biosystems engineering 101 (2008) 417–424
Available at www.sciencedirect.com
journal homepage: www.elsevier.com/locate/issn/15375110
Research Paper: PHdPostharvest Technology
Shape-based methodology for multivariate discrimination among Italian hazelnut cultivars Paolo Menesattia, Corrado Costab,*, Graziella Pagliaa, Federico Pallottinoa, Stefano D’Andreab, Valentina Rimatoria, Jacopo Aguzzic a
CRA-ING Agricultural Engineering Research Unit of the Agriculture Research Council, Via della Pascolare, 16, C.A.P. 00016 Monterotondo, Roma, Italy b ENAMA – Ente Nazionale per la Meccanizzazione Agricola, Via Venafro, 5 – 00159 Roma, Italy c Institut de Cie`ncies del Mar (ICM-CSIC), Passeig Marı´tim de la Barceloneta 37-49, 08003 Barcelona, Spain
article info Cultivar discrimination during on-line quality selection is required by high quality food Article history:
industries. The aim of this work was to evaluate the potential use and efficacy of shape-
Received 11 April 2008
based techniques in order to discriminate among four traditional Italian cultivars (Tonda di
Received in revised form
Giffoni, San Giovanni, Mortarella and Tonda Romana). Tonda di Giffoni and Tonda Romana
18 July 2008
are very similar having a spherical shape, while the other two cultivars are elongated. Color
Accepted 16 September 2008
RGB images of about 400 hazelnuts were analysed with a morphological method based on
Available online 6 November 2008
the elliptic Fourier approximation to closed contours in a two-dimensional plane. This method was applied on the three outlines obtained by the polar, lateral and random plane positioning view of in-shell and unblanched kernel. The coefficients of the harmonic equations were analysed via Partial Least Square Discriminant Analysis (PLSDA) multivariate classification and mean outline for each group was graphically extracted. Results show higher percentage of correct classification for the lateral view (from 77.5% to 98.8% in the independent test). Also the random positioning view, in particular for in-shell kernels between the two rounded cultivars and between the two oblong cultivars, showed good classification results (respectively, 95.1 and 97.6). This preliminary study demonstrates the potential of modern multivariate techniques using shape-based methods on digital images to achieve high efficiency performance in fruit grading and classification. ª 2008 IAgrE. Published by Elsevier Ltd. All rights reserved.
1.
Introduction
According to FAO statistics the main countries producing hazelnut (Corylus avellana, L. 1753) are Turkey, Italy, Spain, USA and Greece. Turkey covers respectively, 70% and 82% of the world’s production and export, with Italy following with nearly 20% of the production and 15% in terms of export (FAO, 2000). The world hazelnut production shows fluctuations depending
on climatic conditions from year to year. In Italy, the most important cultivars are the native ‘‘Tonda Gentile Romana’’ in Lazio region, ‘‘Tonda di Giffoni’’ ‘‘Mortarella’’, ‘‘San Giovanni’’, ‘‘Camponica’’, ‘‘Riccia di Talanico’’, ‘‘Tonda Bianca’’ and ‘‘Tonda Rossa’’ in Campania, ‘‘Tonda Gentile delle Langhe’’ in Piemonte and ‘‘Santa Maria di Gesu`’’ in Sicily. Shelled hazelnut accounted for 79% on the total amount of world’s hazelnut production while in-shell hazelnuts
* Corresponding author. E-mail address:
[email protected] (C. Costa). 1537-5110/$ – see front matter ª 2008 IAgrE. Published by Elsevier Ltd. All rights reserved. doi:10.1016/j.biosystemseng.2008.09.013
418
biosystems engineering 101 (2008) 417–424
