CC BY
33
Технологии и технические средства механизированного производства продукции растениеводства и животноводства_
15 января 2015 г. руководителем ФАНО и президентом РАН утвержден Регламент взаимодействия по осуществлению и развитию международного научного и научно-технического сотрудничества. Однако взаимодействие согласно данному документу затруднительно. Аналогичные проблемы и с публикациями научно-исследовательских разработок в зарубежных журналах.
В заключение хотелось бы отметить, что инвестиции в эту сферу обещают высокую окупаемость (30-75%) и долгосрочную выгоду.
КОРНЕЛ САЛАИ, PhD, научный сотрудник; ЛАСЛО ФЕНВЕЗИ, профессор, PhD, научный руководитель, Национальный центр сельскохозяйственный исследований и инноваций, Институт сельскохозяйственной инженерии
ПРЕДВАРИТЕЛЬНЫЕ РЕЗУЛЬТАТЫ КЛАССИФИКАЦИИ ПШЕНИЦЫ НА ОСНОВЕ ОСОБЕННОСТЕЙ СПЕКТРАЛЬНЫХ СВОЙСТВ
Много-и гиперспектральные приборы уже давно рассматриваются как инструменты неразрушающего распознавания и анализа с минимальными затратами времени и средств. Актуальность их применения в современном сельском хозяйстве в лабораторных и полевых условиях, а также в сельскохозяйственной авиации растет быстрыми темпами.
Измерение в широком спектре электромагнитного излучения в массиве соседних спектральных полос дает возможность описать и определить различные материалы, растительные покровы и сорта растений по их спектральным характеристикам. Теоретически в достаточной степени чувствительная и точная система может идентифицировать и различить даже определенные сорта растений.
В данной статье авторы представляют предварительные результаты проекта, в котором изучались особенности спектра пяти сортов озимой пшеницы с целью разработки процедуры идентификации, основанной на использовании модели.
Ключевые слова: спектральный анализ, сорта пшеницы, классификация
K. SZALAY, PhD, associate researcher; L. FENYVESI, Prof., PhD, research supervisor
ISSN 0131-5226. Сборник научных трудов. _ИАЭП. 2015. Вып. 87._
PRELIMINARY RESULTS OF SPECTRAL FEATURE BASED WHEAT CLASSIFICATION
Multi- and hyperspectral tools have long been considered as time and cost-effective, non-destructive detection and analysis tools. The relevance of their laboratory, in-field and airborne application in modern agriculture rapidly grows.
The wide range of EM radiation detected on several hundred adjacent wavelength bands provides the opportunity to describe and identify different materials, vegetation covers and plant species upon their spectral characteristics. In theory, an adequately sensitive and accurate system can identify and separate even certain plant varieties.
In this paper the authors are introducing the preliminary results of a project where five winter wheat variety's spectral features have been studied to develop a model-based identification procedure.
Key words: spectral analysis, wheat varieties, classification Introduction
Hungary had been one of the biggest seed exporters of Europe for a long time. However its role still important, the area and scale of seed production decrease. Due to the high seed prices which originate in high production costs farmers are less and less concerned about purchasing and sowing certified seeds. As a result of using uncertified seeds quality and quantity of crops often drop below expectations. The lessening demand towards certified seeds threatens the plant breeding and seed production activity and further increases the prices.
There has been a need for a quality insurance system that can balance the licence fee of certified seed, seed production costs, seed prices and bread wheat prices.
A possible solution for regulating the certified seed utilisation is a monitoring system that can be operated on large agricultural areas in economic way.
Remote sensing technologies have the potential to develop a wheat variety identification procedure which can form the basis of a national regulation system.
An important fundament of non-destructive analytical technologies is the spectroscopy, which studies the interaction between electromagnetic radiation and matter (Figure 1).
