Tomato (Solanum lycopersicon) is one of the major commercial vegetable crops in the world (Johansen et al., 2019). Tomato is considered to be beneficial to human health, and they are widely grown and consumed all over the world (Chaudhary et al., 2018), with great commercial and economic values (Brunetti et al., 2019). However, the production of this crop is being threatened by weeds (Qasem, 2019). The competition between weeds and crops for resources such as sunlight and nutrients is one of the main factors affecting crop yields (Hamuda et al., 2016). Early weed control is one of the most important initiatives to prevent significant crop yield losses (Su et al., 2020a).

The most prominent method of weed management is herbicide spraying throughout the crop area (Pérez-Ortiz et al., 2016), but large-scale spraying of herbicides has a negative impact on the environment (Rodrigo et al., 2014). Moreover, incomplete degradability of herbicides can leave residues in plants (Rojas et al., 2017), creating health risks if the products with herbicide residues were consumed by humans and animals (Manh et al., 2001). Additionally, available herbicides for vegetable crops are very limited (Westwood et al., 2018). As a conventional weeding method, mechanical cultivating is not only able to loosen the soil but also can remove weeds between rows (Wang et al., 2019). Nevertheless, this method does not work for intro-row weeds. Intro-row weeds have traditionally been removed by manual weeding. However, manual weeding is inefficient and particularly prone to errors (Slaughter et al., 2008). This approach is increasingly costly, unsustainable and uneconomical in the long run due to labor shortages (Fennimore et al., 2017).

With the increase of the global population and the improvement of living standards, the demand for healthy food continues increasing. Predicting biomass through various types of sensing technologies has become the focus of precision agriculture and smart farming in recent years (Johansen et al., 2020). Precision weeding is considered as one of the most important measures for sustainable vegetable production (Raja et al., 2020a). There is a need to develop an efficient autonomous weeding robot that can intelligently identify weeds and enable precise herbicide application. The primary task of developing a smart weeder is to identify the vegetable crops and weeds accurately and timely.

Detection methods based on characteristics of crop plants and weeds such as color (Gupta and Ibaraki, 2014), size (Lamm et al., 2002), and spectral reflectance (Borregaard et al., 2000) have been proposed to classify crops and weeds. Blasco et al. (2002) developed a robotic system for weed control in transplanted lettuce, which identified the plants with the accuracies of 84% and 99% for weeds and lettuce based on the plant size, respectively. But the probability of misidentification increases significantly when weeds and crops are of similar size. Vrindts et al. (2002) developed a hyperspectral machine vision system to identify weeds in sugar beet fields. However, their method could not be directly used for the real-time detection of the plants. Lin et al. (2017) established a method combining plant spectrum, shape and texture features to discriminate maize and weed with over 95% accuracy, but the spatial location of the plant was not considered in their study. With the development of computer vision technology, more sophisticated object detection methods based on imaging and machine learning have been developed to distinguish crops from weeds (Hall et al., 2017; Lottes et al., 2017; Milioto et al., 2018; Tang et al., 2018; Rehman et al., 2019; Li et al., 2022). Ahmed et al. (2012) proposed a support vector machine method to classify crops and weeds, which achieved above 97% accuracy. dos Santos Ferreira et al. (2017) developed a ConvNets-based software for detecting weeds in soybean, yielding an average accuracy of 99.5%. Subeesh et al. (2022) applied the InceptionV3 model to identify weeds in sweet peppers with an accuracy of 97.7%. However, the above methods need to obtain a large number of samples in advance to label them accurately for model training.

The up-to-date technology called crop signaling was proposed to simplify weed identification (Raja et al., 2019b). This technology creates a machine-readable signal on plants that allows the marked crops to be readily identified by a computer vision system. Crop signaling technology has been successfully used to identify different target plants in weeds (Nguyen et al., 2017; Vuong et al., 2017; Raja et al., 2019a; Su, 2020). For example, Raja et al. (2020b) developed a device containing two cameras to automatically detect lettuce and weeds based on crop signaling. The classification accuracies of lettuce plants and weeds were 99.75% and 83.74%, respectively. However, the use of two cameras to obtain the location information of plants increased the complexity of image processing. In another study, Su et al. (2020b) applied systemic signaling markers to tomato and lettuce plants, allowing the treated plants to be efficiently detected by a single imaging system. Although their technology enabled the detection of exogenous signals applied to crops, the development of the equipment that can be used for the online detection of plant markers has not yet been reported.

