An adaptive spatiotemporal tensor reconstruction method for GIMMS-3g+ NDVI

Satellite-derived normalized difference vegetation index (NDVI) is inevitably contaminated by clouds and aerosols, causing large uncertainties in depicting the seasonal and interannual variations of terrestrial ecosystems, and potentially misrepresents their responses to climate change and climate extremes. Although various methods have been developed to reconstruct NDVI time series using the similarity in time, space or their combination, they typically require known and accurate data quality information. It is still challenging to effectively reconstruct high-quality NDVI from Global Inventory Modeling and Mapping Studies-3rd Generation V1.2 (GIMMS-3g+), which is one of the longest observation records but lacks reliable data quality information. This study introduces an adaptive spatiotemporal tensor reconstruction algorithm that leverages the spatial and temporal patterns of vegetation to produce high-quality long-term NDVI datasets without the need of data quality information. The reconstruction process employs two different tensor completion models to satisfy the low-rank constraints. These two models can effectively remove the high-frequency noises originating from atmospheric contamination, while preserving the abrupt or low-frequency changes attributable to disturbances such as drought, even in the absence of data quality information. The resultant NDVI shows good consistency with observations from geostationary satellites. Regions that show a strong correlation (r > 0.7) with geostationary satellite NDVI increased from 46.7 % (original GIMMS-3g+) to 62.2 % and 62.3 % (two reconstructions results) for East Asia, and from 41.4 % to 58.0 % and 59.0 % for Amazon. Our method also demonstrates superior performance to traditional methods such as Whittaker, HANTS, SG-filter, and comparable performance with the state-of-the-art ST-Tensor method when the fraction of contaminated observations is low. The proposed method can also be applied to other datasets such as EVI, LAI, etc., without additional data quality inputs. The resultant vegetation index dataset has the potential to improve plant phenology retrievals and evaluation of ecosystem responses to extremes.