Технологии и технические средства механизированного производства продукции растениеводства и животноводства_
У Scattering
Irradiation
Figure 1: EM interactions (NOAA Coastal Service Center (2007) modified) Рис. 1. Электромагнитные взаимодействия (Центр береговой службы Национального управления исследования Мирового океана и атмосферы (2007)
с изменениями)
The method of evaluating the spectral characteristics of different biotic or abiotic materials and surfaces originates in the laboratory spectroscopy, where it is generally used in physical and analytical chemistry hence atoms and molecules have unique spectra. Emission is characteristic to atomic bonds while absorption refers to molecular bonds. It provides opportunity to obtain quantitative relationships between the environmental and physiological parameters of organic or non-organic samples, plant parts, soil quality parameters and the features of reflectance spectra (Csorba 2012; Tolner 2011;Tarnawa et al. 2011) (Figure 2).
Figure 2: Distinctive spectral features of different samples Рис. 2. Отличительные особенности спектральных характеристик различных
образцов 19
Characteristic wavelengths can indicate changes in moisture content and other relevant parameters (Kaszab et al. 2008;Milics et al. 2010). Spectral evaluations were proved to be useful in detecting and diagnose certain fungal diseases (Firtha et al. 2008;Neményi 2013). There are studies which proved that spectral information can be used to identify or separate certain plant species or characteristic varieties (Turza et al. 1998; Yong et al. 2007; Esteban-Diez et al. 2004; Seregély et al. 2004; He et al., 2005). Though, the processing of multi- orhyperspectralimages is a very complex procedure (Milics et al. 2008).
Materials and Methods
On the experimental production sight of SzentIstván University, Faculty of Agriculture and Environmental Sciences homogeneous wheat plots were sowed by using genetically uniform and certified seeds. Alfold 90, MvCsárdás, MvMagdaléna, MvSuba, MvToborzó varieties were examined at three different level of nutrient supply (0, 80, 120 kg/ha N broadcasted at tilleringphenofase). The differentiated nutrient broadcast resulted in wide range of nutrient-reactions, yield and quality of wheat. Infield measurements were carried before harvesting.
An ASD FieldSpec 3 Max portable spectroradiometer was used. The device extends the range of visible light (Williams et al. 2010;Lágymányosi and Szabó 2009) to NIR (near infrared) and the SWIR (shortwave infrared) region and covers the range of 350 to 2500 nm (Lágymányosi and Szabó 2011).
Figure 3: In-field spectral measurement. Measuring position is perpendicular to the
incident sunlight.
Рис. 3. Полевые спектральные измерения. Точка измерения перпендикулярна
падающему солнечному свету.
Технологии и технические средства механизированного производства продукции растениеводства и животноводства_
Five spectral libraries were built from varieties containing 60 spectra each. After performing a complex mathematical pre-processing on the dataset a PLS DA (Partial Least Squares Regression Discriminant Analysis - executed under Mathlab PLS Toolbox environment) classification procedures were used. In this process known classes (varieties) were created and defined for the algorithm to find characteristic statistical signatures adequate for variety-based classification.
One individual spectrum contains 2151 variables (reflectance factor measured on 2151 bands) so the number of variables must be reduced in order to identify those few which describes the best the variance in the dataset - the difference between predefined classes (varieties). The PLS DA algorithm also contains a dimensionality reduction. As part of the dimensionality reduction the algorithm generates a number of latent variables (LV) which creates the statistical correlation between the classes and sample spectra.
Model calibration was performed on the dataset to see if the varieties can be separated statistically. The model was cross-validated on a statistically divided sample set by using venetian-blind cross-validation method.
Results
300 spectra were collected under field conditions. Despite of the atmospheric absorption resulting in disturbance around 1450, 1850 and 2450 [nm] the dataset proved to be appropriate for the further analysis (Figure 4).
Figure 4: Reflectance spectra collected in-field Рис. 4. Спектры отражения, полученные в полевых условиях
21
Among the variables generated by the PLS DA algorithm five LV were selected. These five LV described nearly 15 [%] of the total variance in the dataset, resulting minimal classification error in case of each classes (varieties) during the cross-validation (Table 1).
