Classification And Recognization For Urban Scenes In High Spatial-resolution Hyperspectral Image

Hyperspectral remote sensing is one of the significant achievements in theearth observation technology of recent decades. The hyperspectral images with highspectral resolution can provide so much spectral information that they are usedwidely in many areas, but the application on analyzing the urban scenes is limiteddue to the low spatial resolution. The urban scenes are distributed densely and theirareas are generally small. It’s difficult to distinguish them with low spatial-resolution images. With the development of the hyperspectral sensors, the spatialresolution of images improves a lot that many scenes can be described by pixels,which makes the analysis on the urban scenes possible using the hyperspectralimages with high spatial-resolution. In the paper, we will take advantage of thehyperspectral images with high resolution to make research on the classificationand recognition of the city scenes. The dissertation is arranged as the followingthree aspects:Firstly, hyperspectral images are investigated for feature extraction accordingto their own characteristics. Based on the traditional methods-the local fisherdiscriminant analysis (LFDA) and neighborhood preserving embedding (NPE), weproposed a semi-supervised local discriminant analysis (SLDA) which combinesthe above two methods. The SLDA method takes the discriminant information oflabeled samples and structure information of unlabeled samples into consideration.We will classify the urban scenes using the support vector machines (SVM) andmaximum likelihood classification (MLC) with the features extracted by theproposed SLDA, and compared its performance with other methods to test thevalidation of SLDA.Secondly, to make full use of the high-resolution character, spatial information,like morphological features and the shape feature, will be extracted. Then we willmake research on the combined way of spectral and spatial features. There are threeways: combining the two features directly; based on the kernel function, combiningthe different features in the kernel-transformed space; based on the multi featurescombining (MFC) framework, making dimensionality reduction on differentfeatures and combining them by preserving information as much as possible.Classify the images using SVM, and the result shows that the performance ofspectral-spatial feature is better than single feature.Finally, aimed to solve the problem of lacking training samples which may exist in the classification of urban scenes, we will make research on supervisedclassification combined with active learning. Based on the analysis of traditionalactive learning methods, we propose a method of determining the candidatesamples set making use of the labeled samples and initial classification result. Theadvantage of the proposed method is avoiding the manual labeling of the newsamples selected by the active learning algorithm. Moreover, to get the output ofprobability, we research on a multinomial logistic regression (MLR) classificationmethod based on the logistic regression model. Then we make a test on theproposed method of determining candidate set by classifying the urban scenes. Theresult demonstrates that the proposed method not only saves manpower, butimproves the classification accuracy effectively in the case of lacking labeledsamples.