Analyzing the uncertainties of ground validation for remote sensing land cover mapping in the era of big geographic data

Abstract

Ground validation and accuracy assessment on remote sensing land cover classification is a vital process before the product may be used by the end-user. The common approach towards the validation is based on field investigation and manual or automatic image interpretation using the original or higher-resolution images. The ground reference, which is often regarded as “ground truth”, however, contains errors, especially when a large amount of such ground references is expected with the speculation of coming era of big geographic data. In this study, we aim to analyze the uncertainties in the process of ground validation. By taking accuracy assessment of land cover mapping in Central Asia as an example, a two-tier sampling scheme with a collection of more than 27 thousand samples was adopted. Ground references were sampled by manual image interpretation as well as by field investigation. The reference data were then cross-validated. Misclassification and scale issues are highlighted in the analysis. Result indicates that misclassification of ground reference data by image interpretation is common and the errors in the reference data would make misleading accuracy assessment on remote sensing classification. A new evaluation system of data quality is therefore required.