Exploring The Impact of Noise On Hybrid Inversion of PROSAIL RTM On Sentinel-2 Data
Exploring The Impact of Noise On Hybrid Inversion of PROSAIL RTM On Sentinel-2 Data
Article
Exploring the Impact of Noise on Hybrid Inversion of
PROSAIL RTM on Sentinel-2 Data
Nuno César de Sá 1, * , Mitra Baratchi 2 , Leon T. Hauser 1 and Peter van Bodegom 1
Abstract: Remote sensing (RS) of biophysical variables plays a vital role in providing the information
necessary for understanding spatio-temporal dynamics in ecosystems. The hybrid approach to retrieve
biophysical variables from RS by combining Machine Learning (ML) algorithms with surrogate data
generated by Radiative Transfer Models (RTM). The susceptibility of the ill-posed solutions to noise
currently constrains further application of hybrid approaches. Here, we explored how noise affects
the performance of ML algorithms for biophysical trait retrieval. We focused on synthetic Sentinel-2
(S2) data generated using the PROSAIL RTM and four commonly applied ML algorithms: Gaussian
Processes (GPR), Random Forests (RFR), and Artificial Neural Networks (ANN) and Multi-task
Neural Networks (MTN). After identifying which biophysical variables can be retrieved from S2
using a Global Sensitivity Analysis, we evaluated the performance loss of each algorithm using the
Mean Absolute Percentage Error (MAPE) with increasing noise levels. We found that, for S2 data,
Carotenoid concentrations are uniquely dependent on band 2, Chlorophyll is almost exclusively
dependent on the visible ranges, and Leaf Area Index, water, and dry matter contents are mostly
Citation: de Sá, N.C.; Baratchi, M.; dependent on infrared bands. Without added noise, GPR was the best algorithm (<0.05%), followed
Hauser, L.T.; van Bodegom, P. by the MTN (<3%) and ANN (<5%), with the RFR performing very poorly (<50%). The addition of
Exploring the Impact of Noise on noise critically affected the performance of all algorithms (>20%) even at low levels of added noise
Hybrid Inversion of PROSAIL RTM (≈5%). Overall, both neural networks performed significantly better than GPR and RFR when noise
on Sentinel-2 Data. Remote Sens. 2021, was added with the MTN being slightly better when compared to the ANN. Our results imply that
13, 648. https://doi.org/10.3390/
the performance of the commonly used algorithms in hybrid-RTM inversion are pervasively sensitive
rs13040648
to noise. The implication is that more advanced models or approaches are necessary to minimize the
impact of noise to improve near real-time and accurate RS monitoring of biophysical trait retrieval.
Academic Editor: Katja Berger
Received: 7 December 2020
Keywords: radiative transfer models; PROSAIL; sensitivity analysis; inversion; biophysical variables;
Accepted: 8 February 2021
Published: 11 February 2021
machine learning; multi-output; Bayesian optimization; Sentinel-2
best possible inversion of RTMs, which is the focus of our study. This may serve as an
initial step towards global applications for biophysical trait retrieval independent of sensor,
time, and location.
2. Methods
2.1. General Methodology
Our approach consisted of four main steps, and a clear overview is given in Figure 1.
The first step consists of identifying which biophysical variables can potentially be detected
by S2 based on a GSA and band correlation. To perform the GSA, we used an Enhanced
Fourier Amplitude Sensitivity Analysis (EFAST) [31] on the simulated S2-like data to
identify which traits can be detected. The second step was finding the best hyperparameters
of each model to ensure the best possible version of each ML modelling approach. This was
performed using the Tree-structured Parzen Estimator (TPE), which uses Bayesian inference
to search for the best combination of parameters within a pre-set search space [32,33].
Sensitivity Hyperparameter
RTM Inversion Impacts of noise
analysis tuning
Figure 1. The first step consists of identifying which biophysical variables can be detected by S2 while the second step
uses Bayesian optimization to find the best parameters for the various models. Lastly, the third and fourth steps test the
performance of the various algorithms for inverting simulated S2-like data using Radiative Transfer Models in pure and
noisy conditions.
The third part consists of testing the ability of the most commonly used algorithms
for RTM inversion: Random Forests [34], Artificial Neural Networks [35], and Gaussian
Processes [36]. Each of is representative of the different families of algorithms commonly
used for RTM inversion [26]. The final step consists of evaluating the resilience of the above
models to increasing levels of noise. Before describing each of the steps in more detail, we
first describe the selected RTM, i.e., PROSAIL, [37], and the selected algorithms for RTM
inversion.
2.2. PROSAIL
PROSAIL is the most common RTM used for biophysical trait retrievals, and the
most used in Hybrid approaches. PROSAIL is actually the coupling of two core models:
PROSPECT [38–40] and SAIL [37]. PROSPECT uses well-established physical principles
of light-matter interaction to simulate how light interacts with a single leaf. SAIL stands
for Scattering by Arbitrarily Inclined Leaves and is based on the plate model initially proposed
by Suits [41]. The basic principle of SAIL is that vegetation canopy can be represented
by a “plate” that interacts with directional incoming radiation. PROSAIL integrates both
models by characterizing the vegetation canopy using SAIL and the leaves by using
PROSPECT [42]. The result of the model is a description of the reflectance of the vegetation
layer as dependent on the relative positions of the sun, sensor, canopy and soil character-
istics. The ranges of parameters used for this research are described in Table 1 and were
intentionally broader than what is commonly found in other publications (e.g., [42]) to
ensure the full range of possible values of these parameters [43].
Remote Sens. 2021, 13, 648 4 of 20
Table 1. Overview of PROSAIL input parameters used. Trait combinations were sampled representatively using a Latin
hypercube sampling approach across the range identified. For the ML sections, Car was fixed at 10.
where p R and p L are the proportion of data for each node, it is the node impurity before
splitting and i(t L ) and i(tR ) are the measures of impurity after splitting. While in CART, a
single decision tree (DT) is used, in RFR multiple (K) DT are trained and fitted to a set of
data x, and the final prediction results from either a majority (classification) or the mean
(regression). Formalizing, RFR is an ensemble of K decision trees DT, fitted to a set of data x:
K
1
RF( x) =
K ∑ DTx (2)
k =1
Furthermore, RFR uses the bootstrap aggregating technique to reduce the correlation
between trees and increases the accuracy and stability of the final model [34]. An advantage
of RFR is its ability to provide an estimate of prediction error without the need for external
data since, in each iteration, data non-contributing for model construction is used for
internal testing (out-of-bag error, obb) and to control model overfitting [34].
