Remote Sensing–driven ensemble smoother assimilation of LAI for regional sugarcane yield estimation

Sugarcane plays a critical role in global sugar and ethanol production, demanding accurate and timely yield monitoring. While integrating process-based crop models (PBMs) with data assimilation (DA) has shown potential for enhancing yield estimation, a comprehensive DA framework applied at regional scales remains underexplored. This study addresses this gap by improving sugarcane yield estimation in Brazil’s leading production region through the assimilation of remotely sensed Leaf Area Index (LAI) data into a stochastic PBM. Using over 167,000 LAI observations derived from Landsat 7 ETM + imagery, data were assimilated into the DSSAT/SAMUCA model via the Ensemble Smoother (ES) method to reduce model uncertainty. The analysis spanned the 2003–2013 period and employed the Technology Extrapolation Domains (TEDs) framework to categorize sugarcane-growing areas in São Paulo State according to their biophysical attributes. Results demonstrated substantial improvements in yield estimation accuracy. Across all TEDs, Root Mean Square Error (RMSE) was reduced from 43.98 to 17.83 Mg ha⁻¹ and Mean Absolute Error (MAE) from 41.85 to 11.65 Mg ha⁻¹, corresponding to average reductions of 57% and 72%. Notably, some individual TEDs showed even larger improvements, with MAE reductions reaching up to 96% in the best-performing cases. All values refer to pixel-wise simulations aggregated at the TED level. These outcomes highlight the potential of incorporating remote sensing data into process-based modeling frameworks to improve regional-scale yield estimation and strengthen retrospective assessment of crop performance.