Weeding is the process of removing unnecessary plants from farming and agricultural lands to prevent them from competing with desired ones over natural resources. These plants are generally recognized by their unpredictable growing locations along with their ability to reproductively proliferate without human intervention (Gao and Su, 2024). The existence of weeds adversely affects crop yield and leads to irregular maturation of crops, thereby complicating the harvesting process due to their disruption of consistent crop growth (Tshewang et al., 2016; Yu et al., 2019). For instance, weeds result in considerable yield losses in wheat, with winter wheat losses in the United States averaging 25.6%. This, combined with a 23.4% loss in Canada, culminates in a potential annual economic deficit of approximately $2.19 billion. Furthermore, for spring wheat, the losses average 33.2% in the United States and 19.5% when combined with Canada, incurring costs up to $1.39 billion. This situation emphasizes the escalating threat posed by weeds (Flessner et al., 2021). Effective weed management is imperative to alleviate the challenges presented by these plants and to enhance crop productivity while sustaining agricultural profitability. Although these practices are labor-intensive and costly, often necessitating substantial manpower and resources, there are instances where the expenses related to weed management may exceed the economic benefits derived from crop production. This highlights the necessity to assess traditional weed management practices and their inherent limitations (Dhakal et al., 2024).

The selection of weeding practices and implements is influenced by factors such as crop type, soil characteristics, and field conditions. Hand-weeding is commonly adopted for smaller landholdings, provided there is sufficient labor availability (Abebe, 2024). Broadly, traditional weeding methods can be categorized into three distinct approaches: physical, chemical, and biological. Each approach offers specific limitations, based on their operational efficiency, environmental impact, and cost-effectiveness (Bond et al., 2003; Gao and Su, 2024). Physical weeding techniques involve thermal technologies, such as laser and flame weeding, effectively controlling weeds but also come up with several challenges. The high temperatures generated in these processes can ignite dry materials in the field, creating fire hazards. Additionally, these methods can pose risks to humans and animals nearby. Flame weeding, in particular, significantly contributes to greenhouse gas emissions, raising environmental concerns due to its detrimental ecological impact (Sivesind et al., 2009). Mechanical weeding is another physical weeding technique and is a widely used traditional practice, with reported average weed removal efficiencies of approximately 80% based on reductions in both weed density and biomass (Jabran and Chauhan, 2015; Liu et al., 2023). However, it also has certain drawbacks. Tools like chain harrows, used for inter-row weeding, can physically damage crops, leading to bruising and stem breakage, which may inhibit plant growth and increase vulnerability to pathogen infestations. Moreover, the heavy design of mechanical weeding equipment can cause soil compaction, adversely affecting soil aeration and root development. Biological weed control employs living organisms, such as insects, fungi, or bacteria, or their byproducts, to suppress weed populations and lessen their impact on crops. Although this method aligns well with environmental principles, its practical application is limited by slow activation and response times, alongside the short half-life of these biological agents. Additionally, the use of bioherbicides derived from these organisms poses potential risks to human and animal health and may have unintended environmental repercussions, thus limiting their widespread adoption in sustainable agricultural systems. Chemical weed control remains the most effective and widely used method for managing weeds, primarily due to the effectiveness of herbicides against various weed populations. However, its extensive application brings significant drawbacks. Chemical herbicides can be quite expensive, increasing the financial burden on agricultural production. Prolonged exposure to these chemicals presents serious health risks to both humans and animals. For example, glyphosate, a widely used herbicide, was classified as “probably carcinogenic to humans” by the World Health Organization in 2015 (Van Bruggen et al., 2018). Furthermore, the significant use of chemical herbicides leads to environmental pollution, including soil contamination, water runoff, and harm to non-target organisms, raising concerns about their sustainability in agricultural practices.

