Map-Adjusted Area Estimation Equals Calibrated Expected Area

Overview. Two workflows estimate cropland area from the same ingredients: a model score for every pixel and a small set of labeled reference points. The remote-sensing workflow stratifies on the map, samples, and reports the area-adjusted (error-matrix) estimate. The machine-learning workflow calibrates the model on the same labels and sums calibrated probabilities over the map. When calibration is histogram binning with one bin per sampling stratum, the two workflows return the same number for every possible sample. This demo steps through both on a synthetic landscape. Navigate with the buttons, the dots, or ←/→ (or a/d).

References: Olofsson, P., Foody, G. M., Herold, M., Stehman, S. V., Woodcock, C. E., & Wulder, M. A. (2014). Good practices for estimating area and assessing accuracy of land change. Remote Sensing of Environment, 148, 42–57. · Zadrozny, B., & Elkan, C. (2001). Obtaining calibrated probability estimates from decision trees and naive Bayesian classifiers. Proceedings of ICML.