A Tower-to-Satellite Machine Learning Framework for Daily Gross Primary Productivity Estimation in a Seasonally Dry Dipterocarp Forest, Thailand

Dry dipterocarp forests exhibit strong seasonal constraints on carbon uptake, but satellite predictive value remains uncertain. We evaluated a tower-to-satellite framework for daily gross primary productivity (GPP) at Sakaerat, Thailand, combining 698 days of eddy covariance data and meteorology with Sentinel-2 vegetation indices from non-overlapping 10-day median composites within an approximately 90% climatological flux footprint. Composites were assigned to EC days by backward-only matching (maximum age: 45 days). Four algorithms underwent nested purged blocked cross-validation and an independent 184-day holdout (February–August 2026). In 2025, daytime-partitioned (DT) GPP was 2302.62 g C m−2 yr−1, with 79% during the rainy season, whereas nighttime-partitioned (NT) GPP was 60.11% higher. During development, Sentinel-2 reduced error for random forest and XGBoost but increased it for support vector regression (SVR) and ridge regression under DT; under NT, all four improved. M4-SVR had the lowest mean outer-fold RMSE (1.582 vs. 1.586 g C m−2 d−1 for M1-SVR) but a higher pooled OOF RMSE (1.751 vs. 1.649). Holdout performance declined (RMSE = 3.313 g C m−2 d−1; R2 = 0.404), with negative within-season R2 and high-GPP underestimation. These results indicate conditional satellite gains beyond seasonal and meteorological baselines, with limited within-season daily skill; full-year and cross-site generalization remain untested.