生成式人工智能赋能水稻系统农场级农技咨询的综述

Abstract

Agriculture is increasingly characterized by a data paradox, while the sector generates massive volumes of genomic, climatic, remote sensing, and field data. Translating this information into actionable, farm-level insights remains a critical bottleneck. Traditional advisory mechanisms cannot operate at the spatial scales or provide the context-specificity needed for climate adaptation and food security. The work presents GenAI as a transformative interface that makes advanced agricultural science, including crop simulation models and remote sensing diagnostics, accessible to rice farmers in India through natural language, provided that challenges related to data sovereignty, infrastructure gaps, and the need for human oversight in advisory systems are effectively addressed. We trace the shift from static, rule-based expert systems to dynamic foundation models capable of complex reasoning and multimodal analysis. The paper critically reviewed the rise of crop-specific LLM agro-advisory model architectures, such as SeedLLM-Rice and IPM-AgriGPT, which outperform general-purpose models by utilizing specialized scientific corpora to reduce hallucinations. Main attention is on the hybrid integration framework of LLMs with Knowledge Graphs (KGs) for factual grounding, and connecting with process-based simulators (e.g., DSSAT, APSIM) to make biophysical modelling more accessible through natural language. Furthermore, the manuscript examines the fusion of satellite, drone, and smartphone imagery with vision-language models, enabling real-time, context-aware diagnostics for smallholder farmers. The review concludes that while GenAI has the potential to support more equitable forms of precision agriculture, it can only do so under specific conditions. Its success depends on resolving infrastructure gaps, ensuring data sovereignty, and maintaining a “human-in-the-loop” architecture to guarantee scientific rigour and social inclusion.

Graphical Abstract

1 Introduction

1.1 Background

Agriculture is operating under the combined pressures of ensuring food security, adapting to climate variability, and reducing environmental impacts, which will shape food systems in the coming decades (Sapkota et al., 2025). By 2050, the world’s agricultural systems will need to grow food for an estimated 10 billion people apart from adapting to unpredictable changes in climate patterns and achieving sustainable intensification that preserves natural resources and the environment (Wise, 2013). This convergence of demographic pressure, environmental uncertainty, and sustainability imperatives demands transformative innovation in how agricultural knowledge is generated, disseminated, and applied across diverse farming contexts. Traditional agricultural advisory systems face fundamental limitations despite their historical contributions (Manimozhi and Krishnamoorthy, 2025). Extension services operate at a limited scale with delayed responses to emerging threats due to low advisor-to-farmer ratios (Bonilla et al., 2024). Generic recommendations fail to address diverse farm conditions, particularly affecting underserved smallholder farmers in developing regions (Tumwebaze et al., 2025; Prajapati et al., 2025). The digital revolution has transformed agriculture’s data landscape through IoT sensors, satellite imagery, and genomic sequencing (Birner et al., 2021; Marudamuthu et al., 2023). However, translating these massive, varied datasets into practical, actionable insights for farmers remains extraordinarily difficult.

Generative AI (GenAI) fundamentally differs from traditional machine learning approaches in agriculture (Sai et al., 2025). While conventional systems focus on prediction and classification, including forecasting yields, detecting diseases, and optimizing irrigation (Paudel et al., 2021; Bharadiya et al., 2023), GenAI employs large language models (LLMs), diffusion models, and generative adversarial networks to generate content, facilitate reasoning, and enable natural conversations (Shaikh et al., 2025; Yang et al., 2025a). Unlike standard deep learning models that merely estimate disease likelihood from images, GenAI systems engage farmers conversationally, explain reasoning, integrate multi-source knowledge, provide personalized advice, and generate synthetic training data for data-limited conditions (Sai et al., 2025; Chirra et al., 2025).

