| Popis: |
The objectives of this study were to experiment with and extend current methods of Synthetic Aperture Rader (SAR) image classification, and to design and implement a prototype intelligent remote sensing image processing and classification system for land use and land cover mapping in wet season conditions in Bangladesh, which incorporates SAR images and other geodata. To meet these objectives, the problem of classifying the spaceborne SAR images, and integrating Geographic Information System (GIS) data and ground truth data was studied first. In this phase of the study, an extension to traditional techniques was made by applying a Self-Organizing feature Map (SOM) to include GIS data with the remote sensing data during image segmentation. The experimental results were compared with those of traditional statistical classifiers, such as Maximum Likelihood, Mahalanobis Distance, and Minimum Distance classifiers. The performances of the classifiers were evaluated in terms of the classification accuracy with respect to the collected real-time ground truth data. The SOM neural network provided the highest overall accuracy when a GIS layer of land type classification (with respect to the period of inundation by regular flooding) was used in the network. Using this method, the overall accuracy was around 15% higher than the previously mentioned traditional classifiers. It also achieved higher accuracies for more classes in comparison to the other classifiers. However, it was also observed that different classifiers produced better accuracy for different classes. Therefore, the investigation was extended to consider Multiple Classifier Combination (MCC) techniques, which is a recently emerging research area in pattern recognition. The study has tested some of these techniques to improve the classification accuracy by harnessing the goodness of the constituent classifiers. A Rule-based Contention Resolution method of combination was developed, which exhibited an improvement in the overall accuracy of about 2% in comparison to its best constituent (SOM) classifier. The next phase of the study involved the design of an architecture for an intelligent image processing and classification system (named ISRIPaC) that could integrate the extended methodologies mentioned above. Finally, the architecture was implemented in a prototype and its viability was evaluated using a set of real data. The originality of the ISRIPaC architecture lies in the realisation of the concept of a complete system that can intelligently cover all the steps of image processing classification and utilise standardised metadata in addition to a knowledge base in determining the appropriate methods and course of action for the given task. The implemented prototype of the ISRIPaC architecture is a federated system that integrates the CLIPS expert system shell, the IDRISI Kilimanjaro image processing and GIS software, and the domain experts' knowledge via a control agent written in Visual C++. It starts with data assessment and pre-processing and ends up with image classification and accuracy assessment. The system is designed to run automatically, where the user merely provides the initial information regarding the intended task and the source of available data. The system itself acquires necessary information about the data from metadata files in order to make decisions and perform tasks. The test and evaluation of the prototype demonstrates the viability of the proposed architecture and the possibility of extending the system to perform other image processing tasks and to use different sources of data. The system design presented in this study thus suggests some directions for the development of the next generation of remote sensing image processing and classification systems. |