Remote Sens. 2021, 13, 648 5 of 20
where w, b represent the weight and bias of each ith neuron, and x, y represent the input
and output of the hidden layer. Furthermore, the output y of each neuron can also be
plugged into a function, known as activation functions, which can be used to enforce a
specific behavior to each neuron/layer. Each set of interconnected layers of neurons, except
for the input and output sets, are known as hidden layers. The procedure for training a
supervised ANN consists of minimizing the difference of the current model prediction
against the desired outcome by propagating small changes in w and b of all neurons on all
layers in each iteration (epoch).
For this research, we used two different neural networks: a densely connected and
feed-forward ANN with back-propagation training. This is one of the most used archi-
tectures and is used also by the ESA SNAP toolbox [46]. The second architecture is a
feed-forward, Multi-task Neural Network (MTN) with back-propagation training. This
second type of architecture is increasingly more common thanks to the possibility of
partitioning the learning tasks [45].
where x represents sets of observed points, and k a prior covariance function enforced
for designing the covariance matrix. Iσy2 represents noise being added along the diagonal
of the covariance function. This formulation implies that f ( x ) is a linear combination of
underlying sets of Gaussian Processes and predictions on any given location are interpola-
tions of the observed data. The covariance functions are known as kernels and provide the
expected correlation between different observations. The kernel functions tested in this
research were: radial basis function (RBF), rational Quadratic, matérn, and the dot product.
The main benefits of GPR lie in its versatility [47] and the ability to produce good
results even with a limited number of training data [20]. Furthermore, GPR can provide an
empirical estimate of the uncertainty of the prediction given the evidence x and the joint
distribution defined by the covariance matrix k. The caveat is that they are computationally
expensive when using a large data set [15], and that the model requires storing the entire
covariance matrix information which increases the memory load as the covariance matrix
grows quadratically (n2 ) as a function of n data points.
Remote Sens. 2021, 13, 648 6 of 20
Table 2. Overview of the hyperparameters tested during the optimization. The final set of parameters used are given in
Tables 4 and 5.
∑2500
i =400 wi · Ri
RS2 = (5)
∑2500
i =400 wi
where for each independent RS2 band, the weights (w) were obtained from the spectral
responses provided by ESA for S2A [51], and the Ri represents the reflectance values
obtained from PROSAIL for each band at wavelength i. Finally, only the band products
provided at 10–20 m in the final level 2A product were used: i.e., bands 2, 3, 4, 5, 6, 7, 8A,
11 and 12.
To evaluate the performance of the models, we opted to use the Mean Absolute
Percentage Error (MAPE) because it provides an intuitive interpretation and is easily
comparable between all traits. This is important for this case since the different biophysical
variables have very different ranges of values (see Table 1), and this way we ensure a
proper and intuitive interpretation of the results across the board. MAPE is calculated by:
n
1 Re f i − Predi
MAPE =
n ∑ Re f i
(6)
i =1
where Re f and Pred are the reference and predicted data, respectively, and n the number
of data points used for validation. In the absence of noise, the next step was to partition
the data into the three sets needed for cross-validation: 10% of the total data was left out
Remote Sens. 2021, 13, 648 8 of 20
for estimating the validation error; the remaining 90% of the total data was used for both
the algorithm training and to estimate the training error through a 10 k-fold leave-one-out
approach. All tested algorithms were trained under identical conditions and using the
same training data in each run. To identify how many samples are needed for inverting
PROSAIL RTM, we incremented from a minimum of 50 samples to 4000 samples by steps
of 250. This experiment allowed us to identify both the best algorithm for PROSAIL RTM
inversion as well as their data needs.
Combined : Rnoise (λ) = Rsim (λ)[1 + N (0, 2σ (λ)] + N (0, σ (λ)) (7)
Inverted : Rnoise (λ) = 1 − (1 − Rsim (λ))[1 + N (0, 2σ(λ)] + N (0, σ (λ)) (8)
where Rnoise , Rsim represent the noisy signal and pure simulated spectra, and λ the wave-
length. While both models add noise as a function of wavelength, the Inverted model will
affect more bands on the visible and SWIR regions of the spectra, while the Combined
model will affect more the IR region. Using both types of Gaussian models, we tested the
effect of the following increasing levels of noise added to the training data: 1%, 3%, 5%, 10%,
15%, 20%, 25%, 30%, 40%, and 50%. This experiment aims to explore the pervasiveness of
noisy data in the performance of the ML RTM inversion.
Biophysical trait
table
90%
Spectral K-fold
convolution procedure
10%
Noise model
Holdout-out Validation
set error
Simulated S2
data
Figure 2. Flowchart representing the training and validating procedure of the experiments. All
algorithms were trained with the same data to exclude any random effect. The noise is only added in
the second part of the experiment when testing its effect on the error.
3. Results
3.1. Sensitivity Analysis
The Spearman rank correlation coefficients are shown in Table 3, while the scatterplots
are shown in Figure 3. Many significant ρ were found between multiple pairs of bands and
traits. These relationships are, however, not linear as seen in Figure 3. Cab was shown to
have a negative correlation with all visible bands and had a negligible correlation with the
remaining biophysical parameters (Table 3). Somewhat inversely, Cw is strongly negatively
correlated with the two bands in the longest wavelengths (B11 and B12). Cm and LAI are
both strongly correlated with the NIR and SWIR bands, while Cm has a positive correlation
(Table 3). It is also seen that LAI has a negative correlation, indicating a high potential for
interaction in these bands (Table 3). Finally, Car is only significantly correlated with the
Remote Sens. 2021, 13, 648 9 of 20
blue band. Cw was found to be almost exclusively associated with B11 and B12, which
indicates that there might be also some degree of correlation between Cw and Cm (Figure 3
and Table 3).
Table 3. Summary of Spearman’s ρ correlation coefficients for different biophysical variables and S2
bands. The significance levels are represented as: p ≤ 0.001 (***); p ≤ 0.01 (**); p ≤ 0.05 (*); p ≤ 0.1 (.);
p > 0.1: no symbol.