To overcome the limitations of conventional weeding methods, intelligent weeding technologies have emerged as a key component of modern precision agriculture. Unlike traditional Integrated Weed Management (IWM), which combines general mechanical and chemical strategies, intelligent weeding integrates automation, sensing and artificial intelligence to achieve targeted, data-driven weed control (Riemens et al., 2022). Depending on the mode of actuation, intelligent weeding can be categorized into intelligent chemical, intelligent physical and intelligent mechanical approaches (Gerhards et al., 2022; Jiao et al., 2024). Intelligent chemical weeding employs site-specific or variable rate herbicide applications guided by machine vision, thereby reducing chemical usage (Upadhyay et al., 2024a). Intelligent physical weeding utilizes non-chemical energy sources – such as lasers or thermal radiation – to destroy weeds relying on imaging technologies (Bajwa et al., 2015; T. Jin and Han, 2024; Li et al., 2022). Meanwhile intelligent mechanical weeding physically removes or disrupts weeds using robotic end effectors and machine vision, offering an eco-friendly and residue free alternative (Jiang et al., 2023b; Visentin et al., 2023). Collectively, these intelligent systems align with sustainability and resource optimization objectives by minimizing chemical inputs, improving accuracy, and reducing labor requirements.

Despite the benefits of intelligent weed management, the escalating occurrence of herbicide resistant weeds has intensified the need for sustainable, non-chemical control strategies (Délye et al., 2013). Although technologies such as variable-rate spraying, site specific delivery and see and spray systems have improved herbicide efficiency, they can inadvertently accelerate the evolution of resistant weed biotypes (Chang et al., 2023; Upadhyay et al., 2024a). This growing resistance combined with the limited pipeline of new herbicides, underscores the urgency of adopting other non-chemical solutions (Nath et al., 2024). In this context, intelligent mechanical weeding stands out as a promising direction, integrating the precision of robotics with the selectivity of machine vision to address both inter- and intra-row weeds effectively. As a non-chemical method, it not only mitigates herbicide resistance, but also reduces soil and water contamination, fuel consumption and overall environmental burden (Machleb et al., 2020).

Mechanical weed control involves the physical removal or destruction of weeds through direct interaction with the soil and vegetation. The primary techniques employed in mechanical weeding include tillage, cutting, and pulling (Cloutier et al., 2007; Hussain et al., 2018; McCool et al., 2018). These methods disrupt weed growth by turning the soil and uprooting root systems thereby inflicting lethal injuries that prevent regrowth (Zawada et al., 2023). Their effectiveness largely depends on the method, timing, and intensity of operation (Machleb et al., 2020). Despite their effectiveness, conventional mechanical weeding approaches often lead to crop damage followed by soil compaction, adversely affecting soil aeration and root development (Pannacci et al., 2017). Moreover, the performance of mechanical weeders is influenced by field conditions and weather variability, which limit their adaptability across diverse agricultural contexts. Additional issues such as higher energy, fuel consumption and labor demand also reduce their economic feasibility (Hussain et al., 2018).

Recent advances have integrated automation, robotics, and computer vision into mechanical weeding, giving rise to intelligent mechanical weeders. These systems employ sensors and AI algorithms for real-time weed detection and selective removal, minimizing crop disturbance and operator dependency. Specialized end effectors such as finger weeders, rotary cultivators, and elastic comb mechanisms effectively manage coplex intra-row weeds that were difficult to control manually (Chang et al., 2021). For instance, the SMART CULTIVATOR by Stout Industrial Technology, uses True Vision software for crop recognition and adaptive blade control. It handles diverse crops including artichokes, broccoli, cabbage, and pumpkins—with 99% identification accuracy, 1-2 acres/h field coverage, and up to 96% labor reduction compared with manual weeding (Stout, 2020). These systems perform best at early weed-growth stages, roughly two to three weeks after sowing, when weeds are spatially distinct (Jabran and Chauhan, 2015). Under dense weed canopies, their precision declines, favoring hybrid strategies that pair conventional tools (e.g., harrows or hoes) for inter-row control with AI-based weeders for intra-row precision-maximizing efficiency, minimizing environmental impact and advancing sustainable resource optimized agriculture.

This review therefore aims to critically examine the potential and limitations of intelligent mechanical weed control, focusing on its integration with advanced sensing, perception, and actuation systems. It further identifies the technological gaps, operational bottlenecks, and future research pathways required to develop scalable, efficient, and sustainable robotic solutions that align with the broader goals of precision agriculture.

The recent reviews in weed management have predominantly adopted a comprehensive approach, encompassing the entire spectrum of available weed management techniques. These reviews typically provide an overview of various methods, highlighting their advantages, use cases, and technical findings, offering a holistic perspective on weed management challenges and solutions. However, most of these studies are dedicated to exploring the use of ground robots and UAVs for general weed management, often covering a broad range of aspects without delving deeply into specific techniques. Additionally, these reviews frequently separate the discussion of weed removal and weed detection, even though an integrated approach addressing both domains is critical for developing effective, sustainable solutions.

