Remote Sensing Classification By Combing Multiple Classifiers

Concerning the practical needs of obtaining land use/cover information rapidly, automatically and accurately, automatic classification of remote sensing images has become an important research direction to extend the applicability of remote sensing data. It provides a new way for more accurate classification results that fully exploiting the potentials and using the diversity of the existing classifiers by combining multiple existing classifiers, while most efforts were made to develop new classifiers. Therefore, it is theoretically and practically significant to study the method of combining multiple classifiers and explore its application in automatic classification of remote sensing images.In this paper, multi-classifier combination method and optimization of combined classifier based on diversity measure of classifiers is studied and is applied to land use/cover classification of remote sensing images. The main conclusions are as follows.(1) Remote sensing classification by combing multiple classifiersIn order to generate the weights in combining classifiers by majority the method is proposed that makes use of user accuracy of in sub-classifiers classification as the voting weights to improve classification accuracy. In this paper, the performances of sub-classifiers are validated by classification of remote sensing images firstly, and the user accuracy of land use/cover classes are calculated in this process. Then, the outputs of sub-classifiers classification are combined by voting method that uses user accuracy as weighted values of vote. In the result, class which has maximum vote is given to the pixel undefined.The results show that the overall accuracy of combined classifiers are higher than those of sub-classifiers, when any two of the maximum likelihood, neural network and support vector machine classifier are combined in land use/cover classification. However, the overall accuracy of combined classifier which combines all of the three sub-classifiers is higher than the lowest sub-classification accuracy, but not necessarily higher than the most accurate sub-classification. The method of combining multiple classifiers developed in this paper has the same performance in classification of remote sensing images which have different landscape characteristics.It depends on the stability of sub-classification that whether the results of combined classification are affected by the distribution of training samples. If the distribution of training samples only had little influence on the sub-classification, the combined classifiers would have stable performances. In this paper, the classification by combining maximum likelihood and support vector machine, which is little affected by the distribution of training samples, has a better result than sub-classification.(2) The optimization of combined classifier based on diversity measure of classifiersFor the optimization of combined classifier, Entropy-base Pair-wise Diversity measure (EPD) is introduced to measure the diversity of sub-classifiers in combining classifiers and its relationship with the combination accuracy is analyzed later. The availability of EPD in optimizing combined classifier is discussed finally.The results show that the EPDs of sub-classifiers are positively correlated with the dominance of combination classification, which is defined the difference between Kappa of combined classifier and the largest Kappa of sub-classifiers. Therefore, EPD can predict combination performance of sub-classifiers to a certain extent. However, the correlation between EPDs and combination classification results are not strong enough, so the diversity measure methods in classifier combinations need further study.