Assessment Method Of Remote Sensing Classification Based On Sub-Pixel Unmixing And Uncertainty Analysis

With the development of these techniques in earth space observation, theinformation processing of remote sensing has transformed from a quantitativeanalysis into a qualitative analysis, and the processing system of remote sensing hastranslated from functionality into high quality. The classification of land-cover playsan important role in the processing and application of remote sensing. Under theuniform criterion, the issue of verifying the classification algorithm for land-coverand assessing the accuracy of classification has become one of the fundamentalproblems in the field of remote sensing.In the thesis, due to the objective of "Constituting the Platform of StandardRemote Sensing Image Data" as one of the national "863" projects, we will discusson the constitution of the platform for the land cover classification which is used toverify the validity of the latest algorithms and assess the accuracy of classification,the systemic research on some key approaches including the pixel-level classificationassessment, the algorithm of sub-pixel unmixing, and the subpixel-level classificationassessment especially for several representative classification processing methods forland cover. Through the research on these fields, the characteristic and innovativeachievements in scientific research have obtained as follows:First, we constitute a platform for the land cover classification applied in thevalid verification of algorithms and the accuracy assessment for classificationalgorithms using types of remote sensing data from satellites, which provides not onlythe uniform test data for the land cover classification algorithms, but also thecorresponding ground truth data for the accuracy assessment of classification.Second, based on the complete research on the uncertainty which exists in theprocessing of the land cover classification, the thesis introduces a pixel-levelassessment approach for classification algorithms based on the uncertainty analysis.Through the quantitative evaluation for some typical classification algorithms in threepopular remote sensing processing softwares, our approach has proved its validity.Third, considering the mixed pixels existing in remote sensing data, the thesisproposes the concept of End-Pixel and the algorithm of extracting End-Pixel, which isused to unmix the image. In addition, we develop a novel subpixel-level assessmentapproach for the land cover classification, which is based on the theory of sub-pixelunmixing and the research of fuzzy confusion matrix. This method extends theconcept of pure pixel in hyperspectral data to a variety of remote sensing data,including not only hyper/multispectral data but also panchromatic data. As aquantitative assessment approach, it can achieve more objective and accurate resultsthan the pixel-level assessment approach.Finally, the thesis designs and implements a software assessing system whichconsists of these functions such as the assessment for the preprocessing, thepixel-level assessment for the land cover classification, and the subpixel-levelassessment for the land cover classification. Consequently, an assessment system forremote sensing processing techniques under the uniform criteria has been developed.