Artificial Intelligence and Agronomy: An Introductory Reflection on Reducing Herbicide Dependence in Weed Management

Abstract

The crop production sector faces the critical challenge of effectively managing weeds while reducing herbicide dependence, which aligns with environmental and economic sustainability. This chapter explores the shift toward site-specific weed management (SSWM), accelerated by artificial intelligence (AI) and digital technologies. Also, it addresses the often-neglected complexities of weed-seed bank germination. We propose an integrated approach, combining AI-enhanced weed detection, cover crop strategies to limit weed seedling emergence, cost-effective spot spraying, and the application of large language models to enrich decision-making under an integrated weed management (IWM) scheme. This helps ensure varied management tactics and weed resistance prevention. We present findings from our Chilean case study, which provide insights into real-world challenges and successes, and highlight the study’s limitations, such as the specific agroecological conditions and limited sample size, which may affect the generalizability of the results to other contexts. We draw comparisons with global AI-driven weed management advancements. This chapter underscores the potential of such integrated strategies to lower herbicide reliance and contribute to sustainable, technologically advanced weed control, fostering environmental stewardship and economic viability in the face of climate change.

1. Introduction

The challenge of managing agricultural weeds is rapidly evolving, increasingly intersecting with the critical needs for environmental sustainability and economic efficiency. Traditional methods, heavily reliant on broad-spectrum herbicide applications, face growing scrutiny. These methods have raised concerns due to their significant ecological, health, and economic implications [1]. Within this dynamic and challenging landscape, the introduction of artificial intelligence (AI) and digital technologies marks a pivotal shift, offering an opportunity for innovation in weed management strategies [2].

As weeds regularly present a spatial distribution or clusters on fields, site-specific weed management (SSWM) [3, 4], a concept deeply rooted in precision agriculture principles, is at the forefront of this transformation. SSWM advocates for applying the necessary weed control measures precisely where and when they are needed, thus ensuring optimal use of inputs [5, 6]. By leveraging the capabilities of AI and digital technologies, SSWM promises a targeted approach to weed control, significantly reducing herbicide use and minimizing environmental impacts [3]. However, the journey toward the effective implementation of SSWM is fraught with challenges, mainly due to the resilience of weed seedbanks and the swift germination rates of weed seedlings, which pose substantial barriers to effective management using localized treatments [4].

The above-described elements are introduced into an integrated weed management (IWM) framework [7]. This synergistic approach blends the strengths of AI-enhanced weed detection, the ecological benefits of cover crops, and the precision afforded by spot spraying technologies. SSWM into IWM not only aims to curtail herbicide reliance but also ensures the maintenance of effective weed suppression, thereby striking a delicate balance between agricultural productivity and sustainability [5].

Another important aspect we consider is addressing the complexities of IWM, especially in the context of commonly limited access to expert guidance in weed science to the grower; this chapter delves into the emerging role of large language models (LLMs). With their profound ability to process and interpret extensive datasets, LLMs emerge as indispensable allies in the quest for enhanced decision-making within IWM frameworks. They have the potential to give growers nuanced insights, enabling more informed and precise agricultural decisions.

Empirical evidence from a pioneering case study in Chile highlights AI’s tangible potential in detecting weeds and leveraging cover crops to mitigate weed pressures. This exploration offers invaluable insights into the practical application and effectiveness of AI-driven weed management strategies, shedding light on the real-world challenges and triumphs of embracing such innovative technologies [7].

A comparative analysis with global advancements draws a comprehensive landscape of AI’s current and potential future in agronomy. This examination showcases global progress and delineates the challenges and opportunities on the horizon, underscoring the indispensable need to weave technological innovation with agronomic expertise [8, 9].

The chapter is structured as follows: Section 2 provides background and rationale on the prevalence and impact of weeds, the implications of herbicide overreliance, and an introduction to integrated weed management (IWM). Section 3 discusses the advancements in AI for agronomy, including deep learning and computer vision techniques. Section 4 delves into site-specific weed management (SSWM) and its integration with digital technologies. Section 5 explores spot spray technology’s role in advancing precision weed management. Section 6 outlines the conditions for establishing SSWM, including weed population reduction, seed bank management strategies, and the utilization of cover crops. Section 7 examines the potential role of large language models (LLMs) under an integrated weed management (IWM) scheme. Finally, Section 8 discusses future directions and challenges in AI-driven weed management.

2. Methodology

2.1 Literature search and selection

For this chapter, a comprehensive literature search used the customed search agents “Consensus” and “SciSpace” operating in the ChatGPT platform (https://chat.openai.com/). In addition, the AI-driven assistant software Scite (https://scite.ai/assistant) was used. The keywords used for the search included terms related to weed management, artificial intelligence, precision agriculture, and cover crops, as mentioned throughout the chapter. The criteria for paper selection involved relevance to the research topic, publication in peer-reviewed journals, and the presentation of original research findings or significant reviews. Papers that did not meet these criteria or were not published in English were excluded from the analysis.

