Abstract: In strawberry seedling weed control, the harmful effects of herbicide residues have made spraying methods a hot research topic. This paper proposes a machine vision-based herbicide spraying method. After image segmentation and background removal, weed pixels are deleted using an opening operation, resulting in a processed image containing only strawberry pixels. The strawberry image is then divided into 6 rows and 8 columns of sub-regions, and the presence or absence of strawberry pixels in each sub-region determines whether herbicide should be sprayed. Experimental results show that this method effectively controls weeds in strawberry seedlings and saves approximately 50% of herbicide usage. Keywords: Weed, Machine vision, Herbicide spray, Strawberry [b][align=center]Strawberry seedling weed control with machine vision Hu Bo[/align][/b] Abstract: The potential impact of residual herbicide pollution has stimulated research into spray method for strawberry seedling weed control. A novel spray method based on machine vision has been proposed. After the weed image is segmented to distinguish plants and background, weed elements are deleted with open operation. So only strawberry elements remain in the image processing results. Then each strawberry image is divided into 6 rows by 8 columns of spray grids. Whether each grid is sprayed is determined by whether any strawberry elements are present. The result shows that the new method has a good effect on strawberry weed control, reducing herbicide by nearly 50%. Key words: Weed, Machine vision, Herbicide spray, Strawberry, Strawberry is rich in various nutrients and is an important fruit. With the adjustment of China's planting structure, the strawberry planting area has been expanding year by year. Due to the large amount of base fertilizer applied to strawberries and their high water requirement, weeds grow vigorously. In addition, strawberry plants are short and densely planted, making weeding difficult. Studies have shown that weed damage can reduce strawberry yield by 15% to 20%, so weed control has become a major task in strawberry production. Strawberry field weeding can be carried out through a combination of measures such as manual weeding, mulching and weeding, and crop rotation. James et al. pointed out in 2002 that crop rotation and intercropping can reduce strawberry weeds[1]. Haar et al. studied the weed control effect of using the fumigant trichloronitromethane before strawberry transplanting in 2003[2]. These measures have reduced the damage of weeds, but at present, the control of weeds in the strawberry seedling stage still cannot be separated from chemical control. Herbicides have a great impact on the quality and yield of strawberries, especially with the introduction of the requirements for pollution-free strawberry production, which has further increased the requirements for herbicide control. Manual weeding is labor-intensive, and this method also faces challenges due to my country's aging population and urbanization of rural population. In recent years, the use of machine vision to analyze field images, automatically control nozzles, and intelligently spray herbicides has become the main direction of weed control in modern agriculture. Therefore, using machine vision to control strawberry weeds is an important way to reduce pesticide pollution, increase yield, and reduce manual labor. 1 Control strategy for strawberry seedlings Among the existing algorithms for identifying weeds using machine vision, shape features are the main identification features. In 2003, Onyango et al. identified crops and weeds by shape features, and the recognition rates in the experiment reached 82% to 92% and 68% to 92%, respectively[3]. In 2003, Aitkenhead et al. used artificial neural networks to study shape features, and the recognition rate for carrot seedlings, ryegrass and quinoa exceeded 75%[4]. In 2005, Søgaard used shape templates to identify weeds and achieved a recognition rate of more than 65% to 90%[5]. After obtaining the recognition results, the existing related studies generally divide an image into several sub-regions and spray herbicides according to the weed situation in the sub-regions. In 2002, when Tian Lei developed a precise herbicide spraying system, he used multiple nozzles arranged in a row, and each nozzle was responsible for a sub-region while moving[6]. In 2003, Gillis et al. developed an automatic weeding device, dividing the target area into 15cm×15cm sub-areas, with a nozzle set in the center of each area[7]. Through this operation, the herbicide is sprayed onto the weeds as much as possible, and less or no herbicide is sprayed on the crops and