Estimating Green Area Index (GAI) from radiative transfer model: application of the Bayesian theory to account for crop-specificities

Abstract:

Estimating crop Green Area Index (GAI) from optical remote sensing receives an extensive interest in monitoring crop structure and accessing to its functioning. Using radiative transfer models (RTM) to retrieve biophysical variables such as GAI from spectral reflectance offers the advantage to exploit our knowledge on the physical mechanisms governing light interaction within the canopy. Nevertheless, the full inversion of RTM presents several limitations including the high dimensionality of the problems, convergence problems due to possible equifinality, or the dependence of some RTM variables on the crop-specific architecture.

The objective of this study is to evaluate the performances of an RTM inversion approach based on the Bayesian theory to overcome some of these limitations. The PROSPECT+SAILH model is first inverted over a comprehensive global dataset of ground GAI measurements for maize, wheat, and rice species to derive crop-specific model input variables, allowing to closely simulate canopy reflectance observed by Landsat8, Sentinel2, and Quickbird satellites for the given measured GAI values. The Hamiltonian Monte Carlo (HMC) algorithm was used to construct the posterior distributions of these crop-specific variables. They were then used as prior information in the inversion process to estimate GAI values from the reflectance collected over observations not used previously while knowing the corresponding species.

Validation results showed that the approach succeeds to retrieve GAI with respectively Root Mean Square Error of 0.96, 1.79, and 1.63 for maize, wheat, and rice. Some discrepancies were found for particular observations where in situ GAI were probably erroneous. The approach presented here permits to efficiently combine reflectance from different sensors and bands to derive the crop-specific variables within a single process. Additionally, the HMC provides as well the uncertainties of the GAI estimates, which are necessary for biophysical modelling applications (e.g. data assimilation). It also has promising implications for the simultaneous retrieval of other variables including chlorophyll content.