Food Quality and Preference 21 (2010) 987997

Contents lists available at ScienceDirect

Food Quality and Preference

journal homepage: www.elsevier.com/locate/foodqual

Sensory proles and preference analysis in ornamental horticulture: The case of the rosebush
Rachid Boumaza a,b,*, Lydie Huch-Thlier a, Sabine Demotes-Mainard a, Eric Le Coz a, Nathalie Leduc a, Sandrine Pelleschi-Travier a, El Mostafa Qannari c, Soulaiman Sakr a,b, Pierre Santagostini b, Ronan Symoneaux d, Vincent Gurin a
a

UMR 462 SAGAH (AGROCAMPUS OUEST/INRA/Universit dAngers), IFR Qualit et Sant du Vgtal, BP 60057, 49071 Beaucouz, France AGROCAMPUS OUEST, 2 rue Le Ntre, 49045 Angers, France ONIRIS, rue de la Graudire, BP 82225, 44322 Nantes, France d Laboratoire GRAPPE, Groupe ESA, 55 rue Rabelais, BP 30748, 49007 Angers, France
b c

a r t i c l e

i n f o

a b s t r a c t
The context of ornamental horticulture is considered in order to extend the techniques of sensory and preference evaluation by taking the rosebush as a plant model. In a preliminary study (Boumaza, Demotes-Mainard, Huch-Thlier, & Gurin, 2009), a sensory evaluation was conducted in order to set up a list of attributes. Subsequently, this list was adapted to assess 10 rosebushes. After the control of the panel performance using a multivariate strategy of analysis, the average scores were used in product mapping. The evaluation of the preferences with regard to these rosebushes was undertaken: 253 subjects were asked to rank the products by decreasing order of liking. Thereafter, the preference data were subjected to an internal preference mapping and a cluster analysis. Six homogeneous segments of consumers were eventually retained. By way of performing an external preference mapping, the average ranks were regressed upon the sensory attributes using principal component regression: the preferences of 67% of the consumers were satisfactorily explained by the attributes. 2010 Elsevier Ltd. All rights reserved.

Article history: Received 5 August 2009 Received in revised form 3 May 2010 Accepted 4 May 2010 Available online 7 May 2010 Keywords: Rose Rosa hybrida L. Visual quality Panel performance Preference Cluster analysis

1. Introduction In ornamental horticulture, visual aspect and fragrance are important elements of overall plant quality (Dijkshoorn-Dekker, 2002; Habib, Tribo, Gnard, & Le Bail, 1997; Heuvelink, Tijskens, & Kang, 2004) that can be analysed by sensory analysis methods (Boumaza et al., 2009). These qualities are as important as the gustative quality of food products and must be considered in addition to other quality criteria, such as disease resistance, not dealt with herein. In this study, we evaluated only the visual characteristics of plants, focusing on rosebushes. We have not considered herein the study of the ower fragrance since it requires a complete study of its own. Roses were chosen as a model as they are leading products in ornamental horticulture, with about 80 million pot plants and 220 million garden rosebushes sold each year (Roberts, Debener, & Gudin, 2003).

* Corresponding author at: AGROCAMPUS OUEST, Institut National dHorticulture et de Paysage, 2 rue Le Ntre, 49045 Angers cedex 1, France. Tel.: +33 241225481; fax: +33 241731557. E-mail addresses: rachid.boumaza@agrocampus-ouest.fr, rachid.boumaza@ inh.fr (R. Boumaza). 0950-3293/$ - see front matter 2010 Elsevier Ltd. All rights reserved. doi:10.1016/j.foodqual.2010.05.003

In its endeavour to generate the novelty and diversity required to support the market, the horticultural sector is continually trying to improve the visual characteristics of plants. This process involves the creation of new varieties or changes in cultivation practices. Both approaches are slow and costly. For example, breeding process requires the selection over several generations of the best lineage from sexual crosses, prior to the large scale propagation of the new variety. This may last on average 10 years before a new variety of rosebush could be released on the market. Therefore, it is essential that efforts made by breeders to improve the visual characteristics of the plants should be aimed towards meeting consumer expectations from the very beginning of the breeding process. However, assessments of consumer preferences concerning the visual qualities of pot-grown roses and rosebushes have so far mostly been based on expert opinion or sales data. Up to our knowledge, no analysis of preferences associated with sensory attributes has ever been published for roses. In this paper, we present a rst attempt to set up experimental conditions to evaluate consumer preferences in ornamental plants. Such an approach would provide rose breeders and producers with a tool for evaluating the acceptability of the products they plan to release onto the market.

988

R. Boumaza et al. / Food Quality and Preference 21 (2010) 987997

A preliminary evaluation of rosebushes based on sensory analysis methods (Boumaza et al., 2009) was carried out to generate a list of visual attributes of bush roses. Subsequently, this innovative ornamental horticulture study was extended in several directions, three of which form the object of this work: the evaluation of products on the basis of appropriate attributes by a panel of assessors, the description of products in terms of the selected attributes and the analysis of consumer preferences by means of internal and external mapping. This approach is a routine in sensory analysis (Lawless & Heymann, 1999; MacFie & Thomson, 1994; Meilgaard, Civille, & Carr, 2007; Stone & Sidel, 2004). The main aim of this study was to assess the relevance of this approach to products of ornamental horticulture, according to the three main lines of research outlined hereafter and detailed in the materials and methods section. We decided to use photographs of plants as stimuli, rather than real plants. The panel of assessors developed a list of 16 attributes, based on analyses of photographs of rosebushes and adapted from the list developed by Boumaza et al. (2009) for living plants. The attributes selected by the panel related to the entire plant. They concerned both the form of the plant and its aerial organs easily distinguishable on photographs. The aim was to provide as an exhaustive description as possible with a limited number of attributes. Thereafter, the panel evaluated the stimuli which were considered for the present study. One of our concerns was to check the repeatability and reproducibility of the assessors and the panel, and their ability to discriminate between the products. Several relevant control techniques have been developed (Meullenet, Xiong, & Findlay, 2007 for review). In this study, we present the results obtained with two established techniques (Dijksterhuis, 1995; Rossi, 2001) and assess the discriminating power of the attributes using a classical method and an original global method based on cluster analysis. A panel of subjects ranked the same rosebush photographs by order of preference. These subjects were recruited among the exhibitors and visitors of an annual international ower show (Salon du Vgtal Angers, France, 2008), one of the most important European shows frequented by more than 16,000 visitors. These subjects, named hereinafter consumers, were also asked to provide some socio-demographic information about themselves. The aim was to investigate the relationship between their preferences and the visual characteristics of the rosebushes. We start by describing the methods used for data collection and processing. Thereafter, we present results for the evaluation of 10 rosebushes and nally discuss further possible investigations.

