Key Technologies of Intelligent Weeding for Vegetables: A Review

Abstract

Vegetables are an essential part of people’s daily diet, and weeds can cause serious losses in vegetable yield and quality. Intelligent weeding technology for vegetables will be one of the mainstream technologies in modern agricultural development. This article reviews the current research status of intelligent weeding technology for vegetables, including vegetable and weed detection technology, weeding actuators, and weeding robots. Firstly, the vegetable and weed detection technology was introduced in detail from three aspects: global weed detection, crop-rows detection, and vegetable/weed precise recognition technology. The research results of some researchers were summarised, and the vegetable/weed precise recognition technology, including machine learning and proximal sensor technology, was introduced. Secondly, the weeding actuators and robots were introduced, including intelligent chemical weeding, mechanical weeding, physical weeding, and integrated weed management methods. Some weeding actuators and robots developed by researchers and agricultural companies were showcased. Finally, the challenges and future development directions of intelligent weeding technology were discussed and analysed. Intelligent weeding technology for vegetables is still mainly limited by natural conditions and a lack of technology. In the future, it will be possible to develop in the direction of multi-algorithm and multi-sensor fusion technologies. It is necessary to improve the applicability of intelligent weeding equipment for various environments, crops, and weeds. This article can provide a reference for future research in the field of intelligent weeding for vegetables.

1. Introduction

1.1. Background and Motivation

Vegetables with extremely a high nutritional value can provide a variety of essential vitamins and minerals for the human body [1,2,3,4,5]. Nutrients such as chemicals and trace elements contained in vegetables can effectively prevent various diseases that affect human health [6,7,8,9]. Therefore, there is a great demand for vegetables around the world. According to the official statistics from the Food and Agriculture Organization of the United Nations (FAO) [10], By 2022, the global vegetable planting area was approximately 5.83 × 10⁷ hm², with a yield of about 2.01 × 10⁴ kg/hm². Compared with 2010, the vegetable planting area increased by 12% and the yield increased by 8.9%. The world population is expected to exceed 9 billion by 2050. Therefore, the planting area and yield of vegetables will continue to increase according to people’s demand. Vegetable cultivation, as one of the main income sources of farmers, can also improve economic and social benefits [11,12,13]. However, harmful weeds, which compete with vegetables for sunlight, air, water, and nutrients [14,15,16], can affect the growth and development of vegetables [17] and exacerbate the severity of the pests and diseases of vegetables [18,19,20,21], leading to serious losses in vegetables’ yield, quality, and nutritional substance [22,23]. Therefore, the weed control in vegetable fields is of great importance.

The main methods of weed control in vegetables include manual weeding [24,25], chemical weeding [26,27,28,29], mechanical weeding [30,31,32], biological weeding [33], and physical weeding [34,35,36]. Chemical weeding, as the most widely used weed control method worldwide, has the advantages of a low cost and high efficiency, but it can cause environmental pollution, and the herbicide residues are harmful to human health [37,38,39,40]. The widespread use of herbicides can also cause problems such as increased weed resistance [41,42,43], soil degradation, and so on [44,45]. Several countries around the world have adopted various measures to reduce the use of herbicides [46]. In addition, with the improvement of peoples’ living standards, organic vegetables have received more and more attention due to their good quality, high safety, and high nutrition [47,48,49,50,51]. Therefore, how to reduce the use of herbicides is urgent.

Non-chemical weeding methods include manual weeding, mechanical weeding, and physical weeding. Manual weeding has good weeding effects, but it is labour intensive and inefficient [52,53]. Mechanical weeding, which is carried out by mechanical structural devices for weeding [54,55,56,57], can loosen the soil, increase soil permeability, and accelerate the mineralisation of soil organic matter and nutrient uptake [58,59], but it is difficult to remove intra-row weeds, which can easily cause damage to crops. Physical weeding methods mainly include mulching [60,61], flame weeding [36,62,63,64,65], electric weeding [66,67,68,69], thermal weeding (hot oil, hot steam, and hot foam) [35,70,71,72], and laser weeding [73], etc. The widespread use of physical weeding has disadvantages such as a high cost and high resource consumption.

With the development of agricultural technology, plant detection and automatic control technologies have been applied to weed management. These advanced technologies are combined with chemical, mechanical, or physical weeding equipment to form intelligent weeding technology [71,72,73,74]. Intelligent weeding is the weeding process in which chemical, mechanical, or physical weeding actuators are controlled by automatic control technology for weeding, which is performed after crops or weeds are identified by machine vision technology. Intelligent weeding can remove weeds according to the needs of farmers [75,76,77]. It can accurately and selectively remove harmful weeds and retain beneficial weeds by identifying them at specific locations [78,79] and can identify crops while performing large-scale weed control to avoid damaging the crops [73,74]. It can also carry out specific site weed management (SSWM) using weed maps. Compared with the large-scale use of chemical weeding, intelligent weeding technology can reduce the use of herbicides. It can easily remove intra-row weeds, avoid the damage to crops caused by only applying mechanical or physical weeding actuators [55] and reduce the cost and energy consumption of physical weeding [69,70,71,72]. Therefore, intelligent weeding technology will become one of the mainstream technologies for vegetable weeding.

Intelligent weeding technology that targets vegetables and weeds integrates various advanced technologies, such as image recognition, sensor, communication, and automatic control, etc. Although the intelligent weeding technology for vegetables has been deeply and extensively studied in many countries around the world, and a series of mature equipment and products have been developed, the research results have not been summarised in detail and systematically, which is not conducive to the development of key technologies. Therefore, this article reviewed the relevant literature and research materials and provided a comprehensive overview of the research status of intelligent weeding key technologies for vegetables.

1.2. Related Surveys

We investigated the relevant literature on intelligent weeding technology, which provided references for this paper. In this section, some similar studies are introduced and discussed. Roberts et al. [80] introduced the application of weed detection in the agricultural field and analysed the benefits, challenges, and limitations of various detection methods, imagery, and sensor systems. Their review also showcased and described emerging weed control technologies, such as automatic chemical weeding, mechanical weeding, electric weeding, laser weeding, and thermal weeding, and explored the challenges and opportunities for their application in the field. This review does not specifically introduce the weed detection process and associated methods. Xiang et al. [81] reviewed intelligent mechanical weeding technologies, including crop and weed detection technologies (contact sensors, non-contact sensors, and machine vision) and mechanical weeding executive parts (hoes, spring teeth, fingers, brushes, swing, and rotary executive parts). The limitations and future development directions of intelligent mechanical weeding were also discussed. In this paper, only the intelligent mechanical weeding technology has been introduced. Coleman et al. [82] reviewed the imperatives of integrated weed management (IWM) in vegetable production. Their review investigated the effectiveness of prevailing weed control methods and evaluated emerging weeding technologies for vegetables. This review can provide useful information for advanced vegetable production systems globally and smallholder vegetable production in developing economies. Li et al. [83] showcased typical weeding robots from the past thirty years and provided a detailed introduction to weed detection techniques that are based on machine learning and deep learning. It provided a guide for the researchers and practitioners. Zhang et al. [84] introduced weed detection methods based on machine learning and deep learning and some emerging weeding robots and discussed the limitations and development trends of current systems.

These reviews can provide valuable information for researchers. However, there are few literature reviews that can provide detailed information on both weed detection and weed control techniques in the vegetable field. In this paper, the literature and research results in the field of intelligent weed control for vegetables are investigated and summarised. Weed detection technologies, including global weed detection, crop-rows detection, and the precise identification of vegetables and weeds are comprehensively introduced. Researchers can learn about the research and application of weed detection technology in the field of vegetable weed control through this review and can also discover which aspects of the technology are relatively lacking. The knowledge regarding the image acquisition technology, image processing methods, detection algorithms, and detection precision of vegetables and their associated weeds can all be obtained from this article. In addition, weeding actuators and robots (including intelligent chemical weeding, mechanical weeding, and physical weeding) that can be used for vegetable weeding are also presented in this article. Researchers or vegetable growers can learn about which weed control actuators and robots have been applied in the field of vegetable weed control. It is important for researchers to know which vegetables and weeds have been studied for detection and weed control techniques. The challenges and development directions of vegetable detection and weed control technologies are also reflected at the end of the paper. In conclusion, this review can provide a reference for researchers’ and vegetable growers’ research and practice in the field of vegetable weed control.

1.3. Literature Indexing Methods

The literature in this review were retrieved from the Web of Science and EI database, which includes journals published by Elsevier, SpringerLink, Wiley, MDPI, IEEE Xplore, and others. The aim of the literature search was to obtain the latest research results in the field of intelligent weed control for vegetables in the last decade (2014–2024), including reviews and research papers. The search keywords were set based on the review content of each section. In Section 3, we initially executed the following search queries: {“vegetable” AND “weed detection”}; {“vegetable” AND “weed map”}; {“vegetable” AND “crop row detection”}; {“vegetable” AND “machine learning” OR “deep learning”}; {“vegetable” AND “sensor detection”}; {“vegetable” AND “spectral” OR “fluorescence” OR “ultrasonic” OR “LiDAR”}; {“vegetable” AND “crop signal” OR “plant label”}; etc. In Section 4, we initially executed the following search queries: {“vegetable” AND “weeding”}; {“vegetable” AND “intelligent weeding” OR “chemical weeding” OR “mechanical weeding” OR “physical weeding”}; etc. In addition, we also searched for some major vegetables, such as sugar beet, potatoes, tomatoes, cabbage, lettuce, carrots, and so on. We filtered the retrieved results by reading the abstracts, removed duplicate studies, and determined them to be in the agriculture field. A total of 402 articles were obtained, including 39 reviews and 357 research papers. A total of 270 articles were used in this review.

1.4. Paper Organisation

This paper is organised as follows: Section 1 elaborates on the research background and motivation, related surveys, and literature indexing methods of intelligent weeding technology for vegetables; Section 2 provides a brief introduction to the types of vegetables and weeds; Section 3 introduces vegetable weed detection technology from three aspects: global weed detection, crop-rows detection, and precise identification of vegetables/weeds; Section 4 presents the weeding execution mechanism and weeding robot, including intelligent chemical weeding, mechanical weeding, physical weeding, and comprehensive weed management methods; Section 5 discusses the problems and future development of intelligent weeding technology; and Section 6 provides a summary of the entire paper.

