4.2 Uncertainty estimation of remote sensing predictions and its
application in data assimilation
Like with the regression performance of GPR for leaf-level traits like
specific leaf weight (Wang et al., 2019) and chlorophyll content
(Verrelst et al., 2013a), our results showed high predictive performance
for the traits W leaf,N leaf and LAI, at canopy level, not only in the
training dataset (R 2 > 0.95), but
also in the testing dataset (R 2 >
0.81) (Fig. 4). Agreeing with the threshold proposed by Global Climate
Observing System for the ecological application of remote sensing
observations (GCOS, 2011), the relative uncertainties of predicted leaf
traits in our results were below 20% (Fig. 5). The predicted low
uncertainty was in line with other related GPR research (Verrelst et
al., 2016; Wang et al., 2019). Similar to the results of Wang et al.
(2019), relatively high uncertainty always came with high N addition in
the vegetative phase (Fig. 5a-c). We also found that the uncertainty
tended to increase after entering the reproductive phase, especially for
the treatments with low N input, caused by early leaf senescence (Fig.
5d-e). Meanwhile, as GPR captured temporal and spatial variations in
crop growth well, DArs performed better than
DAfm (Fig. 7), which agrees with the results of yield
forecasting by assimilating LAI into the APSIM-Wheat model (Zhang et
al., 2021b). The estimated uncertainty of remote sensing predictions
likely affected inflation factor estimation as well, as the inflation
factor is determined from the updated posteriors based on the
observation errors (Whitaker and Hamill, 2012). The contribution of the
GPR model should be further evaluated given that the uncertainty of
remote sensing observations is commonly overlooked in DA (Huang et al.,
2019) and arbitrarily set based on a general guess (e.g., Kang and
Özdoğan (2019); Ines et al. (2013); Fig. 8b).