Monte Carlo uncertainty analysis from top-of-atmosphere reflectance to plant functional type distributions

Uncertainty in the trends and variations of climate variables in climate data records is as important to understand as climate trends and variations themselves. Metrology provides the framework for assessing and budgeting uncertainty but the application of metrology to climate data records derived from Earth Observation is a scientific and technical challenge and a matter of research. We applied Monte Carlo methodology to demonstrate the end-to-end uncertainty budget for the quantitative variables (seasonal land surface spectral reflectance and plant functional type fractional coverage) of the Land Cover project within ESA’s Climate Change Initiative on the example of one year of Sentinel-3 OLCI remote sensing data and study cases in Africa, Europe, and South America. The budget considers most important sources of errors and takes account of uncorrelated and fully correlated random error structures. The interquartile range of relative standard uncertainty per datum of yearly land surface spectral reflectance is 0.050–0.108 at 490 nm, 0.015–0.046 at 560 nm, 0.007–0.062 at 665 nm, and 0.008–0.024 at 885 nm. The uncorrelated random component of seasonal land surface reflectance uncertainty diminishes with the duration of the season. Spectrally anti-correlated errors in seasonal land surface reflectance composites were attributed to a maximum spectral index selection criterion used for daily image composition. The typical range of standard uncertainty per datum of plant functional type fractional area coverage is 0.3 to 30.8 percent and depends on type abundance. Up to 3.5 percent of fractional coverage uncertainty is attributed to random fluctuation, higher uncertainty is caused by the variation of land cover classes. Errors in plant functional type fractional area coverage are typically anti-correlated. Confusion between natural and managed grass drives the uncertainty in African savannah.