2. Types of Vegetables and Weeds

The intelligent weeding technology introduced in this article mainly targets vegetables and their associated weeds. Therefore, a brief introduction to the types of vegetables and weeds is given here. There is a wide variety of edible vegetables around the world, with hundreds of them, and more than a hundred commonly used vegetables for cultivation. There is no unified and authoritative classification system for the types of vegetables around the world. In this study, according to edible parts, vegetables were divided into rhizomes, leafy vegetables, cabbage, fruits and melons, scallions and garlic, sprouted vegetables, beans, aquatic vegetables, and fungi, as shown in the

Table 1. Types of vegetables.

Categories	Description	Examples
Root vegetables	Vegetables with tuberous roots and tubers as edible portions	Carrots, turnips, root beets, white radishes, potatoes, taro, yams, sweet potatoes, winter bamboo shoots, etc.
Leafy vegetables	Vegetables with fresh and tender green leaves, petioles, and tender stems as edible portions	Lettuce, spinach, coriander, water spinach, sunflower, celery, chrysanthemum, amaranth, celery, rapeseed, water spinach, etc.
Cabbages	Vegetables with leaf bulbs, tender stems, flower bulbs, and tender leaf clusters as edible portions	Chinese cabbage, head cabbage, bulbous cabbage, cauliflower, baby cabbage, broccoli, pickled cabbage, head mustard, etc.
Fruits and melons	Vegetables with fruit as edible portions	Pumpkin, golden pumpkin, winter melon, bitter gourd, breast gourd, cucumber, luffa, bergamot, watermelon, Hu gua, gourd, watermelon, tomato, eggplant, etc.
Scallions and garlics	Vegetables with bulbs, pseudostems, tubular leaves, or strip-shaped leaves as edible portions	Onions, scallions, scallions, garlic, chives, etc.
Sprouted vegetables	Cultivate edible “sprouts” from seeds of various grains, beans, and trees	Pea sprouts, Chinese toon sprouts, radish sprouts, buckwheat sprouts, peanut sprouts, ginger sprouts, soybean sprouts, mung bean sprouts, bean sprouts, etc.
Beans	Vegetables with bean kernels or tender pods as edible portions	Kidney beans, cowpeas, peas, string beans, knife beans, lentils, green beans, mung beans, broad beans, kidney beans, eyebrow beans, four winged beans, etc.
Aquatic vegetables	Vegetables that grow in water and are edible	Lotus root, water chestnut, water celery, water chestnuts, cigu, water bamboo, etc.
Fungi	Fungi that are non-toxic to the human body and can be safely consumed	Black fungus, white fungus, ground fungus, stone fungus, shiitake mushroom, shiitake mushroom, shiitake mushroom, monkey head mushroom, etc.

. Some vegetable crops, especially those of the same species, have similar biological characteristics and planting patterns, which makes intelligent weeding technology more widely used, but it is also difficult to distinguish crop types. In addition, the wide variety of vegetables also poses a challenge for researchers in the study of weeding techniques.

There are many more types of weeds than vegetables. There are over 1000 types around the world. According to the biological characteristics, weeds can be divided into categories such as annual and perennial weeds [85]. According to the needs of chemical control, they can also be divided into grasses, broad-leaved weeds, and sedges. Due to the wide variety of weed species and their similar biological characteristics to vegetables, the research of weed identification technology will meet great difficulties. Many researchers only detect one or a few types of weeds, or indirectly identify weeds by identifying crop-rows [74,76,77].

3. Vegetable and Weed Identification Technology

Identification technology can identify vegetables and weeds in the field, serving weed control mechanisms. The identification of vegetables and weeds has the following advantages: (1) it can detect the precise location and quantity of vegetables and weeds to achieve precise weed control [72,85]; (2) it can clarify the types of vegetables and weeds to select appropriate weeding methods [15,76,77,78]; (3) it can know the growth period of vegetables and weeds to select the optimal weeding time [86]. Researchers have conducted in-depth research on the identification technology, which mainly includes global weed detection, crop-rows detection, and the precise identification of vegetables and weeds [86,87].

3.1. Global Weed Detection (GWD)

The appearance, distribution, and growth density of weeds in vegetable fields can be obtained by GWD technology, which is beneficial for farmers to carry out SSWM and formulate reasonable and effective weeding measures. Researchers have conducted research on GWD techniques. Fernández et al. [87] reviewed weed detection techniques, including GWD for SSWM, introduced image acquisition platforms and application scenarios, and evaluated the applicability and usability of various acquisition methods. The image acquisition platforms used for GWD mainly include drones and satellite remote sensing, etc. The image acquisition sensors include red–green–blue (RGB) cameras, multispectral sensors, and hyperspectral sensors, etc.

Table 2. Research methods and results of global weed detection for some crops.

Platforms	Crops	Weeds	Sensors	Resolution Ratio/mm·pixel−1	Recognition Algorithms	Precision	References
UAV	Wheat	Broadleaf herba, Gramineae	RGB	2.7	Random forest (RF)	72%	Anderegg et al. [95]
		Gahnia tristis	RGB/Multispectral	10/30	UNET-ResNet	>90%	Fraccaro et al. [96]
			Multispectral	11.6	RF	>93%	Su et al. [92]
		Alopecurus myosuroides	RGB, Multispectral	32	RF	87% (RGB); 61% (Multispectral)	Lambert et al. [97]
			Multispectral	82.7	Convolutional Neural Networks (CNN)	82.5%	Lambert et al. [98]
		Matricaria chamomilla L., Papaver rhoeas L., Veronica hederifolia L., and Viola arvensis ssp. Arvensis	RGB	-	Deep residual convolutional neural network (ResNet-18)	94%	Camargo et al. [93]
	Maize	Cyperus rotundus L., Cynodon dactylon L., Malva sylvestris L., Artemisia vulgaris L., Polygonum aviculare L.	Multispectral	50/80/90	Support vector machine (SVM)	61%	Castaldi et al. [99]
		Convolvulus canariensis, Chenopodium album, Digitaria sanguinalis	RGB	1.78	Hough transform (HT) and object based image analysis (OBIA)	94.5%	Gao et al. [100]
	Sugar beet	Cirsium arvensis L.	Multispectral	5.2	PLS-DA analysis model	95% (weeds); 89% (sugar beet)	Garcia-Ruiz et al. [101]
		Galinsoga spec., Amaranthus retroflexus, Atriplex spec., Polygonum spec., Gramineae, Convolvulusarvensis, Stellaria media, Taraxacum spec.	Multispectral	82	Improved Deep Nueral Network (DNN)	>78%	Sa et al. [102]
	Marigold	Setaria viridis, Asclepias syriaca, Cistothorus stellaris	RGB	5	U-net	>93%	Zou et al. [103]
	Pea, Strawberry	Eleusine indica	RGB	3	Semi-Supervised Generative Adversarial Network (SGAN)	90%	Khan et al. [104]
	Oat	Chamaemelum nobile, Cirsium arvense	RGB	35	RF, K-means	>87%	Gašparović et al. [105]
	Cotton, Sunflower	Conyza crispa, Fallopia japonica, Chenopodium retusum, Phalaris canariensis, Convolvulus arvensis	RGB	12–24	OBIA	81% (Cotton); 84% (Sunflower)	de Castro et al. [106]
	Sunflower	Chenopodium retusum, Brassica juncea var. megarrhiza, Convolvulus arvensis, Chenopodium album Linn.	RGB, Multispectral	-	OBIA	>85%	López-Granados et al. [107]
	Sunflower, maize	Chenopodium retusum, Brassica juncea var. megarrhiza, Convolvulus arvensis, Batis maritime	RGB	14	OBIA	95% (Sunflower); 79% (maize)	Pérez-Ortiz et al. [108]
	Onion	Cyperus rotundus, Sorghum halepense, Chenopodium album, Xanthium strumarium, Amaranth, Conyza bonariensis L., Solanum L.	RGB	5	Maximum Likelihood and SVM	>85%	Rozenberg et al. [109]
	Sorghum	Chenopodium album L., Cirsium arvense L.	RGB	1	Deblurring and segmentation models	83.7%	Genze et al. [110]
	Tomato	Solanum nigrum, Amaranthus retroflexus	Hyperspectral	-	Multiclassifier architecture	95.8%	Zhang et al. [111]
	Rice	Leptochloa chinensis Nees, Cyperus L.	RGB	3/5	Fully convolutional network (FCN)	>90%	Huang et al. [112,113]
Aviation aircraft	Pea, broad bean and winter wheat	Diplotaxis spp., Sinapis spp.	RGB, NIR	250	Vegetation indices, Maximum Likelihood	>85%	de Castro et al. [114]
QuickBird satellite	Wheat	Brassica napus	Multispectral	2.4 × 103	Vegetation indices, Maximum Likelihood	>89%	de Castro et al. [115]
		Avena sterilis	Multispectral	2.4 × 103	Maximum Likelihood, SVM	>91%	Castillejo -González et al. [116]
Sentinel-1A and Sentinel-2A	Beans, beetroot, maize, potato, and wheat	-	Multispectral, MSI	10/20/60 × 103	Kernel-based extreme learning machine (KELM)	96.8%	Sonobe et al. [94]
Sentinel-2 satellite	Potato	Triticum aestivum	Multispectral, NIR, SWIR	-	Unsupervised classification	86%	Revathy et al. [117]
	Maize	Davilla strigosa	Multispectral	10 × 103	Guided regularized random forest (GRRF)/Machine learning discriminant	>85%	Mudereri et al. [118,119]
					RF/Subpixel multiple endmember spectral mixture analysis (MESMA)	88%
		Richardia brasiliensis, Chenopodium album, Cyperus esculentus, Megathyrsus maximus		10/20/60 × 103	RF	95%	Mkhize et al. [120]
PlanetScope and Sentinel-2		Davilla strigosa		3/10 × 103	GRRF	92%(PS); 88%(S2)	Mudereri et al. [121]

shows some research achievements of researchers in GWD techniques in recent years. Unmanned aerial vehicles (UAVs) have been widely used in GWD, due to their high flexibility and efficiency [86,88,89,90,91], but their endurance is poor, and they are greatly affected by lighting and wind. Compared with UAVs, the resolution of images acquired by aviation aircraft and satellites are relatively lower, although they have a wide detection range. RGB and multispectral sensors have also been widely used in GWD. Compared with RGB and multispectral sensors, hyperspectral sensors have a high spectral resolution, but they are applied less because of their high cost and complex data processing [90]. Machine learning and deep learning algorithms with high recognition accuracy are used in GWD. As shown in Table 2, some researchers have created a weed distribution map with high accuracy by GWD [92,93,94]. Researchers have conducted extensive research on food crops and grains in GWD, but there is a lack of research on vegetables. Therefore, researchers should increase research on vegetables in GWD to fill the gap in this area. The research results shown in Table 2 can provide a reference for the study of vegetables in GWD.