2.2 Weed detection methodology

For weed detection presented in Section 6.4.2, YOLO (You Only Look Once) v4 models were employed [10]. These state-of-the-art deep learning models were trained on a dataset of annotated images capturing various weed species and crop types under different environmental conditions. The models were optimized for accuracy and real-time performance, enabling precise weed identification and localization within the field. The detected weed instances were then used to generate high-resolution weed density maps, which served as the basis for targeted weed management interventions.

2.3 Cover crop trials

A field trial was conducted to assess the effectiveness of cover crops in suppressing weed growth and reducing weed seed bank levels (Section 7.5). The cover crop selected for this study was rye (Secale cereale), cultivated from May 2023 to December 2023 on a one-hectare plot near Chillan city, Chile. Two nitrogen fertilization treatments were applied: a high rate (150 kg N/ha) and a low rate (50 kg N/ha). The cover crop was terminated at two growth stages: the tillering stage (using glyphosate application) and the anthesis stage (using a roller-crimper). Weed pressure, assessed as the density of weed seedlings per square meter, was evaluated before cover crop termination and at regular intervals after the establishment of the subsequent cash crop. The data collected were subjected to statistical analysis to determine the significance of differences in weed pressure among the treatments.

3. Background and rationale

3.1 Prevalence and impact of weeds in agriculture

The prevalence of weeds within agricultural contexts significantly hampers crop yields, necessitating the development of robust weed management strategies. The complexity of these challenges varies across different agricultural ecosystems, highlighting the need for dynamic management approaches. Studies comprehensively review the global weed spectrum and its evolving challenges, illustrating the dynamic nature of weed interactions within agricultural ecosystems [11, 12]. Different weed species present distinct challenges, competing with crops for resources, engaging in allelopathy, and serving as vectors for pests and diseases. For example, the presence of Chenopodium album in cornfields can result in yield reductions of up to 40%, showcasing the aggressive competition for nutrients, light, and water [13]. Similarly, Glycine max, often termed volunteer soybean when it grows outside its intended cultivation area, can diminish soybean yields by approximately 35%, underscoring the complexities of managing weeds closely related to the crops [14].

The impact of weeds varies depending on the crop type, environmental conditions, and weed density. A comprehensive analysis of how environmental factors influence weed-crop interactions emphasizes the importance of timely weed management [15]. Furthermore, the stage of weed emergence in the crop plays a crucial role in the level of competition and potential yield loss, with early emerging weeds posing significant threats to crop establishment [16].

This situation underscores the need for an integrated approach to weed management. Traditional methods primarily focused on chemical control are increasingly complemented by integrated weed management (IWM) strategies that incorporate ecological and biological aspects of weed control. Reviews stress the significance of adopting a holistic and sustainable combination of cultural, mechanical, biological, and chemical methods [16, 17]. The move toward IWM signifies a growing awareness of the necessity for environmentally sustainable and economically viable weed control strategies.

3.2 Environmental, economic, and social implications of herbicide usage

Using herbicides while effectively controlling weeds carries significant environmental, economic, and social implications. From an environmental standpoint, herbicides contribute to soil degradation, water pollution, and harm to non-target plant and animal species [18]. Economically, herbicide dependency escalates farmers’ costs, particularly when faced with herbicide-resistant weed strains. Socially, there are heightened health concerns for consumers and agricultural workers exposed to these chemicals, prompting a reevaluation of the long-term sustainability of intensive herbicide use [19].

3.3 Introduction to integrated weed management (IWM) and its significance

Integrated weed management (IWM) marks a pivotal shift in confronting these challenges. IWM adopts a comprehensive approach encompassing various weed control techniques, including mechanical, cultural, biological, and chemical strategies [20, 21]. Selecting specific weed management approaches within an IWM framework depends on the weed species present, crop type, environmental conditions, and available resources [22]. Its importance stems from its ability to lessen the dependency on herbicides, thereby mitigating their environmental, economic, and social impacts [23]. Furthermore, IWM seeks to bolster the sustainability of agricultural operations by promoting biodiversity, enhancing soil health, and maintaining the long-term ecological equilibrium of farming ecosystems [24].

4. Advancements in artificial intelligence for agronomy

Integrating AI into agronomy, the science and practice of crop production and soil management, promises a transformative era, offering innovative strategies that enhance precision and sustainability in agricultural practices. AI refers to developing computer systems capable of performing tasks that typically require human intelligence, such as visual perception, speech recognition, decision-making, and language translation. This paradigm shift is particularly evident in the realm of precision agriculture and sustainable weed management, where AI technologies have the potential to revolutionize traditional approaches [25]. Despite a growing number of studies, there are few practical applications at the farming level [26].

