Prototype of a Vision Based System for Measurements of White Fly Infestation

Transcription

1 Prototype of a Vision Based System for Measurements of White Fly Infestation C. Bauch and T. Rath Institute of Horticultural and Biosystems Engineering, University of Hannover Herrenhäuser Str. 2, D Hannover Germany Keywords: image analysis, pest density, white fly, automatic pest detection, selective crop protection Abstract To better respond to ecological and economical aspects of plant production than has been the case in the past, plant protection measures have to be better adapted to reflect actual pest densities in cultivated plant populations. Because the currently applied methods to determine entomological pest densities are time-consuming manual methods, there is a general need to automate this process. A system has been developed to measure the density of an entomological pest, the white fly (Trialeurodes vaporariorum and Bemesia tabaci), within plant stands (Lycopersicon lycopersicum). The system combines automatic pest insect extraction with a computer image analysis unit, and has now reached the prototype state. A mobile aspiration mechanism with a newly developed mechanical filter system is applied to collect adult white flies. The collection unit can extract pest individuals from the plant population and feed them to an optical recognition system. Particles which are loosely attached to the upper section of the plant stand are aspirated, filtered out of the air stream, put on a conveyor belt and transported in front of a color CCD camera. Digital image analysis is then performed to classify the captured objects into "target organisms" (white fly) and biotic or abiotic "error items" (foreign organisms, dust, plant parts). Shape and color parameters serve as differentiation criteria. Finally, these characteristics are compared to a classification dataset for the white fly and then evaluated. INTRODUCTION Selective crop protection is one of the basic requirements in horticulture and agriculture. To better respond to ecological and economical aspects of plant production, specific situational phytosanitary measures which can be adapted to the existing pest density stand in the center of attention. In practice, the cultivator determines the density of entomological pests. He evaluates the crop at several selected points determining the degree of infestation. The need for crop-protection measures is determined for the entire area based on the gained data. According to varying environmental conditions, pests tend to a heterogeneous distribution within the agricultural acreage (Begon et al., 1998). This heterogeneity often creates a patch-like appearance of the pest. As a result, a defined damage threshold is in many cases only exceeded at certain points in the plant population. An over all, nonsituational measurement of crop protection, as currently common in practice, is still far from being the ecological and economical optimum. Based upon this fact, there is an urgent need to adapt phytosanitary measures to the actual, regional existing pest density. The aim of the discussed project is to work out a methodical approach, which makes automatic detection of the infestation density possible by means of computer image analysis. The focus is set on the development of a suitable recognition system. In an initial trial, the system s functioning in pest detection, identification and quantification has been verified. Proc. IC on Greensys Eds.: G. van Straten et al. Acta Hort. 691, ISHS

2 MATERIALS AND METHODS The developed system combines an automatable, classical scoring method (Southwood and Henderson, 2000) with a computer image analyzing unit. In order to realize this approach it is based upon the following theoretical process sequence (Fig. 1). In the first step, by means of an aspiration mechanism, an unknown quantity of pest insects and particles having a low attaching force have to be detached from the upper section of the crop, aspirated, and then transported in front of a vision input module. The collected objects have to be pictured and then, by using digital image analysis, be classified in "target organisms" (pest insects) and biotic or abiotic "non target organisms" (dust, plant parts). Finally the collected data have to be converted into a quantitative statement about the actual existing pest density. Pests For this experiment the white fly (a mixed population of Trialeurodes vaporariorum and Bemesia tabaci) serves as entomological pest. The criteria for this selection were: availability of a wide entomological and phytosanitary knowledge bank, a horticultural relevant pest with a high level of damage potential, significant morphological appearance of the adult individuals, simple management, little mechanical anchoring of the adult individuals in the crop. Experimental Crop The tomato (Lycopersicon lycopersicum var. Vollendung) was selected as experimental crop because of its horticultural relevance. The individual plants have been cultivated in a greenhouse in 10 l containers containing mineral-peat soil. A specific characteristic of the culture management is the fact that the plants were not pinched off. At an average plant height of approx. 40 cm, the shoot axis of the principal axis is removed. By means of this cultivation measure, a bushy habit and therefore a denser crop grow. The experiment started at an average plant height of approx. 50 cm. Experimental Setup The tomato plants are arranged in rows and grouped by five. Each group is covered with a 2.0 m long, 1.0 m high, and 0.7 m wide cage. The cage construction itself is covered with an LS Econet SF protective insect screen produced by Ludvig Svensson Company, which represented an entrance and escape barrier for the adult individuals of the white flies. This constructive configuration makes it possible to create individual factor levels with different pest densities. During the first experiment, the pest density factor is classified as follows: no infestation (0 individuals/cage), medium infestation (75 individuals/cage), high infestation (150 individuals/cage). At five plants per cage there is an average pest density of 0, 15, and 30 individuals per plant with four repetitions per factor level. The experiment starts two days after having spread the adult white flies to ensure a lifelike individual distribution of the pest in the crop. To guarantee the reproducibility of the automated scoring, each aspiration process is performed following a specified pattern. In order to simplify the access to the entire cage, each cage has two long openings in the roof. Via these openings the aspiration unit is put 60 cm deep into the crop and is then, at a constant height, carefully moved x- wise through the upper section of the crop. The duration of the movement per track is fixed to 4 sec. Immediately after the fist x-wise aspiration process, a second likewise movement is performed. Then, the second half of the cage is treated that way. 774

