Precision Weed Control Using Unmanned Aerial Vehicles and Robots: Assessing Feasibility, Bottlenecks, and Recommendations for Scaling

Abstract

Weeds cause significant yield and economic losses by competing with crops and increasing production costs. Compounding these challenges are labor shortages, herbicide resistance, and environmental pollution, making weed management increasingly difficult. In response, precision weed control (PWC) technologies, such as robots and unmanned aerial vehicles (UAVs), have emerged as innovative solutions. These tools offer farmers high precision (±1 cm spatial accuracy), enabling efficient and sustainable weed management. Herbicide spraying robots, mechanical weeding robots, and laser-based weeders are deployed on large-scale farms in developed countries. Similarly, UAVs are gaining popularity in many countries, particularly in Asia, for weed monitoring and herbicide application. Despite advancements in robotic and UAV weed control, their large-scale adoption remains limited. The reasons for this slow uptake and the barriers to widespread implementation are not fully understood. To address this knowledge gap, our review analyzes 155 articles and provides a comprehensive understanding of PWC challenges and needed interventions for scaling. This review revealed that AI-driven weed mapping in robots and UAVs struggles with data (quality, diversity, bias) and technical (computation, deployment, cost) barriers. Improved data (collection, processing, synthesis, bias mitigation) and efficient, affordable technology (edge/hybrid computing, lightweight algorithms, centralized computing resources, energy-efficient hardware) are required to improve AI-driven weed mapping adoption. Specifically, robotic weed control adoption is hindered by challenges in weed recognition, navigation complexity, limited battery life, data management (connectivity), fragmented farms, high costs, and limited digital literacy. Scaling requires advancements in weed detection and energy efficiency, development of affordable robots with shared service models, enhanced farmer training, improved rural connectivity, and precise engineering solutions. Similarly, UAV adoption in agriculture faces hurdles such as regulations (permits), limited payload and battery life, weather dependency, spray drift, sensor accuracy, lack of skilled operators, high initial and operational costs, and absence of standardized protocol. Scaling requires financing (subsidies, loans), favorable regulations (streamlined permits, online training), infrastructure development (service providers, hiring centers), technological innovation (interchangeable sensors, multipurpose UAVs), and capacity building (farmer training programs, awareness initiatives).

1. Introduction

With the world’s population projected to reach 9.0 billion by 2050, marking a substantial 32% increase since 2015, the demand for food is set to surge [1]. Consequently, agricultural output must rise 25% to 75% compared with current levels to safeguard food security. However, weed management is a pivotal concern in modern agriculture, given its significant impact on yields and production costs [2]. Globally, 1800 weed species contribute to a 31.5% yield loss, translating to an annual economic loss of USD 32 billion [3]. Depending on the region and the severity of weed infestations, crop yield reductions vary between 10% and 80% [4]. For instance, weeds cost grain growers AUD 3.3 billion annually in Australia, translating to a nationwide loss of 2.7 million metric tons (MT) of grain [5]. Similarly, in India, yield losses range from ~ 14 to 36% across ten major crops, equating to USD 11 billion annually [6]. In the USA, weeds caused losses of USD 33 billion in 2005 [7] and USD 8 billion in 2017 [8]. China loses around 3 MT of grain annually due to weed infestation [9], while in Africa, yield reductions typically reach around 30%, with certain regions of sub-Saharan Africa experiencing losses exceeding 50% [10]. These losses amount to 2.2 MT annually, valued at USD 1.45 billion, roughly half the region’s rice imports [11].

1.1. Challenges in Conventional Weed Management Approaches

Herbicides have been instrumental in modern agriculture, enabling efficient weed control and increased crop yields. However, indiscriminate herbicide use, including blanket applications and multiple seasonal treatments, leads to herbicide resistance development in weeds [12], biological invasions, and herbicide hazards to farm workers [13]. Weed control costs have eroded farm profitability, especially in developed countries, due to increasing herbicide resistance. Globally, 269 herbicide-resistant weed species, comprising 154 dicots and 115 monocots, have been reported across 99 crops in 72 countries, with the highest number of herbicide-resistant weeds reported in the U.S. and Canada [14].

The adoption of herbicide-tolerant crops like rice, corn, cotton, canola, soybean, sugar beet, alfalfa, and brassica has offered a solution to address the limitations of conventional weed control methods, marking a significant shift in weed management practices worldwide, particularly in countries like the United States, Canada, and Australia [15]. However, the widespread adoption of herbicide-tolerant crops (monoculture) and overreliance on a single herbicide accelerated the development and spread of herbicide-resistant weeds [16]. The resistant weed populations are less affected by once-effective herbicides, necessitating higher doses and more potent herbicides, resulting in higher weed control costs, environmental damage, and adverse impacts on beneficial organisms [17]. Moreover, only a few herbicides with a new mode of action have been discovered in the past 30 years, making them a scarce resource [18].

Glyphosate, a wonder herbicide of modern farming, primarily for its quick and non-selective weed control, now faces increased scrutiny due to its potential health and environmental concerns. Several countries, including members of the European Union (EU), have restricted or considered restricting the use of glyphosate [19]. Such restrictions on glyphosate could have significant repercussions on agricultural production. For example, without glyphosate, in the EU, wheat losses could reach 24 MT, valued at EUR 10.5 billion, and potato production could drop by 10.4 MT, amounting to a loss of EUR 2 billion [19]. Furthermore, the wine industry could lose 4.7 MT of grapes, worth EUR 4.2 billion, affecting both production and exports. Moreover, restrictions on glyphosate might prompt an increase in soil tillage as an alternate method for weed control. This shift could bring other adverse environmental consequences, such as soil structure degradation, soil erosion, reduced soil biodiversity, and increased greenhouse gas (GHG) emissions, with estimates of up to 3.8 MT of CO₂ equivalent per year across the EU [19].

1.2. Weed Menace in Organic Agriculture

Global organic farming has expanded significantly, from 17.16 million hectares (Mha) in 2000 to 76.4 Mha by 2023 [20]. Australia leads with 35.7 Mha, followed by Argentina with 3.7 Mha and Spain with 2.4 Mha [21]. However, weeds remain a major challenge in organic farming, contributing substantially to production costs and yield losses [22]. Weed dockage in organic produce reduces its economic value and increases shipping costs [23]. Manual hand-pulling and hoeing are the common weed control methods in organic farming. However, these methods are constrained by their limited efficiency in effectively controlling weeds. Moreover, hand-weeding is labor-intensive, expensive, and time-consuming, especially for large-scale weed management [24]. The use of mechanical hoeing and rotating blades leads to increased soil erosion and nutrient loss [25], heightened organic matter degradation, weed seed dormancy breaking, and new weed growth [26].

1.3. Shrinking Agricultural Labor Force and Increasing Labor Wage

As economies progress, there is a notable decline in the proportion of people employed in agriculture. Rural youth increasingly choose jobs in industry and services due to their perceived higher social status. This shift has led to an aging farming workforce, higher labor wages, and reduced productivity [27]. In India, agricultural employment fell from 64.6% in 1993–1994 to 42.5% in 2018–19. Similarly, in Asia and sub-Saharan Africa, agricultural employment dropped from 52.6% and 68.4% in 1990–1999 to 45% and 41.2% in 2000–2013 [28]. Manual weeding has traditionally been the most common weed control method, but rising labor costs and shortages have led to increased reliance on herbicides [29].

1.4. Precision Weed Control

Precision weed control (PWC) is an advanced method for managing weeds by targeting herbicides or mechanical measures to specific field areas where weeds are present, unlike traditional methods that apply treatments uniformly across entire fields [30]. PWC using robots and unmanned aerial vehicles (UAVs) allows farmers to identify and treat weed-infested areas with higher accuracy [31], optimizing resource use like herbicides and labor, reducing waste, and cutting costs [32]. Such precision application minimizes herbicide exposure in non-targeted areas, thereby reducing the environmental impact and the risk of soil and water pollution while cutting costs [32,33]. Smart sprayers like See-and-Spray use AI-powered cameras to detect weeds in real time, reducing herbicide use by 70–90% in row crops [34]. Directed energy systems, such as laser weeding and electric pulse technology, provide non-chemical alternatives ideal for organic farming [35]. Precision mechanical tools like GPS-guided inter-row hoes physically uproot weeds, eliminating herbicides in targeted zones [36,37,38]. Hybrid systems combine lasers and mechanical tools, merging chemical precision with mechanical efficiency.

In contrast to traditional blanket herbicide applications, which often contribute to herbicide resistance, runoff, and environmental pollution [34]. The excess water (e.g., rain, irrigation) carries unabsorbed herbicides by soil or plants into nearby waterways, often due to sloped terrain or compacted soil. Apart from increasing yields and maximizing resource use efficiency, PWC technologies also serve as drivers for economic growth by creating jobs and supporting rural development [39]. Robotic systems enable precise and automated weed control. Lasers attached to the robotic weeder provide non-chemical weed control [40]. UAVs facilitate aerial monitoring and mapping of weeds and site-specific herbicide spraying [41,42]. These technologies, designed to minimize pesticide use, align with environmentally sustainable and friendly alternatives in weed management.

While a growing body of research explores the application of UAVs and robotic systems for weed control, a critical gap remains in understanding the practical feasibility and inherent bottlenecks hindering their widespread adoption across large-scale agricultural operations. The existing literature predominantly focuses on proof-of-concept studies and small-scale trials, often overlooking the significant challenges associated with scalability to cover extensive farmlands efficiently within critical operational windows. Furthermore, the economic viability of deploying and maintaining large fleets of these advanced systems, considering the substantial upfront investment and ongoing operational costs, has not been thoroughly investigated. Logistical hurdles concerning infrastructure requirements for deployment, data management, and reliable communication across vast areas also remain largely unexplored. The influence of the evolving regulatory landscape surrounding UAV operations in agriculture and potential technical limitations in areas like robust weed detection and reliable robotic performance in diverse field conditions further contribute to this knowledge gap. Finally, the crucial aspect of farmer adoption, including training needs and integration with existing farming practices at a large scale, has yet to be comprehensively addressed. To address these knowledge gaps, we conducted a review of 155 published articles, organizing the findings into four sub-headings:

• Artificial intelligence in PWC: We examined the role of computer vision (CV), machine learning (ML), and deep learning (DL) algorithms in crop–weed classification and mapping.
• PWC using Robots and UAVs: This section covers the efficiency, advantages, and disadvantages of these methods.
• Bottlenecks for adoption: Identifies existing barriers to large-scale use of robots and UAVs in agriculture.
• Interventions for scaling PWC: Propose innovative solutions to overcome these challenges.

2. Methodology of Literature Selection for the Review

This study reviews scholarly articles on UAVs, robotics, laser technology, and PWC. The relevant literature is identified based on the following criteria.

Relevance to scope: Articles focused on advancements, methodologies, applications, and case studies related to UAVs, robotics, laser weeding, and PWC.

Publication date: Articles published within the last two decades (2001–2025) were considered to ensure the inclusion of recent developments. This period marks a significant increase in research and development in both UAV technology and agricultural robotics, coinciding with advancements in sensor technology, AI, and machine learning that are crucial for PWC.

Publication type: Only peer-reviewed articles from reputable journals, conferences, and academic proceedings were included.

Language: Only literature published in the English language was included.

Outcomes of interest: Studies evaluating the technical feasibility, economic viability, adoption challenges, or performance of these technologies.

Databases: A thorough search was conducted across reputable academic databases, including but not limited to PubMed, IEEE Xplore, ScienceDirect, and Google Scholar

Keywords: “Unmanned Aerial Vehicle” OR “UAV” OR “drone” AND “Weed Control” OR “Precision Weed Control” OR “Weed Control Efficiency” AND “Robotics” OR “Automated Weeding” OR “Laser Weeding” AND “Machine Learning” OR “Deep Learning” OR “Artificial Intelligence” AND “RGB Sensors” OR “Multi-spectral” OR “Hyper-spectral” AND “Cost-benefit” OR “Economics” OR “Feasibility” OR “Bottleneck”.

