Planta

, Volume 236, Issue 6, pp 1943–1954

Extraction of quantitative characteristics describing wheat leaf pubescence with a novel image-processing technique

Authors

  • Mikhail A. Genaev
    • Laboratory of Evolutionary Bioinformatics and Theoretical Genetics, Department of Systems BiologyInstitute of Cytology and Genetics SB RAS
  • Alexey V. Doroshkov
    • Laboratory of Evolutionary Bioinformatics and Theoretical Genetics, Department of Systems BiologyInstitute of Cytology and Genetics SB RAS
  • Tatyana A. Pshenichnikova
    • Department of the Genetic Resources of Experimental Plants, Sector of Genetics of Grain QualityInstitute of Cytology and Genetics SB RAS
  • Nikolay A. Kolchanov
    • Laboratory of Evolutionary Bioinformatics and Theoretical Genetics, Department of Systems BiologyInstitute of Cytology and Genetics SB RAS
    • Chair of Informational BiologyNovosibirsk State University
    • Laboratory of Evolutionary Bioinformatics and Theoretical Genetics, Department of Systems BiologyInstitute of Cytology and Genetics SB RAS
    • Chair of Informational BiologyNovosibirsk State University
Emerging Technologies

DOI: 10.1007/s00425-012-1751-6

Cite this article as:
Genaev, M.A., Doroshkov, A.V., Pshenichnikova, T.A. et al. Planta (2012) 236: 1943. doi:10.1007/s00425-012-1751-6

Abstract

Leaf pubescence (hairiness) in wheat plays an important biological role in adaptation to the environment. However, this trait has always been methodologically difficult to phenotype. An important step forward has been taken with the use of computer technologies. Computer analysis of a photomicrograph of a transverse fold line of a leaf is proposed for quantitative evaluation of wheat leaf pubescence. The image-processing algorithm is implemented in the LHDetect2 software program accessible as a Web service at http://wheatdb.org/lhdetect2. The results demonstrate that the proposed method is rapid, adequately assesses leaf pubescence density and the length distribution of trichomes and the data obtained using this method are significantly correlated with the density of trichomes on the leaf surface. Thus, the proposed method is efficient for high-throughput analysis of leaf pubescence morphology in cereal genetic collections and mapping populations.

Keywords

Common wheatComputer image analysisHigh-throughput phenotypingLeaf pubescenceTrichomes

Abbreviations

MAE

Mean absolute error

MAPE

Mean absolute percentage error

Introduction

Pubescence in wheat is due to unicellular unbranched epidermal outgrowths (trichomes). Leaf pubescence plays an important biological role in resistance to pests and adaptation to the environment and displays wide phenotypic variation (Vavilov 1990). It has been reported that highly pubescent wheat cultivars are much more resistant to the cereal leaf beetle (Oulema melanopus L.) (Ringlund and Everson 1968; Schillinger and Gallun 1968) and the Hessian fly (Mayetiola destructor) (Roberts et al. 1979). It has been demonstrated that pubescence is often associated with a more rational use of water, is important for water retention in spring common wheat and the trichome number changes during droughts (Johnson 1975; Hameed et al. 2002; Likhenko 2007). Krupnov and Tsapaikin (1990) reported that drought-resistant cultivars in the steppe ecological group develop abundant pubescence, while cultivars confined to humid climates, by contrast, develop little pubescence.

As have been found earlier by different methods of genetic analysis (Ringlund and Everson 1968; Leisle 1974; Maystrenko 1976; Doroshkov et al. 2011), leaf pubescence is often controlled by more than one gene. Division of hybrid populations into as many phenotypic classes as there are underlying genotypes is probably the most difficult part of any genetic experiment aimed at studying the inheritance of this trait. To date, two genes controlling wheat leaf pubescence have been mapped: Hl1 on chromosome 4B and Hl2 on chromosome 7B (Taketa et al. 2002; Dobrovolskaya et al. 2007; Pshenichnikova et al. 2007). Methods with which to quantify pubescence for further high-throughput analysis in various cultivars and hybrid populations with a view to gaining a better insight into the genetic control of this trait are yet to be developed. Currently, the typical methods used for exploring trichome morphology in wheats are qualitative visual evaluation and visual counting using lenses (Sharma and Waines 1994; Dobrovolskaya et al. 2007; Pshenichnikova et al. 2007). The drawback associated with the use of these methods includes a high labor consumption, and hence their limited applicability for analysis of a large number of samples in hybrid and mapping populations. Another popular approach for analysis of trichome morphology in plants is scanning electron microscopy (SEM) (Luo and Oppenheimer 1999; Nagata et al. 1999; Perazza et al. 1999). SEM allows the trichomes to be studied to the finest detail. However, this approach is expensive, time-consuming and unsuitable for high-throughput analysis.

These limitations can be overcome using computer methods based on automatic analysis of digitized images. A variety of such approaches had previously been proposed for analysis of leaf pubescence in A.thaliana. Kaminuma et al. (2008) proposed that digitized 3D images obtained using microfocus X-ray computed tomography could be effectively used for analysis of such quantitative characteristics describing leaf pubescence as trichome number, trichome size and the distance between trichomes. That method has proven to be quite informative; however, it requires special equipment. Bensch et al. (2009) developed an algorithm for counting trichomes on A. thaliana leaf blades, which allows trichome positions to be analyzed on images obtained using 3D confocal laser scanning microscopy. To study the development of trichomes, the authors analyzed the time series of images (4D data) of plant leaves, on which the trichomes were tagged with the trichome-specific green fluorescent protein.