Fig. 1. The panel of 10 rosebushes displayed at the Salon du Vgtal 2008 at Angers for the collection of preference data. These photographs of rosebushes were also used for the collection of sensory data.

on the basis of their visual diversity, without reference to their variety. 2.2. Sensory data The products were evaluated by a jury of 14 experts, 2 men and 12 women. Most of them were from the scientic and administrative staff of the QUASAV (quality and health of the plant) Federative Research Institute of Angers (France). They all have a good knowledge about plants but no one was acquainted with neither the varieties of the 10 rosebushes nor their origins. They were mainly selected on their availability and motivation. These assessors had previously been trained in seven sessions, each lasting 90 min to 2 h, and had already participated in a jury of this type. That allowed us to assess their ability to carry out such an evaluation. During training, the assessors selected and specied the following 16 attributes from an initial list of 44 attributes (Boumaza et al., 2009) using a consensus method (Lawless & Heymann, 1999): top sided shape (Top_Sided_Shape), foliage thickness (Foliage_Thickness), plant symmetry (Plant_Symmetry), stem vigour (Stem_Vigour), number of stems (N_Stem), branching level (Branching_Level), number of owers (N_Flower), staggering of owering (Flower_Staggering), ower enhancement (Flower_ Enhancement), ower size (Flower_Size), number of faded owers/fruits (N_Faded_Fruit), number of oral buds (N_Bud), density of ower petals (Petal_Density), intensity of ower colour (Flower_ Colour_Intensity), leaf size (Leaf_Size), and darkness of leaf colour (Leaf_Colour_Darkness). Indeed, Boumaza et al. (2009) studied a primary list of 44 attributes and reduced it to a list of 18 attributes by discarding 26 attributes. This reduction was based on three criteria: unambiguity,

2. Materials and methods 2.1. Samples This work was based on the observation of photographs rather than of real plants. This is motivated by the fact that plants are living products and their growth modies their shape. Rose development is rapid, particularly as concerns the transformation of ower buds into open owers, the withering of the owers and fruit production. Therefore, the choice of photographs as stimuli ensures that repeated evaluations of plants were performed in equivalent conditions at each scoring session and for each product. This mode of operation allowed the analysis to be carried out over several days and, more importantly, provided the jury with stimuli that were similar to that presented to the consumers. In the evaluation considered in this work, 10 photographs of rosebushes (Fig. 1) of different cultivars or forms were selected

R. Boumaza et al. / Food Quality and Preference 21 (2010) 987997

989

discriminating power and independence. The authors have also mentioned the possibility to reintroduce an eliminated attribute providing that its meaning is specied to achieve a consensus among the assessors. From the list of 18 attributes, we discarded the attributes of the organs which were not present on most of the roses to be assessed or not visible on their photos, as bud or petal shape, thorn density and leaf condition. From the list of 26 discarded attributes, we reintroduced and specied more clearly the attributes plant symmetry and branching level and substituted two attributes, namely the number of stems and foliage density, to the attribute plant compactness. Thereafter, a large image database of rosebushes gave an opportunity to improve the performance and agreement of the assessors. This database was also used to select images that represented extreme anchors for each attribute. Fig. 2 shows these extreme anchors for four attributes. The extreme anchors of the 12 other attributes can be found in Boumaza et al. (2009) where the questionnaire submitted to the assessors is also given. For the nal evaluation, each assessor evaluated three times each of the 10 rosebushes for the 16 attributes on a structured nine-level scale. The rosebushes were presented in a random order for each scoring session. The three scoring session took place 2 7 days apart, depending on the availability of the assessors, in a room dedicated to sensory evaluation (Meilgaard et al., 2007). The sensory data thus obtained consisted of 420 (10 products 14 assessors 3 sessions) 16-dimensional statistical units, corresponding to the scores attributed by the jury for the 16 attributes. 2.3. Preference data The 10 rosebushes were presented on a panel (Fig. 1) displayed during the annual international ower show Salon du Vgtal (Angers, France, 2008) attended mostly by professionals from the horticultural industry. A sample of 253 potential consumers recruited on a voluntary basis ranked these rosebushes by order of preference, with no ties allowed. The ranks assigned to the various products by each consumer were considered to indicate the degree of liking of these products.

Consumers ranked the products in descending order of preference. For convenience, we reversed the ranking, so that the most preferred product had the highest rank. Consumers were also asked to provide information about their age, their gender and their sector of activity. 2.4. Methods of analysis 2.4.1. Concordance between assessors 2.4.1.1. Repeatability, reproducibility and consonance. Meullenet et al. (2007) classied techniques for controlling the performance of assessors and panels both by product and attribute using univariate techniques and also, for the entire set of products, either by attribute or with the entire list of attributes using multivariate techniques. We measured the performance of the panel by calculating the consonance by attribute introduced by Dijksterhuis (1995). This consists in considering each attribute in turn and setting up a data table whose rows correspond to the products and columns to the scores given by the various panellists for the attribute under consideration. Thereafter, this data table is subjected to principal component analysis (PCA). From the examination of the loadings associated with the rst principal component it is possible to single out some assessors who did not agree with the rest of the panel with respect to the attribute under consideration. Ideally, these loadings should be of the same sign and of the same magnitude and the percentage of variance explained by the rst principal component should be close to 100%. This percentage of total variance accounted for by the rst principal component is tightly linked to the consonance (Dijksterhuis, 1995). We also computed the repeatability and reproducibility measures introduced by Rossi (2001) to assess the overall panel performance. For each product and each attribute, the repeatability measure is the square root of the pooled variance of the scores and the reproducibility measure is the standard deviation of the mean scores attributed by the assessors. The lower the repeatability and reproducibility measures, the higher the overall panel performance is. 2.4.1.2. Discriminating power of the attributes. The discriminating power between products of an attribute can be assessed by the

Fig. 2. Rosebushes corresponding to the extreme scores of 4 attributes among the 16 used by the assessors: Foliage_Thickness, Plant_ Symmetry, N_Stems and Branching_Level.

