A novel method to reconstruct normalized difference vegetation index time series based on temporal-spatial iteration estimation

Abstract

Reconstructing normalized difference vegetation index (NDVI) time series datasets is essential for monitoring long-term changes of the terrestrial surface. Here, a temporal-spatial iteration (TSI) method was developed to estimate the NDVIs of contaminated MODIS13Q1 pixels based on reliable MODIS13Q1 data. NDVIs of contaminated pixels were firstly computed through linear interpolation of adjacent high-quality pixels in the temporal series. Then, undetermined NDVIs of contaminated pixels were derived using the NDVI of the high-quality pixel that reflected the most similar land cover within the same ecological region, based on the weighted trajectory distance algorithm. These two steps were repeated iteratively, taking the estimated NDVIs as high-quality NDVIs to estimate other undetermined NDVIs of contaminated pixels until all NDVIs of contaminated pixels were estimated. The accuracies of estimated NDVIs using TSI were clearly higher than the asymmetric Gaussian, Savitzky-Golay, and window-regression methods; root mean square error and mean absolute percent error decreased by 14.0–104.8% and 19.4–47.3%, respectively. Furthermore, the TSI method performed better over a variety of environmental conditions. Variation of performance by the compared methods was 8.8–17.0 times than that of the TSI method. The TSI method will be most applicable when large amount of contaminated pixels exist.