We had previously proposed a method named LHDetect for counting trichomes on transversely folded leaves using their 2D images obtained by transmitted light microscopy (Doroshkov et al. 2009; http://wheatdb.org/lhdetect). The LHDetect performs image segmentation into a leaf and a background region, location of the leaf/background boundary, deployment of “ghost” leaf/background boundaries variously offset from the leaf/background boundary into the background region and counting the crossings between each “ghost” leaf/background boundary and trichomes. The resulting numbers of crossings between “ghost” leaf/background boundaries and trichomes are recalculated into a vector of the distribution of trichome lengths. A drawback of the method is that it is impossible to obtain information on the size of separate trichomes and a low accuracy in trichome number identification if trichomes in an image cross each other.

In this paper, we present the recommendations for photomicroscopy of the transverse folds of wheat leaves and a novel algorithm, which allows such images to be used for rapid quantitative evaluation of leaf pubescence. This algorithm has been implemented as the LHDetect2 software program, which comes in two flavors: as a console application and as a Web service. It has been demonstrated that LHDetect2 performs about 20 % better than LHDetect when counting trichomes. An advantage is that LHDetect2 is a high-throughput phenotyping method: it takes just 1 day to obtain detailed quantitative characteristics describing pubescence in dozens of plants.

We provide an example of analysis of leaf pubescence in wheat cultivar Hong-mang-mai and wheat line 102/00i, the leaves of which cannot be distinguished by touch. Differences in leaf pubescence between these genotypes and their F2 progeny have been identified.

Materials and methods

Photomicroscopy of folded wheat leaves for quantitative evaluation of leaf pubescence

The recommended steps for photomicroscopy of a wheat leaf are as follows. Detach a leaf from the plant. On the leaf blade surface, choose a location at which pubescence is to be analyzed. At the chosen location, fold the leaf transversely so that the adaxial surface is presented to view. Once the leaf is folded, place it on a slide and fix firmly with an adhesive tape. Take a photograph of the fold under a light microscope. A detailed description of imaging recommendations, including the required microscope settings, is provided in Additional file 1.

Typical photomicrographs obtained using these recommendations are presented in Fig. 1. In these images, the leaf region is to the right, the background region is to the left and the leaf/background boundary bisects the image vertically. Trichomes are identified as filamentous outgrowths narrowing from their bases (which are where they are attached to the leaf surface) toward their tips. In this work, penultimate leaves were used on all occasions.
https://static-content.springer.com/image/art%3A10.1007%2Fs00425-012-1751-6/MediaObjects/425_2012_1751_Fig1_HTML.gif
Fig. 1

Typical images of fold lines on transversely folded wheat leaves collected from a poorly pubescent (a) and a higher pubescent (b) plant. The scale bar is 500 µm long

The algorithm for the extraction of quantitative characteristics describing wheat leaf pubescence

The following are the characteristics that we considered in the images of each transverse fold of wheat leaves: the number of trichomes in the image (it tells about trichome density), trichome length and its distribution (it tells about trichome size), average trichome length (it tells about average trichome size) and the distance between the current trichome and the next on the leaf fold.

The LHDetect2 algorithm that we developed for the evaluation of these characteristics consists of several stages that are as follows:
  1. 1.

    conversion of an image into a grayscale and noise rejection;

     
  2. 2.

    segmentation of the image into the leaf region containing trichomes and the background region;

     
  3. 3.

    identification of the boundary between the leaf and the background;

     
  4. 4.

    identification of single trichomes and their groups as image objects;

     
  5. 5.

    analysis of the size and shape of the objects.

     

A detailed description of the algorithm, its parameters and performance is provided in Additional file 2.

A photomicrograph of the leaf in any of the PNG, TIFF, BMP, JPEG or GIF graphic format is input into LHDetect2 that processes the image according to the parameters set. The output data contain information on each particular object with trichomes: the ID number of the object which defines the object; the number of pixels within the contour; the indication of whether the object is plain or composite; the number of trichomes in the object; the number of trichome tips identified within the object; the coordinates of the trichome base border pixels; the distance between the center of the current trichome base and the center of the next on the leaf fold; the coordinates of each tip and the length of the trichome to which the tip belongs. The length of each trichome was approximated by the length of its centerline, which was reconstructed for each plain object using the algorithm proposed by Leifer et al. (2011) and for each composite object using our original algorithm explained in Additional file 2 on page 4.

The information on the objects is further converted into a vector, n, of the length distribution of trichomes using a script written in Perl. This vector is a numerical representation of the trichome length distribution histogram. Each element, ni, of this vector is the number of trichomes in the image that has the lengths falling within the user-specified ith discrete interval (bin). Note that the user may specify the units of measurement of trichome size other than pixels, for example, micrometers. For example, if the user sets the length interval at 42 µm (which is at 20 pixels on the scale we use) and the first three elements of the vector are n1 = 3, n2 = 5 and n3 = 2, then it means that LHDetect2 identifies three trichomes with lengths within the interval ]0;42], five trichomes with lengths within the interval ]42;84] and two trichomes with lengths within the interval ]84;126]. If, additionally, the maximum trichome length in the image is 500 µm, then the vector is of dimension 12 and the last interval is ]462;504].

This representation is convenient for comparing pubescence in different specimens as points in a multidimensional space with coordinates equal to the elements of the vector, with the dimension of the space being equal to the number of bins on the histogram. Only vectors obtained for identical length intervals can be compared.