Exclusion: Article published in a language other than English, non-peer-reviewed article, magazine, and not freely available.

A total of 155 articles were systematically reviewed to extract key information on UAVs, robotics, laser technology, and PWC. The complete literature screening process is displayed in Figure 1.

3. Artificial Intelligence in Precision Weed Control

PWC primarily relies on precise field scouting to identify prevalent weed species and map their spatial distribution. Artificial intelligence (AI) enables machines to mimic human intelligence, while CV replicates human visual perception and understanding [43,44]. The integration of AI and CV techniques enhances weed classification and mapping accuracy in real time and applies precise interventions, such as spraying, mechanical weeding, or laser-based weed ablation, improving overall weed management strategies. AI-enabled machines can also adapt their strategies in response to dynamic factors like shifting weed populations, genetic variations, and environmental changes, thereby significantly improving the success of weed management efforts [45]. State-of-the-art CV methodologies, including deep belief networks (DBNs) and convolutional neural networks (CNNs), have demonstrated remarkable effectiveness in weed detection, even under challenging conditions such as complex backgrounds, variable lighting, diverse capture angles, and the presence of weeds in various forms and colors [46].

Similarly, AI systems utilize ML and DL to analyze techniques to analyze large datasets encompassing crop and weed images, environmental conditions, and historical crop management practices. As a result, AI empowers farmers to make data-driven decisions and implement precise interventions tailored to specific field conditions [44,47,48]. While both ML and DL are integral to these systems, they differ fundamentally in their operational approaches. ML typically involves multiple stages, like preprocessing, feature extraction, classification, and output generation. In contrast, DL streamlines the process by taking an input image, processing it through deep neural networks, and directly producing the desired output (Figure 2). This distinction highlights the efficiency and scalability of DL in handling complex agricultural tasks, further solidifying its role in modern precision agriculture.

3.1. Image Classification Using AI Algorithms

The algorithm’s ability to accurately classify a pixel as either a crop or a weed is referred to as detection capacity [49]. Numerous ML/DL algorithms have been developed and evaluated for their accuracy in discriminating weeds within crop fields. These algorithms encompass a diverse set, including convolutional neural networks (CNNs), cluster analysis (CA), artificial neural networks (ANNs), random forest (RF), general discriminant analysis (GDA), linear discriminant analysis (LDA), discriminant analysis (DA), site-specific classifiers (SSCs), support vector machines (SVMs), decision trees (DTs), naive bayes (NB), one-class self-organizing maps (OCSOMs), one-class support vector machines (OCSVMs), neural networks (NNs), mixture of Gaussians (MOGs), self-organizing maps (SOMs), partial least square discriminant analysis (PLSDA), and one-class principal component analysis (OCPCA). Newer and more advanced models such as MOGs, SOMs, and generative adversarial networks (GANs) are showing high accuracy in weed identification and classification [50,51].

A review of existing research highlights that no single algorithm consistently excels in discriminating weeds across all crop types. Different algorithms demonstrate varying degrees of suitability depending on the specific crop under consideration. For example, CA and DA models show 100% accuracy in barley and rocket salad, respectively, while other models like Bayesian and RF exhibit lower accuracies. Certain models are highly effective for specific crops, such as OCSVM for milk thistle (96.1%) [52], and RF with Markov random fields for sugar beet (95%) [53]. These variations in performance across different crops necessitate tailored model architectures or training data for specific crops to achieve optimal weed discrimination accuracy. AI models such as CNNs, ANNs, RF, and DA have demonstrated exceptional accuracy in crop and weed discrimination across different crops, often surpassing 90% and reaching 100% in some cases (Table 1). This underscores the efficacy of AI in precisely detecting weeds in diverse agricultural environments.

CNNs consistently outperformed other models and emerged as a dominant and highly effective model type, with accuracies often reaching or exceeding 98%, particularly in crops like barley (99%), cotton (99.3%), and wheat (98.2%) (Figure 3). One widely adopted CNN is You Only Look Once (YOLO), renowned for its robustness and suitability across diverse applications demanding rapid detection. YOLO has demonstrated success in the identification of weeds, contributing to significant advancements in precision weed control [54]. By integrating UAV imagery with AI, researchers have achieved high accuracy rates (over 90%) in identifying and mapping weeds across different crops, largely due to the high-resolution imagery provided by UAVs (Table 1). Integrating UAV-based weed mapping with other precision agriculture technologies, such as variable-rate applications and automated guidance systems, could further enhance weed control practices.

Kargar and Shirzadifar [55] developed a machine vision weed spot sprayer that achieved over 90% accuracy in distinguishing between grass leaves and maize plants using image segmentation and feature extraction, leveraging the broader width of corn leaves compared with grass leaves. Similarly, Zhang et al. [56] developed a hyperspectral imaging technique paired with a heated oil microspray application system, achieving identification accuracy of 95% for tomatoes, 94% for Solanum nigrum, and an impressive 99% for Amaranthus sp. within the tomato seed lines. Therefore, rigorous testing with independent datasets ensures the model’s accuracy in distinguishing between crops and weeds in agricultural fields. Despite these advancements, developing a robust and versatile classifier capable of distinguishing between soil, weeds, and crops across diverse conditions remains a challenge [30,48,57]. There is still room for improvement in interpreting images captured by UAVs and robots. To augment the capabilities of AI for weed detection and mapping, continued research and development efforts are imperative.

3.2. Challenges to Using AI in Image Analysis for Weed Mapping

The application of AI in image analysis for weed mapping presents several significant challenges, which can be broadly categorized into data-related and technical issues. These challenges must be addressed to fully realize the potential of AI in precision agriculture.

3.2.1. Data-Related Challenges

The quality and quantity of training data significantly influence model performance [58]. AI models rely heavily on large volumes of high-quality, diverse, and well-annotated datasets for effective training and validation [59,60]. However, creating such datasets is a labor-intensive and costly process that demands domain expertise [58]. The inherent variability in weed and crop characteristics, such as differences in growth stages, shapes, sizes, and colors, adds complexity to the modeling process, making it difficult to develop a universal model [57,61]. Additionally, the close resemblance between certain weeds and crops often leads to misclassification, further complicating the task. Environmental factors, including variations in lighting conditions, shadows, and seasonal changes, also affect image quality and consistency, posing additional hurdles for model training [62]. Distinguishing weeds from soil, debris, and other non-weed elements further increases the complexity of the task. Moreover, AI models often struggle with generalization, performing well on training data but failing to adapt to new or unseen field conditions. Imbalanced datasets, where weeds are underrepresented compared with crops, can lead to biased models that underperform in detecting less-represented classes [63]. Beyond technical issues, concerns related to data privacy and algorithmic bias also hinder the widespread adoption of AI in agriculture [64].

3.2.2. Technical Challenges

Real-time image processing for weed detection and mapping is computationally intensive, often resulting in high latency that impedes timely agricultural interventions. Training ML and DL models requires substantial computational resources, including powerful GPUs, which may not be readily accessible to all stakeholders [58]. Deploying these models on UAVs or robots, which typically have limited processing power and battery life, presents additional challenges [65]. Ensuring seamless integration and operation of ML/DL models with UAVs, robots, and other sensors under diverse field conditions is another significant hurdle [66]. Furthermore, continuous monitoring, maintenance, and updating of models are necessary to adapt to evolving field conditions and emerging weed species, but these tasks are resource-intensive and complex [67]. Finally, it is essential to ensure that the economic and operational benefits of using ML/DL for weed mapping justify the associated costs, particularly for small- and medium-scale farmers [68].

Addressing these challenges requires a multidisciplinary approach that combines advancements in ML/DL algorithms, improvements in data collection and processing techniques, and better integration of technology into farming practices. Additionally, robust economic and educational support for farmers is crucial to facilitate the adoption of AI-driven solutions.

3.3. Interventions to Overcome Challenges in AI-Based Weed Mapping

3.3.1. Data-Related Interventions

To foster wider adoption of AI in weed mapping, it is imperative to address the foundational challenge of data quality and diversity. High-resolution, comprehensive datasets that encompass a wide range of weed species, growth stages, and environmental conditions must be developed to ensure robust model performance. Techniques resilient to variations in lighting, scale, and rotation should be prioritized, alongside the integration of multi- and hyperspectral sensors, which enhance the ability to distinguish between plant species with greater precision [57]. UAVs can play a pivotal role in continuous data collection, significantly reducing manual labor and associated costs while ensuring consistent and scalable data acquisition. To mitigate inconsistencies in input data, automated tools should be employed to correct shadows, lighting variations, and seasonal changes, thereby normalizing datasets for more reliable analysis. Advanced annotation tools can further enhance the precision of labeling weed instances and features in images, ensuring high-quality training data. Engaging farmers and stakeholders in contributing labeled data can enrich dataset diversity, capturing a broader spectrum of real-world conditions.

Addressing data scarcity is another critical step, which can be achieved through the generation of synthetic data or the augmentation of existing datasets using techniques such as rotation, flipping, and lighting adjustments [69]. Additionally, re-weighting techniques should be applied during model training to balance underrepresented classes, thereby improving the fairness and accuracy of AI models [70]. Standardized protocols for image capture under varying conditions, such as fixed times of day or controlled lighting, can further enhance dataset consistency. By investigating the impact of environmental factors, such as lighting conditions and crop density, researchers can develop more robust and adaptable AI solutions for weed management. Advanced methods like GANs can be leveraged to simulate underrepresented weed species and growth stages, addressing gaps in real-world data. Finally, ethical considerations must be prioritized by enforcing data ownership policies, anonymizing sensitive information, and continuously monitoring models to prevent algorithmic bias. This ensures fair treatment of all data types and promotes trust in AI-driven weed mapping solutions.

3.3.2. Technical-Related Interventions

Deploying on-field edge computing can enable faster data processing and real-time decision-making, while a hybrid computational model can balance efficiency by offloading resource-intensive tasks to the cloud and handling time-sensitive operations locally. Combining multiple AI models, such as through ensemble learning techniques, can enhance prediction accuracy by addressing challenges like crop–weed similarity and environmental variability. Additionally, developing lightweight ML and DL algorithms tailored for low-resource environments ensures accessibility for farmers in resource-constrained settings. Establishing regional AI hubs equipped with high-performance computing infrastructure can provide farmers with centralized access to advanced computational resources, data processing, and model training. Investment in high-performance hardware, such as GPUs and TPUs, alongside scalable cloud platforms, will facilitate efficient large-scale model training and support the development of more sophisticated AI systems.

Equipping UAVs and agricultural robots with energy-efficient AI chips can enable onboard data processing, reducing reliance on external servers and enhancing operational autonomy. Optimizing data transfer protocols between UAVs/robots and remote servers will ensure seamless offloading of computationally heavy tasks, while developing AI systems compatible with existing farm hardware minimizes the need for costly upgrades. Standardizing AI algorithms, sensor technologies, UAVs, and robotics will promote interoperability and scalability across agricultural systems, reducing barriers to adoption. Implementing systems for automatic model updates and continuous monitoring for model drift will ensure AI systems remain accurate and reliable as field conditions evolve. Finally, developing affordable, tailored AI solutions for small- and medium-sized farms will ensure equitable access to advanced weed mapping technologies, fostering wider adoption and supporting sustainable agricultural practices.

Table 1. Machine learning models and their performance accuracy in crop/weed discrimination in different crops.