accounted for 21%. The food industry applies a qualitative selection only to shelled nuts. Kernels for confectionery should be plump and free from shrivelled kernels, mould and extraneous matter. In Italy, the confectionery industry and high quality markets have increased their demand for entire shelled nuts of the same cultivar to emphasize the traditional product or in order to guarantee local production. Quality improvement plans have primarily pointed to genetic and cultural improvements of products grown in plantations which seek to produce more homogeneous fruits and are not scattered among other plantations. The improvement in quality was identified in the conception of the improvement plan scheme as an important target which would pave the way to improved competitiveness (SEC, 2002). For this reason cultivar discrimination during on-line quality selection is required by high quality food industries. Actually, non-destructive on-line selection of nuts using image analysis and opto-electronic techniques mainly leads to damage identification or sorting by merceological characteristics, for example opening extent of pistachio shells (Ghazanfari et al., 1997; Pearson and Toyofuku, 2000). These systems are able to detect external fruit damages (especially insect damage and moulds) (Kim and Schatzki, 1998). Shapebased analysis using machine vision could help in grading and selecting morpho-types (Nagata and Cao, 2000). Various methods for quantitatively evaluating shapes have been suggested in biological and agronomical context. The most common is based on elliptic Fourier descriptors, which have been successfully applied to the evaluation of several plant organs such as leaves (White et al., 1988; McLellan, 1993; Iwata and Ukai, 2002; Jensen et al., 2002; Neto et al., 2006), leaflets (Furuta et al., 1995), kernels (Ohsawa et al., 1998), roots (Iwata et al., 1998), flowers (Uga et al., 2003; Yoshioka et al., 2004) and fruits (Goto et al., 2005). This method describes the entire shape mathematically by transforming coordinate information concerning the contours into Fourier coefficients (Rohlf and Archie, 1984). The aim of this work was to evaluate the potential use and efficacy of shape-based techniques in order to discriminate among 4 traditional Italian cultivars (Tonda di Giffoni, San Giovanni, Mortarella and Tonda Romana). Tonda di Giffoni and Tonda Romana are very similar and are classified by FAO (2000) as round shaped, while the other two cultivars are classified as oblong. Multivariate classification methods (Partial Least Square Discriminant Analysis, PLSDA) were applied on elliptic Fourier analysis (EFA) coefficients extracted from the digital images of hazelnuts either in-shell or as kernels. The main goal was to demonstrate an innovative
combined approach based on opto-electronic techniques (image analysis, morphometry and multivariate statistical analysis) that could lead to future application, at industrial level, of rapid, effective and non-destructive cultivar selection.
2.
Materials and methods
About 100 hazelnuts per cultivar (Tonda di Giffoni, San Giovanni, Mortarella and Tonda Romana) (Table 1) were randomly chosen from three producing plants. All the fruits were provided by the cultivar collection of the Fruit Grown Research Institute of Caserta and harvested in the production season 2006. Being a methodological study, in order to avoid the influence of several factors, which can affect the size and shape of the kernels, samples came from a single growing location, field, crop year and from a selected research station that certified the genotype and the growing conditions – the present study would need to be extended to a more extensive range of material before it could be applied in practice. RGB images were analysed using a morphological method based on the elliptic Fourier approximation to closed contours in a two-dimensional plane. Digital images of hazelnuts were acquired using a background illumination in order to maximize the object outline. Digital images, with high optical resolution, were acquired for each fruit in-shell (I) and unblanched kernel (K) from different viewpoints: polar view (PV), lateral view (LV) and the random plane positioning view (RV) (Fig. 1).
2.1.
