Study On The Application Of The Agent-Based Approach For Remote Sensing Image Classification

Image classification is an important method in remote sensing information extraction. Post-processing of image classification is a necessary approach to improve the image quality of initial classification result. After reviewing and analyzing the existing remote sensing image classification researches and considering the demand of actual production, the author found the following deficiencies of current researches: 1. Majority of the researches emphasized on the improvement and application of classification method, but rarely involved in the post-processing method of image classification; 2. Some of the proposed post-processing methods were only for the image itself, while the information contained in the images were kept underutilized. 3. Manual post-processing of classification depended on experiences heavily and cost too much work. Though automatic tools offered by current commercial software improved the image quality of initial classification to a certain extent, there is still excessive clustering phenomenon.Computational Agent is a kind of abstract model in dynamic environment, which has a high level of intelligence in digital image processing. This paper attempts to introduce the Agent theory and model into remote sensing image classification. After studying the concepts and theories of Agent, the author constructed the Agent and multi-Agent system for the post-processing of image classification. According to the characteristics of initial classified images and the abundant enhancement information of objects contained in the remote sensing images, the author constructed a multi-Agent system which consists of three types of cooperating Agents, namely classification Agent, decision making Agent and integrated smart Agent. The Agents could improve deficiencies in initial classification image automatically by sensing, reasoning and utilizing information in images and environments. Then the author developed the core module in post-processing tools of Agent classification by IDL programming language to realize the post-processing classification workflow which based on Agent system.The Paper used the maximum likelihood method, neural network method and spectral angle method to implement supervised classification on the preprocessed Landsat 8 OLI image of Beijing. Meanwhile, nine features including NDVI, brightness and greenness were extracted from the image. Auto-adjustment on the initial classified images were implemented through post-processing tools of Agent classification. The validity of the method were verified by statistical accuracy and visual interpretation methods, and also compared with the classification tools of ENVI software.Conclusions: 1. The Agent-based mode proposed in this research for post-processing of image classification can process remote sensing images automatically and suppress the salt-and-pepper noise effectively. The overall classification accuracy can be improved by 5.5% at most. 2. The Agent-based post-processing tools combined the information of initial classification images and remote sensing images, thus can avoid excessive cluster when based on filtering process. Meanwhile, the method is still available when supporting information missing. 3. When the accuracy of initial image classification is relatively low, the Agent-based post-processing method can improve the classification accuracy much more. 4. The core function module developed by IDL programing language can be integrated with ENVI or further developed to be a stand-alone system easily, which can be used widely.