AI technologies are increasingly important in advancing agronomy, providing sophisticated solutions that range from crop monitoring to disease detection and resource optimization. At the forefront are machine learning (ML) and deep learning (DL) algorithms, which have the potential to contribute significantly to the precision and efficiency of agricultural practices. These algorithms, capable of analyzing extensive and complex datasets, enable the prediction of crop health and the detection of diseases with remarkable accuracy, facilitating targeted interventions that enhance crop management while minimizing the need for chemical inputs. Deep learning, employing convolutional neural networks, has been critical in extracting accurate information from complex images, marking a revolution in field application developments during the last decade. The application of computer vision technology, which, through drone and satellite imagery, allows for the early detection of anomalies and opens the possibility for targeted management strategies that significantly reduce herbicide use, paving the way for more sustainable agricultural practices [27, 28].

The advent of robotics and autonomous systems represents another potential leap toward the automation of agricultural practices. These systems, powered by AI, undertake various tasks, including planting, harvesting, and precise weed control, thus reducing labor costs and minimizing chemical dependencies [29]. Furthermore, predictive analytics, driven by AI models, offer valuable insights that assist farmers in making well-informed decisions. This predictive capability supports optimizing resource use and maintaining crop health, showcasing AI’s potential to impact agricultural outcomes significantly [30].

Several impactful case studies demonstrate the practical application of AI in agronomy. Precision farming projects employing AI-driven drones and sensors have markedly reduced the indiscriminate application of herbicides by accurately identifying areas requiring treatment [31]. Additionally, AI for pest detection has facilitated the development of systems capable of early pest infestation detection. Through machine learning algorithms, automated systems enable targeted pest control measures that significantly reduce chemical treatments [32]. Large language models (LLMs, i.e., adjusted versions of ChatGPT by OpenAI or Claude by Anthropic) promise to revolutionize agricultural extension services by providing real-time, customized advice to farmers, enhancing productivity and sustainability [33].

While AI presents significant opportunities for agronomy, its widespread integration encounters clear technical obstacles. The primary challenges are the high costs of implementing AI solutions and the complexity of these technologies, which can inhibit their adoption across the agricultural sector [34]. Furthermore, the necessity for stringent data privacy and security measures demands the establishment of comprehensive data governance frameworks to safeguard sensitive agricultural data [35]. As an example, an essential component of effectively applying AI within agriculture is ensuring compatibility between AI systems and existing agricultural infrastructure. In this way, the success of AI applications relies on their ability to seamlessly integrate with current practices, highlighting the need for interoperability and standardization [36]. Addressing these technical challenges is vital for leveraging AI to improve agricultural efficiency and sustainability.

5. Site-specific weed management and AI-driven precision weed control

5.1 SSWM background

Site-specific weed management (SSWM) strategically employs herbicides for precise weed control, enhancing efficiency while minimizing environmental impacts and safeguarding non-target species [37]. This concept started with photoelectric diodes in the 1970s and 1980s and evolved through the 1990s with the adoption of machine learning and camera technologies for refined weed identification [38]. Nowadays, some SSWM initiatives incorporate advanced methods such as laser and electrical weeding, capitalizing on computer vision and deep learning with deep convolutional neural networks (CNN) for unparalleled precision in weed detection.

The role of artificial intelligence (AI) in SSWM is transformative, significantly enhancing weed detection and identification accuracy through advanced machine learning algorithms and computer vision techniques. This precision allows for targeted herbicide application directly onto weeds, substantially reducing broad-spectrum herbicide use [39]. Consequently, this focused strategy not only diminishes herbicide consumption but also mitigates the emergence of herbicide-resistant weed strains [40]. Thus, AI-driven SSWM represents a harmonious blend of technological innovation and environmental stewardship, establishing a new standard for sustainable and cost-effective weed management practices.

5.2 Integration of digital technologies

Digital technologies, including drones, sensors, and geographic information systems (GIS), are integral to SSWM, bolstering real-time monitoring and decision-making capabilities [41]. Drones with high-resolution cameras and AI algorithms provide comprehensive field scans, identifying weed infestations with remarkable accuracy [13]. Field-deployed sensors offer valuable environmental data, aiding in predicting weed emergence patterns [42]. GIS applications synthesize these data, producing detailed maps that guide targeted weed control measures, effectively optimizing SSWM strategies [43].

5.3 Advantages of SSWM

SSWM presents numerous advantages over traditional herbicide application methods. It significantly reduces the volume of herbicides needed, decreasing the environmental footprint of weed management and promoting biodiversity and soil health [44]. Economically, SSWM yields considerable savings for farmers by optimizing herbicide usage and reducing labor requirements [43]. Furthermore, SSWM boosts crop yield and quality by eliminating competition with weeds for resources [45].

