Implications of land-cover misclassification for parameter estimates in global land-surface models: An example from the simple biosphere model (SiB2)

One of the primary applications of the global 1-km land-cover DISCover product is to derive biophysical and ecological parameters for a range of land-surface models, including biosphere-atmosphere, biogeochemical, and ecological models. The validation effort reported in this special issue enables a realistic assessment of the implications of misclassification errors for parameter estimates within the models. In most land-surface models, cover types are aggregated to coarser groupings than the 17 IGBP classes for estimating parameters, with aggregation schemes varying with individual models and individual parameters within each model. Misclassification errors are consequential only when they occur between cover types that are not aggregated by the model. We use examples of two biophysical parameters-leaf area index and surface roughness-as estimated for use in the Simple Biosphere Model (SiB2) and other modeling applications to quantify the effects of misclassification on parameter estimates. SiB2 relies on satellite data as well as land-cover information for estimating the biophysical parameters. Consequences of misclassification are likely to be greater for those models that do not use satellite data. Mean class accuracy based on those sites for which a majority of interpreters agreed (percentage of validation pixels classified correctly out of total number of validation pixels, averaged over all classes), adjusted by area of each cover type in the IGBP DISCover product, is 78.6 when all misclassification errors are included. By excluding misclassification errors when they are inconsequential for leaf area index and surface roughness length estimates, mean class accuracies are 90.2 and 87.8. respectively. The results illustrate that misclassification errors are most meaningfully viewed in the context of the application of the land-cover information.