Spatially constrained inversion of radiative transfer models for improved LAI mapping from future Sentinel-2 imagery

The robust and accurate retrieval of vegetation biophysical variables using radiative transfer models (RTM) is seriously hampered by the ill-posedness of the inverse problem. With this research we further develop our previously published (object-based) inversion approach [Atzberger, 2004, RSE 93: 53–67] and evaluate it against simulated Sentinel-2 data. The proposed RTM inversion takes advantage of the geostatistical fact that the biophysical characteristics of nearby pixels are generally more similar than those at a larger distance. This leads to spectral co-variations in the n-dimensional spectral features space, which can be used for regularization purposes. A simple two-step inversion based on PROSPECT + SAIL generated look-up-tables is presented that can be easily adapted to other radiative transfer models. The approach takes into account the spectral signatures of adjacent pixels in gliding (3 × 3) windows. Using a set of leaf area index (LAI) measurements (n = 26) acquired over alfalfa, sugar beet and garlic crops of the Barrax test site (Spain), it is demonstrated that the proposed regularization using neighbourhood information yields more accurate results compared to the pixel-based inversion. With the proposed regularization, the RMSE between ground measured and Sentinel-2 derived LAI is 0.54 m2/m2 and hence significantly lower compared to the RMSE of the standard inversion approach (RMSE: 1.46 m2/m2) and also of higher accuracy compared to a scaled NDVI model (RMSE: 0.90 m2/m2). At the same time, a positive effect on the modelled leaf chlorophyll content (Cab) is noticed, albeit too few field measurements were available for deriving statistically sound results. For the other retrieved biophysical parameters such as leaf dry matter content (Cm), soil brightness (αsoil) and average leaf angle (ALA) the proposed algorithm yields more plausible and spatially consistent results. Altogether the findings confirm the positive effect of regularizing the model inversion using spatial constraints. As for any other inversion strategy, the approach requires a RTM well suited for the crop under study. For three additional crops (maize, potatoes and sunflower), the forward modelling with field measured LAI did not match the observed signatures. Consequently, for these canopies both the standard and the object-based inversion failed.