6. Spot spray technology: advancing precision in weed management

6.1 Theoretical underpinning and significance of spot spray technology

Spot spray technology marks a significant advancement within the site-specific weed management (SSWM) framework. It enables selective herbicide application on weeds, markedly reducing the use of agrochemicals [46]. This development is increasingly critical amid growing calls for environmental conservation and economic efficiency. The fusion of this technology with AI-driven weed detection represents a leap toward achieving unmatched precision and effectiveness in herbicide application [47].

6.2 Technological innovations and their implications for targeted herbicide delivery

The evolution of spot spray technology has seen the introduction of systems equipped with sophisticated sensing and AI algorithms [48]. These innovations facilitate the precise differentiation between crops and weeds, allowing for targeted herbicide applications [49]. The deployment of drones and autonomous vehicles for spot spraying highlights this technology’s adaptability and scalability, providing versatile solutions for various agricultural contexts [50].

6.3 Environmental and economic ramifications of spot spray technology adoption

Spot spray technology brings significant environmental benefits, notably reducing herbicide runoff and protecting soil and aquatic ecosystems [51]. Economically, it offers the potential for cost savings through reduced herbicide use and increased crop yields by efficiently managing weed competition. Additionally, it addresses the challenge of herbicide resistance, ensuring the sustainable management of weeds [52].

6.4 AI integrations in Chilean agriculture for weed management

6.4.1 Initiatives to incorporate AI methods to weed management

Since 2021, Chile has embarked on innovative initiatives led by the Instituto de Investigaciones Agropecuarias (INIA) and the Centro Nacional de Inteligencia Artificial (CENIA) and supported by the Ministry of Agriculture. These efforts represent a concerted move toward incorporating artificial intelligence (AI) into agriculture, with a specific focus on leveraging large language models (LLMs) and advanced weed detection technologies. Central to these initiatives is the development and application of AI as the cornerstone for advancing SSWM and associated technological innovations. The agricultural landscape in Chile is markedly diverse, encompassing a broad spectrum of crops across various regions. This diversity, coupled with relatively smaller farm sizes compared to countries like the USA, Argentina, and parts of the European Union, presents unique challenges and opportunities for implementing AI-driven solutions. The initiatives aim to tailor AI technologies to meet the specific needs of Chilean agriculture, ensuring relevance and applicability across different scales and types of farming practices.

Also, a notable challenge in the region is the scarcity of weed science specialists, intensifying the need for innovative, personalized weed management solutions. The shortage of experts in this field has heightened the urgency for deploying AI technologies that can offer scalable, efficient, and accessible support to farmers. The collaborative efforts between INIA and CENIA are directed at filling this critical gap, utilizing AI to democratize access to expert knowledge and decision-support tools in weed management.

6.4.2 First experiences

Efforts have been concentrated on testing methodologies for various weed-crop combinations, image types (aerial or terrestrial), lighting conditions, and the phenological development of the detected plants. This was achieved using deep learning models for object detection. The results indicate that using artificial intelligence models makes it possible to quantify different weed species across diverse environmental and phenological conditions of the populations studied. In these cases, the overall prediction accuracy (F1 score) exceeds 80%, with a peak of 96%. Some images of these predictions are shown in Figure 1(a–h).

Figure 1.

Weed detection results for models for various crops in Central-Southern Chile: This figure presents representative results of weed detection across different agricultural settings, including (a) wheat fields, (b) autumn chemical fallow, (c) maize cultivation, (d) lentil fields, (e) industrial chicory plants, (f) wild carrot inflorescences, and (g, h) Echinochloa crus-galli inflorescences. Each detection is highlighted with squares, indicating the species identified, accompanied by their respective EPPO codes.

Despite these promising results, various aspects require improvement. A primary concern is integrating all versions of the artificial intelligence model into a single or a few models, for example, according to phenological states, capable of predicting across a broader range of weed/crop scenarios. The second challenge is enabling these models to function at machinery operational speeds under field conditions. Progress has been made toward real-time weed detection under variable field conditions and for diverse crop and weed combinations, where more than 24 frames per second can be analyzed (see Video 1, https://drive.google.com/file/d/1zTZml60zujauSTZrt0zbyOhYsSTLfabE/view?usp=drive_link; Video 2, https://drive.google.com/file/d/1CXafkJYgPevsLxVmSSYVhjMxjQIZufKd/view?usp=drive_link; and Video 3, https://drive.google.com/file/d/1N82wKsl4qJasLHNM-2nQW3-eCq7NKBV2/view?usp=drive_link).

The strategic decision to focus on first place on developing computer vision models for SSWM is predicated on their foundational role across various applications, from weed mapping to the operation of site-specific weed control machinery. This emphasis arises from the need for robust databases that capture diverse scenarios of weed and crop interactions and adaptable models for widespread application. Once deployed, these models require iterative fine-tuning to accommodate to new field scenarios, such as changes in crop-weed dynamics, lighting conditions, and phenological stages. Consequently, the initial models will be refined over time, learning and evolving through their operational lifespan. Streamlining this process to ensure robustness and efficiency poses a significant challenge that must be addressed in forthcoming agricultural seasons.

