Terrain Classification Of Remote Sensing Images Via Soft Computing And Mutual Information Theory

Remote sensing is characterized by covering a wide range area, possessing large amount of information, and obtaining the information quickly. The images obtained by remote sensing have been widely applied in many fields of the national security and the national economy. Terrain classification of remote sensing images is the key technology of remote sensing image interpretation. It is a current research hotspot attracting wide attention of researchers. In this thesis, the synthetic aperture radar(SAR) image terrain classification and hyperspectral remote sensing image terrain classification are studied. Considering the characteristics of these two types of images, the classification algorithms are proposed accordingly.As a representative of microware remote sensing, SAR can work under all-time and all-weather conditions. However, due to its coherent imaging system, SAR images have the speckle inevitably. The presence of the speckle may hinder the interesting details and targets. The intensive study has been implemented in the challenging problems, including serious speckle, poor edge location, and the lack of labeled samples involved in SAR image classification. In this thesis, robust SAR image classification algorithms are proposed.Recently, since the hyperspectral image of the high spectral resolution can accurately and detailedly distinguish different land covers, it has become another frontier filed of remote sensing images. However, a large number of spectral bands bring heavy computation burden for hyperspectral image terrain classification and produce “Hughes phenomenon”. The thesis focuses on the research of the supervised, unsupervised, and semi-supervised band selection of hyperspectral images by the soft computing and mutual information(MI) theory. The effectiveness of the proposed band selection methods are demonstrated experimentally and theoretically.The main contributions can be summarized as follows:(1) For the construction of visual words in bag-of-visual-words(BOV), the traditional clustering algorithms are sensitive to initialization and easy to trap in the local optima. Moreover, they are separate from the final classification result. To address these problems, a BOV based on clonal selection algorithm is proposed for SAR image feature extraction and classification. In the proposed algorithm, a new criterion based on k-fold cross validation is defined. It can obtain discriminative visual words by directly estimating the final classification performance. Additionally, clonal selection algorithm(CSA) is devised to optimize this criterion, which overcomes the shortcoming of the traditional clustering algorithms. Experiments demonstrate that the proposed algorithm obtains the satisfying terrain classification results and has the good robustness to the SAR speckle.(2) To overcome the sensitivity to the noises involved in fuzzy c-means(FCM), many improved FCM methods have been proposed. They are robust for additive noises, but have limited effort on tackling the multiplicative speckle in SAR images. Moreover, they over-smooth edge regions when they make the noises robust. To address the problem, a non-local FCM algorithm with improved edge preservation is proposed for unsupervised terrain classification of SAR images. In the proposed algorithm, a new non-local similarity based on ratio distance and a statistical-based edge preservation method are devised, which classify SAR image guaranteeing speckle insensitiveness and edge detail preservation simultaneously. Compared with representative unsupervised classification algorithms, experimental results on synthetic SAR images and real SAR images demonstrate that the proposed algorithm obtains the better classification results and shows better robustness to the speckle.(3) MI has been widely used as the measure criterion of hyperspectral band selection for its nonlinear and nonparametric characteristics. Traditional MI-based criteria commonly use bivariate MI to approximate the ideal MI-based criterion. However, these criteria may miss the bands having discriminative ability, and do not give the condition of the approximation. To address the problem, a new band selection method based on trivariate MI and CSA is proposed for hyperspectral image terrain classification. In the proposed algorithm, a novel trivariate MI based criterion is used to measure the redundancy for classification. From the MI theory perspective, the proposed trivariate MI based and traditional bivariate MI based criteria are proved as the low-order approximations of the ideal criterion. Compared with them, a more relaxed assumption condition is required for the trivariate MI based criterion. Additionally, to search an appropriate band subset, a new CSA is proposed. In CSA, adaptive clone and mutation operators are devised to speed up the convergence.Experimental results demonstrate the potential of the proposed algorithm for hyperspectral band selection.(4) How to select the spectral bands in hyperspectral images effectively without the guidance of class labels is a very challenging subject. To select the bands maximally preserving the information of original bands, a maximum joint MI criterion is defined. But it is hard to accurately calculate. Therefore, a new criterion based on high information and low redundancy is proposed as its low-order approximation. Moreover, from MI theory perspective, a theoretical proof is given that many existing unsupervised feature selection criteria are also its low-order approximations. Compared with them, the proposed criterion has more extensive application. Experimental results on hyperspectral terrain classification demonstrate the effectiveness of the proposed algorithm.(5) In hyperspectral images, how to improve the performance of band selection by using both limited labeled samples and a large number of unlabeled samples is a worth researching topic. In this thesis, two kinds of semi-supervised band selection algorithm are proposed. Inspired from the former supervised and unsupervised algorithms, a semi-supervised criterion based on high discrimination, high information, and low redundancy is defined. Compared with existing semi-supervised feature selection methods, experiments demonstrate the effectiveness of the proposed algorithm. To overcome the sensitivity of trivariate MI to the discriminative abilities of bands and high time complexity caused by search strategy, another semi-supervised algorithm based on normalized trivariate MI and affinity propagation is proposed. In the proposed algorithm, a normalized trivariate MI is defined. Its upper and lower bounds are derived. According to its different ranges, it can consider not only band redundancy, but also band synergy. Moreover, a new statistical-based method is devised by using the continuity property of bands. It can remove noise bands automatically to avoid the disturbance of noise bands for the clustering-based algorithms. Compared with existing semi-supervised feature selection methods, experiments on hyperspectral terrain classification demonstrate the effectiveness and efficiency of the proposed algorithm.