Elliptic Fourier analysis
The overall shell shape was studied through EFA of the contour coordinates (Rohlf and Archie, 1984). This method consists of decomposing a curve into a sum of harmonically related ellipses (e.g. Lestrel, 1997; Loy et al., 2000). The EFA was applied on the three series of outlines. The profiles were extracted with the software TPSdig2 (Rohlf, 2006) obtaining 180 equally angularly spaced points for each individual (Fig. 2). There are several approaches that can be used to deal with outline data; they involve fitting some type of curve to the outline and then using the parameters of the curve for subsequent analysis (Rohlf, 1996). Of these, elliptic Fourier decomposition (Kuhl and Giardina, 1982) has been shown to be a powerful taxonomic descriptor (Rohlf and Archie, 1984; Ferson et al., 1985; Loy et al., 2000; Sheets et al., 2006) and has several advantages over other methods such as that it does not require explicit definition of a biologically homologous or
Table 1 – Numbers and weights of the hazelnuts analysed for each cultivar Cultivar Mortarella San Giovanni Tonda Giffoni Tonda Romana
Label
Number of fruits analysed
Weight in-shell (g) SD
Weight kernel (g) SD
MORTARELLA GIOVANNI GIFFONI ROMANA
104 104 102 104
1.93 0.35 2.43 0.55 2.58 0.45 2.85 0.54
0.84 0.21 1.08 0.38 1.22 0.31 1.28 0.39
biosystems engineering 101 (2008) 417–424
419
Fig. 1 – Examples of the three different views. (A) Polar view of the in-shell; (B) lateral view of the in-shell; (C) random plane positioning view of the in-shell; (D) polar view of the kernel; (E) lateral view of the kernel; (F) random plane positioning view of the kernel.
mathematically determined outline centroid. Given an outline approximated as a polygon described by a series of xy-coordinates, Fourier analysis is used to reduce the dimensionality of the dataset and to eliminate the redundant information. This redundancy is determined by the high correlation between adjacent coordinates in the raw dataset. EFA is based on the separate Fourier decomposition of the incremental changes of the x- and y-coordinates as functions of the cumulative chordal length of the outline polygon. It yields the spectrum of the closed contour in terms of harmonically related trigonometric curves. For each harmonic equation, two Fourier coefficients are computed for both the
x- and y-projections, and thus the total number of coefficients is 4n, where n is the number of harmonics fitted to the outline (Crampton, 1995). The correct number of harmonics was calculated using the method proposed by Crampton (1995). The Fourier series was truncated for both the view and the merceological status (I or K) at the value of k at which the average cumulative power was 99.99% of the average total power. As recommended by Rohlf and Archie (1984), Elliptic Fourier coefficients were mathematically normalized in order to avoid bias in results due to the different size, location, rotation and starting position of specimens. The coefficients of the harmonic equations were extracted with the software Morpheus (Slice, 1998) and analysed via multivariate classification. Mean outline for each group was graphically extracted using Morpheus.
2.2.
Fig. 2 – Average cumulative power of the average total power at different k harmonics for the random plane view of the unblanched kernel (LV-I). The gray line represents the value of 99.99%. The outline, constructed on 180 equiangular spaced coordinates (one every 28), is also shown.
Statistical classification and modelling
In order to build models discriminating between the cultivars based on different view and status, a multivariate classification analysis PLSDA was applied. PLSDA (Sjo¨stro¨m et al., 1986; Sabatier et al., 2003) consists of a classical partial least squares (PLS) analysis regression where the response variable is a categorical one (Y-block; replaced by the set of dummy variables describing the categories) expressing the class membership of the statistical units. Therefore, PLSDA does not allow for response variables other than those that define the groups of individuals. As a consequence, all measured variables play the same role with respect to the class (cultivar) assignment. PLS components result for the compromise between two purposes: describing
420
biosystems engineering 101 (2008) 417–424