3.2. Crop-Rows Detection (CWD)

CWD is one of the techniques used to achieve SSWM. CWD technology can be used for vehicle navigation [122,123], inter-row weed detection, and inter-row weeding [124,125], in which it can control the travel routes of the vehicle and weeding device and prevent damage to crops. Most researchers conduct research on automatic navigation by CWD [123]. Shi et al. [126] reviewed the relevant methods and applications of CWD technology in agricultural machinery navigation and discussed and summarised the advantages and disadvantages of current mainstream CWD methods. CWD technology has great potential to identify inter-row weeds, although there is limited research in this area.

Table 3. Research methods and results of crop-rows detection for some vegetables.

Platforms	Crops	Sensors	Resolution Ratio	Detection Algorithms	Precision	References
UAV	Grapevine, pear, and tomato	RGB	0.1 m	Thresholding algorithms, classification algorithms, and Bayesian segmentation	>90%	Ronchetti et al. [127]
	Sugar beet, maize	Multispectral	2.5 cm	CRowNet—a model formed with SegNet (S-SegNet) and a CNN based Hough transform (HoughCNet)	93.58%	Bah et al. [128]
	Spinach, legumes	RGB	0.35 cm	Hough transform; simple linear iterative clustering (SLIC); ResNet-18 (CNN)	>90%	Bah et al. [125]
Robot	Sugar beet, maize	RGB	-	Deep learning (DL)	90.25%	Silva et al. [129]
	Cabbage	RGB	1920 × 1080 px	A method of cabbage crop row localization and multi-row adaptive ROI extraction based on the limited threshold vertical projection	95.75%	Han et al. [130]
	Sugar beet	RGB-D	1280 × 720 px	A semantic segmentation approach for crop row detection based on U-Net	85%	Silva et al. [131]
	Sugar beet, canola, and leek	3D laser and vision camera	-	Pattern Hough transform	-	Winterhalter et al. [132]
	Kohlrabi, Chinese cabbage, and pointed cabbage	GNSS and colour camera	5-megapixel	Localization—beyond crop row following	83%	Winterhalter et al. [133]
	Cucumber	RGB	5-megapixel	Prediction point Hough transform algorithm	-	Chen et al. [134]
	Cabbage and red cabbage	RGB/Multispectral	1280 × 960 px	CNN Vegetables Detection—characterization method	90.5%	Cruz Ulloa et al. [135]
	Maize	Hyperspectral	2 nm	Self-supervised method	>80%	Wendel et al. [136]
Tractor	Soybean	RGB	1280 × 480 px	Image analysis method	-	Ospinad et al. [137]
	Rice, rape, and wheat	RGB	480 × 640 px	Shear-binary-image algorithm; least square method	96.7%	Zhang et al. [138]
Handheld camera/UAV	Cabbage, kohlrabi, and rice	RGB	4032 × 3024 px	Improved YOLO-v8 and threshold-DBSCAN	98.9%/97.9%/100%	Shi et al. [139]
Mobile phone/Moving vehicle	Potato	RGB	1280 × 720 px	Deep learning (VGG16) and feature midpoint adaptation	>90%	Yang et al. [140]
Handheld camera	Sugar beet, maize	Multispectral	6 mm	Unsupervised classification algorithm based on spatial information (Hough transform) and spectral information	100%	Louargant et al. [141]
Public data	Sugar beet, legumes	Multispectral	-	CNN (autoencoder) and simple linear iterative clustering (SLIC)	88%	Bah et al. [142]
	Corn, celery, potato, onion, sunflower, and soybean	RGB	-	Hough transform and DBSCAN clustering analysis	63.3%	Zhao et al. [143]
	Maize, celery, potato, onion, sunflower, and soybean	RGB	2560 × 1920 px	New crop row detection algorithm	84%	Rabab et al. [144]
				New method	74.7%	Vidović et al. [145]
	Garlic, corn, oilseed rape, rice, and wheat	RGB	-	Seedling crop row extraction method based on regional growth and mean shift clustering	98.18%	Wang et al. [146]
	Lettuce and cabbage	RGB	768 × 1024 px	Automatic accumulation threshold of Hough Transformation; K-means clustering method	97.1%	Chen et al. [147]
	Cauliflower	RGB	2704 × 1520 px	Kalman filtering and Hungarian algorithm	99.34%	Hamuda et al. [148]

shows the research results of some researchers on CWD for vegetables. The collection platforms that are available for CWD include UAV, robot, tractor, moving vehicle, handheld camera, etc. The collection sensors include RGB, multispectral, hyperspectral, LiDAR (Light Detection and Ranging Sensor), etc. The recognition algorithms of crop rows include Hough transform, linear regression, vertical projection in combination with Hough transform, a random sample consensus, frequency analysis, the exploration of horizontal strips, a clustering algorithm, deep learning, etc. Hough transform, a wide image analysis method, can extract various shapes such as lines, circles, and ellipses from images. The detection principle is to convert the spatial coordinate points into curves in parameter space and detect the shape in the image by statistically analysing the intersection points of the curves in parameter space. As shown in Table 3, researchers have studied about thirty types of vegetables in CWD using different methods, and most of the detection accuracies reached over 80% [127,128,129,130,131]. Research methods and results can provide references for the study of other vegetables in CWD.

3.3. Vegetable Weed Precise Identification

3.3.1. Machine Learning (ML)

Machine vision technology, which is widely used in the agriculture field, can identify crops and weeds by image acquisition technology, image recognition technology, etc. Image acquisition technology can collect crop and weed images by various sensors such as RGB, multispectral, LiDAR, ultrasound, and so on. Image recognition technology includes traditional image recognition and ML technology. Traditional image recognition uses segmentation models to distinguish crops and weeds based on plant features such as their colour, shape, texture, geometric parameters, and so on [149]. ML utilises classification models with data learning capabilities, which are more accurate and efficient in handling complex data. ML, which can be divided into traditional machine learning (TML) and deep learning (DL), involves five key steps: plant image acquisition, image pre-processing, image segmentation, plant image feature extraction, and plant classification [150,151,152].

Researchers have conducted extensive research on plant detection based on ML methods [150]. Some studies have reviewed and introduced ML methods and detection steps [150,151]. Al-Badri et al. [152] introduced traditional machine learning and deep learning weed detection methods and processes, analysed the advantages and limitations of ML, and discussed the future development trends of weed detection. Wang et al. [153] described the process of ML and detection, and provided a detailed introduction to image processing, image segmentation, plant feature extraction, and plant classification. Some review papers have also introduced DL methods and processes for plant detection [85,154]. Hu et al. [155] reviewed research on weed detection based on deep learning, presented common datasets and evaluation metrics, introduced deep learning architecture, and discussed future challenges and potential solutions. Qu et al. [156] reviewed the application of deep learning in vegetable and weed detection, provided detailed introductions to image data acquisition, image processing, image segmentation, image feature extraction, and plant classification algorithms, and presented the research results of some researchers.

Table 4. Research methods and results of traditional machine learning for some vegetables.

Crops	Weeds	Platforms	Sensors	Image Pre-Processing Methods	Image Segmentation Methods	Feature Extraction Methods	Classification Algorithms	Accuracy	Precision	References
Carrot, maize, and tomato	-	Robot	RGB	Colour space transformation from RGB to CIE Luv	Based on learning method	Neighbourhood pixel information	SVM	-	91%/88%/89%	Rico-Fernández et al. [157]
Carrot	-	Public data	RGB	Technique of contrast limited adaptive histogram equalisation (CLAHE); dimensionality reduction	-	Laws’ texture masks	RF	94%	-	Kamath et al. [158]
	Impatiens balsamina Linn., Humifuse Euphorbia Herb, Eleusine indica	-	Hyperspectral	Data normalization, soil background removal	Normalized Vegetation Index (NDVI)	Band selection, wavelet transform	Fisher linear discriminant analysis (LDA) and SVM	-	>85%	Liu et al. [159]
Sugar beet	Thistle	Ground vehicle	RGB	-	Modified excess green (ExG) index	Local features based on affine invariant regions and scale invariant keypoints, colour vegetation indices	Bag-of-Visual-Word scheme with SVM	99.07%	-	Kazmi et al. [160]
	-	Robot	RGB + NIR	Normalization processing	Near-infrared information	Statistical and shape features	RF	-	>90%	Lottes et al. [161]
Spinach	-	Robot	RGB	-	ExG	Image filtering to extract colour and area features	A classification based on area	-	>80%	Pulido-Rojas et al. [162]
Lettuce	Chenopodium serotinum, Polygonum lapathifolium	Robot	RGB	-	Based on colour features	Histogram of oriented gradient (HOG), local binary pattern (LBP), and grey-level co-occurrence matrix (GLCM)	SVM and image blocking	-	94.73%	Zhang et al. [163]
Tomato, cotton	Solanum nigrum, Abutilon theophrasti	-	RGB	Random rotation and zooming, Gaussian, Poisson, and salt–pepper noise addition, and contrast and brightness shifts	Normalized difference vegetation index, OTSU threshold	-	Fine-tuned densenet and SVM	-	99.29% (F1)	Espejo-Garcia et al. [164]
Broccoli, lettuce	Bromus, Chenopodium retusum, Chenopodium album Linn., Dysophyllastellata Benth., Echinochloa crusgalli L., Convolvulus arvensis, Impatiens balsamina Linn., Trifolium repens	Remote-controlled ground vehicle	RGB-D	Removed invalid pixels and noise pixels in point clouds	Colour and depth information	Canopy and leaf features	ML; Colour-depth fusion-segmentation algorithm	-	96.6%/92.4%	Gai et al. [165]
	-	Tractor	RGB	Illumination correction for an increased uniformity; transformation of the colour coordinates from RGB to Hue saturation value; extraction of the H and S channels; filtering for noise reduction	-	Image area and shape factors	Shape analysis and colourimetric k-nearest neighbor (k-NN) clustering	-	65–79%/69–96%	Pallottino et al. [166]
Lettuce, cauliflower, and maize	-	Tractor	RGB	-	ExG	-	A novel method calculating lateral offset	>95%	-	Li et al. [167]

and

Table 5. Research methods and results of deeping learning for some vegetables.

