuracy. The proposed approach is compared with the maximum likelihood estimator (MLE) based on the complex Wishart distribution and the results show the superiority of the proposed method, with the average classification accuracy by the proposed method (95.4%) higher than that of the MLE (93.77%). Use of PCA to reduce the dimensionality of the feature vector helps reduce the memory requirements and computational cost, thereby enhancing the speed of the process.
It is very important to efficiently represent the target scattering characteristics in applications of polarimetric radar remote sensing. Three probability mass functions are introduced in this paper for target representation: using similarity parameters to describe target average scattering mechanism, using the eigenvalues of a target coherency matrix to describe target scattering randomness, and using radar received power to describe target scattering intensity. The concept of cross-entropy is employed to measure the difference between two scatterers based on the probability mass functions. Three parts of difference between scatterers are measured separately as the difference of average scattering mechanism, the difference of scattering randomness and the difference of scattering intensity, so that the usage of polarimetric data can be highly efficient and flexible. The supervised/unsupervised image classification schemes and their simplified versions are established based on the minimum cross-entropy principle. They are demonstrated to have better classification performance than the maximum likelihood classifier based on the Wishart distribution assumption, both in supervised and in unsupervised classification.