Method for classifying seeds

Abstract

The present invention relates to a method for classification and/or sorting of seeds with the help of terahertz time-domain spectroscopy, use of terahertz time-domain spectroscopy for classification and/or sorting of seeds and seeds classified and/or sorted with terahertz time-domain spectroscopy.

Claims

1. A method for classifying seeds by using radiation in the terahertz range, comprises: (a) applying an input terahertz pulse to the seed; (b) measuring a signal generated by the input terahertz pulse after transmission and/or reflection by the seed; (c) comparing the measured signal with the input terahertz pulse to determine a difference; (d) determining one or more of the parameters based on the difference between the measured signal and the input terahertz pulse, the parameters being selected from the group consisting of a change in amplitude, a time delay in passing through the seed, additional echo pulses as a result of reflections within the seed, and field oscillations following the input terahertz pulse; (e) assigning the seed to a predetermined seed class based on the one or more parameters obtained in step (d) as compared to a corresponding reference value.

2. The method according to claim 1, wherein the method has a sorting accuracy of at least 75%.

3. The method according to claim 1, wherein the method is automated.

4. The method according to claim 1, wherein the terahertz pulse has a duration in the range of 3 picoseconds (ps) to 25 ps.

5. The method according to claim 1, wherein the terahertz pulse has a delay of 0-104 ps with an interval of 0.05 ps.

6. The method according to claim 1, wherein the seed is assigned to the predetermined class based on a determination of an embryo status of the seed, and wherein the embryo status of the seed is determined based on the one or more parameters determined in step (d).

7. The method according to claim 6, wherein the embryo status of the seed is determined based on the change in amplitude, the time delay in passing through the seed, the additional echo pulses as a result of reflections within the seed, and the field oscillations following the input terahertz pulse.

8. The method according to claim 1, wherein the terahertz radiation of the terahertz pulse has a frequency of 0.1 THz-10 THz.

9. The method according to claim 8, wherein the terahertz radiation of the terahertz pulse has a frequency of 0.1-2 THz.

10. The method according to claim 1, wherein the method involves the use of a terahertz time-domain spectrometer.

11. The method according to claim 10, wherein the seed is sorted according to its predetermined seed class after taking it out of the measurement range of the terahertz time-domain spectrometer.

12. The method according to claim 10, wherein the seed is calibrated prior to classification.

13. The method according to claim 10, wherein the method comprises a step of introducing the seed into the measurement range of the spectrometer prior to step (a) and/or a step of taking the seed out of the measurement range of the spectrometer after step (d) and/or after step (e).

14. The method according to claim 1, wherein the seed is selected from the group consisting of vegetable seeds, cereal seeds, grains of pitted fruit, grains of berries, nuts, and grains of Amaranthaceae family.

15. The method according to claim 14, wherein the seed is the seed of the genus selected from the group consisting of Beta, Allium, Helianthus, and Capsicum.

16. The method according to claim 15, wherein the seed is from the species selected from the group consisting of Beta vulgaris, Allium cepa, and Helianthus annuus.

Description

(1) Designs and embodiments of the present invention are described in an exemplary form with respect to the accompanying figures, in which:

(2) FIG. 1: shows time-domain spectra of sugar beet seeds. This figure shows the average spectra of the respective classes.

(3) FIG. 2: shows time domain spectra of sugar beet seeds. Detail of FIG. 1 in the range of 12-24 ps. This figure shows the average spectra of the respective classes.

(4) FIG. 3: shows time domain spectra of sugar beet seeds. Detail of FIG. 1 in the range of 45-70 ps. This figure shows the average spectra of the respective classes.

(5) FIG. 4: shows phase spectra of sugar beet seeds. This figure shows the average spectra of the respective classes.

(6) FIG. 5: shows phase spectra of sugar beet seeds. Detail of FIG. 4 in the range of 0.2-2 THz. This figure shows the average spectra of the respective classes.

(7) FIG. 6: shows absorption coefficient spectra of sugar beet seeds. This figure shows the average spectra of the respective classes.

(8) FIG. 7: shows absorption coefficient spectra of sugar beet seeds. Detail of FIG. 6 in the range of 0.1-0.6 THz. This figure shows the average spectra of the respective classes.

(9) FIG. 8: shows refractive index spectra of sugar beet seeds. This figure shows the average spectra of the respective classes.

(10) FIG. 9: shows refractive index spectra of sugar beet seeds. Detail of FIG. 8 in the range of 0.1-2 THz. This figure shows the average spectra of the respective classes.

