Leveraging large language models (LLMs) for GeoAI-enabled digital agro-advisory

Abstract

Digital agriculture has undergone a profound transformation driven by rapid advances in Earth observation, unmanned aerial vehicles and advanced data processing techniques. Despite this progress, climate-smart agriculture management faces a critical challenge: a farmer-centric agro-advisory system. Farmers and extension workers have limited access to GeoAI-based agro-advisory services, primarily due to limited data access and insufficient technical support. This study synthesises the potential of Large Language Models to address this challenge by translating complex GeoAI data into actionable and human-centric advice. The review examines gaps, recent advancements, and the model architectures required to adapt general-purpose LLMs for remote-sensing-based crop monitoring and management. We distinguish between experimentally validated capabilities of LLMs and the broader prospective applications proposed for agricultural remote sensing, noting that many current advances are primarily driven by multimodal foundation models, computer vision, and GeoAI systems rather than standalone LLMs. The study highlights the importance of advanced Retrieval-Augmented Generation and Supervised Fine-Tuning in agronomic science and mitigating the risk of misinterpretation. Further, we examine the emerging capabilities of multimodal approaches, which can seamlessly integrate visualisation and textual reasoning to support stakeholders in assessing crop health conditions and biophysical anomalies. The representative case studies, spanning multiple geographies and including voice-based advisory systems, demonstrate the shift from static advisory tools to dynamic, interactive recommendation systems. We highlight current challenges, such as the need for region-specific fine-tuning, data governance, and operation in low-connectivity environments, and advocate for a “human-in-the-loop” approach to decision-making. Through this, the LLMs will function as co-pilots assisting multiple stakeholders rather than as autonomous decision-makers vulnerable to biased output or hallucination problems. The review concludes by outlining a forward-looking research scope and closed-loop systems that iteratively learn from field outcomes.

1 Introduction

Over the past decade, digital agriculture has undergone a profound transformation driven by rapid advances in Earth observation, unmanned aerial vehicles (UAVs), Internet of Things (IoT), and proximal sensing technologies. These developments have enabled continuous, multiscale, and unprecedented volumes of spatiotemporal data to be generated for monitoring crops, soils, hydro-ecology, and agroecosystem processes (Gebbers and Adamchuk, 2010). Operational satellite missions such as Landsat, Sentinel, MODIS, LISS, Cartosat, WorldView, PlanetScope, combined with very high-resolution UAV imagery, now form the backbone of precision agriculture, supporting applications including crop monitoring, yield estimation, cropping practices and the implementation of climate-smart agricultural practices (Fritz et al., 2019; Thenkabail et al., 2021). Despite the growing maturity of remote sensing technologies, the increasing availability of data has exposed a fundamental limitation in digital agriculture: the interpretation and convergence bottleneck. Remote sensing data are typically expressed as spectral reflectance or backscatter values, vegetation indices, and derived biophysical parameters, which require substantial agronomic and contextual expertise to interpret and apply to improve crop management. Translating these quantitative indicators into actionable recommendations requires an understanding of crop type, phenological stage, soil characteristics, climatic conditions, and management history, which is complex (Atzberger, 2013). This complexity constrains the timely delivery of decision-support information, particularly in smallholder-dominated agricultural systems where access to trained extension services remains limited and costly (Thornton et al., 2014). Recent advances in Large Language Models (LLMs) offer a promising framework for addressing these challenges. LLMs such as the Generative Pre-trained Transformer (GPT), Large Language Model Meta AI (LLaMA), and Pathways Language Model (PaLM) are based on transformer architectures that model long-range dependencies, support contextual reasoning, and enable large-scale knowledge synthesis (Brown et al., 2020; Vaswani et al., 2017). Unlike conventional machine learning models, which are optimised for narrowly defined prediction tasks, LLMs are designed to interpret, explain, and adapt to complex information across domains. These capabilities position LLMs as an effective interface layer for translating remote-sensing-derived indicators into human-interpretable, decision-relevant guidance for agronomists, extension agents, and farmers.

The potential of LLM-enabled agricultural advisory systems is further enhanced by recent advances in multimodal foundation models and vision-language frameworks that integrate visual, textual, and geospatial information. It is important, however, to distinguish between standalone LLMs and multimodal computer vision systems. Standalone LLMs are primarily designed for language understanding, knowledge synthesis, reasoning, and natural-language generation, whereas remote sensing image interpretation is largely performed by computer vision, vision-language, and multimodal foundation models. Recent vision-language models (VLMs) and geospatial foundation models have demonstrated increasing capability in analysing remote sensing imagery, field photographs, and UAV observations for applications such as crop stress detection, disease identification, canopy trait assessment, phenological monitoring, and interpretation of spectral anomalies (Li J. et al., 2024; Weng et al., 2025). These models generate structured visual insights that can subsequently be translated into actionable recommendations through LLM-based reasoning and conversational interfaces. This distinction is particularly important given the exponential growth of Earth observation imagery, driven by the expansion of satellite constellations and the widespread adoption of UAV platforms for agricultural management. While multimodal foundation models and GeoAI systems increasingly perform perception and analytical tasks associated with image understanding, LLMs primarily contextualise these outputs, integrate agronomic knowledge, retrieve relevant information from external knowledge bases, and communicate recommendations in a human-interpretable form. Therefore, the emerging opportunity lies not in replacing remote sensing analytics with LLMs, but in integrating multimodal perception systems with language-based reasoning frameworks to support end-to-end agricultural decision-making. Despite this rapid progress, a systematic understanding of how LLMs can be effectively integrated into remote sensing-based agricultural workflows remains lacking. The existing literature predominantly focuses on remote sensing for crop monitoring or the application of LLMs in broader agricultural contexts, leaving a crucial gap in current knowledge. To date and to our knowledge, no comprehensive review has synthesised how LLMs specifically enhance the interpretation, communication, and decision-support functionalities of remote sensing in digital agriculture. Moreover, critical deployment challenges such as the risk of model “hallucination”, the necessity for region-specific fine-tuning, and strategies for effective implementation remain largely unexplored. Existing studies tend to focus either on remote sensing-based crop monitoring and modelling (Atzberger, 2013; Weiss et al., 2020) or on the application of LLMs in broader agricultural advisory and knowledge systems (Hansen et al., 2019) with relatively little synthesis across these domains. As agricultural monitoring systems increasingly transition from descriptive assessment to operational advisory services, the need for intelligent interpretation frameworks becomes more acute. LLMs are efficient in data mining through historical observations and practices, textual reports, and agronomic knowledge bases, thereby reducing human effort. The Earth observation foundation models highlight the growing integration of geospatial AI, multimodal learning, and LLM-assisted agricultural analytics. For instance, the IBM-NASA Prithvi model uses multi-temporal satellite imagery and transformer-based learning to support applications such as crop monitoring, land-use mapping, and environmental assessment (Jakubik et al., 2023). Similarly, the European Space Agency (ESA) WorldCereal initiative applies AI-driven analysis of Sentinel satellite data to generate global crop maps for major cereals, including wheat and maize (Van Tricht et al., 2023). Although these systems are not standalone LLMs, they demonstrate how foundation models and multimodal geospatial AI are advancing scalable agricultural monitoring and advisory systems. Future integration of these Earth observation models with conversational LLM frameworks could further improve farmer-centric decision support and agro-advisory services (Xiao et al., 2025). Therefore, a coherent understanding of how LLMs can be systematically integrated into remote sensing-driven agricultural workflows will improve farm management and decision support.

This review aims to address the following vital gaps: (a) To synthesise the potential for integrating advanced remote sensing tools/techniques and LLMs for precision/digital agriculture and advisory generation. (b) To present representative case studies illustrating the integration of LLMs with satellite- and UAV-based crop monitoring workflows, focusing on rice-based cropping systems. (c) To critically examine technical, operational, and governance challenges, including reliability, validation, and ethical considerations associated with LLM-based advisory. (d) To outline a forward-looking roadmap/framework that prioritise multimodal model development, regionally adaptive fine-tuning, and edge or low-bandwidth deployment strategies.

By bridging remote sensing analytics with LLM-based knowledge representation, this review argues that foundation models have the potential to significantly enhance the scalability, contextual relevance, and inclusivity of digital agriculture. When appropriately grounded in agronomic knowledge and validated against observational data, LLMs can transform geospatial information from an expert-centric analytical resource into a human-centred and actionable advisory ecosystem that supports informed decision-making (Thornton et al., 2014). Figure 1 illustrates the knowledge translation pipeline from remote sensing data acquisition through biophysical indicator derivation to LLM-driven farm-level advisory, highlighting the interpretation gap that LLMs are positioned to bridge.

