Ensemble Learning Based Full Polarimetric SAR Image Classification

Due to its advanced capabilities such as active, day-and-night, full wether and range look imaging capabilities, also the captured signal carrying abundant and specific physical meanings, since the commercialize operation high resolution spaceborne polarimetric SAR systems such as RADARSAT-2, ALOS PALSAR, TerraSAR-X, COSMOS-SkyMed and Sentinal-1 have been successfully launched, spaceborne polarimetric SAR images have been more widely used in the contex of remote sensing application for disaster monitoring, forestation and deforestation, urbanization, agricultural mapping etc. From the technical point of view, the spaceborne polarimetric SAR image classification and interpretation are among most important ones. Generally, developing new classifiers and optimial feature extraction methods are the two main directions in polarimetric SAR image classification. Although, plenty of classifiers have been proposed in the last couple decades, yet no one can guarantee that one or some classifiers have generalized capabilities that suit for all polarimetric SAR image classification circumstances. Besides, polarimetric decomposition methods can’t successfully separate multiclass objects in complex environment. In addition, speckle filters, feature extractors and feature selectors are the key components in polarimetric SAR image classification task, but the complementarity information between various filters, extractors and selectors has not been extensively investigated, especially under the framework of ensemble learning. New machine learning methods such as extreme learning machine, kernel extreme learning machine, and their ensemble learning based extended versions need to be studied in polarimetric SAR image classification. To this end, also according to the National Natural Science Foundation of China under Grant No.41171323, needs for classifying polarimetric SAR image more accurately and reliably, utilizing the complementarity information between various speckle filters, feature extractors and selectors, and investigating the capabilities of extreme learning machines in ensemble learning framework, is practically meaningful for extending the ensemble learning in polarimetric SAR image classification. Finally, according to the experimental results the obtained conclusions are:(1) A filter diversity based ensemble learning framework was proposed. In details, this framework including single filter single window multiple classifiers ensemble, same filter multiple windows same classifiers ensemble and multiple filters multiple windows multiple classifiers ensemble, are proposed based on the diversity criterion in ensemble learning. Finally, all the experimental results confired their advanced performance on classisification of PolSAR image, especially, multiple filters multiple windows multiple classifiers ensemble always capable of reaching the highest classification accuracy with more special details.(2) In consideration of the diversity among various features such as polarimetric, spatial and decomposed ones, feature extractors and selectors were used to construct feature space diversity based ensemble for a better classification accuracy purpose. The results showed that morphological profiles, Shannon entropy, eigenvalues of coherence matrix, polarimetric anisotropy, volume and surface scattering are the best for polarimetric SAR image classification. The feature selectors, extractors and classifiers diversity baed ensemble methods are capable of obatain better results. More in details, feature selection diversity baed ensemble better than feature extraction diversity based ensemble, the best classification results shown by the feature selection and extraction diversity based ensemble.(3) Extreme learning machie, kernel extreme learning machines based PolSAR image classification were investigated, better results reached by their extensions such as bagging based extreme learning machines, adaboost based extreme learning machines and the boosted multiple kernel extreme learning machines.(4) Experimental results on the Shang Ku Li research area and otherwidely used test PolSAR images confirmed the advantages of utilizing the properties of different features and and classifiers, that is ensemble learning capable of leading to better classification performance than the conventional polarimetric decomposition and supervised Whishart classifier based methods. In conclude, the proposed methods in this paper, the filters diversity based ensemble, features diversity based ensemble and ensemble learning based extreme learning machines have a great prospects in the context of PolSAR image classification.