(11) FIG. 10: main component analysis of sugar beet seeds: score diagram of the main components 1 and 2.

(12) FIG. 11: shows time-domain spectra of onion seeds. This figure shows the average spectra of the two classes.

(13) FIG. 12: shows time-domain spectra of onion seeds. Detail of FIG. 11 in the range of 12-24 ps. This figure shows the average spectra of the two classes.

(14) FIG. 13: shows time-domain spectra of onion seeds. Detail of FIG. 11 in the range of 45-70 ps. This figure shows the average spectra of the two classes.

(15) FIG. 14: shows time-domain spectra of pepper seeds. This figure shows the average spectra of the two classes.

(16) FIG. 15: shows time-domain spectra of pepper seeds. Detail of FIG. 14 in the range of 12-24 ps. This figure shows the average spectra of the two classes.

(17) FIG. 16: shows time-domain spectra of pepper seeds. Detail of FIG. 14 in the range of 45-70 ps. This figure shows the average spectra of the two classes.

(18) FIG. 17: shows the main component analysis of onion seeds: score diagram of main components 1 and 2.

(19) FIG. 18: shows the main component analysis of pepper seeds: score diagram of main components 1 and 2.

(20) FIG. 19: shows the time-domain spectra of sunflower seeds. This figure shows the average spectra of the two classes.

(21) FIG. 20: shows the main component analysis of sunflower seeds: score diagram of main components 1 and 2.

SEED SAMPLES

(22) The quality of the seed classification and sorting method according to the invention is demonstrated below on seeds of various types of crops, such as sugar beet (B. vulgaris), onion (A. cepa), pepper (Capsicum) and sunflower (H. annuus). The evaluation of the suitability of the method according to the invention is based on an X-ray diagnostic procedure performed previously in the laboratory to classify the seeds. The classification achieved based on the X-ray diagnostic procedure is referred to below as the actual class.

(23) Sugar Beets:

(24) The seeds of the sugar beet are divided into 6 calibers based on the grain geometry and grain density. The seed grains of each caliber are subjected to X-ray diagnostics classifying the individual seed grains into the classes 1 through 6 described above. Class 1 is referred to below as empty, class 2 as full, class 3 as slightly shrunken, class 4 as severely shrunken, class 5 as twin and class 6 as bigerm.

(25) Onions and Peppers:

(26) Seeds of the onion and pepper varieties are divided into two calibers based on the grain geometry and grain density for each variety. The seed grains are also subjected to an X-ray diagnostic procedure. In the case of onion and pepper it is advisable to classify the seeds into the classes full and shrunken, where the class shrunken could contain grains of the classes empty, slightly shrunken and severely shrunken.

(27) Sunflower:

(28) In contrast with the two varieties of vegetables described above, with sunflowers there is no classification of the seeds in calibers. The seed grains were also subjected to the X-ray diagnostic procedure. As was the case with onion and pepper, it is advisable in the case of sunflower to classify the seeds in the classes full and shrunken or shriveled.

(29) Specifications for the Terahertz Spectrometer:

(30) The measurements were performed with the TERA K15 terahertz time-domain spectrometer from Menlo Systems GmbH, Martinsried, Germany. The TERA K15 spectrometer consists of a laser source with fiber optics, a delay zone with a scanning range of 0-300 ps, a TERA15-TX-FC terahertz emitter antenna and a TERA15-RX-FC terahertz receiver antenna fitting with the former, terahertz-capable optics, a lock-in TERA-C amplifier and a computer with a measurement program for data acquisition and analysis.

(31) The spectrometer offers a dynamic range of >70 dB and a spectral range of >3.5 THz. The built-in femtosecond laser is linearly polarized, emits at 1560 nm with a repeat frequency of 100 MHz with an average laser output power of >60 mW. The pulse length after 2.5 m fiber optics is <90 fs.

(32) To measure the seed grains of all varieties of vegetables, the following measurement settings and parameters are selected: a spectral measurement range of 0.01-10 THz and a delay of 0-104 ps with an interval of 0.05 ps. The integration time pro delay position is 30 msec, which yields a total measurement time of approx. 1 minute per seed grain.