FIGURE 1

1.1 Scope and methodology of this review

This review follows a structured narrative and conceptual review approach to map the intersection of LLMs, remote sensing, and agricultural advisory systems. The literature search was conducted across Web of Science, Scopus, Google Scholar, and the arXiv preprint repository using Boolean combinations of key terms including “Large Language Model,” “LLM,” “Generative AI,” “remote sensing,” “digital agriculture,” “precision farming,” “crop advisory,” “farm advisory chatbot,” “RAG agriculture,” and “multimodal geospatial AI.” The search covered publications from 2010 to 2025, with emphasis on post-2020 literature to capture the rapidly evolving LLM landscape. Inclusion criteria required that studies: (i) explicitly addressed the application of LLMs or conversational AI in agricultural contexts, (ii) incorporated remote sensing data or geospatial analytics, or (iii) presented technical architectures (SFT, RAG, RLHF, multimodal fusion) with relevance to agro-advisory generation. Studies that focused exclusively on conventional machine learning for crop classification, without an advisory or interpretive component, were excluded. Grey literature, including technical reports from CGIAR centres, FAO, and national agricultural extension bodies, was also consulted to capture operationally deployed systems not yet represented in peer-reviewed literature. A total of over 330 sources were screened, of which 97 were retained for detailed synthesis (Figure 2). The review is organised thematically rather than chronologically, progressing from the remote sensing foundation (Section 2) through LLM adaptation pathways (Section 3), representative field deployments (Section 4), deployment challenges (Section 5), and future research frontiers (Section 6).

FIGURE 2

2 Overview of remote sensing in digital agriculture and associated challenges

Remote sensing has become a cornerstone of modern digital agriculture, providing the spatial and temporal multi-dimensional information needed to monitor, assess, and manage complex agroecosystems. The expansion of various satellite platforms, UAV-based imaging systems, and sensor-integrated field networks has revolutionised our ability to observe crop conditions, map spatiotemporal variability, and optimise resource use (Mulla, 2013; Thenkabail et al., 2009). These advances fuel precision agriculture, enabling data-driven decisions from crop suitability, fertilisation and pest management, crop health monitoring, water resource utilization, and yield forecasting. Although a large proportion of the latest remote sensing-based crop management uses machine learning models, most of these trained models operate in silos and require human intervention for consistent interpretation and translation into actionable advice for farmers, extension workers, and policymakers.

2.1 Landscaping the diverse remote sensing applications in agriculture

Satellite remote sensing remains the backbone of national, regional, and global agricultural monitoring systems, providing consistent, spatially explicit observations of cropland dynamics over large areas. A diverse constellation of publicly accessible and commercial Earth observation platforms, including multispectral optical sensors such as Sentinel-2, Landsat-8/9, and PlanetScope, MODIS, WorldView, LISS, as well as synthetic aperture radar (SAR) systems such as Sentinel-1, EOS-04, and hyperspectral data such as HySI and PRISMA, and satellite data-derived products such as soil moisture (by SMAP), now deliver high-frequency data streams essential for operational agricultural applications (Belgiu and Csillik, 2018; Singh et al., 2021; Torres et al., 2012). Despite these advances, a key limitation of satellite remote sensing lies in its spatial, spectral, and temporal resolutions. Most available data have revisit times of 5–6 days–20 days and moderate spatial resolution, which limits their application during the crucial crop-monitoring phase. While freely available sensors typically operate at spatial resolutions of 10–30 m, which are often insufficient to capture within-field variability in small farmlands with field sizes below 1 ha (Fritz et al., 2015; Guilpart et al., 2017). As a result, satellite products are highly effective for strategic monitoring and policy-level assessments in homogeneous fields. However, they are less suited for heterogeneous small field-level decision-making without costly high-resolution imagery.

To overcome the spatial resolution limitations of satellite-based observations, UAVs have emerged as a critical component of field-scale agricultural intelligence. UAV platforms provide centimetre-level spatial detail with flexible deployment schedules, allowing crop condition monitoring at user-defined temporal frequencies in key phenological stages (Maes and Steppe, 2019; Wang et al., 2024). Advanced UAV systems can be equipped with a diverse suite of lightweight sensors, each contributing complementary information on crop status. High-resolution RGB cameras generate detailed orthomosaics that support plant stand assessment, lodging detection, weed mapping, and canopy structure analysis (Pon Arasan et al., 2025). Recent advances in deep learning have further expanded the capabilities of UAV and mobile vision systems beyond traditional vegetation index mapping. Convolutional neural networks and real-time object detection frameworks, particularly the YOLO family of models, have enabled automated crop detection, fruit localisation, segmentation, counting, and growth-stage recognition at field and greenhouse scales. For example, Paul et al. (2024) developed an integrated YOLO-based system for capsicum detection, segmentation, growth-stage classification, fruit counting, and real-time mobile identification. Such perception-oriented computer vision models provide detailed crop-level intelligence that can complement satellite-derived biophysical indicators and serve as important upstream information sources for multimodal advisory systems. In future GeoAI-enabled agricultural platforms, outputs from UAV-based detection and crop-stage recognition models could be integrated with LLM reasoning frameworks to generate more context-aware recommendations on harvesting, crop management, and resource optimisation. In parallel, thermal infrared sensors enable the mapping of canopy temperature, which serves as a proxy for plant water status and transpiration efficiency, supporting irrigation scheduling and drought stress detection at fine spatial scales (Maes and Steppe, 2019). These high-fidelity datasets are already reshaping agricultural research and innovation pipelines. For farmers and extension agents, a detailed crop health or thermal anomaly map may indicate variability. However, it does not inherently convey why stress is occurring or what management action should follow. Bridging this interpretive gap remains a challenge in translating UAV-based insights into timely, actionable agricultural decisions.

Recent advances in UAV and mobile vision systems have further extended field-scale monitoring capabilities through deep learning-based crop detection, segmentation, counting, and growth-stage classification. Paul et al. (2024) demonstrated the application of YOLO-based frameworks for real-time capsicum detection, segmentation, growth-stage identification, and mobile deployment, highlighting the growing convergence between computer vision and precision horticultural management.

Remote sensing observations from satellite and UAV platforms are rarely interpreted in their raw spectral form. Instead, they are systematically transformed into standardized biophysical and agroecological indicators that provide an interpretable description of crop condition, canopy functioning, and landscape-scale processes. Optical satellite sensors capture surface reflectance across visible, near-infrared, and shortwave infrared wavelengths, enabling the development of widely adopted vegetation indices. Metrics such as the Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), Soil-Adjusted Vegetation Index (SAVI), Normalized Difference Red Edge Index (NDRE), Photochemical Reflectance Index (PRI), and Normalized Difference Moisture Index (NDMI), are routinely used to characterize crop vigour, biomass accumulation, canopy development, and leaf water content throughout the growing season (Huete et al., 2002). Beyond spectral greenness, structural and radiative parameters offer a more process-oriented perspective on crop growth. Leaf Area Index (LAI) characterises canopy structure by quantifying leaf surface area available for light interception, while the fraction of absorbed Photosynthetically Active Radiation (fAPAR) represents the efficiency with which vegetation captures incoming solar energy. These variables form the foundation of many crop growth models and yield estimation approaches, directly linking remote sensing observations to biomass accumulation and productivity dynamics (Running et al., 2004). Thermal and moisture-related indicators add a critical dimension by capturing crop-water interactions. Land Surface Temperature (LST), derived from thermal infrared observations, is widely used to infer crop water stress, evapotranspiration, and surface energy balance, particularly under heat and moisture-limited conditions (Anderson et al., 2011).

Complementing optical observations, SAR sensors provide a critical capability for agricultural monitoring in cloud-prone, monsoon-dominated environments. Operating independently of solar illumination and atmospheric conditions, SAR backscatter measurements enable continuous tracking of crop presence, phenology, water inundation, and soil moisture dynamics (Torres et al., 2012). These characteristics make SAR data particularly valuable in rice-based production systems and humid tropical regions, where persistent cloud cover often limits the usability of optical imagery during key crop growth stages (Zhao et al., 2021). The microwave data provides an alternative for crop damage assessment during extreme rainfall conditions. Table 1 illustrates the remote-sensing-derived biophysical indicators that underpin nearly every operational workflow in digital agriculture.

Although remote sensing-derived indicators provide valuable information for agricultural monitoring, their reliability is influenced by multiple sources of uncertainty that can propagate through downstream decision-support systems. Spectral observations are affected by atmospheric scattering, aerosol loading, sensor calibration errors, bidirectional reflectance effects, cloud contamination, and mixed-pixel conditions, all of which can introduce uncertainty into vegetation indices and biophysical parameter retrievals (Weiss et al., 2020; Li J. et al., 2024). Temporal uncertainty further arises from revisit limitations, cloud-induced data gaps, and asynchronous observations that may fail to capture rapidly evolving crop stress events during critical phenological stages. These uncertainties become increasingly important when remote sensing products are integrated into Retrieval-Augmented Generation (RAG) frameworks for agro-advisory generation. In a typical RAG pipeline, remote sensing-derived indicators such as NDVI, LAI, soil moisture, or crop stress metrics are first transformed into structured descriptions, which are subsequently used to retrieve relevant agronomic documents, management recommendations, or historical case records. Errors or uncertainty in the original remote sensing observations may therefore influence retrieval quality by directing the system toward partially relevant or inappropriate knowledge sources. For example, cloud-induced underestimation of vegetation condition may trigger retrieval of drought-management recommendations, whereas the actual field condition may be associated with nutrient deficiency or transient atmospheric artefacts.