Crops	Weeds	Platforms	Sensors	Image Pre-Processing Methods	Image Segmentation Methods	Feature Extraction/Data Annotations Methos	Classification Algorithms	Accuracy	Precision	References
Chinese white cabbage	-	-	RGB	Data augmentation, brightness, rotation, flips, colour variations, and image definition	ExG	Bounding box annotation	YOLO-v3	-	97.1%	Jin et al. [168]
Chinese cabbage	-	Mobile trolley	RGB	Gaussian noise, random rotation, random cropping, and random flip	-	Image annotation	Improved U-Net	98.24%	94.56%	Ma et al. [169]
Potato	Chenopodium album	Smart sprayers	RGB	Adjusting image size	-	Image annotation	DCNNs (GoogLeNet, VGG-16, and EfficientNet), PyTorch/TensorFlow	>90%	>90%	Hussain et al. [170]
	Malva neglecta, Portulaca oleracea, Chenopodium album L, Secale cereale L., Xanthium strumarium	Moving platform	RGB	-	Based on threshold	Based on the gray level cooccurrence matrix (GLCM), colour features, spectral descriptors of texture, moment invariants and shape features	Hybrid artificial neural network-ant colony (ANN-ACO), hybrid artificial neural network-simulated annealing (ANN-SA), and hybrid artificial neural network-genetic algorithm (ANN-GA)	98%	-	Sabzi et al. [171]
	-	-	RGB	Data enhancement		Image annotation	Improved YOLOv4 model (MC-YOLOv4	98.52%		Zhao et al. [172]
Sugar-beet, carrot, soybean	Heuchera americana, Chenopodium album Linn., Chenopodium retusum, Rapistrum rugosum; Mentha requienii, Daucus carota; Poaceae Barnhart, Broadleaf herba	Public data	RGB	-	ExG	Bag-of-visual-words-augmented histogram of oriented gradients	Backpropagation neural network	>90%	>92%	Abouzahir et al. [173]
Cassava	-	UAV	RGB	Image-space and colour-space transformations and the novel mosaic data augmentation	-	Bounding box annotation	YOLOv5	-	96.7%	Nnadozie et al. [174]
Strawberry, pea	Eleusine indica	UAV	RGB	-	-	Bounding box annotation	Faster R-CNN	94.73%	-	Khan et al. [175]
Cabbage	Acalypha australis Linn., Chenopodium album, Humulus scandens, Ivy glorybind, Capsella bursa-pastoris	-	RGB	Cropped	ExG	Bounding box annotation	YOLOv5/YOLOv8	-	98.6%/ 97.3%	Sun et al. [176]
Sugar beet	Solanum tuberosum	Robot	RGB	Adjusting image size; data enhancement	-	-	AlexNet, VGG-19, GoogLeNet, ResNet-50, ResNet-101 and Inception-v3	>90%	-	Suh et al. [177]
		Robot	RGB	Translation, rotation, shear, scale, reflection, saturation, brightness, Aadjusting image size	-	Image annotation	YOLOv3	-	84%	Ruigrok et al. [178]
	Fumaria indica	UAV	Multispectral	Small patches classification	-	Image annotation	VGG-BEET network	-	93.4%	Moazzam et al. [179]
	Chenopodium retusum, Chenopodium album Linn., Eruca vesicaria, Rapistrum rugosum	-	RGB	Removal of duplicate images and low-quality images, data annotation using LableIMg	Based on threshold	Fourier descriptors and moment invariant features	Artificial neural network-ANN/SVM	-	>92%	Bakhshipour et al. [180]
	-	Public data	RGB	-	-	Bounding box annotation	CCCS-YOLO	79.5%	81.3%	Zhang et al. [181]
	-	Public data	RGB	-	-	Bounding box annotation	WeedNet-R	-	>85%	Guo et al. [182]
	-	Public data	RGB	Data augmentation and background segmentation	-	-	TIA-YOLOv5 network	90%	80%	Wang et al. [183]
Bok choy	-	-	RGB	Colour, brightness, rotation, and image definition, data enhancement	Based on colour	Bounding box annotation	CenterNet	-	95.6%	Jin et al. [184]
Tomato	Monocotyledonous, Dicotyledonous	Public data	RGB	Manually labelled, adjusting image size	-	Bounding box annotation	RetinaNet, YOLOv7 and Faster-RCNN	>90%	-	López-Correa et al. [185]
Cauliflower and broccoli	-	Robot	RGB	The images were corrected for vignetting and flash patterns and down sampled to a resolution of 1080 × 720 pixels	-	Bounding box annotation	Faster R-CNN architecture with a Resnet-101 feature extractor	-	94.79%/ 94.3%	Bender et al. [186]
Maize, tomato	Solanum nigrum L., Portulaca oleracea L., Setaria Verticillata L.	-	RGB	Applying data augmentation to the images, consisting of scaling, horizontal flips, and random adjustments to the exposure and saturation of the images.	ExGR	Bounding box annotation	Yolov 8l and RetinaNet algorithms	-	93%	Moreno et al. [187]
Brinjal	-	Public data	RGB	Flipping the image vertically and horizontally and adding some noise; adjusting image size; normalization processing	-	Bounding box annotation	ResNet-18, YOLOv3, CenterNet, and Faster RCNN	88%	85%	Patel et al. [188]
Chicory, sugar beet	Mercury Herb	UAV/Public data	RGB	Marking, cropping	-	Bounding box annotation	YOLOv7		61.3%	Gallo et al. [189]
Okra, bitter gourd, sponge gourd	Conyza canadensis, Paris quadrifolia, Gramineae	Robot	RGB	Scaling	-	Bounding box annotation	YOLOv5	88%	83%	Fatima et al. [190]
Strawberry, tomato	Eleusine indica	-	RGB	Annotation of the entire plant and annotation of partial sections of the leaf blade	-	Bounding box annotation	YOLOv3-tiny	87%/77%	74%/38%	Sharpe et al. [191]
Crops	-	-	RGB	Gabor filtering and short-term Fourier transformation for image enhancement	-	-	Elman Neural Network Model (ENN) and Modified Barnacles Mating Optimization (MBMO)	98.99%	96.13%	Albraikan et al. [192]
Carrot, sugar beet, rice	-	Public data	RGB	Data augmentation of random horizontal flip, normalisation, and random Gaussian blur	-	-	Improved deep convolutional neural network (DCNN) algorithms	-	>91%	Janneh et al. [193]
Vegetables	-	Public data	RGB	Data augmentation and filtering	-	-	Improved YOLOv3	96%	96%	Madanan et al. [194]
Beet, parsley and spinach	-	UAV	RGB	Data augmentation and normalisation processing	-	Bounding box annotation	Visual transformers (ViT-16)	-	>97%	Reedha et al. [195]
Carrot	Crabgrass, Musa spp., Pale persicaria, Cephalanoplos	-	RGB	Data augmentation	-	Bounding box annotation	YOLO v4-weeds	88.46%	-	Ying et al. [196]
Cauliflower	Chenopodium album, Capsella bursa-pastoris, Lemna minor	-	RGB	Reduce image size; Gaussian fuzzy filtering; HSV colour space conversion	Based on colour	Bounding box annotation	PASCAL visual object classes method		99.04%	Hamuda et al. [197]

show the research findings of some researchers in vegetable and weed detection based on TML and DL.

As shown in Table 5, researchers have used different methods to study the precise detection of various types of vegetables and weeds. Some researchers chose to identify weeds [163,164,165], while others chose to identify crops [161,162,171]. Most of them achieved high detection accuracy and precision. It can be seen that researchers are currently using deep learning methods more than traditional machine learning methods in the study of precise detection for vegetables. The literature and research findings in Table 4 and Table 5 can provide references and inspiration for future research in this field. The detection processes of traditional machine learning and deep learning methods are briefly introduced as follows:

Traditional Machine Learning (TML)

(1) Plant image acquisition

Some researchers collect crop and weed images under different natural conditions (different lighting and windy weather, etc.) by autonomous collection [157,160,161,162]. Image acquisition platforms generally include drones, ground vehicles, robots, or handheld cameras, etc. Acquisition sensors include RGB, spectral, near-infrared, and LiDAR, etc [198,199]. RGB sensors have been widely used due to their low cost and ability to clearly capture spatial backgrounds, plant colours, leaf morphology, textures, and other features, as shown in Table 4 and Table 5. Some researchers have used public datasets containing images from different scenarios and images that have undergone the processes of background segmentation and image annotation, which provide great convenience to researchers [158,173,181,182,183]. Deng et al. [200] investigated image datasets that can be used for weed detection, collected a total of 36 public datasets, and presented a new dataset. These datasets can be found on the website and are updated in real-time [201,202]. The dataset is expanded by various methods (such as rotation, scaling, trimming, flipping, spatial colour, etc.), which can reduce the risk of overfitting, solve the problem of data scarcity, improve the generalisation ability, robustness, and classification performance of classification models, and improve data balance. When identifying plants, the dataset is divided into training, validation, and test sets in certain proportions. The training set is used to construct the model, the validation set is used to determine the network structure or parameters that control the complexity of the model, and the test set is used to evaluate the performance of the final optimal model.

(2) Image pre-processing

The collected images are pre-processed to improve the image quality. Image pre-processing methods generally include resizing, image enhancement, normalisation processing, spatial colour conversion, and denoising. The image size of the original images is adjusted appropriately, which can remove excess information and minimise computational cost. Image enhancement is used to enhance and adjust the contrast and other parameters of the original image to solve the problems such as lighting and shadows [164]. Normalisation processing refers to the process of transforming images into a fixed standard form, which can avoid interference caused by uneven lighting. Spatial colour conversion can convert image colours according to different purposes. The colour space models, including RGB, Hue–Saturation–Intensity (HIS), Hue–Saturation–Value (HSV), Lab (Luminosity (L) and related colours a and b), YCrCb (luminance, tone, and saturation), can be obtained from RGB transformation functions [153]. Image noise can be removed by image filtering, which can smooth, sharpen images, and increase the resolution of images [166]. Common filtering methods include mean filtering, median filtering, bilateral filtering, Gaussian filtering, blur filtering, and homomorphic filtering.