3 The aspirated adult white fly individuals are counted manually and by computer image analysis to be able to make a statement about the pest acquisition and quantification performance of the developed system. RESULTS Collection Unit Figure 3 shows the experimental setup of the collection unit prototype. The developed system can be subdivided into seven functional assembly groups: 1. lower vacuum chamber 2. upper vacuum chamber 3. aspiration tube 4. conveyor belt made of filter fabric 5. progressive-scan color camera with macro lens 6. LED illumination 7. photoelectric sensor for triggering The aspiration tube of an industrial vacuum cleaner (Nilfisk, 1000W) is connected to the lower vacuum chamber, generating a vacuum by drawing off air. It is possible to compensate the caused deficit of air volume by in taken air which flows through a 2 cm wide and 15 cm long channel in the tightly sealed base plate. Thereby, a vacuum is produced in the upper vacuum chamber which is as well almost air tightly connected to the base plate. The air volume required for this pressure compensation continues flowing via a 3 m plastic tube with a pipe diameter of 10 cm into the upper chamber. A bellshaped nozzle with a diameter of 18 cm and a depth of 10 cm is fixed to the extreme end of the pipe in which an aspiration effect is produced by the inflowing air. The aspiration tube nozzle is inserted into the crop as described above and moved through the upper section of the crop performing a well-defined, x-shaped movement. Due to the slight movements of the leaves some adult white flies are stimulated to leave the leaves and are aspirated by the system. The effective aspiration power of the system is sufficient enough to extract and pick up the loosely attached insects and particles from the upper section (10 15 cm) of the crop. The entrained objects cross the upper vacuum chamber and reach the system s transport unit, an air permeable conveyor belt with a particle filter function construction. The conveyor belt consist of a black, very fine-meshed insect screen (laboratory model) produced by Ludvig Svensson Company and can filter particles with a diameter of > 0.5 mm out of the inflowing air stream. The filtered objects are pressed to the conveyor belt by the dominating aspiration power that goes through the belt and are transported with a constant speed of 8 cm/s out of the upper vacuum chamber passing the vision input device. The belt movement causes the objects to leave the fixing vacuum zone and they are actively removed from the conveyor belt by a brush mechanism. A 1-chip CCD color camera with progressive scan mode, type DxD 4023 (Theimagingsource, 2001), is used as optical acquisition unit. The used resolution is 646 x 515 pixels and the color bit depth is 24 bits. Therefore it is possible to perform the image acquisition with a frequency of up to 3 fps. A photoelectric sensor type E3T-FD13 (Omron, 2001) triggers the image acquisition. The markers used to actuate the trigger signal are placed in intervals of 27.8 mm (maximum linear expansion of the captured image area) on the conveyor belt. Picturing of the entire conveyor belt surface, even during varying conveyor speeds, is ensured by this mode of image capturing. A zoom lens with a focal length of mm in combination with a 20 mm focal length extension is applied as macro lens. The captured area has a dimension of 27.8 x 23 mm. Four LED arrays are used to illuminate the area. Each array consists of 12 white LEDs (Liteon, 2002), with light output of 7000 mcd each. The LED arrays arranged in a ring around the macro lens are positioned at a height of 4 cm over the acquisition area. In order to 775