the set of explanatory variables and predicting the response ones. The PLS-based classification benefits from such a property as it builds typologies with intrinsic prediction power. The model includes a calibration phase and a validation phase for which residual errors (root mean square error) were calculated (RMSEC, RMSECV). The prediction ability of PLSDA depends also on the number (k) of the LV used in the model. The optimal k value was determined by predicting the results for independent sets of samples (test set) for different values of k, and determining the value of k for which the highest percentage of correct classification was found in the test set. PLSDA calculates a ‘‘prediction probability’’ and a classification threshold for each class modelled. These are calculated using a Bayesian method which takes the predicted y values from the PLSDA model, fits a normal distribution to them, then uses that to calculate the probability of observing a given y-value. The PLSDA analysis provides the percentage of correct classification and the loadings of each species on each LV. This analysis expressed also the statistical parameters indicating the modelling efficiency indicated by sensitivity and specificity parameters. The sensitivity is the percentage of the species of a category accepted by the class model. The specificity is the percentage of the species of the categories different from the modelled one, rejected by the class model. This analysis was performed using Matlab (rel. 7.1, PLSToolbox Eigenvector rel. 4.0) on the shape variables (Xblock; harmonic coefficients). The X-block (EFA coefficients) values were pre-processed with an abs procedure (takes the absolute values of the data). Each dataset, consisting of the harmonic coefficients of each view and the merceological status (I or K), was divided into two subsets: the first, containing 80% of fruits, was used for the class modelling and
Table 2 – k Values at different view and status View Polar Lateral Random plane Polar Lateral Random plane
Status
Label
k
In-shell In-shell In-shell Unblanched kernel Unblanched kernel Unblanched kernel
PV-I LV-I RV-I PV-K LV-K RV-K
5 7 6 12 10 10
validation; the second (83 fruits) was used for the independent test. To optimally select the 20% test set, the Kennard and Stone (1969) algorithm was applied. This algorithm belongs to the family of space-filling algorithms and is based on Euclidean distances between data. These algorithms select objects without the a priori knowledge of a regression model. The hypothesis is that as the true model is rather complex, it requires an uniform distribution of objects in the information space (see also Costa et al., in press for further detail on methodology). PLSDA was performed in order to discriminate each viewpoints and merceological status of: (a) the 4 studied cultivars (Tonda di Giffoni, San Giovanni, Mortarella and Tonda Romana); (b) the differences between the two rounded shape cultivars (Tonda di Giffoni and Tonda Romana); (c) the differences between the two oblong shape cultivars (San Giovanni and Mortarella); (d) The differences between rounded and oblong cultivars.
3.
Results
The total number of harmonics (the ‘Nyquist frequency’) is equal to 90. The Fourier series was truncated at different k
Fig. 3 – Mean configurations of the different cultivars with different view and status.
421
biosystems engineering 101 (2008) 417–424
Table 3 – PLSDA summarized results: (A) Discrimination between the four cultivars (cv); (B) Discrimination between the two rounded shaped cv (Tonda Romana and Tond Giffoni); (C) Discrimination between the two oblong shaped cv (San Giovanni and Mortarella); (D) Discrimination between the rounded and oblong cv PLSDA
cv-view
N LVs
% Classif Model
% Classif Test
Cumulated % explained variance (X-block)
Mean RMSECV
Mean sensitivity
Mean specificity
A
4C_LV-I 4C_LV-K 4C_PV-I 4C_PV-K 4C_RV-I 4C_RV-K
10 11 8 6 11 10
83.2 74.5 59.9 53.0 79.2 65.7
97.5 77.5 59.8 34.7 87.8 69.0
99.98 99.95 99.99 99.87 99.99 99.94
0.29 0.33 0.38 0.40 0.31 0.34
0.90 0.79 0.72 0.63 0.86 0.82
0.88 0.85 0.75 0.74 0.85 0.78
B
RO_LV-I RO_LV-K RO_PV-I RO_PV-K RO_RV-I RO_RV-K
4 11 6 5 7 4
85.6 84.7 75.2 78.9 78.8 74.5
90.0 82.4 73.2 67.6 95.1 50.0
99.96 99.96 99.98 99.85 99.97 99.86
0.35 0.35 0.43 0.41 0.40 0.42
0.86 0.84 0.74 0.76 0.74 0.76
0.86 0.85 0.73 0.77 0.76 0.75
C
OB_LV-I OB_LV-K OB_PV-I OB_PV-K OB_RV-I OB_RV-K
8 10 5 9 11 9
97.0 87.2 76.6 80.5 97.0 91.3
97.6 89.2 87.8 73.0 97.6 91.9
99.98 99.94 99.97 99.92 99.99 99.93
0.24 0.33 0.41 0.40 0.24 0.30
0.96 0.85 0.76 0.80 0.99 0.91
0.96 0.87 0.78 0.82 0.99 0.90
D
OR_LV-I OR_LV-K OR_PV-I OR_PV-K OR_RV-I OR_RV-K
11 10 7 10 10 6
91.1 91.3 78.6 77.0 92.5 87.4
98.8 95.8 75.6 60.0 97.6 94.4
99.98 99.94 99.99 99.93 99.98 99.89
0.27 0.28 0.40 0.41 0.29 0.33
0.95 0.91 0.78 0.76 0.92 0.86
0.94 0.92 0.78 0.76 0.93 0.88
values, depending on the viewpoints and merceological status (I or K), at which the average cumulative power reached 99.99% (Table 2). Fig. 2 represents an example of cumulative power of the Fourier series for the random plane view of the unblanched kernel (LV-I) where the Fourier series were truncated at k=7 harmonic equations. In Fig. 3 the mean configurations of the different cultivars with different view and status are represented. The number of harmonics for each view and status depends on the one calculated and reported in Table 2. It is possible to observe general differences between round shaped hazelnuts (Tonda di Giffoni and Tonda Romana) and oblong shaped hazelnuts
(San Giovanni and Mortarella); meanwhile round shaped cultivars appear similar.