990

R. Boumaza et al. / Food Quality and Preference 21 (2010) 987997

F-ratio, or equivalently the p-value, for product effects in the twoway analysis of variance (ANOVA) carried out with this attribute and the two factors: assessor and product. We also propose another strategy of analysis to assess the discriminating power of the attributes in a multivariate setting. As stated above, the original sensory data were considered as consisting of 420 statistical units in a 16-dimensional space, each dimension corresponding to an attribute. We should expect that units pertaining to the same product are very close to each others. Moreover, if the attributes have good discrimination ability, then we should expect that units from different products are far removed from each others. In the ideal case, a partition of the units into 10 (the number of products) clusters using a cluster analysis procedure will lead to the statistical units corresponding to the same product being grouped into a single class, with the 10 classes corresponding to the 10 products. This constitutes the rationale behind our strategy of analysis which consists in four steps. Step 1. Partitioning of the 420 statistical units into 10 classes. The hierarchical tree derived from the ascending hierarchical cluster analysis using Wards criterion is cut into 10 classes (corresponding to the number of products). Step 2. Each class is named according to the majority product it contains and a confusion matrix is computed by cross-tabulating the actual classes (products) and the classes obtained by cluster analysis in Step 1. The general term Mij of this confusion matrix designates the number of statistical units corresponding to a product i assigned to class j. Step 3. The percentage of statistical units correctly classied, which we shall refer to as the panel discrimination index, is computed. This index is maximal (100%) if the products differ markedly for all the attributes considered and the assessors display a high level of repeatability and reproducibility in their scoring for each attribute considered separately (Mandel, 1991; Rossi, 2001). Step 4. Listing of all the incorrectly classied statistical units which can be subjected to further investigations. In an approach similar to that proposed by Rossi (2001) facilitating the visualisation of atypical measurements, atypical statistical units are detected almost automatically for further analysis. 2.4.2. Mapping of products and of preferences Once the validity of the sensory data has been checked, as described above, the products attributes matrix of average scores (over panellists and repetitions) was computed and subjected to PCA in order to identify correlations between attributes and achieve a synthetic description of the products. This PCA also made it possible to construct new uncorrelated variables, the principal components, for use in external preference mapping. As to the preference data, we studied the rosebushes in terms of the mean preferences of all consumers, then in terms of the mean preferences for each group of consumers with the same sociodemographic characteristics. Thereafter, we generated both an internal map of consumer preferences, with the aim of identifying homogeneous segments of consumers, and subsequently an external map (Greenhoof & MacFie, 1994) to identify the visual attributes of rosebushes accounting for the mean preferences of each segment of consumers, based on regression analysis on principal components. The study was carried out in two main stages. The rst involved mapping consumer preferences to determine the principal directions of these preferences. Thereafter, the consumers were divided into homogeneous preference segments using a cluster analysis approach and, for each segment, the relationships between preferences and sensory data were investigated using multiple linear regressions of the mean ranks of the segment against the rst ve principal components. Then the regression coefcients were re-expressed in terms of the sensory attributes. High absolute values of these coefcients indicate attributes with

a high effect on the preferences of the concerned consumer segment. Moreover, the signs of these coefcients reect whether the consumers in the segment under consideration positively (positive signs) or negatively (negative signs) react to the attributes. 2.5. Statistical analysis All statistical analyses were carried out in the R environment (R Development Core Team, 2009) with the cluster package, except for ANOVA, which was carried out with Statgraphics (1997) software. 3. Results 3.1. Sensory data 3.1.1. Concordance between assessors 3.1.1.1. Repeatability, reproducibility and consonance. The overall panel performance measured by panel repeatability and panel reproducibility (Table 1) indicates a satisfactory repeatability and reproducibility for most of the attributes except the attribute Branching_Level which concerns the architecture, and the attributes N_Flower, Flower_Enhancement and Flower_Colour_Intensity which concern the owers. The consonance measure (Table 2) indicated satisfactory agreement between the assessors for most of the attributes, but the level of agreement was low for the attributes Branching_Level and Flower_Enhancement. Unfortunately, we were unable to identify outlier assessors on the basis of the loadings (not shown), because different assessors were singled out as outliers depending on which attribute is considered. 3.1.1.2. Discriminating power of the attributes. Using the cluster analysis approach discussed in Section 2.3.1, we found that 87.6% of the statistical units were correctly classied. The percentage of correct classication (Table 3) of statistical units by product varied from 76% to 100%, except for product A, for which the percentage was 48%. In total, 52 of the 420 statistical units were incorrectly classied, and 38 of these incorrect classications involved products A (22), B (10) and F (6). Fourteen of the 42 statistical units relating to product A were attributed to class I and eight were attributed to class F. This may be due to the resemblance between product A and products I and F (Fig. 1). The other major misclassications concerned products F and B, which also share substantial similarities. Thus, the high rate of correct classication (87.6%) may be interpreted as a measurement of the discrimination capacity of the attributes. It may also be interpreted as a measurement of the overall ability of the assessors to differentiate between the products. By way of validating the general strategy of analysis, we applied it to the 368 correctly classied statistical units, the percentage of correct classication increased to 99.8%, giving an almost diagonal confusion matrix. We carried out all subsequent calculations both for all 420 statistical units and also for only the 368 correctly classied statistical units. The results obtained with these two sets of calculation were not signicantly different and, therefore, we retained the whole dataset for the subsequent analyses. Let us now consider the discriminating power of each attribute measured by the F-ratio from the product effects (Table 4). All these F-ratios are high (p < 105) indicating a good discrimination ability of all the attributes. Moreover, it is worth noting that for the majority of the attributes the within variation were small except for the attributes Branching_Level, N_Flower, Flower_Stagerring, Flower_Enhancement. This highlights some difculties in using these latter attributes.