As depicted in Figure 2, the progression from model development stages (Figure 2a–c) to deployment in specialized field-ready computers (Figure 2d) lays the groundwork for the generation of prescriptive maps (Figure 2e). This progression also facilitates the integration of these models into machinery equipped for executing SSWM actions in real time within the field (Figure 2f). Such advancements signify an advance in SSWM, where machine learning enhances weed management efficacy and becomes a continuous learning process, adapting and improving with each crop cycle.

Figure 2.

This schematic illustrates the integrated process of employing computer vision models for site-specific weed management (SSWM), beginning with (a) aerial and terrestrial data acquisition over crops via drones. The workflow progresses to (b) image processing, where computer vision techniques identify weeds. Subsequently, (c) machine learning algorithms, represented by neural networks, learn to distinguish between crops and weeds. The trained model is then deployed on (d) a portable computer tailored for field conditions, which supports generating (e) prescriptive maps for targeted interventions. Finally, (f) machinery equipped with precision spraying technology executes SSWM actions in real time, as directed by the insights gained through advanced AI modeling.

6.4.3 Next steps

The momentum behind these initiatives in Chile is set to be further propelled by a new project titled “Weed Management Using Artificial Intelligence for Sustainable Agriculture in Wheat, Rice, and Legumes in Chile” (2024–2026), supported by the Foundation for Agrarian Innovation (FIA) agency. The main goal of this project is to develop and implement an AI system for mapping weeds and providing sustainable recommendations and controls, thereby enhancing the efficiency and sustainability of wheat, rice, and legume agriculture in the central-southern zone of Chile. Its specific objectives include:

• Developing AI models that are applicable under our productive conditions for weed recognition.

• Implementing an AI-based weed mapping system.

• Developing a language model trained to give sustainable weed management recommendations.

• Integrating the mapping system and the language model into a user-friendly tool for farmers and service companies.

• Implementing integrated weed management, combining AI mapping, personalized recommendations, and variable control under real conditions.

• Conducting extension and dissemination activities toward farmers, technical assistance services, and service companies.

This project embodies a comprehensive approach to embedding AI into Chile’s agricultural practices, focusing on critical crops such as wheat, rice, and legumes. By leveraging AI for precise weed detection and management, the initiative aims to revolutionize traditional practices, making them more sustainable and efficient. This endeavor demonstrates the potential for AI to address specific agricultural challenges and sets a precedent for how such technologies can be adapted to different environmental and agricultural contexts. The focus on creating an accessible tool for farmers underscores the commitment to ensuring that these advancements are practical and beneficial at the grassroots level, further contributing to Chile’s overarching goal of sustainable agriculture.

Spot spray technology and SSWM use in Chile is part of a worldwide trend toward technologically improved and sustainable weed management methods. This approach emphasizes the significance of precision agriculture in increasing productivity while maintaining a healthy environment. It also underscores the need for global cooperation and knowledge exchange to promote the implementation of AI and SSWM [53].

One of the main topics we consider in our workflow in Chile is the adaptability of future spot spray technology and SSWM in different agricultural settings, which will be vital for its widespread use [11, 54]. The fact that this technology can be customized according to specific management, economic, environmental conditions, and crop species shows its potential for local adaptation.

7. Conditions for establishing site-specific weed management

Identifying and implementing foundational conditions conducive to effective deployment is critical within the SSWM domain. This section presents a part of the strategy necessary for SSWM’s successful application, emphasizing the management of weed populations per square meter and the significant role of cover crops within an integrated weed management strategy. These elements are crucial in enhancing the accuracy and efficiency of weed control measures in precision agriculture, a consideration not extensively covered in technical literature.

7.1 Reduction of weed populations per square meter

Reducing weed populations per square meter is pivotal for adopting SSWM strategies. This tactic mitigates immediate competitive stress exerted upon crops and aims to curtail the potential of the weed seed bank, thereby addressing a critical aspect often overlooked in the literature.

From our analysis, addressing this concern is imperative; neglecting to do so could negate the advantages of an SSWM scheme. If weed populations exceed a certain density threshold (e.g., one seedling per square meter), the necessity for blanket application methods may be maintained, as used in traditional weed management approaches. Such a scenario not only diminishes the specificity and efficiency gains offered by SSWM but also incurs increased application costs, potentially detracting from the scheme’s appeal and feasibility. Consequently, managing seedling densities per square meter presents an agronomic challenge that necessitates a long-term, holistic perspective, encompassing the entire crop rotation cycle and leveraging all available agronomic tools (cultural, mechanical, etc.) throughout successive growing seasons. In this sense, integrating advanced precision technologies, including drone-assisted surveillance and AI-enabled identification, and, most importantly, proper agronomical management through rotation is vital for effectively managing these populations [55, 56].

