| Currently, hyperspectral remote sensing (HRS) has become a cutting-edge field of remote sensing (RS) area. Hyperspectral image classification has become one of the hottest research fields of the remote sensing area. The hyperspectral classification technology has very high value both in application and theory research. However, hyperspectral image processing also introduced many problems. The problems are: how to solve the problem of high-dimensional with small amout of samples, how to use the information ignored in current research, such as structural information, how to use the vast amounts of unlabeled samples that contain information, how to choose the fittest classifier. In HRS applications, the requirements of hyperspectral classifiers performance are increasingly. The corresponding hyperspectral classification technology research should focus on the topic, connecting with the specific application of hyperspectral remote sensing. This research includes the following aspects:First, in order to deal with the problems of strong spectral correlation and data redundancy in HRS, dimensionality reduction strategies are intorduced for the high-dimensional data. Considering with spectral characteristics of the HRS data, two different dimension reduction methods are presented. A novel feature extraction algorithm which is called subspace of the bands (SOB) feature extraction algorithm is proposed. The method divides the spectral space into several sub-band set according to the local relevance of spectral bands. Then the feature is extracted in the space of each subspace of bands. The hybrid feature is contructed by feature combination, and used in the classification procedure. The SOB feature extraction method can effectively reduce the impact of local spectral correlation in the classification. In order to reduce the impact of data transformation on the original spectral data, a novel band feature selected method is presented. The algorithm is based on combinatorial optimization algorithm. The method integrates the genetic operators into particle swarm optimization (PSO). Thus the proposed method utilizes the advantages of different evolutionary algorithms. It can overcome the shortcomings of most optimization algorithms, such as slow convergence, weak search capabilities of global and local space. The performance of band selection algorithm is increased.Second, most current classification algorithm only makes use of spectral information, ignoring the information of hyperspectral images in the shape, texture and other structural features. So novel hyperspectral image classification techniques are proposed based fusing spectral and spatial domain characteristics. The proposed methods use the spatial features to improve classifier, and enhance the classifier performance. In order to fuse the spectral and spatial features, two different data fuse strategies are introduced, which are combined features strategy and combined kernel functions strategy. First feature extraction method is executed in hyperspectral image. Then determine which principal components (PCs) should be retained by virtual dimension (VD) estimation methods. Then extract the morphological features of the image by executing mathematical morphology on the retained principal components images. So construct the expansion of morphological characteristics of the feature vector. The combined features vector is constructed by combining the morphological features with the spectral-domain characteristics. Then the combined vector is used in the classification operation, to enhance the performance of the classifier. So the effectiveness of feature fusion strategy is verified. The dimension of the hyperspectral is very high. In order to avoid the problem of "curse of dimensionality", this paper presents a novel classification algorithm by combination of kernel fuctions. It fuses the features by using kenel functions combination strategy. The input data is transformed into high-dimensional feature space through non-linear mapping transformation. So problem of the "curse of dimensionality" in the traditional classifiers is avoided. The efficiency of complementary diversity in the different characteristics is increased. The classification accuracy is enhancing.Then there are fewer labeled sample points in hyperspectral image classification compared with the mass unlabeled samples. Beceause the labeled samples is difficult to obtain and colected very costly. So the small amount of labeled samples can not fully represent the statistical distribution of the sample space. The conventional supervised classification algorithm does not have enough training samples. Thus the resulting classifier performance does not meet the actual application requirements. To solve this problem, this paper proposed the hyperspectral classification methods based on semi-supervised learning strategies. The proposed methods make use of the information contained in the mass unlabeled sample points to improve the classification accuracy. Based on the size of sample set and the actual application in hyperspectal classification, three different semi-supervised classification algorithms are presented. First, The spectral weighting transductive support vector machine (TSVM) semi-supervised classification method is proposed. Different spectral bands have different identifying capabilities. Through the analyzing the information for hyperspectral classification contained in different bands, different weight estimation strategies are designed. Respectively, the strategies are designed for two and multi-class classification. Using the estimated weight of the specral bands, the kernel fuction is modified. The non-convex objective function of the TSVM is opimized by Concave-Convex Procedure (CCCP). The classifier performance is enhanced. Second, based on clustering assumption that in the clustering the samples of same class tend to be divided in the same class, a novel semi-supervised classification algorithm is presented. The method is based on clustering kernel function. The bagging kernel function is constructed through the results of several times clustering. Then the hybrid kernel funcion is constructed by different computing strategies, which combines the bagging kernel funcion with the standard kernel function. The hybrid kernel function intorduces the information contained in the unlabeled samples into the classifier training process. So the classification accuracy is improved. Finally, graph-based semi-supervised classification algorithm is studied. The graph is constructed by the degree of similarity between sample points. The distance between two sample points is calculated, based on density-sensitive manifold distance instead of the Euclidean distance. The distance is used to measure the degree of similarity between sample points. The method also uses the priori information of original type to modify the classifier. The algorithm has good mathematic explanation and learning performance. It also can avoid convergence to local optima.Finally, a novel hyperspectal classification method is introduced based on support vercor machine and ensemble learning (EL). The proposed algorithm can deal with the problem of classifiers selection, especially when the prior knowledge is not enough. First the method use independent component analysis (ICA) method for hyperpectral image preprocessing. The resulting independent components are used as the features for the classication. Then severl classifiers are constructed by ensemble learning. Thereby the probability of selection improper classification, which is inroduced by the lack of prior information, is reduced. The method receives better generalization performance than single classifier. The proposed algorithm combines several different classification models into one model through changing the form of training sample set. The classification model's generalization performance is improved by the differences between multi based models.In the final, the paper summarizes the work, and points out the need for further research.
|
PDF Full Text Request