Recent studies have highlighted the importance of uncertainty-aware Earth observation systems and confidence estimation for operational agricultural monitoring (Jakubik et al., 2023; Xiao et al., 2025). Future GeoAI-LLM advisory systems should therefore incorporate confidence metadata, quality flags, and uncertainty estimates alongside remote sensing products before ingestion into retrieval pipelines. Such approaches would enable confidence-aware prompting, probabilistic retrieval, and transparent communication of advisory reliability. Rather than generating deterministic recommendations, future systems could provide confidence-annotated guidance that explicitly communicates the quality of the underlying observations and associated decision uncertainty.

2.2 Need for LLMs in remote sensing-based agroecological applications

Despite the widespread application of geoinformatics in agriculture, the established approaches are not intrinsically actionable when adopted in isolation. Although the various remote sensing-based indicators offer strong biophysical interpretability, their agronomic reliability is not absolute. The gap between quantitative remote sensing indicators and practical decision-making remains a central challenge. Translating data-driven metrics into actionable guidance requires contextual interpretation, integrating agronomic knowledge, temporal dynamics, local hydroclimatic conditions, applied resources, and optimum yield. Moreover, the relationships between remote sensing-based spectral metrics and crop conditions are often scale-dependent, vary across cultivars and agroecological zones, and are influenced by varied management practices. Although multimodal approaches that fuse optical, thermal and microwave signals help mitigate single-indicator uncertainty by capturing vegetation vigor, water stress, and canopy heat response, they still require contextual information to produce valid agronomic recommendations. Consequently, the process remains semantically inert without expert interpretation, underscoring the critical need for knowledge-grounded reasoning systems. In such diverse and complex scenarios, LLMs can translate multi-source data into operational recommendations tailored to local farming realities, even in heterogeneous landscapes like India.

Moreover, analysing the large volume of multi-sensor data requires substantial computing resources and interdisciplinary technical expertise spanning data science, remote sensing, and agronomy. Cloud-based platforms such as Google Earth Engine (GEE) and the Open Data Cube, along with national digital initiatives, including AgriStack in India, have substantially lowered technical barriers (Gorelick et al., 2017; Killough, 2018). Despite these technological advances, a persistent and critical gap remains between the generation of analytical outputs and their practical use at the farm level. This gap is commonly referred to as the “last-mile” problem, arising from challenges in translating data into timely, accessible, user-friendly communication through systematic channels. LLMs, which emerged as a critical enabler, can assist various stakeholders, from line departments, data analysis experts, extension workers, and farmers, and connect them through a common platform.

Unlike conventional rule-based decision support systems, LLMs are designed to reason across heterogeneous information sources and communicate insights in natural language. In the context of geoinformatics-driven agriculture, LLMs can contribute in three critical ways. First, they can interpret and explain complex geospatial outputs, such as spectral indices, in clear, non-technical language beneficial to farmers and extension agents (Brown et al., 2020). Second, they enable contextualised advisory generation by integrating remote sensing data with contextual information, including crop calendars, hydroclimatic conditions, historical management practices, and short-term weather forecasts (Li J. et al., 2024). Third, LLMs can deliver this guidance in regional languages and dialects, addressing linguistic barriers that have historically limited the reach of digital advisory services in many parts of the Global South (Kshetri et al., 2024). By embedding reasoning, contextualisation, and communication capabilities within a broader AI-assisted framework, LLMs can support the interpretation of remote sensing outputs and facilitate farm-level advisory generation. In practice, data acquisition and primary geospatial processing are typically performed by remote sensing platforms, GeoAI pipelines, and multimodal analytical systems, while LLMs function mainly as interpretive and decision-support components. Figure 3 presents an LLM-assisted geospatial workflow illustrating how multi-source data can be integrated with an LLM-based interpretive layer to generate farm-level advisory outputs under human-in-the-loop governance.

FIGURE 3

The human-in-the-loop framework emphasises that LLM-assisted agro-advisory systems should operate as collaborative decision-support tools rather than autonomous recommendation engines. The remote sensing observations and other environmental conditions are interpreted through LLM-based reasoning layers, while agronomists, extension personnel, and farmers provide contextual validation, local adaptation, and feedback for refinement of the advisories. The periodic human oversight is essential for improving reliability, reducing hallucination risks, and ensuring regionally appropriate farm-level recommendations.

3 LLMs and their adaptation for agricultural remote sensing

LLMs represent a fundamental shift in artificial intelligence, extending beyond conventional structured prediction tasks toward advanced interpretation, reasoning, and natural language generation. Built on the transformer architecture (Vaswani et al., 2017), LLMs are trained on massive, diverse text corpora to learn complex statistical relationships in language. This training endows them with a unique capacity to generate coherent explanations, respond to nuanced queries, and synthesise information across multiple contexts (Bommasani et al., 2021; Brown et al., 2020). These capabilities are particularly relevant for agriculture, where data-driven insights must be interpreted through the lenses of agronomy, environment, and socioeconomics. In contrast to traditional machine learning models, which typically operate as black-box predictors, LLMs excel at articulating why a given pattern or anomaly matters and how it should inform management decisions. This explanatory capacity positions LLMs as a natural interface between complex remote sensing analytics and human decision-making. The self-attention mechanism enables models to dynamically weight the relevance of different elements within a sequence, regardless of their relative positions (Vaswani et al., 2017). The LLM can take inputs from proprietary models (e.g., GPT-4, Gemini) and open-source models (e.g., LLaMA, Mistral, Falcon). However, open-source models are more user-friendly, enabling fine-tuning, on-premises deployment, and adaptation to defined languages and contemporary knowledge systems (Touvron et al., 2023).

3.1 LLM architecture and agricultural insight

Despite their reasoning ability, general-purpose LLMs require domain-specific knowledge to provide advisory services aligned with local and regional agronomic practices. Several approaches have been proposed to calibrate user-specific LLMs for agricultural science. Supervised Fine-Tuning (SFT) involves training an LLM using curated, verified datasets, such as domain-specific expert knowledge (e.g., fertiliser use), pest management guidelines, remote-sensing-based inputs for management, or soil and water resource management based on weather conditions. This equips the model with terminologically and domain-specific logical decisions, thereby improving its relevance and consistency (Luo et al., 2024). The utilization of Reinforcement Learning with Human Feedback (RLHF) refines model predictions iteratively by providing explicit expertise feedback. The Retrieval-Augmented Generation (RAG) is well-suited for agricultural remote sensing applications, which retrieve information from trusted databases during inference. The model then optimises the response based on verifiable data, improves transparency and traceability of information, and reduces the risk of hallucination (Gao et al., 2024; Lewis et al., 2021). For operational agricultural deployment, the effectiveness of RAG systems depends not only on retrieval accuracy but also on the design of the underlying retrieval infrastructure. Agricultural knowledge bases often contain heterogeneous information sources, including agronomic guidelines, crop calendars, weather advisories, remote sensing products, extension bulletins, and scientific literature. These resources are typically converted into vector embeddings and indexed within vector databases such as FAISS, Milvus, Pinecone, or Chroma to enable efficient semantic retrieval (Gao et al., 2024). Retrieval performance is influenced by embedding dimensionality, chunk size, indexing strategy, and database scalability. Excessively large knowledge chunks may dilute retrieval relevance, whereas overly small chunks can fragment contextual information and reduce advisory coherence. Recent studies suggest that hierarchical chunking and metadata-aware retrieval strategies improve retrieval precision while maintaining contextual integrity (Zhao et al., 2024; Gao et al., 2024). Operational systems must also consider retrieval latency and scalability, particularly when supporting large numbers of farmer interactions during critical cropping periods. As knowledge repositories grow, vector search efficiency, embedding storage requirements, and query response times become important constraints. Hybrid retrieval architectures combining dense semantic retrieval with keyword-based filtering, along with dynamic knowledge-base updating, can improve responsiveness while maintaining the relevance of advisory outputs. These implementation-level considerations are essential for translating RAG-based agricultural advisory systems from research prototypes to operational deployment. Figure 4 provides a comparative overview of these three domain adaptation pathways: SFT, RLHF, and RAG, illustrating their mechanisms, strengths, and limitations for agricultural remote sensing advisory.

FIGURE 4

Table 2 summarises these domain adaptation pathways, highlighting that fine-tuning on static corpora is insufficient for operational use and must include inputs from dynamic climate and market conditions. Table 3 further compares general-purpose and agricultural-domain LLMs across key dimensions including parameter count, fine-tuning feasibility, remote sensing data support, multilingual capability, and edge deployment.

