Synergetic Classification Of Hyperspectral Data-based Multi-source Remote Sensing Images

With the rapid development of sensor technology and the wide application of computer technology,conventional single type remote sensing data turns to be unable to satisfy the increased application requirement.Thus,the comprehensive utilization of multi-source remote sensing images attracts more and more attention of human researches.Different spectrums and resolutions usually reflect the attributes and features from different aspects,thus,taking full use of the complementarity and differences between multi-source remote sensing images will maximize the advantages of multi-source remote sensing images.Hyperspectral(HS)image has become the most important technical method in multi-source remote sensing images.Depending on its fine spectral resolution and the feature of combining image with spectrum,HS image presents its unique advantages in land-cover classification and target detection.Nevertheless,the deficiency of spatial resolution,and the problem of “the same object but different spectrum” and “the same spectrum but different objects” of HS image lead to the failness in solving lots of classification problems with high-accuracy demand.With the emphasis on HS image,this dissertation focuses on the synergetic classification of multi-source images including high-resolution panchromatic,multispectral,hyperspectral,and infrared hyperspectral images,and analyze the the problems and challenges in multi-source image synergetic processing.In summary,this dissertation has remarkable theoretical significance and reference value in enhancing the ability of data utilization and information acquisition,as well as the extension of the remote sensing application.The main research content of this dissertation includes the following parts.First,since the HS image generally has lower spatial resolution,this dissertation proposes a spatial resolution enhancement method based on spectral modulation for HS image.Due to the higher spectral resolution and wider spectral range of HS image,conventional multispectral pan-sharpening methods are turned out to be unsatisfactory.Therefore,we start with the multispectral image pan-sharpening,combining multi-level inter-band structure model with non-subsampled contourlet transform to effectively enhance the spatial details of the multispectral image.Furthermore,this dissertation investigates the spectral modulation-based multispectral and HS image fusion method to enhance the spatial resolution of the HS image by employing spectral unmixing method and the spectral-spatial information of the two images.After that,we also investigate the classification methods for HS image.Experimental results indicate that the proposed approach can effectively minimize the spectral distortion of the fused image,and improve the classification results.Second,due to the problem of “the same object but different spectrum” and “the same spectrum but different objects” of HS image,this dissertation proposes a synergetic classification method of HS and panchromatic images based on semi-supervised rotation forest.For panchromatic image,we use the object-oriented spatial feature extraction method,to overcome the limitation when using only spectral feature to describe landcovers.Then,random subspace-based ensemble learning method is employed to conduct spatial-spectral feature classification to overcome the potential feature-redundancy and curse of dimensionality problem for multi-source images.Furthermore,since the principal component analysis of existing rotation forest method commonly neglect the discriminative information of labeled samples,we use the proposed semi-supervised rotation forest algorithm to enhance the classification accuracy.Experiments on four simulated and real data sets show that the proposed method can take advantage of both the discriminative and structure information to improve the discriminative ability of the classifiers and obtain better results,compared with some state-of-the-art ensemble learning methods as well as some other typical classification methods.Particularly,this dissertation proposes a semi-supervised self-learning-based synergetic classification algorithm for HS and panchromatic images to tackle the small sample size problem of HS image classification.Due to the insufficiency of the training samples in RS images,and the high cost of human labeling,this dissertation discusses the characteristics of semi-supervised learning and active learning algorithms respectively,and take advantage of the two techniques,and integrate panchromatic image segmentation technique into HS image classification.By combining the predicted results of the classifier and spatial-spectral features,the algorithm can select and discriminate samples automatically.Thus,no extra cost of human expertise is required for labeling the selected pixels when compared with conventional active learning methods,which realizes the synergetic classification task of HS and panchromatic images.Furthermore,we also investigate the diversity of samples to improve the efficiency of the semi-supervised learning method.Experiments indicate that the proposed method can utilize the spatialspectral features,and effectively improve the classification accuracy through an iterative style.Last,we propose context-based sparse representation classification fusion method for infrared HS image and visible multispectral image.Based on the conventional HS image processing method,this dissertation further analyzes the features of the infrared HS image,and investigates the synergetic classification methods with visible multispectral image.Infrared HS image differs from the traditional HS image,as it is obtained according to the emissivity information of land-covers.Therefore,by analyzing the individual characteristics of the infrared HS and visible multispectral images,we propose to adopt a hierarchical processing strategy.After extracting the spectral,spatial,temperature and emissivity features respectively,and applying sparse representation,and joint sparse representation classifications,the classification results are merged by decision level fusion techniques based on context information of the images.Experiments indicate that the proposed method can take full use of the complementary information of images with different spectral ranges,provides higher accuracies and more interpretive results.