Crops	Models and Performance Accuracy (%)		Reference
Barley	CNN	99	[71]
	CA	100	[72]
Blueberry	Color co-occurrence matrices	94	[73]
Broad bean	ANN	100	[74]
Cabbage	Spectral angle mapper	100	[75]
	Bayesian	84.3	[76]
Canola	ANN	94	[77]
Carrot	RF	93.8	[47]
Chickpea	GDA	95	[76]
Cotton	CNN	99.3	[30]
	RF	85.8	[78]
	CNN	97	[79]
	LDA	>90	[80]
	RF	94.4	[81]
	DA	100	[82]
Lettuce	SSC	90.3	[83]
Maize	RF	94.5	[84]
		94	[84]
	SVM, LDA	>98.4	[85]
	MOGs	100	[50]
	MOGs, SOM	>96	[51]
	SVM	93	[86]
		97	[87]
		81.6	[88]
	PLSDA	>94.8	[89]
	LDA	94	[90]
	CA	100	[72]
	CNN	95.6	[84]
	SVM	91.5	[84]
	CNN	98.2	[91]
Pea	ANN	94	[77]
Peanut	CNN	95.6	[92]
Rice	RF, SVM	100	[51]
	DT	98.2	[93]
	SVM	95.3
	NB	93.1
	DA	100	[75]
	CNN	>94	[94]
Rocket salad	DA	100	[95]
Milk thistle	OCSVM	96.1	[52]
	OCSOM	94.7
	Autoencoders	94.3
	OCPCA	90
Soybean	CNN	92.9	[84]
		97	[96]
	Haar mother wavelet	100	[97]
	NN	100	[88]
	LDA	>90	[80]
Sugar beet	CNN	98.2	[84]
	RF with Markov random field	95	[53]
	CA	100	[72]
	RF	96	[98]
	SVM, ANN	96.7	[91]
	LDA	97.3	[99]
Sugarcane	Fuzzy real-time classifier	92.9	[100]
	RF	97	[89]
Sunflower	SVM	91.5	[84]
Tomato	Bayesian	95.9	[90]
		95.8	[90]
		92.2	[101]
		95	[101]
	CNN	99.3	[30]
Wheat	ANN	94	[77]
	CNN	98.2	[84]
		95.6	[92]
	CA	100	[72]
	SVM	91	[73]
	ANN	100	[74]
	PLSDA	85	[73]
	GDA	95	[76]

CNN—convolutional neural network; CA—cluster analyses; ANN—artificial neural network; RF—random forest; GDA—general discriminant analysis; LDA—linear discriminant analysis; DA—discriminant analysis; SSC—site-specific classifier; SVM—support vector machine; DT—decision tree; NB—naive Bayes; OCSOM—one-class self-organizing map; OCSVM—one-class support vector machine; NN—neural network; MOGs—mixture of Gaussians; SOM—self-organizing map; PLSDA—partial least square discriminant analysis; OCPCA—one-class principal component analysis.

Table 1. Machine learning models and their performance accuracy in crop/weed discrimination in different crops.

Crops	Models and Performance Accuracy (%)		Reference
Barley	CNN	99	[71]
	CA	100	[72]
Blueberry	Color co-occurrence matrices	94	[73]
Broad bean	ANN	100	[74]
Cabbage	Spectral angle mapper	100	[75]
	Bayesian	84.3	[76]
Canola	ANN	94	[77]
Carrot	RF	93.8	[47]
Chickpea	GDA	95	[76]
Cotton	CNN	99.3	[30]
	RF	85.8	[78]
	CNN	97	[79]
	LDA	>90	[80]
	RF	94.4	[81]
	DA	100	[82]
Lettuce	SSC	90.3	[83]
Maize	RF	94.5	[84]
		94	[84]
	SVM, LDA	>98.4	[85]
	MOGs	100	[50]
	MOGs, SOM	>96	[51]
	SVM	93	[86]
		97	[87]
		81.6	[88]
	PLSDA	>94.8	[89]
	LDA	94	[90]
	CA	100	[72]
	CNN	95.6	[84]
	SVM	91.5	[84]
	CNN	98.2	[91]
Pea	ANN	94	[77]
Peanut	CNN	95.6	[92]
Rice	RF, SVM	100	[51]
	DT	98.2	[93]
	SVM	95.3
	NB	93.1
	DA	100	[75]
	CNN	>94	[94]
Rocket salad	DA	100	[95]
Milk thistle	OCSVM	96.1	[52]
	OCSOM	94.7
	Autoencoders	94.3
	OCPCA	90
Soybean	CNN	92.9	[84]
		97	[96]
	Haar mother wavelet	100	[97]
	NN	100	[88]
	LDA	>90	[80]
Sugar beet	CNN	98.2	[84]
	RF with Markov random field	95	[53]
	CA	100	[72]
	RF	96	[98]
	SVM, ANN	96.7	[91]
	LDA	97.3	[99]
Sugarcane	Fuzzy real-time classifier	92.9	[100]
	RF	97	[89]
Sunflower	SVM	91.5	[84]
Tomato	Bayesian	95.9	[90]
		95.8	[90]
		92.2	[101]
		95	[101]
	CNN	99.3	[30]
Wheat	ANN	94	[77]
	CNN	98.2	[84]
		95.6	[92]
	CA	100	[72]
	SVM	91	[73]
	ANN	100	[74]
	PLSDA	85	[73]
	GDA	95	[76]

CNN—convolutional neural network; CA—cluster analyses; ANN—artificial neural network; RF—random forest; GDA—general discriminant analysis; LDA—linear discriminant analysis; DA—discriminant analysis; SSC—site-specific classifier; SVM—support vector machine; DT—decision tree; NB—naive Bayes; OCSOM—one-class self-organizing map; OCSVM—one-class support vector machine; NN—neural network; MOGs—mixture of Gaussians; SOM—self-organizing map; PLSDA—partial least square discriminant analysis; OCPCA—one-class principal component analysis.

4. Execution Platform in Weed Control

PWC, driven by cutting-edge technologies such as robots and UAVs, is rapidly revolutionizing weed management in agriculture. Two primary approaches are used in PWC to achieve site-specific variable rate applications (VRA). Firstly, the map-based VRA approach utilizes data derived from digital field property maps to administer precise herbicide quantities [102]. This approach is used in UAVs for PWC. Secondly, the sensor-based VRA approach utilizes real-time sensor data to dynamically adjust herbicide dosages, aligning them with the unique requirements of the cultivation environment (Figure 4). This approach is mostly used in robots. The success of precision spray systems hinges on several critical factors, including accurate crop–weed differentiation, the creation of precise weed infestation maps, awareness of the nozzle tip’s position relative to the target weed, precise herbicide placement, and effective control of spray drift [103].

4.1. Weed Control Through Robots

Autonomous robots equipped with various sensors, cameras, and mechanical implements to detect, target, and remove weeds are used in several countries. Robots offer a technologically advanced alternative to traditional manual or chemical-based weed control. The key components of robotic weed control systems are sensors, data analysis, mechanical tools, and autonomous navigation. Sensors and cameras capture images of the crop fields, while advanced algorithms analyze the data in real time to determine weed presence, distribution, size, and growth stages, forming the basis for treatment plans [104]. Mechanical implements such as robotic arms and precision spraying mechanisms target weeds without damaging other crops. Onboard GPS enables robots to maneuver through intricate field terrains, adapt to diverse crop structures, and operate in real time.

By leveraging weeding robots, farmers can reduce their reliance on chemical herbicides and expensive human labor, while also benefiting from round-the-clock operation and increased weed control efficiency [55,105]. Robotic weed control reduces health risks by limiting worker exposure to adverse conditions and hazardous chemicals and reallocates human labor to higher-value tasks [106]. Moreover, robotic weed control systems assess treatment effectiveness by gathering data at regular intervals after the execution of weed control treatment, improving overall efficiency [107]. Robots used for weeding are classified into three categories: spraying robots, laser robots, and mechanical weeding robots (Figure 5).

4.1.1. Spraying Robot

Herbicide spraying robots are equipped with spraying mechanisms, typically nozzles, that accurately dispense herbicides onto weeds while minimizing the application on crop plants (Figure 6). The use of herbicide spraying robots offers several benefits, such as reduced herbicide usage, increased weed control efficiency, minimized drift, and environmental harm. Spot spraying systems demonstrated significant potential for reducing pesticide usage, with reductions ranging from 5% to over 90%, depending on weed spatial and temporal distribution [108,109]. Autonomous spraying robots have achieved precise spot spraying, resulting in 96% weed control with just 10% crop damage [110]. Additionally, an autonomous precision spraying robot demonstrated sub-centimeter accuracy, delivering lower doses (2.5 μL) to targets [111]. Lamm et al. [112] demonstrated an 88.8% weed control rate in cotton fields by selectively applying herbicide only to targeted weeds, facilitated by a CV algorithm while moving at a speed of 0.45 m/s. A robotic sprayer utilizes machine vision algorithms, namely the foliage detection algorithm and the grape clusters detection algorithm, to automatically regulate the pesticide spray valve. This innovative technology leads to a 45% reduction in pesticide consumption [113].

A YOLO-based model (YOLOv5n) was used for variable-rate targeted spraying, utilizing DL for tobacco plant recognition. Compared with conventional broadcast spraying, the variable-rate targeted spraying achieved impressive reductions of up to 60% in agrochemical usage [33]. Commercial spraying systems such as See and Spray (Blue River Technology, USA) and H-Sensor (Agricon GmbH, Germany) are AI-powered solutions capable of discerning crop plants from various weeds [34]. Furthermore, a Raspberry Pi-based intelligent robot, incorporating sprayer and Pi Camera components along with software, holds the potential to enhance weed removal accuracy and precision in the field [114]. Weeds are identified from provided images, achieving a high recall of 0.99 and an accuracy of 82.13% through the utilization of semi-supervised learning, reducing the reliance on manually annotated images [115]. Spraying robots equipped with CV for targeted herbicide applications have dramatically reduced herbicide usage by up to 97% [116]. Steward et al. [117] developed an advanced control robot that achieved a 91% accuracy rate in weed targeting, operating at speeds ranging from 0.9 to 4.0 m/s. Midtiby et al. [118] modified a microsprayer with an inkjet printer head, successfully achieving a 94% weed targeting rate.

In carrot farming, a robotic system equipped with a precision-enhancing Drop-on-Demand (DoD) mechanism for glyphosate application proved highly effective in managing various weed species such as Tripleurospermum inodorum, Chenopodium album, Poa annua, and Stellaria media. The system efficiently administered 5.3 μg of glyphosate per plant (equivalent to 191 g/hectare), compared with the 7.56 μg per plant typically required (equivalent to 540 g/ha to 2880 g/ha) in conventional applications. This translated to substantial herbicide savings of 73–95%. Such outcomes underscore the DoD system’s immense potential in optimizing weed control and curbing glyphosate usage, a vital endeavor for preserving both the environment and human well-being [119]. Robots with RGB cameras and laser sensors quickly identify weeds and measure their shape, size, and distance using machine vision. The automated spraying system then applies herbicide precisely based on weed diameter, reducing herbicide use by 45% through targeted application [113].

The solar-powered RIPPA robot uses RGB and NIR cameras to identify weed targets and GPS for locomotion, operating at a speed of 0.4–1.2 m/s. It can spray herbicide on up to 13 targets per second with sub-2 mm precision [120]. Another spraying robot, AVO, utilizes GPS, LiDAR, and visual sensors for locomotion, achieving a coverage rate of 10 hectares per day. Powered by solar energy and rechargeable batteries, it reduces chemical usage by up to 95% [121]. The Asterix robot presents a compelling solution for targeted weed control. Utilizing an Nvidia Jetson TK1 GPU platform and GPS + 4MP camera navigation, it operates at a speed of 0.8 m/s. Its hybrid power system (48V DC, 4-stroke generator) supports herbicide spraying with a DoD system, achieving 100% weed treatment while demonstrating a tenfold reduction in herbicide usage [119]. A variety of spraying robots equipped with sensors and AI available for weed control in different crops is presented in

Table 2. Commercial spraying robot in agriculture.

