Error reporting for continuous variables is often published at the full scale of the model or analysis performed. This can sometimes encompass a very large and variable region. These products are often used in research whose study area intersects with part of the modeled area. If the subset of the model used does not represent the full variability of model errors demonstrated in the original product, the error information will not be as useful.
Using domain estimation we demonstrate a method to report model errors for continuous remote sensing variables using smaller subsets of the data, or domains. These can be established as graduated intervals of the variable itself. For example, below we show the number of samples used for validation in 10% Tree Canopy Cover (TCC) domains in a confusion matrix.
The domains don't have to be related to the variable in question, as demonstrated in the table below where we depict error metrics for the same TCC model using National Land Cover Database (NLCD)-derived landcover domains.
