Investigation of Continous Imaging Analysis of Grain Quality on a Combine Harvester Janine Berberich 1 *, Hilke Risius 2, Markus Huth 3, Jürgen Hahn 4 1 Humboldt University of Berlin, Faculty of Agriculture and Horticulture, Department of Crop and Animal Sciences, Divison of Biosystems Engineering, Philippstr.13, 10115 Berlin, Germany *Corresponding author. E-mail: janine.berberich@agrar.hu-berlin.de Abstract EU legislation establishing limits and sampling plans for mycotoxins has come into force recently. Thresholds for mycotoxins emphasise the necessity of food safety monitoring at the beginning of the grain processing chain. The availability of rapid detection methods of mycotoxin contamination in agricultural commodities is still limited. Thus improved methods for grain quality monitoring and processing should be established. Imaging techniques in combination with NIRS are expected to be applicable for the inline analysis of Fusarium spp. and mycotoxin contamination. Key words: On-combine monitoring, Imaging Techniques, Grain Quality 1 Introduction Mycotoxins are secondary metabolites of fungi. The major fungal species producing mycotoxins include Aspergillus, Fusarium and Penicillium. Wheat kernels infected with Fusarium head blight show a wrinkled coat, a broadened crease, and the appearance is pink and chalky white. Rapid methods for the analysis of mycotoxins (Zheng, Richard et al. 2006) are still unavailable for in-stream sensing of grain quality. Work on early detection of Fusarium infection in wheat using hyper-spectral imaging was reported by Bauriegel, Giebel et al. (2010; 2011), however the availability of sensors in practice is limited. The use of Near Infrared Spectroscopy (NIRS) for the determination of mycotoxins in cereals has been described by Pettersson and Aberg (2003), Quilitzsch and Schütze (2006) and Rasch, Kumke et al. (2010). Publications show that NIRS is suitable both for real-time analysis of grain quality oncombine and for grain flow control based on defined quality thresholds (Risius, Hahn et al. 2008; Risius, Hahn et al. 2010; Risius, Hahn et al. 2011) (Fig. 1). The objective of the study described herein is to investigate the feasibility of imaging techniques in combination with NIRS for the detection of Fusarium spp. infestations and mycotoxin contamination (Pettersson and Aberg 2003; Kamaranga Peiris, Pumphrey et al. 2009). The image processing software is calibrated both to detect colour alterations for the monitoring of Fusarium spp. infestation and to obtain data for the assessment of spectral data quality in future.
Figure 1 On-Combine Grain Quality Analysis and Processing: NIRS, Imaging Techniques and Sampling System and Online Grain Monitoring and Quality Processing 2 Material and methods 2.1 Grain samples Individual wheat kernels are selected manually according to alterations in kernel colour. The grain samples were taken during field trials in 2011. The samples were stored dark, dry and airproof to avoid metabolic changes during storage. 2.2 Experimental Setup The experimental setup consists of a colour camera (D5000, Nikon), a zoom lens of 70-300 mm focal length (Nikon AF-S NIKKOR 70-300 mm 1:4,5-5,6G ED VR) and a micro conversion lens (DCR 250, raynox) mounted to a telescopic base with a 40 W light source on both sides. In order to avoid casting of shadows both angle and height of the light sources are adjustable. Single kernels were fixed with blue coloured modelling material. The colour contrast enhances the detection of pink colour alterations by the image processing software. Images of both upper and lower surface of each individual kernel are taken and consecutively numbered. An infrared remote-control release is used to avoid definition lack as a result of vibration.