In this study, an intelligent sensing system equipped with six color mark sensors and a color camera is expected to be developed for automatic identifications of weed and tomato plants. The color mark sensor is a photoelectric sensor used to quickly detect a specific color label based on the difference between the gray value of the target label and the reference. The color mark sensor is a reverse device, which realizes detection by receiving and analyzing the scattered light of the detected object. The detection principle of the color mark sensor mainly involves three steps. First, the color mark sensor emits monochromatic light (or white light) with the same intensity to the surface of the measured label. Then, it receives the diffuse reflection from the surface of the measured label. Finally, the label is identified according to whether the intensity of the diffuse reflection is consistent with the preset reference value. Prior to detection, the color mark sensor detects and records the amount of the diffuse reflection from the reference and target, respectively. This amount is used as the preset reference value. When an equivalent amount is detected, the object is the desired target. Although the color mark sensor has been widely used in packaging, printing, spinning and other industries (Yang et al., 2009), it has not been used in plant recognition yet.

The captured images were converted from red-green-blue (RGB) space to hue-saturation-value (HSV) space for color segmentation in this study. The reason is that the images acquired in the natural environment are sensitive to illumination variation, and all three components of the RGB space are closely related to illumination. If the illumination of the image changes, it may not be possible to completely separate the plant pixels from the background using RGB space (Hamuda et al., 2016). In contrast, the HSV space is robust to illumination changes (Siogkas and Dermatas, 2006). Additionally, HSV space is more in line with human color perception (Saxe and Foulds, 1996). This makes HSV space more suitable for segmenting plants and backgrounds (soil and residues) and has been widely used in various computer vision applications (Huang et al., 2015).

The main objective of this study was to develop a novel detection method for smart differentiation of tomato plants from packchoi (Brassica chinensis L.) plants. The specific objectives were 1) developing a smart sensing system containing several color mark sensors and a color digital camera, 2) comparing experimental results of arranging color mark sensors and plant labels in different schemes to determine the best combination. 3) evaluating effects of the signal tolerance zone size on the performance of the system in locating weeds and tomato plants to determine the final solution. As far as we know, this was the first study to realize real-time identification and localization of tomato from background plants using a composite sensing system developed by combining computer vision and color mark sensing.

In this study, color mark sensors and cameras were combined in a pioneering way for plant detection. The development of an integrated sensing system that detects plant labels and calculates their location is the main innovation of this study. In order for the developed system to reliably classify different plants, the parameters of the color marker sensors and plant tags need to be determined, which is very important and necessary. The tomato plants were labeled using plant labels to give them a machine-readable signal. Thus, the algorithm to distinguish tomato plants from another unlabeled plant was simplified to improve the efficiency of computer execution. The system successfully located tomato and packchoi plants in real-time by using integrated sensing technology.

However, there are also a few plants that are not correctly localized by this system. The tomato plants were classified as weeds probably due to the low frame rate of the camera used in this system. After detecting a plant label, the next frame in the video stream was acquired by the computer as the image for calculating the plant location. Thus, the camera shutter did not open at the same moment as the sensor detected the plant labels in this image. The desynchronization between the camera shutter and the sensor caused incorrect predicted position of the plant label, as shown in Figure 9 . In the future, high-speed cameras should be considered to improve the accuracy of plant classification. The errors introduced by the low frame rate of the camera and the delay of the command transmission were compensated by tolerance zones. The suitable size of tolerance zone facilitates the system to classify weeds and tomato plants precisely, as shown in Figure 10 . In addition, weeds incorrectly classified as tomato plants possibly due to the occlusion caused by the tomato leaves and stem or weeds located too close to the tomato plants, as shown in Figure 11 . In the future, an algorithm for segmenting overlapping weeds and tomato plants in the image should be developed to improve the accuracy of weed detection.