LHDetect2: implementation and availability

The proposed algorithm is implemented in the LHDetect2 software program. LHDetect2 is a console application written in C using the OpenCV 2.0 library (Bradski and Kaehler 2008). LHDetect2 as a console application is available at https://github.com/genaev/LHDetect2 under the BSD license. LHDetect2 as a Web service is accessible at http://wheatdb.org/lhdetect2.

Assessment of LHDetect2 accuracy

LHDetect2 was assessed for performance using a test set of 70 images of folded leaves from wheat plants of two genotypes with contrasting leaf pubescence (cultivar Saratovskaya 29 and line 102/00i). These images were not used for the assessment of LHDetect2 parameters, but had similar ratios between poorly and highly pubescent samples.

To assess how accurately LHDetect2 counts trichomes, we manually drew circles, each being 15 pixels in diameter, around trichome tips on each image. If the tip of a trichome identified with LHDetect2 fell within the region, it was assumed that the trichome had been identified correctly. For each image, we calculated the true positive value (i.e., the number of correctly identified trichomes) (tp), the false negative value (i.e., the number of trichomes with no tips identified within a selected region) (fn), and the false positive value (that is, the number of misidentified trichomes) (fp). These values were further used for the calculation of the precision, pr = tp/(tp + fp), the recall, re = tp/(tp + fn), and the F score, F = 2(pr × re)/(pr + re) (Van Rijsbergen 1979). The higher the F score, the more accurate is the trichome identification. When all the trichomes are identified correctly and the false positive is zero, the F score is 1.

The F score was calculated for each image as described above and was then averaged for all the test images. Additionally, we calculated Pearson’s correlation coefficient for the number of trichomes on each test image determined using LHDetect2 and the corresponding manual count.

To assess the accuracy in trichome length measurements, we used 20 test images and compared the length of each trichome in pixels determined with LHDetect2 and that measured manually using the Segmented Line Tool of the image-processing program ImageJ (http://rsbweb.nih.gov/ij/). A total of 231 trichomes were analyzed. Pearson’s correlation coefficient was used as a measure of similarity between length estimates.

We also compared the performance of LHDetect2 and LHDetect when counting trichomes. We calculated the mean absolute error (MAE) and the mean absolute percentage error (MAPE) for the number of trichomes in the image:
$$ {\text{MAE}} = \frac{1}{M}\sum\limits_{j = 1}^{j = M} {|N_{j} - N^{\prime}_{j} |} $$
$$ {\text{MAPE}} = \frac{100\,\% }{M}\sum\limits_{j = 1}^{j = M} {\left( {\frac{{|N_{j} - N^{\prime}_{j} |}}{{N_{j} }}} \right)} $$
where M is the number of images in the test selection; \( N_{j} \) is the number of trichomes identified in the jth image manually; \( N^{\prime}_{j} \) is the number of trichomes identified with LHDetect2 or LHDetect. The higher the MAE and MAPE, the higher is the error in the number of trichomes identified as such.

With LHDetect, the measure of trichome number was the number of crossings between the trichomes and the “ghost” leaf/background boundary offset from the leaf/background boundary by 6 μm (3 pixels). With this offset, LHDetect had the highest accuracy in trichome number identification.

Comparison of the number of trichomes on the leaf fold and on the leaf surface

To demonstrate that the number of trichomes on a leaf fold line can serve as an adequate estimate of hair density (i.e., the number of trichomes per leaf surface area unit), we compared the average number of trichomes per 1 mm of the leaf fold line (nf) and the average number of trichomes per 1 mm2 of the leaf surface (ns). The estimates were obtained using 200 pairs of images of leaf fold lines and leaf surface fragments located 1–3 cm away from the leaf fold line (Fig. 2). For each plant, pairs of images were obtained at four locations at various distances from the leaf base. A total of 50 plants with various genotypes ranging from poorly pubescent to highly pubescent were used (for details, see “Plant material and growing conditions”). The leaf folds were imaged according to the recommendations laid down in Additional file 1.
https://static-content.springer.com/image/art%3A10.1007%2Fs00425-012-1751-6/MediaObjects/425_2012_1751_Fig2_HTML.gif
Fig. 2

A photomicrograph of a 102/00i plant leaf fold (a) and a fluorescent photomicrograph of the leaf surface (b) near the fold line. Five trichomes were identified on the leaf fold line (1.83 trichomes per 1 mm of the fold line). Twenty-one trichomes were identified on the leaf surface (3.63 trichomes per 1 mm2 of the leaf surface). The scale bar is 500 µm long

To count trichomes on the leaf surface, a leaf fragment was placed on a slide so that its adaxial surface was uppermost. The image of the surface was obtained using a Carl Zeiss Axioskop 2 Plus microscope. Trichomes were identified by autofluorescence of the leaf’s epidermis. The best image quality was achieved using a Zeiss filter set No. 02. All the other microscope settings were as those used for leaf fold imaging (Additional file 1).

The image area for the fold line and surface photomicrographs was 2.730 × 2.163 mm, which amounted to an area of 5.776 mm2. The density of trichomes on the leaf surface in these plants varied from 0.4 to 110 trichomes/mm2. The trichomes visible in the fold line and in the leaf surface were counted manually. Data were further normalized as follows: the fold line counts were normalized to the length of the fold line visible in the photomicrograph and the leaf surface counts were normalized to the area of the leaf surface visible within the photomicrograph.

