Several influential Earth system science studies in the last three decades were based on Normalized Difference Vegetation Index (NDVI) data from Advanced Very High Resolution Radiometer (AVHRR) series of instruments. Although AVHRR NDVI data are known to have significant uncertainties resulting from incomplete atmospheric correction, orbital drift, sensor degradation, etc., none of these studies account for them. This is primarily because of unavailability of comprehensive and location-specific quantitative uncertainty estimates.

The first part of this dissertation investigated the extent of uncertainty due to inadequate atmospheric correction in the widely used AVHRR NDVI datasets. This was accomplished by comparison with atmospherically corrected AVHRR data at AErosol RObotic NETwork (AERONET) sunphotometer sites in 1999. Of the datasets included in this study, Long Term Data Record (LTDR) was found to have least errors (precision=0.02 to 0.037 for clear and average atmospheric conditions) followed by Pathfinder AVHRR Land (PAL) (precision=0.0606 to 0.0418), and Top of Atmosphere (TOA) (precision=0.0613 to 0.0684).  

` Although the use of field data is the most direct type of validation and is used extensively by the remote sensing community, it results in a single uncertainty estimate and does not account for spatial heterogeneity and the impact of spatial and temporal aggregation. These shortcomings were addressed by using Moderate Resolution Imaging Spectrometer (MODIS) data to estimate uncertainty in AVHRR NDVI data. However, before AVHRR data could be compared with MODIS data, the nonstationarity introduced by inter-annual variations in AVHRR NDVI data due to orbital drift had to be removed. This was accomplished by using a Bidirectional Reflectance Distribution Function (BRDF) correction technique originally developed for MODIS data.

The results from the evaluation of AVHRR data using MODIS showed that in many regions minimal spatial aggregation will improve the precision of AVHRR NDVI data significantly. However temporal aggregation improved the precision of the data to a limited extent only.

The research presented in this dissertation indicated that the NDVI change of ~0.03 to ~0.08 NDVI units in 10 to 20 years, frequently reported in recent literature, can be significant in some cases. However, unless spatially explicit uncertainty metrics are quantified for the specific spatiotemporal aggregation schemes used by these studies, the significance of observed differences between sites and temporal trends in NDVI will remain unknown.