0.04
0.05 0.2 0.5 0.6 0.4
0.10 0.6 0.20
0.04 0.4 0.5 0.5
Cw 0.03 0.3 0.15
0.06 0.03 0.1 0.3 0.4 0.4
0.3 0.3 0.10
0.02 0.2 0.2
0.02 0.02
0.005 0.010 0.005 0.010 0.005 0.010 0.005 0.010 0.005 0.010 0.005 0.010 0.005 0.010 0.005 0.010 0.005 0.010
0.04
0.05 0.2 0.5 0.6 0.4
0.10 0.6 0.20
0.04 0.4 0.5 0.5
Cm 0.03 0.3 0.15
0.06 0.03 0.1 0.3 0.4 0.4
0.3 0.3 0.10
0.02 0.2 0.2
0.02 0.02
0.005 0.010 0.005 0.010 0.005 0.010 0.005 0.010 0.005 0.010 0.005 0.010 0.005 0.010 0.005 0.010 0.005 0.010
0.04
0.05 0.2 0.5
0.10 0.6 0.6 0.4 0.20
0.04 0.4
LAI 0.03 0.5 0.5
0.3 0.15
0.06 0.03 0.1 0.3 0.4 0.4
0.02 0.3 0.3 0.10
0.02 0.2 0.2
0.02
5 10 5 10 5 10 5 10 5 10 5 10 5 10 5 10 5 10
0.04
0.05 0.2 0.5 0.6 0.4
0.10 0.6 0.20
0.04 0.4 0.5 0.5
Car 0.03
0.06 0.4 0.3 0.15
0.03 0.1 0.3 0.4
0.3 0.3 0.10
0.02 0.2 0.2
0.02 0.02
30 60 30 60 30 60 30 60 30 60 30 60 30 60 30 60 30 60
Figure 3. Scatterplot between each trait and band. To improve visualization, only a fraction of 100 samples were used in
this plot.
We found that for around 2000 samples the variation on the estimate of GSA stabilized;
therefore, we used this number of samples to analyse the sensitivity (Figure 4). The final
GSA estimates (Figure 5) show that LAI was most associated with the three bands on
Remote Sens. 2021, 13, 648 10 of 20
the red-edge region B6, B7 and B8A, and most of the sensitivity on each band was of
first-order. Cw was strongly dependent on bands B11 and B12, with some interaction effects
occurring on B12. Cm was shown to be sensitive to variations in multiple bands from the
infrared region (B06, B07 and B8A) to the short-wave infrared region (B11 and B12). Some
interaction between Cm and Cw can be observed in B12. Car was found to be exclusively
dependent on the blue band (B2), which is of concern due to the high degree of interaction
with Cab . This, combined with the lack of variation in reflectance (see Figure A2), indicates
that Car cannot be detected by S2 data simulated by PROSAIL; therefore, we excluded it
from the remainder of the analysis.
●
●
● ● ● ●
0.75 ●●
●● ●
●● ● ● ● ● ● ● ●
●
● ● ● ● ● ●
● ● ● ●
0.50 ●
● ● ● ●
● ●
● ●
● ● ● ●
● ● ●
● ● ● ● ● ●
●
●
0.25 ● ● ● ●
●● ● ● ● ●
●● ● ● ● ●
● ● ● ● ●
● ●
● ● ● ● ●●
Sensitivity index
●
● ● ● ●
●● ● ● ● ● ●
0.00 ●● ●
●● ● ●
● ●
● ●
● ● ●
●● ● ● ● ●
●● ● ● ● ● ●
●●
● ● ●
● ●
● ●
● ●● ● ● ● ● ●● ● ● ● ● ● ●
●●
● ● ● ● ●● ● ● ● ● ●● ● ● ● ●
●
●
● ● ● ●
●
●
0.75 ●● ● ● ● ● ● ●
●
● ● ●
●
●
● ● ● ●
● ● ● ● ●
●
● ●
● ●
● ●
● ●
●● ● ● ●
0.50 ●
●
● ● ●
● ● ● ● ●
●● ● ● ● ●
●● ● ● ● ●
● ● ● ●
0.25 ● ● ● ●
●●
●
●
● ●
● ● ● ●
● ●● ● ●
●● ● ● ● ● ● ●
● ● ● ● ●● ●
● ●
● ●
● ●
● ● ● ● ● ● ● ● ● ● ● ● ● ●
● ● ● ● ● ● ● ● ●● ● ● ●
0.00 ●
● ● ● ●
● ●●
● ●
● ●
● ●
● ●
● ● ●
● ● ● ●
● ●
● ●● ● ● ● ● ● ● ●● ● ●
● ● ● ●● ●
0 1000 2000 3000 0 1000 2000 3000 0 1000 2000 3000 0 1000 2000 3000 0 1000 2000 3000 0 1000 2000 3000 0 1000 2000 3000 0 1000 2000 3000 0 1000 2000 3000
Samples
● Cab ● Car ● Cm ● Cw ● LAI
LAI
Cw
Biophysical traits
Cm
Car
Cab
0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00
Sensitivity index
B02 B04 B06 B8A B12
B03 B05 B07 B11
Figure 5. The sensitivity index as a function of each trait. The first-order sensitivity is the sensitivity associated with each
trait while the total order is the sum of the first order with the interactions.
shows the task-specific parameters for the MTN. For the RFR, 850 trees were selected with a
negligible minimum number of sample being used for node or leaf splitting. For GPR, it was
found that 80 optimization procedures were necessary, using a rational quadratic kernel.
Table 4. Summary of selected hyperparameters. * on the multi-task network are related to the shared
layers; task-specific parameters are given in Table 5.
In the case of the neural networks, for all cases, the maximum number of hidden layers
was selected and with a high number of neurons (Tables 4 and 5). Non-linear activation
functions were selected with different functions for the ANN and MTN (shared layers).
For the MTN, less complexity was found on the layers associated with Cab and Cw , while
more complexity was needed for Cm , and especially for LAI, as shown by the need for
more neurons in each hidden layer (Tables 4 and 5). This is of particular interest since it is
well established that reflectance often saturates at higher levels of LAI (Tables 4 and 5).
of approximately 2%. The MTN converged faster and showed a significant improvement
in accuracy in all biophysical variables by achieving approximately 2% error in all target
traits. GPR was by far the best performing model with a negligible MAPE in all biophysical
variables, even with fewer than 1000 samples.
60
40
20
0
0 1000 2000 3000 4000 0 1000 2000 3000 4000 0 1000 2000 3000 4000 0 1000 2000 3000 4000
Nr of samples
ANN Training error GPR Training error MTN Training error RFR Training error
ANN Validation error GPR Validation error MTN Validation error RFR Validation error
Figure 6. Model performance on pure PROSAIL RTM simulation. While each model has its own color, the darker colors
represent the training error. The data used for these plots is accessible in the Github repository provided at the end of
this manuscript.
