Bayesian inversion of a forest reflectance model using Sentinel-2 and Landsat 8 satellite images

The inversion of reflectance models is a generalizable tool to obtain estimates on forest biophysical parameters, such as leaf area index, with theoretically little information need from a study area, instead relying on the knowledge about physical processes in the forest radiation regime. The use of prior information can greatly improve the reflectance model inversion, however, the literature does not yet provide much information on the selection of priors and their influence on the inversion results. In this study, we used a Bayesian approach to invert the PARAS forest reflectance model and retrieve leaf area index from Sentinel-2 MSI and Landsat 8 OLI multispectral satellite images. The PARAS model is based on the theory of spectral invariants, which describes the influence of wavelength-independent parameters on forest radiative transfer. The Bayesian inversion approach is highly flexible, provides uncertainty quantification, and enables the explicit incorporation of prior knowledge into the inversion process. We found that the choice of prior information is crucial in inverting a forest reflectance model to predict leaf area index. Regularizing and informative priors for leaf area index strongly improved the predictions, relative to an uninformative prior, in that they counteracted the saturation effect of the optical signal occuring at high values for leaf area index. The predictions of leaf area index were more accurate for Landsat 8 than for Sentinel-2, due to potential inconsistencies in the visible bands of Sentinel-2 in our data, and the higher spectral resolution.Peer reviewe