(3) Image segmentation

Image segmentation is the process of dividing an image into specific and unique regions and defining regions of interest. In weed detection, crops and weeds are mainly separated from the background by image segmentation methods that include colour indices, thresholding, and learning segmentation [203]. The colour indexes include Normalised Difference Index (NDI), Excess Green Index (ExG), Excess Green (EG), and Excess Green minor Excess Red Index (ExGR), etc. [159,160,161,162,163]. The threshold segmentation method, which compares and classifies the pixel collection points of the processed images with a preset feature threshold to obtain the region of interest, is the most basic and widely used in weed detection [164,171,180]. Common methods include fixed threshold (using a fixed grayscale value as a threshold, comparing the pixel values of the image with the threshold, and classifying the pixels into two categories based on the comparison results), Ostu method (an automatic threshold segmentation technique that automatically selects the optimal global threshold based on the greyscale histogram of the image to maximise interclass variance or minimise intraclass variance), dynamic threshold (automatically adjusts the threshold based on changes in the data to adapt to fluctuations and changes in the data), automatic threshold (calculates local thresholds based on the brightness distribution of different regions in the image to adapt to the local characteristics of the image), histogram entropy (calculates histogram entropy based on image histogram information to determine optimal threshold for image segmentation), etc. Learning methods use machine learning algorithms to classify image pixels and achieve the segmentation of plants and backgrounds [157]. Compared with segmentation methods based on colour index and thresholds, learning methods have better classification performance but require a large amount of computation. The segmented image may contain undesirable noise, misclassified pixels, and connected areas, and polishing is necessary to improve image quality. Common measures include morphological opening and closing, median filtering, and pattern filtering.

(4) Image feature extraction

There are several characteristic differences between crops and weeds, which can be divided into four categories: texture features, biological morphology features, spectral features, and spatial background. Extraction methods for texture features (such as leaf vein texture [165], smoothness, and roughness, etc.) include statistical feature methods (such as grey level co-occurrence matrixs (GLCM), etc.) [157,171], structure methods (such as local binary patterns (LBP), etc.) [168], model methods (such as fractal models, autoregressive models (AR), Markov random field models, etc.), and transformation methods (such as curve transformations, wavelet transformations, etc.) [200]. Texture feature methods have the characteristics of high accuracy, adaptability, and robustness. Biological morphological features include region shape parameters (such as area, perimeter, length, and width, etc.) [167], region shape indices (such as eccentricity, circularity, convexity, rectangularity, etc.), boundary-based shape descriptors (such as moment invariants (MI), beam angle statistics (BAS), Fourier descriptor (FD), tensor scale descriptor (TSD), etc.) [161]. Biological morphology methods have strong independence and noise resistance but are strongly influenced by environmental factors. When there is a significant difference in spectral reflectance between crop and weed leaves, spectral features (colour indices) can be used for feature extraction [160,162]. The spatial background can be applied to weed identification, and some researchers indirectly identify inter-row weeds by identifying crop rows [125]. In addition, multi-feature fusion methods are also commonly used to distinguish crops from weeds, which can complement each other’s strengths, improve detection accuracy and stability, and resolve some interferences factors under non-ideal conditions.

(5) Plant classification

Plant classification classifies plant features by machine learning algorithms. TML classification algorithms, which can be divided into supervised and unsupervised learning, can make extracting image features form a data feature set and then use classification models to train data features. The trained model can classify and combine different feature data to distinguish between crops and weeds. Supervised learning algorithms, including Support Vector Machine (SVM) [157,160], Linear Classifier (LC) [159], Decision Tree (DT), Random Forest (RF) [158,161], and K-Nearest Neighbour (K-NN) [166], etc., train and classify labelled feature data. Unsupervised learning algorithms, including K-means clustering and principal component analysis (PCA), do not require the labelling of data and can automatically classify similar features or attributes in the data.

Deep Learning (DL)

DL algorithms can learn the inherent patterns and representation levels in data, extract the multi-scale and multi-dimensional spatial semantic feature information of plants, and improve the accuracy of weed identification. TML requires manual feature extraction from image data based on experience and professional knowledge, while DL is a hierarchical learning approach that uses representational learning to automatically extract deep features from images, with stronger representational capabilities and unique network feature structures. Compared with TML, DL can effectively avoid the subjectivity brought by the feature extraction process and improve the accuracy of plant identification. The detection process based on DL is similar to TML, and the main difference is the classification algorithm. Hasan et al. [204] reviewed the application of DL in weed detection, evaluated the advantages of DL over TML, introduced the detection process, and also presented some datasets.

DL includes supervised learning, unsupervised learning, and semi-supervised learning. Supervised learning needs to conduct image annotation in image pre-processing to distinguish different plants. Common annotation methods include bounding box annotation [174,175,176], pixel-level annotation, image annotation [169,170], polygon annotation, and composite annotation, etc. Unsupervised learning does not need to conduct image annotation, which can automatically extract discriminative information and features and divide them into separate groups. As shown in Table 5, common DL architectures include Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), Fully Convolutional Networks (FCN), Deep Belief Networks (DBN), Generative Adversarial Networks (GAN), and Graph Convolutional Networks (GCN), etc.

3.3.2. Sensor Detection Technology

Some researchers have used near-end sensors (such as spectral sensors, ultrasound, LiDAR, and thermal imaging) and corresponding recognition algorithms for plant detection [149]. Palottino et al. [205] reviewed the main types of sensors and their applications in plant detection, including greyscale/RGB imaging, visible near-infrared (VIS-NIR) and near-infrared (NIR) sensors, stereoscopic vision, thermal imaging, combination sensors, phenotype sensors, etc. Singh et al. [149] reviewed existing technologies that can be used to develop ground sensor systems, including multispectral and hyperspectral recognition, and fluorescence imaging recognition techniques for weed identification. Fernandez-Quintanilla et al. [87] described the current research status and applicability of remote sensing and ground detection technologies and introduced the limiting factors and future development directions of weed detection systems. The research methods and results in sensor detection technology can provide reference and inspiration for future research in vegetable detection.

Spectral Detection

Differences in spectral reflectance can be used for plants identification. Spectral sensors, including multispectral and hyperspectral, can collect spectral images that reflect the reflectance information of visible, infrared, near-infrared, and ultraviolet light on plant leaves. Multispectral sensors can usually obtain spectral information from 3 to 10 bands, with a relatively small amount of captured information, while hyperspectral sensors can obtain information from hundreds or thousands of bands with a high resolution and rich information content, but with a high cost, complex information processing, slow computing speed, and the need for specific processing algorithms. The steps of spectral plant detection include image acquisition, image pre-processing, image binarisation, image segmentation, feature extraction, and decision/weed detection [149]. Spectral sensors are often used in remote sensing systems to detect plants in the field, due to the high resolution. Table 6 shows some research methods and results of vegetable and weed detection using spectral sensors. It can be seen that the technology has high detection accuracy for vegetables and weeds, but its application for vegetables is limited.

In addition, Mohammadi et al. [206] developed a new type of multispectral camera that can be used for crop and weed detection. The multispectral camera was based on Fabry-Pérot technology and can cover visible and near-infrared wavelengths ranging from 380 to 950 nm. A genetic algorithm was used to optimise spectral bands, and a total of 16 spectral bands suitable for agricultural applications were selected. Tao et al. [207] designed a multispectral sensor for winter rape detection and weed detection to eliminate the influence of astigmatism on spectral sensors. Experimental verification results showed that the system can achieve high reflectivity, measurement accuracy, and efficiency. Duncan et al. [208] designed a low-cost multispectral plant detection sensor and the experimental results showed that the near-infrared bands of 810 and 860 nm had the best response to vegetation and soil.

Table 6. Research methods and results of spectral detection for some vegetables and weeds.

Platforms	Crops	Weeds	Sensors	Resolution Ratio	Bands	Detection Algorithms	Precision	References
Handheld	Sorghum	Amaranthus, Chenopodium retusum, Malva sylvestris, Cyperus rotundus Linn., Alloteropsis cimicina	Hyperspectral	-	20 bands, 325–1075 nm	Stepwise linear discriminant analysis (SLDA)	80–100%	Che Ya et al. [209]
UAV			Multispectral	0.87 mm	440, 560, 680, 710, 720, and 850 nm	OBIA	93%
Aircraft	Corn	Sorghum halepense, Xanthium strumarium, Abutilon theophrasti	Hyperspectral	2 m	21 bands, 456–1650 nm	Spectral angle mapper and Spectral mixture analysis	60–80%	Martin et al. [210]
Tractor	Lettuce	Broadleaf herba	Multispectral	0.95 mm	525, 650, and 850 nm	Otsu’s multi-level threshold	>81%	Elstone et al. [211]
Flat bed truck	Field pea, spring wheat, canola	Avena fatua Linn., Amaranthus retroflexus L., Chenopodium retusum	Hyperspectral	1.25 mm	61 bands, 400–1000 nm	Modified Chlorophyll Absorptance Reflectance Index; Principal component analysis and stepwise discriminant analysis	88–94%	Eddy et al. [212]

Table 6. Research methods and results of spectral detection for some vegetables and weeds.

Platforms	Crops	Weeds	Sensors	Resolution Ratio	Bands	Detection Algorithms	Precision	References
Handheld	Sorghum	Amaranthus, Chenopodium retusum, Malva sylvestris, Cyperus rotundus Linn., Alloteropsis cimicina	Hyperspectral	-	20 bands, 325–1075 nm	Stepwise linear discriminant analysis (SLDA)	80–100%	Che Ya et al. [209]
UAV			Multispectral	0.87 mm	440, 560, 680, 710, 720, and 850 nm	OBIA	93%
Aircraft	Corn	Sorghum halepense, Xanthium strumarium, Abutilon theophrasti	Hyperspectral	2 m	21 bands, 456–1650 nm	Spectral angle mapper and Spectral mixture analysis	60–80%	Martin et al. [210]
Tractor	Lettuce	Broadleaf herba	Multispectral	0.95 mm	525, 650, and 850 nm	Otsu’s multi-level threshold	>81%	Elstone et al. [211]
Flat bed truck	Field pea, spring wheat, canola	Avena fatua Linn., Amaranthus retroflexus L., Chenopodium retusum	Hyperspectral	1.25 mm	61 bands, 400–1000 nm	Modified Chlorophyll Absorptance Reflectance Index; Principal component analysis and stepwise discriminant analysis	88–94%	Eddy et al. [212]

Fluorescence Detection

Plant stems and leaves with inherent characteristics will emit a portion of the light energy in a fluorescent manner under the irradiation of specific light. This is why fluorescence technology with high sensitivity is used for plant detection. The two most commonly used types of fluorescence, which can be induced by laser, ultraviolet, or red–blue light, include blue–green fluorescence in the range of 400–600 nm and chlorophyll fluorescence in the range of 650–800 nm. Fluorescence detection is a unique detection technique, but it still has certain limitations and can only be used for some plants.

