3.3 Observations from remote sensing prediction
The leaf traits, like W leaf,N leaf and LAI, were predicted using the GPR modelling of hyperspectral image data, as leaf traits are more likely observed and thus able to be robustly predicted from collected remote sensing features, compared with other observations such asW above, W grain,N above and N grain. The performance of their site-specific predictions was evaluated in the training (Fig. 4a-c) and testing (Fig. 4d-f) datasets. We tried to reduce the overfitting by averaging the values of estimated hyperparameters from the repeated training subsets (Section 2.2.2). The predicted leaf traits in the training set performed better than those in the testing set. The R 2 values of the predicted leaf traits in the training set were higher than 0.95, while those in the testing set were lower than 0.88. The NRMSE of predicted leaf traits in training set ranged from 0.110 to 0.172, while theNRMSE in the testing test increased and varied from 0.175 to 0.336 (Fig. 4). Compared with N leaf (Fig. 4b, e),W leaf (Fig. 4a, d) and LAI (Fig. 4c, f) fitted with their measurements better in both the training set and the testing set.
Maps of the predicted leaf traits at the experimental site were generated, in which the temporal and spatial differences of crop growth at different growing stages were predicted (Fig. 5). In the treatments with low N input, the mean values of predicted leaf traits were low at the stem-elongating stage (Fig. 5a-c), and it later caused early senescence, which was reflected by the low values of the predicted leaf traits at the grain-filling stage (Fig. 5d-f). The mean and STD of the predicted leaf traits tended to be higher in experimental plots with higher N input at the stem-elongating stage (Fig. 5a-c). However, the predicted STD for leaf traits with lower predicted mean values at grain-filling stage tended to be even higher than those with higher predicted mean values at the stem-elongating stage (Fig. 5). The diverse uncertainties across the different growing seasons and the spatial variance between and within the experimental plots demonstrated the necessity of assimilating in situ observations into crop growth simulations for accurate forecasting of crop growth status.