Influence of landscape spatial heterogeneity on the non-linear estimation of leaf area index from moderate spatial resolution remote sensing data

The monitoring of earth surface dynamic processes requires global observations of the structure and the functioning of vegetation. Moderate resolution sensors (with pixel size ranging from 250 m to 7 km) provide frequent estimates of biophysical variables to characterize vegetation such as the leaf area index (LAI). However, the computation of LAI from moderate resolution remote sensing data induces a scaling bias on the LAI estimate if the moderate resolution pixel is heterogeneous and if the transfer function that relates remote sensing data to LAI is non-linear.This study provides a model to evaluate and correct the scaling bias. The model is built first for a univariate semi-empirical transfer function relating LAI directly to NDVI. The scaling bias is a function of (i) the degree of non-linearity of the transfer function quantified by its second derivative and (ii) the spatial heterogeneity of the moderate resolution pixel quantified by the variogram of the high spatial resolution (20 m) NDVI image. Then, the model is extended to a bivariate transfer function where LAI is related to red and near infrared reflectances. The scaling bias depends on (i) the Hessian matrix of the transfer function and (ii) the variograms and cross variogram of the red and near infrared reflectances.The scaling bias is investigated on several distinct landscapes from the VALERI database. Adjusting for scaling bias is critical on crop sites which are the most heterogeneous sites at the landscape level. Regarding the univariate transfer function, the magnitude of the scaling bias increases rapidly with pixel size until this size is larger than the typical spatial scale of the data. For the bivariate transfer function, it results from the addition of several components that may add up or cancel each other out. It is thus more difficult to analyze.The accuracy of the model to estimate the scaling bias is discussed. It depends mainly on the ability of the variograms and cross variogram to represent the local dispersion variances and covariance within the moderate resolution pixel. The model is generally highly accurate at 1000 m spatial resolution for the univariate transfer function and less accurate for the bivariate transfer function.