3.1. PLSDA to discriminate the four cultivars (Tonda di Giffoni, San Giovanni, Mortarella and Tonda Romana) In Table 3 (Section A) it is possible to observe the higher percentage of correct classification in the independent test for the in-shell (I) views, particularly for lateral view (97.5%) and random plane view (87.8%). In Fig. 4 the scatter plot of the first 3 LVs scores obtained by the PLSDA on the four cultivars on the LV-I is reported. Differences between cultivars are clearly
Fig. 4 – Scatter plot of the first 2 LVs obtained by the PLSDA on the four cultivars on the LV-I. Tonda di Giffoni (black circle); San Giovanni (black triangle); Mortarella (white triangle); Tonda Romana (white circle).
422
biosystems engineering 101 (2008) 417–424
Fig. 5 – Scatter plot of the first 2 LVs obtained by the PLSDA on the rounded cultivars on the RV-I. Tonda di Giffoni (black circle); Tonda Romana (white circle).
observable. The PLSDA model for LV-I has high sensitivity (0.90) and specificity (0.88).
3.2. PLSDA to discriminate the two rounded shape cultivars (Tonda di Giffoni and Tonda Romana) In Table 3 (Section B) it is possible to observe a higher percentage of correct classification in the independent test for the in-shell (I) views, particularly random plane view (95.1%) and lateral view (90.0%). In Fig. 5 differences between cultivars are not very visible on the first three axes even though the efficacy of discrimination of the PLSDA model with 7 LV is very high. However this model doesn’t have high values of sensitivity (0.74) and specificity (0.76) indicating a limited efficacy and efficiency in modelling even with high performance of classification.
3.3. PLSDA to discriminate the two oblong shape cultivars (San Giovanni and Mortarella) In Table 3 (Section C) it is possible to observe the higher percentage of correct classification to discriminate between the two oblong cultivars in random plane view (RV-I ¼ 97.6%; RV-K ¼ 91.9%) with respect to the other views, except LV-I (97.6%). In Fig. 6 differences between the two cultivars are clearly observable. The PLSDA model for RV-I has very high values of sensitivity (0.99) and specificity (0.99).
3.4. PLSDA to discriminate rounded and oblong cultivars In Table 3 (Section D) it is possible to observe the higher percentage of correct classification to discriminate between oblong and rounded cultivars in both lateral view (LVI ¼ 98.8%; LV-K ¼ 95.8%) and random plane view (RV-I ¼ 97.6%; RV-K ¼ 94.4%) with respect to the polar view. In Fig. 7 differences between the two cultivars are clearly observable. The PLSDA model for RV-K has high values of sensitivity (0.86) and specificity (0.88). In general the lateral view and random plane view were better for classification, particularly for the in-shell hazelnuts.
4.
