K. Miyamoto et al., J. Near Infrared Spectrosc.

6, 267–271 (1998) 267

Classification of High Acid Fruits by Partial Least Squares

K. Miyamoto et al., J. Near Infrared Spectrosc. 6, 267–271 (1998)

Classification of high acid fruits by partial

least squares using the near infrared
transmittance spectra of intact satsuma
mandarins

Kumi Miyamoto
Wakayama Fruit Tree Experiment Station, 751-1 Oki, Kihi, Arida, Wakayama, Japan.

Miyuki Kawauchi and Toshitaka Fukuda

Foundation Saika Technological Institute, 75-2 Kuroda, Wakayama, Japan.

The feasibility of on-line measurement of citric acid content using NIR transmittance spectroscopy was investi-
gated. The citric acid content was determined by titration with NaOH. The second derivative absorption values of
transmittance spectra through the fruit equator were autoscaled by mean-centring and weighing by 1/S.Dev. The
data were analysed by PLS1 using the UNSCRAMBLER software. The spectra of peeled fruits measured by a
NIRSystems Model 6250 were analysed, and the model composed of 12 factors provided the highest accuracy;
R = 0.93, Bias = –0.013% and SEP = 0.146%. In the next step, the citric acid content in intact satsuma mandarins
was regressed by the same method using an on-line instrument. The model composed of five factors explained
68.2% of total Y variation. The accuracy of prediction result was rather low; R = 0.83, Bias = 0.024% and
SEP = 0.147%. But, it was possible to classify non-destructively the high acid fruits with about a 20% error rate.
The absorption bands of citric acid contributed heavily to this PLS model.

Keywords: satsuma mandarin, fruit, citric acid, NIR transmittance spectrum, PLS.

Introduction which contribute much to the taste of the fruit, in sat-

Fruit grading in packing houses is necessary for
quality assurance to consumers. On-line grading ac-
cording to external appearance, such as fruit shape,
peel colour and defects using a RGB camera is estab-
lished in most packing houses. Recently, the control
of internal fruit quality, especially taste, has become
a very important subject of sales promotion for Japa-
nese fruit growers. It is necessary to develop a non-
destructive measurement of sugar and acid content,
suma mandarins.
Non-destructive measurement of sugar content in
satsuma mandarins using NIR transmittance was re-
ported by Kawano et al.1 for the first time. We have
studied the application of this method for practical
use, and developed an instrument for on-line mea-
surement of sugar content in satsuma mandarins in
1994. In the next step, the feasibility of simultaneous
measurement of citric acid content by the same
method was investigated.

© NIR Publications 1998, ISSN 0967-0335

268 Classification of High Acid Fruits by Partial Least Squares

Materials and methods

Peeled fruits
The transmittance spectra through the fruit equa-
tor were measured at 20–24°C by a NIR
spectrophotometer (NIRSystems Model 6250,
USA) to which was attached a fibre optic connecting
to the light source. The energy of the transmitted
light was very weak and it was difficult to detect pre-
cisely at wavelengths longer than 850 nm. Peeled
fruits were used as samples to obtain the transmit-
tance spectra as precisely as possible in the region of
680 nm to 960 nm. 128 samples were selected as the
calibration set. The second derivative transmittance
spectra were computed with a 20 nm segment size Figure 1. Diagram of an on-line instrument developed for
and 0 nm gap size, and the values were autoscaled by NIR transmittance measurement.
mean-centring and weighing by standardisation
(I/S.Dev.). These pretreated spectral data were used
as X variables; the total number was 279 in the re- with a linear array silicon photodiode flat field
gion 710 nm to 930 nm. The citric acid contents, detector. Figure 1 shows a diagram of this measure-
which were determined by titration with NaOH solu- ment method. The wavelength range of the measure-
tion, were autoscaled by the same method mentioned ment was 700–1000 nm at 1 nm intervals.
above and used as Y variables. These data were ana- 689 samples were selected as a calibration set,
lysed by PLS regression using the and the spectral data and the citric acid content
UNSCRAMBLER software (Version 5.5, Camo AS, pretreated as described above. The autoscaled sec-
Norway). The accuracy of the regression equation ond derivative absorption values at 730–930 nm
was evaluated by a test sample set (n = 108). were used as X variables; the total number was 201.
The data were analysed by PLS regression and the
Intact fruits prediction accuracy of the regression equation was
The transmittance spectra of intact satsuma man- evaluated by a test sample set (n = 548).
darin fruits were measured at room temperature The citric acid content of each sample set used for
(5–20°C) by the on-line model instrument that was the analysis is shown in Table 1.
developed for the fruit grading process based on
sugar content in a packing house in 1995. The con- Satsuma mandarin juice
veyer speed was 50 m min–1 and it could measure Transmittance spectra of the juice samples were
6–10 fruits s–1. This instrument uses a multichannel measured by a NIRSystems Model 6250 using a
grating polychromator of the post-dispersive type quartz glass cell with 20 mm pathlength at 20°C.

Table 1. Distribution of citric acid content in sample sets (g dL–1).