Analysis of pubescence in cultivar Hong-mang-mai, line 102/00i, and F2 progeny from crosses between them

To demonstrate the power of the LHDetect2 method, we compared plants of two genotypes, their leaves being indistinguishable visually or to the touch. Those genotypes were cultivar Hong-mang-mai and line 102/00i. First, we compared the length distribution of trichomes between plants of the indicated genotypes (20 plants of each genotype) and progeny from their crosses (76 plants). For each leaf fold image, LHDetect2 yielded the vector (n) of the distribution of trichome lengths, the number of trichomes (N), the average trichome length (L) and the maximum trichome length (Lmax). The length interval used was 42 µm (20 pixels). There were a total of 40 intervals, the maximum bin width being 1,680 µm. The pubescence characteristics obtained for the kth plant were averaged, and so produced nk, Nk, Lk and Lmax k. This procedure was also run on 76 Hong-mang-mai × 102/00i F2 plants.

In search of groups of plants in the F2 population with similar pubescence within each such group, we applied hierarchical clustering to the 76 vectors ns. For each pair of plants, k and l, we calculated the Euclidean distance between the vectors nk and nl,
$$ d_{kl} = \left( {\sum\limits_{i = 1}^{i = 40} {({n}_{ik} - {n}_{il} )^{2} } } \right)^{\frac{1}{2}}, $$
where i is the number of the bin. This distance tells us how different the trichome length distribution in one plant is from that of another, and therefore how different is the pubescence of these plants. The greater the distance, the stronger is the difference in pubescence, and if dkl = 0, then the corresponding trichome length distributions are identical. Based on the pairs of distances so obtained, we applied complete linkage clustering (Sneath and Sokal 1973) to the plants. As a result, we built a binary tree with leaf nodes representing 76 plants from our experiment and edge lengths representing the degree of differences in pubescence between these plants.

Clustering of the vectors ns for the F2 plants and their parental populations was used as a control. The tree that was built following that clustering allows pubescence to be compared between F2 plants and their parents.

Secondly, we analyzed the distribution of maximum trichome lengths. The data used in this analysis were obtained as follows: the maximum trichome length was determined in each plant and the distribution of these values in the parents and progeny was analyzed. For each genotype, a dataset was created from the longest leaf trichomes found on all its plants. This analysis was performed to determine a possible interplay in the progeny between genes involved in trichome length.

Finally, to demonstrate that LHDetect2 allows useful information to be extracted not only on trichome length or trichome number, but also on their arrangement, we analyzed the arrangement of trichomes on the leaf folds in Hong-mang-mai and 102/00i plants. To this end, we used LHDetect2 outputs: the distance between the center of the current trichome base and the center of the next on the leaf fold. We converted those data into the distances between all possible pairs of trichomes in the image using a script written in Perl. Furthermore, we built the distribution of the distances between the closest short (no longer than 100 µm) trichomes and between the closest long (longer than 100 µm) trichomes in 60 images for each cultivar Hong-mang-mai and line 102/00i.

Plant material and growing conditions

In this work, we used spring wheat plants of various genotypes with a wide range of leaf pubescence variation. The plants with rare short trichomes were Rodina and Diamant 2. The plants with dense short and rare long trichomes were Saratovskaya 29, Hong-mang-mai, and line 102/00i derived by crossing Rodina and the wild wheat relative Aegilops speltoides Tausch (Lapochkina et al. 2003) and carrying the introgressed gene for leaf pubescence (Pshenichnikova et al. 2007). Pubescence in line 821 derived by crossing Saratovskaya 29 and Triticum timofeevii (Leonova et al. 2001) is similar to that in its tetraploid parent. Also, we used several plants from an F2 Diamant 2 × ANK-7B hybrid population. ANK-7B is an isogenic line derived from cultivar Novosibirskaya 67 (Koval et al. 2001) and has a gene for pubescence from k-46444, a Chinese variety (Vavilov Institute of Plant Industry, St. Petersburg, Russia).

The plants were grown in the hydroponic greenhouse of the Institute of Cytology and Genetics of the Siberian Branch of the Russian Academy of Sciences according to a single protocol. The substrate was expanded clay aggregate; the hydroponic solution was Knopp’s solution.

Statistical data analysis

Data analysis was performed using Microsoft Excel (Microsoft) and STATISTICA (StatSoft, Inc.).

Results

LHDetect2 accuracy

The comparison of trichome numbers estimated with LHDetect2 and manual counts is presented in Fig. 3a and Additional file 3. LHDetect2 has a tendency to underestimate trichome number, as on average it overlooks three to four true trichomes in every image it is run on. The coefficients of the linear regression forced through the origin yields y = 0.887x, r = 0.986 (P < 0.01), and the F score was 0.87. Trichome number underestimation occurs because LHDetect2 ignores any epidermal outgrowth that is not longer than 12.6 μm (6 pixels). At the same time, some of them were visually identified as being trichomes. However, their share in the overall distribution is quite low and it is therefore quite safe to state that a good correspondence is observed between the number of trichomes identified with LHDetect2 and the manual counts.
https://static-content.springer.com/image/art%3A10.1007%2Fs00425-012-1751-6/MediaObjects/425_2012_1751_Fig3_HTML.gif
Fig. 3

Comparison of estimates of pubescence characteristics obtained through LHDetect2 and manual countings for the number of trichomes (a) and the length of trichomes (b) (in pixels, 1 pixel = 2.1 µm) in a fold line. The solid line is the x = y line

A similarly good correspondence has been observed between trichome length estimates obtained using LHDetect2 and manual measurements (Fig. 3b, and Additional file 3). The number of pixels comprising the centerline of a trichome estimated by LHDetect2 is compared to the length of a broken line created manually using ImageJ to represent an approximation of the centerline. The coefficients of the linear regression forced through the origin yield y = 0.980x, r = 0.999 (P < 0.01). The average difference between the lengths obtained using LHDetect2 and measured manually is 3 pixels (6.3 µm), which makes up, on average, 6 % of the trichome length.