Discussion
Describing fruit shape is often necessary in agricultural research for a range of different purposes including cultivar description in applications for plant variety rights or cultivar registers (Paulus and Schrevens, 1999; Beyer et al., 2002; Paglia et al., 2008), evaluation of consumer preference (Kays, 1999), investigating hereditability of fruit shape traits (Cannon and Manos, 2001), or analyzing shape abnormalities (Brewer et al., 2007). Considering the reduced sample size and the controlled conditions of growth and harvest, the proposed methodology suggested high classification performance. In this study two
Fig. 6 – Scatter plot of the first 2 LVs obtained by the PLSDA on the oblong cultivars on the RV-I. San Giovanni (black triangle); Mortarella (white triangle).
biosystems engineering 101 (2008) 417–424
423
Fig. 7 – Scatter plot of the first 2 LVs obtained by the PLSDA on the oblong (black triangle) vs rounded (gray circle) cultivars on the RV-K.
different approaches were applied. The first is based on the appropriate and regular positioning of the fruit in the lateral and polar view. The second is based on the random positioning of the fruit on the plane.
(Cannon and Manos, 2001; Beyer et al., 2002) or flowers (Dominguez et al., 1998; Yoshioka et al., 2004).
4.1.
The implementation of modern multivariate techniques, such as PLSDA, on shape-based methods, such as EFA, on digital images, demonstrates, in this preliminary study, a high efficiency and performances in fruit grading and classification. The use of an image analysis system based on common chromatic digital standard (RGB or grayscale), instead of spectroscopic analysis, matches the general agro-industrial aim to develop selection machines based on consolidated and cheaper technology.
Lateral and polar view
In this study the best results on classification were obtained with the lateral view positioning in particular of the in-shell fruit (LV-I; from 90.0% to 98.8%), followed by the lateral view of the kernel (LV-K; from 77.5% to 95.8%). Meanwhile the polar view positioning in hazelnuts did not produce good results (PV; from 34.7% to 87.8%). Shape is considered a fundamental property of a fruit together with size, colour, condition and absence of defects (Kays, 1999). Cultivar certification is based also on a general and qualitative shape description (see for example Paulus and Schrevens, 1999). This study has demonstrated the technical feasibility of a quantitative shape description and determination of four hazelnuts cultivars. For example, the application of similar methods on different fruit leads to cultivar description (Paulus and Schrevens, 1999; Dubey et al., 2006) and fruit grading (Tao et al., 1995; Ghazanfari et al., 1997).
4.2.
Random plane view
In this study the results obtained with the random plane view positioning are good especially for the in-shell fruits (RV-I; from 87.8% to 97.6%). This approach has the potential to be implemented in the future to develop new on-line sorting systems with high speed and minimum damage. However, it is necessary to undertake a validation on a wider sample size to cover the possible variations expected from growing hazelnuts in different regions, crop years and agropedo-climatic conditions. Among the opto-electronic non-destructive methods, to our knowledge, no spectroscopic techniques have been used for cultivar selection, while there are few shape-based studies for the same purpose (Tao et al., 1995; Paglia et al., 2008). In other research fields, especially botany and agronomy, there have been some shape-based studies that describe different varieties and species through analysis of leaves (Chi et al., 2002; Jensen et al., 2002; McLellan, 2005; Neto et al., 2006), fruits
5.
Conclusion
Acknowledgements The research project ‘‘FRUMED’’ was supported by the Italian Ministry of Agriculture and Forestry, Paper no. 30. The authors would like to thank Mrs. Iliana Niciarelli and Mr. Matteo Cegna for the support during the laboratory operation.
