Ana səhifə

Defra pvsd funded research final report


Yüklə 72.27 Kb.
tarix24.06.2016
ölçüsü72.27 Kb.

NIAB

DEFRA PVSD FUNDED RESEARCH – FINAL REPORT


PROJECT TITLE:

Development of the Image Analysis system as a powerful tool for determining distinctness in Oilseed Rape (Brassica napus L.)

PROJECT START DATE:

01/04/04


PROJECT END DATE:

31/03/06


NIAB PROJECT LEADER:

Carol Norris/ David Smith


PROJECT COSTS:

£30K


_____
SUMMARY OF PROJECT ACTIVITIES AND RESULTS:
Further Development of Image Analysis for Determining Distinctness in Oilseed Rape


  1. Introduction

Winter Oilseed Rape is one of the largest of the DUS testing workloads at NIAB with over 2000 plots sown in Autumn 2005, including over one hundred candidate varieties. Currently Image Analysis (IA) of oilseed rape is performed on cotyledons only. A preliminary evaluation of the importance of existing characteristics and their ability to distinguish varieties has been carried out using historical data. The IA-measured characteristics have been found to be particularly discriminatory for the DUS process and the use of this technology has greatly increased accuracy and precision since being introduced, and has therefore improved the establishment of distinctness of new varieties. A further statistical evaluation has now been completed to include the combined additional characteristics from cotyledon measurements carried out by IA. Results of these show mostly high F1 values and lambda values near to 1 for the combined characteristics (Table 1). High F1 values show good discriminatory powers and lambda values of near to 1 show little variation between years.


The objectives of this project are to extend the use of IA to allow more precise measurement of oilseed rape pods, petals and leaves in the same way as has been proven with cotyledon measurements. The introduction of IA in these areas could considerably improve current procedures and lead to a more robust testing approach, which would be both more efficient and more accurate. A computerised and more streamlined system could also lead to better discrimination between varieties and automation of processes is likely to increase throughput of sample recording.
Table 1: Table showing F1 and lambda values of characteristics currently recorded on WOSR with IA cotyledon characteristics highlighted in blue (anther spot and growth habit are no longer recorded).

To achieve these objectives the following points were investigated:




  • The use of IA for measurement of pod, flower and leaf characteristics.

  • Assessment of the practicalities of using IA as a routine method for these characteristics.

  • Whether IA for these measurements will improve the accuracy and efficiency of recording methods.

  • Improvements to the current imaging software.

  • Statistical evaluation of the effectiveness of characters measured by IA in determining distinctness.



  1. Camera set up and standardisation



Camera set-up
Prior to sampling plant material, the use of an Olympus E-1 SRL digital camera was investigated This was mounted on a Raiser® stand, which allows the camera to be adjusted to any height. The camera height was set at the appropriate height above the sample according to the object being photographed and the following camera settings were used:
Film setting Super High Quality

Contrast Default

Shape Default

Saturation CS3

SQ 1280 x 960 1/8

AF Illumination ON


White Balance 6000K

Colour Space RGB

Zoom setting Auto Focus

Ring Setting 20

Lens Zuiko digital 14-54mm f2.8-3.5

Zuiko digital 50mm 1:2 macro


A 1cm black and white scale was fixed to the light box underneath the camera and a colour scale placed on the right hand side (Annex 1, Figure 1).

Colour Standard

Inconsistencies in the colour of digitally captured images stem largely from:




  1. Differences in the colour response between different digital cameras

  2. Fluctuations in ambient lighting conditions

Until recently IA at NIAB has used grey-scale images. Images for DUS testing purposes are now starting to be captured in colour, for example the slide collection of Chrysanthemum images that has been digitised to create a database for reference varieties (‘Digimum’). It is important therefore to ensure consistency in the capture of colour images by using a suitable standard colour reference with every image.


The most thorough method for measuring colour would be to use a reflectance spectrophotometer, which measures the intensity of the reflected light at a number of frequencies. While this approach is desirable, it would be time-consuming, expensive and perhaps excessive for our purposes. Our approach is to include several colour standards in each photo (Figure 1). As all imaging is carried out under controlled conditions, differences in colour within the images of each crop are relatively small, and the inclusion of colour standards allows these small differences to be accounted for. Furthermore, they provide a measure of the consistency of the imaging process and evidence that any differences in the colour of crop samples between images are genuine. The colour reference used in colour matching for ornamental crops at NIAB is the Royal Horticultural Society (RHS) colour chart. After taking photos of the 76 colours at 3 white balance settings, it was determined that the following colours should be used at 6000K white balance:
1a yellow, 43a red, 96a blue, 140a green, 82a magenta and 125a cyan plus greys 202a, 202b, 202c and 202d. To reduce any effects of fading caused by light exposure, the cards were covered when not in use.
The greys were chosen to give roughly equal spacing from very dark to very light, and the colours were chosen for their high saturation and proximity to “pure” red, green etc. in terms of the response of the camera and the resulting pixel values.
At present colour assessment has not been developed for use with any of the oilseed rape characteristics measured by IA as it was not within the remit of the project, however by incorporating a colour chart into every image this will allow the option to use colour assessments in the future after further development.