Table 1: Number of latent variables used in the classification
Таблица 1
Число скрытых переменных, использованных в классификации
\ X-Block Captured [%1 X-Block Total [%1 CV Class. Err. Alföld 90 CV Class. Err. Mv Csardas CV Class. Err. Mv Magdalena CV Class. Err. Mv Suba CV Class. Err. Mv Toborzö
LV 1 8,97 8,97 0,285 0,269 0,406 0,444 0,019
LV2 2,02 10,99 0,071 0,021 0,263 0,412 0,025
LV3 1,68 12,67 0,000 0,025 0,033 0,363 0,021
LV4 1,46 14,13 0,000 0,000 0,000 0,025 0,021
LV5 0,66 14,79 0,000 0,000 0,000 0,033 0,004
Based on the selected latent variables the authors could calibrate a model that fits on the predefined classes without classification error and high R2 values (Table 2). The self-test of the model was flawless (Table
3).
Table 2: Statistical result of self-test
Таблица 2
_Статистический результат самотестирования_
Alföld 90 Mv Csârdâs Mv Magdaléna Mv Suba Mv Toborzo
Class. Err. (Cal) 0,000000 0,000000 0,000000 0,000000 0,000000
RMSEC 0.0782365 0.0835833 0,0859306 0,0934484 0,0914896
Bias -3.05311e-016 -3.88578e-016 -2.77556e-017 4.996e-016 1.66533e-016
R Calibration 0.961744 0,956336 0,95385 0,945421 0,947685
Table 3: Self-test of model
Таблица 3
_Самотестирование модели_
Confusion Table (Calibration/Self-test)
___________________ Alföld 90 Mv Csârdâs Mv Magdaléna Mv Suba Mv Toborzo
Predicted as Alföld 90 60 0 0 0 0
Predicted as Mv Csardas 0 60 0 0 0
Predicted as Mv Magdalena 0 0 60 0 0
Predicted as Mv Suba 0 0 0 60 0
Predicted as Mv Toborzo 0 0 0 0 60
Технологии и технические средства механизированного производства продукции растениеводства и животноводства_
The most expressive illustration of sample distribution was reached through representing the spectra in a three-dimensional space described by LV 2, LV 4 and LV 5. The distribution of varieties is rather promising (Figure 5).
Figure 5: Distribution of samples in a three-dimensional data space defined by three
selected latent variables Рис. 5. Распределение выборки в трехмерном пространстве данных, определенное по трем выбранным скрытым переменным
After the self-test every fifth sample was selected, removed from calibration dataset and used as a validation dataset to check the model classification. The statistical results (Table 4) of cross-validation indicated low classification error in case of MvSuba variety which resulted in two false classifications (Table 5).
Table 4: Statistical result of cross-validation
Таблица 4
_Статистический результат перекрестной проверки_
Class. Err. (CV) Alföld 90 0,000000 Mv Csardas 0,000000 Mv Magdaléna 0,000000 Mv Suba 0,033333 Mv Toborzo 0,004167
RMSECV 0,112048 0,131951 0,133935 0,181933 0.171339
CV Bias 0.00204009 -0.0077842 0.00137336 -0.00736565 0.0117364
R2 Cross-validation 0,941248 0,900749 0,919763 0,820191 0,810157
Table 5: Cross-validation of the model
Таблица 5
_Перекрестная проверка модели_
Confusion Table (CV)
~ .................. Alföld 90 Mv Csardas Mv Magdaléna Mv Suba Mv Toborzo
Predicted as Alfold 90 60 0 0 0 0
Predicted as Mv Csardas 0 60 0 0 0
Predicted as Mv Magdalena 0 0 60 0 0
Predicted as Mv Suba 0 0 0 58 0
Predicted as Mv Toborzo 0 0 0 2 60
Conclusions
However results were validated on a dependent data divided from the original data set the study shows promising achievement that underpins the theoretical possibility of variety classification and the existence of variety-specific spectral signatures. In-field spectral measurements and spectral library based PLS DA classification method can have the potential in identifying different varieties. In the future it is recommended to test the procedure on independent datasets. Acknowledgement
Authors would like to thank the generous support of the National Agricultural Research and Innovation Center Institute of Agricultural Engineering and the Univesity of Szentlstvan. The processing of the spectral data was carried out at the External Agro-Technical Department of University of Szentlstvan.
