A comparative study of multiple classifier combination methods in remote sensing

Due to the growing volume of images originated from the variety of sensors available, new data processing techniques are required to allow the information to be processed promptly and accurately. Although the range of image processing techniques has been greatly expanded, fr om classical statistical approaches to neural network methods, there is no single classification algorithm capable of deriving generic products from remotely sensed data. The performance of these algorithms is strongly dependent upon data selection and on the efforts devoted to the design phase. In this paper, Mr propose a more systematic investigation into the problem of combining multiple classifiers in the context of remote sensing. Four methods of combining the outputs of multiple classifiers ar-e used. These are: voting principles, Bayesian formalism?, evidential reasoning and neural networks. Preliminary results indicate that improvements can be obtained in difficult pattern recognition problems by combining or integrating the outputs of complementary multiple classifiers. The concept of combining multiple classifiers is suggested as a new direction for the development of reliable recognition systems. This approach potentially has a wide variety, of applications and can successfully promote the transition from? data to information.