Scale

In the current DUS IA system, the scale of captured images is measured by using reference objects. In this process, a circular object of known size is imaged before crop sample images are taken. While this gives an accurate scale measurement, it means that the scale and sample images are separate and discrepancies cannot be re-assessed. The new system includes a grid type scale in every image, in both the vertical and horizontal directions, so the scale is tied to the sample image.


3. SOFTWARE DEVELOPMENT
All imaging software was developed by Bob Farrell (PsiSoft Scientific Programming, Glasgow UK). The Olympus Studio software was developed to allow capture of images to be taken directly from the computer by the press of a button. The user has the option to select a species and the type of image to be captured and analysed. Each species and type of image has specific camera settings associated with it which are automatically set when the species and type are selected. A “quick preview” option is available to allow the captured image to be viewed before continuing with a large batch of image capturing. Once a plot number is entered then the setting automatically changes to “high resolution” capture and a photograph is taken by the camera via the computer. The image is then stored in a pre-selected folder (“Source Images” is set as default) and as a .tif file. Before analysing a batch of images, the files must be backed up manually. Images can be analysed singly or in large numbers allowing batch analysis to be carried out overnight or without the need for an operator. The software allows the user to select which files are to be analysed and gives an option to automatically close an image once it has been analysed. This ensures that a batch of images can be analysed without the computer running out of memory. The Analyse Image software allows the analysed images to be viewed and checked for outliers or errors. If a single sample in an image is outside the normal range for that variety, or two of the samples are touching, the analysed image will highlight the sample as an outlier. The user can then choose whether to accept the outlier as part of the data set. After checking for outliers the data file is then updated.
For analysis of pods, the batch analysis is carried out and there is an option available to line up the samples in the image on-screen. The beak and the peduncle must be then be identified manually by clicking on the image in the correct position. This option was developed due to difficulties encountered in identification of the beak and peduncle areas on the image by programming. Further work would be needed to automate recognition of the beak and peduncle as surface features of the image.
The data output from the analysed images is then stored as .IA files and the combined characteristics stored as .DER files. These are then ready for collation and routine statistical analysis.

4. IA OF LEAVES

Sampling of leaves

Plant material was taken from the DUS hybrid backup trial at Thornaugh, near Peterborough. A minimum of 22 leaves from 3 replicates of 80 varieties were picked from oilseed rape plants at growth stage 6-12 leaves and the 5th leaf selected from each plant. Leaves picked included the whole petiole and were fully expanded with no discoloration or damage. Leaves were placed in labelled plastic bags and refrigerated until measurements were taken.


Manual measurement of leaves
Leaf width (measurement between the two widest points) and total leaf length (mm), including stalk, was measured from 20 leaves from each plot using a standard cm ruler and the data recorded onto Allegro dataloggers. Data were downloaded and collated for statistical analysis.
IA of leaves
After manual measurements were taken, the same leaves were laid out dorsal side up onto white card (with an identifying plot number) and placed onto the light box with the colour standard chart (Annex 1, Figure 4). Digital photographs were then taken of each plot of 20 leaves and analysed using the Image Analysis software.

Results
Leaves proved to be the most problematic plant part to image due to their size and turgidity. It was not possible to fit 20 leaves onto one image so the programme was modified to accept several images for one plot as an option. Leaves had to be left for an hour to wilt so that they could be laid flat and didn’t create a shadow which the computer programme would measure as part of the leaf edge.
The analysis programme measures length and width (experimental characteristics in the CPVO TP 36/1: Figure 2), and the following novel combined characteristics:


  1. Area

  2. A/ (length x width)

  3. WTOL (width to length ratio)

  4. BTWPTOL (base to wide point/length)

  5. BTCTOL (base to centre/length)

  6. Shape (4 x Pi x area/perimeter/perimeter)

  7. Symmetry (2 x area of overlap/total area - Figure)

Leaf data were analysed statistically using a student’s T test for comparisons at the 5% level between the hand measurements and the IA measurements. Table 2 shows numbers of pairs separated for the manual method and for the IA method. For both length and width the IA method separates fewer variety pairs than the manual method. Some of the shape measurements show potential for having good discrimination power and separate many more pairs than the absolute measurements.