The results of a comparison of the LHDetect2 and LHDetect performance are presented in Table 1 and Additional file 3. We compared these methods with reliance on two parameters, which characterize the accuracy of identification of trichome number on an image, namely the MAE and MAPE (see “Materials and methods” and the “Assessment of LHDetect2 accuracy”). Additionally, we visually classified the 70-strong selection of images into plain (those containing few crossing trichomes) and composite (those containing a large number of trichomes and a large number of crossings between them). The sample contained 44 plain and 26 composite images. The average number of trichomes was 15.64 per plain image and 48.42 per composite image. The results of the comparison of the respective MAE and MAPE produced by LHDetect2 and LHDetect demonstrate that the error with LHDetect2 is about 0.8 as much as that with LHDetect. Interestingly, for plain images, the error is about the same, no matter which algorithm. Stronger differences are observed in composite images, which contain a large number of trichomes and their crossings. LHDetect MAE is twice as high as LHDetect2 MAE, and LHDetect MAPE is one-and-a-half times as high as LHDetect2 MAPE.
Table 1

Comparison of trichome number identification with LHDetect2 and LHDetect (Doroshkov et al. 2009)

Sample

LHDetect2

LHDetect

MAE

MAPE (%)

MAE

MAPE (%)

Entire dataset (70 images)

3.33

11.65

4.54

14.22

Plain images (44 images)

1.82

11.23

1.64

11.60

Composite images (26 images)

5.88

12.38

9.46

18.64

Presented are the respective mean absolute errors (MAEs) and mean absolute percentage errors (MAPEs). The entire dataset, the subset of plain images (those with few crossing trichomes) and the subset of composite images (those with a large number of trichomes and a large number of crossings between them) are considered separately

The number of trichomes on the transverse fold line of a leaf as a measure of leaf pubescence density

The relationship between trichome number on the leaf fold (nf) and leaf pubescence density (ns) can be described using a simple model (Fig. 4). The picture of the leaf fold displays trichomes on a part of the leaf surface (Sf) that is roughly equal to the area of a rectangle with width the thickness of the fold edge (h) and the length of the leaf fold line. Therefore, the number of trichomes per leaf surface area unit (ns) should be roughly equal to nf/h. However, if the pubescence density is low and h is small, no trichome may occur on the Sf area with a high probability. This implies that the regression line of ns on nf must intersect the y axis at a point >0.
https://static-content.springer.com/image/art%3A10.1007%2Fs00425-012-1751-6/MediaObjects/425_2012_1751_Fig4_HTML.gif
Fig. 4

The view of the leaf fold and the leaf surface area used for comparing the numbers of trichomes on them. A leaf surface unit area defined by a square with side S = 1 and the fold edge with area Sf and thickness h are shown. Triangle-headed arrows point at schematics of the leaf fold and the leaf surface unit area

This model is consistent with observed data. The results of a comparison of the manual count, nf, of trichomes per leaf fold line length unit and the manual count, ns, of trichomes per leaf surface unit at the location closest to the leaf fold line (see “Materials and methods” and the “Comparison of the number of trichomes on the leaf fold and on the leaf surface”) are presented in Fig. 5. As can be seen from this figure, at nf > 3 the dependence of ns on nf is linear (the correlation coefficient is significant, P < 0.01). Because ns was normalized to 1 mm2, the thickness of the fold edge, given the regression coefficient in this model equal to 3.120, is h = 1/3.120 ≈ 0.320 mm, which is approximately double the thickness of the wheat leaf.
https://static-content.springer.com/image/art%3A10.1007%2Fs00425-012-1751-6/MediaObjects/425_2012_1751_Fig5_HTML.gif
Fig. 5

Number of trichomes per 1 mm of the leaf fold line (nf) compared to the average number of trichomes per 1 mm2 of the leaf surface (ns), estimated for 200 images from 50 plants of a variety of genotypes (see "Materials and methods" and the “Comparison of the number of trichomes on the leaf fold and on the leaf surface”). Shown are the equation of the function ns(nf) at nf > 3, the correlation coefficient and the regression line. Gray is an uncertainty zone created by nf values less than or equal to 3, at which the regression of ns on nf is not satisfied

When nf ≤ 3, situations often arise where nf values are close, however, the corresponding ns values are strongly different, which characterizes this range of nf values as an uncertainty range (gray in Fig. 5). Within this range, there is a strong variation across leaf pubescence density values obtained from a single image, and that is why we recommend the user perform averaging over several different images from the same leaf (see the recommendations laid down in Additional file 1).

The analysis performed demonstrates that the number of trichomes in the leaf fold photomicrograph can serve as an estimate of leaf pubescence density on the wheat leaf.

Analysis of leaf pubescence in cultivar Hong-mang-mai, line 102/00i and F2 progeny from crosses between them

In this work, we have conducted a detailed study of leaf pubescence morphology in cultivar Hong-mang-mai and line 102/00i that have densely pubescent leaves. Their pubescence is due to short trichomes and occasional long trichomes and is indistinguishable between these genotypes visually or to the touch.

We scrutinized trichome length distributions in the Hong-mang-mai, line 102/00i plants and F2 progeny from crosses between them. Note that all the F2 plants did develop trichomes on their leaves and their leaf pubescence was similar both visually and to touch. The trichome length distribution vectors, trichome number, average trichome length and maximum trichome length inferred for each plant are presented in Additional file 4.

