Two-stage cluster sampling reduces the cost of collecting accuracy assessment reference data by constraining sample elements to fall within a limited number of geographic domains (clusters). However, because classification error is typically positively spatially correlated, within-cluster correlation may reduce the precision of the accuracy estimates. The detailed population information to quantify a priori the effect of within-cluster correlation on precision is typically unavailable. Consequently, a convenient, practical approach to evaluate the likely performance of a two-stage cluster sample is needed. We describe such an a priori evaluation protocol focusing on the spatial distribution of the sample by land-cover class across different cluster sizes and costs of different sampling options, including options not imposing clustering. This protocol also assesses the two-stage design's adequacy for estimating the precision of accuracy estimates for rare land-cover classes. We illustrate the approach using two large-area, regional accuracy assessments from the National Land-Cover Data (NLCD), and describe how the a priorievaluation was used as a decision-making tool when implementing the NLCD design.
