Fig. 8: Trade-off between model overfitting and extrapolability observed in this study.

Variation of the prediction deviation (i.e., the RMSE difference between the training and test sets) at different VHT alteration rates, based on the (a) non-overfitted and (b) slight-overfitted GPR models. The interpolation and extrapolation ranges are roughly divided according to the data distribution in the VHT dataset. Note that * represents the unit of the scaled VHT alteration rate, i.e., g/(m2·d).