Algorithms	Sensor	Locomotion	Robot Name	Crop	Reference
Faster Region-CNN	RGB	-	Prototype	Cotton	[122]
Excess Green minus Excess Red	Lidar, spherical camera, hyperspectral and thermal infrared cameras	Four-wheel steering	LadyBird	Lettuce, Cauliflower, and Broccoli	[123]
ML algorithms	RGB, hyperspectral and thermal	Four-wheel steering	RIPPA	Beetroot, onion and other vegetables	[120]
FDA and GDA	RGB, and laser	Four-wheel driving	Robotic sprayer	Grapes	[113]
2D and 3D algorithms	Lidar, radar, and cameras	Four-wheel driving	AgriRobot, SAVSAR	Vineyard	[124]
DL based YOLOv5n	RGB camera		Differential-Drive Mobile Robot	Tobacco	[33]
NA	Laser, RGB, and multispectral	Four-wheel steering	BoniRob	Sugar beet	[125]
Bayesian formulation, unscented Kalman filter	RGB, Visual Odometry	Two-wheel steering	Asterix	Carrot	[119]
AI powered auto navigation system	RGB camera	95% and 10 ha/day	AVO	Rapeseed, beans and cotton	[121]

Abbreviations: Red Green Blue—RGB; Deep Learning—DL; Not Available—NA; RIPPA—Robot for Intelligent Perception and Precision Application; FDA—Foliage Detection Algorithm; GDA—Grape clusters Detection Algorithm.

.

4.1.2. Mechanical Weeding Robot

Mechanical weeding robots are equipped with mechanical implements or tools such as blades, brushes, or rotating wheels that physically remove (uproot or cut) weeds from the soil (Figure 6). Mechanical weeding robots contribute to the preservation of soil health, water quality, and biodiversity by avoiding the use of herbicides. Also, enhance the productivity and sustainability of organic farming practices [36,37]. For example, “Robovator”, a mechanical weeding robot, has been shown to remove 18% to 41% more weeds and reduce weeding times by 20% to 45% compared with manual hand weeding [36]. The Robovator operates at speeds ranging from 1 to 3 mph, depending on plant and weed density, while consuming low power, typically only 5 kW. This innovative technology has been used both in the United Kingdom and the United States since 2015. Mechatronic weeding enhances weeding speed by at least 1.6 times. Notably, intra-row weeding in vegetable crops has shown an impressive 2.57-fold increase in speed [38]. A plethora of robots equipped with hoeing blades for mechanical weed management are presented in

Table 3. Mechanical robots for agricultural weed control: Sensor, accuracy, speed, and suitable crops.

Sensor	Name	Crop	Algorithm	Power Source	Accuracy	Speed	Locomotion	References
RGB	AgBotII	Thistle, Feather top, and Rhodes	LBP and covariance feature	Battery	92.3%	5–10 km/h	Four-wheel steering	[126]
RGB	Dino	Vegetables in row and on beds	-	Battery	90.6%	3 to 4 km/h and 3 to 5 ha/day	Four-wheel steering	[115]
RGB	K-Weedbot	Rice field	Hough transform	Battery and solar	80%	0.2 m/s	Four-wheel steering	[127]
RGB	Weeding robot 1	Broccoli and lettuce	RANSAC	Battery and solar	-	20 m/min	Four-wheel driving	[128]
RGB, LiDAR	Oz	Vegetables, Nurseries, and Horticulture	-	Battery	cm level	1000 m2/h	Four-wheel driving	[125]
RGB	Ted	Grape	-	Solar	-	5 km/h and 5 ha/day	Four-wheel steering
RGB	VITIROVER	Soil grass	-	Primarily solar, but also utilizes batteries	2–3 cm from the trunk	500 m/h	Four-wheel driving	[125]
Laser, RGB, and multispectral	BoniRob	Sugar beet	-	Battery and solar	91%	5 km/h	Four-wheel steering	[125]
Capacitive and azimuth sensors	Tertill	Parks and Gardens	-	Solar	80–95%	0.5–1.5 ac/h	Four-wheel driving	[105]
Capacitive sensors	AIGAMO-ROBOT	Rice	-	Battery and solar	±35 mm or 90–95%	±35 mm	Track	[129]
Capacitive sensors	Weeding robot 2		-	-	71%	500 m/s	Four-wheel driving	[128]
Capacitive sensors	Weeding robot 3		-	-	-	-	Boat

Abbreviations: RGB—Red Green Blue; RANSAC—Random Sample Consensus; LBP—Local Binary Pattern.

.

The BoniRob robot uses both mechanical tools and herbicide spraying to eliminate weeds. Similarly, AgBot II, developed in Australia, employs three different tools, viz. hoe, toothed, and cutting tool, to remove weeds. It relies on CV techniques, such as Local Binary Pattern and Covariance Features, to distinguish weeds from images captured by its RGB camera [126]. Autonomous robots like Oz, Dino, and Ted, developed by Naio Technologies, are widely used in commercial-scale vegetable and wine farms for weeding. These robots are equipped with mechanical tools and powered by lithium batteries, enabling them to operate for up to 8 h [125]. The K-Weedbot, designed for rice fields, uses gears for efficient weed removal and employs advanced image processing techniques, including the Otsu method, for precise navigation and collision avoidance. Its RGB camera and row detection algorithm allow the robot to stay on course with a deviation of less than 1 inch [127].

4.1.3. Laser Weeding Robot

Laser weeding holds significant promise as a more sustainable and efficient approach to weed control compared with traditional methods like herbicides. This method is both selective and environmentally friendly, as it eliminates weeds without causing harm to crops or the environment [130]. In laser weeding, a laser beam is utilized to target and eliminate weeds precisely. Laser beams deliver high-density energy to targeted spots, heating the plant tissue by focusing on the leaves or stems of the weed, ultimately rupturing the cells and leading to the demise of the plant [131]. This beam is generated through a process called stimulated emission of electromagnetic radiation using optical amplification [132]. The process of laser weeding involves several crucial steps. The process starts with a camera capturing the field of view. This visual data is then fed into a detection model that analyzes the image based on texture, color, and shape characteristics. Subsequently, CV technologies are utilized to identify and precisely locate the weeds. Once identified, precise laser targeting is employed to selectively eliminate weeds, ensuring minimal impact on the surrounding crops (Figure 7). Thus, for accurate weed killing, laser weed control systems rely on recognition systems that utilize AI to distinguish between weeds and crops [133,134].

The efficiency of laser weed control depends on factors like wavelength, exposure time, spot size, laser power, and weed species. For example, Mathiassen et al. [135] reported that Stellaria media and Tripleurospermum inodorum are more susceptible than Brassica napus. The energy needed to kill a plant varies depending on the plant’s diameter, growth stage, and weed species. In general, laser energy ranges from 5 W to 90 W are used (Table 4). The use of a lower-energy laser (1 W) proved insufficient to eradicate all weed plants and necessitated excessively long exposure times. While high-energy laser (5 W) exhibited greater efficiency and required less exposure time. However, higher energy laser poses a risk of crop damage if the laser beam splits into two during the weeding process [40]. Laser weeding is most effective at early weed stages. Later stages require more energy and may lead to regrowth. Dicot weeds are ideal targets due to their apical meristem (Table 4). In robotic weed control, seven distinct types of laser beams, such as CO₂, diode, fiber, pulsed, ultraviolet (UV), near-infrared (NIR), and light detection and ranging (LiDAR), are employed for precise and targeted weed removal. The selection of the laser beam type depends on various factors, including the weed species, the desired level of precision, and the type of robotic weed control system being utilized.

CO₂ lasers are widely utilized for weed control because they emit a high-powered infrared laser beam, capable of effectively burning weeds upon contact, owing to their high energy levels [132,136]. However, their relatively larger size necessitates higher voltage and specialized cooling arrangements for outdoor use. CO₂ lasers, typically operating at a wavelength of 10,600 nm, are utilized to burn tissues due to their strong absorption by the water present in biological cells [137,138]. Kaierle et al. [139] examined the effect of laser treatments (dose and wavelength) on Amaranthus retroflexus L. at early growth stages. Results demonstrate that a wavelength of 10,600 nm CO₂ laser with a dose of 125 J energy per weed was required to kill. In comparison, the use of fiber, diode, and solid-state lasers with wavelengths of 1908, 940, and 532 nm, respectively, required a higher dose of 230, 237, and 1400 J energy per weed, respectively, to kill the weed. Larger laser spot diameters required higher energy levels due to less concentrated heat, and a critical temperature level was only achieved in limited areas. Smaller spots needed less energy but required precise targeting. Both larger spots and later weed stages needed higher doses for effective control (Table 4). Marx et al. [35] evaluated a CO₂ laser with a wavelength of 10,600 nm to control Echinochloa crus-galli and found that a laser energy of at least 54 J per plant was required to kill the plant. High weed density can hinder efficient energy coupling during laser weeding. However, areas without crops can be effectively cleared using thermal methods.

Diode lasers emit a continuous wave of laser light and are frequently used for selective weed ablation. They can be tuned to specific wavelengths to target certain types of weeds. Diode lasers offer the advantage of compactness, widespread availability, and affordability, requiring minimal voltage and current [140]. Fiber lasers are often integrated into robotic systems for weed control because of their ability to deliver a focused laser beam with high efficiency and precision. Both CO₂ and fiber lasers pose challenges in terms of portability and cost. Fiber lasers have emerged as the most powerful, albeit expensive option. The 2 μm fiber laser emits energy that is primarily absorbed by the water inside the plant, resulting in heating of a larger area of the plant. A thulium-doped fiber laser operating at a 2 μm wavelength is found to be more effective for weed control as the radiation penetrates through the epidermis, rather than being solely absorbed on the surface of the plant [138].

Pulsed lasers emit short bursts of high-energy laser pulses for precision weed removal by rapidly heating and vaporizing targeted weed tissues. Efficient energy coupling can be achieved by using mid-infrared wavelengths, low power with long irradiation times, pulsed lasers or zigzag patterns, small laser spots compared with target tissue, and significant overlap between target spots and the effective laser area [35]. UV lasers produce short-wavelength light and are effective at breaking down plant cell walls, suitable for weed ablation while posing minimal risk to crop damage [132]. NIR lasers are occasionally employed for weed detection rather than direct weed control. They primarily serve to distinguish between weeds and crops based on the reflectance properties of plant tissues [141]. LiDAR lasers are used for 3D mapping and navigation in robotic systems, aiding in the detection of the position and height of weeds, thereby enabling PWC. A list of robots equipped with different types of lasers, along with their specifications for PWC, is presented in Table 4.

4.1.4. Economics of Weeding Robots

The economic viability of weeding robots in agriculture is a multifaceted issue, heavily influenced by factors such as crop type, farming practices (conventional vs. organic), weeding intensity, robot utilization, labor costs, herbicide expenses, and the initial investment in the robotic system. A key economic driver for adopting weeding robots is their potential to significantly reduce labor costs, which can be substantial in weeding operations, particularly in organic farming. Sørensen et al. [142] study on mechanical weeding robots in organic agriculture reported a potential 85% reduction in labor use for organic sugar beet farming and 60% for organic carrot production, assuming 100% weeding efficiency. This substantial decrease in manual labor translates directly into significant cost savings, especially in regions with high labor costs or limited labor availability. Similarly, Pérez-Ruíz et al. [143] demonstrated that a mechanical weeding co-robot system could replace approximately 60% of hand hoeing labor for intra-row weed control in tomato. Sørensen et al. [142] further explored the Maximum Acquisition Value (MAV), representing the highest price a farmer could profitably pay for a mechanical weeding robot. They found MAV hinges on weeding intensity and annual robot use. At high intensity/utilization (300 operational hours per year), MAV reached EUR 110,000, suggesting that for larger farms with significant weeding demands and the ability to maximize the robot’s operational hours, a substantial investment in weeding robots can be economically justifiable. Conversely, with low weeding intensity and low utilization (180 operational hours per year), MAV dropped below EUR 40,000, underscoring that economic viability depends critically on the robot’s deployment frequency and efficiency.