Table 2: Student’s T test used in comparisons of manual and IA methods of leaf measurement at 5% level (based on a sample set of 80 varieties)


MANUAL




IMAGE ANALYSIS




Characteristic

No. of pairs of varieties separated

Characteristic

No. of pairs of varieties separated

4 llen

667

4 llen

661

5 lwid

861

5 lwid

723







90 Area

714







91 A/(l*wd)

695







92 WTOL

1147







93 BTWPTOL

1211







94 BTCTOL

900







95 Shape

1110







96 Symmetry

723




Figure 1: OSR leaf sample image with scale and colour standards.



Figure 2: Analysis of leaves showing length and width measurements



  1. IA OF PETALS


Background
Following a visit in June 2004 to the Bundessortenamt in Germany, methods for displaying petals to the camera were investigated. IA is routinely used to measure petal characteristics at the Bundessortenamt, and a method of sticking petals onto self-adhesive folex photographic paper has been successfully developed. Oilseed rape petals are small, very fragile and can dry out quickly so the method of fixing has to be fast and relatively easy to carry out. Fixing the petals and covering them with a sheet of adhesive plastic stops them from drying out, allowing them to be stored for over a year. This will provide flexibility at a busy time of year, and the possibility of taking measurements at a later date if necessary.
Method
Plant material was taken from the DUS hybrid backup trial at Thornaugh, near Peterborough. A minimum of 22 flowers from 3 replicates of 20 varieties were picked from oilseed rape plants at growth stage. Flowers picked were fully opened with no discoloration or damage. Flowers were placed in labelled plastic bags and refrigerated until measurements were taken.
20 petals were measured by hand according to methods for measuring routine characteristics on the current oilseed rape UK protocol and recorded onto Allegro dataloggers. The same undamaged petals were then placed on a blue background (Figure 3) on self-adhesive folex photographic paper. The reason for using blue as a background is because it is the opposite of yellow on the chromatic scale and provides the best contrast for the images. These were then photographed using the same process and camera set-up as used with the leaves, with a slight adjustment to camera height. Initial tests indicated that reflection from the covering plastic material was a problem when analysing the images, however this was overcome by using Folex matt translucent film which is less reflective.

Results
Measuring petals by IA proved to be the most straight-forward of the plant parts to develop for imaging due to their similarity in shape and size to cotyledons for which routine IA was developed in the 1990s. The method of presenting the petals took approximately the same amount of time as measuring them manually. The analysis programme measures petal length and width (required characteristics in the CPVO TP 36/1), and the following novel combined characteristics:


  1. Area

  2. A/ (length x width)

  3. WTOL (width to length ratio)

  4. BTWPTOL (base to wide point/length)

  5. BTCTOL (base to centre/length)

  6. Shape (4 x Pi x area/perimeter/perimeter)

  7. Symmetry (2 x area of overlap/total area – Figure 4)

Petal data were analysed statistically using a student’s T test for comparisons at the 5% level between the hand measurements and the IA measurements. Table 1 shows numbers of pairs separated for the manual method and for the IA method. The IA method separates more varieties for the measurements “petal width” and “petal length” than the manual method, and the new shape characteristics show potential for separating more varieties, particularly “width to length ratio”.



Table 3: Student’s T test used in comparisons of manual and IA methods of petal measurement at 5% level (based on a sample set of 20 varieties)


MANUAL




IMAGE ANALYSIS




Characteristic

No. of pairs of varieties separated

Characteristic

No. of pairs of varieties separated

52 petl

64

52 petl

106

53 petw

123

53 petw

144







90 Area

119







91 A/(l*wd)

85







92 WTOL

157







93 BTWPTOL

99







94 BTCTOL

93







95 Shape

63







96 Symmetry

54




Figure 3. OSR petals showing length and width measurements. One outlier has been identified.



Figure 4: Calculation of symmetry of an OSR petal.

