The Application Of Rough Neural Network Based On Fuzzy Rule In Remote Sensing Image Classification

Rough sets which is a mathematic tool for processing uncertain and incomplete information, has become an important branch of uncertain computing , as well as the neural network which shows powerful non-linear mapping, self-learning, adaptive, associative memory and fault-tolerant ability. Due to the complementarity between rough sets and the neural network in anti-noise capability and generalization ability, etc., rough neural network, which integrates both these technologies, solves some difficult problems of traditional integration system in a certain extent. Because rough neural network accords with the character of human's intellection, it has become the main technology of present integration system. And it has been applied in many fields successfully.Because the spectral information of remote sensing image is abundant and the bands usually correlate with each other, moreever, both fuzzy uncertain information and rough uncertain information are always in existence, therefore, classifying remote sensing image by traditional neural network classifier result in complex structure, slow training rate, poor generalization, low classification accuracy. This paper utilizes the fuzzy clustering to partition the univers, and rough set analysis theory is applied to extract the fuzzy rules. Thus the size of neural network structure is greatly reduced and generalization ability is improved.The defects of rough sets theory in processing continuous attributes are analysed. A method to partition the universe of discourse based on fuzzy clustering is proposed to solve the partition problem in the process of designing rough neural network. Considering traditional clustering algorithm has the problem of easily fall into local optimum, a modified PSO algorithm with crossover and mutation operators is combined with FCM algorithm. Thus a new fuzzy clustering algorithm (CMPSO-FCM) is proposed. The searching capability and clustering effectiveness are improved by this new algorithm.Two popular algorithms of attribute reduction, which are based on attribute dependencies and attribute entropy respectively, are discussed according to the center and fuzzy membership matrix which can be achieved by fuzzy clustering algorithm CMPSO-FCM. And the main steps to extrac fuzzy decision rules are also introduced. Finally, a rough neural network can be built under these decision rules. The rough neural network model presented above is applied in pattern classification and real remote sensing image classification. Experiments results of UCI data sets and remote sensing image classification in the area of Washington show that, compared with traditional rough neural network, this method has superiorities at the aspect of structure, classification precision and generalization. It has good application value and prospect.At last, the main research contents and results are summarized, and the problems, which should be investigated further, are discussed.