Remote Sensing Image Fusion For Classification Based On Hybrid Intelligent System

With the development of aerospace and sensor technologies, the humans can get multi-platform, multi-types of remote sensing images. These images make us cognize one object from multiple characteristics including multi-scale effects, geometry, texture distribution, multi/hyper spectral, infrared radiation, backscatter and so on. This provides an excellent opportunity for making a decision more accuracy and reliability. However, because the classification process is affect by a number of factors, the accuracy and reliability of decision-making are reduced. These effections including as follows:1) the factors such as the sensor radiometric errors, unstable atmospheric conditions, the complexity of the surface environment and others, will lead to the roughness in image data itself and the fuzziness in classification information, which would seriously affect the accuracy of the decision;2) due to the influence of noise, the dependence between features will be changed, and classification accuracy subsequently will be changed too, for example, the bigger the correlation between the features and decisions is, the greater the impact of the noise on the classification accuracy;3) high-dimensional features from multiple sources remote sensing images may contain unuseful redundant features, which not only increases computing cost, but also affect the classification judgment;4) generative classification model and judgment based classification model do not take full advantage of the relationship between features, while in classification decisions, only sufficient the use of multi-source remote sensing images dependencies between features, a reasonable classification model to be able to effectively identify judgment. These factors bring about challenges for decision-making based on remote sensing image fusion.Most of existing approaches of processing multi-source remote sensing images stay at the pixel level fusion. However, mechanistic differences between the multi-source imaging result in a fundamental difficulty in pixel-level image fusion. In addition, the superimposed results of the pixel level fusion highlights some particular features of an object, which is conducive to human interpretation, but these are not usefull for the automatic classification, and difficult to improve the efficiency of automatic classification. While, the most of existing feature and decision level fusion algorithm are only starting from the application of certain characteristics, and can solve the problem in certain areas. Considering the complexity issues of multi-source remote sensing image fusion, including many types of uncertainty handling, effective feature selection, classification models, most of available methods are insufficient to these complex problems. From the above analysis, we can result that the existing methods are inadequate to solve the problem of decision-making based on remote sensing image fusion.In order to overcome the series of complex difficulties in the decision-making based on remote sensing image fusion, including image registration, property function estimation, feature selection and classification decisions under many uncertainties, this paper construct Gaussian fuzzy rough set and intuitionistic fuzzy rough set model based on fuzzy set and rough set theory, and combining particle swarm optimization, genetic algorithm, ant colony algorithm and Bayesian network model, to form the hybrid intelligent systems, which include the hybrid intelligent property function estimation algorithm, feature selection based on Gaussian fuzzy rough sets, feature selection based on intuitionistic fuzzy rough sets and hybrid intelligent decision making algorithm based on Bayesian network. Finally we can realize decision-making based on remote sensing image fusion by hybrid intelligent systems. The main work of this paper can be specifically summarized as follows:1) Construcing robust image registration algorithm for complex remote sensing images based on structure information. For solving the difficulties caused by repeated patterns and gray changes beween remote sensing images, this article studies the shape context feature and its improved method, combines it with point feature descriptor, and builds hybrid feature descriptor that can guarantee robust matched points in the initial matching. And, the paper studies graphics transform matching (GTM) algorithm which is based on the hypothesis that the neighborhoods of one point is invariance after transformation. GTM can remove outliers effectively. Experiments show that the registration process based on structure information can effectively improve registration accuracy under the complex conditions.2) Proposing the hybrid intelligent method for property function estimation. From the connections between fuzzy sets and Gaussian kernel function, this paper defines fuzzy probability, and establishes a symmetric uncertainty for fuzzy conditions, and combines the data-driven method and intelligent optimization approach to estimate the parameters. We give a hybrid intelligent kernel parameter estimation method, which can effectively calculate the optimized parameters and improve the classifier performance.3) Establishing the feature selection method based on Gaussian fuzzy rough sets. In order to solve the feature selection problem under many uncertainties such as the data roughness, classification fuzziness, redundancy features, noise features, firstly, this paper use Gaussian kernel function to fully exploit the fuzzy relationship among the data, and establish the Tcos fuzzy equivalence relations, then, build the Gaussian fuzzy rough set model; secondly, this paper fully considers the impact of redundancy in selected feature subset, and constructs feature evaluation criterion taking into account the redundancy among features and the relevancy between features and classes; finally, the feature subset is achieved under many uncertainties. Experimental results show that this method can select feature subset with a small number for specific classification tasks, and can maintain the classification accuracy.4) Establishing the feature selection approach based on intuitionistic fuzzy rough sets. In order to model many uncertainties more comprehensive, this paper extends the Gaussian fuzzy relationship to an intuitionistic fuzzy equivalence relationship with max-min transitive, and thus the fuzzy rough set is expanded to intuitionistic fuzzy rough sets. Taking full account of the class information, the decision dependent correlation (DDC) and the decision independence correlation (DIC) are used to propose the new feature evaluation criterion which is taking into account the redundancy and correlation at the same time. Finally, under the framework of hybrid intelligent optimization, the feature subsets for particular classification tasks are selected under many uncertainties.5) Construcing the hybrid intelligent decision approach based on Bayesian network. To solve these problems such as high-dimension features, data roughness, classification fuzziness, redundancy features, modeling relevant for classification, this article carefully analysis the approaches for learning Bayesian network classifiers including network structure learning and parameter learning, and combining the intuitionistic fuzzy rough sets and Bayesian network classifiers to construct the selective expansion of Bayesian network classifier. This model estimates the Bayesian network parameters by post probability through SVM methods which can be applied to continuous variables. Experiments show that this classification model is more efficient and higher classification accuracy.In summary, this paper firstly solves the preprocesses of image fusion such as the image registration under complex conditions and classification property function estimation; Secondly, in order to solve the feature selection problem under many uncertainties, this paper carefully discuss the uncertain modeling, feature evaluation criterion and intelligent optimization algorithms and other dependent aspects to establish the hybrid intelligent feature selection method; Finally, this paper use the selective expansion Bayesian network classifiers to model the classification problem, this model can not only improve the accuracy of the classification but also improve the computational efficiency. Research works in this paper cover the basic process of remote sensing image fusion, and can be used for analyzing the redundancy and correlation among the features scientifically, and can be used for selecting effective feature subset for specific classification task, and can be used for modeling the correlation among features. These works can improve the accuracy of the decision-making, and provide a reference for remote sensing image fusion.