3.2 Multimodal LLMs for visual and geospatial data

The latest foundation models, which represent multimodal LLMs, modernise language understanding for visual and structured data modalities. Models such as GPT-4V and open-source counterparts, including LLaMA and BLIP-2, can process text, images, charts, and diagrams simultaneously, enabling integrated reasoning across heterogeneous inputs (Li et al., 2023; Liu et al., 2023). This approach introduces transformative opportunities in agricultural remote sensing, utilising image and index interpretation and spatial pattern recognition for decision-making and recommendations. Such models can analyse field photos to identify discernible stress, pest attack, and disease patterns, for example, chlorosis due to nitrogen deficiency. The perception layer of multimodal advisory systems is often built upon deep-learning-based object detection and crop-stage recognition models. Recent studies have demonstrated the effectiveness of YOLO architectures for agricultural monitoring. For example, Verma et al. (2024) successfully applied a fine-tuned YOLOv8 framework for cotton growth-stage detection, enabling automated phenological assessment and field intelligence generation. Such perception models can serve as upstream information sources for LLM-assisted advisory systems by providing structured crop-condition information for downstream reasoning and recommendation generation. Moreover, these models articulate possible agronomic factors in natural language and bridge the gap between image interpretation and management insights. Although multimodal LLMs can identify the presence of crops in an image, they lack intrinsic significance in geographic (coordinate and reference) systems, spatial topology, scale, and neighbourhood relationships (Reichstein et al., 2019). Thus, effective deployment of LLMs requires integrating spatial context into inference and recommendations. These multimodal advisory systems increasingly rely on upstream perception models, including YOLO-based object detection, segmentation, and crop-stage classification frameworks, which transform raw imagery into structured agronomic information before LLM-based interpretation and recommendation generation.

Recent advances in geospatial foundation models have sought to address these limitations through coordinate-aware transformers, spatial tokenization, and geospatial embedding mechanisms. Unlike conventional LLMs that treat location information as text tokens, coordinate-aware architectures explicitly encode latitude-longitude information, spatial relationships, and neighbourhood context within the model representation. Spatial tokenization strategies partition geographic space into structured units that preserve spatial hierarchy and topology, while geospatial embeddings transform location-specific information into machine-readable representations that can be integrated with transformer-based reasoning. Emerging models such as GeoChat and RemoteCLIP demonstrate how spatially grounded representations can improve interpretation of remote sensing imagery and location-aware querying (Kuckreja et al., 2024; Liu et al., 2024). These developments provide an important foundation for future advisory systems capable of generating recommendations that are sensitive to local agroecological conditions, landscape context, and spatial variability.

While many AI-enabled agricultural monitoring studies focus on cereal crops, horticultural systems present unique challenges that require finer-scale perception and decision-making capabilities. High-value horticultural crops often demand precise monitoring of fruit development, disease symptoms, canopy structure, and harvest readiness at the individual plant level. Recent advances in computer vision, multimodal perception, and robotic systems have enabled automated detection, maturity assessment, and harvesting of horticultural crops under greenhouse and field conditions. These developments demonstrate the broader applicability of multimodal AI beyond large-scale crop monitoring and highlight the growing convergence of computer vision, robotics, and intelligent agricultural management systems (Paul and Machavaram, 2025a).

Traditional CNN-based approaches remain highly effective for crop classification, object detection, segmentation, and other perception-oriented remote sensing tasks. However, recent transformer-based multimodal systems extend beyond image understanding by integrating heterogeneous data sources and supporting contextual reasoning, knowledge retrieval, and advisory generation (Xiao et al., 2025). Table 4 summarises the complementary strengths and limitations of these two paradigms in agricultural applications.

While Earth observation foundation models (EOFMs) such as Prithvi, Clay, and SatMAE have demonstrated remarkable capabilities for extracting spatial representations from large-scale satellite imagery, their outputs are typically feature embeddings rather than directly actionable recommendations. Consequently, an important research challenge involves connecting geospatial foundation models with downstream LLM reasoning layers. A typical workflow begins with satellite, UAV, and sensor observations that are processed by EOFMs to generate geospatial embeddings representing crop condition, vegetation dynamics, soil moisture, or environmental stress patterns. These embeddings can subsequently be transformed into structured descriptors, semantic tokens, or retrieval queries that are integrated with agronomic knowledge bases through Retrieval-Augmented Generation (RAG) pipelines. LLMs then utilize the retrieved contextual information together with geospatial summaries to perform reasoning, explain observations, evaluate management options, and generate farmer-centric recommendations. This layered architecture enables a clear separation between perception (foundation models), knowledge retrieval (RAG systems), and reasoning (LLMs), thereby supporting more reliable and interpretable agricultural advisory systems (Jakubik et al., 2023; Xiao et al., 2025).

A critical component of multimodal agricultural AI systems is the fusion of heterogeneous information sources, including satellite imagery, UAV observations, IoT sensor measurements, weather data, and textual agronomic knowledge (Table 5). Several fusion strategies have been proposed to integrate these diverse data streams. Early fusion combines multiple data modalities at the input level by concatenating features prior to model training, enabling joint representation learning but often requiring harmonized data structures. Late fusion integrates modality-specific outputs at the decision stage, offering greater flexibility when data sources differ in spatial, temporal, or semantic characteristics. More recently, cross-attention mechanisms employed in transformer architectures have enabled dynamic interactions between modalities by allowing the model to selectively attend to relevant information across image, sensor, and text inputs (Li X. et al., 2024; Weng et al., 2025). Graph-based fusion approaches provide an additional framework for representing relationships among fields, sensors, management zones, weather stations, and agronomic knowledge networks. By explicitly modelling spatial and semantic dependencies, graph neural networks and graph-enhanced transformers can improve contextual understanding and support more robust decision-making. Future GeoAI-LLM systems are likely to combine cross-attention and graph-based fusion architectures to integrate Earth observation data, in situ measurements, and domain knowledge within unified reasoning frameworks.

3.3 Supporting farm-level advisory and climate risk management

To date, the majority of digital agriculture relies on decision-rule-based approaches and requires significant human interpretation for advisory, limiting their application to farm-level field conditions across a large number of land parcels in India. A profound feature of LLMs is their ability to interact with users (line departments, extension agents, and farmers) through conversational interaction. This allows the users to ask follow-up questions, seek clarification, or explore alternative management options. Such interactive conversational capability enhances feedback loops between observation, interpretation, and action, improving the user’s confidence. It can be delivered through multiple communication channels, such as dedicated applications and messaging services (Agnihotri et al., 2022).

Extreme weather events, including floods, droughts, heat/cold waves, hailstorms, and cyclones, cause massive damage to agricultural systems in India. Climate models and remote sensing play a crucial role in prediction and detection; however, the timely translation of these predictions/observations into protective/mitigation action remains a significant challenge. LLMs in such cases can be transformative, providing actionable recommendations based on users’ interests and time constraints. For instance, an LLM connected to near-real-time satellite-derived soil moisture and weather-forecasting APIs can proactively alert farmers to impending waterlogging or drought stress and recommend preemptive actions, such as drainage management or supplementary irrigation. Thus, timely information sharing will provide robust early warning and enhance farming communities’ capacity to absorb climate shocks.

It is instructive to compare LLM-based advisory with conventional Decision Support Systems (DSS) that have served agriculture for decades. Traditional DSS, such as DSSAT (Jones et al., 2003) and APSIM (Holzworth et al., 2014), are process-based simulation models that generate recommendations through deterministic crop-growth equations parameterised with local soil, weather, and management data. While these systems offer high agronomic rigour and mechanistic interpretability, they demand substantial input data preparation, expert parameterisation, and are typically limited to specific crops and regions for which they have been calibrated. They also lack natural-language interfaces, making them inaccessible to most farmers and field-level extension agents. In contrast, LLM-based systems sacrifice mechanistic precision for versatility, contextual reasoning across heterogeneous data sources, and conversational accessibility. An LLM can simultaneously interpret a satellite-derived NDVI anomaly, cross-reference it with recent rainfall data, and generate a plain-language recommendation capabilities that a traditional DSS cannot provide without extensive middleware development. However, LLMs carry risks of hallucination and lack the quantitative crop-growth modelling rigour of process-based DSS. The optimal deployment pathway, therefore, is not replacement but complementarity: LLMs serving as the interpretive and communication layer atop quantitative models, with RAG pipelines retrieving process-based simulation outputs to ground LLM reasoning in empirically validated predictions.

