Method for land cover classification accuracy assessment considering edges

Abstract

Land cover, as the direct description of the Earth surface, has close relationships to the circle of global substances and energy, climate change, and economic activities of human society. The acquisition of land cover products is usually based on image classification. The accuracy of image classification is highly important to the monitoring and investigation of global environment as well as in decision-making support. Thus, classification accuracy assessment of land cover products is an important procedure. Given that many mixed pixels are located in between different classes, the accuracy of edge pixels tend to be lower than that of interior pixels. This scenario leads to the increment of heterogeneity to the classification accuracy of each class and to the increase of uncertainty of accuracy assessment. This study presents a method named stratified sampling considering edges (SSCE) based on traditional stratified sampling (SS) to optimize the process of classification accuracy assessment. Theoretical derivations and experimental results indicate that SSCE has high stability and accuracy on the estimation of overall accuracy and kappa. SSCE has high accuracy on estimating classification accuracy when only a few sampling points exist. SSCE needs less sampling points than SS under the same tolerance of error. The higher the difference on accuracy is and the more equal the areas between edge regions and interior regions are, the more accurate SSCE is on accuracy assessment. In a word, SSCE costs minimal sampling points. It also has high accuracy and stability on the assessment of land cover classification accuracy.