On the performance of remote sensing time series reconstruction methods – A spatial comparison

Abstract

The satellite observed Normalized Difference Vegetation Index (NDVI) time series, which describe the temporal and spatial variability of global terrestrial vegetation, are inevitably contaminated by clouds, aerosol, snow and ice cover. In general all these conditions yield negative deviations in the time series of NDVI. Many time series reconstruction models have been developed to eliminate effect of the negative deviations and most of them perform differently in different applications and regions. The Harmonic Analysis (HA), Double logistic (DL), Asymmetric Gaussian (AG), Whittaker smoother (WS) and Savitzky–Golay filter (SG) are five of the most widely used time series reconstruction models owing to their simplicity of implementation or the capability to extract phenological metrics from the time series. The performance of these models varies with the NDVI signal and the noise distribution, however, and until now there is no consensus on which method outperforms all others under all situations. Since the NDVI signal and the noise distribution are highly dependent on regional climate and land cover, the reconstruction performance is expected to be spatially variable. Thus this study compared the five reconstruction models at pixel scale to provide practical and biome – specific recommendations for future time series reconstruction applications. Specifically, the 14 years raw daily reflectance data and ancillary Quality Assessment (QA) information from the MODIS sensor were used to generate pixel reference series and noisy series. Then the five candidate models were applied to both reference series and noisy series and three reconstruction performance metrics i.e. Overall Reconstruction Error (ORE), Fitting Related Error (FRE), and Normalized Noise Related Error (NNRE), were calculated. Finally, the performance of the five candidate reconstruction models was evaluated by applying the three metrics. The preliminary results showed that when considering ORE only, the Asymmetric Gaussian model outperforms other models over most areas of high latitude boreal region, while the Savitzky–Golay model gives the best reconstruction performance in tropical and subtropical regions. The FRE and the NNRE helped to reveal the main error sources in the reconstruction in different regions. The comparison method developed and applied in this study led to suggest adaptive selection of the best reconstruction model for specific NDVI signals and noise distribution.