0
0 1000 2000 3000 4000 0 1000 2000 3000 4000 0 1000 2000 3000 4000 0 1000 2000 3000 4000
Nr of samples
ANN Training error GPR Training error MTN Training error RFR Training error
ANN Validation error GPR Validation error MTN Validation error RFR Validation error
While most models produced errors below 2.5% on Cab , the RFR was the worst overall
model by being unable to properly predict Cm or Cw , and not being able to predict LAI
with less than 10% error (Figure 7). The GSA analysis showed that these three biophysical
variables are likely correlated and interacting (see Table 3 and Figure 5) and thus RFR
was not able to address this effect. MTN outperformed ANN both in terms of the MAPE
error as well as in terms of the number of samples necessary for accurate predictions.
Remote Sens. 2021, 13, 648 13 of 20
Although the actual MAPE error is small for both ANN and MTN, it was also several
orders of magnitude larger than the MAPE error of GPR. For the remainder of the analysis,
2000 samples were used given that the error margins did not improve significantly with a
larger sample size (Figure 7).
Cab Cm Cw LAI
100
Mean absolute percentage error (%)
Combined
50
100
Inverted
50
0
0 10 20 30 40 50 0 10 20 30 40 50 0 10 20 30 40 50 0 10 20 30 40 50
Noise level (%)
ANN Training error GPR Training error MTN Training error RFR Training error
ANN Validation error GPR Validation error MTN Validation error RFR Validation error
Figure 8. Effect of noise on model performance. While each model has its own color, the darker colors represent the training
error. The data used for these plots is accessible in the Github repository provided at the end of this manuscript.
Cab Cm Cw LAI
50
40
Mean absolute percentage error (%)
Combined
30
20
10
0
50
40
Inverted
30
20
10
0
0.0 2.5 5.0 7.5 10.0 0.0 2.5 5.0 7.5 10.0 0.0 2.5 5.0 7.5 10.0 0.0 2.5 5.0 7.5 10.0
Noise level (%)
ANN Training error GPR Training error MTN Training error RFR Training error
ANN Validation error GPR Validation error MTN Validation error RFR Validation error
Inverted noise affected the prediction accuracy more than Combined noise. This effect
is particularly evident for LAI, Cm and Cw , even at low levels of noise (Figure 9). For
low values of added noise, the neural networks were the best performing model for most
traits/noise models, even if the absolute error increased significantly. Furthermore, MTN
performed better than ANN and showed more resilience to noise.
4. Discussion
The results of our sensitivity analysis provided clear insights into which biophysical
variables can be measured using S2 data (see Figure 5 and Table 3), and furthermore,
provided insights into the effects of their interaction with spectral responses. Our results
on the band correlation (Table 3) and sensitivity analysis (Figure 5) are in line with previous
results [52], and in particular, when compared with recent results by Brede et al. [27] also
on S2 data, and Gu et al. [49] on Landsat TM data. Similarly, both studies found that Cab
was associated with the visible range in Landsat TM bands, and that Cm , Cw and LAI was
associated with NIR and SWIR bands (see Figure 5). The GSA analysis also shows that Cm ,
Cw and LAI have a degree of interaction between, them which has also been experienced
in previous research [15,52,53]. These interactions can have serious implications for the
accuracy of the retrievals, as specific traits can be more easily inverted due to a higher
number of dedicated bands (e.g., Cab ), while others are more vulnerable to interactions (e.g.,
Cm and Cw ). On top of that, from a biological perspective, it is also likely that some traits
vary together (e.g., Cab and LAI). The main difference for S2 is on which bands are sensitive
to which biophysical variables: in our case, the red-edge (5 and 6) and NIR (8A) were most
strongly associated with LAI, which can be indicative of this sensor being more suited for
LAI estimates but also possessing a higher potential for correlations and interactions. A
significant output of this analysis is that Car is hardly measurable by S2 because only band
B2 showed sensitivity with a high degree of interaction with Cab (Figure 5), and almost no
variation is detectable (see Figure A2).
The use of the Tree-structured Parzen Estimator showed that deeper and more complex
architectures yield better results. This is notable when comparing the ANN with MTN and
observing that a different number of neurons and activation functions were selected as
a function of the objective trait (see Tables 4 and 5). In the case of MTN, and consistent
with sensitivity analysis (Figure 5), more complex architectures were selected by the op-
timization procedure for the traits with higher degrees of interactions (see Figure 5 and
Table 5). This is also in line with other results in other fields of ML where it has been
shown that automatic procedures can greatly improve the model performance and infer
specific domain knowledge [49]. Together, these results can have significant implications
for the commonly used architectures, which are often simpler (e.g., ESA SNAP) and sug-
gest that more complex architectures can improve the retrieval of biophysical variables.
Furthermore, encoding hyperparameter tuning approaches into retrieval procedures can
increase the model versatility by adapting it to specific location or trait-spaces. In addition,
although these adaptations would be hindered by the high computational cost of hyper-
parameter tuning [54], recent developments in the field of AutoML are addressing this
through the concept of meta-learning [55] and multi-layer stacking [56] which minimize
the computational costs for finding suitable models.
Not all ML algorithms were found to be equally capable of hybrid inversion of pure
S2 data (Figures 6 and 7). As with most recent research, [10,20,57] found the Gaussian
Processes produced the best results for pure signals by significant levels of magnitude. This
is of particular importance when considering that the use of Gaussian Processes for this task
is less common than the use of other methods such as ANN [15] or RFR [17]. While in our
case, we found that RFR is unfit for hybrid inversion of S2 data, Campos-Taberner et al. [19]
successfully mapped biophysical variables using this algorithm on MODIS data, implying
that these results are dependent also on the sensors used - and intersecting with the
concept of meta-learning [55], model pipelines trained on sensors with different bands can
potentially provide pre-set model hyperparameters that are transferable between sensors.
Remote Sens. 2021, 13, 648 15 of 20
Model performance was severely affected even at low levels of noise (Figures 8 and 9).
This is of particular concern given that every data has noise, and previous research has
shown that the Sen2Cor atmospheric correction algorithm produces between 5% and 10%
noise [27,52]. Moreover, and in contrast to results obtained on pure data, both neural
networks generally performed better than GPR when noise was added, which indicates
that further developments with GPR are needed to be able to deal effectively with noisy
observations. Recently, Svendsen et al. [58] showed how GPR can be extended to more
complex, structures similar deep learning neural networks and is worth exploring in
future research.
As expected, noise did not affect all bands equally [52], which is consistent with the
different band selection and interactions according to the sensitivity analysis (see Figure 5).