India’s rice production faces critical sustainability challenges that threaten the nation’s food security (Gathorne-Hardy et al., 2016). Kumar et al. (2021) documented that long-term intensive rice cultivation in northern Indian plains has led to severe deterioration of natural resources, declining factor productivity, multiple nutrient deficiencies, depleting groundwater, and labor scarcity. The rice-wheat cropping system in the Indo-Gangetic Plains shows particularly alarming trends, with groundwater tables declining approximately one meter per year while crop productivity stagnates (Khedwal et al., 2023). Rice demand in India is expected to rise sharply by mid-century, while resource degradation and environmental stresses threaten sustainable production. GenAI presents transformative solutions to address these multifaceted challenges. The AgriLLM initiative, developed by Consultative Group on International Agricultural Research (CGIAR) in partnership with AI71, provides region-aware, role-specific responses supporting farmers, advisors, and decision-makers with multilingual capabilities (CGIAR, 2025). GenAI is not here to replace what farmers already know, but to be a powerful helper that makes expert farming advice accessible to everyone. It breaks down technical barriers, allowing small farmers who may not read well or use technologies to ask questions in their own language and receive informed answers (Kumar et al., 2025; Sai et al., 2025; Chirra et al., 2025). It does not replace local agricultural advisors, but it helps them reach more farmers more quickly with up-to-date information and tools.

1.2 Scope and objectives

This review addresses the central research question: What is the current landscape of generative AI applications for rice systems in India, and how do these applications bridge agronomic science with farm-level advisory delivery? This scoping review establishes clear boundaries, providing focused yet comprehensive coverage of generative AI applications in agriculture and food systems. While the agricultural challenges faced are universal, this analysis examines global trends with a particular focus on rice production systems. Rice serves as the staple food for more than half of humanity, predominantly in Asia, Africa, and Latin America, where smallholder farming remains the dominant production model (Asma et al., 2023; Mohapatra et al., 2025). The recent development of SeedLLM, a specialised large language model for rice cultivation, exemplifies how crop-specific AI systems are emerging to address the unique agronomic complexities of individual commodities (Lu et al., 2024; Yang et al., 2025a). This focus allows us to explore both the technical innovations and practical implementation challenges that characterize the frontier of agricultural AI development.

The geographic scope of this review is centred on Indian agricultural contexts but draws on international evidence where Indian-specific literature is sparse, including from related institutional bodies such as the International Rice Research Institute (IRRI), CGIAR, and the Food and Agriculture Organization (FAO) CABI (2024). Of the 183 sources included in the scoping synthesis, 25.7% directly address Indian agricultural contexts, while the remainder contribute methodological, architectural, or international policy perspectives that inform a synthesis with implications for Indian smallholder rice systems. This India-centred framing with transferable international evidence is consistent with the realities of an emerging research field in which crop-specific generative AI work for any single national context remains, by necessity, a small fraction of the global literature.

Temporally, this work concentrates on the transformer era in AI, beginning with the architectural breakthroughs that enable modern LLMs and extending through the present moment. Since around 2017, model capabilities have proliferated, foundation models have emerged, and methods for domain adaptation have advanced significantly. By focusing on this transformative period, the analysis demonstrates how advances in general-purpose AI were adapted and refined for agricultural applications. Earlier rule-based expert systems and traditional machine-learning methods are mentioned only to provide historical context and to highlight the major shift introduced by generative artificial intelligence.

This manuscript outlines five objectives illustrating agricultural AI development and future directions. It traces evolution from early rule-based systems through machine learning, deep learning, to modern generative AI, emphasizing how innovations have transformed agricultural knowledge while highlighting risks of losing valuable earlier approaches. The review examines crop-specific LLMs like SeedLLM for rice versus broad foundation models, discussing domain adaptation, data requirements, and specialization trade-offs. It presents integration frameworks combining AI tools, knowledge graphs, sensing systems, and crop simulation models, and acknowledging development complexities. The global policy environment is evaluated, with particular emphasis on India’s agricultural AI, offering lessons on regulation, investment, digital access, and smallholder farmer benefits. Finally, it identifies transformative opportunities like personalized advisory systems and accelerated crop improvement, while addressing challenges including data gaps, infrastructure limitations, technical issues, and adoption barriers. Together, these objectives map the current landscape of generative AI in agricultural advisory, expose integration gaps across sensing, knowledge, model, interface, and governance layers, and identify the institutional and policy questions that will shape responsible deployment in Indian rice systems.