R. Boumaza et al. / Food Quality and Preference 21 (2010) 987997

991

Table 1 Panel repeatability and reproducibility measures (Rossi, 2001). The values greater than 2 are pointed out in bold characters. The smaller these measures are, the most efcient the panel is. Attribute Panel repeatability Top_Sided_Shape Foliage_Thickness Plant_Symmetry Stem_Vigour N_Stem Branching_Level N_Flower Flower_Staggering Flower_Enhancement Flower_Size N_Faded_Fruit N_Bud Petal_Density Flower_Colour_Intensity Leaf_Size Leaf_Colour_Darkness Panel reproducibility Top_Sided_Shape Foliage_Thickness Plant_Symmetry Stem_Vigour N_Stem Branching_Level N_Flower Flower_Staggering Flower_Enhancement Flower_Size N_Faded_Fruit N_Bud Petal_Density Flower_Colour_Intensity Leaf_Size Leaf_Colour_Darkness A 0.90 1.11 0.76 1.01 1.11 1.47 1.18 0.91 1.88 1.65 0.94 0.87 1.36 0.71 1.23 1.43 0.58 0.59 1.08 1.25 1.65 1.67 1.82 1.44 2.36 1.65 0.54 1.39 1.91 0.83 0.97 0.87 B 0.67 1.10 1.19 1.00 0.76 1.34 0.76 0.98 0.69 0.60 1.04 1.46 0.76 0.77 0.79 0.82 0.58 0.77 1.21 0.75 1.22 1.91 0.83 1.18 1.02 0.53 1.27 1.13 1.12 1.68 0.92 0.75 C 0.79 0.80 0.95 0.98 0.60 1.33 0.53 0.83 0.46 1.05 1.32 1.31 0.95 0.91 0.96 0.95 0.70 0.48 0.87 1.63 0.60 2.43 0.50 1.12 1.08 1.23 1.02 1.05 0.89 1.13 0.86 0.92 D 0.94 0.93 0.53 1.07 0.85 1.20 0.77 0.87 1.10 0.79 1.08 0.94 1.31 1.00 0.85 1.18 0.77 0.83 0.47 1.51 0.85 1.54 1.71 1.36 1.52 0.62 1.47 1.10 1.44 1.22 1.09 1.10 E 1.05 0.89 0.96 0.87 1.21 0.69 1.23 1.85 1.41 1.39 0.72 0.93 1.65 0.49 0.76 1.10 0.65 0.96 0.84 1.33 1.01 1.87 1.36 1.38 0.93 0.74 1.22 0.87 1.09 0.95 1.30 1.27 F 0.71 0.85 1.38 0.96 0.58 1.13 1.02 1.04 1.29 1.06 0.77 1.08 1.24 1.13 1.14 0.83 0.83 0.81 1.33 0.76 1.09 2.07 2.13 1.15 1.62 0.70 0.64 1.21 1.24 1.41 1.08 0.71 G 0.38 0.89 0.69 0.62 0.67 1.16 0.96 0.74 0.96 0.93 0.15 1.25 1.10 0.85 0.69 1.22 0.39 1.12 1.09 1.08 1.01 1.62 1.06 1.07 0.98 0.62 0.45 1.41 1.13 1.06 0.67 1.11 H 0.80 1.00 1.08 1.21 0.79 0.80 1.29 1.13 1.48 1.10 1.07 1.25 0.93 0.60 0.99 1.20 0.85 0.76 1.51 0.85 1.02 1.93 2.21 0.88 1.50 1.11 1.26 1.41 1.03 0.84 0.94 1.14 I 0.91 1.07 0.67 1.24 0.98 0.86 1.46 1.29 1.64 0.96 0.86 1.11 1.54 2.42 0.87 0.98 1.12 0.88 0.72 1.59 1.17 1.58 2.22 1.93 2.46 0.51 1.39 1.21 0.86 2.11 0.97 1.18 J 0.94 0.90 0.56 1.27 0.72 0.96 0.77 1.33 1.02 1.10 1.28 1.13 1.09 0.85 0.90 0.85 1.11 1.19 0.54 1.79 0.66 2.00 0.43 1.32 1.08 1.03 1.37 1.34 1.63 1.04 0.98 0.97

Table 2 Attribute consonances: percentage of the total variance (%) accounted for by the rst principal component of the mean products assessors matrix. The low levels of agreement between the assessors are pointed out in bold. Attribute Top_Sided_Shape Foliage_Thickness Plant_Symmetry Stem_Vigour N_Stem Branching_Level N_Flower Flower_Staggering Flower_Enhancement Flower_Size N_Faded_Fruit N_Bud Petal_Density Flower_Colour_Intensity Leaf_Size Leaf_Colour_Darkness Consonance 93 92 87 68 86 36 75 73 48 85 91 81 75 79 84 72

Table 3 Confusion matrix obtained following the partitioning of the 420 statistical units into 10 classes each class being named after its predominant product. The last column shows the correct classication rates (in%), by product and for all products. Product Class name A A B C D E F G H I J 20 32 41 39 6 1 37 36 42 1 41 1 38 Total 3 42 2 4 1 B C D E F 8 7 G 3 1 H I 14 J 47.6 76.2 97.6 92.9 88.1 85.7 100 97.6 90.5 100 87.6 %

3.1.2. Mapping of the products The rst three principal components in the standardized PCA of the mean products attributes matrix (Table 4) respectively accounted for 35%, 22% and 17% of the total variance, corresponding to 74% in total. An analysis of PCA outcomes made it possible to cluster the attributes into three segments on the basis of their correlations with these rst three principal components. The rst dimension is related to attributes of overall shape, vigour, leaf size, thickness associated with the number of branches and oral or-

gans. The second dimension characterises the greenness of the leaves and the staggering of owering associated with the number of fruits. The third dimension is linked to attributes of owers. From the rst principal component (Fig. 3, PC1) three classes of products can be distinguished: C and J: wide, dense rosebushes, with many frail branches carrying large numbers of oral organs and small leaves; G, D and E, with the opposite characteristics: top sided rosebushes, with only a few vigorous branches carrying few owers and large leaves; A, F, I, B and H with intermediate characteristics.