Beyond labor savings, spot spraying robots offer the potential for significant reductions in herbicide usage, leading to substantial cost savings and environmental benefits. Pedersen et al. [144] present a comparison of a microspraying robotic system with conventional spraying in Danish sugar beet farming and estimated a 90% reduction in herbicide use. This dramatic decrease in input costs significantly impacted the overall profitability of farming operations. Their economic feasibility assessment concluded that robotic weeding was more profitable than conventional systems, with the potential to reduce operating costs by up to 24%. This finding suggests a strong economic case for adopting spot spraying robots, particularly in systems where herbicide costs are a significant component of operational expenses. While the long-term operational cost savings appear promising, the initial investment cost of weeding robots remains a crucial factor influencing their adoption. Pedersen et al. [144] estimated an initial cost of nearly EUR 65,000 for their microspraying robotic system. Sørensen et al. [142] MAV analysis also implicitly reflects the significant upfront investment required. The high initial capital outlay can be a major barrier, especially for smaller farms or those with limited access to capital. The economic viability, therefore, hinges on the payback period, which is determined by the magnitude of labor and herbicide savings relative to the initial cost and operational expenses (including maintenance, energy, and potential repairs).

Several factors will continue to shape the economics of weeding robots. Maturing technology is expected to lower initial costs and improve efficiency, speed, and accuracy, enhancing economic attractiveness. While current economic studies favor large farms with high utilization, developing affordable and scalable robots suitable for small and medium-sized farms is crucial for wider adoption [145]. Government incentives and subsidies supporting sustainable agriculture, including financial aid for robot purchases, can play a significant role in overcoming the initial cost barrier. Finally, the emergence of robot-as-a-service (RaaS) models has the potential to reduce upfront investment, thereby increasing accessibility for a broader range of farmers.

Table 4. Details of different laser weeding robots.

Weeds	Energy	Laser	Spot Diameter	Target Region and Weed Stage	Efficiency (%)	Reference
Sinapis arvensis, C. album	50 W (2.3 J)	CO2 laser	0.6 mm2	Apical meristem	90	[146]
Tripleurospermum inodorum, S. media, B. napus	5 W, 532 nm and 90 W, 810 nm	Continuous-wave diode laser	0.9 and 1.8 mm and 1.2 and 2.4 mm	Apical meristem at cotyledon stage	5 W, 532 nm laser and 18 mm spot diameter most effective	[135]
Elytrigia repens	5 W 450 nm (35 J)		0.55 mm	35 days old	86	[40]
Lolium rigidum	25 W, 975 nm (76.4 J mm−2)	Fiber-coupled diode laser	5 mm	Three-leaf stage	100	[147]
Amaranthus palmeri	6.1 W 450 nm for 1.5 s (9.25 J)	Diode lasers	0.8 and 2.65 mm	2 weeks after emergence	80	[140]
Amaranthus retroflexus	125 J, 10,600 nm	CO2 laser	3.0, 4.2, and 6.0 mm	3 stages (Cotyledons completely unfolded, two true leaves and >4 true leaves), Meristem	100	[139]
	230 J, 1980 nm	Fiber laser
	237 J, 940 nm	Diode laser
	1400 J, 532 nm	Solid-state laser

Table 4. Details of different laser weeding robots.

Weeds	Energy	Laser	Spot Diameter	Target Region and Weed Stage	Efficiency (%)	Reference
Sinapis arvensis, C. album	50 W (2.3 J)	CO2 laser	0.6 mm2	Apical meristem	90	[146]
Tripleurospermum inodorum, S. media, B. napus	5 W, 532 nm and 90 W, 810 nm	Continuous-wave diode laser	0.9 and 1.8 mm and 1.2 and 2.4 mm	Apical meristem at cotyledon stage	5 W, 532 nm laser and 18 mm spot diameter most effective	[135]
Elytrigia repens	5 W 450 nm (35 J)		0.55 mm	35 days old	86	[40]
Lolium rigidum	25 W, 975 nm (76.4 J mm−2)	Fiber-coupled diode laser	5 mm	Three-leaf stage	100	[147]
Amaranthus palmeri	6.1 W 450 nm for 1.5 s (9.25 J)	Diode lasers	0.8 and 2.65 mm	2 weeks after emergence	80	[140]
Amaranthus retroflexus	125 J, 10,600 nm	CO2 laser	3.0, 4.2, and 6.0 mm	3 stages (Cotyledons completely unfolded, two true leaves and >4 true leaves), Meristem	100	[139]
	230 J, 1980 nm	Fiber laser
	237 J, 940 nm	Diode laser
	1400 J, 532 nm	Solid-state laser

4.1.5. Challenges to the Adoption of Robotic Weed Control

Robotic weed control faces numerous challenges, including weed recognition [134], navigational complexity [148,149], power supply, scalability, data management, high cost, and maintenance [67,150]. Weeding robots encounter challenges in the accurate identification of weeds among crops due to variations in weed species, growth stages, environmental conditions (e.g., lighting, soil color), and overlapping vegetation. Current vision systems and algorithms often struggle with accurately distinguishing between crops and weeds, especially at early growth stages, due to the diverse appearance of both crops and weeds, leading to misidentification and potential crop damage [134]. For instance, laser weeding, though precise, risks damaging crops if calibration is inaccurate or there is an overlap between weed and crop areas [40,130]. Furthermore, the mechanical weeding efficacy of robots is affected by soil texture and weed density, requiring adaptable tools and control algorithms. Additionally, managing and analyzing the large amounts of data collected by these robots is a challenge without proper tools. The effectiveness of robotic weed management depends on real-time data, but connectivity issues (poor or no mobile and internet connectivity) in remote areas hinder data access.

Weeding robots must seamlessly integrate with existing farm infrastructure, including irrigation systems, crop layouts, and other IoT devices. The diverse cropping systems and typically small farm sizes due to fragmented landholding prevalent in many developing countries often present significant challenges for the integration of farm infrastructure with weeding robots. As a result, its adoption is very minimal in all developing countries except in protected cultivation. Safety concerns also limit adoption, as lasers pose risks to operators and bystanders due to their potential to cause eye injuries or burns [68]. Furthermore, limited digital literacy in developing countries requires training for farmers to use and trust robotic systems [151,152]. Farmers are hesitant to rely on complex robotic systems, particularly if they lack the technical expertise to operate and maintain them. Mechanical weed control methods, such as cutting or uprooting, are energy-intensive, reducing the robot’s efficiency. Compared with mechanical or chemical methods, laser systems cover less area per unit of time due to their narrow working width (2–3 m) and slow speed (4–6 km/h), making them inefficient for large-scale operations [153]. The large number of weed patches (>100 m²) slows them down further as they must locate and differentiate between weeds and crops in real-time [154]. Furthermore, laser weeders struggle with certain types of weeds, such as those with thick or waxy leaves that resist heat damage [146].

Most weed control robots rely on batteries, limiting their operational time, while the energy demands of laser systems drive up operating costs, particularly in areas with high fuel or electricity prices [40]. Additionally, laser weeding presents a fire risk, similar to flame weeding, as laser beams can ignite dry organic materials like straw, and dry leaves [132,155]. Navigating uneven terrain while avoiding obstacles such as rocks and irrigation equipment remains a technical hurdle, demanding sophisticated path planning and obstacle avoidance capabilities [149]. Economically, the high initial cost of weeding robots is a major barrier to adoption. Compared with traditional methods, the investment required for these advanced machines is substantial, making them inaccessible to many farmers, especially smallholders [68]. Regular maintenance and repairs add to the overall cost of ownership. For example, high initial costs for equipment like lasers and targeting systems pose a significant barrier, especially for small-scale farmers [150]. The return on investment depends heavily on factors such as farm size, labor costs, and the robot’s operational efficiency. The long-term cost-effectiveness of robotic weeding compared with conventional methods needs further investigation and clear demonstration to incentivize adoption. In many regions, there is limited government or institutional support for weed control robots. Addressing these challenges is essential for widespread adoption and effective implementation of robotic weed control in agriculture.

4.1.6. Key Recommendations for Scaling Robotic Weed Control in Agriculture

Scaling robotic weed control in agriculture requires a multifaceted approach that integrates technological innovation, energy efficiency, and precision engineering. A key priority is the development and refinement of advanced weed detection algorithms and machine learning models capable of accurately distinguishing between crops and weeds, even under complex and variable field conditions [156]. High-resolution sensors, coupled with sophisticated image processing techniques and improved calibration technologies, are essential to minimize crop damage and ensure precise weed eradication [157]. Additionally, advancements in laser targeting systems can significantly enhance the precision and speed of weed control, particularly for challenging weed types with thick or waxy leaves, making these systems highly suitable for large-scale agricultural operations. Energy efficiency is another critical factor in scaling robotic weed control. Innovations in power management systems, alongside the exploration of alternative energy sources such as solar power or advanced battery technologies, can address the limitations posed by short battery life and high operational costs [158,159]. Optimizing the energy consumption of mechanical weeding robots will extend their operational time and coverage, making them more viable for widespread adoption.

Scaling robotic weed control in agriculture, particularly for small to medium-sized farms, demands a holistic strategy that addresses affordability, accessibility, education, and infrastructure. Developing cost-effective robotic systems tailored to the needs of smaller farms is a critical step toward broader adoption [67,160]. Modular robot designs, which allow farmers to incrementally add functionalities as needed, can reduce upfront costs and provide flexibility. A survey of 174 farmers in Bavaria, Germany, conducted by Spykman [161], highlighted a preference for small robots over autonomous tractors, underscoring the importance of scalable and adaptable solutions. To further alleviate financial barriers, subsidies and tax exemptions, particularly for organic and small-scale farmers, can offset initial investment costs and incentivize adoption. Additionally, shared service models, where multiple farms collectively share the cost and use of robotic systems, can make advanced technology more accessible. Agricultural cooperatives or service providers can manage and maintain these shared resources, ensuring efficient utilization. Spykman et al. [161] found that non-purchase options, such as contractor services and machinery sharing, are preferred by farmers, indicating the potential of collaborative approaches to drive adoption.

Equally important is enhancing digital literacy, especially in developing regions, to ensure the successful integration of robotic systems into farming practices. Improved digital skills not only empower farmers to operate these technologies effectively but also contribute to rural income growth [151,152]. Comprehensive training programs, offered through agricultural extension services, universities, and private organizations, can build farmer confidence by addressing operational and maintenance concerns while emphasizing safety protocols and the benefits of robotic systems. Demonstration projects and pilot farms can serve as practical showcases, illustrating the real-world advantages of robotic weed control and stimulating interest among farmers. Infrastructure development is another critical enabler for scaling robotic weed control. Investing in rural broadband and mobile network connectivity is essential to support the deployment of robotic technologies in remote areas. Public–private partnerships can play a pivotal role in building this infrastructure, ensuring that even farms in underserved regions can access and benefit from advanced agricultural technologies.

4.2. Unmanned Aerial Vehicles

Equipped with different types of sensors/cameras, UAVs capture high-resolution aerial imagery. This UAV imagery is subsequently processed using ML algorithms to produce precise weed infestation maps. By uploading the weed infestation map to a UAV, it can fly autonomously and apply herbicide only in weed-infested areas, reducing herbicide use, labor, and environmental impact, and significantly enhancing the efficiency and effectiveness of weed management in agriculture.

