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.