992

R. Boumaza et al. / Food Quality and Preference 21 (2010) 987997

Table 4 For each attribute and each product, mean and standard deviation (between parenthesis) of the 42 scores given by the 14 assessors in three replicates. The extreme mean values and the standard deviations greater than 2 are pointed out in bold. The last column shows the product F-ratios (two-way ANOVA with product and judge factors). A Top_Sided_Shape Foliage_Thickness Plant_Symmetry Stem_Vigour N_Stem Branching_Level N_Flower Flower_Staggering Flower_Enhancement Flower_Size N_Faded_Fruit N_Bud Petal_Density Flower_Colour_Intensity Leaf_Size Leaf_Colour_Darkness 6.2 (0.9) 6.5 (1.1) 7.0 (1.2) 6.2 (1.5) 6.3 (1.9) 3.8 (2.0) 5.3 (2.0) 3.6 (1.6) 5.2 (2.8) 2.7 (2.1) 1.2 (0.9) 7.0 (1.5) 2.4 (2.2) 1.8 (1.0) 5.8 (1.4) 5.9 (1.5) B 8.2 (0.8) 6.1 (1.2) 5.9 (1.5) 6.6 (1.1) 6.0 (1.3) 5.0 (2.2) 7.7 (1.0) 5.9 (1.4) 7.6 (1.2) 7.1 (0.7) 2.3 (1.5) 4.6 (1.6) 6.3 (1.3) 5.5 (1.8) 5.5 (1.1) 7.2 (1.0) C 2.3 (0.9) 7.8 (0.8) 2.1 (1.2) 3.3 (1.8) 8.0 (0.8) 4.6 (2.6) 7.9 (0.7) 6.4 (1.3) 6.9 (1.1) 4.2 (1.5) 2.0 (1.5) 7.0 (1.5) 7.3 (1.2) 4.0 (1.3) 3.1 (1.2) 5.5 (1.2) D 7.6 (1.1) 2.8 (1.1) 8.3 (0.6) 6.8 (1.7) 3.4 (1.1) 4.5 (1.8) 4.2 (1.8) 7.8 (1.5) 6.7 (1.7) 7.4 (0.9) 5.9 (1.7) 3.4 (1.3) 6.3 (1.8) 2.5 (1.5) 6.5 (1.3) 5.2 (1.4) E 6.1 (1.1) 3.8 (1.2) 7.2 (1.1) 5.9 (1.5) 4.8 (1.4) 3.6 (1.9) 5.8 (1.7) 7.0 (2.0) 6.8 (1.5) 3.8 (1.4) 7.0 (1.3) 2.8 (1.1) 4.0 (1.7) 8.2 (1.0) 6.6 (1.4) 4.0 (1.5) F 7.9 (1.0) 7.3 (1.1) 5.8 (1.7) 7.3 (1.1) 7.8 (1.2) 4.0 (2.2) 5.5 (2.2) 5.4 (1.4) 5.1 (1.9) 6.7 (1.1) 1.5 (0.9) 7.6 (1.5) 6.2 (1.6) 4.9 (1.7) 5.5 (1.4) 7.1 (1.0) G 8.6 (0.5) 3.1 (1.3) 7.0 (1.2) 7.6 (1.2) 2.8 (1.1) 2.9 (1.9) 4.1 (1.3) 4.4 (1.2) 7.1 (1.2) 6.0 (1.0) 1.1 (0.5) 4.0 (1.7) 5.6 (1.4) 6.4 (1.2) 7.2 (0.9) 6.1 (1.5) H 7.0 (1.1) 3.6 (1.1) 4.4 (1.7) 4.7 (1.3) 6.6 (1.2) 5.3 (2.0) 5.9 (2.4) 8.2 (1.3) 6.4 (1.9) 4.6 (1.4) 6.0 (1.5) 6.3 (1.7) 3.7 (1.3) 7.8 (1.0) 3.8 (1.2) 4.5 (1.5) I 4.4 (1.3) 6.6 (1.2) 7.9 (0.9) 6.1 (1.9) 6.8 (1.4) 4.1 (1.7) 4.4 (2.5) 6.9 (2.2) 3.8 (2.8) 1.3 (0.9) 4.5 (1.5) 5.6 (1.5) 1.3 (1.5) 3.3 (2.9) 5.0 (1.2) 6.8 (1.4) J 3.9 (1.3) 6.9 (1.4) 8.4 (0.7) 4.8 (2.0) 7.9 (0.9) 4.4 (2.1) 7.8 (0.8) 7.6 (1.7) 7.3 (1.4) 4.1 (1.4) 3.9 (1.7) 5.4 (1.6) 4.8 (1.8) 2.6 (1.2) 3.4 (1.2) 4.1 (1.2) F-ratio 188 148 125 37 113 10 51 52 21 104 145 64 67 95 80 41

A F I

A and F: rosebushes on which the oral organs were all at the same stage, buds for A and open owers for F, with dark green foliage for F. The third principal component (not shown) retained for this mapping contrasted rosebush I (no open owers) with rosebushes B and, to a lesser extent, C, which had much more apparent owering (many large and open owers, with a high petal density or well presented owers). Thus, the rst principal component was linked to the architectural characteristics of the rosebushes and the second and third principal components described their owering. Rosebushes A, F and I were located close together in the plane formed by the rst two principal components (Fig. 3), thus conrming what was observed in the discussion of classication errors in the confusion matrix (Table 3). 3.2. Preference data
2 4

2
G

PC2 (22%)

Stem_Vigour

Leaf_Colour_Darkness N_Bud Foliage_Density

Leaf_Size Plant_Symmetry N_Stem Top_Sided_Shape Flower_Size B Petal_Density N_Flower Flower_Colour_Intensity Branching_Level Flower_Enhancement N_Faded_Fruit Flower_Staggering

J D

PC1 (35%)
Fig. 3. Biplot of standardized PCA of the mean products attributes matrix.

The second principal component (Fig. 3, PC2) gave two extreme classes and a heterogeneous intermediate class. The two extreme classes contain rosebushes of the following types: H and E: rosebushes with an extended owering period with fruits/owers/buds simultaneously present and with pale green foliage;

3.2.1. Mean ranks Internal preference mapping (Greenhoof & MacFie, 1994) was achieved by carrying out PCA (Fig. 4) on the preference score data matrix with products as rows and consumers as columns. The rst two axes of this internal map, explaining only 43% of the total variance, indicated a preference for rosebushes J, A and C and the rejection of rosebushes G, H, E and F. One-way ANOVA of rank data, with product as the only factor, identied signicant differences in the classication of products (Table 5): rosebush A was the preferred rosebush and roses H and G, which had similar ranks, were the least preferred. Obviously, we did not consider the assessors as a source of variation in the ANOVA because we dispose of rank data. Statistical procedures more suitable for the analysis of rank

R. Boumaza et al. / Food Quality and Preference 21 (2010) 987997

993

10

15

PC2 (19%)

with an occupation related to horticulture (74%) which was precisely the target of the present preference study. With each rosebush considered separately, a three-way ANOVA (factors: age, gender and horticultural activity) on preference scores showed that the factors considered had only a small effect on preferences. Indeed, for the 10 analyses carried out (60 p-values, not shown), few signicant differences were identied with a 5% signicance threshold:
A

H G

10

15

10

10

PC1 (24%)
Fig. 4. First plane of the internal preference map showing the consumers (as vectors) and the products (AJ).

