Gaussian processes uncertainty estimates in experimental Sentinel-2 LAI and leaf chlorophyll content retrieval

Abstract

ESA’s upcoming Sentinel-2 (S2) Multispectral Instrument (MSI) foresees to provide continuity to land monitoring services by relying on optical payload with visible, near infrared and shortwave infrared sensors with high spectral, spatial and temporal resolution. This unprecedented data availability leads to an urgent need for developing robust and accurate retrieval methods, which ideally should provide uncertainty intervals for the predictions. Statistical learning regression algorithms are powerful candidats for the estimation of biophysical parameters from satellite reflectance measurements because of their ability to perform adaptive, nonlinear data fitting. In this paper, we focus on a new emerging technique in the field of Bayesian nonparametric modeling. We exploit Gaussian process regression (GPR) for retrieval, which is an accurate method that also provides uncertainty intervals along with the mean estimates. This distinct feature is not shared by other machine learning approaches. In view of implementing the regressor into operational monitoring applications, here the portability of locally trained GPR models was evaluated. Experimental data came from the ESA-led field campaign SPARC (Barrax, Spain). For various simulated S2 configurations (S2-10m, S2-20m and S2-60m) two important biophysical parameters were estimated: leaf chlorophyll content (LCC) and leaf area index (LAI). Local evaluation of an extended training dataset with more variation over bare soil sites led to improved LCC and LAI mapping with reduced uncertainties. GPR reached the 10% precision required by end users, with for LCC a NRMSE of 3.5–9.2% (r2: 0.95–0.99) and for LAI a NRMSE of 6.5–7.3% (r2: 0.95–0.96). The developed GPR models were subsequently applied to simulated Sentinel images over various sites. The associated uncertainty maps proved to be a good indicator for evaluating the robustness of the retrieval performance. The generally low uncertainty intervals over vegetated surfaces suggest that the locally trained GPR models are portable to other sites and conditions.