4 suppress any possible extraneous light sources the system is covered with light proof material. Image Analysis The gained image data are transmitted to a PC and buffered for the following digital computer image analysis. The relevant objects are segmented from the buffered picture series. The segmenting follows a suitable color space transformation using statistical grey-scale thresholds. The use of fixed threshold values is possible because of the artificial light system and the black conveyor belt that provide a constant object background contrast. The object classification has been implemented using shape and color parameters. Only the object area is used as shape parameter. The use of the absolute area as criterion for the object classification is possible because of the constant object resolution inherent in the construction design. Additional shape parameters (e.g. contour length, convexity, etc.) cannot be used since the shape of the adult white flies varies significantly after the impact on the conveyor belt (Fig. 2). It has been observed that the position of the wings in relation to the bodies is very unequal. In some cases, even the loss of wings has been detected. The ratios of the average grey-scale values of red/green, red/blue and green/blue channels in the RGB color model serve as color parameter. Therefore, the typical whiteyellow object coloration has been included as a decision criterion during object classification. The results shown in Table 1 have been gained by applying a combination of the above mentioned classification parameters. According to these results, 83.1% of the white flies were classified correctly. On average, 16.9% of the adult white files were not recognized as target objects during automatic counting. The type 2 error (non-target classified as white fly) was approx. 2%. Performance of the Acquisition Unit The results of the computer image analysis and the manual count over the factor levels are shown in Figure 4. It is recognizable that the number of counted individuals escalates with increasing infestation densities. Significant differences between the factor levels of the high and medium infestation densities, respectively zero level, could be proven. A significant difference between the factor level medium and no infestation could not be shown. Figure 4 also illustrates that the manual and computer assisted image evaluation led to comparable results. DISCUSSION Collection Unit During the evaluation of the recorded image data, it has been observed that some adult white flies lost their wings. This circumstance can be derived from the fact that the forces generated by the air turbulence in the suction device and the sudden impact of the insect on the filter conveyor belt are very strong. Since these modifications of the object shape highly complicate the automatic object detection, it will be important to optimize this point in future investigations. Image Analysis The shown classification results are based on the combination of shape and color parameters. The resulting recognition rate of 83% for correctly classified white flies is not yet sufficient. Given the case of a low infestation degree and a resulting little number of aspirated individuals, possible erroneous classification will lead to disproportional large errors in subsequent pest density quantification. Therefore, additional research has to be performed to examine the suitability of the texture parameters and morphological characteristics of adult white flies in order to improve the automatic object recognition. 776

5 Performance of the Acquisition Unit In general, it is proven that the presented system is able to differentiate between various infestation degrees. The fact that no significant difference between medium and no infestation factor level could be exhibited, can be explained by the low number of repetitions and the therefore resulting high standard deviations. Furthermore, the presented results allow for the statement of an existing relation between the collected number of adult white flies and the existing pest density. The comparability of these results has to be proven in future researches. Subsequent experiments should additionally research the quantitative relation between the number of individuals that are detected by the system and the actual pest infestation density. The smallest possible pest density that can still be detected by the system is yet another point that has to be researched on. CONCLUSIONS In conclusion, it is shown that using this detection unit it is possible to extract a representative number of individuals from the crop. By means of the computer image analysis, the aspirated pest insects can be automatically detected and counted, and it is possible to differentiate between the various pest densities. Consequently, with this system it is possible to show a relation between actual pest densities in the crop and the automatically acquired pest insects. Literature Cited Begon, M.E., Harper, J.L. and Townsend, C.R Ökologie. Spektrum Akademischer Verlag Heidelberg, Berlin. Southwood, R. and Henderson, P.A Ecological Methods. Blackwell Science, Oxford. Tables Table 1. Number of objects for the manual count (MM) and the automatic count (AM) method. MM AM target non target Σ target non target Σ

6 Figurese Fig. 1. Process steps. Fig. 2. Variation of the object shape of white flies on the conveyer belt. 778

7 Fig. 3. Experimental setup of the developed collection unit: Lower (1) and upper (2) lowpressure chamber, aspiration tube (3), conveyor belt made of filter fabric (4), LED modules (5), progressive scan color CCD camera with macro-lens (6) and photoelectric sensor (7). Fig. 4. Relation between the pest density and image analysis/manual count of the aspirated adult white flies; equal characters stand for no significant difference between the groups with a probability of 0.05, level 0 = 0, level 1 = 75, level 2 = 150 individuals/cage. 779

8 780