Practical global sampling methods for estimating area and map accuracy of land cover and change

Recent advancements in data storage and computing, particularly cloud-based processing, enable mapping global land cover and change relatively quickly and easily. Multiple versions of a map could be produced within a matter of days with various adjustments of selected parameters of machine learning models. Sample-based validation is then required to establish correspondence between these map prototypes and the real world, thus turning them from algorithm data outputs into sources of information with quantified errors. Implementing global probability sampling of geographic data for the purposes of area estimation and map accuracy assessment presents multiple challenges, primarily linked to the way these geographic data are stored (coordinate systems and projections) and the objectives of the specific project. Here we summarize various approaches to global sampling aimed at assessing accuracy of global land cover and change maps and producing unbiased estimators of area along with the standard errors associated with these estimates for the target land cover classes. We provide a unified set of estimators that accommodate a variety of sampling designs by explicitly accounting for the area of each sample unit, as well as code and technical details necessary to implement the presented methods. While we do not compare relative precision of the presented sampling design options, our aim is to help practitioners select an appropriate sampling design and estimators for their specific data format and project objectives, and to facilitate the correct implementation and increased reproducibility of global sampling methods within the land cover mapping community.