Assessment of vegetation indices for regional crop green LAI estimation from Landsat images over multiple growing seasons

Abstract

There is an increasing need to monitor the dynamics of green LAI of field crops through the growing season. A simple approach is to use a regression model to estimate crop LAI from a vegetation index derived from optical remote sensing data. However, variations of interference factors in the signal path could induce variations in spectral reflectance, leading to uncertainty in LAI estimation. A semi-empirical equation was implemented to estimate green LAI of field crops from Landsat-5/7 data using a few vegetation indices, including the normalized difference vegetation index (NDVI), the optimized soil adjusted vegetation index (OSAVI), the two band enhanced vegetation index (EVI2) and the modified triangular vegetation index (MTVI2). Data were collected during several growing seasons, from 1999 to 2006, over corn, soybean, and spring wheat fields in an experimental farm in Ottawa (ON, Canada). LAI estimated for corn, soybean and wheat from Landsat data using the vegetation indices was compared to ground LAI. Except for NDVI, comparable results were obtained from the other three vegetation indices, with a coefficient of determination above 0.83 and a root mean square error (RMSE) not more than 0.60. The performance of NDVI was less satisfactory (RMSE>0.66). The uncertainties in LAI estimation induced by variations in soil reflectance, leaf optical properties, canopy structure, and atmospheric conditions were assessed through a global sensitivity analyses using the PROSPECT leaf model coupled to the SAIL canopy model along with the 6S atmospheric transmission model. The sensitivity analyses show that different indices are affected differently by the various interference factors. Comparatively, NDVI is the most influenced by leaf chlorophyll but the least affected by leaf inclination, OSAVI and the narrow band MTVI2 are more efficient in reducing soil effects, and EVI2 has a better performance in reducing aerosol perturbation. At high LAI, the uncertainty of NDVI is the smallest, but the uncertainty propagated to LAI estimation is the largest due to saturation. In this case, vegetation indices that are less prone to saturation should be considered, such as EVI2 and MTVI2. When MTVI2 is used on multispectral data, its ability to reduce soil and leaf chlorophyll perturbation is similar to EVI2 but weaker than when it is used on hyperspectral data. These results show that vegetation indices can be used in a simple regression model to generate baseline green LAI product for seasonal crop growth monitoring, however it is important to be aware of the sources of uncertainty and their relative amplitudes when using the product.