references
Beyer M; Hahn R; Peschel S; Harz M; Knoche M (2002). Analysing fruit shape in sweet cherry (Prunus avium L.). Scientia Horticulturae, 96, 139–150. Brewer M T; Moyseenko J B; Monforte A J; van der Knaap E (2007). Morphological variation in tomato: a comprehensive study of quantitative trait loci controlling fruit shape and development. Journal of Experimental Botany, 58(6), 1339– 1349. Cannon C H; Manos P S (2001). Combining and comparing morphometric shape descriptors with a molecular phylogeny: the case of fruit type evolution in Bornean Lithocarpus (Fagaceae). Systematic Biology, 50(6), 860–880. Chi Y T; Chien C F; Lin T T (2002). Leaf shape modelling and analysis using geometric descriptors derived from Bezier curves. Transactions of the ASAE, 46(1), 175–185. Costa C; Menesatti P; Aguzzi J; D’Andrea S; Antonucci F; Rimatori V; Pallottino F; Mattoccia M. External shape differences between sympatric populations of commercial clams Tapes
424
biosystems engineering 101 (2008) 417–424
decussatus and T. philippinarum. Food and Bioprocess Technology, in press. Crampton J S (1995). Elliptic Fourier shape analysis of fossil bivalves: some practical considerations. Lethaia, 28, 179–186. Dominguez C A; Eguiarte L E; Nunez-Farfa J; Dirzo R (1998). Flower morphometry of Rhizophora mangle (Rhizophoraceae): geographical variation in Mexican populations. American Journal of Botany, 85(5), 637–643. Dubey B P; Bhagwat S G; Shouche S P; Sainis J K (2006). Potential of artificial neural networks in varietal identification using morphometry of wheat grains. Biosystems Engineering, 95(1), 61–67. FAO (2000). Inventory of hazelnut research, germoplasm and references. REU Technical Series No.56. Ferson S; Rohlf F J; Koehn R K (1985). Measuring shape variation of two-dimensional outlines. Systematic Zoology, 34, 59–68. Furuta N; Ninomiya S; Takahashi S; Ohmori H; Ukai Y (1995). Quantitative evaluation of soybean (Glycine max L., Merr.) leaflet shape by principal component scores based on elliptic Fourier descriptor. Breeding Science, 45, 315–320. Ghazanfari A; Irudayaraj J; Kusalik A; Romaniuk M (1997). Machine vision grading of pistachio nuts using Fourier descriptors. Journal of Agricultural Engineering Research, 68, 247–252. Goto S; Iwata H; Shibano S; Ohya K; Suzuki A; Ogawa H (2005). Fruit shape variation in Fraxinus mandshurica var. japonica characterized using elliptic Fourier descriptors and the effect on flight duration. Ecological Research, 20, 733–738. Iwata H; Ukai Y (2002). SHAPE: a computer program package for quantitative evaluation of biological shapes based on elliptic Fourier descriptors. Journal of Heredity, 93, 384–385. Iwata H; Niikura S; Matsuura S; Takano Y; Ukai Y (1998). Evaluation of variation of root shape of Japanese radish (Raphanus sativus L.) based on image analysis using elliptic Fourier descriptors. Euphytica, 102, 143–149. Jensen R J; Ciofani K M; Miramontes L C (2002). Lines, outlines, and landmarks: morphometric analyses of leaves of Acer rubrum, Acer saccharinum (Aceraceae) and their hybrid. Taxon, 51, 475–492. Kays S J (1999). Preharvest factors affecting appearance. Postharvest Biology and Technology, 15, 233–247. Kennard R W; Stone L A (1969). Computer aided design of experiments. Technometrics, 11, 137–148. Kim S W; Schatzki T F (1998). Detection of insect damages in almonds. Proceedings of SPIE, 3543, 101–110. Kuhl F P; Giardina C R (1982). Elliptic Fourier features of a closed contour. Computer Graphics and Image Processing, 18, 236–258. Lestrel P E (1997). Fourier Descriptors and Their Applications in Biology. Cambridge University Press. Loy A; Busilacchi S; Costa C; Ferlin L; Cataudella S (2000). Comparing geometric morphometrics and outlines fitting methods to monitor fish shape variability of Diplodus puntazzo (Teleostea, Sparidae). Aquacultural Engineering, 21(4), 271–283. McLellan T (1993). The roles of heterochrony and heteroblasty in the diversification of leaf shape in Begonia dregei (Begoniaceae). American Journal of Botany, 80, 796–804. McLellan T (2005). Correlated evolution of leaf shape and trichomes in Begonia dregei (Begoniaceae). American Journal of Botany, 92(10), 1616–1623. Nagata M; Cao Q (2000). Study on grade judgment of fruit vegetables using machine vision. Japan Agricultural Research
