Multi-scale Assessment For Uncertainty Of Classification Of Remote Sensing Image Based On Information Theory And Rough Set

Thematic category information obtained by remote sensing image classification has beenwidely used in various fields. To evaluate the thematic category information can be used in thefollow-up study, and to analyze the impact of thematic category information in subsequent use,reliability is provided with thematic category information at the same time. Therefore, how tocomprehensively, accurately evaluate the classification uncertainty is the focus of this paper.In order to comprehensively, accurately measure the attribute uncertainty of classificationof the remote sensing image, this article establishes a uncertainty assessment system forremote sensing image classification in three scale, pixel-landcover class-whole image. Thetest used IKONOS multispectral images in Shihezi of Xinjiang as the data source. Then, toclassify the data using minimum distance classifier and support vector machine classifier.Finally, using the above assessment system evaluate the classification results.At the pixel scale, the author studies uncertainty evaluation theory for the classification ofremote sensing image based on the information theory. Probability vector obtained fromclassification as the breakthrough point, using probability entropy model assesses theuncertainty of classification at pixel scale. Not only won the classification uncertaintyinformation of each pixel, but also visually display the location information of classificationuncertainty.At the scale of landcover class, the author studies uncertainty evaluation method for theclassification of remote sensing image based on the rough set. After analyzing the problems oforiginal method, the author modifies the original model and proposes the boundary regionbased modified rough entropy model. First of all, the author from the theory proved that theimproved model is more objective in measuring the uncertainty caused by classificationknowledge. Then, using the modified rough entropy model and the boundary region based modified rough entropy calculate the classification uncertainty of each landcover class.At the whole image scale, the author studies uncertainty evaluation method for theclassification of remote sensing image based on the rough set. The lower and upperapproximation sets get from the probability vector as the breakthrough point, using theapproximation classification accuracy and the approximation classification quality assess theuncertainty of classification at the whole image scale. The classification accuracy of two kindsof classifiers is compared.Conclusions are obtained from the experimental results:First, at the pixel scale, classification uncertainty larger areas are mainly distributed inboundary region between the different categories. Comparison of the two classificationalgorithms can be seen that boundary ranges of the minimum distance classification issignificantly larger than the SVM classification.Second, whether using the minimum distance or support vector machine classifier, forany classes, uncertainty evaluation results in the use of boundary region based modified roughentropy is less than that in the use of modified rough entropy model. So in view of theinfluence of the boundary region, boundary region based modified rough entropy proposed byauthor can objectively describe the uncertainty caused by classification knowledge.Third, the evaluation indicator which is chosen by the author is full use of all the pixels inthe image, not relying on the pure test sample. Evaluation method based on rough set theorymore fully considers the influence of boundary pixels on the classification results. Therefore,assessment results are closer to the true situation.Based on the probability vectors obtained in the processing of classification, usinginformation theory and rough set theory, the author studies multi-scale assessment foruncertainty of classification of remote sensing image. Thus, with different scales, differentangles, we could obtain all kinds of uncertainty information of image, and solve the problemof incomplete expression of uncertain information, when only using a single scale to evaluatethe classification uncertainty.