This paper aims to address the existing research gap by focusing on a critical aspect of weed management—advancements and emerging trends in mechanical weed removal techniques for precise and intelligent weed removal. It demonstrates the use of imaging technologies for accurately localizing and positioning weeds, a foundational component of intelligent mechanical weeding systems. Furthermore, the review incorporates detailed case studies that illustrate prevailing trends and methodologies, along with a comparative analysis of the performance of platforms employing these technologies. By adopting this focused approach, the paper emphasizes the potential of integrating weed detection and removal into a cohesive system, offering valuable insights into contemporary precision agriculture practices.

In the context of a literature review, two primary approaches are commonly employed: systematic review and narrative review. The systematic approach follows a structured methodology that addresses specific technical questions by systematically analyzing relevant research papers. It provides a comprehensive synthesis of results, methodologies, and key findings from existing studies. Conversely, the narrative approach focuses on providing a theoretical background and conceptual understanding of the subject matter, primarily adopting a qualitative perspective rather than a quantitative one (Rother, 2007). This review adopted a systematic approach guided by the PRISMA 2020 guidelines to ensure transparency, reproducibility, and methodological rigor (Page et al., 2021). This approach was chosen over a narrative approach to enable structured identification, screening, and synthesis of peer-reviewed studies addressing both traditional and smart mechanical weed control systems.

The overall workflow of the review from topic definition through article extraction and synthesis is summarized in Figure 1. To retrieve relevant literature, two primary databases-Google Scholar and Web of Science were used between October 2024 to January 2025. These databases were selected for their broad interdisciplinary coverage spanning precision agriculture, agricultural robotics, and artificial intelligence. Cross verification confirmed that most relevant studies, indexed in other major engineering/computer-science databases, were already represented within these two databases ensuring comprehensive inclusion while minimizing redundancy.

A Boolean keyword-based strategy was adopted to capture the widest possible scope of research activity: (“mechanical weeding” OR “mechanical weed control” OR “mechanical weeder”) AND (“computer vision” OR “machine vision” OR “image processing” OR “deep learning” OR “artificial intelligence” OR “vision”).

The initial research retrieved 11,275 publications (Google Scholar = 7430; Web of Science = 3845). Duplicate records were removed through automatic deduplication followed by manual verification yielding 9862 unique articles. Titles and abstracts were screened according to predefined inclusion and exclusion criteria (Table 1). The process involved multiple iterations to verify relevance.

After screening, 2705 articles were retained for full-text evaluation. Studies not written in English, review papers, purely economic assessments or those unrelated to mechanical or intelligent mechanical weeding were excluded. Ultimately, 176 technical papers directly addressing mechanical or automated weeding were included for quantitative analysis. From this dataset, 33 focal studies were selected for in-depth synthesis based on quality scoring rubric that assessed five dimensions: (i) design clarity, (ii) details of sensor/actuator integration, (iii) availability of algorithmic or performance data, (iv) validation method and (v) completion report. Each study was rated on a five-point scale (1 = low detail to 5 = comprehensive) with the highest scoring studies forming the analytical subset for further discussion. The complete identification and selection process has been illustrated in the PRISMA flow diagram, Figure 2.

Data extracted from the selected papers included mechanical actuation type (rotary, finger, blade, linear, etc.), sensing and detection technology (RGB, LiDAR, multispectral, hyperspectral), (c) robotic actuation configuration and key performance indicators such as precision, accuracy, operational speed etc. Potential sources of bias were qualitatively evaluated by examining the transparency of experimental design, availability of performance data and validation consistency across studies. To contextualize the growth of research in this domain, Figure 3 illustrates the trend in peer-reviewed publications on traditional and mechanical weeding between 2000 and 2024. The data show an increasing trend in work related to intelligent mechanical weeding after 2019 reflecting the rapid adoption of smart technologies in agricultural automation.