Rosebush A was less popular with assessors under the age of 25 years; Rosebush I was more popular with men than with women; Rosebush F was slightly more popular with those working in the horticultural sector; Preferences for rosebush C were the most complicated: sector of activity had no signicant effect, whereas the other two factors, age and gender, and their interaction were signicant. Among assessors between the ages of 35 and 45, this rosebush was markedly preferred by men than by women whose mean preferences (and standard error) are respectively 8.0 (S.E. = 0.6) and 5.4 (S.E. = 0.7). These preliminary results on global preferences and preferences as a function of socio-demographic category are interesting for rose breeders and producers. However, considering that the amount of total variance explained by the rst two principal components is not very large (43%), we have performed a segmentation of the consumers using cluster analysis. Indeed, the preference map (Fig. 4) indicates that several different segments of consumers could be identied. 3.2.2. Segmentation of consumers Hierarchical cluster analysis using Wards criterion was performed in order to segment the consumers. The hierarchical tree was cut into six segments of consumers. The numbers of consumers in each segment varied from 24 to 63 (Table 7). Fig. 5 shows a modied version of the internal preference map from Fig. 4 in which the consumers are represented separately for each segment. With the exception of segment 5, the consumer segments were homogeneous as all the members of within each segment displayed a similar preference pattern. The preferences of the consumers in segment 5 were not well accounted for by the rst two principal components and are better accounted for by subsequent components (not shown). Information about the products preferred or rejected by the consumers of each segment, including segment 5, is provided in Table 7. 3.2.3. External preference analysis We investigated the relationship between the preference scores of each segment of consumers and the sensory attributes, by regressing the mean scores of each segment against the rst ve components of the sensory data, which accounted for about 90% of the total variance. The determination coefcients of these regressions are shown in Table 8 and reect, for each segment of consumers, the percentage of the variance of preference scores accounted for by the sensory characteristics. The coefcients for consumer segments 4 and 5 are not particularly high. By contrast, for the other four segments, which correspond to two thirds of the panel, the coefcients obtained seem to be sufciently high to assess that sensory data account for the preference scores. The regression coefcients re-expressed as a function of the sensory attributes are plotted (Fig. 6). This latter gure, together with Fig. 7 which depicts the correlations of sensory attributes with the rst two principal components from internal preference mapping, reveal several trends:

Table 5 Preferences for rosebushes, in descending order. Rosebushes with different letters in the pairwise comparison column are signicantly different (p < 0.05; S.E. = 0.17). Product A J B C D E I F H G Mean preference 7.0 6.5 6.3 6.1 5.7 5.3 5.0 4.8 4.3 4.0 Pairwise comparison a b bc bc cd de e e f f

data (Friedman test) were also performed and led to the same conclusions. Table 6 gives the distribution of consumers by age, gender and occupation (related to horticulture or not). Consumers aged less than 25 years were more numerous and half of them were students (results not shown). Consumers aged more than 56 years were less numerous (13%). They were mainly formed of persons still in activity (not retired). The panel of consumers comported slightly more males than females and as expected, it contained more consumers

Table 6 Distribution of the background characteristics of the participants in the consumer test. Age 1625 years 2635 3645 4655 P56 Gender Male Female Sector of activity Horticultural Non-horticultural % 29.1 16.7 19.9 21.1 13.2 53.6 46.4 73.5 26.5

994

R. Boumaza et al. / Food Quality and Preference 21 (2010) 987997

Table 7 In the rst two lines, number of consumers in the six segments obtained by cutting the ascending hierarchical tree of consumer preferences (Fig. 5) into six segments. In the subsequent lines mean (standard error) preferences for each rosebush. Segment n % A B C D E F G H I J 1 63 25 8.1 (0.2) 5.3 (0.3) 8.8 (0.2) 3.7 (0.2) 3.8 (0.3) 4.9 (0.3) 2.2 (0.2) 4.3 (0.3) 6.9 (0.3) 7.1 (0.3) 2 42 17 4.8 (0.4) 7.5 (0.4) 6.6 (0.5) 5.1 (0.4) 6.5 (0.4) 4.3 (0.3) 6.7 (0.3) 6.6 (0.3) 1.8 (0.4) 5.3 (0.4) 3 40 16 7.0 (0.4) 5.5 (0.3) 2.8 (0.3) 6.5 (0.3) 5.1 (0.3) 4.4 (0.3) 2.6 (0.3) 3.7 (0.4) 8.8 (0.2) 8.8 (0.3) 4 50 20 7.5 (0.3) 7.4 (0.3) 8.4 (0.3) 7.0 (0.2) 5.5 (0.2) 4.0 (0.3) 3.5 (0.4) 3.4 (0.3) 2.8 (0.3) 5.7 (0.4) 5 34 13 7.6 (0.4) 8.2 (0.3) 3.4 (0.4) 6.9 (0.4) 5.2 (0.4) 7.1 (0.4) 4.5 (0.4) 3.6 (0.4) 3.7 (0.4) 4.8 (0.5) 6 24 9 Total 253 100 6.5 (0.6) 3.1 (0.3) 2.8 (0.5) 6.6 (0.4) 7.5 (0.4) 4.5 (0.5) 6.8 (0.5) 4.4 (0.5) 5.9 (0.5) 6.9 (0.5)

Segment 1 (25%) focused principally on the shape of the rosebush. The consumers in this segment liked dense, highly branched rosebushes and did not like top sided rosebushes. Segment 2 (17%) focused principally on owering characteristics. The consumers in this segment preferred rosebushes with lots of large owers with a high density of brightly coloured petals. They also seemed to prefer asymmetric rosebushes. Conversely, the consumers of segment 3 (16%) seemed to prefer, above all, symmetric rosebushes. They also appreciated pastelcoloured roses and a staggered owering period. The consumers of segment 6 (9%) were fewer in number than the members of the other segments and their preferences were more difcult to determine. However, they seemed to prefer symmetric, open rosebushes with large leaves. Thus, the preferences of two thirds of consumers could be accounted for by sensory attributes and three attributes number of faded owers and fruits, staggering of the owering period and branching level did not contribute to these preferences. Finally, it is worth noting that only segment 2 and segment 3 preferences are related to owering attributes.

