Uncertainty assessment of Sentinel-2-retrieved vegetation spectral indices over Europe

Abstract

ABSTRACT Vegetation spectral indices (VIs) from multispectral remotely sensed imagery provide useful information in several sectors, especially if longing for change detection analyses or land monitoring. In this context, estimating uncertainty of VI values is crucial to recognize significant differences in both space and time domains. Unexpectedly, most applications reported in literature and involving VI do not take care about this issue, thus making unreliable a significant part of deductions. In this work, authors present an approach aimed at mapping in time and space the theoretical uncertainty of some widely used VIs basing their approach on the so-called variance propagation law (VPL). VPL can be consequently used to get an estimate of the theoretical VI uncertainty, starting from one of the bands involved in VI computation. VI uncertainty all along the year 2020 was then mapped at pixel level by Google Earth Engine over the whole Europe to test seasonal trends. Uncertainty of VI differences, as possibly resulting from a change detection approach, was tested by comparing monthly composites of VI and computing the expected uncertainty of differences along the year. An example was reported involving two NDVI maps (June–September) proving that about 30% of ΔVI were not significant.