Calibration set Prediction set

Fruit type n Range Mean S.Dev. n Range Mean S.Dev.
Peeled fruit 128 0.42–2.33 1.03 0.384 108 0.50–2.23 0.95 0.331
Intact fruit 689 0.56–1.85 1.16 0.290 548 046–1.50 0.95 0.267
K. Miyamoto et al., J. Near Infrared Spectrosc. 6, 267–271 (1998)
Figure 2. Results of PLS regression for citric acid content in peeled satsuma mandarin fruits. (a) Calibration result of 128
samples and (b) prediction result of 108 samples.
The absorption bands of citric acid were investi-
gated by the addition of citric acid to satsuma man-
darin juice filtered through double gauze. The
spectral changes were observed by adding citric
acid: 6.25, 12.5, 25 and 50 g per 100 mL juice. In ad-
dition, the spectral changes of juice by adding su-
crose were observed at the same time, because sugar
is the second major constituent after water.
Results and discussion
Peeled fruits
As the result of PLS1 analysis, the accuracy of the
regression equation was highest when using 12 fac-
tors. These 12 factors explained 99.9% and 85.0% of
the X or Y variation of the calibration set, respec-
tively. The reference versus calculated value plots
for calibration and prediction are shown in Figure 2.
The prediction accuracy was not so high; standard
error of prediction (SEP) was 0.1463% and bias was
–0.0128%. However, it was possible to use this PLS
model for classification of high acid fruits (over
1.1% citric acid content). The ability of classifica-
269
tion was evaluated by the precision rate (no. of ac-
tual high acid fruits in numerator / no. of high acid
fruits judged by the equation × 100) and the recall
rate (no. of high acid fruits judged by the equation in
numerator / no. of actual high acid fruits × 100). In
this case, the precision rate was 81.5% and the recall
rate was 84.6%.
The contribution of each X variable to this PLS
model was interpreted from the loading weight and
the standard regression coefficient. The wave-
lengths in the 890–903 nm region, 917–928 nm re-
gion, 841–864 nm region, 811–836 nm region and
724–738 nm region made positive contributions to
the regression equation according to the citric acid
content (Figure 3). Through observation of juice
samples, the absorption bands of citric acid were re-
cognised in the region of 850 nm and 900 nm. The
loading weights in the 840–865 nm region in factor
1, 865–930 nm region in factor 2 and 880–914 nm
region in factor 3 were large. It was suggested that
the wavelengths in the absorption bands of citric
acid contributed heavily to this PLS model. Figure 4
shows the changes of the difference spectra (each
second derivative spectrum minus a juice spectrum)
270
Figure 3. Standard correlation coefficients of the PLS
model composed of 12 factors for measurement of citric
acid content in peeled satsuma mandarin fruits.
Figure 4. The difference spectra of satsuma mandarin
juice with added citric acid or sucrose. The difference
spectra were calculated by subtracting the second deriva-
tive value of the juice from that of the modified ones.
by adding citric acid or sucrose. The second deriva-
tive absorption value at 900 nm was constant regard-
less of the sucrose content and was varied by the
citric acid content. The standard coefficient of the
equation in the region of 900 nm was especially
large. The same phenomenon has already been re-
ported by Fujiwara et al.2 in 1995, at the wave-
lengths of 1718 nm and 2290 nm, in a study of the
NIR measurement of citric acid content in satsuma
mandarin juice.
Intact fruits
The PLS model composed of five factors showed
the best result, but the accuracy was rather low; SEP
Classification of High Acid Fruits by Partial Least Squares
Figure 5. Prediction result of PLS regression for citric acid
content in intact satsuma mandarins (548 test samples).
Figure 6. Loading weights of the five factors of the equa-
tion for the measurement of citric acid content in intact
satsuma mandarins.
was 0.1467% and bias was 0.0235%. The prediction
result of the test set is shown in Figure 5. It was pos-
sible to classify high acid fruits at the allowable
level for practical use in spite of the fact that the
fruits were being measured very quickly. The preci-
sion rate and the recall rate of selection of high acid
fruits having more than 1.1% citric acid content
were 75.8% and 79.0%, respectively.
The loading weights at each wavelength in the
730–930 nm region of five factors are shown in Fig-
ure 6. The absorption in the region of 765 nm,
860 nm and 910 nm, specifically 890–930 nm, con-
tributed positively to factor 1 according to citric acid
content. These wavelengths are in the absorption
K. Miyamoto et al., J. Near Infrared Spectrosc. 6, 267–271 (1998) 271

bands correlated to citric acid, and factor 1 ex- gested that factors 3, 4 and 5 explained the variation
plained mainly the variation of citric acid content. In due to fruit temperature.
factor 2, the absorption in the region of 820 nm and
910 nm contributed positively to the equation and in
the region of 735 nm, 787 nm and 840 nm contrib-
uted negatively according to citric acid content. The References
absorption values at these latter three wavelengths 1. S. Kawano, T. Fujiwara and M. Iwamoto, J. Ja-
correlated well with optical path length (Miyamoto pan. Soc. Hort. Sci. 62, 465 (1993).
et al.3 1995). It was suggested that factor 2 explained 2. T. Fujiwara and T. Honjo, Nippon Shokuhin
the variation related to optical path length. In the Kogaku Kaishi 42, 109 (1995).
case of factors 3, 4 and 5, the maximum and mini- 3. K. Miyamoto and Y. Kitano, J. Near Infrared
mum peaks of each loading weight plot were recog- Spectrosc. 3, 227 (1995).
nised in the region of 790 nm, 835 nm and 870 nm.
There was high correlation between the absorption Received: 15 September 1997
at these wavelengths and fruit temperature Revised: 23 April 1998
(Miyamoto et al.3 1995). Therefore, it was sug- Accepted: 9 June 1998