4.2.1. UAVs for Precision Weed Control

Among the remote sensing platforms, UAVs stand out due to their ability to capture images with millimeter-level spatial resolution by flying at low altitudes, providing real-time data, a capability unmatched by satellite platforms [162]. Equipped with advanced cameras and sensors, UAVs quickly capture high-resolution field images, which are processed using ML algorithms to generate detailed weed infestation maps [52,78,104,163]. A resolution of less than 0.5 cm/pixel is required for precise weed mapping [164]. Accurate mapping of individual weed species in mixed stands of crops and weeds is challenging with an image resolution of 2–8 cm/pixel [58]. The identification of weed hotspots allows for precise, site-specific herbicide applications, reducing the quantity of herbicide needed and minimizing environmental impact (Figure 8). The capabilities of UAVs are further enhanced by the integration of Global Navigation Satellite System (GNSS) or Global Positioning System (GPS) technologies.

Studies have shown successful mapping of weeds in row crops such as maize and sunflowers using UAV imagery [165,166]. The UAV-based site-specific weed management process begins by capturing imagery with sensors chosen based on the weed species, field conditions, and management objectives. The captured images are then processed, including orthorectification and stitching, to ensure high data quality for decision-making. Various algorithms are then employed to classify crops and weeds in the collected imagery (Table 1). UAVs are mostly mounted with three types of cameras for weed mapping, viz., RGB (red, green, and blue), multispectral, and hyperspectral cameras. The RGB cameras are the most common due to their low cost and lightweight design. UAVs commonly use cameras with 3–12-megapixel resolution, due to their affordability and easy availability [58]. Multispectral sensors, which capture more spectral bands than RGB cameras, enable the estimation of a larger number of vegetation indices and accurate weed mapping, especially at later growth stages [167,168].

Hyperspectral sensors capture hundreds to thousands of narrow radiometric bands, primarily in the visible and infrared spectra. Selecting the appropriate band and radiometric range is key for hyperspectral weed mapping. Each band, or a set of bands, can identify specific features within a field due to their narrow spectral width, holding the potential to yield exceptionally precise weed maps during both early and late phenological stages, despite high operational costs [51,85,167]. The data in

Table 5. Leveraging UAVs, spectral sensors, and AI for weed detection and mapping.

Crop	Weeds Identified	Algorithm	Accuracy (%)	References
RGB
Barley, Sugar beet	Echinochloa crus-galli, Avena fatua, Alopecurus myosuroides, Chenopodium album	CA	100	[72]
Chickpea	Broadleaf weeds, grass weeds	GDA	95	[76]
Cotton	Echinochloa crus-galli	DA	100	[82]
Maize	Ambrosia artemisiifolia, Amaranthus retroflexus, Echinochloa crus-galli, Setaria glauca, Capsella bursa-pastoris, Chenopodium album, Panicum capillare, Digitaria sanguinalis, Avena fatua, Alopecurus myosuroides, Beta vulgaris	CA	100	[72]
Rice	Echinochloa crus-galli	SVM	97	[87]
Soybean	Eleusine indica, Alternanthera philoxeroides, Amaranthus palmeri	NN	100	[88]
Sugarcane	Commelina benghalensis, Brachiaria brizantha, Brachiaria decumbens, Megathyrsus maximus, Convolvulus arvensis, Ageratum conyzoides, Crotalaria juncea	RF	97	[89]
Soybean	Amaranthus palmeri, Echinochloa crus-galli, Digitaria sanguinalis,	LDA	>90	[94]
Multispectral
Barley	Brassica napus, Raphanus sativus	CNN	99	[71]
Cotton	Amaranthus palmeri, Leptochloa chinensis	RF	85.8	[78]
Maize, Sugar beet	Cirsium arvense, Xanthium strumarium, Geranium sylvaticum, Amaranthus graecizans, Sorghum halepense, Humulus japonicus, Xanthium sibiricum, Amaranthus palmeri, Convolvulus arvensis, Chenopodium album, Digitaria sanguinalis, Brassica nigra, Batis maritima, Alopecurus myosuroides, Sinapis arvensis, Galium aparine, Stellaria media, Apera spica-venti, Tripleurospermum inodorum, Capsella bursa-pastoris	SVM, ANN	96.7	[91]
		CNN	98.2
Peanut, Wheat	Chenopodium album, Humulus scandens, Xanthium strumarium, Xanthium sibiricum	CNN	95.6	[92]
Cotton, Tomato	Cyperus iria, Leptochloa chinensis	CNN	99.3	[30]
Soybean	Digitaria sanguinalis, Convolvulus arvensis, Echinochloa crus-galli, Amaranthus palmeri	CNN	97	[96]
Sugarcane	Ipomoea alba, Convolvulus arvinse, Coccinia grandis, Trianthema portulacastrum, Amaranthus viridis, Cyanotis axillaris, Physalis minima, Commelina bengalensis, Cyperus rotundus	Fuzzy real-time classifier	92.9	[100]
Sunflower	Tripleurospermum inodorum, Capsella bursa-pastoris, Apera spica-venti, Geranium sylvaticum, Alopecurus myosuroides, Sinapis arvensis, Galium aparine, Stellaria media, Chenopodium album, Amaranthus palmeri, Brassica nigra, Convolvulus arvensis, Batis maritima	CNN	92.9	[84]
Hyperspectral
Broad bean	Cruciferous weeds	ANN	100	[74]
Canola, Pea, wheat	Amaranthus retroflexus, Avena fatua	ANN	94	[77]
Cotton	Pennisetum villosum, Sonchus oleraceus, Avena fatua	CNN	97	[79]
	Lolium perenne	LDA	>90	[80]
	Amaranthus palmeri	RF	94.4	[81]
Maize	Convolvulus arvinse, Rumex crispus, Echinochloa crus-galli, Digitaria sanguinalis, Cirsium arvense, Trapa natans, Ipomoea spp., Polymeria spp., Ranunculus repens, Amaranthus retroflexus, Chenopodium album, Sinapis arvensis, Stellaria media, Taraxacum spp., Poa annua, Medicago lupulina, Ranunculus repens, Cyperus esculentus, Sonchus oleraceus, Polygonum persicaria, Urtica dioica, Oxalis europaea	SVM, LDA	>98.4	[85]
		MoG	100	[50]
Rice	Echinochloa crus-galli, Oryza sativa f. spontanea	RF, SVM	100	[51]
Soybean	Ipomoea purpurea, Lolium perenne	Haar mother wavelet	100	[97]
Sugar beet	Polygonum convolvulus, Equisetum arvense, Setaria viridis, Stellaria media	LDA	97.3	[99]

highlight the promising potential of UAVs equipped with various spectral sensors (RGB, multispectral, and hyperspectral) for precise weed mapping. The effectiveness of different spectral sensors varies depending on weed species and crop type, indicating that a tailored approach is often required for optimal weed mapping.

Factors such as flight altitude, speed, camera resolution, and the ML algorithms used greatly influence the accuracy and resolution of the resulting weed maps [92]. Flights conducted at wind speeds < 5 km/h greatly improve weed detection accuracy [58]. Similarly, flying at lower altitudes (under 50 m), which corresponds to a 15 mm per pixel resolution, ensures precise weed mapping [163]. At higher altitudes, soil pixel signatures tend to dominate, making it difficult to distinguish vegetation edges from the soil background [58]. Another key factor in weed mapping is image overlap, with optimal settings being a frontal overlap of 70% and a side overlap of 60–70% [169,170]. The currently available UAV spray units on the market lack GPS-based, site-specific application capabilities, relying instead on manual triggering by the pilot. Although this method works well for small areas, it is not feasible for large farms due to its dependence on the pilot’s skill. A UAV spray unit that can automatically trigger spray applications at predetermined locations needs to be developed.

The analysis of studies employing RGB, multispectral, and hyperspectral imagery for weed classification reveals strong potential for accurate weed classification across all sensor types. RGB imagery, utilized across a broad range of crops, consistently achieved accuracies between 95% and 100%, with CNNs being a frequently employed and effective algorithm. Multispectral imagery also demonstrated strong performance (85.8–99%), with CNNs again showing promising results. Hyperspectral imagery exhibited high accuracy (92.9 to >98.4%) and saw the application of a more diverse set of algorithms. While direct crop-by-crop comparisons across all three sensor types are limited in this dataset, the evidence suggests that all three sensors hold significant potential for accurate weed detection and classification in agriculture.

Currently, UAV image preprocessing and information extraction for PWC are typically performed externally using powerful GPUs and advanced equipment operated by skilled personnel [58]. This process involves transferring the captured images to external processing units for analysis to detect and map weed infestations [43]. Alternatively, images can be uploaded to the cloud for processing using proprietary software available on a subscription or payment basis, with costs determined by the volume of images processed [171]. However, relying on external processing introduces significant delays in implementing timely weed control measures, compromising the effectiveness of weed management strategies. To overcome these delays, real-time, onboard image processing within the UAV is essential. This requires integrating advanced computing hardware capable of handling real-time analysis, similar to what has been achieved with ground-based weed sensor sprayers [66]. Moreover, reliable network or Wi-Fi connectivity in the field is critical to ensure smooth data transfer and communication between UAVs and control systems. Since many agricultural areas lack sufficient connectivity, establishing a strong network infrastructure is crucial for implementing real-time, onboard image processing in PWC systems [58].

4.2.2. Advantages of Using UAVs in Weed Management

UAVs are often the preferred choice for rapid and precise remote sensing due to their accessibility, versatility, and ease of use. Equipped with various sensors, they perform a wide range of tasks, with camera systems offering spatial resolutions of less than one centimeter per pixel, ideal for weed mapping [172]. In crops like wheat, sugar beet, and barley, site-specific herbicide applications using UAVs have achieved over 90% weed control efficiency across several countries [41,172,173]. This approach has also resulted in 26% to 50% herbicide savings compared with blanket applications (Table 6). The savings are due to UAV’s precision spraying, minimizing overlap, and ensuring site-specific herbicide use [66,174]. Operating at 3 to 5 m/s speeds, UAVs cover large areas quickly without disrupting the field’s topography or pressing sown seeds into the soil, a common issue with manual or boom spraying [175]. Additionally, UAVs reduce workers’ exposure to toxic chemicals, improving the health and safety of agricultural laborers [106,176].

The use of UAVs for herbicide spraying has substantially reduced overall costs, cutting labor expenses by 50% and lowering the risk of musculoskeletal injuries compared with boom and power sprayers [32,93]. In maize cultivation, UAVs offer cost savings ranging from 16 to 45 EUR/ha, reduced herbicide usage by 14–39.2%, and increased weed control efficiency [177]. Nikolić et al. [178] reported similar benefits, with savings of 15.91 EUR/ha to 32.93 EUR/ha, a 42.5% reduction in herbicide usage, and 98.1% weed control efficiency. In rice, UAV spraying reduced costs by 1050 INR/ha (12 USD/ha), achieving a weed control efficiency of 78.8% [179]. UAV use in wet direct-seeded rice also increased net returns, benefit-to-cost ratio, energy-use efficiency, and energy productivity over traditional knapsack sprayers [179]. Quan et al. [180] survey of 2000 farmers across 11 Chinese provinces indicated that using UAVs led to a per-hectare revenue increase of roughly USD 434–USD 488 and a reduction in pesticide application time by 14.4–15.8 h per hectare. Ruzlan et al. [181] found that UAV spraying is economically advantageous for areas over 3000 hectares, with potential savings of 4% to 28%. It also significantly improved operational efficiency, reducing working hours by 37%, water consumption by 91%, and human labor costs by 81% compared with traditional methods like mist blowers and knapsack sprayers.

Cavalaris et al. [182] found UAV spraying costs 1.45–2 times more than traditional methods due to high initial investment and short economic life. They suggest smaller UAVs with longer battery life and expanded annual use could make UAV spraying economically competitive. On the contrary, Umeda et al. [32] found that using UAVs halved labor costs compared with conventional pest control. Furthermore, due to the lower physical workload, 63.86% of UAV-related tasks were identified as having a low risk of musculoskeletal injury. They suggest smaller UAVs with longer battery life and expanded annual use could make UAV spraying economically competitive. Herbicide spraying through UAV is a scalable technology that benefits both small- and large-scale farmers. Most importantly, PWC through UAV reduces the risk of herbicide-resistant weed development, limits herbicide spread, conserves biodiversity, and minimizes environmental harm [66,162]. Most of the studies reported that UAV application led to economic benefits through direct cost savings and indirectly through reduced input costs, particularly herbicide usage. The high weed control efficiency reported across various crops and countries indicates the potential for improved yields and reduced losses due to weed competition, further contributing to economic advantages.