Several previous review studies have investigated different aspects of weed management, including weed control techniques, sensor-based weed detection, robotic weeding systems, and artificial intelligence-driven weed detection approaches (Gao and Su, 2024; Machleb et al., 2020; Rai et al., 2023; Upadhyay et al., 2024b). However, given the breadth of existing literature, the present review adopts a more specialized focus on smart mechanical weeding technologies, emphasizing recent advancements, their integration with imaging sensors for automation and actuation and the technical challenges involved.

Mechanical weeding, a practice that employs tools, implements, and machinery for weed control, has been a cornerstone of agricultural weed management since ancient times, complementing manual hand-pulling. This method has demonstrated high effectiveness in eliminating weeds while ensuring no chemical residues are left on crops (Pannacci et al., 2017; Zimdahl, 2018). The mechanical approach primarily involves processes such as cutting, burying, or uprooting weeds, effectively destroying these undesired plants (Hussain et al., 2018).

The primary objective of a precise mechanical weeding system is to effectively eliminate and manage weeds while minimizing environmental impact, reducing crop injury, and ultimately enhancing net yield. These goals form the foundation for designing and developing an efficient weed management system. Moreover, economic feasibility and platform versatility are crucial factors that influence the adoption of these systems by farmers and stakeholders. These elements play a vital role in motivating users to invest in advanced weed management technologies. Based on an extensive literature review and analysis of documented research examples, it is clear that significant progress is still needed to optimize the performance of such systems to fully realize their potential.

Currently, the most commercially available mechanical weeding systems are tractor-mounted implements designed for traditional weed management methods. However, the adoption of intelligent and autonomous mechanical weeding technologies remains in its early stages. Farmers and end-users are gradually becoming familiar with these advanced systems. Despite significant progress in autonomous agricultural systems, including deep learning and machine vision-based weed detection, several limitations and challenges persist.

Technical challenges in autonomous mechanical weeding systems include precise identification and localization of weeds in real-time scenarios. Although cutting-edge deep learning models, such as the latest variants of the YOLO (You Only Look Once) framework, have shown excellent performance in weed detection and classification, several challenges remain. These challenges encompass issues related to dataset quality and requirements, variations in lighting conditions, morphological similarities between crops and weeds, overlapping images in datasets, and the dynamic nature of environmental conditions. Additionally, the different growth stages of weeds further complicate detection tasks. Beyond these data-related challenges, computational costs, hardware limitations, and potential misidentifications or incorrect detections present significant hurdles. For instance, inadequate datasets or insufficient model training can lead to performance degradation in real-world scenarios, particularly under varied environmental conditions (Rai and Sun, 2024; Saiwa, 2024; Wang et al., 2019).

To address these limitations, deep learning models should be trained on larger and more diverse datasets enhanced through augmentation techniques, such as brightness adjustments, noise addition, and rotation, to simulate real-world variations. Furthermore, the model architecture implemented on edge devices must prioritize lightweight designs to reduce computational complexity while preserving accuracy. Striking a balance between model performance and processing speed is essential in selecting the optimal model for real-time weed detection tasks. By ensuring this trade-off, the models can achieve higher accuracy and efficiency in field operations, ultimately leading to more effective and reliable weed management solutions.

Another major challenge in autonomous mechanical weeding systems is the potential damage to crops caused by the physical interaction of mechanical tools with the soil during weeding operations. This issue is critically important, serving as a key selection criterion for any mechanical weeding system. While the primary aim of these systems is to eliminate weeds, unintended harm to crops can reduce yield, thereby undermining the system’s effectiveness. Although inter-row weeders generally operate between crop rows and pose minimal risk to crops, the concern increases with intra-row weeders, which operate in closer proximity to the soil and crops.

The weeding end effector analyzed in this review exhibit a diverse array of operational methodologies, each characterized by distinct levels of efficiency and an associated percentage of crop injury. At present, crop injury and inadvertent interactions, arising from misidentifications, variations in weed morphology, and diverse soil types, continue to pose challenges that have not yet been thoroughly addressed. The design of end effector capable of precise targeting with minimal impact on crops represents an area of ongoing scholarly inquiry. Approaches designed to tackle these challenges encompass the advancement of high-resolution weed mapping techniques, facilitating the creation of an accurate real-time distribution of both weeds and crops, thereby clearly delineating the operational territory for the end effector. Furthermore, the designs of end-effectors ought to be adaptable to various weed growth stages and soil conditions.

