Propagating Sentinel-2 Top-of-Atmosphere Radiometric Uncertainty Into Land Surface Phenology Metrics Using a Monte Carlo Framework

Abstract

Time series of optical imagery allow one to derive land surface phenology metrics. These metrics are only complete with a statement about their uncertainty. A source of uncertainty is the radiometry of the sensor. We propagated radiometric uncertainties within a Monte Carlo framework into phenological metrics using the TIMESAT approach based on time series of the normalized difference vegetation index (NDVI), three-band enhanced vegetation index (EVI), and green leaf area index (GLAI) derived from radiative transfer modeling. In addition, we studied the effect of propagated uncertainties on scene preclassification. We focused on Sentinel-2 multispectral imager top-of-atmosphere data since quantitative estimates of radiometric uncertainties are available. Propagation was carried out for a growing season over an agricultural region in Switzerland. Propagated uncertainties had little impact on the classification except for spectrally mixed pixels. Effects on the spectral indices and GLAI were more pronounced. In detail, the GLAI was more uncertain due to the ill-posedness of radiative transfer model inversion (median relative uncertainty for all crop pixels and Sentinel-2 scenes: 4.4%) than EVI (2.7%) and NDVI (1.1%). Regarding phenology, metrics exhibited largest uncertainties in the case of GLAI. The magnitude of uncertainty in the metrics depends on the interscene error correlation, which we assumed to be either zero (uncorrelated) or one (fully correlated) since the actual correlation is unknown. If uncertainties are fully correlated, uncertainties in metrics are small (two to three days) but take values up to greater ten days under the uncorrelated assumption. Thus, our work provides guidance for the interpretation of phenological metrics.