First of all, we compared the quantitative characteristics of leaf pubescence in the parents (Table 2). The Hong-mang-mai plants have denser leaf trichomes than the 102/00i plants (the average trichome number of the Hong-mang-mai plants is twice as large as that in the 102/00i plants). At the same time, the average trichome length in the Hong-mang-mai plants is nearly one-third as large as that in the 102/00i plants, which suggests that a higher short-to-long trichome ratio is found in Hong-mang-mai than in 102/00i plants. Additionally, the line 102/00i plants show a strong variation in the average trichome length. Interestingly, the average maximum trichome lengths in the plants of these genotypes are quite close, even though their respective average trichome lengths are strongly different. For either of the parents, no significant correlation was found between average trichome number and average trichome length measured on the 20 samples (Table 2).
Table 2

Comparison of pubescence between Hong-mang-mai and 102/00i plants

Genotype

Hong-mang-mai

102/00i

Sample size

20

20

\( \bar{N} \)

28.91

9.03

Var(N)

24.17

7.51

\( \bar{L} \)

136.79

202.02

Var(L)

447.62

1435.20

\( \bar{L}_{\max } \)

863.10

861.00

r (N, L)

−0.01 (P < 0.95)

–0.15 (P < 0.53)

The table contains the number of plants analyzed (sample sizes), the average number of trichomes \( \bar{N} \), the average trichome length L, their variances Var(N) and Var(L), the maximum trichome length Lmax, the coefficients of correlation between \( \bar{N} \) and \( \bar{L} \) and their P values

Furthermore, we compared pubescence in the F2 progeny. Analysis of the trichome number distribution demonstrates that the F2 plants are heterogeneous in relation to this parameter (see Additional file 5, Fig. A5.1). Analysis of the distribution of average trichome lengths in the F2 population (see Additional file 5, Fig. A5.2) also revealed heterogeneity.

To characterize the Hong-mang-mai × 102/00i F2 population in more detail, we used hierarchical cluster analysis comparing the vectors of the trichome length distributions among the F2 plants (see “Materials and methods”). The results are presented as a dendrogram (Additional file 5, Fig. A5.3) and revealed two groups among the F2 plants: group F2G1 containing 43 plants and group F2G2 containing 33 plants. Estimates of the quantitative characteristics describing their pubescence are presented in Table 3 (see also Additional file 5, Fig. A5.4, A5.5). As can be seen from this table, the F2G1 plants have denser pubescence (\( \bar{N} \) = 35.86) than the F2G2 and even the Hong-mang-mai plants. At the same time, the average trichome number in the F2G2 plants is nearly twice as large as that in 102/00i plants, but lower than that in the Hong-mang-mai plants. It is also interesting that the average trichome length between these groups is nearly equal, which suggests that so are the ratios of long to short trichomes. As far as average trichome length is concerned, the F2 plants are closer to the 102/00i than to the Hong-mang-mai plants. No significant correlations have been found between average trichome number and average trichome length within each of the F2G1 and F2G2 plants.
Table 3

Comparison of the quantitative characteristics describing leaf pubescence between Hong-mang-mai × 102/00i F2 plants in groups F2G1 and F2G2 revealed by hierarchical cluster analysis and having different pubescence morphology

F2 group

F2G1

F2G2

Sample size

43

33

\( \bar{N} \)

35.86

19.08

Var(N)

78.54

31.96

\( \bar{L} \)

190.41

194.39

Var(L)

639.71

907.69

\( \bar{L}_{\max} \)

1170.63

997.55

r (N, L)

0.11 (P < 0.48)

–0.01 (P < 0.95)

The table contains the number of plants analyzed (sample sizes), the average number of trichomes \( \bar{N} \), the average trichome length L, their variances Var(N) and Var(L), the maximum trichome length Lmax, the coefficients of correlation between \( \bar{N} \) and \( \bar{L} \) and their P values

To elucidate the details of the differences between hairiness in groups F2G1 and F2G2, we developed, for each group, a profile of the distribution of trichome lengths averaged over the entire group members. A comparison of these profiles with those developed for the parental genotypes was used as a control. The results are presented in Fig. 6.
https://static-content.springer.com/image/art%3A10.1007%2Fs00425-012-1751-6/MediaObjects/425_2012_1751_Fig6_HTML.gif
Fig. 6

Analysis of trichome lengths in 20 Hong-mang-mai (HMM) plants, 20 102/00i plants and 76 F2 progeny (43 in group F2G1 and 33 in group F2G2). On the x axis are the trichome lengths of 0–630 µm, the length of the intervals being 42 µm; on the y axis are the numbers of trichomes of each particular length, whiskers representing standard errors within the 95 % confidence interval for the trichome number of each particular trichome length

As can be seen, all the four distributions are unimodal, with maxima within the length interval of 0–100 µm. As the trichome length values increase, the number of trichomes drops abruptly and, for length values upward of 630 μm, the differences in trichome number among the different genotypes are within the 95 % confidence intervals (not shown on the figure).

The Hong-mang-mai plants have more 20–170 μm trichomes than the 102/00i plants. The difference in the number of trichomes of each particular length between these genotypes is much larger than the standard error within 95 % confidence intervals.