Table 7. Research methods and results of fluorescence detection for some vegetables and weeds.

Crops	Weeds	Induced Light Source	Collecting Spectral Bands	Optimal Spectral Bands	Analysis Methods	Accuracy	References
Corn	Monocotyledonous weeds and dicotyledonous weeds	Ultraviolet	400–760 nm	-	Principal component analysis and linear discriminant analysis	91.8%	Longchamps et al. [213]
				400–425 nm and 425–490 nm	A partial least squares discriminant analysis (PLS-DA)	>91.7%	Panneton et al. [214]
Crop	Weeds, fake leaves, soil, and plant residues	Blue, red, and white light	650–800, 675–825 and 500–950 nm	685 nm(blue) and 740 nm(red)	Mean normalisation and standard normal variate transformation (SNV); principal component analysis combined with linear discriminant analysis (PCA–LDA) and SVM	100%	Wang et al. [215]
Green plants, simulated green plants, and soil			650–850 nm	731.1 nm(white), 730.76 nm(blue), 731.1 nm(red)	Soft independent modelling of class analogy (SIMCA) and linear discriminant analysis (LDA)	>92%	Wang et al. [216]
Winter wheat	Herbicide resistant weeds	Blue light	>680 nm	-	-	95%	Wang et al. [217]
Corn	-	Laser light	350–820 nm	680 and 740 nm	-	-	Lednev et al. [218]

shows some research methods and results of vegetables and weeds detection by fluorescence detection technology. It can be seen that the detection accuracy of the technology is relatively high, exceeding 90%, but its application in vegetables is also limited.

Ultrasonic Detection

Ultrasonic sensors measure the distance between the sensor and the plant by emitting ultrasonic waves and receiving reflection signals, thereby detecting the height information of the plant. Zhao et al. [219] used ultrasound technology to detect the canopy height of wheat and studied the detection accuracy of ultrasound at different operating speeds. The results showed that the average detection deviation for five types of wheat canopy densities was 0.14 m when the moving speed was between 0.3–1.2 m/s. Ultrasonic sensors, which can also be used to assess plant biomass, covered amount, weed density, etc., have a fast detection speed, low cost, and stability, and can relatively simply characterise plant structures in complex field conditions. However, ultrasonic sensors are greatly affected by leaf obstruction and cannot specifically distinguish between plant species [220].

LiDAR Detection

LiDAR, which can emit high-frequency laser beams to plants and receive signals reflected back from the plants, can obtain information about the posture, shape, distance, and height of plants, and present them in the form of point clouds. Compared with ultrasonic sensors, LiDAR has higher measurement frequency, accuracy, and measurement area, and has great potential in plant phenotype and target detection [221,222].

Table 8. Research methods and results of LiDAR detection for some vegetables and weeds.

Platforms	Crops	Weeds	Sensors	Angle Resolution	Frequency	Analysis Algorithms	Precision	References
Tractor	Cabbage, leek, potato, and wheat	Leymus chinensis Tzvelev	LMS111 LiDAR	0.5°	50 Hz	Based on Euclidean distance	>85%	Krus et al. [223]
Fixed bracket	Wheat	Avena fatua, Sonchus oleraceus L.	MRS6000 LiDAR	0.13°	10 Hz	Euclidean clustering	100%	Shahbaz et al. [224]
Unmanned ground vehicles	Crisphead lettuce, wild lettuce, romaine, stem lettuce, butterhead lettuce, and loose-leaf lettuce	-	VLP-16 LiDAR	0.1°–0.4°	-	Random sampling consistency (RANSAC), Euclidean clustering, and K-means clustering algorithm	-	Cai et al. [225]
Robot	Corn	-	LMS111 2D-LiDAR	-	-	Plant detection with Euclidian clustering (PDEC)/Iterative plant clustering Method (IPCM)	73.7%/100%	Reiser et al. [226]
Moving platform		-	SICK LMS4121R-13000 LiDAR	0.1°	600 Hz	RF	89.4%	Forero et al. [227]
A movable tripod	Cabbage, tomato, and eggplant	-	Terrestrial laser scanner	-	-	Deep convolution neural network (CNN) model-CropPointNet	90% (cabbage, eggplant); 70% (Tomato)	Jayakumari et al. [228]
Tractor	Tomato	-	LMS 111 LiDAR laser scanner	0.5°	50 Hz	K-means clustering	96%	Martinez- Guanter et al. [229]

shows some research methods and results of crop and weed detection by LiDAR detection.

Combination Sensors Detection

Combination sensors use multiple sensors for plant detection. Guo et al. [230] developed a high-throughput platform for plant crop phenotype analysis with multi-sensor fusion, which utilises LiDAR sensors, high resolution cameras, thermal cameras, and hyperspectral image sensors. The platform can obtain the parameters of plant shape and vegetation index. Zhang et al. [231] designed a LiDAR and multispectral sensors detection system based on drones, which was used to evaluate the leaf area index of snap beans and can also be used for sugar beets, soybeans, and winter wheat. Weis et al. [232] used LiDAR, ultrasonic sensors, RGB cameras, and spectral sensors to detect the plant characteristics of spring barley and oilseed rape under different conditions (planting density and herbicide stress) and compared two different data fusion methods.

Table 9. Research methods and results of combination sensors for some vegetables.

Platforms	Crops	Sensors	Detection Algorithms	Precision	References
Tractor	Sugarcane	BA2M-DDT Photoelectric sensor; HC-SR04 Ultrasonic sensor	Decision Tree (DT)	>90%	Maldaner et al. [233]
Moving platform	Maize	RGB and LiDAR sensors	Deep learning-SSD	>90%	Wang et al. [234]
Y-12 aircraft	Shelter forest, cereal crops (maize, wheat), and vegetables (leek, lettuce, cauliflower, potato, watermelon, and pepper)	Compact Airborne Spectrographic Imager (CASI) and Leica ALS70 LiDAR	OBIA	90.33%	Liu et al. [235]
Fixed platform	Different types of blades	Reflection and fluorescence spectra and LiDAR	RF	90.69%	Chen et al. [236]
Remote-controlled ground vehicle	Broccoli and lettuce	RGB and depth sensors	Colour-depth fusion algorithm, ML	96.6% (broccoli), 92.4% (lettuce)	Gai et al. [165]

shows some research methods and the results of crop and weed detection using combination sensors. It can be seen that the detection technology has high precision for vegetable detection, but combination sensors need to achieve synchronous cooperation between sensors and obtain a large amount of data, require more computing power.

3.3.3. Plant Modification Technology

Plant modification technology, which alters the appearance, internal substances, or genes of plants by physical, chemical, or biological techniques, can be used to detect crops and weeds. Su et al. [237] introduced crops signal technology, reviewed the research progress of plant signal technology in crop and weed detection, and discussed the advantages, challenges, and application prospects of the technology. As shown in

Table 10. Research methods and results of plant modification technology for some crops.

Crops	Modification Methods	Marking Materials/ Substances	Marking Position	Detection Methods	Analysis Algorithms	Precision	References
Tomato	Physical	An environmentally friendly straw and an environmentally friendly paint	Main stem	Colour mark sensor	Method based on colour and threshold	95.19%	Li et al. [243]
Soybean	Chemical	Rhodamine B (Rh-B)	Seeds	Green light-induced fluorescence method	Improved Multi-View Positioning Algorithm	96.7%	Jiang et al. [238]
Snap bean	Chemical	Rh-B	Seeds	White/green/UV light-induced fluorescence method	Method based on colour and threshold	-	Su et al. [244,245,246]
Tomato, lettuce	Chemical	Rh-B	Roots	Sunlight/UV-induced fluorescence method	Machine vision (MV)	-	Raja et al. [28,239,240,241]
	Biological	Green fluorescent protein	Leaves	Blue light-induced fluorescence method	Method based on colour and threshold	-
	Biological	Lc maize anthocyanin regulatory gene	Seeds	Sunlight-induced fluorescence method	Method based on colour	-
	Physical	Maize-based plastic straws (water-based latex fluorescent green/orange striping paint)	Independent labels	Sunlight/UV-induced fluorescence method	MV	97.8%
	Chemical	Water-based latex fluorescent paint	Stems	UV-induced fluorescence method	MV	98–100%
	Chemical	Water-based synthetic polymer-based paints	Foliage	White light/UV-induced fluorescence method	MV	99.7%
	Physical	Biodegradable beverage straws (green or orange fluorescent water-based paint)	Independent labels	Sunlight/UV-induced fluorescence method	MV	-	Kennedy et al. [74]
	Chemical	Green or orange fluorescent water-based paint	Lettuce foliage, tomato seedlings
Soybean	Chemical	Rh-B	Seeds	Green light-induced fluorescence method	Method based on colour and threshold	97%	Su et al. [242]
Celery	Chemical	Rh-B	Seedlings			-	Su et al. [247]

, plant modification technologies can be divided into physical modification, chemical modification, and biological modification. Physical modification involves setting physical labels (such as biodegradable plastics, fluorescent coatings, etc.) on or near the surface of plants, and then using machine vision technology for crop recognition. Chemical modification involves applying compounds (such as fluorescent compounds, tracers, etc.) to plant seeds, roots, stems, or leaves for global or local labelling, and then detecting crops by inducing fluorescence methods or others technologies. Biological modification can genetically modify plants to produce substances that can be used for fluorescence detection. Some research results indicate that the application of plant modification technology in plant detection is feasible and has a high detection accuracy, but it requires a significant amount of labour, cost, or time [238,239,240,241,242,243,244,245,246].

4. Weeding Actuators and Robots

Researchers and agricultural companies have developed a series of weeding actuators that can be divided into chemical weeding, mechanical weeding, physical weeding, and combination weeding for weed control operations [248,249,250]. Roberts et al. [80] reviewed new technologies for weed control, such as chemical, mechanical, electrical, laser, and thermal weed control, and discussed the challenges and future development trends of these new technologies in weed control. Fennimore et al. [248,249,250] introduced some automated and mechanised weed control technologies that can distinguish between crops and weeds for selective weed control.