In line with this, Cab estimates were shown to be more resilient to noise, which can be
explained by the high number of bands associated with this trait as well as their lack of
interaction. The opposite was observed for LAI, Cm and Cw , which share many of the same
band sensitivities and behaved similarly when noise was added. Nevertheless, the error
dependence on each trait implies that specific care to each trait should be given when using
a hybrid approach for RTM inversion in S2. Two potential approaches to minimize the trait
specific error are (1) to encode more hyperparameter optimization effort towards inverting
the more vulnerable traits (2) to when using neural networks, enforce task-sharing and
task specific layers on these known interactions.
For practical purposes, in the case of global predictions of biophysical variables,
both GPR and the two neural networks (ANN and MTN) have specific advantages and
disadvantages. For example, ANN are models that can be continuously updated without
significant increases in terms memory requirements as new training implies only changing
the weights and biases of the neurons and hidden layers. This implies that it is trivial to
share pre-trained neural networks for the entire globe which can be used as meta-learning
models for warm-starting the hyperparameter tuning optimization procedures tailored
for each location/sensor. On the other hand, GPR requires a very low number of samples
to stabilize its prediction and offers Active Learning opportunities that combine field
measurements with hybrid inversion within the model itself [20,59]. The caveats regarding
GPR are its tendency to overfit if given too much data, its training time scaling cubically in
function of the number samples [47,60], thus increasing its computational load and cost.
Another important aspect regarding GPR is that optimizing this model for tasks implies
either prior knowledge to impose a specific behaviour to the model or complex methods
for structure discovery due to the additive and multiplicative properties of the kernels [47].
A possible solution to address these challenges which has not yet been fully explored
could be to combine Decision Trees or Random Forests with Gaussian Processes [60], by
clustering the data into similar groups and applying GPR on the leaf nodes.
A final concern is that most approaches for hybrid inversion work on the assumption
that there are no known correlations between the biophysical variables. In nature, it is
virtually impossible to observe random combinations of traits as were used for model
training, e.g., a leaf with LAI of 10 but no Cab . The lack of real-world correlations between
biophysical variables likely aggravates the ill-posedness problem and future research on
exploring approaches that can include known correlations between biophysical variables or
exclude unrealistic combinations will likely contribute to improvements without necessarily
losing generalization [61].
Further potential avenues for improving the retrieval of biophysical parameters using
RTM inversion include the ingestion of a priori knowledge of the ecosystem and trait-space
understudy. Such ancillary information can help reduce the range of input parameters
used which in turn limits the solution space and related degree of ill-posedness [62]. In
dealing with ill-posedness, Rivera et al., 2013 [63] has suggested the use of an ensemble
approach which could prove helpful in reducing uncertainties in real-world applications.
Band selection methods and the use of spectral indices (ratio) in addition to spectral bands
to train inversion models can be also be explored as earlier results have shown [10,64].
Remote Sens. 2021, 13, 648 16 of 20
In many cases, RTMs are used as surrogate information given the lack of field data
due to these campaigns being particularly labour intensive and costly. If more data were
available, more data-driven approaches could be implemented, and the requirement of
RTM-based inversions can be minimized. While plant trait database conglomerates such as
TRY [43] or continental-scale observatory networks such as NEON [65] are a vital step in
the right direction, further steps are needed to make use of the unprecedented availability
of EO data and cloud computing solutions that are increasingly available for researchers.
5. Conclusions
This study shows that biophysical variables can be obtained successfully from S2 data
using a hybrid inversion of PROSAIL RTM. Within such inversion, Gaussian Processes
outperform all other models when using pure data. However, with added noise, this is
no longer true, and ANN, MTN, and GPR have similar results, with the neural networks
being slightly better. Moreover, while GPR can be trained quicker and with less data,
neural network architectures offer a lot of opportunities for improvement and less compu-
tational load. Therefore, more complex neural network architectures should be explored
to improve the accuracy and generality of biophysical trait retrieval on global scale. Such
future explorations may also benefit from the integration of ecological knowledge into the
training procedure, e.g., by imposing these relationships on the training data itself and
from automatic hyperparameter tuning based on sensor, location, and objects. Other future
opportunities can arise with the increasing number of publicly available trait databases and
sensors which can allow the development of meta-learning approaches that could improve
both the model performance as well as reduce the computational cost. In combination,
this study provides a major encouragement for using hybrid RTM inversion to retrieve
biophysical variables from satellite data, e.g., for monitoring global change impacts.
Author Contributions: N.C.d.S. was responsible for the conceptualization, implementation, execu-
tion and preparation of this manuscript. M.B. was responsible for the conceptualization, assisted in
the implementation and contributed for the preparation of this manuscript. L.T.H. contributed for
the conceptualization and preparation of the manuscript. P.v.B. contributed for the conceptualization
and preparation of the manuscript. All authors have read and agreed to the published version of
the manuscript.
Funding: This research was partly funded through the Leiden University Data Science Research
Program.
Code Availability: https://github.com/nunocesarsa/RTM_Inversion (accessed on 2 December 2020).
Acknowledgments: The authors would like to acknowledge the three reviewers for their valuable
comments and suggestions that greatly improved the manuscript. The authors would also like to
acknowledge Andrew Sorensen for his review of the English quality of this manuscript.
Conflicts of Interest: The authors declare no conflict of interest.
Remote Sens. 2021, 13, 648 17 of 20
Appendix A
Appendix A.1. Spectral Variability in Response to Trait Variability
Cab Cm
●
● ●
7.5 ●
●
●
●
●
20
●
5.0
Mean absolute percentage error (%)
●
● ●
●
●
●
10
● ●
● ●
●
2.5 ●
●
● ● ●
● ● ● ● ●
●
● ●
● ●
●
●
● ●
●
● ●
● ● ● ●
● ● ●
● ● ●
● ●
● ● ● ●
● ● ●
●
● ●
0
Cw LAI
25
● 30 ●
20 ●
●
●
15 20 ● ●
●
10 ●
●
●
●
● ●
●
●
●
10 ●
●
●
●
● ●
5 ● ●
● ●
●
●
● ● ●
●
● ● ●
●
● ●
● ● ● ●
● ● ● ● ● ●
● ● ● ●
● ● ●
0 0
0 1000 2000 3000 4000 0 1000 2000 3000 4000
Epochs
●
ANN Training ●
ANN Validation ●
MTN Training ●
MTN Validation
Figure A1. Neural networks training and validation error in function of epochs. Error bars represent the standard deviation
of a 5-fold cross-validation procedure.
