Unification of NDVI Definitions and Areal Fraction Models in Remote Sensing of Vegetation

Vegetation greenness and areal coverage are important parameters for understanding the global carbon cycle and hydrological cycle by the terrestrial biosphere. Many studies that monitor greenness and coverage are based on the Normalized Difference Vegetation Index (NDVI) and various derivative models of coverage. Varieties of NDVI definition and model inconsistency hinder our understanding of the carbon and water cycle and bias carbon sequestration and evapotranspiration characterization. To clarify these issues, we analyzed the consequences of using a variety of NDVI definitions for several commonly used multi- and hyper-spectral satellite sensors ALI, IKONOS, ASTER, ETM+, HRVIR, Hyperion, and then examine the performance of different models of areal vegetation coverage - the linear reflectance model, the linear NDVI model, and the quadratic NDVI model. We conclude that: (1) spectral reflectance should be used in the calculation of NDVI. Using spectral radiance to calculate NDVI defined by spectral reflectance can always underestimate the NDVI and result in an absolute error as high as 0.14 for ALI, 0.15 for IKONOS, 0.16 for ASTER, and 0.20 for ETM+ and HRVIR; (2) using DN value to calculate NDVI should always be avoided because data from different sensors are generally different in radiometric resolutions and thus their respective DN values carry different levels of information and can not be directly compared. Besides, the atmospheric correction to spectral DN values is often difficult because the physics in the atmospheric correction to spectral DN value is vague and often confusing. Using DN to calculate NDVI defined by spectral reflectance can result in an error as high as 0.23 (overestimate) for ALI, 0.17 (overestimate) for ASTER, 0.47 (underestimate) for ETM+, 0.39 (underestimate) for HRVIR, 0.15 (underestimate) for Hyperion, 0.02 (underestimate) for IKONOS. (3) in deriving the vegetation areal coverage, the linear reflectance model generally gives results more consistent with in situ measurements for shrub biomass zone than the other two models; and (4) the linear NDVI model outperforms the other two models in piñon-juniper biomass zone. These observations are consistent with the fact that non- linear effect is less important in shrub land than in piñon-juniper woodland. The linear NDVI model seems to be more capable of capturing non-linearity in the spectral analysis than the other two models, though more tests are needed to quantify its capability.