On estimating the gross primary productivity of Mediterranean grasslands under different fertilization regimes using vegetation indices and hyperspectral reflectance

We applied an empirical modelling approach for gross primary productivity (GPP) estimation from hyperspectral reflectance of Mediterranean grasslands undergoing different fertilization treatments. The objective of the study was to identify combinations of vegetation indices and bands that best represent GPP changes between the annual peak of growth and senescence dry out in Mediterranean grasslands. In situ hyperspectral reflectance of vegetation and CO2 gas exchange measurements were measured concurrently in unfertilized (C) and fertilized plots with added nitrogen (N), phosphorus (P) or the combination of N, P and potassium (NPK). Reflectance values were aggregated according to their similarity (r 90 %) in 26 continuous wavelength intervals (Hyp). In addition, the same reflectance values were resampled by reproducing the spectral bands of both the Sentinel-2A Multispectral Instrument (S2) and Landsat 8 Operational Land Imager (L8) and simulating the signal that would be captured in ideal conditions by either Sentinel-2A or Landsat 8. An optimal procedure for selection of the best subset of predictor variables (LEAPS) was applied to identify the most effective set of vegetation indices or spectral bands for GPP estimation using Hyp, S2 or L8. LEAPS selected vegetation indices according to their explanatory power, showing their importance as indicators of the dynamic changes occurring in community vegetation properties such as canopy water content (NDWI) or chlorophyll and carotenoids = chlorophyll ratio (MTCI, PSRI, GNDVI) and revealing their usefulness for grasslands GPP estimates. For Hyp and S2, bands performed as well as vegetation indices to estimate GPP. To identify spectral bands with a potential for improving GPP estimates based on vegetation indices, we applied a two-step procedure which clearly indicated the short-wave infrared region of the spectra as the most relevant for this purpose. A comparison between S2- and L8-based models showed similar explanatory powers for the two simulated satellite sensors when both vegetation indices and bands were included in the model. Altogether, our results describe the potential of sensors on board Sentinel-2 and Landsat 8 satellites for monitoring grassland phenology and improving GPP estimates in support of a sustainable agriculture management