Therefore, it is essential to view SSWM and its associated technologies not as standalone solutions but as integral components of an IWM program. Such integration is indispensable, particularly under conditions observed in Chile, where managing cover crops emerges as a promising strategy to alleviate seedling pressure, as detailed in Section 6.5. Incorporating cover crops effectively into IWM schemes underscores the potential to significantly reduce weed seedling emergence, thereby enhancing the efficacy of SSWM strategies in maintaining weed populations below critical thresholds and ensuring sustainable agricultural productivity.

7.2 Weed seed bank management strategies

The weed seed bank poses a significant threat to agronomic productivity by storing seeds capable of causing future infestations. Effective seed bank management involves strategies to reduce seed viability and prevent germination. Techniques such as soil solarization, strategic tillage, and pre-emergent herbicides can be part of a comprehensive approach to minimizing the seed bank’s impact [57]. Furthermore, employing models to forecast and model seed bank dynamics can enhance the timing and effectiveness of these strategies [58].

7.3 Cover crops: a dual role in weed and seed bank management

Cover crops play a pivotal role within the SSWM framework, offering dual benefits of weed emergence suppression through competitive exclusion and soil health improvement [52]. Their ability to occupy space and resources deters weed proliferation [43]. Additionally, certain cover crops exhibit allelopathic properties, releasing chemicals that inhibit weed seed germination and growth [59]. Integrating cover crops into crop rotation and management plans highlights their value in comprehensive weed management strategies [60].

7.4 Synergistic integration of cover crops with SSWM

The synergy between cover crops and SSWM techniques significantly enhances weed control efficiency. This combination leverages the ecological benefits of cover crops with the precision and efficacy of SSWM, promoting herbicide reduction and environmental conservation. Digital mapping and AI analyses aid in identifying optimal cover crop species and planting strategies to support specific SSWM practices, providing customized solutions for each farm’s unique needs [61].

7.5 Necessity of cover crops implementation in Chile: initial steps

In Chile’s primary zone for annual crop production, weed infestation from the seed bank significantly exceeds the thresholds for adopting a site-specific weed management (SSWM) system, with infestations often surpassing 200 plants per square meter in various conditions (Figure 3a). This challenge coincides with the urgent need to reduce herbicide applications during the fallow period, especially the use of glyphosate and the excessive reliance on tillage for weed-free soil. In response to these challenges, adopting cover crops has been identified as a strategic initiative to mitigate weed pressure, including annual and complex perennial species previously unutilized in our cropping areas. Cover crops offer a scalable and economically viable strategy that complements the careful use of pre-emergent herbicides.

Figure 3.

(a) High-density weed infestations are common in the irrigated Central Valley of Chile (e.g., Cyperus esculentus in wheat). Initiatives have been launched to incorporate cover crops to reduce the germination weed burden. These initiatives have started with treatments of varying establishment densities, with low densities (e.g., <40 kg of seed/ha) (Figure 3b) compared to high densities (e.g., 150 kg/ha of rye), (Figure 3c) where a noticeable decrease in weed pressure (Raphanus spp.) was observed in the latter case. Additionally, field experiments using different nitrogen doses (Figure 3d) have been conducted, where rye at both high fertilization and lower doses was able to suppress germinating weeds. However, at the high N dose, the rye showed lodging (Figure 3e). Finally, initial experiences with terminating the cover crop (Figure 3f) and sowing the following spring crop in the rotation (beans, Figure 3g) have also been carried out.

Trials with winter rye have been initiated to explore the impact of sowing densities (Figure 3b and c ) alongside evaluations of nitrogen fertilization levels (Figure 3d and e), the termination of cover crops (Figure 3f), and the establishment of spring crops (i.e., dry bean, Figure 3g) following the cover crop phase.

Early results suggested a marked reduction in weed pressure in fields (Figure 3f and g) that previously faced high weed densities, particularly with problematic species such as Ambrosia artemisiifolia and Cyperus esculentus. Implementing cover crops within Chile’s diverse agricultural landscapes aims to validate their practicality and benefits within the SSWM framework. The goal is to integrate cover crops into Chilean agricultural systems effectively, thereby improving weed suppression, soil structure, and the overall yield and quality of crops [62]. This initiative is expected to highlight the adaptability and efficiency of cover crops and SSWM practices across varied agricultural ecosystems [63, 64, 65].

8. Integrated weed management and large language models

8.1 The evolution and impact of large language models in agriculture

The introduction of (LLMs) represents a pivotal advancement in artificial intelligence, significantly impacting various fields, including agriculture [66, 67]. LLMs are defined as deep learning algorithms capable of understanding, generating, and interacting with human language at an unprecedented scale [31]. The global impact of LLMs was notably amplified with the advent of ChatGPT, a variant of the generative pre-trained transformer models, which demonstrated the potential of LLMs to revolutionize not only technology sectors but also industries far afield, such as agriculture [68]. This technology’s ability to process and generate text akin to human writing has revolutionized decision-making, predictive analytics, and automation in agricultural practices [69]. Recent developments in LLMs have led to the creation of sophisticated algorithms that interpret complex agricultural data, offering insights and strategies for effective farm management [70, 71, 72].