(33) Performing the Terahertz Measurements:

(34) For the measurement, a seed grain as described above is placed in the focal point of the terahertz beam. A terahertz pulse is then applied to the seed grain. The terahertz pulse is generated by the transmitter antenna, passes through the lens, is focused and interacts with the seed grain. The detector antenna records the signal generated by the terahertz pulse after transmission and/or reflection by the seed grain. Multiple terahertz pulses are typically are applied to the seed grain, even with different delay times. The signal of the detector antenna is read out by a computer using a measurement program and the transmission-determined and/or reflection-determined amplitude, the time delay, the phase and/or the spectrum of the signal is/are determined. In additional steps the absorption coefficient and the refractive index of the seed grain can be calculated.

(35) Evaluation and Results

(36) In the following embodiment the method for classification is carried out using terahertz radiation on the basis of a caliber.

(37) Sugar Beets:

(38) FIGS. 1-3 shows the time-domain spectra of the seed of an exemplary caliber (3.16-3.50 mm), where an average spectrum has been formed for each class. FIG. 1 shows the complete spectrum, but FIGS. 2 and 3 show details to illustrate differences in the spectra. The spectra of the six classes show definite differences, for example in the range of 15-22 ps.

(39) Furthermore the spectra of the phase, the absorption coefficient and the refractive index also disclosed differences in a comparison of the classes within a caliber, as shown in FIGS. 4-9. FIGS. 4, 6 and 8 each show the complete spectrum and FIGS. 5, 7 and 9 each show an enlarged detail of the phase, the absorption coefficient and the refractive index.

(40) The information on the time domains, the phase, the absorption coefficient and the refractive index and/or combinations thereof can be used for the classification. In the embodiment in the present case, for example, only the time-domain spectrum is used for classification; the situation is similar with the other types of fruit listed below.

(41) To illustrate the differences in the time-domain spectra and to further investigate the data, the analysis is calculated by performing a main component analysis (PLS Toolbox Version 7.9.1, Eigenvector Research, Inc., Wenatchee, United States of America based on Matlab 2014b, The MathWorks GmbH, Ismaning, Germany). FIG. 10 shows the score diagram of the main components 1 and 2. The classes empty, slightly shrunken, greatly shrunken, twin and bigerm occupy the main component space in quadrants II, III and IV and are thus clearly differentiated from the classes full and another group of the class empty in quadrant I. The boundary between the classes full and empty in quadrant I is not sharp, i.e., some samples of these two classes may overlap.

(42) To emphasize the differences in the data, the raw data is optimized for further processing for example by means of Savitzky-Golay [method] (A. Savitzky; M. J. E. Golay (1964). Smoothing and Differentiation of Data by Simplified Least Squares Procedures; Analytical Chemistry 36 (8): 1627-1639) or moving average smoothing filter (The Scientist and Engineer's Guide to Digital Signal Processing, Chapter 15Moving Average Filters, by Steven W. Smith, Nov. 17, 2001), wherein the smoothing can also be supplemented with a derivation. In the second step of data processing the entire data set is split into two independent data sets by means of the Kennard-Stone algorithm (R. W. Kennard, R. W. and L. A. Stone (1969) Computer aided design of experiments. Technometrics 11(1), 137-148) (also PLS Toolbox Version 7.9.1, Eigenvector Research, Inc. Wenatchee, United States of America based on Matlab 2014b, The MathWorks GmbH, Ismaning, Germany). The data selected using the Kennard-Stone algorithm is then used to create the classification model and the remaining data is validated with this data set.

(43) Various algorithms such as the k-nearest neighbor algorithm (abbreviated KNN) can be used for the classification (N. S. Altman (1992), An introduction to kernel and nearest-neighbor nonparametric regression, The American Statistician 46 (3): 175-185. doi:10.1080/00031305.1992. Ser. No. 10/475,879)) or the support vector machine (abbreviated SVM) (N. Cristianini and J. Shawe-Taylor (2000), An Introduction to Support Vector Machines and Other Kernel-Based Learning Methods, First Edition (Cambridge: Cambridge University Press)) may be used for the classification.

(44) The evaluation of the classification is assessed by means of the following quality factors: true positive, true negative, false positive and true positive. Since, as described above, the separation of the class full from the remaining classes is of primary concern, then the quality factors are defined as follows: in the case of true positive, seed assigned to the actual class full is correctly assigned to the class full; in the case of a true negative, seed of another actual class is correctly not assigned to the full class; in the case of a false negative, seed from the actual full class is falsely assigned to another class; and with a false positive seed from another actual class is falsely assigned to the full class.