Scalability and computational economics represent critical considerations for operational deployment. Cloud-hosted proprietary models (GPT-4o, Gemini) offer high reasoning quality but incur per-query API costs that may become prohibitive at the scale of millions of farmer interactions per season. Open-source alternatives (LLaMA-3 8B, Mistral-7B) significantly reduce inference costs and can be deployed on modest GPU hardware or even edge devices, albeit with some reduction in reasoning depth. Quantisation techniques (4-bit, 8-bit) and model distillation further compress models for deployment on low-cost hardware, making on-device inference feasible in bandwidth-constrained environments. For large-scale national advisory programmes, a tiered architecture is recommended: lightweight on-device models handle routine queries locally, while complex or high-stakes queries are routed to cloud-hosted larger models for deeper reasoning. Such hybrid architectures balance cost, latency, and advisory quality, making LLM-based advisory economically viable at scale. A key requirement for practical deployment is the ability to operate under low-resource rural environments characterized by intermittent connectivity, limited computational infrastructure, and low-cost mobile devices. Recent advances in model compression techniques, including quantization, pruning, and knowledge distillation, have enabled the deployment of lightweight LLMs on edge devices with reduced memory and energy requirements. Edge-compatible architectures leveraging mobile processors, Edge TPUs, and hybrid cloud-edge frameworks can support near real-time inference while minimizing communication latency and bandwidth dependence. Such approaches are particularly relevant for agricultural advisory systems operating in remote regions where continuous cloud connectivity cannot be guaranteed.

4 Case studies: From theory to field application

This section presents a few case studies on the application of LLMs with geoinformatics in digital agriculture, from India and Kenya, illustrating different agricultural systems (Table 6). Figure 5 presents the chronological evolution of LLM-enabled agricultural advisory systems from 2018 to 2025, tracing the progression from rule-based systems (Era 1) through LLM + RAG integration (Era 2) to multimodal GeoAI platforms (Era 3).

FIGURE 5

4.1 Case study 1: FarmChat - A speech-based conversational advisory in India

4.1.1 The challenge

In several regions, including Jharkhand, smallholder farmers have limited access to extension services. Moreover, text-based interactions are often ineffective due to low literacy levels and poor contextual localisation, which restricts their adoption.

4.1.2 The intervention

FarmChat is one of the early conversational AI systems that aims to provide agronomic advisory to farmers through voice-based interaction in India (Jain et al., 2018). The system was designed to reduce language barriers by integrating local languages. FarmChat was trained on over eight million Kisan Call Centre queries and expert advisory. Using intent and entity recognition, the mobile application delivered crop advisories for potato crop management, allowing interaction in the local language (Hindi).

4.1.3 Impact and significance

Field trials with farmers reported confidence, usability, and utility of voice-based advisory. Although remote sensing data-derived parameters and advisory services were unavailable, it established a scalable advisory paradigm shift to develop modern LLM-based and geospatially enhanced agricultural advisory systems.

4.2 Case study 2: Virtual Agronomist for kenyan maize farmers

4.2.1 The challenge

Rainfed maize systems in Kenya are highly vulnerable to climate change and extreme weather events, especially for smallholder farmers, due to limited access to timely agronomic advice. These constraints are further compounded for women farmers, who often face additional barriers to accessing extension services, digital tools, and locally appropriate communication.

4.2.2 The intervention

The Virtual Agronomist, a conversational-based advisory application, was developed using a RAG framework. The system integrates satellite observations (Sentinel-2 vegetation indices), publicly available weather data (rainfall and soil moisture), and a curated, verified database of maize management practices. This system allows farmers to interact in local languages (Kiswahili) and English, receive recommendations on crop management (planting time, fertiliser use, and pest control), and receive alerts based on remote-sensing-derived stress conditions.

4.2.3 Impact and significance

The application translates complex satellite remote sensing observations into field-level advisories in a short span of time. This application demonstrates the strength of LLMs to provide farm-level agronomic advisory using earth observation and climate data in sub-Saharan Africa. Beyond cereal-based systems, horticultural crops represent an important Frontier for AI-enabled agricultural automation because they require fine-scale monitoring, disease detection, fruit maturity assessment, and precision harvesting. Recent advances combine computer vision, robotic manipulation, and AI-driven decision support to automate harvesting operations in protected cultivation systems. For example, Paul and Machavaram (2025b) developed a vision-guided 6-DOF robotic arm for greenhouse capsicum harvesting using hybrid AI optimization and kinematic modelling, demonstrating the potential of integrating multimodal perception and intelligent control systems for precision horticulture. Such systems highlight future opportunities for coupling GeoAI, computer vision, robotics, and LLM-enabled advisory frameworks within greenhouse environments.

Table 6 synthesises several similar applications deployed for crop advisories in agricultural and horticultural systems in several countries.

The selected case studies collectively represent major agricultural production systems discussed throughout this review, including maize, rice, potato, orchard crops, and mixed farming systems. While the degree of remote sensing integration varies, these examples demonstrate how LLMs and related AI technologies can support crop-specific monitoring, advisory generation, irrigation management, pest control, and decision support across diverse agricultural contexts. Nevertheless, further research is needed to evaluate system performance and advisory effectiveness across a broader range of crop systems, particularly horticultural and high-value crops.

An emerging Frontier in digital agriculture involves the integration of LLMs with agricultural digital twins. Digital twins are virtual representations of agricultural systems that continuously assimilate data from remote sensing platforms, IoT sensors, weather observations, and management records to simulate crop growth and environmental interactions in near real time. Crop simulation models such as APSIM and DSSAT provide the foundation for many digital twin frameworks by enabling scenario analysis of crop performance under alternative management and climate conditions (Jones et al., 2003; Holzworth et al., 2014; Verdouw et al., 2021). In this context, LLMs can serve as conversational interfaces that translate complex simulation outputs into actionable recommendations, support what-if scenario exploration, and facilitate interaction between farmers, agronomists, and simulation environments. Such integration could enable adaptive advisory systems that combine real-time observations, predictive modelling, and natural-language reasoning to improve decision support.

Although these applications illustrate the transformative potential of LLM-assisted agricultural advisory systems, most remain experimental and have not yet achieved widespread operational adoption. Challenges related to data quality, model reliability, infrastructure constraints, regulatory oversight, and farmer trust continue to limit deployment at scale. Therefore, current systems should be viewed as emerging decision-support tools rather than fully autonomous agricultural advisory solutions.

4.3 Cross-cutting synthesis of deployed advisory systems

A comparative analysis of the eleven systems in Table 6 reveals several cross-cutting patterns and persistent gaps. First, a clear maturation trajectory is visible: early systems (FarmChat, Krushi, Agribot, Paddy Mitra) relied on curated knowledge bases and rule-based or retrieval-only architectures with no remote sensing integration, whereas more recent platforms (Farmer.Chat, Krishi Sathi, Virtual Agronomist, Smart Agriculture Assistant) employ generative LLMs, RAG pipelines, and progressively incorporate satellite and UAV-derived data. Second, only two of the eleven systems, the Virtual Agronomist and the multimodal system of Zhao et al. (2024), achieve full integration of remote sensing imagery into the advisory loop; the remainder either exclude geospatial data entirely or incorporate it indirectly through pre-processed documents or structured databases. This underscores that the convergence of LLMs and remote sensing remains nascent. Third, multilingual and voice-first delivery has emerged as a critical success factor: FarmChat, AgriVerseAI, and Farmer.Chat all report that accessibility gains were driven primarily by local-language and voice interfaces rather than by underlying model sophistication. Fourth, evaluation metrics across deployed systems are strikingly heterogeneous, ranging from user satisfaction scores and intent accuracy to precision/recall and farmer trust assessments, with no system yet reporting rigorous agronomic outcome data such as yield improvements or input savings through randomised controlled trials (RCTs). This absence of standardised, outcome-oriented evaluation represents a critical gap that future deployments must address to move beyond usability demonstrations toward evidence-based impact assessment. Although several case studies demonstrate the potential of GeoAI–LLM advisory systems, the degree of actionable specificity varies considerably across applications. Most current systems provide general recommendations, alerts, or explanatory guidance rather than precise field-level prescriptions such as fertiliser dosage, irrigation timing, or pest-control interventions. Generating such specific advisories requires reliable ground-truth observations, calibrated crop and soil parameters, local weather data, crop-stage information, and farmer management histories. Therefore, current LLM-based systems should be viewed mainly as decision-support and interpretive frameworks rather than fully autonomous advisory engines. Future research should focus on validating whether these systems can produce agronomically reliable, location-specific recommendations across multiple seasons and agroecological contexts.

4.4 Agricultural robotics and intelligent automation

Beyond advisory generation, multimodal AI systems are increasingly being integrated into agricultural robotics and intelligent automation platforms. Advances in computer vision, sensor fusion, and AI-driven decision-making have enabled applications such as robotic harvesting, autonomous crop monitoring, greenhouse automation, and precision crop manipulation. In these systems, vision-based perception models identify crop location, maturity status, and harvesting readiness, while AI algorithms support navigation, manipulation, and task optimization. For example, Paul and Machavaram (2025b) developed a vision-guided 6-DOF robotic arm for capsicum harvesting in greenhouse environments using kinematic modelling and hybrid AI-based optimization. Such systems demonstrate how multimodal perception, robotics, and intelligent decision-making can be combined to automate labor-intensive agricultural operations. Future developments may further integrate LLMs as natural-language interfaces and reasoning modules that support robot supervision, task planning, and adaptive decision-making in controlled and open-field environments.