Figure 2 Experimental Setup 2.3 LabVIEW-Vision Builder The software includes two parts: a group of programs for image acquisition and a software package for image analysis. The image analysis software is the essential part in experimental procedure. National Instruments Vision Builder for Automated Inspection (Vision Builder AI) is configurable software for building, benchmarking, and deploying applications and does not require programming. A built-in deployment interface is included for rapid identification application. The ability to set up complex pass/fail decisions to control digital I/O devices and communicate with serial or Ethernet devices is also included. For system calibration, a database is set up to identify wheat kernels with pink colour alterations. Therefore a query is generated in the Vision Builder to display all available images of the wheat kernel sample set. The first decision procedure allows for a qualitative determination of colour alterations. A Region of Interest (ROI) is defined to limit image analysis to the kernel surface as well as to decrease computing capacity requirements. The classification algorithm 1 is calibrated to identify pink colour alteration in the defined ROI. Low variations in colour and structure are neglected. The search routine is closed at a maximum amount of 100 pink discolourations and an Excel spreadsheet is generated automatically. The spreadsheet indicates the amount of discolorations of each sample. In addition, charts show the recognition rate and a corresponding recognition rate threshold. The computing procedure is stopped herewith and the classification process is repeated for of the training observations. In addition to qualitative analysis, a quantitative analysis is implemented. The second classification algorithm is based on several templates and includes additional classification criteria of pink colour alterations, modelling material and image background. The individual templates are associated with unique categories to allow for the classification of image background, modelling material and healthy and damaged wheat kernels respectively. Results are transferred to an Excel spreadsheet as well. In this case, the ratio of colour features (red, grey, blue and kernel-coloured is computed. A chart indicates the classification accuracy.
1 6 11 16 21 26 31 36 41 46 51 56 61 66 71 76 81 86 91 96 Total of Wheat Kernel Samples = 100 % 3 Results and discussion 3.1 Image Analysis The first classification procedure to identify colour alterations of wheat kernels (Figure 3) selected 100 colour features (Figure 4) with a recognition rate of nearly 70 %. Figure 3 Image of Wheat Kernel before Processing Figure 4 Image of Wheat Kernel after Processing The dependency of recognition rate and the selected threshold emphasises the necessity of an accurate colour feature selection. Figure 5 shows that merely a low number of kernels were misclassified. 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% Threshold: Number of Selected Colour Features Figure 5 Results of Imaging Analysis, Classification Method 1(N=96) False Positive False Negative True Negative True Positive Recognition Rate The identification of colour alterations using classification method 2 (Figure 6) shows promising results though computation time requires optimisation. Figure 6 Classification Method 2 Analysis Results: ROI, 1 = Modelling Material; 2 = Healthy Wheat Kernel; 3 = Pink Colour Alterations
1% 6% 11% 16% 21% 26% 31% 36% 41% 46% 51% 56% 61% 66% 71% 76% 81% 86% 91% 96% Total of Wheat Kernel Samples = 100 % In comparison to classicfication method 1, Figure 7 shows a sufficient recognition rate at a percentage of 11 % of colour alterations. 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% False Positive False Negative True Negative True Positive Recognition Rate Threshold: Number of Selected Colour Features Figure 7 Results of Imaging Analysis, Classification Method 2(N=96) 3.2 Discussion The key aspect of colour alteration classification methods presented in this study is colour information. The identification of wheat kernel colour alterations is predominantly depending on subjective decisions whereas these decision criteria depend on various factors. In addition to camera settings, the colour rendering of the monitor is essential. The use of different monitors showed different classification of colour alterations, though monitor calibration can minimize this lack of accuracy. Fusarium pathogens cause yield and quality losses due to sterility of the florets and formation of discoloured, withered and light test-weight kernels and