Abstract
As part of the Grassland Research and Innovation Lab within the Horizon Europe project ScaleAgData (https://scaleagdata.eu/en), we assess the potential of Sentinel-1 (S1) SAR data to enrich the time series of grassland biophysical parameters derived from Sentinel-2 (S2) optical data in South Tyrol and Trentino, in North-Eastern Italy. Many studies have demonstrated the potential of SAR data in accurately estimating biophysical parameters of crops. However, over grasslands, there are fewer applications because of the difficulties in disentangling the different effects that influence the SAR signal, including vegetation composition, soil moisture, and terrain roughness. Further research is needed to optimize the estimation of grassland biophysical parameters from SAR data, including Leaf Area Index (LAI) and above-ground biomass, thus enhancing its usability for grassland monitoring and management. In contrast to the optical S2 data, S1 SAR is independent of the frequent cloud coverage in the Alps. By penetrating the grass canopy and by its dual-polarisation mode it provides useful information about vegetation structure, biomass, and moisture content. The acquisition in both descending and ascending orbits by the constellation (2016-2021) provides the high temporal resolution needed especially for managed meadows, where mowing leads to rapid changes in biomass content. These advantages make S1 SAR imagery promising for data fusion with S2-based LAI maps. Still, the topographic effects of the Alps must be corrected as they lead to foreshortening and layover which causes gaps with no data coverage. In this study we exploit different machine learning regressors to estimate LAI from S1 data, utilizing S2 LAI derived by the S2 SNAP biophysical processor as the target variable, and day of the year (DOY), soil moisture data, and various features derived from S1 SAR backscattering as predictors. Different indices and polarisations as well as descending versus ascending mode are examined. We cluster the meadows in the Alps by altitudinal zones and assess the performance of non-linear regressors, such as Random Forest (RF) and Gaussian Process Regressor (GPR) across these zones. Preliminary results show that for both regressors especially DOY, combinations of VV and VH polarisation and soil moisture lead to highly accurate predictions on a test set (R² = 0.91, RMSE = 0.29 for RF, R² = 0.83, RMSE = 0.4 for GPR). In addition to the modelling experiments, we collect ground-derived key biophysical parameters, including LAI, above-ground biomass, vegetation composition, and soil moisture at eight forage production meadows over South Tyrol and Trentino during the growing season in 2023. This data will be used to validate the LAI product obtained from the S1-S2 data fusion and to derive a model to estimate grassland above-ground biomass from the gap-filled LAI. Finally, these estimations will serve to improve a drought index that Eurac Research has developed to estimate mountain grassland yield losses due to drought conditions for insurance purposes.