Fusion And Classification For Hyperspectraland SAR Image Based On Machine Learning

With the speedy evolution of remote sensing technology observing the earthsurface from a long distance and the wide application of many different kinds ofsatellite sensors, there are more and more remote-sensing images coming from thesame region which have their own characteristics and drawbacks. And nowadays itis already difficult to satisfy the complicated requirement in actual practice simplybased on only one image data source. Therefore image fusion becomes one of themost significant topics in remote sensing researches, mainly by comprehensivecombining different data sources to obtain enough reliable information.In this dissertation, hyperspectral and SAR images are applied in the researchon image fusion and classification. In particular, hyperspectral images containrelatively high spectral resolution and a large number of bands, including richspectral features of land surface. As for polarimetric SAR images, the sensors couldkeep working day and night without the time and weather restriction to getabundant polarized scattering information and texture features. By making full useof these complementary advantages based on machine learning theory, some specialalgorithms are studied to realize the precise classification of land cover, and at thesame time to gain the natural overall recognition of interesting targets. Thedissertation is arranged as the following three aspects:Firstly, hyperspectral and SAR images are investigated for feature extractionaccording to their own characteristics, which provides accurate data basis forfollowing researches. Here traditional PCA and LDA methods are applied forhyper-spectral images and compared with SVM-PP algorithm which combines theSupport Vector Machine and Projection Pursuit theory reasonably, especially forsolving the problem of small training set in classification. Particularly, it isproposed to use genetics algorithm to select the best feature set from the relativelyencoding table. In addition, H/A/α decomposition, Freeman decomposition andYamaguchi decomposition are applied in feature extraction and selection forpolarimetric SAR images.Secondly, the study about the feature level combination is based on machinelearning. In order to solve the problems of insufficient number of training samplesand traditional classifiers unable to effectively fuse the features from hyperspectraland polarimetric images, one adaptive classification scheme called MGBL-AL isproposed as follows: firstly, the MGB-LOOC basic classifier is built based onMeta-Gaussian joint distribution and Leave-one-out-covariance method to deal with multi-source image data; then active learning theory could actively select the mostinformative unlabeled samples based on different query rules, and add them into thetotal training set to further improve the classification accuracy.Finally, the research on decision level combination is based on the thought ofensemble of classifiers learning. Considered the highest decision level to mergehyperspectral and polarimetric images, the idea of ensemble learning is utilized tocombine the results from different classifiers to make the final classificationdecision, to achieve more reliable recognition. And the Gentle AdaBoost.MHalgorithm based on simple decision tree classifier is improved and analyzed indetail, when compared with other methods based on the Random Forests thought.