The trichome length distribution profiles for the F2 population display the strongest differences within the interval of 20–170 μm, which is similar to the behavior of the profiles developed for their parental forms. As with their parents, the difference in the number of trichomes of such length between F2G1 and F2G2 plants exceeds the standard error at the 95 % confidence level. It is well observed that the F2G1 plants developed more 0–200 μm trichomes than the F2G2 plants and that the Hong-mang-mai plants developed more such trichomes than the 102/00i plants (see Fig. 6). Within the interval of 0–150 μm, the differences in trichome number between the F2G1 and Hong-mang-mai plants are within the 95 % confidence interval. However, the F2 plants have more 120–250 μm trichomes than the Hong-mang-mai plants. Leaf pubescence in the F2G2 plants is strongly different from that in each of the F2G1, Hong-mang-mai and 102/00i plants. The F2G2 trichome numbers are somewhere in the middle between the Hong-mang-mai and 102/00i values and differ significantly from them.

Thus, the F2G1 and Hong-mang-mai plants have a very similar pubescence; pubescence is poorer in the F2G2 plants than in Hong-mang-mai plants, yet richer in the F2G2 plants than in the 102/00i plants. This conclusion is supported by the clustering diagram for the parental and F2 plants (Additional file 5, Fig. A5.6). In this diagram, 11 out of 20 Hong-mang-mai plants are within a large cluster mostly composed of F2G1 plants. At the same time, 16 out of 20 102/00i plants form a cluster, which is close to the cluster mostly composed of F2G2 plants.

Note that the longest trichomes in the F2 plants were 1,659 μm in length in group F2G1 and 1606.5 μm in length in group F2G2. In the parental plants, the longest trichomes were much shorter: 997.5 μm in the 102/00i plants and 1,134 μm in the Hong-mang-mai plants. We compared the distributions of the maximum trichome length in the F2 plants and their parental genotypes as described in “Materials and methods”. The results are presented in Fig. 7. As can be seen from this figure, the maximum trichome length values are higher for F2G1 and F2G2 plants than otherwise (F2G1 plants developing the longest trichomes). This may be indicative of an additive effect of the respective genes for pubescence in Hong-mang-mai and 102/00i.
https://static-content.springer.com/image/art%3A10.1007%2Fs00425-012-1751-6/MediaObjects/425_2012_1751_Fig7_HTML.gif
Fig. 7

Histograms of the distribution of the maximum trichome length in Hong-mang-mai (HMM), 102/00i, and two groups of their F2 progeny (F2G1 and F2G2). On the x axis are the trichome lengths ranging from 400 to 2,200 µm, the length of the intervals being 300 µm; on the y axis are the numbers of trichomes of each particular length. The legend is at the bottom right

Finally, we have perused trichome arrangements on the leaf blades in the plants of both parental genotypes. The results are presented in Fig. 8. As can be seen, the distributions of the distances between each of two closest short (<100 µm) trichomes and between each of two closest long (>100 µm) trichomes are different (significant with P < 0.01 according to the Kolmogorov–Smirnov test for both genotypes). These differences are seen particularly well in the Hong-mang-mai plants and are especially clear within the interval ]0;300] µm (Fig. 8b). It is possible that these results are due to differences that long and short trichomes have in location preferences. Long trichomes mainly occur along leaf veins (see Fig. 2b), which is likely to account for the high frequency of long trichomes ]200;300] µm apart. Thus, the data on trichome arrangements obtained using LHDetect2 are relevant to wheat leaf morphology, even though the images we have analyzed contain no information whatsoever on vein arrangement or any other morphological detail of the leaf.
https://static-content.springer.com/image/art%3A10.1007%2Fs00425-012-1751-6/MediaObjects/425_2012_1751_Fig8_HTML.gif
Fig. 8

Analysis of the distances between the base centers of the closest shortest (shorter than 100 µm) trichomes and the base centers of the longest (longer than 100 µm) trichomes. a The histogram of the distribution of such distances in the 102/00i plants. Dark bars stand for the distribution of the base-to-base distances between two long trichomes; light bars stand for the distribution of the base-to-base distances between two short trichomes. On the x axis are the distances ranging from 0 to 2,200 µm, the length of the intervals being 100 µm. On the y axis are the numbers of trichome pairs, for which the corresponding distance was observed. b The histogram of the distribution of such distances in the Hong-mang-mai plants. On the x-axis are the distances ranging from 0 to 1,200 µm, the length of the intervals being 100 µm. On the y axis are the numbers of trichome pairs, for which the corresponding distance was observed. The other designations are as in a

Discussion

The growing need for rapid and accurate approaches for large-scale assessment of phenotypic characters in plants becomes more and more obvious in the studies looking into relationships between genotype and phenotype (Benfey and Mitchell-Olds 2008; Houle et al. 2010). Modern approaches seek to take advantage of automated phenotyping, which warrants a much more rapid data acquisition, higher accuracy of the assessment of phenotypic features, measurement of new parameters of these features and exclusion of human subjectivity from the process (Montes et al. 2007; Hartmann et al. 2011). Additionally, automation allows measurement data to be rapidly loaded into computer databases, which reduces data processing time (Vankadavath et al. 2009; Joosen et al. 2010; Lu et al. 2011).

One of the methods that improve plant leaf hairiness phenotyping throughput is by using analysis of digital images (Kaminuma et al. 2008; Bensch et al. 2009). In this work, we propose a method for evaluating quantitative characteristics describing leaf pubescence in wheat by analysis of the number and geometric properties of trichomes on the leaf fold photomicrograph. The main idea underlying its development was to use relatively inexpensive equipment and to make analyses simple, rapid, reasonably accurate and informative. All this is especially important when it comes to analysis of mapping populations, which comprise dozens and hundreds of cereal lines. Analyses like these should take as little time to complete as possible while the plants are at an identical stage of development and are not affected by environmental factors.

