Development of an error matrix algorithm (EMA) for calibrating land cover data

Yue Wu, Purdue University

Abstract

Reference data and error matrix are commonly used to calculate a variety of accuracy statistics for thematic maps. In this thesis, we presented an algorithm called EMA (Error Matrix Algorithm) to increase the accuracy of land cover data based on the error matrix. The method is based on an assumption that classification errors often occur on shared edges of land cover types. The goal is to correct that type of misclassifications according to the statistics of the error matrix and neighborhood relationships. Theoretically, EMA has more potential to increase accuracy comparing with some commonly used filter, such as the majority function in ArcGIS. We obtained an optimal threshold of 1/8 for EMA by carrying out experiments in Lafayette area of Indiana assuming a perfectly classified map. The preliminary experiment in Lafayette area supported that EMA can be used to increase the accuracy of land cover data by using reference data that covers a sub-area. The overall accuracies of both NLCD 2001 and GAP in Kansas were increased significantly by applying EMA either on subareas or the whole state. The Kappa statistic also shows that both data were improved, with NLCD 2001 more significantly. Meanwhile, the consistency between the two data was increased by carrying out Z-test of Kappa. The more convincing evaluation of EMA is to randomly select a part of reference data to run EMA and use the rest to assess it. Statistical data shows that EMA works well for NLCD 2001 but not so well for GAP. In general, EMA is helpful to increase classification accuracy of land cover data, especially for those with relatively low accuracy.

Degree

M.S.

Advisors

Shao, Purdue University.

Subject Area

Land Use Planning|Remote sensing

COinS