Quarterly, 32(4). Available from: http://www.jircas.affrc.go.jp/ kankoubutsu/jarq/32-4/nagata/nagata.htm. Neto J C; Meyer G E; Jones D D; Samal A K (2006). Plant species identification using Elliptic Fourier leaf shape analysis. Computers and Electronics in Agriculture, 50(2), 121–134. Ohsawa R; Tsutsumi T; Uehara H; Namai H; Ninomiya S (1998). Quantitative evaluation of common buckwheat (Fagopyrum esculentum Moench) kernel shape by elliptic Fourier descriptor. Euphytica, 101, 175–183. Paglia G; Antonucci F; Costa C; De Giorgio D; Menesatti P (2008). Studio sulle differenze di forma in frutti di cultivar di mandorlo (Prunus amygdalus Batsch) attraverso l’analisi ellittica di Fourier. In: VIII Convegno Nazionale ‘‘La biodiversita` – una risorsa per sistemi multifunzionali’’, 21–23 Aprile 2008 Lecce. Paulus I; Schrevens E (1999). Shape characterization of new apple cultivars by Fourier expansion of digitized images. Journal of Agricultural Engineering Research, 72, 113–118. Pearson T; Toyofuku N (2000). Automated sorting of pistachio nuts with closed shells. Applied Engineering in Agriculture, 16(1), 91–94. Rohlf F J (1996). Fitting curves to outlines. Proceedings of Michigan Morphometric Workshop, Special Publication, 2, 167–177. Rohlf F J (2006). TpsDig Ver. 2.10, Digitalized Landmarks and Outlines. Department of Ecology and Evolution, State University of New York, Stoony Brook. Rohlf F J; Archie J W (1984). A comparison of Fourier methods for the description of wing shape in mosquitoes (Diptera: Culicidae). Systematic Zoology, 3, 302–317. SEC (2002). Analysis of the nut sector. Brussels, 10.7.2002. Commission staff working paper SEC (2002) 797 Sabatier R; Vivein M; Amenta P (2003). Two approaches for discriminant partial least square. In: Between Data Science and Applied Data Analysis (Schader M; Gaul W; Vichi M eds). Springer-Verlag, Berlin. Sheets H D; Covino K M; Panasiewicz J M; Morris S R (2006). Comparison of geometric morphometric outline methods in the discrimination of age-related differences in feather shape. Frontiers in Zoology, 3, 15. Sjo¨stro¨m M; Wold S; So¨derstro¨m B (1986). PLS discrimination plots. In: Pattern Recognition in Practice II (Gelsema E S; Kanals L N eds). Elsevier, Amsterdam. Slice D E (1998). Morpheus et al.: Software Fore Morphometric Research. Revision 01-30-98-Beta. Department of Ecology and Evolution, State University of New York, Stony Brook, New York. Tao Y; Morrow C; Heinemann P H; Sommer H J (1995). Fourier-based separation technique for shape grading of potatoes using machine vision. Transactions of the ASAE, 38(3), 949–957. Uga Y; Fukuta Y; Cai H W; Iwata H; Ohsawa R; Morishima H; Fujimura T (2003). Mapping QTLs influencing rice floral morphology using recombinant inbred lines derived from a cross between Oryza sativa L. and Oryza rufipogon Griff. Theoretical and Applied Genetics, 107, 218–226. White R J; Prentice H C; Verwijst T (1988). Automated image acquisition and morphometric description. Canadian Journal of Botany, 66, 450–459. Yoshioka Y; Iwata H; Ohsawa R; Ninomiya S (2004). Analysis of petal shape variation of Primula sieboldii by elliptic Fourier descriptors and principal component analysis. Annals of Botany, 94, 1–8.