4.1. Is the assessment protocol on photos satisfactory? As we stated in the section devoted to the presentation of the samples (Section 2.1), plants are living products. Their development modies their shape and their owering which are the main aspects of their visual quality. Boumaza et al. (2009) have proposed to carry out the assessment on photos of the products rather than on actual plants. This makes it possible to: give sufcient time to the sensory leader to organize testing sessions and preference assessments; present the panellists and the consumers with similar (images of) products; undertake sensory analysis in a room dedicated to sensory evaluation (Meilgaard et al., 2007). Obviously, this is not possible with real rosebushes. This assessment mode is also not completely disconnected from the reality of the market. Indeed, plants such as rosebushes are usually sold during the winter when they have lost their foliage and owers, and consumers rely on the picture attached to the naked plant to decide whether or not they buy the product. The present study corroborates the ndings of Boumaza et al. (2009). However, the study also highlighted some limitations. For instance, the repeatability and reproducibility of some attributes such as Level_Branching, and to a lesser extent Flower_Enhancement, Flower_Stagerring and N_Flower were not satisfactory. The assessment on the basis of photos is certainly responsible of this pitfall. In order to counteract this problem for at least some attributes such as Branching_Level, we can propose to present each assessor with two photos (instead of one photo) of each rosebush: one photo presenting the whole plant and another photo presenting the defoliated plant. It is also worth mentioning that with the advent of the technology, we may consider to present a video sequence of each rosebush showing the various facets of the same rosebush. Further improvements should also concern the denition of the attributes, the training of the panel or the use of instrumental mea-

4. Discussion and perspectives It emerges from the present study that the techniques of sensory analysis and preference studies can be successfully adapted to ornamental horticulture. However, several issues can be raised and we propose to address some of them in the following. In addition, it is worth noting that the data demonstrate a proof-of-concept, but results should be viewed as indicative and interpreted with caution as one of the referees of a previous version of the present manuscript rightly stated. Indeed, the rosebushes considered in this study although they represented a wide range of rosebushes in terms of variety, shape, colour and growth stage, were mainly used to serve the purpose of showing the interest of performing sensory techniques and preference assessment in ornamental horticulture.

R. Boumaza et al. / Food Quality and Preference 21 (2010) 987997

995

Segment 1
15
C

Segment 2
C

10

A H G F

H G

10

Segment 3
15
C

Segment 4
C

10

A H G F

H G

10

Segment 5
15
C

Segment 6
C

10

A H G F

H G

10

15

10

10

15

10

10

Fig. 5. Internal map of preferences, showing the rosebushes (AJ) and the consumers in each segment.

surements instead of some attributes. Regarding the denition of the attributes, Table 4 may be very useful in order to rene the scoring scale by showing, in addition to the photos reecting the extreme rosebushes, photos that show intermediate prototypes. It is worth noting that there exists an international norm which provides a standard scale for some attributes. For instance for assessing the size of the owers, UPOV (1990) recommends a structured nine-level scale. This scale denes the odd levels of the scale by referring to specic varieties. This denition makes it possible to provide references independently of the region of

production of the rosebushes. In current investigations, we are comparing the merits of the two kinds of scales. The aim is to replace the assessment of some sensory attributes by instrumental measurements which are less time consuming and less costly than sensory analysis on photos. In particular, we are currently exploring this possibility, with a view of replacing the attributes of ower and leaf dimensions, for example, by direct measurements of these features on plants. For this project, the rst measurements were performed in 2009, subsequent measurements are planned in 2010 and 2011.

996

R. Boumaza et al. / Food Quality and Preference 21 (2010) 987997

Table 8 For each consumer segment, determination coefcients for linear regressions of the mean preference of the segment upon the rst ve principal components of the sensory mean score matrix. Segment 1 2 3 4 5 6 R2 (%) 92 89 90 50 57 89 R2 adjusted (%) 82 75 77 11 3 76

4.2. Is the explanation of the preferences by sensory attributes satisfactory? The consumers who took part in the preference study were mainly professionals in horticulture. As such, they play a major role in directing and structuring the supply of ornamental plants. It is worth noting that a similar study was undertaken with non professionals in horticulture. The results (not shown herein) indicated that the attributes related to the owering have a high impact on the preferences of the consumers. Therefore, this comparison indicates that the panel of consumers considered in the present study is very likely biased towards professionals and may not be representative of all consumers of rosebushes. In the results section, we noted that the preferences of two thirds of consumers could be accounted for by the sensory attributes. This result is by itself very interesting. However, one may wonder if it is possible to increase this ratio. Firstly, we noted that three attributes number of faded owers and fruits, staggering of the owering period and branching level did not contribute to explaining the preferences. The fact that most of the consumers worked in the horticultural sector, may potentially explain why these owering attributes did not signicantly explain the preferences. Conversely, the lack of effect on preferences of branching level is much more surprising in a panel of professionals, as the number of branches at the base of the rosebush is one of the key elements in the classication of bush roses (AFNOR, 1990). For example, category 1 roses have at least two stems sprouting up at less than 10 cm above the grafting point generally located at the base of the rosebush. This nding raises several questions: is the attribute branching level really irrelevant and, therefore, not useful in this type of analysis? Were the roses not well differentiated with respect to this attribute or, was this attribute not consistently used by the assessors, as it was noted in the study of the concordance between assessors and the above discussion about the stimuli used. These questions cannot be answered on the basis of the results obtained with this single case study. Secondly, we think that the description of the rosebushes with only 16 attributes is not sufcient to explain all the diversity in preferences. We could be more exhaustive in the description of the products by adding visual attributes as the ower colour, or non visual attributes as fragrance. . . However, one must bear in mind that similarly to other products the ornamental quality of a plant is a non-objective multicriterion-based phenomenon that cannot be easily measured. It also depends on location (countries, regions. . .) and time (fashions. . .), on groups of consumers (gender, age. . .) as well as on individual conceptions.

