Pattern Classification Research Of Remote Sensing Image Based On Artificial Neural Network

Remote sensing (RS) is a technology by adjudging, measuring and analyzing character of targets at long bowls. It realizes the collection, process, recognition and classification of image information on the earth. RS technology has good advantages on dynamic cycle, data, abundant information and easily acquisition as a tool in age of information, so it is the most effective technology and means of obtaining space-time information. RS image is mostly used to map relief maps, make orthograph and thematic maps by professional interpretation, which can be saved in the databases of geographical information system (GIS) or updated the GIS databases.Pattern classification is a key technique in remotely sensed image processing. Although the research history of pattern classification techniques is quite long, users require for more accurate classification result and smaller computing load now. So there is an urgent need for modern pattern classification methods to solve the modern remote sensing applications.In recent years, with the development of the theory about artificial neural network system, the neural network technology is becoming increasingly an effective means of classification processing of remote sensing images. Compared with classification of the traditional Bayesian statistics, the results show it has not only the highest accuracy but also the fastest speed of classification.Based on the conclusion of existing research fruit, the thesis discusses some artificial neural network methods, such as BP, SOFM, LVQ neural network, which is applied to RS image classification. In this thesis, the following aspects have been researched:At first, this thesis reviews some principled problems about the practical application of the methods to remote sensing data classification. BP neural network is widely used for classification of remote sensing image data nowadays. And then the thesis practices supervised classification with BP algorithm on the base of the clustering image supported by ERDAS software.After analyzing self-organizing feature map algorithm and learning vector quantization algorithm developed by Kohonen, this thesis proposes a hybrid learning vector quantization algorithm combining the modified SOFM algorithm with the LVQ2 algorithm, then establishes a general HLVQ-based classification model for remote sensing image .Compared with the conventionally statistical method and LVQ2 classifier, the HLVQ classifier has more advantages on recognition rate and general classification performance.Lastly after comparing learning vector quantization algorithm with generalized learning vector quantization algorithm, this thesis establishes a general GLVQ-based classification model for remote sensing image. With the experimental applications of land-over classification by the presented model, the GLVQ classifier not only significantly increases the convergence rate of processing, but also the highest accuracy of classification.