5 Challenges and limitations: The path to responsible deployment

FIGURE 6

Embedding LLMs within human-centric advisory will ensure technological advancement without undermining the reliability and effectiveness of the digital advisory system. Collectively, LLMs should not be viewed as replacements for agronomists or geospatial experts. Instead, the conservative and practical approach is to deploy them as co-pilots processing complex, multi-source data for rapid information collection and advisory. Keeping humans in the loop with localised knowledge for decision-making could be a responsible deployment pathway. Table 7 provides a detailed assessment of the challenges, risk levels, and mitigation strategies.

Despite the rapid advancement of LLMs and multimodal AI systems, most agricultural advisory applications remain at the prototype, pilot, or proof-of-concept stage. While several studies have demonstrated promising results under controlled conditions, evidence of large-scale operational deployment and long-term agronomic impact remains limited. Many reported evaluations rely on small datasets, simulated environments, expert assessments, or short-duration field trials rather than comprehensive multi-season validation across diverse agroecological regions. Consequently, claims regarding scalability, reliability, and economic benefits should be interpreted cautiously until supported by rigorous field-based evaluations and longitudinal studies (Li X. et al., 2024; Singh et al., 2024).

5.1 Limitations of this review

Several limitations of this review should be acknowledged. First, the field of LLM applications in agriculture is evolving rapidly, and systems reported in preprints or grey literature may not have undergone rigorous peer review, potentially affecting the reliability of reported performance metrics. Second, this review focuses primarily on crop advisory applications in rice, maize, wheat, and potato systems; horticultural, livestock, and aquaculture applications are underrepresented and warrant dedicated synthesis. Third, the case studies examined are geographically concentrated in South Asia and East Africa, limiting generalizability to other agroecological contexts such as Latin America, Central Asia, or temperate European farming systems. Fourth, due to the absence of standardised evaluation frameworks across deployed systems, direct quantitative comparison of advisory quality, hallucination rates, or agronomic impact was not possible. Finally, the rapid pace of model development means that some architectures discussed (e.g., GPT-4V, Gemini 1.5) may have been superseded by the time of publication, and readers are encouraged to consult the latest model documentation for current capabilities. Most demonstrated advances in agricultural remote sensing are primarily driven by ML, deep learning, computer vision, and multimodal foundation models for tasks such as crop monitoring, yield prediction, and disease detection. In contrast, the direct use of standalone LLMs for remote sensing analysis remains limited and largely exploratory. Current evidence indicates that LLMs are more effective as complementary reasoning and decision-support components within multimodal or vision-language frameworks rather than as independent geospatial image analysis systems. Therefore, the role of LLMs in agricultural remote sensing should be interpreted cautiously and regarded as an emerging research direction rather than a fully established operational capability (Xiao et al., 2025; Li J. et al., 2024; Zhou et al., 2025).

Although this review includes representative applications across multiple geographic regions and crop systems, the current literature on LLM-assisted agricultural remote sensing remains concentrated in selected regions and crop contexts, particularly smallholder cereal-based systems. Studies involving horticultural crops, plantation systems, and diverse agroecological environments remain comparatively limited. Future research should therefore focus on expanding geographically diverse, crop-specific, and regionally validated implementations to improve the scalability and generalizability of LLM-assisted agro-advisory systems.

6 Future directions: Toward AI-augmented agroecosystems

Early systems such as FarmChat (Jain et al., 2018), Krushi (Momaya et al., 2021), and Paddy Mitra were trained on expert-knowledge and farmer-expert interaction databases to disseminate timely advice with no remote sensing-based indicators (Table 6). While platforms like Agribot (Jain et al., 2025) and AgAsk (Koopman et al., 2022) added remote-sensing-derived inputs via recorded documents, the recent LLMs such as Farmer.Chat (Singh et al., 2024), Krishi Sathi (Vijayvargia et al., 2025), and AgriVerseAI (Birje et al., 2025) demonstrated a leap with added contextualization, adaptive advisory based on weather feeds, geospatial intelligence, and market data. At the forefront, hybrid GeoAI frameworks such as Smart Agriculture Assistant (Jagtap et al., 2025) and the multimodal systems (Zhao et al., 2024), use explicit remote sensing data, IoT sensors, and knowledge graphs as reasoning layers that interpret multi-source complex attributes into farmer-centric grounded advice.

Although existing systems have demonstrated significant advances, most applications rely primarily on static databases, pre-curated content, or IoT sensor feeds (Zhao et al., 2024). The robust utilisation of remote sensing observations, including multimodal crop signals from satellites, UAVs, and thermal and hyperspectral imaging, remains fragmented and largely experimental. Past evaluation predominantly focuses on usability, farmer confidence, and linguistic suitability rather than on yield gains, irrigation efficiency, disease prevention, and climate risk mitigation. This highlights the need for rigorous improvement and assessment to measure advisories verified with notable economic benefits.

As remote sensing data becomes more deeply integrated, RAG-based applications will increasingly incorporate geospatial evidence into their recommendations. Future platforms will replace top-down knowledge transfer with the co-production of “living knowledge,” where verified advisories from agronomists and farmers’ feedback will continuously enrich model performance. Moreover, given the uncertainty in communication and advisories, expert validation should be considered as a foundational system requirement, not an optional safeguard. To provide a structured basis for assessing the effectiveness of these emerging systems, Table 8 presents a proposed evaluation framework for GeoAI-LLM agricultural advisory systems. A persistent challenge in evaluating LLM-assisted agro-advisory systems is the absence of standardised metrics for measuring interpretability and advisory quality. While conventional remote sensing and machine learning systems are typically evaluated using predictive accuracy measures, LLM-based systems require additional assessment dimensions related to explainability, consistency, reliability, and factual correctness. Emerging evaluation frameworks for generative AI increasingly incorporate metrics such as advisory explainability, semantic consistency, hallucination frequency, and domain-specific reliability indicators (Min et al., 2023; Gao et al., 2024; Li X. et al., 2024). In agricultural applications, these metrics are particularly important because advisory outputs directly influence management decisions, resource allocation, and risk mitigation. Future evaluation frameworks should therefore combine technical performance measures with agronomic validity, uncertainty communication, and user comprehension to provide a holistic assessment of advisory quality.

The proposed human-in-the-loop framework positions LLMs as decision-support and interpretive systems rather than autonomous agricultural advisors. In this framework, farmers, extension personnel, agronomists, domain experts, AI developers, and policymakers collectively contribute to validating, contextualising, and refining advisory outputs generated from remote sensing and GeoAI workflows. Domain experts and developers play a critical role in model fine-tuning, calibration, bias mitigation, and region-specific adaptation, while farmers and extension workers provide field-level feedback essential for improving operational reliability and practical relevance. Such iterative human oversight is necessary to reduce hallucination risks, improve trustworthiness, and ensure that advisory recommendations remain agronomically valid and locally actionable.

A major challenge in comparing existing agricultural AI systems is the lack of standardized evaluation criteria. Current studies report heterogeneous metrics, including classification accuracy, precision-recall statistics, user satisfaction scores, expert assessments, and qualitative outcomes, making cross-system comparison difficult. To facilitate consistent evaluation and benchmarking, a normalized assessment framework is proposed that integrates technical performance, advisory quality, user-centered outcomes, and operational deployment considerations (Chen et al., 2024; Min et al., 2023).

6.1 Research frontiers for geospatial LLMs in agriculture

Despite rapid advances, the full potential of GeoAI-based LLMs for agricultural advisory requires progress along five interconnected research frontiers.

6.1.1 Frontier 1 - Advanced geospatial reasoning

Future LLMs must incorporate coordinate awareness, spatial topology, multi-resolution remote sensing data handling, and neighbourhood relationship inference into their architectures. Current models treat spatial data as unstructured text; purpose-built geospatial encoders and spatially aware attention mechanisms would enable LLMs to reason about field boundaries, zone-level variability, and scale-dependent phenomena, moving beyond the limitations identified with current systems (Section 3.2). Promising research directions include coordinate-aware transformers, hierarchical spatial tokenization, geospatial foundation models, and spatial embedding frameworks capable of integrating multi-scale Earth observation data with LLM-based reasoning systems. Future AI ecosystems may combine remote sensing, field robotics, greenhouse automation, and LLM-enabled reasoning to support closed-loop agricultural decision-making and intervention.

Future GeoAI-LLM ecosystems may increasingly incorporate agricultural digital twins that dynamically integrate remote sensing, sensor networks, crop simulation models, and management records. Such systems could support closed-loop decision-making through continuous monitoring, scenario simulation, adaptive recommendation generation, and feedback-driven model updating.

Another promising research direction involves using generative AI to generate synthetic agricultural datasets. Generative AI models can be used to create synthetic imagery, crop disease imagery, agronomic dialogues, field scenarios, and multilingual advisory datasets that complement real-world observations and improve model training, validation, and generalization. Such approaches may help address data scarcity, reduce annotation costs, and enhance the representation of underrepresented crops, management practices, and geographic regions (Paul and Machavaram, 2025a). However, ensuring realism, minimizing bias, and maintaining agronomic validity remain important challenges for operational adoption.

