Thermal time dominates satellite predictors for within-season wheat biomass estimation under spatial cross-validation

Francisco Zambrano , Abel Herrera, Mauricio Molina-Roco

Satellite imagery combined with machine learning (ML) is widely used to estimate crop above-ground biomass (AGB), but model accuracy at sites not used for training is rarely compared with simple phenological baselines. We evaluated how much Sentinel-1 (S1) backscatter, Sentinel-2 (S2) and PlanetScope (PS) spectral predictors, precipitation, and soil moisture add to accumulated growing degree days ($\sum$GDD) for within-season wheat (*Triticum aestivum*) AGB estimation across four Mediterranean field-seasons in central Chile (153 destructive samples tracking biomass progression from sowing to ripening, 2020 to 2023). Fourteen predictor sets were evaluated with linear regression, random forest (RF), and XGBoost under leave-one-site-season-out cross-validation (LOSO-CV) with fixed hyperparameters, and the model with all predictors was interpreted with held-out permutation importance and accumulated local effects. A linear model on $\sum$GDD alone reached R² = 0.53 (RMSE = 6.24 t/ha) and outperformed days after sowing (R² = 0.19, RMSE = 7.85 t/ha). Models based on a single sensor transferred worse than this baseline (RMSE = 8.22 t/ha for S2, 8.57 t/ha for PS, and 10.92 t/ha for S1 with RF); S1 results should be interpreted cautiously as data were used without speckle filtering or incidence angle normalization. The best-performing configuration in this sample, RF with S1 and $\sum$GDD, reached R² = 0.58 (RMSE = 5.80 t/ha, MAE = 4.10 t/ha) and improved on the baseline in three of four held-out site-seasons (mean RMSE reduction of 5.9%), but this gain was not reproduced with XGBoost. RF with all predictors (R² = 0.51, RMSE = 6.77 t/ha) did not improve on the baseline, and its most important optical predictors were cumulative indices that behaved as measures of season progress. With only four site-seasons, statistical power was limited, and differences among the leading configurations were smaller than the variability among held-out site-seasons. For within-season biomass tracking, most of the transferable AGB signal was captured by thermal time, and satellite predictors provided at most modest, configuration-dependent gains that may reflect both satellite information and model nonlinearity. This sampling design favors predictors that track phenological progression; satellite sensors may demonstrate greater value for detecting condition differences at the same growth stage across fields or seasons. AGB studies should report phenological baselines evaluated under site-level validation before attributing skill to remote sensing predictors.