also produce significant levels of deleterious mycotoxins including trichothecenes (Goswami and Kistler 2005). Therefore, the kernel geometry has to be investigated as infected grain significantly differ from healthy grain in size, surface structure and shape (Williams 2006). An algorithm for the detection of shriveled grain has to be developed as well. In combination with the classification method for detecting discolouration, it could provide a qualitative assessment of Fusarium infestation of grain. Since no uniform structure of discolouration could be detected yet, it is important to consider possible changes in analysis results if i.e. the rotation of single templates is disabled and thus differences in the structure are prevented. To this effect, experiments show significantly different results from those with active rotation. Since recognition rates differ only marginally, this aspect is of secondary importance. Whether the reproducibility of results is depending on the hardware used needs further investigation. 4 Outlook Based on current scientific knowledge, the lack of rapid detections methods for mycotoxin producing fungi in the processing of grain could be solved by combining Near-Infrared Spectroscopy and imaging techniques, since NIRS has been evaluated for the real-time determination of mycotoxin contamination of cerals by observing changes in protein, carbohydrate
and moisture content and imaging techniques are appropriate to detect discolorations and differences in size of grain kernels. First results of the laboratory experiments show promising results regarding analysis of the grain stream for Fusarium spp. Image processing in addition to NIRS analysis is both applicable for the monitoring of spectral data quality and therefore for the analysis of chemical properties of grain by Near Infrared Reflectance Spectroscopy (NIRS). In future, an optimised on-combine analysis of grain quality by means of NIR analysis and imaging techniques will be established for the segregation of grain into fractions of high and low quality. Acknowledgments This cooperative research project is funded by the Federal Agency for Agriculture and Food (Bundesanstalt für Landwirtschaft und Ernährung, BLE) of the Federal Republic of Germany. Reference list Bauriegel, E., A. Giebel, et al. (2010). "Early detection of Fusarium infection in wheat using hyper-spectral imaging." Computers and Electronics in Agriculture 75(2): 304-312. Bauriegel, E., A. Giebel, et al. (2011). "Hyperspectral and Chlorophyll Fluorescence Imaging to Analyse the Impact of Fusarium culmorum on the Photosynthetic Integrity of Infected Wheat Ears." Sensors 11(4): 3765-3779. Goswami, R. S. and H. C. Kistler (2005). "Pathogenicity and In Planta Mycotoxin Accumulation Among Members of the Fusarium graminearum Species Complex on Wheat and Rice." Phytopathology 95(12): 1397-1404. Kamaranga Peiris, H. S., M. O. Pumphrey, et al. (2009). "NIR absorbance characteristics of deoxynivalenol and of sound and Fusarium-damaged wheat kernels." Journal of Near Infrared Spectroscopy 17(2): 213-221. Pettersson, H. and L. Aberg (2003). "Near infrared spectroscopy for determination of mycotoxins in cereals." Food Control 14(1): 229-232. Quilitzsch, R. and W. Schütze (2006). Comparison of spectroscopic and chromatographic methods for mycotoxin determination in samples of winter wheat. 3rd Int. Seed Health Conference, Bydgoszsc, Polen. Rasch, C., M. Kumke, et al. (2010). "Sensing of Mycotoxin Producing Fungi in the Processing of Grains." Food and Bioprocess Technology 3(2): 1-9. Risius, H., J. Hahn, et al. (2008). Near Infrared Spectroscopy for Sorting Grain according to Specified Quality Parameters on a Combine Harvester. 67th International Conference on Agricultural Engineering LAND.TECHNIK AgEng, Stuttgart-Hohenheim. Risius, H., J. Hahn, et al. (2010). Monitoring of grain quality and segregation of grain according to protein concentration threshold on an operating combine harvester. Book of Abstracts XVII.th World Congress of the International Commission of Agricultural and Biosystems Engineering (CIGR SCGAB), Québec City, QC,Canada Risius, H., J. Hahn, et al. (2011). In-line-Sensing of Grain Quality on a Combine Harvester. 69th International Conference on Agricultural Engineering LAND.TECHNIK AgEng 2011, Hannover, Germany, VDI-Verlag. Williams, P. (2006). Near-Infrared Spectroscopy of Cereals. Handbook of Vibrational Spectroscopy. J. M. Griffiths and V. R. Chalmers, John Wiley & Sons, Ltd. Zheng, M., J. Richard, et al. (2006). "A Review of Rapid Methods for the Analysis of Mycotoxins." Mycopathologia 161(5): 261-273.