Cross-validation performance of machine-learning models for global cropland GPP upscaling under observation-, year-, site-, and region-holdout schemes (1994–2014 FLUXNET training data)

<p dir="ltr">This dataset contains the full cross-validation results underlying the model-performance assessment reported in Observation-constrained assessment of cropland productivity trends and drivers at the global scale<br></p><p dir="ltr">Three machine-learning algorithms (random forest, boosted regression trees, artificial neural network) were trained on monthly FLUXNET2015 GPP (GPP_NT_VUT_MEAN, QC > 0.8; 9,567 observations from 166 sites, 1994–2014), using monthly meteorological variables and one of three satellite vegetation indices (GIMMS3g NDVI, MODIS13C2 V6 NDVI, MODIS13C2 V6 EVI) as predictors.</p><p dir="ltr">Four holdout schemes of increasing stringency were applied: leave-observation-out (10% of records withheld at random), leave-year-out (10% of years withheld), leave-site-out (10% of sites withheld), and leave-region-out (one of six continents withheld at a time).

The first three were repeated ten times with independent random draws.</p><p dir="ltr">Reported metrics are the correlation coefficient (R, dimensionless), predictive coefficient of determination (R², dimensionless), root mean square error (RMSE, g C m⁻² month⁻¹) and bias (g C m⁻² month⁻¹).</p>