8.2 LLMs’ role in weed management strategies

LLMs have the potential to revolutionize weed management strategies by providing intelligent, data-driven solutions to the complex challenges faced in integrated weed management (IWM) and site-specific weed management (SSWM). These advanced AI models can process and analyze vast amounts of data from various sources, such as satellite imagery, weather patterns, soil conditions, and historical weed occurrence records. By leveraging this wealth of information, LLMs can generate highly personalized and context-specific weed control recommendations, tailored to individual farms’ unique needs or even specific field sections.

LLMs’ ability to provide expert-level guidance is particularly valuable in regions with limited access to weed management specialists. By encapsulating the knowledge and experience of top experts in the field, these models can serve as virtual advisors, offering farmers real-time, actionable insights on the most effective weed control strategies for their specific situations. This empowers farmers to make informed decisions and promotes the wider adoption of IWM and SSWM practices, even in areas with limited access to professional support.

However, it is important to note that the application of LLMs in weed management is still in its early stages. When writing this chapter, researchers and companies are actively exploring and developing prototypes that showcase the immense potential of these models.

Current foundational LL models, such as GPT-3, BERT, and T5, have demonstrated remarkable capabilities in various natural language processing tasks. However, these models are not specifically trained or fine-tuned for the agricultural domain, particularly in weed science. To harness the full potential of LLMs in addressing weed management challenges, it is crucial to adapt these models to the specific language, terminology, and context of agriculture. This requires a collaborative effort between AI researchers and domain experts in weed science to curate relevant datasets, define domain-specific tasks, and fine-tune the models accordingly. By incorporating the knowledge and expertise of weed scientists into developing agricultural-specific LLMs, we can create powerful tools that provide accurate, reliable, and actionable insights for farmers and land managers. This is an urgent task that we, as domain experts, must undertake to bridge the gap between AI advancements and the agricultural community’s practical needs, ultimately promoting sustainable and effective weed management practices.

While the current implementations may be limited in scope and scale, the rapid advancements in AI and the growing availability of agricultural data suggest that LLMs will play an increasingly pivotal role in shaping the future of weed management. As these models continue to evolve and mature, they are poised to become indispensable tools for farmers, enabling them to tackle the complexities of IWM and SSWM with unprecedented precision and efficiency [73].

8.3 Addressing agricultural challenges with LLMs

Farmers must navigate a complex web of factors, including unpredictable weather conditions, soil health, pest and disease outbreaks, and market fluctuations, all while striving to optimize crop yields and maintain economic viability. The scarcity of domain experts, particularly in specialized areas like weed science, exacerbates these challenges, leaving many farmers without access to the knowledge and guidance needed to make informed decisions.

This is where LLMs can play a transformative role. By serving as virtual consultants, LLMs can bridge the knowledge gap and provide farmers with the insights and recommendations they need to overcome their myriad challenges. These models can analyze vast amounts of data from multiple sources, including scientific literature, historical records, and real-time sensor data, to generate comprehensive and actionable advice tailored to each farmer’s specific circumstances.

By functioning as virtual experts, LLMs can democratize access to specialized knowledge and empower farmers to make data-driven decisions that optimize their operations and mitigate risks. This is particularly crucial in regions where access to agricultural specialists is limited, as LLMs can fill the void and provide farmers with the support they need to navigate the complexities of modern agriculture. As LLMs continue to advance and integrate with other precision agriculture technologies, they can transform farmers’ problem-solving and decision-making, ultimately leading to more sustainable, resilient, and productive agricultural systems.

8.4 Future prospects of LLMs in agriculture

The future integration of LLMs within the agricultural sector promises further refinement, with models becoming increasingly customized to address specific farming challenges. Anticipated advancements include the capacity of LLMs to assimilate real-time data, offering dynamic, personalized advice to farmers.

Incorporating LLMs IWM and SSWM heralds a transformative era for agriculture, helping to address its complexities and uncertainties. As these technologies continue to evolve, they are set to play a pivotal role in enhancing sustainable weed management practices, providing vital support to farmers worldwide. This synergy marks a significant stride toward greater efficiency, sustainability, and resilience in agricultural systems.

9. Future directions and challenges

9.1 Emerging trends in AI for weed management

Innovations at the cutting edge of agricultural technology, such as autonomous weed control robots, drone-based surveillance with hyper-spectral imaging, and AI-driven models for predicting herbicide resistance, are redefining the landscape of SSWM. These advancements promise to enhance the precision and efficacy of weed management strategies [74, 75, 76].

