Cross Validation

Please see Tramontana et al. (2016) for details and a thorough discussion!

​ Different ML approaches were applied to RS and RS+METEO setups using the same sets of predictor variables, and a thorough 10-fold ‘leave-towers-out’ cross-validation was conducted. Due to the computational expense of the RS+METEO setup, only one method representing each “family” – RF, MARS, ANN and KRR – was trained.

capturing variations the across-sites, seasonal and the deviations from the mean seasonal cycle variability

Coefficients of determination (R2) from the comparison of overall time series across-sites, mean seasonal cycle, and the anomalies, in particular: the determination coefficients between predictions by the ensemble median estimate of RS setup and observation (dark grey bars), between predictions by the ensemble median estimate of RS+METEO setup and observation (light grey bars), and between the two ensembles median estimate (white bars). Whiskers were the higher and lower R2 when the comparisons were made among the singular ML. The comparison of output by the multiple regressions was also shown (black points)

• General performance: Rn > H > LE > GPP > TER > NEE.
• Striking consistency between RS and RS+METEO.
• Between-sites variability was in general well captured by machine learning methods (best for GPP and TER). It suggests that machine learning methods are suitable to reproduce the spatial pattern of mean annual fluxes.
• Less predictive capability of the anomalies.