Weeding robots that integrate plant recognition and weed control technologies has also been developed by researchers and agricultural companies. Among them, some developed countries (such as the United States, Germany, the United Kingdom, and Australia, etc.) have developed a series of mature weeding robots and have been widely used them locally, while weeding robot technology in other countries is relatively rudimentary [83,84,251]. The walking platforms used by weeding robots generally include drones, robots, and tractors. Li et al. [83] introduced the weeding robots based on machine vision that has been developed by researchers in the past thirty years; Zhang et al. [84] reviewed different weed detection methods and weed control robots for precise weed management, summarised the development trends in this field, discussed the limitations of existing weeding systems, and proposed ideas for future research directions. Gerhards et al. [251] compared seven new weeding robots and systems and conducted experiments in sugar beet and winter rapeseed fields to evaluate the weed control effectiveness, work efficiency, and cost of each weed control system. The experimental results indicate that weeding robots have great potential in the weeding of vegetables or other crops and can reduce the use of herbicides.

4.1. Intelligent Chemical Weeding

With the development of machine vision and other technologies, intelligent chemical weeding technology, which can achieve precise spraying of weeds under the premise of weed detection technology and apply different types of herbicides according to weed resistance, is gradually becoming widely used in green agricultural production. The technology can reduce the use of herbicides and has a high weeding efficiency, but it has not completely eliminated the use of herbicides. Allmendinger et al. [252] reviewed the precise chemical weeding technology for precision patch spraying and spot spraying based on convolutional neural network detection and introduced tractor spray and autonomous robots. Özlüoymak et al. [253] designed a novel camera-integrated spraying needle nozzle for precise weed control. The camera can detect the position of weeds, and the nozzle can spray automatically. The experimental results showed that the system had a small spraying error and good spraying effect.

Table 11. Intelligent chemical weeding robots.

Robots	Photos	Platforms	Crops	Actuators	Detection Methods	Spraying Methods	Work Efficiency	Save Pesticide Dosage	Developers/Companies
Mobile robot	-	Self-developed robot	-	Lechler standard flat fan nozzle	MV	Micro dose directional spraying	-	95%	Özlüoymak et al. [254]
Smart variable rate sprayer	-		Potato	Spraying nozzle	DL	Variable spraying	-	>40%	Hussain et al. [255]
Robot	-		Peucedani radix, tomato, and eggplant	Parallel robotic arms and circular nozzles	DL	Precise spraying	-	-	Zhang et al. [256]
Sprayer	-		Lettuce	Precision spray assembly	Fluorescence detection	Precise spraying	-	-	Raja et al. [257]
-	-	Tractor	Carrot	Injector nozzle	MV	Real-time, variable-rate herbicide application	-	>30%	Dammer et al. [258]
Asterix		Robot		Drip irrigation nozzle	MV	Drip irrigation	-	90%	Utstumo et al. [259]
Smart sprayer		Tractor	Sugar beet, sunflower, and maize	Spot sprayer	MV	Targeted spraying	-	10–55%	Spaeth et al. [260]
ARA smart sprayer		Tractor	Pastures, field vegetables, large crops, and lawns	A strip with 156 nozzles, spaced apart every 4 cm	Multi-camera vision system	Ultra-precise spraying	4 hm2/h	95%	Ecorobotix (ecorobotix.com, URL (accessed on 29 June 2024))
AVO		Solar powered platform	-	-	-	Smart and ultra-ecological spraying	10 hm2/d	95%
Aviro D15 Pro		UAV	-	1 to 12 nozzles options with cone/jet options	AI prescription map	Point to point spraying	4.08 hm2/h	-	Avirtech (avirtech.co, URL (accessed on 29 June 2024))
XAG XP 2020			-	Rotary atomisation nozzle		Variable rate spraying	10 hm2/h	-
WEED-IT		Tractor	-	-	WEED-IT detection sensor	Spot spraying/Dual spraying/Variable rate spraying	90 hm2/h	-	WEED-IT (weed-it.com, URL (accessed on 29 June 2024))
WeedSeeker 2		Tractor	-	Fixed nozzle	Infrared sensor and high-resolution blue LED-spectrometer	Spot spraying	-	90%	Trimble Agriculture (ptxtrimble.com, URL (accessed on 29 June 2024))
AX-1		Robot	Carrots, parsley root, spinach, radish, rocket, baby leaves, and celeriac	A new type of nozzle	RTK	Selective spraying	-	-	Kilter (kiltersystems.com, URL (accessed on 29 June 2024))

shows some intelligent chemical weeding robots and weeding methods developed by researchers and agricultural companies. In addition to the weeding robots shown in Table 11, some intelligent chemical weeding robots have also been developed by agricultural companies, such as Garford (garford.com, URL (accessed on 28 June 2024)), Agrifac (agrifac.com, URL (accessed on 28 June 2024)), AgriCon (agriconagroproducer.com, URL (accessed on 28 June 2024)), Fobro Mobil D49/D92 (fobro-mobil.com, URL (accessed on 28 June 2024)), Greeneye Technology (greeneye.ag, URL (accessed on 28 June 2024)), Goldacres (goldacres.com.au, URL (accessed on 28 June 2024)), Agrointelli (agrointelli.com, URL (accessed on 28 June 2024)), Blue-River Technology (bluerivertechnology.com, URL (accessed on 28 June 2024)), Bilberry, Amazone, etc. It can be seen that intelligent chemical weeding robots based on drones, tractors, and robots have been developed for the weeding of some vegetables, which use different weed detection methods, weeding actuators, and spraying methods. Some robots have good work efficiency and have achieved good effects in saving herbicides.

4.2. Intelligent Mechanical Weeding

Intelligent mechanical weeding, which can overcome the problem of crop damage caused by traditional mechanical weeding and can also remove intra-row weeds, has gradually become widely used in vegetable weeding. Xiang et al. [81] reviewed intelligent mechanical weeding technology, introduced various mechanical actuators, and discussed their limitations and future development trends. Some researchers have developed and tested intelligent mechanical weeding actuators [261,262,263], but the mechanisms have not been applied in actual production.

Table 12. Intelligent mechanical weeding robots.

Robots	Photos	Platforms	Crops	Actuators	Detection Methods	Inter-Row/Intra-Row Weeding	Work Efficiency	Weeding Rate	Developers/Companies
Intelligent intra-row robotic weeding system		Mobile robot platform	Maize	Vertical disc weeding knife	DL	Intra-row	-	85.91%	Quan et al. [264]
FD20		Solar powered platform	Sugar beets, beetroots, onions, spinach, rapeseed, and different herbs	Hoeing wires inter-row, cutting knifes in-row	RTK-GPS records exact position of each crop seed	Inter-row and intra-row	-	-	FarmDroid (farmdroid.co.uk, URL (accessed on 30 June 2024))
ROBOTTI 150D		Automatic navigation platform	Sugar beets, basil, lettuce, maize, onions, faba beans, parsnips, white cabbage, etc.	Hoeing inter-row and finger or horizontal hoeing in-row	Deep learning and bigdata	-	0.5–2 hm2/h	-	Agrointelli (agrointelli.com, URL (accessed on 30 June 2024))
Bonirob		Robot	Carrot	Rotating shovel or tooth	ML	-	-	90%	Deeofield Robotics (deepfield-robotics.com, URL (accessed on 30 June 2024))
Farming Revolution W4		Robot	Bok choy, broccoli, cabbage, carrot, cauliflower, lettuce, corn, garlic, onion, soy bean, pumpkin, sugar beet, etc.	Side-cut knives inter-row, rotary hoes in-row	Multispectral and CNN	Inter-row and intra-row	70 hm2/week	100%	Farming Revolution (farming-revolution.com, URL (accessed on 30 June 2024))
Fobro Mobil B38		Self-developed tractor	-	Hoeing wires inter-row and finger weeding in-row; FOBRO row hoeing brush	-	Inter-row and intra-row	-	-	Fobro Mobil (fobro-mobil.com, URL (accessed on 30 June 2024))
-		Tractor	Sugar beet, lettuce, broccoli, cabbage, carrot, corn, etc	Horizontal hoeing	MV	Inter-row and intra-row	1.2 hm2/hour	-	FarmWise (farmwise.io, URL (accessed on 29 June 2024))
Robocrop InRow Weeder		Tractor	Lettuce, cabbage, celery, etc.	Hoeing inter-row; rotating notched disc in-row	Robocrop video image analysis techniques	Inter-row and intra-row	-	-	Garford (garford.com, URL (accessed on 29 June 2024))
WEAI		Battery powered Robot	Onions, beetroots, carrots, etc.	-	Crop and weed detection based on AI	Intra-row	10 hm2/time	-	Ekobot (ekobot.se, URL (accessed on 29 June 2024))
Mechanical Robovator		Tractor	Lettuce, onion, etc.	Automatic hoeing in the row	MV	Intra-row	-	-	Visionweeding (visionweeding.com, URL (accessed on 29 June 2024))

shows some intelligent mechanical weeding robots developed by researchers and agricultural companies. In addition, there are other agricultural companies that have developed similar robots: Kverneland (ien.kverneland.com, URL (accessed on 29 June 2024)), Naïo-technologies (naio-technologies.com, URL (accessed on 29 June 2024)), Steketee IC-Weeder AI, etc. The weeding robots can identify vegetables or weeds by machine vision and sensor technology and then control the mechanical weeding actuators to remove the inter-row and intra-row weeds.

Table 13. Mechanical weeding tools.

Weeding Tools	Photos	Inter-Row/Intra-Row Weeding	Actuated Methods	Developers/ Companies
V-shaped hoeing and finger		Inter-row and intra-row	Soil resistance	KULT Kress (kult-kress.com, URL (accessed on 29 June 2024))
Hoeing		Intra-row	Hydraulic or electric power
Rotating notched disc		Intra-row	Electric power	Fennimore et al. [265] Tillett et al. [266]
Torsion bar		Intra-row	Soil resistance	Vandereide et al. [267]
Hoeing brush		Inter-row and intra-row	PTO 540 rpm or hydraulic	Fobro Mobil (fobro-mobil.com, URL (accessed on 29 June 2024))
Horizontal hoeing		Intra-row	Hydraulic	FarmWise (farmwise.io, URL (accessed on 29 June 2024))
Vertical disc weeding knife		Intra-row	Servo motor	Quan et al. [264]
Weeding knife		Intra-row	Pneumatic cylinders	Raja et al. [246]

shows some mechanical weeding tools developed by researchers and agricultural companies. In addition, other companies have also developed mechanical weeding components, such as KULT Kress (kult-kress.com, URL (accessed on 29 June 2024)), which has developed comb teeth and the habicht hoe, etc. These weeding tools and robots can provide references for the development of intelligent mechanical weeding technology.