Table 6. Weed control efficiency and herbicide saving in various crops using UAVs.

Country	Crop	Cost Saving	Percent Saving of Herbicide	Percent Weed Control Efficiency	Reference
Italy	Maize	16–45 EUR/ha	14–39.2	No difference	[177]
Italy	Maize	-	37	-	[174]
China	Wheat	-	50	98.45	[173]
USA	Surrogate weed	-	-	94	[66]
USA	Maize	-	26	78.4	[78]
China	Wheat	-	-	98	[41]
Germany	Maize Sugar beet	-	-	96 80	[172]
Iran	Volunteer Barley	-	-	94.3	[42]
India	Wheat	-	-	44.8	[183]
India	Rice	1050 INR/ha	-	78.8	[179]
Italy	Maize	15.91 EUR/ha to 32.93 EUR/ha	42.5	98.1	[178]

Table 6. Weed control efficiency and herbicide saving in various crops using UAVs.

Country	Crop	Cost Saving	Percent Saving of Herbicide	Percent Weed Control Efficiency	Reference
Italy	Maize	16–45 EUR/ha	14–39.2	No difference	[177]
Italy	Maize	-	37	-	[174]
China	Wheat	-	50	98.45	[173]
USA	Surrogate weed	-	-	94	[66]
USA	Maize	-	26	78.4	[78]
China	Wheat	-	-	98	[41]
Germany	Maize Sugar beet	-	-	96 80	[172]
Iran	Volunteer Barley	-	-	94.3	[42]
India	Wheat	-	-	44.8	[183]
India	Rice	1050 INR/ha	-	78.8	[179]
Italy	Maize	15.91 EUR/ha to 32.93 EUR/ha	42.5	98.1	[178]

UAV spraying significantly reduces water usage compared with traditional manual methods. While manual spraying typically requires 200 to 400 L of water per hectare to cover the entire field, UAVs use only 25 to 50 L per hectare, about 10 times less [179]. This is due to the smaller droplet size and uniform spray coverage. Moreover, the spot spraying for escaped weed control would require even lower water volume. UAVs produce finer droplets, enabling them to cover larger areas with the same amount of water. Multipurpose UAVs, capable of spreading seeds, fertilizers, and pesticides, as well as providing surveying and mapping functions, have boosted their popularity [176]. The introduction of fast-charging batteries (0 to 80% in 15 min) has addressed flight duration issues, and modern UAVs now have increased payload capacities (25 to 40 kg), accelerating their adoption in agriculture.

4.2.3. Variation in UAV Regulations Across Countries and Their Impact on Adoption

The pace of UAV adoption is closely tied to regulatory simplicity for registration and licensing, the availability of training infrastructure, and age-related accessibility. Countries with well-defined regulatory bodies and clear guidelines have demonstrated higher UAV adoption rates. For example, the United States and China, with their robust regulatory systems, have issued hundreds of thousands of UAV pilot licenses and registered millions of UAVs, reflecting widespread adoption (

Table 7. Country/region-wise number of UAV training and certification agencies.

Country/ Region	Regulatory Body	Approx. Number of UAV Training and Certification Organizations	Number of UAV Pilot Licenses Issued	Guideline for UAV Registration	Guideline for UAV Pilot License	Minimum Age to Become UAV Pilot
India	Directorate General of Civil Aviation (DGCA)	157	12,432 remote pilot certificates (RPC) issued as of Feb 2025.	All drone except nano category Total 29,756 UAVs registered	UAV weighing > 250 g	18
USA	Federal Aviation Administration (FAA)	800	427,335 certificated remote pilots	UAVs weighing > 250 g drone >1 million UAVs registered as of January 2025	Remote pilot certificate to fly a drone for commercial purposes	16
China	Civil Aviation Administration of China (CAAC)	200+	194,000 people hold drone pilot certificates till 2023.	UAVs weighing > 250 g 1.875 million registered UAVs till June 2024	Any drone weighing >7 kg	16
Australia	Civil Aviation Safety Authority (CASA)	Dozens	Remote Pilot License (RePL)	All UAVs used for business regardless of its weight	UAV weighing > 2 kg	16
European Union	European Union Aviation Safety Agency (EASA)	Hundreds (across member states)	Certificate based on risk category	UAVs weighing > 250 g	UAV weighing > 250 g	16
Philippines	Civil Aviation Authority of the Philippines (CAAP)	40	NA	UAVs weighing 7 kg and above	UAV weighing > 7 kg	18
Thailand	Civil Aviation Authority of Thailand (CAAT)	NA	NA	UAVs weighing 2 kg and above 130,000 registered UAV till October 2024	UAV weighing 25 kg and above	20
Japan	Japan Civil Aviation Bureau (JCAB)	NA	NA	UAVs weighing 100 g	UAV weighing > 25 kg	16
South Korea	Ministry of Land, Infrastructure, and Transport (MOLIT)	NA	Approx 31,300 certified UAV pilot in 2021	UAVs weighing > 250 g 138,000 registered UAVs as of May 2023	UAV weighing 12 kg	18
South Africa	South African Civil Aviation Authority (SACAA)	23	83 registered operators and 1818 UAV pilots	Any drone that is used for commercial purposes Approx 50,000 UAVs registered	UAV weighing > 7 kg	18
Nigeria	Nigerian Civil Aviation Authority (NCAA)	NA	NA	UAVs weighing > 250 g	UAV weighing > 25 kg	16
Kenya	Kenya Civil Aviation Authority (KCAA)	12	1200 licensed UAV pilots	All UAVs used for commercial purposes	UAV weighing > 25 kg	18
Brazil	National Civil Aviation Agency (ANAC)	NA	NA	UAVs weighing > 250 g	UAV weighing > 25 kg	18
Argentina	National Civil Aviation Administration (ANAC)	NA	NA	UAVs weighing > 500 g	UAV weighing > 500 g	18

NA—not available.

). In contrast, many African, Asian, and South American countries have shown slower adoption, as evidenced by fewer registered UAVs and issued pilot licenses due to the slowed establishment of UAV regulations. The presence of UAV training and certification organizations plays a pivotal role in driving adoption. Nations like the United States (800 agencies), India (157 agencies), and China boast a higher number of such institutions, which likely contributes to their larger pool of certified pilots and registered UAVs. Conversely, countries with limited training infrastructure, such as the Philippines, Thailand, and Nigeria, exhibit lower adoption rates due to restricted access to necessary skills and certifications.

Another critical factor influencing UAV adoption is the minimum age requirement for UAV pilots. In most developed countries, including the United States, China, Japan, the European Union, and Australia, the minimum age is set at 16 years. In contrast, countries like India, South Korea, South Africa, Kenya, the Philippines, Brazil, and Argentina require pilots to be at least 18 years old. Thailand stands out with a higher minimum age requirement of 20 years, which may hinder adoption compared with countries with a lower age threshold (Table 7). Additionally, countries such as Kenya, Nigeria, Japan, Brazil, and Thailand have streamlined UAV adoption by mandating registration only for UAVs weighing more than 25 kg. This approach reduces regulatory burdens and accelerates the integration of UAV technology compared with countries requiring registration for UAVs as light as 250 g. To address security concerns, however, these nations often require UAV operators to register their flights in a web portal before operation, ensuring accountability without imposing excessive regulatory hurdles.

4.2.4. Real-Time Object Detection

UAVs with onboard sensors and processors enable rapid data processing for reliable real-time object detection. RGB sensors are commonly used in real-time object detection due to their high resolution, light weight, low cost, and ease of use. Similarly, hybrid-wing UAVs are favored for real-time detection, as they combine the benefits of fixed-wing and rotary-wing designs [184]. Three paradigms exist for real-time object detection: embedded systems, cloud computing, and edge computing. Most studies favor edge computing, with fewer exploring embedded and cloud solutions [184]. Cloud computing offers powerful processing and storage for large UAV datasets, accessible to users anytime, anywhere, with high reliability and dynamic resource allocation. However, its limitations for real-time UAV tasks include network bandwidth constraints hindering data upload. Edge computing addresses this by bringing cloud services closer to the UAV, reducing latency, improving network efficiency, and user experience. NVIDIA Jetson Nano, NVIDIA Jetson TX2 (a popular, balanced choice), and NVIDIA Jetson Xavier NX are common GPU platforms for edge computing [184]. Real-time UAV detection algorithms fall into three categories: traditional methods, ML (often SVM), and DL [51,73,84,85,86]. Linear classifiers require few computing resources, which enables real-time object detection on lightweight computing platforms mounted on drones. However, they require manual feature extraction and extensive preprocessing, whereas deep learning methods automatically extract features but are more resource-intensive [184].

DL is key for real-time detection due to its success in image detection, and its integration into edge computing is anticipated to be crucial for real-time detection. Various YOLO versions are frequently employed, alongside other lightweight deep learning architectures like MobileNetV2, U-Net, ResNet18, AlexNet, FCN, and SegNet [33,63,185]. AI-powered UAVs autonomously detect and treat weed pests using a YOLOv10 model (F1 scores: 0.78 for Araujia sericifera, 0.80 for Cortaderia selloana) in organic orange groves. UAVs use GPS/RTK to create KML files, guiding drones for precise organic treatments. Economic analysis revealed a potential savings of EUR 1810–2650 per hectare [185]. Tetila et al. [186] evaluated YOLOv5 models for real-time weed detection in soybean fields using drone imagery and a new dataset of 4129 annotated samples. The best-performing model, YOLOv5s6, achieved high accuracy in classification and detection (MAE: 1.14, RMSE: 1.67, R²: 0.93), demonstrating the system’s potential for targeted herbicide application, promising reduced costs and environmental benefits.

4.2.5. Bottlenecks for UAV Adoption in Agriculture

UAV spraying operations face regulatory hurdles due to the complexity of obtaining permits and the limited availability of authorized drone pilot training institutes [187,188]. Additionally, safety standards are also inadequately followed due to a lack of awareness. Countries with more flexible, clear, and supportive regulations, such as the United States and Australia, tend to see higher adoption rates. In contrast, countries with more restrictive regulations, in India (no permission, no takeoff) and some EU member states (the complexity of obtaining operational authorizations), have slower adoption. In regions with stringent visual line-of-sight regulations, the benefit of extended battery life for covering larger areas is diminished. Harmonizing regulations across countries and addressing common challenges could facilitate the broader adoption of UAVs in agriculture (

Table 8. Bottleneck with potential solutions and the primary responsible actor(s) for implementation.

Bottleneck	Potential Solutions	Responsible Actors
Complex and restrictive UAV regulations	Streamline permitting processes, establish online licensing platforms, harmonize regulations with international standards.	Policy/Government
Limited availability of authorized drone pilot training institutes	Increase funding for and accreditation of training centers, develop online training modules, incentivize private sector involvement in training.	Policy/Government, Training/Education Institutions, Industry Associations
Inadequate adherence to safety standards	Develop and disseminate clear safety guidelines and SOPs, implement stricter enforcement, integrate safety training into pilot programs.	Policy/Government, Industry Associations, Training/Education Institutions
Limited payload capacity	Invest in R&D for lighter and more efficient materials, optimize UAV design, explore multi-UAV systems for larger payloads.	Engineering/Manufacturing, AI Research/Development
Short battery life	Advance battery technology (higher energy density, faster charging), explore alternative power sources (e.g., fuel cells), optimize flight planning and energy management.	Engineering/Manufacturing, AI Research/Development
High initial investment and operational costs	Implement targeted subsidy programs, promote UAV leasing/hiring services, incentivize local manufacturing, reduce maintenance requirements through improved design.	Policy/Government, Engineering/Manufacturing, Financial Institutions, Service Providers
Weather dependency	Develop more robust UAV designs for varied weather, integrate advanced weather forecasting into mission planning, optimize spraying schedules (early morning/late evening).	Engineering/Manufacturing, AI Research/Development
Concerns about chemical drift	Develop advanced spray nozzle designs, optimize flight parameters (lower altitude, controlled speed), utilize drift prediction algorithms, implement buffer zones.	Engineering/Manufacturing, AI Research/Development, Industry Associations, Farmers/End-Users
Lack of specialized skills among farmers	Develop accessible training programs (workshops, online courses, demonstrations), integrate UAV operation into agricultural extension services, target youth.	Training/Education Institutions, Policy/Government, NGOs
Lack of SOPs	Facilitate collaboration among researchers, industry experts, and regulatory bodies to develop crop- and region-specific SOPs, disseminate best practices.	Industry Associations, Policy/Government, Research Institutions

).

