Sentinel-2 temporal upscaling of eddy-covariance GPP in Brazilian Pampa grasslands under hydroclimatic variability (2019–2022)

Grassland productivity is strongly influenced by hydroclimatic variability, yet the temporal transferability of Sentinel-2-based gross primary productivity (GPP) models under contrasting climatic conditions remains poorly understood. This study quantified hydroclimatic variability and ecosystem productivity dynamics in a managed native grassland of the Brazilian Pampa during 2019–2022 and evaluated statistical, machine-learning, and process-inspired approaches for upscaling 5-day eddy-covariance GPP (GPPEC) using Sentinel-2 spectral information and meteorological drivers under strict temporal validation (training: 2019–2021; testing: 2022). The study period was characterized by recurrent meteorological dryness, with negative annual precipitation anomalies throughout 2019–2022. Nevertheless, local precipitation and radiation increased in 2022 relative to the two preceding years, coinciding with the highest annual ecosystem productivity observed during the study period. During the independent 2022 test year (N=39), models combining Sentinel-2 and climatic predictors consistently outperformed the corresponding single-source configurations. GAM achieved the highest predictive coefficient of determination (Rpred2=0.53) and the lowest RMSE (1.62 g C m −2 d −1), whereas Elastic Net achieved the lowest MAE (1.30 g C m −2 d −1). SVM-RBF showed the strongest observed–predicted association (Pearson r=0.90), highlighting that correlation-based association and predictive accuracy capture complementary aspects of model performance. These results demonstrate the value of integrating canopy-state information from Sentinel-2 with short-term environmental forcing for GPP estimation. At this EC site, the results show that integrating Sentinel-2 observations with climatic information improves short-term GPP prediction and supports site-specific temporal transferability across contrasting hydroclimatic conditions. The findings also emphasize the importance of strict independent-year validation when assessing the temporal generalization of ecosystem productivity models in subtropical grasslands.