Evaluating the effectiveness of smoothing algorithms in the absence of ground reference measurements

Time series of vegetation indices like NDVI are used in numerous applications ranging from ecology to climatology and agriculture. Often, these time series have to be filtered before application. The smoothing removes noise introduced by undetected clouds and poor atmospheric conditions. Ground reference measurements are usually difficult to obtain due to the medium/coarse resolution of the imagery. Hence, new filter algorithms are typically only (visually) assessed against the existing smoother. The present work aims to propose a range of quality indicators that could be useful to qualify filter performance in the absence of ground-based reference measurements. The indicators comprise (i) plausibility checks, (ii) distance metrics and (iii) geostatistical measures derived from variogram analysis. The quality measures can be readily derived from any imagery. For illustration, a large SPOT VGT dataset (1999–2008) covering South America at 1 km spatial resolution was filtered using the Whittaker smoother.