The color mark sensor selected in this study distinguished the target from other objects by using the difference in their reflection of red light. Based on this photometric principle, the color mark sensor showed different detection capabilities for objects of different colors or materials (Yang et al., 2008). Due to significant optical differences, the red light emitted by the color mark sensor can hardly be reflected by the stems of tomatoes and weeds, while it can be reflected by special plant labels. Thus, the plant labels were more easily detected by color mark sensors. Further, the weed stem is slightly lighter in color than the tomato stem, so it can reflect more red light. This might be the reason for the lower recognition accuracy of red and white plant labels when weed stem was used as a reference. When the tomato stem was used as the reference, the white plant labels were more readily detected than red and green plant labels. The reason for this may be that white labels reflect any wavelength of red light, while tomato stems reflect little red light. Although red labels can reflect red light, the wavelength range of the red light reflected did not exactly match the wavelength range of filter in the sensor, so only a portion of the red light reflected by the red labels received by the photosensitive element of the color mark sensor (Bayer-Krucsay, 1937). There was no significant difference in the light reflected by the green label and the green tomato stem, thus the color mark sensor could not distinguish between them.

A comparative analysis of the various weed and crop classification systems that have been developed are presented in Table 3 . Comparison of crops, sensors, methods, detection speed and accuracy was performed to evaluate the performance of this system. Although only tomato plants were studied in this paper, the methods proposed in the study are applicable to most of the crops that need to be transplanted. Furthermore, although the studies presented in Table 3 targeted different crops, they all used computer vision to identify tomato and packchoi plants. Thus, it is of great interest to compare the results of the present study with them.

The performance of the system is mainly reflected in speed and detection accuracy. Lee et al. (1999) earlier proposed a Bayesian-based classifier to classify tomato plants and weeds, which achieved a speed of 344 ms f ^-1. However, its accuracy was low, with only 47.6% of the weeds and 75.8% of the crops being successfully classified. Lamm et al. (2002) developed a machine vision system to classify crops and weeds in commercial cotton fields based on their sizes, which improved the weed identification accuracy. Bakhshipour and Jafari (2018); Ma et al. (2019) proposed a weed detection method based on the SegNet semantic segmentation method and achieved a high accuracy rate (93.9% for weeds and 93.6% for crops), but the speed of their method was slow (604 ms f ^-1). A method for distinguishing weeds in maize fields based on Mahalanobis distance was proposed by Garcia-Santillan and Pajares (2018). The method achieved an accuracy of 91.8% and a processing speed of 280 ms f ^-1, but this method must obtain crop row information in advance. Then, Raja et al. (2020b) developed a machine vision system based on crop signaling technology to detect weeds in lettuce rows. The system achieved 83.74% weed detection accuracy and 99.75% crop detection accuracy at a processing speed of 160 ms f ^-1. The crop classification accuracy and speed of this system met the requirements of real-time weeding, but the weed classification accuracy of this system was low. The method proposed in this study achieved the highest weed detection accuracy (98.00%) and the second highest crop detection accuracy (94.44%), which proves the effectiveness of the method. In addition, the system established in this study takes only 30 ms to process an image, which is much faster compared to the existing systems that take at least 160 ms to process an image.

An automatic real-time localization system was developed, which successfully located tomato and packchoi plants using integrated sensing techniques based on the crop signaling. With the plant labels applied to tomato plants, the information acquired by a color camera and six color marker sensors integrated by the developed image recognition algorithm. The results demonstrated that tomato stem was a reliable reference for higher accuracy. The detection accuracy of the white physical plant label was significantly higher than that of the red label and the green label. The lower part of the label was identified with the highest accuracy compared to the upper and middle part of the white physical plant label. The topical marker directly applied to the lower part of the plant stem can be more readily detected than the physical plant label. The combination of six color mark sensors is the best solution for detecting plant labels. A tolerance zone with a radius of 7.5 mm maximizes the accuracy of plant classification. Based on the established systematic method, the identification accuracy of tomato plants was 98.00%, and the accuracy of weeds was 94.44%. In addition, the system took only 30ms to process an image, which was faster than existing detection systems.Therefore, the system developed in this study had strong performance for weed and tomato identifications, which could provide prescription maps for weed control in transplanted vegetables at a faster rate.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

W-HS supervised this study, conceived the general idea, provided directions for the system development and revised the manuscript. J-LL developed the system, conducted experiments and analysed the data. W-HS, H-YZ and J-LL drafted the manuscript. W-HS and YP revised the manuscript and provided suggestions for the system development. All authors contributed to the article and approved the submitted version.