on panel was
ce.
one bud with
re clipped. For
> pots together
i leaves were
mical analysis
or each sample
> need for wet
determine the
re steamed for
tivity causing
e drying in an
und using an
mined by the
etry at 540nm
m, 1800-1970
' noise due to
om the data.
I observations
(Araujo et al.,
r regression
the features of
egressions. lt
atent variables
the size of
of dependent
ng found with
1989).
) establish the
ical contents
in the sample
and test data
ate the partial
mance of the
lictions of the
of the PLSR
mination (R2)
SEP, Equation
ns on test data
(1)
:rved value of
on the model
2.6 A hybrid approach (nonlinear regression approach)
For one tea variety growing in the greenhouse, the neural
network approach were applied to build the spectral-chemical
relationship using nonlinear regression way. A one hidden layer
feed-forward, error-back propagation artificial neural network
were adopted in this research, for this algorithm has been
frequently and successfully used in previous studies (Skidmore
et al. 1997). To find the optimal number of nodes in the hidden !
layer, we investigated the training and test accuracies using
different number of neurons (1-20) in the network (the
maximum number was designed no more than 20 to keep the
model parsimony and save the calculation time). Levenberg-
Marquardt optimization method was used to train the networks
in which the parameters of networks were adjusted adaptively
(Lera and Pinzolas, 2002; More, 1978) and an earlier stop
technique was applied in this study to avoid overtraining (Lin
and Chen, 2004).
Before running the neural network model, an effective variable
selection method named successive projections algorithm was
applied to spectral data (350-2500 nm) after pre-processing. It is
a forward selection approach. The purpose of this algorithm is
to select wavebands containing minimally redundant
information, so that collinearity problems caused by
hyperspectral data can be minimized.
The available data (64 samples) were randomly divided into
three groups: the training dataset (n = 32, 50% of the sample),
the validation dataset (n = 16, 25% of the sample) and the test
dataset (n = 16, 25% of the sample). The performance of the
ANN model was evaluated by the root mean square error of
prediction (RMSEP) between the predicted and measured
concentration based on the test dataset (Mutanga et al., 2004).
To speed up the training process of neural networks models, the
input data of chemical concentrations were normalized between
0 and 1(Mutanga et al., 2004).
3. RESULTS
Table 2 shows the measured concentrations for total tea
polyphenols by varieties and soil treatments. All values are
reported on a dry-matter basis. The range of the chemical data
accords with the values which have been previously reported.
For tea polyphenols measured for greenhouse experiment , the
combination of higher level of nitrogen, phosphorus and
potassium resulted in the maximum concentration and vice
Versa.
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B8, 2012
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia
No. Mean Minimum Maximum
; (mgg”)_ (mag) (mag ”)
Six varieties
A 8 176.20 167.46 181.81
B 8 180.67 472.21 195.35
C 8 186.31 173.48 199.36
D 8 208.72 203.90 213.77
E 8 270.92 260.63 288.99
E 8 218.87 201.16 235.94
All 48 206.95 167.47 288.99
Soil treatments
à 8 126.28 118.40 134.62
b 8 132.36 126.57 138.42
8 132.87 125.77 137.27
d 8 133.83 129.83 141.35
e 8 132.59 127.32 138.98
f 9 143.88 137.02 149.53
g 8 133.24 12573 141.01
i h 8 145.95 141.89 149.99
All 64 135.13 129.07 141.40
Table 2. Descriptive statistics of the total tea polyphenols
measured in the laboratory
For different tea varieties, using partial least squares regression,
observed versus predicted concentrations of tea polyphenols for
both training (N=30) and test (N=18) data are shown in Figure 2.
The satisfactory accuracy of prediction was obtained at canopy
level : based on the independent data set, total tea polyphenols
were estimated with high 12 values (> 0.8) and low RMSEP
values (RMSEP = 13.68 mg g-1 , RMSEP/mean = 6.63%).
300
* training
* test .
+
47 x=y À
'O» 250
D
E
© fe
o 9
A
à te à “= 2
D - Tr.
Sos x ;
* Set s r? - 0.84
9
RMSEP = 13.68
150
150 200 250 300
Predicted (mg 9°!)
Figure 2. Scatter plots describing the measured and predicted
total tea polyphenols for training and test using canopy spectra
(mean centred). r^ is coefficient of determination between
model predictions and measured chemical concentrations on test
data set, and RMSEP is the root mean square error of test data
prediction.
Figure 3 presents relationships between the predicted and
measured biochemical concentrations using a hybrid of neural
networks and SPA variable selections (SPA-ANN): on test data
set, using the wavebands selected by the successive projections
algorithm, the neural networks with optimal settings yielded
coefficient of determination r2 of 0.82, for the prediction of
total tea polyphenols in the greenhouse experiment, with a root
mean square error of 4.30 mg g-1 (3.0% of the mean).
Figure 4 shows the optimal choice of the number of wavelength
selected by successive projections algorithm. According to the
criterion of the root mean square error of validation, the best
choice of 12 wavebands has been selected for the prediction of
total tea polyphenols. In an order of importance (from most to
least), wavelengths selected by SPA for the prediction of total
tea polyphenols are 2001 nm, 2206 nm, 1424 nm, 1799 nm,
1439 nm, 1426 nm, 689 nm, 1971 nm, 1428 nm, 1435 nm, 1422
nm and 1502 nm.