4.3. Intelligent Physical Weeding

Physical methods, such as electric shock [268], laser [269], thermal weeding [71], etc., which have good weeding effects, are used for weed control. The widespread use of physical weeding has the problem of high cost, while point-precise physical weeding, which can be achieved by plant recognition technology, can reduce the cost problem to some extent. Physical weeding has the potential to successfully control weeds and has been studied by researchers for a long time [268,269]. Electric weeding is the process of killing plant cells by short bursts of high voltage electricity to achieve the weeding objective [69]. The weeding method involves small areas of contact with plant stems, leaves, or roots, and cannot completely kill weeds. More contact time and energy output are required to completely remove weeds, which undoubtedly increases costs. Bloomer et al. [24] reviewed the traditional and new technologies for vegetable weed management and separately introduced the pulse electric microshock weed control technology, which can be combined with automation technology and artificial intelligence to create an autonomous, ultra-low energy consumption, selective, non-chemical weeding system. Laser weeding kills weeds with laser beams using laser-emitting equipment, including carbon dioxide lasers, diode lasers, and fibre laser equipment, etc. [73]. Thermal weeding, including flame [63], steam [70], hot foam, or hot oil weeding, etc., can kill weeds using high temperatures and energy [71,72].

Table 14. Intelligent physical weeding robots.

Robots	Photos	Platforms	Crops	Weeding Methods	Actuators	Developers/ Companies
AI-powered laser weeding robot		Tractor	100+ crops	Laser weeding	30X 150W CO2 10.6 μm lasers with tracking cameras	Carbon Robotics (carbonrobotics.com, URL (accessed on 29 June 2024))
CLAWS robot		100% battery and solar-powered robot	Brassicas and lettuce	Concentrated light autonomous weeding	Short pulse of concentrated light	Earth Rover (earthrover.farm, URL (accessed on 29 June 2024))
Tensorbot		V2 robot	-	Hot oil weeding	Canola oil 160 °C	Tensorfield Agriculture (tensorfield.ag, URL (accessed on 29 June 2024))
Thermal robovator	-	-	Sugar beets	Flame weeding	Propane flames	Visionweeding (visionweeding.com, URL (accessed on 29 June 2024))
eWeeding		Tractor	-	Electric weeding	High frequency alternating current	RootWave (rootwave.com, URL (accessed on 29 June 2024))
Annihilator 12R30/16R30, Terminator T1/T2/T3			Alfalfa, oats, rye peas, clover	Electric weeding	A copper bar that has voltage on it	The Weed Zapper (theweedzapper.oldschoolmanufacturing.com, URL (accessed on 29 June 2024))
XPS			-	Electric weeding	Fixed and movable electrodes	Zasso (zasso.com, URL (accessed on 29 June 2024))

shows some intelligent physical weeding robots developed by researchers and agricultural companies. The weeding robots can identify vegetables or weeds using machine vision and sensor technology, and then control the physical weeding actuators to remove the weeds. These weeding robots can provide references for the development of intelligent physical weeding technology.

4.4. Integrated Weed Management (IWM)

Traditional weeding robots adopt a single weeding method or equipment type. IWM, which combines multiple weeding methods and strategies, is used for weeding by some researchers to improve the weeding effects and efficiency [270,271,272]. Young [270] introduced the integrated weed management method and discussed the future development direction of the technology. The researchers drew the system devices that could be included in the integrated weed management robot, as shown in Figure 1a. The integrated weed management system arranges reasonable weed control strategies based on the field environment and weed types. Bawden et al. [271] developed a modular robot platform, Agbot II (qut.edu.au, URL (accessed on 29 June 2024)), with heterogeneous agricultural weed control arrays and introduced the weed control mechanism that combined chemical and mechanical weed control methods, as shown in Figure 1b. Wu et al. [272] also designed a modular weed control system that included mechanical and chemical weed control tools, as shown in Figure 1c. With the cooperation of the weed detection system, the weeding system can achieve precise weed control. The combination of mechanical and chemical weeding methods can reduce the use of chemical herbicides to a certain extent, and also improve work efficiency.

5. Challenges and Future Development

Intelligent weed control technology for vegetables is one of the mainstream technologies in modern agricultural development. Researchers have conducted extensive research in this field. Although some achievements have been made, there are still many challenges and problems:

(1) In terms of plant detection (including global weed detection, crop-rows recognition, and precise vegetable/weed identification), there is a lot of research on plant detection technology based on machine learning and deep learning methods, while there is relatively little research on detection technology based on near-end sensors and plant modification. Although the detection technologies based on machine learning and deep learning have high detection accuracy, there are still certain limitations. The resolution and information of images collected by collection platforms (drones, ground vehicles, remote sensing, etc.) and sensors (RGB, multispectral sensors, etc.), which are affected by the natural environment, are poor and incomplete. Therefore, most researchers need to collect thousands, or even more, of field images under various natural conditions, which undoubtedly increases the workload and is inevitable [164,165,166,167,168,169]. Traditional image processing techniques have some problems such as poor accuracy and slow processing speed, while machine learning and deep learning technologies can improve recognition accuracy, but the technologies require more time for image processing and feature training [174,175,176,177,178,179]. In addition, the plant characteristics of vegetables and weeds change with the different growth stages, which has a significant impact on detection accuracy. The existing plant detection technology can only identify individual plants and has a narrow application range. Therefore, it is a major problem to accurately distinguish types of crops or weeds and expand the application range [15,78,79]. According to the types of vegetables and existing detection technologies for vegetables, it can be seen that there is still a lack of research on detection technologies for a large number of vegetables and their associated weeds. In conclusion, the detection technology for vegetables and weeds still needs further in-depth research.

(2) In terms of weeding actuators, intelligent chemical weeding has high efficiency and weeding efficacy [256,257,258,259,260], but the application of herbicides has not been completely eliminated, and there are still problems such as environmental and soil pollution, weed resistance, etc. Intelligent mechanical weeding can remove inter-row and intra-row weeds simultaneously; some weeding actuators have a high weeding rate and efficiency [264,265,266,267]. This may be a better weeding method option for farmers. However, there are also some weeding actuators that have not shown weeding effects, which require further field experimental research. In addition, compared with intelligent chemical weeding, intelligent mechanical weeding has lower working efficiency and requires higher technology investment [81]. Physical weeding also has certain limitations. Laser and electric weeding have a certain effect on weed removal, but they cannot completely kill weeds and have lower efficiency and a higher cost, which is not conducive to their use in large-scale vegetable fields [24,78]. Thermal weeding (such as flames, hot oil, steam, etc.) can kill harmful bacteria and insects while removing weeds, but it can also kill beneficial bacteria and insects [60,70,71,72,73]. There are also problems related to a high cost and low efficiency. Therefore, there are a series of problems with existing weeding machinery, and efficient, environmentally friendly, and cost-effective weeding equipment still needs to be developed and improved.

(3) The weeding robots integrate plant detection and intelligent weeding devices, which need to be combined with information and automatic control technologies to complete the weeding operation. The existing weeding robots are still limited by some problems such as image processing speed and data transmission, resulting in low efficiency and poor real-time performance. Merfield et al. [273] proposed ten characteristics and functions that weeding robots should possess, and this article agrees with these viewpoints. Some developed countries have developed a series of mature weeding robots, and these have been widely used locally, while the weeding robot technology in other countries is relatively rudimentary. The vegetable weeding robot shown in this article can provide reference and inspiration for their research.

Despite the problems mentioned above, intelligent weeding technology for vegetables still has certain development potential. In the future development of modern agriculture, the technology will become one of the mainstream technologies in vegetable production. In response to existing problems, the future development direction is as follows: In terms of plant detection, the impact of the natural environment can be addressed by manual intervention or multi-sensor methods. The collection technology should be combined with actual field conditions, so that the collected images can reflect real field information. For image processing problems, new algorithms or multi-algorithm fusion methods can be used to minimise time costs as much as possible. In precision agriculture, future plant detection technology should be able to accurately identify vegetables and weeds at different growth stages, as well as distinguish vegetable varieties and weed species. In terms of weeding actuators, integrated weed management will become a mainstream technology to overcome the limitations of a single weeding method by integrating multiple weed control methods. As for weeding robots, the adaptability and applicability of weeding robots in complex agricultural environments should be improved, especially in terms of weeding efficiency. With the development of sensor technology, weeding robots will integrate multiple sensors for field operations in the future, enabling them to have multiple functions such as “hands”, “feet”, “eyes”, and a “brain”.

6. Conclusions

This review provides a detailed overview of intelligent weeding technology for vegetables from the perspective of vegetable weed detection technology, weeding actuators, and robots. In terms of vegetable weed detection technology, global weed detection, crop-rows detection, and precise vegetable/weed recognition technologies were introduced. Global weed detection can create a weed map and arrange appropriate weed control strategies based on the weed map; crop-rows detection can be used to detect inter-row weeds and correct weed control trajectories; the precise identification technology for vegetables/weeds can be used for precise weed control to avoid damaging seedlings. Researchers have achieved significant results in the plant detection technology for vegetables using various sensors and classification algorithms, but there are still some problems. The next step should be to conduct research on issues such as the impact on the natural environment, slow processing speed, and poor real-time performance. In terms of weeding actuators and weeding robots, intelligent chemical weeding, mechanical weeding, physical weeding, and integrated weed management methods were described. Although some researchers and agricultural companies have developed various weeding actuators and robots, some still suffer from high costs, low weeding efficiency, and high seedling damage rates. At the end of the article, the problems and future development directions of intelligent weeding technology were discussed and analysed. The systematic analysis method is a popular literature review method nowadays, but this review adopts statistical and descriptive methods, which may be a limitation of this paper. We hope that this review presents the research status of intelligent weeding technology for vegetables from several aspects, which can provide a reference for future research in this field.