The proposed method for preparing wheat leaf folds and photography under a light microscope for further computer analysis considerably reduces the time required to assess quantitative characteristics describing wheat leaf pubescence. Photomicroscopy using a properly set up microscope takes less than 1 min to complete; digital imaging and image processing take a matter of few seconds. Our experience suggests that it takes a single operator only 1 h to complete the recommended wheat leaf folding and processing procedures on ten plants (~60 images). It should be noted that while the phenotyping techniques based on 3D scanning are not invasive (Kaminuma et al. 2008; Bensch et al. 2009), ours is, which prevents its use for studying trichomes in development.

Analysis demonstrates that a significant correlation exists between the number of trichomes per wheat leaf fold line length unit and the number of trichomes per wheat leaf surface area unit. This correlation allows various geometric properties of trichomes visible on leaf fold photomicrographs to be considered as reliable leaf pubescence characteristics.

Noteworthy, our approach affords rather a high performance when counting trichomes: the F score is 0.85 and the mean absolute percentage error is 11.65 %. This implies that, given from 30 to 40 trichomes in an image, the departure of their estimated number from their actual number is three or four trichomes. Our experience shows that the error is lower for specimens with a low number of crossing trichomes. At the same time, LHDetect2 performs better on composite images than LHDetect (Doroshkov et al. 2009) (Table 1). Furthermore, LHDetect2 warrants a high accuracy of trichome length evaluation, with the error being about 6 %.

The proposed method allows the researcher to determine the ratio between shorter and longer trichomes on the leaf surface and to use multivariate statistical analysis methods in comparisons of quantitative characteristics of leaf pubescence. This advantage became clear as we compared leaf pubescence between cultivar Hong-mang-mai and line 102/00i, both having similarly long trichomes on their leaf blades. In a previous study (Dobrovolskaya et al. 2007), no glabrous plants were found in the F2 population, which suggested that both cultivars were allelic to the Hl2 gene controlling the development of long trichomes. In this work, we have found that the Hong-mang-mai × 102/00i F2 plants are heterogeneous for trichome number. However, a better understanding of the mechanisms underlying the genetic control of leaf pubescence requires further testing on a larger number of progeny.

The use of cluster analysis on the length distribution vectors demonstrates their usefulness for a more detailed division of progeny into classes, because it takes account of not only the total trichome numbers, but also of the number of trichomes with each particular length.

Interestingly, we have not observed any correlation between trichome number and trichome length in the genotypes in question, while these measures have been found correlated in some Russian spring wheats (Doroshkov et al. 2011). This is perhaps because the mechanisms underlying trichome growth and setting trichome length limitations in the genotypes we have used in this work may be not the same as in the cultivars that we used previously.

We have also demonstrated that some F2 plants develop longer trichomes than their parental genotypes. This fact might be indicative of an additive effect of the genes for leaf pubescence in Hong-mang-mai and 102/00i. A statistically significant identification of longer trichomes in Hong-mang-mai × 102/00i F2 plants has only become possible with LHDetect2. This observation deserves further research with a view to understanding the mechanisms of leaf pubescence development during cereal plant ontogenesis.

We have also demonstrated that LHDetect2 is a method that provides adequate information not only on trichome dimensions, but also on the distances between trichomes. In particular, the distribution of the distances between the closest long trichomes that was produced by LHDetect2 indicates that long trichomes are largely confined to leaf veins.

Although LHDetect2 cannot rival scanning electron microscopy (Perazza et al. 1999) or 3D analyses (Kaminuma et al. 2008; Bensch et al. 2009) with their brilliantly detailed description of leaf pubescence morphology, it can help rapidly identify the most interesting cases of the inheritance of morphological traits with a view to further analysis using more accurate yet labor-consuming techniques.

Two major genes controlling leaf pubescence have been identified and mapped to date (Taketa et al. 2002; Dobrovolskaya et al. 2007; Pshenichnikova et al. 2007). One more gene has been found in the diploid wheat Triticum monococcum ssp. boeoticum without chromosome assignment (Sharma and Waines 1994). In distant relatives of wheat, a gene for leaf pubescence has been found in the barley Hordeum vulgare ssp. spontaneum on chromosome 3HL (Franckowiak 1997). It is possible that even more genetic factors influencing quantitative characteristics of leaf pubescence in cereals exist. We believe that the proposed method for leaf pubescence quantification in wheat will be helpful in the determination of the mechanisms of the genetic control of this trait.

Acknowledgments

Analyses were performed using equipment at the Interinstitutional Shared Center for Microscopic Analysis of Biological Objects. This work was supported by programs Б.25 and A.II.6 from the Russian Academy of Sciences, Scientific School 5278.2012.4, and RFBR grant 11-04-91397. We are thankful to Vladimir Filonenko for translating this manuscript from Russian to English, and to Mikhail Ponomarenko for helpful comments.

Supplementary material

425_2012_1751_MOESM1_ESM.pdf (160 kb)
Supplementary material 1 (PDF 160 kb)
425_2012_1751_MOESM2_ESM.pdf (246 kb)
Supplementary material 2 (PDF 245 kb)
425_2012_1751_MOESM3_ESM.xls (90 kb)
Supplementary material 3 (XLS 90 kb)
425_2012_1751_MOESM4_ESM.xls (86 kb)
Supplementary material 4 (XLS 85 kb)
425_2012_1751_MOESM5_ESM.pdf (221 kb)
Supplementary material 5 (PDF 263 kb)

Copyright information

© Springer-Verlag 2012