soil. In strawberry weed control, since herbicides have a great impact on the growth of strawberries, the first requirement is to avoid spraying herbicides onto strawberries as much as possible. However, in the seedling stage, when weeds have just appeared, the weed plants are small and difficult to identify; on the other hand, the identification rate of existing identification methods cannot reach 100%, and there are cases where strawberries are identified as weeds. Therefore, using traditional identification strategies to spray herbicides is easy to damage strawberry plants. If strawberries are identified and herbicides are sprayed in areas that are not strawberry plants, strawberries can be protected and weeds can be controlled. Although this method has not significantly improved the problem of pesticide residues in the soil, it has a significant effect on protecting strawberries, especially in the production of pollution-free strawberries. In addition, this method of spraying herbicides also reduces the amount of herbicide used to a certain extent. Therefore, using machine vision to identify strawberries and then spraying herbicides on areas that are not strawberry plants is an effective strategy for controlling weeds during the strawberry seedling stage. 2 Image processing algorithm and herbicide spraying method The original image obtained from the field includes strawberries, weeds and background. First, the background needs to be removed through segmentation. The initial state of the color digital image obtained by the hardware device is an RGB image, as shown in Figure 1(a). In the RGB image, each RGB color pixel is represented by three values: R, G, and B. The value range of the three components is determined by the storage method. The currently used full-color image has 24 bytes per pixel and 8 bytes per component, so each component is divided into 256 (28) gray levels. Since threshold segmentation is the main algorithm in weed image segmentation, the segmentation operation mainly determines the features used in the segmentation. In 1995, Woebbecke et al. proposed the super green feature (2g-rb) for the segmentation of weed images and analyzed other color features such as rb, gb, (gb)/(rg) and H. The results showed that (2g-rb) was the most desirable [8]. In the existing research on weed image segmentation, this result is widely accepted, and color features represented by super green features have become the most important weed segmentation features. However, the problem of segmentation error has not been solved. In 2006, Mao Hanping et al. analyzed the factors affecting segmentation error [9]. Through a large number of experiments, the authors used a genetic algorithm to optimize the segmentation features and obtained the optimized features as: -149R+218G-73B, which reduced the influence of these factors. This paper uses the optimized features to segment the weed image. If the segmented pixel is g, then: where the pixels marked as 1 are strawberries and weeds, and the pixels marked as 0 are the background. T represents the segmentation threshold. After segmentation, the weeds occupy relatively few pixels in the image because they have just appeared and the plant volume is small. By using morphological opening operations, weed pixels in the image can be removed with less change to the strawberry outline, thus obtaining the strawberry image. Therefore, the method for controlling strawberry seedling weeds using machine vision is as follows: STEP 1: Using feature -149R+218G-73B, remove soil and other backgrounds through threshold segmentation; STEP 2: Binarize the segmentation result and perform an opening operation to obtain the strawberry image; STEP 3: Divide the strawberry image into 6 rows and 8 columns, a total of 6×8=48 sub-regions, and determine whether to spray herbicide in each region based on whether there are strawberry pixels. [align=center] Figure 1 Strawberry seedling weed image (a) Original image (b) Segmented image (c) Recognition result (d) Sprayed sub-region (e) Spraying method[/align] As shown in Figure 1(a), in the original image, the weed plants are small and difficult to spot, making manual removal difficult. After segmentation and soil removal, the background in Figure 1(a) becomes white, and the weeds are scattered in the image with obvious color differences, as shown in Figure 1(b). Figure 1(b) was binarized and then subjected to morphological opening to obtain Figure 1(c). The pixels corresponding to weeds were removed, leaving only the black areas representing the strawberry pixels in the image. Figure 1(c) was then divided into 6 rows and 8 columns, totaling 6×8=48 sub-regions, as shown in Figure 1(d). The presence or absence of strawberry pixels in each region determined whether herbicide was sprayed. Therefore, the spraying method obtained from Figure 1(d) is shown in Figure 1(e), where the sub-regions with darker colors were sprayed with herbicide, and the sub-regions with unchanged colors were not sprayed. 