6. IA OF PODS
Method
Plant material was taken from the DUS hybrid backup trial at Thornaugh, near Peterborough. A minimum of 22 pods from 3 replicates of 20 varieties were picked from oilseed rape plants at growth stage. Pods picked were fully developed with no discolouration or damage. Pods were placed in labelled plastic bags and refrigerated until measurements were taken.
20 pods were measured by hand according to methods for measuring routine characteristics on the current oilseed rape UK protocol and recorded onto Allegro dataloggers. The same pods were then placed on double-sided sticky acetates. These were then photographed using the same process and camera set-up as used with the leaves, with a slight adjustment to camera height.
Results
Measuring pods by IA presented a few problems. Pods are more three-dimensional structures than leaves or petals, and often when they are measured by hand they have to be stretched straight to get a true measurement. To overcome this problem the line of length measurement was calculated down the midline of the pod (Figure 5), which is then smoothed slightly to remove discontinuities. The width of the pod was calculated at the point as the diameter of the largest circle that can be drawn inside the pod. Presenting the pods to the camera for imaging is much quicker than hand measuring, however the pods must be secured to the paper to prevent movement, and care must be taken to present each pod in the same way. Pods are also prone to splitting especially when left to dry out, so imaging needs to be carried out quickly. The analysis programme measures total length, pod length, width, beak and peduncle length (identified by manually clicking on the relevant areas on the image): (required characteristics in the CPVO TP 36/1) and the following combined characteristics (not yet numbered):
Total length

Area


WTOPL (width to pod length ratio)

WTOTL (width to total length ratio)



Shape (4 x Pi x area/perimeter/perimeter)
Pod data were analysed statistically using a student’s T test for comparisons at the 5% level between the hand measurements and the IA measurements. Table 2 shows numbers of pairs separated for the manual method and for the IA method. Pod length and pod width both show increased discrimination in the IA method over the manual method. Some of the pods measured by IA had damaged beaks due to excessive handling as they had already been hand measured, and this may be reflected in the results of the beak length measurements which show that IA separates fewer pairs than the manual method. Beak length and peduncle length were both measured in the IA method by manually highlighting the correct area on the image. Accuracy of this judgement may be improved with experience.

Table 3: Student’s T test used in comparisons of manual and IA methods of pod measurement at 5% level (based on a sample set of 20 varieties)


MANUAL




IMAGE ANALYSIS




Characteristic

No. of pairs of varieties separated

Characteristic

No. of pairs of varieties separated

62 Pedlen

84

62 Pedlen

87

64 Beakllen

52

64 Beakllen

45

61 Podw

18

61 Podw

23

63 podlen

73

63 podlen

85

91 podshape

63

91 podshape

65




-

Total length

72




-

Area

52




-

Width to pod length

92




-

Width to total length

71



Figure 6: Pods showing width and length measurements on a chequer-board scale


7. CONCLUSIONS

Leaves

Although leaf width and length are currently not listed as routine characteristics in the CPVO oilseed rape technical protocol, they have been put forward as ‘Research Characteristics’ to be used in an experiment in which several Member States will participate to review the usefulness of these characteristics. The importance of the leaf characteristics is therefore currently less than those of the petals and pods as it is not yet certain as to whether leaf characteristics will be included in the CPVO technical protocol.




Manual measurement and imaging of leaves both presented problems. If the characteristics are taken up as routine, then measuring by IA may be more accurate than measuring by hand but may not be any quicker. Measurement of leaves by hand is crude and results may not reflect true values as it is difficult to judge the widest and longest points by eye and the leaves have to be flattened before measuring. The size of the leaves can vary enormously and the area of workspace taken up with laying the leaves out for imaging is large. Fewer pairs of varieties were separated by IA than by the manual measurements. This is likely to be because of the three-dimensional qualities of the leaves, and at the time that the experiment was carried out there were problems with a shadow effect where the leaf stands proud from the background. This has since been rectified by leaving the leaves to wilt before imaging. Modifications have also been made to the programme whereby the central vein of the leaf is located (Figure 2) and measured rather than taking a straight measurement from the highest to lowest point.

Petals

Petals proved to be the most successful application of IA and will be taken up for routine use in Spring 2006. The IA method is no quicker than the manual method but the results of the pair separation comparison show that the IA method separated far more varieties than the manual method. Measurements previously taken by hand were taken to the nearest mm so the greater accuracy of the IA measurements is no surprise. There is an additional benefit of the seven new combined characteristics which could be used if proved to be useful in discriminating between varieties and found to be biologically meaningful and accepted by the CPVO.


Pods
Pods will also be measured by IA this year. This will be the most cost-effective use for the IA system as hand-measuring pods is a laborious and long process requiring several members of staff over several weeks. Time taken to present the pods is much reduced with the IA system and the batch analysis of the samples can take place overnight if necessary, saving time during working hours. The UK is now the most advanced country of the CPVO Member States in development of pod measurement by IA for DUS of oilseed rape. Measurement of pods by IA will reduce the time taken to record the whole DUS trial by more than a third and at the same time increase the accuracy of the measurements and therefore discrimination between varieties.


Page of 13 14/03/06


Verilənlər bazası müəlliflik hüququ ilə müdafiə olunur ©atelim.com 2016
rəhbərliyinə müraciət