An enhanced vegetation index time series for the Amazon based on combined gap-filling approaches and quality datasets

Vegetation indices from MODIS data are subject to residual atmospheric noise, affecting processes requiring data continuity and analyses. This work reconstructed a time series of MODIS EVI mosaics for the Amazon using a novel combination of curve-fitting and spatiotemporal gap-filling. TIMESAT was used for initial curve fitting and gap filling, using a Double Logistic method and MODIS Usefulness values as weights. Pixels with large temporal gaps were handled by a spatiotemporal gap filling approach. The method scans Julian Days before and after the image being gap filled, searching for a good quality pixel (Pg) at the location of the pixel to be replaced. If Pg is found, a window is defined around it and a search for good quality pixels (Px) with spectral characteristics similar to Pg is performed. Window size increases during processing and pixel similarity uses Euclidean distance based on MOD13A2 reflectances. A good quality EVI value for the image being gap filled and at the location analogous to the minimum distance Px replaces the low quality pixel. Results from the spatiotemporal gap filling were then used in TIMESAT for smoothing. An evaluation strategy of the spatiotemporal approach involved flagging 5,000 randomly selected good-quality pixels as low-quality, running the algorithm and regressing the results with the original EVI values (R(2)= 0.62). The combined strategy was able to find replacement pixels and reduce spikes for images with high cloud cover and was used to rebuild a time series of EVI over the Amazon region for the period 2000-2010.