3. Experimental Results and Analysis Twenty images of strawberry seedlings with weeds were collected at approximately 17:30 on October 22, 2006, in a strawberry field in Xige Township, Liuzhou City, using a Canon A75 digital camera. Image processing, recognition, and spraying results were performed according to the method described in this paper, and are shown in Table 1. Table 1 shows the identification results and spraying methods. Following the method described in this paper, approximately 50% of herbicide can be saved, and virtually no herbicide is sprayed onto the strawberries. Furthermore, areas where herbicide is missed are very small, effectively controlling weeds during the strawberry seedling stage. In cases where herbicide was sprayed on sub-regions containing strawberries, the strawberry area was small, mostly leaf edges. This was mainly due to the removal of some leaf edges during the initial operation. These cases were few in number and primarily involved strawberry leaf edges, so the damage to the strawberries was minimal. In areas where no strawberries were present and no herbicide was sprayed, the weeds had already grown and occupied a relatively large area, making them impossible to remove from the image during the initial operation. Clearly, this method is suitable for the period after strawberry transplanting when weeds have not yet grown large. As weeds continue to grow, the effectiveness of this method gradually decreases. Obviously, during the process of dividing the image into several sub-regions, the size of the sub-regions directly affects the final identification results. If the sub-regions are too large, weeds and strawberries may be in the same area, resulting in missed herbicide spraying and affecting the weeding effect. On the other hand, because the opening operation alters the strawberry edges, if the area is too small, the number of deleted strawberry leaf edges occupying individual sub-regions increases. These sub-regions are then sprayed with herbicides, thus affecting strawberry growth. Through numerous experiments, it was found that for strawberry seedling images, sub-region areas with sizes similar to the strawberry leaves in the image yielded better results. 4. Conclusion This paper proposes a spraying strategy for weed control during the strawberry seedling stage, using machine vision to identify strawberries and then spraying herbicides in areas that are not strawberry plants. Based on this strategy, a corresponding treatment method was developed. Experimental results show that this method minimizes the amount of herbicide sprayed onto the strawberries, and the number of areas missed is also small, essentially achieving weed control during the strawberry seedling stage while saving approximately 50% of herbicide. This method can be applied to weed control in crops like strawberries that have high requirements for pesticide residue control. References [1]James.A.LaMondiaa Wade.H.Elmera Todd.L.Mervoshb Richard.S.Cowles. Integrated management of strawberry pests by rotation and intercropping[J]. Crop Protection 21 (2002):837–846 [2]MJHaara SAFennimorea HAAjwaa CQWinterbottom. Chloropicrin effect on weed seed viability[J]. Crop Protection 22(2003):109–115 [3]Christine.M.Onyango, JAMarchant. Segmentation of row crop plants from weeds using color and morphology[J]. Computers and electronics in agriculture. 39(3):141-155,2003 [4]MJAitkenhead, IADalgetty, CEMullins, AJSM McDonald. Weed and crop discrimination using image analysis and artificial intelligence methods[J]. Computers and electronics in agriculture. 39(3):157-171,2003 [5]Søgaard.HT Weed classification by active shape models[J]. Biosystems engineering, 91(3):271-281,2005 [6]Lei Tian. Development of a sensor-based precision herbicide application system[J]. Computers and Electronics in Agriculture. 36(2):133-149,2002 [7]KPGillis, DKGiles, DCSlaughter, D.Downey. Injection mixing system for boomless, target-activated herbicide herbicide spraying[J]. Transactions of the ASAE. 46(4):997-1008,2003 [8]DMWoebbecke, GEMeyer, K.Von.Bargen, DAMortensen. Color Indices for Weed Identification Under Various Soil, Residual, and Lighting Conditions[J]. Transactions of the ASAE, 1995, 38(1): 259-269 [9] Mao Hanping, Qian Dan. Study on the influencing factors of crop digital image segmentation error[J]. Microcomputer Information. 2006, 10-3: 296-297