6.1.2 Frontier 2 - Quantitative uncertainty communication

Actionable advisories must explicitly communicate confidence levels and data limitations rather than presenting deterministic prescriptions. This requires integrating Bayesian or ensemble-based uncertainty quantification into the LLM inference pipeline and developing prompt engineering strategies that ensure generated advisories always disclose the provenance and reliability of underlying data (e.g., “Based on Sentinel-2 imagery from 3 days ago with 15% cloud cover in your district”). Future research should prioritize uncertainty-aware GeoAI-LLM systems capable of propagating confidence estimates across sensing, retrieval, reasoning, and recommendation-generation stages to support safer and more transparent agricultural decision-making.

6.1.3 Frontier 3 - Localised smallholder training corpora

Developing high-quality, region-specific datasets for fine-tuning and reinforcement learning is essential, particularly for the Global South where existing training data is sparse. This includes digitising indigenous crop management knowledge, vernacular agronomic terminology, and traditional farming calendars that are currently absent from English-centric training corpora. Also, Future research should focus on tighter integration between Earth observation foundation models and LLM-based reasoning systems through shared embedding spaces, geospatial tokenization schemes, and multimodal retrieval architectures. Such developments could enable seamless translation of satellite-derived representations into context-aware agricultural recommendations while preserving spatial fidelity and uncertainty information.

6.1.4 Frontier 4 - Closed-loop feedback systems

Most advisory systems terminate after generating recommendations. A key research priority is the development of closed feedback loops in which post-recommendation field outcome data (yield, input use, crop health observations) is systematically captured and used to reevaluate and iteratively refine subsequent advice. Such “living knowledge” systems, co-produced by agronomists and farmers, would represent a fundamental shift from static to adaptive advisory. Another promising direction involves integrating LLMs with agricultural digital twins. Crop simulation models such as APSIM and DSSAT, together with greenhouse digital twin platforms, can provide real-time representations of crop growth, environmental conditions, and management interventions. Future advisory systems may use LLMs as conversational interfaces that interact with digital twins, interpret simulation outputs, explore management scenarios, and generate adaptive recommendations based on continuously updated field observations and remote sensing data (Paul et al., 2024).

6.1.5 Frontier 5 - Safety, equity, and human-centred governance

LLM-driven advisories introduce risks of bias, hallucination, and unsafe agronomic recommendations, particularly regarding the use of fertilisers and pesticides. Robust governance frameworks must include mandatory agronomist review for high-stakes recommendations, transparent provenance tracking, fairness audits across gender, geography, and crop diversity, and mechanisms for farmer communities to contest or correct advisory outputs. Human-in-the-loop validation should be designed not as an optional add-on but as a structural component of any deployed system.

7 Conclusion

This review has synthesised the emerging intersection of Large Language Models, remote sensing, and agricultural advisory systems, an area of rapid development but limited prior synthesis. Several key findings emerge from this analysis.

First, the interpretation gap between remote sensing-derived biophysical indicators and actionable farm-level recommendations remains the central bottleneck in digital agriculture (Section 2). While satellite and UAV platforms now deliver unprecedented volumes of spatiotemporal data, translating this information into farmer-centric guidance requires contextual reasoning capabilities that conventional machine learning models and rule-based DSS lack.

Second, LLMs offer a transformative interface layer for bridging this gap. Through domain adaptation pathways (Supervised Fine-Tuning, Retrieval-Augmented Generation, and Reinforcement Learning from Human Feedback), general-purpose LLMs can be calibrated for agricultural specificity while retaining their core strengths of natural language communication, multi-source data synthesis, and multilingual delivery (Section 3). The comparison with traditional DSS (Section 3.3) demonstrates that the optimal deployment model is complementarity rather than replacement, with LLMs serving as the interpretive and communication layer atop quantitative agronomic models.

Third, the analysis of eleven deployed advisory systems (Section 4) reveals a clear maturation trajectory from rule-based chatbots toward RAG-enabled, multimodal GeoAI platforms. However, full integration of remote sensing data into the advisory loop remains achieved in only a minority of systems, and standardised, outcome-oriented evaluation (yield impact, input efficiency) is conspicuously absent across the field.

Fourth, the path to responsible deployment requires addressing interconnected challenges spanning hallucination risk, weak geospatial reasoning, low connectivity, data quality, training data bias, and ethical governance (Section 5). The human-in-the-loop paradigm is not merely a safeguard but a structural requirement for systems operating in high-stakes agricultural decision contexts.

Fifth, five research frontiers (advanced geospatial reasoning, uncertainty communication, localised training corpora, closed-loop feedback, and human-centred governance) define the roadmap for translating current prototypes into operationally impactful systems (Section 6).

In summary, LLM-enabled agronomic services should be conceptualised not as autonomous decision-makers, but as interpretive mediators that translate multi-source data and agronomic principles into locally intelligible, confidence-annotated insights. Integrating LLMs into remote sensing workflows is not intended to replace agronomic and geospatial analytics; rather, it enables their broader, more effective use in real-world decision-making. Their impact is most significant where the gap between data generation and decision-making is widest, particularly for smallholder farmers with limited technical capacity. As GeoAI systems mature, the agricultural community has an opportunity to build advisory ecosystems that are globally informed yet locally grounded, technically sophisticated yet humanly accessible, and data-driven yet ethically governed. A key future research direction involves the development of integrated GeoAI-LLM ecosystems that combine Earth observation foundation models, multimodal transformers, agricultural digital twins, intelligent robotics, and edge-deployable AI systems within a unified decision-support framework. In such architectures, satellite and UAV observations would provide continuous environmental monitoring, digital twins would simulate crop responses under alternative management and climate scenarios, multimodal transformers would fuse geospatial, sensor, and agronomic information, and LLMs would generate context-aware recommendations through natural-language interfaces. Agricultural robots and automated greenhouse systems could further act upon these recommendations, enabling closed-loop monitoring, reasoning, and intervention. Coupled with lightweight edge deployment and multilingual interfaces, such systems have the potential to support climate-resilient, resource-efficient, and farmer-centric precision agriculture.

While recent advances demonstrate substantial potential, the transition from experimental prototypes to operationally robust agricultural advisory systems will require extensive agronomic validation, transparent evaluation frameworks, uncertainty-aware decision support, and sustained engagement with farmers, extension agencies, and policymakers.

References

• 1

  AgnihotriA.MadonT.GadgilA. J. (2022). “Introduction to development engineering,” in Introduction to Development Engineering: A Framework with Applications from the Field (Cham: Springer International Publishing), 1–15.

  • Google Scholar
  • View reference in article

• 2

  AhmadU.AlvinoA.MarinoS. (2021). A review of crop water stress assessment using remote sensing. Remote Sens.13 (20), 4155. 10.3390/rs13204155

  • CrossRef
  • Google Scholar
  • View reference in article

• 3

  AndersonM. C.KustasW. P.NormanJ. M.HainC. R.MecikalskiJ. R.SchultzL.et al (2011). Mapping daily evapotranspiration at field to continental scales using geostationary and polar orbiting satellite imagery. Hydrology Earth System Sciences15 (1), 223–239. 10.5194/hess-15-223-2011

  • CrossRef
  • Google Scholar
  • View reference in article

• 4

  AtzbergerC. (2013). Advances in remote sensing of agriculture: context description, existing operational monitoring systems and major information needs. Remote Sens.5 (2), 949–981. 10.3390/rs5020949

  • CrossRef
  • Google Scholar
  • View reference in article

• 5

  AwaisM.AlharthiA. H. S. A.KumarA.CholakkalH.AnwerR. M. (2025). “Agrogpt: efficient agricultural vision-language model with expert tuning,” in 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) (IEEE), 5687–5696.

  • Google Scholar
  • View reference in article

• 6

  BaiY.JonesA.NdousseK.AskellA.ChenA.DasSarmaN.et al (2022). Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv:2204, 05862.

  • Google Scholar
  • View reference in article

• 7

  BelgiuM.CsillikO. (2018). Sentinel-2 cropland mapping using pixel-based and object-based time-weighted dynamic time warping analysis. Remote Sens. Environ.204, 509–523. 10.1016/j.rse.2017.10.005

  • CrossRef
  • Google Scholar
  • View reference in article

• 8

  BirjeS.NandaniK. B.SupritK. L.NaikS. (2025). AgriVerseAI: smart agricultural web application for Karnataka people. IRJMETS. 10.56726/IRJMETS78194

  • CrossRef
  • Google Scholar
  • View reference in article

• 9

  BommasaniR.HudsonD. A.AdeliE.AltmanR.AroraS.von ArxS.et al (2021). On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258.

  • Google Scholar
  • View reference in article

• 10

  BooriM. S.ChoudharyK.ParingerR.KupriyanovA. (2020). Spatiotemporal ecological vulnerability analysis with statistical correlation based on satellite remote sensing in Samara, Russia. J. Environ. Manag.271, 111000.

