Remote Sensing Classification By Adaptive Combing Multiple Classifiers

As a new technology extracting geographic information, Remote Sensing owns many features such as:multi-temporal, multi-band, large-scale. And the data sources of Remote Sensing are acquired repetitive and cyclical, so it is conducive to the dynamic monitoring of land use/cover. The remote sensing has become a main method and technique in land use/cover study. Combining different classification algorithms existing, making use of the advantages of these different classification algorithms, achieving the complementarities between these classification algorithms, and reaching the purpose of improving classification accuracy are important directions in land use/cover research.The content of this paper are mainly two aspects:The Error Analysis and Verification of Combined Classifier; The Construction and Validation of Adaptive Combination Classifiers. These researches are as follows:(1) The Error Analysis and Verification of Combined ClassifierWith further research, the development of combined classifiers was also faced with two problems. On the one hand, compared with the single classifier, combined classifiers can improve the effect of classification and classification accuracy; However, the improving effect is not significantly and the enhance rate of the classification accuracy is not high. On the other hand, Combined Classifiers not only does not improve the classification accuracy and sometimes even reduce the classification accuracy in some case. The classification accuracy value of combined classifiers is higher than the lowest accuracy in single classifiers’, but lower than the highest one. It lies between the lowest and highest accuracy, and is more close to the value of higher accuracy.For these two cases, this article analyses the error of single classifiers in Combined Classifiers, and designs appropriate example which is used to test and verify the results of error Analysis. The results indicated that the precision of combined classifier is related to the position of incorrectly classified pixels in each single. While the incorrectly classified pixels by single classifiers are separated, the accuracy of combined classifier is the highest; While the incorrectly classified pixels by single classifiers are intersected in classified results, the precision of combined classifier is higher than that of each single classifier and the improvement is inversely proportional to the size of error set in combined classified results; While the incorrectly classified pixels by one classifier are included in another classifier, the accuracy of combined classifier is located in between the high and low accuracy, which is near to the higher one.(2) The Construction and Validation of Adaptive Combination ClassifiersThis article introduced the concepts about adaptability of single classifier for every pixel in image and proposed adaptive combination classifiers. Different pixels depend on different single classifiers in classification process, so the incorrectly classified pixels by single classifiers are separated the accuracy of combination classifiers will be the higher than single classifiers’. Greater improvement in the classification results of the combined classifiers can be achieved, and adaptive combination classifiers can improve the classification accuracy of the combination classifier.By designing a case related, this article test and validate the effect of adaptive combination of classifiers. Results show that:adaptive combination of classifiers increase accuracy in remote sensing image classification and achieve the purpose to improve the classification accuracy. When changed training sample sets and the classification algorithms together, the accuracy of combination classifiers is higher. Compared with the method of changing classification algorithm only to produce a single classifier, the adaptive combination classifiers own diversity more and perform better in classification process.