Appendix A.2. Artificial Neural Networks Mean Absolute Percentage Error in Function of
Training Epochs
Figure A2. With this figure it is possible to visualize the variation in PROSAIL simulated S2-like data reflectance spectra
when stabilizing all traits except for one. This shows that variations in Car are likely not measurable by Sentinel-2.
Remote Sens. 2021, 13, 648 18 of 20
References
1. Pereira, H.M.; Ferrier, S.; Walters, M.; Geller, G.N.; Jongman, R.H.G.; Scholes, R.J.; Bruford, M.W.; Brummitt, N.; Butchart, S.H.M.;
Cardoso, A.C.; et al. Essential Biodiversity Variables. Science 2013, 339, 277–278. [CrossRef] [PubMed]
2. Jetz, W.; McGeoch, M.A.; Guralnick, R.; Ferrier, S.; Beck, J.; Costello, M.J.; Fernandez, M.; Geller, G.N.; Keil, P.; Merow, C.; et al.
Essential biodiversity variables for mapping and monitoring species populations. Nat. Ecol. Evol. 2019, 3, 539–551. [CrossRef]
[PubMed]
3. Vihervaara, P.; Auvinen, A.P.; Mononen, L.; Törmä, M.; Ahlroth, P.; Anttila, S.; Böttcher, K.; Forsius, M.; Heino, J.; Heliölä, J.; et al.
How Essential Biodiversity Variables and remote sensing can help national biodiversity monitoring. Glob. Ecol. Conserv. 2017,
10, 43–59. [CrossRef]
4. Pettorelli, N.; Schulte to Bühne, H.; Tulloch, A.; Dubois, G.; Macinnis-Ng, C.; Queirós, A.M.; Keith, D.A.; Wegmann, M.; Schrodt, F.;
Stellmes, M.; et al. Satellite remote sensing of ecosystem functions: Opportunities, challenges and way forward. Remote Sens.
Ecol. Conserv. 2017, 4, 71–93. [CrossRef]
5. Cavender-Bares, J.; Gamon, J.A.; Townsend, P.A. (Eds.) Remote Sensing of Plant Biodiversity; Springer: Berlin/Heidelberg, Germany,
2020. [CrossRef]
6. Ghamisi, P.; Gloaguen, R.; Atkinson, P.M.; Benediktsson, J.A.; Rasti, B.; Yokoya, N.; Wang, Q.; Hofle, B.; Bruzzone, L.;
Bovolo, F.; et al. Multisource and Multitemporal Data Fusion in Remote Sensing: A Comprehensive Review of the State
of the Art. IEEE Geosci. Remote Sens. Mag. 2019, 7, 6–39. [CrossRef]
7. Nock, C.A.; Vogt, R.J.; Beisner, B.E. Functional Traits. eLS 2016. [CrossRef]
8. Asner, G.P. Biophysical and Biochemical Sources of Variability in Canopy Reflectance. Remote Sens. Environ. 1998, 64, 234–253.
[CrossRef]
9. Weiss, M.; Baret, F.; Myneni, R.B.; Pragnère, A.; Knyazikhin, Y. Investigation of a model inversion technique to estimate canopy
biophysical variables from spectral and directional reflectance data. Agronomie 2000, 20, 3–22. [CrossRef]
10. Verrelst, J.; Dethier, S.; Rivera, J.P.; Munoz-Mari, J.; Camps-Valls, G.; Moreno, J. Active Learning Methods for Efficient Hybrid
Biophysical Variable Retrieval. IEEE Geosci. Remote Sens. Lett. 2016, 13, 1012–1016. [CrossRef]
11. Durán, S.M.; Martin, R.E.; Díaz, S.; Maitner, B.S.; Malhi, Y.; Salinas, N.; Shenkin, A.; Silman, M.R.; Wieczynski, D.J.;
Asner, G.P.; et al. Informing trait-based ecology by assessing remotely sensed functional diversity across a broad tropical
temperature gradient. Sci. Adv. 2019, 5, eaaw8114. [CrossRef]
12. Schneider, F.D.; Morsdorf, F.; Schmid, B.; Petchey, O.L.; Hueni, A.; Schimel, D.S.; Schaepman, M.E. Mapping functional diversity
from remotely sensed morphological and physiological forest traits. Nat. Commun. 2017, 8. [CrossRef] [PubMed]
13. Zarco-Tejada, P.; Hornero, A.; Beck, P.; Kattenborn, T.; Kempeneers, P.; Hernández-Clemente, R. Chlorophyll content estimation in
an open-canopy conifer forest with Sentinel-2A and hyperspectral imagery in the context of forest decline. Remote Sens. Environ.
2019, 223, 320–335. [CrossRef] [PubMed]
14. Verrelst, J.; Camps-Valls, G.; Muñoz-Marí, J.; Rivera, J.P.; Veroustraete, F.; Clevers, J.G.; Moreno, J. Optical remote sensing and the
retrieval of terrestrial vegetation bio-geophysical properties—A review. ISPRS J. Photogramm. Remote Sens. 2015, 108, 273–290.
[CrossRef]
15. Verrelst, J.; Malenovský, Z.; der Tol, C.V.; Camps-Valls, G.; Gastellu-Etchegorry, J.P.; Lewis, P.; North, P.; Moreno, J. Quantifying
Vegetation Biophysical Variables from Imaging Spectroscopy Data: A Review on Retrieval Methods. Surv. Geophys. 2018,
40, 589–629. [CrossRef]
16. Staben, G.; Lucieer, A.; Scarth, P. Modelling LiDAR derived tree canopy height from Landsat TM, ETM+ and OLI satellite
imagery—A machine learning approach. Int. J. Appl. Earth Obs. Geoinf. 2018, 73, 666–681. [CrossRef]
17. Moreno-Martínez, Á.; Camps-Valls, G.; Kattge, J.; Robinson, N.; Reichstein, M.; van Bodegom, P.; Kramer, K.; Cornelissen, J.H.C.;
Reich, P.; Bahn, M.; et al. A methodology to derive global maps of leaf traits using remote sensing and climate data. Remote Sens.