To tackle the challenges related to autonomous mechanical weeding systems, the manufacturing and research sectors must follow a comprehensive set of guidelines and standards. These guidelines should act as a benchmark for creating effective and broadly acceptable systems capable of achieving precise weed eradication and sustainable management. The following points outline the essential requirements for the proper functioning and operation of weed management systems:

While mechanical weeding serves as a reliable option for effective weed management, an integrated method that combines mechanical systems with advanced techniques such as laser-based weed removal and “see-and-spray” technologies can enhance performance even further. This hybrid strategy provides intelligent application capabilities while emphasizing environmental safety.

Robotic systems that meet these criteria and follow the guidelines should be prioritized for commercialization and large-scale implementation. These guidelines not only provide a roadmap for developing autonomous and sustainable mechanical weed management systems but also highlight future research directions aimed at tackling the global challenge of weed management. Advanced mechanical weeding approaches should focus on net yield as a critical performance indicator, optimizing system design to maximize crop yields for specific varieties. By aligning system performance with yield outcomes, mechanical weeding systems can become a transformative solution for precision agriculture.

The ongoing research and advancements in the integration of Artificial Intelligence, machine vision and robotics into mechanical weed control systems has advanced significantly, positioning intelligent mechanical weeding systems as an important context in precision agriculture sector. This review examines a broad body of literature synthesizing insights on mechanical weeding methods, sensing and detection technologies, integrated robotic platforms and actuation strategies that define the state of the art in autonomous weed management. The evidence shows that intelligent mechanical weeders leveraging high resolution RGB cameras, sensors and DL models such as YOLO and ResNet, are steadily improving in their capacity to identify and localize weeds under real field conditions. These systems, paired with adaptive end-effectors provide an increasingly reliable means of controlling inter and intra-row weeds with greater precision compared to traditional approaches.

Over the past two decades, intelligent mechanical weeding systems have progressed from conceptual prototypes to increasingly functional field-ready platforms, yet their widespread adoption remains constrained by persistent challenges. Accurate weed–crop differentiation under diverse field conditions is still difficult, as lighting variability, morphological similarities, and occlusion often reduce classification accuracy and compromise real-time actuation. Synchronizing detection models with mechanical tool actuation in real time also requires high computational resources, creating trade-offs between speed and accuracy when deployed on embedded or edge devices. Crop injury risks, inconsistent performance across soil types and terrains, and high system costs further limit confidence among farmers and stakeholders. These obstacles highlight that the next generation of intelligent mechanical weeding systems must be not only technologically advanced but also economically viable, scalable, and easy to use.

Promising research directions are emerging to address these gaps. Lightweight and efficient deep learning algorithms optimized for edge computing platforms will be essential to achieve accurate and real-time detection in resource-constrained environments. Adaptive end-effectors capable of operating across different crop systems, weed growth stages, and soil conditions can minimize crop disturbance while improving robustness. Likewise, reliable navigation will increasingly depend on sensor fusion approaches that combine LiDAR, GPS, IMUs, and computer vision to deliver stable performance in heterogeneous and unstructured terrains. Modular platform designs with intuitive user interfaces will lower adoption barriers by enabling farmers to operate and troubleshoot systems without specialized expertise. Just as important, more long-term and large-scale field trials must be conducted under varying agro-ecological conditions to validate performance, ensure reliability, and strengthen confidence in these technologies.

The future of intelligent weeding will likely rest on hybrid approaches that integrate mechanical control with complementary methods such as selective see-and-spray systems or laser-based techniques. Such integration can enhance precision, reduce herbicide reliance, and provide a more holistic and sustainable weed management solution. By combining the mechanical precision of robotics with the flexibility of other innovative methods, these hybrid platforms can address both environmental sustainability and economic feasibility. In conclusion, while significant challenges remain, the convergence of AI-driven detection, adaptive mechanical tools, and advanced robotic platforms is paving the way for a new era of weed management. Intelligent mechanical weeders hold the potential to reduce chemical inputs, improve crop yields, and promote sustainable farming practices. As research continues to refine detection accuracy, actuation efficiency, navigation reliability, and system affordability, these technologies are poised to become indispensable components of precision agriculture, offering practical and scalable solutions to the pressing challenge of sustainable food production.