1.0

0.5

6 3 1 1 3 6 3 1 2 6 3 6 6 1 2 3 2 1

2 2 3 2 6 1 3 3 1 6 6 1 1 2 3 6 6 3

2 6 2 2 1 3 6 1 1 3 1 2 6

Regression coefficient

0.0

2 6 2 6 3

2 1 2

2 6

1.0

0.5

3
Flower_Col_Intense Top_Sided_Shape Flower_Enhance Branching_Level Plant_Symmetry Foliage_Density N_Faded_Fruit Flower_Staggering Leaf_Col_Dark Petal_Density Stem_Vigour Flower_Size Leaf_Size N_Stem N_Flower N_Bud

Fig. 6. Regression coefcients for the mean preferences of consumer in segments 1, 2, 3 and 6, on sensory attributes using principal component regression analysis (regression on the rst ve principal components of the PCA performed on sensory data).

10

15

B 2 Petal_Density

N_Flower

PC2 (19%)

Flower_Enhancement Flower_Size

N_Bud Branching_Level

Foliage_Density 1 N_Stem

5. Conclusion In this study: We obtained a consensus evaluation of the products by a panel of assessors using a rigorous protocol. The panel selected the attributes, constructed a scoring task and evaluated the roses. This approach is completely novel in the ornamental horticulture sector. We described the rosebushes and classied them according to visual attributes based on plant architecture and the principal aerial organs of the plant. This description related to a small number of products in this study, but could be extended to a much larger number of products. Indeed, we have more recently investigated 21 products in a eld setting, in a study focusing on the effects of nitrogen nutrition on the visual quality of

A
Leaf_Colour_Darkness

Flower_Colour_Intensity

Top_Sided_Shape Flower_Staggering Leaf_Size E D Stem_Vigour N_Faded_Fruit

J 3

10

Plant_Symmetry

15

10

0 PC1 (24%)

10

Fig. 7. Representation of consumer segments and attributes as additional points on the rst plane of the internal preference map (Fig. 4).

R. Boumaza et al. / Food Quality and Preference 21 (2010) 987997

997

eld-grown garden roses. The results of this second study are currently being analysed. We accounted for the preferences of almost two thirds of the panel of consumers. We demonstrated the importance of plant architecture in consumer preferences. This result is particularly pertinent to our ecophysiological studies of the effects of environmental factors on the architecture of ornamental shrubs, using the rosebush as a model. However, these preliminary results concern only one panel of consumers and require conrmation with larger groups including more participants from outside the horticultural sector. This study was limited to rosebushes, but the approach developed could easily be generalised to other products in ornamental horticulture. This would necessarily involve the identication of suitable attributes for the products to be evaluated, the training of panels of assessors and the evaluation of products. Thus, notwithstanding some pitfalls related to the specically visual nature of ornamental plants, the sensory and preference evaluation techniques that have proved highly efcient in product characterisation and development in several domains, such as the food, cosmetics and car industries, can be successfully adapted to ornamental horticulture.

References
AFNOR. (1990). Produits de ppinires. Rosiers. Spcications particulires. Association franaise de normalisation. Norme NF.V12-053. La Plaine Saint-Denis, France: AFNOR Group. Boumaza, R., Demotes-Mainard, S., Huch-Thlier, L., & Gurin, V. (2009). Visual characterization of the esthetic quality of the rosebush. Journal of Sensory Studies, 24, 774796. Dijkshoorn-Dekker, M. W. C. (2002). Crop quality control system: A tool to control the visual quality of pot plants. Thesis, Wageningen University, The Netherlands. Dijksterhuis, G. (1995). Assessing panel consonance. Food Quality and Preference, 6, 714. Greenhoof, K., & MacFie, H. J. H. (1994). Preference mapping in practice. In H. J. H. MacFie & D. M. H. Thomson (Eds.), Measurement of food preferences. Chapman & Hall: Glasgow. Habib, R., Tribo, E., Gnard, M., & Le Bail, M. (1997). La nutrition azote des cultures et la qualit des produits. In Matrise de lazote dans les agro-systmes. Paris: INRA (Les colloques, no. 83). Reims (France), 1920 novembre 1996. Heuvelink, E., Tijskens, P., & Kang, M. Z. (2004). Modelling product quality in horticulture: An overview. Proceedings of the international workshop on models for plant growth and control of product quality in horticultural production. Acta Horticulturae, 654, 1930. Lawless, H. T., & Heymann, H. (1999). Sensory evaluation of food. New York: Springer. MacFie, H. J. H., & Thomson, D. M. H. (Eds.). (1994). Measurement of food preferences. Glasgow: Chapman & Hall. Mandel, J. (1991). The validation of measurement through interlaboratory studies. Chemometrics and Intelligent Laboratory Systems, 11, 109119. Meilgaard, M. C., Civille, G. V., & Carr, B. T. (2007). Sensory evaluation techniques. London: CRC Press. Meullenet, J. F., Xiong, R., & Findlay, C. J. (2007). Multivariate and probabilistic analyses of sensory science problems. Ames: Blackwell Publishing. R Development Core Team (2009). R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. ISBN 3900051-07-0, URL <http://www.R-project.org>. Roberts, A.V., Debener, T., & Gudin, S. (2003). Encyclopedia of rose science (Vol. 1). Amsterdam: Elsevier Academic Press. Rossi, F. (2001). Assessing sensory panelist performance using repeatability and reproducibility measures. Food quality and preference, 12, 467479. Statgraphics. (1997). Statistical Graphics Corporation. Statgraphics for Windows, 19931997. Stone, H., & Sidel, J. L. (2004). Sensory evaluation practices. San Diego: Elsevier Academic Press. UPOV. (1990). Guidelines for the conduct of tests for distinctness, homogeneity and stability. Rose (Rosa L.). UPOV/TG/11/7. Geneva: International Union for the Protection of New Varieties of Plants.

Acknowledgment The authors thank N. Mansion for the quality of the rose photographs, C. Bernard, M.H. Bouvet, D. Chalain, S. Chalain, R. Champenois, O. Douillet, P. Expert, C. Foubert, R. Guisnel, D. Honor, E. Jullien, V. Kapusta, M. Laffaire, D. Lalane, L. Landais, G. Michel, R. Robic, M.H. Sakr and I. Tranchant for their participation in the sensory assessment of the rose bushes. The authors wish to extend their thanks to two anonymous reviewers for their valuable comments.