Environ. 2018, 218, 69–88. [CrossRef]
18. Darvishzadeh, R.; Wang, T.; Skidmore, A.; Vrieling, A.; O’Connor, B.; Gara, T.; Ens, B.; Paganini, M. Analysis of Sentinel-2 and
RapidEye for Retrieval of Leaf Area Index in a Saltmarsh Using a Radiative Transfer Model. Remote Sens. 2019, 11, 671. [CrossRef]
19. Campos-Taberner, M.; Moreno-Martínez, Á.; García-Haro, F.; Camps-Valls, G.; Robinson, N.; Kattge, J.; Running, S. Global
Estimation of Biophysical Variables from Google Earth Engine Platform. Remote Sens. 2018, 10, 1167. [CrossRef]
20. Svendsen, D.H.; Martino, L.; Campos-Taberner, M.; Camps-Valls, G. Joint Gaussian processes for inverse modeling. In Proceedings
of the 2017 IEEE International Geoscience and Remote Sensing Symposium (IGARSS), Fort Worth, TX, USA, 4 December 2017.
[CrossRef]
21. Camps-Valls, G.; Martino, L.; Svendsen, D.H.; Campos-Taberner, M.; Muñoz-Marí, J.; Laparra, V.; Luengo, D.; García-Haro, F.J.
Physics-aware Gaussian processes in remote sensing. Appl. Soft Comput. 2018, 68, 69–82. [CrossRef]
22. Tuia, D.; Verrelst, J.; Alonso, L.; Perez-Cruz, F.; Camps-Valls, G. Multioutput Support Vector Regression for Remote Sensing
Biophysical Parameter Estimation. IEEE Geosci. Remote Sens. Lett. 2011, 8, 804–808. [CrossRef]
23. Annala, L.; Honkavaara, E.; Tuominen, S.; Pölönen, I. Chlorophyll Concentration Retrieval by Training Convolutional Neural
Network for Stochastic Model of Leaf Optical Properties (SLOP) Inversion. Remote Sens. 2020, 12, 283. [CrossRef]
24. Zurita-Milla, R.; Laurent, V.; van Gijsel, J. Visualizing the ill-posedness of the inversion of a canopy radiative transfer model: A
case study for Sentinel-2. Int. J. Appl. Earth Obs. Geoinf. 2015, 43, 7–18. [CrossRef]
Remote Sens. 2021, 13, 648 19 of 20
25. Locherer, M.; Hank, T.; Danner, M.; Mauser, W. Retrieval of Seasonal Leaf Area Index from Simulated EnMAP Data through
Optimized LUT-Based Inversion of the PROSAIL Model. Remote Sens. 2015, 7, 10321–10346. [CrossRef]
26. Verrelst, J.; Rivera, J.P.; Veroustraete, F.; Muñoz-Marí, J.; Clevers, J.G.; Camps-Valls, G.; Moreno, J. Experimental Sentinel-2 LAI
estimation using parametric, non-parametric and physical retrieval methods—A comparison. ISPRS J. Photogramm. Remote Sens.
2015, 108, 260–272. [CrossRef]
27. Brede, B.; Verrelst, J.; Gastellu-Etchegorry, J.P.; Clevers, J.G.P.W.; Goudzwaard, L.; den Ouden, J.; Verbesselt, J.; Herold, M.
Assessment of Workflow Feature Selection on Forest LAI Prediction with Sentinel-2A MSI, Landsat 7 ETM+ and Landsat 8 OLI.
Remote Sens. 2020, 12, 915. [CrossRef]
28. Berger, K.; Caicedo, J.P.R.; Martino, L.; Wocher, M.; Hank, T.; Verrelst, J. A Survey of Active Learning for Quantifying Vegetation
Traits from Terrestrial Earth Observation Data. Remote Sens. 2021, 13, 287. [CrossRef]
29. Upreti, D.; Huang, W.; Kong, W.; Pascucci, S.; Pignatti, S.; Zhou, X.; Ye, H.; Casa, R. A Comparison of Hybrid Machine Learning
Algorithms for the Retrieval of Wheat Biophysical Variables from Sentinel-2. Remote Sens. 2019, 11, 481. [CrossRef]
30. Xie, Q.; Dash, J.; Huete, A.; Jiang, A.; Yin, G.; Ding, Y.; Peng, D.; Hall, C.C.; Brown, L.; Shi, Y.; Ye, H.; Dong, Y.; Huang, W.
Retrieval of crop biophysical parameters from Sentinel-2 remote sensing imagery. Int. J. Appl. Earth Obs. Geoinf. 2019, 80, 187–195.
[CrossRef]
31. Saltelli, A.; Tarantola, S.; Chan, K.P.S. A Quantitative Model-Independent Method for Global Sensitivity Analysis of Model
Output. Technometrics 1999, 41, 39–56. [CrossRef]
32. Bergstra, J.; Yamins, D.; Cox, D.D. Making a Science of Model Search: Hyperparameter Optimization in Hundreds of Dimensions
for Vision Architectures. In Proceedings of the 30th International Conference, International Conference on Machine Learning,
Atlanta, GA, USA, 17–19 June 2013; p. I-115–I-123.
33. Bergstra, J.; Komer, B.; Eliasmith, C.; Yamins, D.; Cox, D.D. Hyperopt: A Python library for model selection and hyperparameter
optimization. Comput. Sci. Discov. 2015, 8, 014008. [CrossRef]
34. Breiman, L. Random Forests. Mach. Learn. 2001, 45, 5–32. [CrossRef]
35. McCulloch, W.S.; Pitts, W. A logical calculus of the ideas immanent in nervous activity. Bull. Math. Biophys. 1943, 5, 115–133.
[CrossRef]
36. Rasmussen, C.E.; Williams, C.K.I. Gaussian Processes for Machine Learning (Adaptive Computation and Machine Learning); The MIT
Press: Cambridge, MA, USA, 2005.
37. Verhoef, W. Light scattering by leaf layers with application to canopy reflectance modeling: The SAIL model. Remote Sens.
Environ. 1984, 16, 125–141. [CrossRef]
38. Jacquemoud, S.; Baret, F. PROSPECT: A model of leaf optical properties spectra. Remote Sens. Environ. 1990, 34, 75–91. [CrossRef]
39. Jacquemoud, S.; Verhoef, W.; Baret, F.; Bacour, C.; Zarco-Tejada, P.J.; Asner, G.P.; François, C.; Ustin, S.L. PROSPECT + SAIL
models: A review of use for vegetation characterization. Remote Sens. Environ. 2009, 113, S56–S66. [CrossRef]
40. Féret, J.B.; Gitelson, A.; Noble, S.; Jacquemoud, S. PROSPECT-D: Towards modeling leaf optical properties through a complete
lifecycle. Remote Sens. Environ. 2017, 193, 204–215. [CrossRef]
41. SUITS, G. The calculation of the directional reflectance of a vegetative canopy. Remote Sens. Environ. 1971, 2, 117–125. [CrossRef]
42. Berger, K.; Atzberger, C.; Danner, M.; D’Urso, G.; Mauser, W.; Vuolo, F.; Hank, T. Evaluation of the PROSAIL Model Capabilities
for Future Hyperspectral Model Environments: A Review Study. Remote Sens. 2018, 10, 85. [CrossRef]
43. Kattge, J.; Diaz, S.; Lavorel, S.; Prentice, I.C.; Leadley, P.; Bönisch, G.; Garnier, E.; Westoby, M.; Reich, P.B.; Wright, I.J.; et al.