Most agricultural UAVs have limited payload capacities, typically ranging from 10 to 20 kg. This restricts the area they can cover in a single flight, increasing the number of flights required to cover per unit area. Increased payload directly reduces battery life and flight duration. Longer flight times, necessary for covering larger fields, are often compromised by the need to carry sufficient treatment solutions or advanced sensor systems. Heavier components like larger spray tanks necessitate more power for lift and maneuvering, leading to shorter operational times and potentially requiring more frequent battery swaps or recharges, impacting efficiency and workflow [176,189]. Similarly, more advanced sensing capabilities come at the cost of reduced flight endurance and area coverage per flight. Farmers find UAVs less cost-effective for large-scale operations due to the need for frequent refills or battery swaps.

Limited battery life inherently restricts the area a drone can effectively cover in a single autonomous flight. Battery-powered UAVs typically operate for less than 20 min, requiring frequent recharging or battery replacement, which further increases costs and operational downtime [176,190]. Extreme temperatures, both high and low, can negatively impact the performance and lifespan of drone batteries. High temperatures can lead to overheating and reduced discharge capacity, shortening flight times. Low temperatures can also reduce battery efficiency and power output, similarly, limiting operational duration. Although quick-charging batteries exist, their high cost remains a barrier. Additionally, most batteries last only 300–400 recharge cycles, necessitating frequent replacements and raising investment costs.

UAV spraying is highly dependent on weather conditions such as wind, rain, and temperature [189]. Adverse weather conditions such as strong winds, rain, and extreme temperatures significantly narrow the operational window for drone flights. While UAVs offer precision, there are concerns about chemical drift beyond target areas, which could negatively impact ecosystems and biodiversity [65]. Factors, such as wind speed, direction, turbulence, spray nozzle design, pressure, and flight height, influence drift hazards [58,175]. Strong winds carry UAV spray droplets away from target areas, causing drift hazards [189]. Sudden wind gusts or turbulence destabilize UAVs, making it difficult to maintain control and precise positioning and affecting the accuracy of pesticide spraying. Drift hazards from UAV spraying can be minimized by using different nozzle designs, increasing droplet size, and flying at lower altitudes (<2.5 m) [106,191]. For example, Lou et al. [192] reported 7.9% drift at 1.5 m spray height compared with 20% at 2 m height. Similarly, avoid spraying in high winds (>6 m/s) to ensure uniform application.

Fighting against strong winds can lead to faster battery depletion as UAVs must expend more energy to counteract wind resistance, limiting operational range and duration. UAVs cannot be used in areas known for strong winds, such as coastal regions, or at times when wind conditions are excessively strong. Spraying during early morning or late evening is ideal due to lower temperatures, which improve plant uptake and reduce evaporation. Morning dew also aids herbicide absorption by maintaining a cooler microclimate in the crop canopy. On the contrary, spraying at midday, especially in tropical regions where the temperatures and evaporation demands are higher, reduces the spraying efficiency and operation efficiency. High temperatures cause batteries to overheat, increasing the risk of failure or even fire. In high temperatures, air density decreases, reducing lift and making it harder for drones to stay airborne, especially at higher altitudes.

UAV effectiveness is dependent on the precision of onboard sensors, GPS accuracy, and the spraying system. Technological limitations, such as sensor malfunction and signal interference, can compromise the treatment quality [190]. UAV operation requires specialized skills, which are limited among farmers in developing countries. Training programs are needed to enhance digital literacy and UAV piloting skills [58,151,152]. The high initial investment, coupled with significant maintenance and operational costs, is a major barrier, particularly for small-scale farmers [58,65,67]. While some countries offer subsidies for UAVs, the number of beneficiaries under these schemes remains lower than expected. Standard operating protocols (SOPs) for UAV spraying are lacking for many crops and regions, hindering broader adoption. Developing these protocols is essential for effective and widespread use. It is crucial to address these challenges to realize the full benefits of UAV spot spraying.

4.2.6. Scaling UAV Technology in Agriculture

Scaling UAV technology in agriculture requires a multifaceted approach that combines financial support, favorable government policies, infrastructure development, technological innovation, and capacity building (

Table 9. Prioritizing Recommendations: A Feasibility–Impact Matrix.

	High Impact	Low Impact
High Feasibility	(Quick Wins) Streamline permitting processes; increase funding for/accreditation of training centers; develop and disseminate clear safety guidelines/SOPs; promote UAV leasing/hiring services; develop accessible farmer training programs; facilitate collaboration for SOP development.	(Low Effort, Low Gain) Increase awareness campaigns (general); explore multi-UAV systems (initial stages without robust coordination); integrate basic weather forecasts into planning.
Low Feasibility	(Strategic Initiatives) Harmonize regulations internationally; invest significantly in fundamental battery technology advancements; implement large-scale subsidy programs; develop fully autonomous, fail-safe BVLOS capabilities; achieve significant breakthroughs in lightweight, high-capacity payloads; develop sophisticated drift prediction and mitigation algorithms.	(Re-evaluate) Relax citizenship requirements for training (potential political/security considerations); enforce existing safety standards more strictly (resource intensive without addressing root causes); promote specific nozzle designs without broader context.

). Financial measures, such as subsidies and incentives, are critical to making UAVs more accessible, particularly for small-scale farmers who may face significant upfront costs [65]. For example, the government of the Philippines allocated 300 million pesos under the Drones4Rice project, offering 2000 pesos per hectare for drone-assisted farming services, including crop establishment, nutrient management, and pest control. Such initiatives demonstrate how targeted financial support can lower barriers to adoption and encourage the widespread use of UAVs in agriculture.

Supporting local service providers, custom hiring centers, and farmer-producer organizations can further reduce costs and enhance accessibility. In India, the Namo Drone Didi scheme, with a budget of INR 261 crores, aims to equip 15,000 women’s self-help groups with drones for rental services between 2023 and 2026, while also providing training in UAV operation [193]. Similarly, the Union Kisan Pushpak scheme offers loans covering up to 75% of the drone cost (up to INR 12 lakhs for two drones) with a three-year repayment period. Additionally, the Production-Linked Incentive (PLI) Scheme incentivizes local drone and component manufacturing, capping individual benefits at 25% of the total annual outlay to maximize participation [194]. Infrastructure development, including the creation of local service providers and custom hiring centers, can further reduce costs and improve accessibility. By enabling farmers to lease or hire UAVs, these models lower the financial burden and make advanced technology more attainable. These initiatives highlight the importance of government-backed programs in fostering a supportive ecosystem for UAV adoption.

Clear and supportive regulatory frameworks are equally essential to ensure safety, privacy, and efficient airspace management while encouraging local UAV manufacturing and startups. A comparative analysis of drone regulations in India, China, the USA, Australia, and Singapore by Singh et al. [187] revealed that India’s requirement of citizenship for drone ownership and pilot licensing may hinder foreign experts from contributing to training and services during the country’s UAV expansion. Relaxing such restrictions could facilitate technical expertise and accelerate adoption. Furthermore, while the USA offers an online system for drone training and licensing, India currently lacks a similar platform. Establishing an online training and certification organization in India would streamline the licensing process, making it more efficient and accessible.

To maximize the potential of UAVs in agriculture, it is essential to equip them with interchangeable sensors capable of performing diverse functions such as crop monitoring, pest detection, and irrigation management. This modular approach enhances scalability and adaptability, allowing farmers to tailor UAVs to their specific needs [177]. Furthermore, developing multipurpose UAVs that integrate functionalities like seed spreading, fertilizer application, pesticide spraying, and crop health monitoring can significantly increase their utility and cost-effectiveness. Innovations such as line seeding mechanisms for UAVs also hold promise for improving precision and efficiency in planting operations. However, advancing these technologies requires substantial investment in research and development to enhance sensor capabilities, automation, and overall system performance.

Capacity building is another critical component, as enhancing farmers’ skills and knowledge in UAV operation and maintenance is essential for successful adoption. Training programs, demonstrations, and pilot projects can build confidence and showcase the practical benefits of UAV technology in agriculture. Collaboration among government organizations, private institutions, and NGOs is critical to the successful development, testing, and refinement of UAVs tailored to local agricultural conditions. Such partnerships foster knowledge sharing, resource pooling, and collective problem-solving, accelerating technological advancements and adoption. For instance, joint initiatives can address region-specific challenges, such as varying crop types, climatic conditions, and farming practices, ensuring that UAV solutions are both practical and effective.

Equally important is the need to strengthen farmers’ skills and knowledge through targeted training programs and agricultural extension services. As UAV technology is relatively new, farmers require specialized training to gain proficiency in operating and maintaining these systems. Government organizations, NGOs, research institutions, and technology developers should collaborate to offer comprehensive training programs, including workshops, online courses, and hands-on demonstrations. These initiatives should particularly target youths, who are often more receptive to adopting new technologies. Additionally, awareness programs, farmer field schools, and on-farm training sessions can play a pivotal role in building confidence and reducing concerns related to UAV operations, such as drift hazards during pesticide spraying.

5. Conclusions

Modern agriculture faces escalating challenges from herbicide resistance, weed flora shifts, herbicide residues, environmental pollution, rising labor costs, and labor shortages, necessitating a shift towards sustainable weed control. Promising technologies like robotics, lasers, and UAVs have emerged from research as potential solutions to reduce reliance on chemical herbicides and manual labor. Weeding robots have demonstrated the capacity to significantly reduce herbicide use and labor demands, thereby minimizing residues and environmental impact. Similarly, UAVs have shown potential for site-specific herbicide application in numerous studies, although commercially available options remain limited. However, it is crucial to differentiate between these research advancements and their practical application, particularly for small and marginal farmers. While these technologies hold great promise, their feasibility is currently constrained by the complexity of diverse weed populations in real-world field conditions. Furthermore, the high initial investment associated with robotics and advanced UAV systems presents a significant barrier to adoption for smaller agricultural operations. These cutting-edge technologies are still in their infancy, and the field is rapidly evolving.

To harness the benefits and facilitate wider adoption, especially among small-scale farmers, several key steps are necessary. Firstly, focused research and development should prioritize creating more cost-effective and versatile robotic and UAV solutions tailored to the scale and needs of smaller farms. This includes exploring modular designs, open-source platforms, and collaborative ownership models. Secondly, investments in education and training programs are essential to enhance digital literacy and technical skills among farmers, enabling them to effectively operate and maintain these advanced technologies. Thirdly, the establishment of local service providers and accessible repair networks is crucial for ensuring the long-term viability and support of these systems in rural agricultural communities. Finally, government policies and subsidies should be strategically implemented to offset the initial costs and incentivize the adoption of sustainable, technology-driven weed management practices, particularly for vulnerable smallholder farmers. Addressing these adoption hurdles will be critical to translating the promising research in agricultural robotics and UAVs into tangible benefits for all scales of farming and fostering a more sustainable agricultural future.