The merger of blockchain with AI in agriculture introduces a new level of data transparency and traceability, crucial for building trust across the supply chain. This fusion facilitates a cooperative platform for seamless data exchange and informed stakeholder decision-making. It propels agriculture toward a smarter, more sustainable future by enabling precise resource use and enhancing operational efficiency [77].

9.2 Limitations, challenges, and barriers to adoption

Introducing AI technologies and infrastructure comes with high costs, posing substantial challenges, especially for smallholder farmers and those in developing regions. Furthermore, the inherent complexity of AI systems necessitates specialized knowledge and extensive training [78, 79].

The lack of structured and comprehensive databases hinders the widespread adoption of large language models (LLMs) and computer vision models in weed science. LLMs require high-quality, domain-specific data to generate accurate insights, while vision models rely on large amounts of labeled images to effectively learn and classify weed species. However, in weed science, there is a scarcity of well-organized and easily accessible datasets that can be used to train and fine-tune these models.

Developing robust LLMs and vision models for weed management necessitates a rich corpus of data encompassing various aspects of weed biology, ecology, and control strategies. Without such comprehensive datasets, these models may struggle to provide reliable and context-specific recommendations to farmers and land managers.

Collaborative efforts among weed scientists, data experts, and AI researchers are crucial to address this limitation. Building and curating structured databases involves collecting, organizing, and annotating data from diverse sources, such as field experiments, remote sensing, and farmer observations. Standardizing data formats and creating user-friendly interfaces for data access and contribution can further facilitate the development of tailored LLMs and vision models.

Establishing data-sharing protocols and incentives for researchers and institutions to contribute to these databases can accelerate the growth and refinement of these resources. By fostering a culture of open data and collaboration within the weed science community, we can create a robust foundation for developing accurate and reliable models that revolutionize weed management practices.

9.2.1 Out-of-domain problem in DL models

From our point of view, addressing the out-of-domain (OOD) problem will be crucial when incorporating AI models into IWM schemes, where deep learning (DL) models often encounter data vastly different from what they were trained on. This discrepancy, particularly pronounced due to the diverse and unpredictable nature of agricultural environments, can significantly impair the models’ performance. For instance, an AI system trained to identify specific weed species might struggle when presented with images or conditions it has not seen before. To overcome this challenge and enhance model reliability, it is essential to develop generalist models. These models are designed to adapt flexibly to new situations by drawing on a broad and varied dataset and incorporating robust architectures to changes. Doing so ensures accuracy across various agricultural scenarios, ensuring effective decision-making and operational efficiency [80].

9.2.2 Finetuning models for dynamic agricultural environments

Given the dynamic nature of agricultural operations, there is a pressing need for the ongoing and efficient finetuning of DL models to ensure their rapid adaptability to changing conditions. Initiatives like the current FIA-Chilean project, “Weed management with the use of artificial intelligence for sustainable agriculture in wheat, rice, and legumes in Chile”, showcase innovative methods to increase data labeling efficiency and enhance model adaptability.

10. Conclusions

This chapter explores the critical challenge of effective weed management while reducing herbicide dependence in crop production. We propose a shift toward AI-driven site-specific weed management (SSWM) as a promising approach to address this challenge.

Our research emphasizes the importance of considering weed-seed banks and seed germination in weed management. We present an integrated approach combining AI-enhanced weed detection, cover cropping, spot spraying, and large language models (LLMs) to ensure varied management tactics and prevent herbicide resistance.

Our Chilean case study provides insights into the real-world implementation of AI-driven SSWM, demonstrating its potential in enhancing weed management and reducing herbicide reliance despite limitations in generalizability due to specific agroecological conditions and limited sample size.

This chapter compares global advancements and underscores the transformative potential of integrated strategies combining AI, precision agriculture, and ecological approaches for sustainable and adaptable weed control in the face of climate change.

However, adopting AI-driven SSWM faces challenges, such as high initial costs, lack of structured databases, and the need for collaboration between AI researchers and weed science experts to fine-tune AI models.

In conclusion, despite limitations and challenges, AI-driven SSWM has significant potential in revolutionizing weed management. An integrated approach combining AI, precision agriculture, cover cropping, and ecological principles can lead to more sustainable and effective weed control strategies. Stakeholders must collaborate and invest in making these technologies and practices more accessible and adaptable to diverse farming contexts, contributing to a more sustainable and resilient future for crop production.

Acknowledgments

We sincerely thank the “Fundación para la Innovación Agraria” (FIA) and the Ministry of Agriculture of Chile for their support of our research. Their commitment has been essential in advancing our work on the integration of artificial intelligence (AI) in weed management, highlighting the significance of collaborative innovation in agriculture. The authors acknowledge the use of ChatGPT, for language polishing of the manuscript.

Conflict of interest

The authors declare no conflict of interest.