TRY—A global database of plant traits. Glob. Chang. Biol. 2011, 17, 2905–2935. [CrossRef]
44. LeCun, Y.; Bengio, Y.; Hinton, G. Deep learning. Nature 2015, 521, 436–444. [CrossRef]
45. Zhang, Y.; Yang, Q. An overview of multi-task learning. Natl. Sci. Rev. 2017, 5, 30–43. [CrossRef]
46. Weiss, M.; Baret, F. S2ToolBox Level 2 Products: LAI, FAPAR, FCOVER; Institut National de la Recherche Agronomique (INRA):
Avignon, France, 2016.
47. Duvenaud, D.; Lloyd, J.R.; Grosse, R.; Tenenbaum, J.B.; Ghahramani, Z. Structure Discovery in Nonparametric Regression
through Compositional Kernel Search. In Proceedings of the 30th International Conference, International Conference on Machine
Learning, Atlanta, GA, USA, 17–19 June 2013; pp. 1166–1174.
48. Sobol, I. Sensitivity Analysis or Nonlinear Mathematical Models. Math. Model. Comput. Exp. 1993, 4, 407–414.
49. Gu, C.; Du, H.; Mao, F.; Han, N.; Zhou, G.; Xu, X.; Sun, S.; Gao, G. Global sensitivity analysis of PROSAIL model parameters
when simulating Moso bamboo forest canopy reflectance. Int. J. Remote Sens. 2016, 37, 5270–5286. [CrossRef]
50. Wang, S.; Gao, W.; Ming, J.; Li, L.; Xu, D.; Liu, S.; Lu, J. A TPE based inversion of PROSAIL for estimating canopy biophysical and
biochemical variables of oilseed rape. Comput. Electron. Agric. 2018, 152, 350–362. [CrossRef]
51. The European Space Agency. Sentinel-2 Spectral Response Functions (S2-SRF). Available online: https://earth.esa.int/web/
sentinel/user-guides/sentinel-2-msi/document-library (accessed on 1 January 2020).
52. Verrelst, J.; Vicent, J.; Rivera-Caicedo, J.P.; Lumbierres, M.; Morcillo-Pallarés, P.; Moreno, J. Global Sensitivity Analysis of
Leaf-Canopy-Atmosphere RTMs: Implications for Biophysical Variables Retrieval from Top-of-Atmosphere Radiance Data.
Remote Sens. 2019, 11, 1923. [CrossRef]
53. Pan, H.; Chen, Z.; Ren, J.; Li, H.; Wu, S. Modeling Winter Wheat Leaf Area Index and Canopy Water Content With Three Different
Approaches Using Sentinel-2 Multispectral Instrument Data. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2019, 12, 482–492.
[CrossRef]
Remote Sens. 2021, 13, 648 20 of 20
54. Yang, L.; Shami, A. On hyperparameter optimization of machine learning algorithms: Theory and practice. Neurocomputing 2020,
415, 295–316. [CrossRef]
55. Feurer, M.; Eggensperger, K.; Falkner, S.; Lindauer, M.; Hutter, F. Auto-Sklearn 2.0: The Next Generation. arXiv 2020,
arXiv:2007.04074.
56. Erickson, N.; Mueller, J.; Shirkov, A.; Zhang, H.; Larroy, P.; Li, M.; Smola, A. AutoGluon-Tabular: Robust and Accurate AutoML
for Structured Data. arXiv 2020, arXiv:2003.06505.
57. Lazaro-Gredilla, M.; Titsias, M.K.; Verrelst, J.; Camps-Valls, G. Retrieval of Biophysical Parameters With Heteroscedastic Gaussian
Processes. IEEE Geosci. Remote Sens. Lett. 2014, 11, 838–842. [CrossRef]
58. Svendsen, D.H.; Morales-Álvarez, P.; Ruescas, A.B.; Molina, R.; Camps-Valls, G. Deep Gaussian processes for biogeophysical
parameter retrieval and model inversion. ISPRS J. Photogramm. Remote Sens. 2020, 166, 68–81. [CrossRef] [PubMed]
59. Gewali, U.B.; Monteiro, S.T.; Saber, E. Gaussian Processes for Vegetation Parameter Estimation from Hyperspectral Data with
Limited Ground Truth. Remote Sens. 2019, 11, 1614. [CrossRef]
60. Fröhlich, B.; Rodner, E.; Kemmler, M.; Denzler, J. Efficient Gaussian process classification using random decision forests. Pattern
Recognit. Image Anal. 2011, 21, 184–187. [CrossRef]
61. Camps-Valls, G.; Svendsen, D.H.; Cortés-Andrés, J.; Moreno-Martínez, Á; Pérez-Suay, A.; Adsuara, J.; Martín, I.; Piles, M.;
Muñoz-Marí, J.; Martino, L. Living in the Physics and Machine Learning Interplay for Earth Observation. arXiv 2020,
arXiv:2010.09031.
62. Combal, B.; Baret, F.; Weiss, M. Improving canopy variables estimation from remote sensing data by exploiting ancillary
information. Case study on sugar beet canopies. Agronomie 2002, 22, 205–215. [CrossRef]
63. Rivera, J.; Verrelst, J.; Leonenko, G.; Moreno, J. Multiple Cost Functions and Regularization Options for Improved Retrieval of
Leaf Chlorophyll Content and LAI through Inversion of the PROSAIL Model. Remote Sens. 2013, 5, 3280–3304. [CrossRef]
64. Richter, K.; Hank, T.B.; Vuolo, F.; Mauser, W.; D’Urso, G. Optimal Exploitation of the Sentinel-2 Spectral Capabilities for Crop
Leaf Area Index Mapping. Remote Sens. 2012, 4, 561–582. [CrossRef]
65. National Science Foundation. National Ecological Observatory Network. Available online: https://data.neonscience.org/home
(accessed on 1 January 2020).