Calibration of satellite classifications of land area

In forestry applications, the classical and the inverse calibration methods are commonly used to correct for misclassification bias in satellite-based areal cover type estimates. Both methods, as applied in forestry, require that the reference data used for calibration be independent of the data used to train the classifier. Further, variance estimates for the calibration methods are valid only if the reference data arise from a simple random sample. A comprehensive simulation study was conducted comparing these two methods for a broad range of cover type distributions, classification accuracies, and reference sample sizes. The estimates of the two methods were compared on the basis of bias and accuracy. Variance estimates for the two methods were also studied. The inverse method consistently outperformed the classical method. Results of the simulation study and possible implications are discussed.