REFERENCES
1. Csorba A, LangV., FenyvesiL., MicheliE. (2012): Reflektanciaspektroszkopiaalkalmazasatalajokszervesszen- esCaCO3-tartalmanakbecsleseben. AgrokemiaesTalajtan. 6/2, 277-290. o.
2. Esteban-DiezJ., Gonzalez-SaizM., PizarroC. (2004): Anevaluationoforthogonalsignalcorrection methods for the characterisation of arabica and robusta coffee varieties by NIRS. AnalyticaChimicaActa, 514(1), 57-67 pp.
3. He Y., Li X. L., Shao Y. N. (2005): Quantitative analysis of the varieties of apple using near infrared spectroscopy by principal component analysis and BP model. Lecture Notes in Artificial Intelligence, 3809, pp. 10531056.
4. Kaszab T., Firtha F., Fekete A. (2008) Influence of Non-ideal Storage Conditions on Carrot Moisture Content Loss. Progress in Agricultural Engineering Sciences 4, 61-75 pp. DOI: 10.1556/Progress.4.2008.4
24
Технологии и технические средства механизированного производства продукции растениеводства и животноводства_
5. Lagymanyosi A. and Szabo I. (2009) Calibration procedure for digital imaging, Synergy and Technical Development, Godollo, Hungary, 30. Synergy2009 CD-ROM Proceedings
6. LagymanyosiA., and Szabo I. (2011) Investigation of chaff surface with imaging (In Hungarian). FiatalMtiszakiakTudomanyosUlesszaka XVI., KolozsvarMtiszakiTudomanyosFuzetek 177-180 pp.
7. Milics G., Burai P., Lenart Cs. (2008): Pre-Harvest Prediction of spring barley nitrogen content using hyperspectral imaging. Cereal Research Communications, AkademiaiKiado, Volume 36, 1863-1866 p. Proceedings of the VII. Alps-Adria Scientific Workshop.Szlovakia, StaraLesna. 2008. aprilis 28. - majus 2.
8. Nemenyi M. (2013): Investigation of Fusarium Ear Rot Symptoms on Maize (Zea Mays L.) Using a Spectroradiometer. In: Nemenyi M., Varga L., Facsko F., Lorincz I. (Eds.) "Science for Sustainability" International Scientific Conference for PhD Students, Sopron: Nyugat-magyarorszagiEgyetemKiado, 2013. pp. 4 (ISBN:978 963 334 103 2)
9. SeregelyZs.,Deak T., BisztrayGy. D. (2004): Distinguishing melon genotypes using NIR spectroscopy. Chemometrics and Intelligent Laboratory Systems, 72 (2), pp. 195-203.
10. Tarnawa A., Nyarai, F., Mate A. (2011) Statistical assessment of climatic impacts on the nitrogen nutrition of maize (Zeamayze L.) crop. 10 th Alps-Adria Scientific Workshop. Novenytermeles.Opatija, Croatia 207210 pp.
11. Tolner I., T, (2011) Investigation of the effects of various acid treatments on the optical reflection spectra of a soil sample. 10 th Alps-Adria Scientific Workshop. Novenytermeles.Opatija, Croatia 207-210 pp.
12. Turza, S., Toth, A., Varadi, M. (1998).Multivariate classification of different soybean varieties.In Davies (Ed.), Proceedings of the 8th international conference AMC. Chichester, UK: NIR Publications pp. 183187.
13. Williams P.C., Manley M., Fox, G. esGeladi, P. (2010): Indirect detection of Fusariumverticillioides in maize (Zea maize L.) kernels by NIR hyperspectral imaging. Journal of Near Infrared Spectroscopy, 18, pp. 49-58, 2010.
14. Yong H., Xiaoli L., Xunfei D. (2007): Discrimination of varieties of Tea Using Near Infrared Spectroscopy by Principal Component Analysis and BP Model. Journal of Food Engineering, 79(4): pp. 1238-1242.