2 Materials and methods

In this scoping study, a comprehensive search approach has been employed to encompass both formal research studies and practical examples of generative AI in agriculture. Because the field is advancing at an exponential rate, important innovations often appear first on preprint servers, blogs, or open-source platforms before they are published in peer-reviewed journals. This indicates that our approach extends beyond conventional academic databases to capture the full range of relevant evidence in this fast-moving field. We conducted structured searches across multiple complementary databases to ensure comprehensive coverage of the literature. Web of Science and Scopus provided access to established peer-reviewed literature, capturing research that has undergone formal quality assessment. Google Scholar has extended its reach to include conference proceedings, technical reports, and publications from regional journals that may not be indexed in primary services. To maintain reproducibility, Google Scholar results were limited to the top 180 relevance-ranked records, consistent with established scoping review practice (Haddaway et al., 2015). Recognizing that cutting-edge AI research often appears first as preprints, we systematically searched arXiv’s computer science and quantitative biology sections, as well as bioRxiv, for agricultural applications with computational components. This multi-database approach enabled coverage across formally vetted academic literature, emerging preprint research, and grey literature (Figure 1).

Figure 1

The keyword strategy struck a balance between specificity and comprehensiveness to identify relevant literature. Core search terms included “large language model agriculture,” “generative AI crop,” “agricultural knowledge graph,” “multimodal agricultural AI,” and “crop simulation LLM.” The related terms-, such as “foundation model farming,” “transformer agriculture,” “conversational AI advisory,” “diffusion model crop breeding,” and crop-specific queries like “rice language model” and “wheat phenotype generation,” were supplemented. Named-system queries were also included to capture specific crop-adapted models reported in the field, such as “AgriLLM,” “AgriGPT,” “SeedLLM,” “AgriLLaVA,” “AgroLLM.” Boolean operators enabled the combination of domain terms (such as agriculture, farming, crop, and agronomy) with AI technology terms (LLM, generative AI, GPT, transformer, and diffusion model) to capture interdisciplinary work. The study primarily focused on literature from 2020 to 2026, a period marked by rapid growth in generative AI capabilities and agricultural applications. This timeframe encompasses the transition from early transformer models to sophisticated foundation models, as well as their domain adaptation for agricultural applications. Selected literature from 1980 to 2019 provides essential background on the evolution from rule-based expert systems to conventional machine learning approaches and establishes the historical context necessary to understand the paradigm shift that generative AI represents. Grey literature formed a crucial component of the evidence base, reflecting the reality that much innovation in applied AI occurs outside traditional academic publishing. Systematically reviewed reports from bodies such as the Ministry of Agriculture and Farmers Welfare, the Department of Science and Technology, Government of India, and the Indian Space Research Institute, which capture policy perspectives and national strategic initiatives, are also part of the detailed scoping.

Sources were eligible for inclusion if they addressed generative AI, large language models, foundation models, diffusion models, or knowledge graph–integrated systems with relevance to agriculture, with priority given to rice, related cereals, and Indian agricultural contexts. Peer-reviewed journal articles, conference proceedings, substantive preprints, book chapters, and institutional grey literature were all included. Studies addressing agricultural AI without generative or foundation-model components, opinion pieces lacking technical substance, non-English publications, and duplicate records were excluded.

Records identified across all sources were exported and de-duplicated, after which a two-stage screening process was applied: title and abstract review against the eligibility criteria, followed by full-text assessment of records passing the first stage. Discrepancies were resolved through discussion among the authors. A total of 632 records were identified, of which 500 remained after duplicate removal. Title and abstract screening excluded 278 records, and a further 32 were excluded at full-text review, yielding 190 sources included in the scoping synthesis. The complete screening process is summarized in Figure 1.

During synthesis, included sources were organized according to a consistent set of characteristics to enable thematic comparison across the evidence base. These characteristics included publication type, the primary AI architecture or approach, the target crop and agricultural task, geographic focus, the stage of evaluation reported (distinguishing benchmark performance, experimental evaluation, and real-world deployment), and the principal limitations noted. Consistent with PRISMA-ScR guidance (Tricco et al., 2018), formal critical appraisal of individual sources was not undertaken, as the review aims to map the breadth of evidence and identify thematic patterns rather than to evaluate study-level methodological quality. Given the heterogeneity of the included sources, a narrative synthesis approach was adopted, supplemented by structured comparative tables and an integrative conceptual framework.

3 An integrative framework for generative AI in rice advisory

Generative AI in agriculture is usually described as a series of separate advances: a crop-specific language model, a knowledge graph, a sensing pipeline, a chatbot, a policy paper. Each is useful alone, but the literature says little about how they connect into a working advisory system. A farmer benefits only when data from the field is structured, reasoned over, and returned as usable guidance. To organise the synthesis that follows, we propose a framework tha