(45) Table 1 shows the classification of the seed grains from the validation data set. The classification achieves quality factors of 100% with a true positive result, 100% with a true negative, 0% with a false positive and 0% with a false negative, based on the full class. If all the classes are included, then 89% of all grains are classified correctly. This result shows that sugar beet seed grains within one caliber can be classified according to the present invention with a high quality.

(46) TABLE-US-00001 TABLE 1 Result of the classification of sugar beet seeds by the method according to the invention (=predicted class). Actual class slightly greatly shrunken empty shrunken s full twin bigerm Predicted Total 4 11 10 32 37 6 class empty 4 0 0 0 0 0 slightly 0 9 3 0 0 0 shrunken greatly 0 1 7 0 0 0 shrunken full 0 0 0 32 0 0 twin 0 1 0 0 37 3 bigerm 0 0 0 0 0 3
Onion and Pepper:

(47) As already described, only the information about time domains will be discussed here as an example. The time-domain spectra of the onion and pepper seeds, each of an exemplary caliber, are shown in FIGS. 11-16, where an average spectrum has been formed for each class. FIGS. 11 and 14 each show the complete spectrum of the onion seeds and/or pepper seeds. However, FIGS. 12 and 13 for onion seeds and FIGS. 15 and 16 for pepper seeds represent details of the complete spectra illustrating the differences in the spectra.

(48) The data preprocessing and analysis of the onion and pepper seeds were performed as done with the sugar beet. First, a main component analysis was performed, then the data was optimized using a smoothing filter and possibly a derivation as also described above and the respective data set was divided into two independent data sets by means of the Kennard-Stone algorithm. However the main component analysis of the pepper seeds was not calculated with the raw data but instead was calculated with the data that had already been optimized.

(49) The diagram for the score of the main component analysis of the onion seeds is shown in FIG. 17. There is a recognizable separation between the two classed full and shrunken but the data points for the two groups are situated close together. The situation looks similar in the main component analysis of pepper, represented in FIG. 18. Here again, there is a recognizable differentiation in the two classes full and shrunken but the two groups of data points are also close together.

(50) The fact that the signal differences in between the onion and pepper seeds turn out smaller than those with the sugar beets can probably be attributed to the size/geometry of the seeds and their embryos. Onion seeds are much smaller than the sugar beet seeds, while the pepper seeds are definitely flatter.

(51) A classification of the seeds of onions and peppers is possible by means of the method according to the invention. Table 2 shows the classification of the validation data set of onion seeds. The quality of the classification is 87.5% true positive, 12.5% false negative, 14.3% false positive and 85.7% true negative. On the whole, 86.7% of the onion seed grains are classified correctly. The classification of the validation data set of the pepper seeds is shown in table 3. All the seed grains were classified correctly, with 100% true positive quality factors, 0% false negative, 0% false positive and 100% true negative. On the whole 100% of the pepper seed grains were classified correctly.

(52) TABLE-US-00002 TABLE 2 Results of classification of onion seeds by the method according to the invention (=predicted class). Actual class shrunken full total 7 8 Predicted shrunken 6 1 class full 1 7

(53) TABLE-US-00003 TABLE 3 Results of classification of pepper seeds with the method according to the invention (=predicted class). Actual class shrunken full total 5 8 Predicted shrunken 5 0 class full 0 8
Sunflower:

(54) The time domain spectra of the sunflower seeds are shown in FIG. 19. Since the seeds were not subdivided according to caliber in the case of the sunflower seeds, FIG. 19 shows the average spectra of the two classes.

(55) The data preprocessing and analysis were performed in the same way as for the sugar beet, pepper and onion seeds. However, the main component analysis of the sunflower seeds was not calculated using the raw data but instead using the data that had already been optimized. FIG. 20 shows the score diagram of both main components 1 and 2. Here again, we see fundamentally a separation in the two classed full and empty but the adjacent data points overlap.

(56) The classification of the validation data set of the sunflower seeds is shown in table 4. The quality of the classification is 98.9% true positive, 1.1% false negative, 0% false positive and 100% true negative. A total of 99.1% of the sunflower seed grains were classified correctly.

(57) TABLE-US-00004 TABLE 4 Results of the classification of sunflower seeds with the method according to the invention (=predicted class). Actual class shrunken full total 23 91 Predicted shrunken 23 1 class full 0 90

(58) These results show that the method can be used not only for sugar beet seeds but also for seeds from other types of fruits, such as those shown here. In addition, it has been shown that depending on the properties of the grains, a classification without a prior division according to caliber is also possible.