Purposive versus random sampling for map validation: a case study on ecotope maps of floodplains in the Netherlands.

ABSTRACT The quality of ecotope maps of five districts of main water courses in the Netherlands was assessed on the basis of independent validation samples of field observations. The overall proportion of area correctly classified, and user's and producer's accuracy for each map unit were estimated. In four districts, the validation samples were selected purposively in the past. For the fifth district, a stratified two-stage probability sample was designed, such that the spatial pattern resembles that of the purposive samples. For the purposively sampled districts, measures of map quality were estimated following a model-based approach, using a model for the spatial variation of classification errors. For the randomly sampled district, measures of map quality were obtained following a design-based approach, using the inclusion probabilities of the sampling points. The overall proportions of area correctly classified vary from 61% to 76%. Both user's and producer's accuracies show large variation among the map units. Model-based predictions of producer's accuracies from two models differed strongly, indicating that with the model-based approach the validation results strongly depend on model assumptions. We demonstrate that stratified two-stage random sampling answers to the same practical and budgetary constraints as purposive sampling. Stratified two-stage sampling combined with a design-based estimation method results in model-free estimates of the overall proportion of area correctly classified, user's and producer's accuracies. This means that the inferential validity of the estimated overall proportion of area correctly classified and user's and producer's accuracies is rigorously established if a design-based approach is followed. Copyright © 2012 John Wiley & Sons, Ltd. [ABSTRACT FROM AUTHOR]