A Terrain Classification Study On Island Remote Sensing Image Based On Artificial Neural Network

The theme of this study comes from national island monitoring system program. As the gradually enhancing of the monitoring system, we have accumulated large amount data of island remote sensing image. In order to excavate the island information which contained in the remote sensing image, we need to make analyses and categorize these images and then promote the level of quantitative analysis on monitoring. Island remote sensing image terrain classification technology is the main method to obtain the nature of island in the daily island work. Yet, with the development of remote sensing and computer techniques and considering the features of remote sensing images: multisource, huge quantity, scale limitation, ambiguity, high space dimension, multi-time phase, etc, the remote sensing image terrain classification skill which based on the traditional Bayesian classifiers haven't already satisfy the demand of interpreting work on island sensing images in certain degree. The aim of this research is applying the artificial neural network to national monitoring system program reference to preceding study results on remote sensing image terrain classification and hoping to find a practical and efficient island remote sensing image interpreting method during work.Artificial neural network is a mathematic model which imitates the memory of human brain. It has the advantages of strong learning capability, good fault tolerance, without presume rate model. What's more, it can adapt to remote sensing images of high ambiguity wonderfully, combine with light spectrum, vein, shape and structure of images easily, recognize and extract the terrain information of images precisely and rapidly and solve the common problems of “homogenizes and different spectrum” and “same spectrum and heterogeneity” in terrain classifications work better and promote the precision of island remote sensing image terrain classifications and work efficiency of island work stuff.In this paper, we used the aerial remote sensing image of Dongyu Island, Hainan province in 2008, to build the "3-8-5" three-layer BP neural network classification model by MATLAB software and select training samples by ENVI software. The model was applied to obtain the classification of the Dongyu Island remotely sensed image. The results showed that this BP neural network classification method provides a more effective island remotely sensed image classification technology and improves the accuracy of classification. The overall accuracy of the BP classification was 87.85%, and the Kappa coefficient was 0.84,i.e., the total classification accuracy increased by 5.53 percentage and the Kappa coefficient increased by 0.07 compared with the Minimum Distance method.BP artificial neural network can obtain terrain information automatically and intelligently. The classification result is much more approach to the visual interpretation. If the national island monitoring system program adopted this method to interpret the remote sensing images, it would further enhance the dynamic island monitoring ability and then promote the orderly processing of island protection, development activities.