Improving retrieval of crop biophysical properties in dryland areas using a multi-scale variational RTM inversion approach

Optical radiative transfer models (RTM) are used to study the relationship between vegetation biophysical properties and corresponding canopy reflectances. In this paper, a RTM is inverted with satellite based surface reflectance observations to estimate key vegetation biophysical properties such as leaf area index (LAI) and leaf chlorophyll content (Cab). The complexity of model inversion makes model optimization challenging, particularly in dryland areas where the vegetation signal may have become confounded by bright soil backgrounds. Add to this, general difficulties in separating vegetation properties contributing to the combined surface reflectance signal. In this study, using a Bayesian approach, the inversion approach is written as a cost function to minimize. The high non-linearity of the RTM makes the analytical resolution of the optimization unpractical. To overcome this problem, a new multi-scale variational inversion approach is proposed. It solves progressively the inversion problem by first simplifying it, then solving it, and then coming back progressively to the original inversion problem. The approach is tested over a dryland irrigated agricultural system composed of fields of alfalfa, Rhodes grass, carrots and maize. Validation is done comparing results to in-situ measurements and other commonly used retrieval methods. The retrieved properties are shown to be in good agreement with in-situ observations of Cab (RMSE = 0.2 mu g cm(-2) and R2 = 89%) and LAI (RMSE = 0.18 m(2) m(-2) and R2 = 92%), which is shown to be an improvement over other traditional variational techniques (Gradient, Newton, LUT and QNT).