  • Google Scholar
  • View reference in article

• 11

  BrownT. B.MannB.RyderN.SubbiahM.KaplanJ.DhariwalP.et al (2020). Language models are few-shot learners. Adv. Neural Inf. Process. Syst.33, 1877–1901.

  • Google Scholar
  • View reference in article

• 12

  ChenJ.YuT.CherneyJ. H.ZhangZ. (2024). Optimal integration of optical and SAR data for improving alfalfa yield and quality traits prediction: new insights into satellite-based forage crop monitoring. Remote Sens.16 (5), 734. 10.3390/rs16050734

  • CrossRef
  • Google Scholar
  • View reference in article

• 13

  ChungH. W.HouL.LongpreS.ZophB.TayY.FedusW.et al (2024). Scaling instruction-finetuned language models. J. Mach. Learn. Res.25 (70), 1–53.

  • Pubmed Abstract
  • Google Scholar
  • View reference in article

• 14

  CorbariC.HuT.PaciollaN.SchlerfM.MallickK.RonellenfitschF. K.et al (2025). High spatial resolution thermal infrared data from airborne acquisition reveal inconsistencies in evapotranspiration and crop water stress retrieval from different models. Remote Sens. Appl. Soc. Environ.38, 101563. 10.1016/j.rsase.2025.101563

  • CrossRef
  • Google Scholar
  • View reference in article

• 15

  DasK.PandeyA. C. (2025). Surface soil moisture retrieval from Sentinel-1 SAR data for semi-arid wheat croplands in India. Geocarto Int.40 (1), 2312456.

  • Google Scholar
  • View reference in article

• 16

  DoddapaneniS.AralikatteR.RameshG.GoyalS.KhapraM. M.KunchukuttanA.et al (2023). “Towards leaving no Indic language behind: building monolingual corpora, benchmark and models for Indic languages,” in Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics, 1, 12402–12426.

  • Google Scholar

• 17

  FengH.TaoH.FanY.LiuY.LiZ.YangG.et al (2022). Comparison of winter wheat yield estimation based on near-surface hyperspectral and UAV hyperspectral remote sensing data. Remote Sens.14 (17), 4158. 10.3390/rs14174158

  • CrossRef
  • Google Scholar
  • View reference in article

• 18

  FritzS.SeeL.MccallumI.YouL.BunA.MoltchanovaE.et al (2015). Mapping global cropland and field size. Glob. Change Biol.21 (5), 1980–1992. 10.1111/gcb.12838

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 19

  FritzS.SeeL.CarlsonT.HaklayM.OliverJ. L.FraislD.et al (2019). Citizen science and the united nations sustainable development goals. Nat. Sustain.2 (10), 922–930. 10.1038/s41893-019-0390-3

  • CrossRef
  • Google Scholar
  • View reference in article

• 20

  GaoY.XiongY.GaoX.JiaK.PanJ.BiY.et al (2024). Retrieval-augmented generation for large language models: a survey. arXiv preprint arXiv:2312.10997.

  • Google Scholar
  • View reference in article

• 21

  GawlikowskiJ.TassiC. R. N.AliM.LeeJ.HumtM.FengJ.et al (2023). A survey of uncertainty in deep neural networks. Artif. Intelligence Review56 (Suppl. 1), 1513–1589. 10.1007/s10462-023-10562-9

  • CrossRef
  • Google Scholar
  • View reference in article

• 22

  GebbersR.AdamchukV. I. (2010). Precision agriculture and food security. Science327 (5967), 825–828. 10.1126/science.1183899

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 23

  GorelickN.HancherM.DixonM.IlyushchenkoS.ThauD.MooreR. (2017). Google Earth engine: planetary-scale geospatial analysis for everyone. Remote Sensing Environ.202, 18–27. 10.1016/j.rse.2017.06.031

  • CrossRef
  • Google Scholar
  • View reference in article

• 24

  GuilpartN.GrassiniP.SadrasV. O.TimsinaJ.CassmanK. G. (2017). Estimating yield gaps at the cropping system level. Field Crops Research206, 21–32. 10.1016/j.fcr.2017.02.008

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 25

  HansenJ. W.VaughanC.KagaboD. M.DinkuT.CarrE. R.KörnerJ.et al (2019). Climate services can support african farmers' context-specific adaptation needs at scale. Front. Sustain. Food Syst.3, 21. 10.3389/fsufs.2019.00021

  • CrossRef
  • Google Scholar
  • View reference in article

• 26

  HolzworthD. P.HuthN. I.deVoilP. G.ZurcherE. J.HerrmannN. I.McLeanG.et al (2014). APSIM–evolution towards a new generation of agricultural systems simulation. Environ. Model. & Softw.62, 327–350. 10.1016/j.envsoft.2014.07.009

  • CrossRef
  • Google Scholar
  • View reference in article

• 27

  HuY.MaiG.CundyC.ChoiK.LaoN.LiuW.et al (2023). Geo-knowledge-guided GPT models improve the extraction of location descriptions from disaster-related social media messages. Int. J. Geogr. Inf. Sci.37 (11), 2289–2318. 10.1080/13658816.2023.2266495

  • CrossRef
  • Google Scholar
  • View reference in article

• 28

  HueteA.DidanK.MiuraT.RodriguezE. P.GaoX.FerreiraL. G. (2002). Overview of the radiometric and biophysical performance of the MODIS vegetation indices. Remote Sensing Environment83 (1-2), 195–213. 10.1016/s0034-4257(02)00096-2

  • CrossRef
  • Google Scholar
  • View reference in article

• 29

  JagtapM.SuryawanshiM.GaneshK. K.YashawantB. S. (2025). Smart agriculture assistant: an AI-powered approach for precision farming and crop management. IJARCCE. 10.17148/IJARCCE.2025.141057

  • CrossRef
  • Google Scholar
  • View reference in article

• 30

  JainM.KumarP.BhansaliI.LiaoQ. V.TruongK. N.PatelS. N. (2018). FarmChat. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies2 (4), 1–22. 10.1145/3287048

  • CrossRef
  • Google Scholar
  • View reference in article

• 31

  JainN.JainP.KayalP.SahitJ.PachpandeS.ChoudhariJ.et al (2025). Agribot: agriculture-specific question answer system. arXiv preprint arXiv:2509.21535.

  • Google Scholar
  • View reference in article

• 32

  JakubikJ.RoyS.PhillipsC. E.FraccaroP.GodwinD.ZadroznyB.et al (2023). Foundation models for generalist geospatial artificial intelligence. arXiv preprint arXiv:2310.18660.

  • Google Scholar
  • View reference in article

• 33

  JiangA. Q.SablayrollesA.MenschA.BamfordC.ChaplotD. S.CasasD.et al (2023). Mistral 7B. arXiv:2310.06825.

  • Google Scholar
  • View reference in article

• 34

  JonesJ. W.HoogenboomG.PorterC. H.BooteK. J.BatchelorW. D.HuntL. A.et al (2003). The DSSAT cropping system model. Eur. Journal Agronomy18 (3-4), 235–265. 10.1016/s1161-0301(02)00107-7

  • CrossRef
  • Google Scholar
  • View reference in article

• 35

  KakwaniD.KunchukuttanA.GollaS.NCG.BhattacharyyaA.KhapraM. M.et al (2020). “IndicNLPSuite: monolingual corpora, evaluation benchmarks and pre-trained multilingual language models for Indian languages,” in Findings of the association for computational linguistics: EMNLP 2020, 4948–4961.

  • Google Scholar

• 36

  KarmakarP.TengS. W.MurshedM.PangS.LiY.LinH. (2024). Crop monitoring by multimodal remote sensing: a review. Remote Sens. Appl. Soc. Environ.33, 101093. 10.1016/j.rsase.2023.101093

  • CrossRef
  • Google Scholar
  • View reference in article

• 37

  KarpukhinV.OğuzB.MinS.LewisP.WuL.EdunovS.et al (2020). Dense passage retrieval for open-domain question answering. arXiv:2004, 04906.

  • Google Scholar
  • View reference in article

• 38

  KilloughB. (2018). Overview of the open data cube initiative. IGARSS, 2018.

  • Google Scholar
  • View reference in article

• 39

  KlerkxL.JakkuE.LabartheP. (2019). A review of social science on digital agriculture, smart farming and agriculture 4.0. Wageningen J. Life Sci.90-91, 100315.

  • Google Scholar
  • View reference in article

• 40

  KoopmanB.MouradA.LiH.van der VegtA.ZhuangS.GibsonS.et al (2022). AgAsk: an agent to help answer farmer's questions from scientific documents. arXiv preprint arXiv:2212.10762.

  • Google Scholar
  • View reference in article

• 41

  KshetriN.DwivediY. K.DavenportT. H.PanteliN. (2024). Generative artificial intelligence in marketing. Int. J. Inf. Manag.75, 102716.

  • Google Scholar
  • View reference in article

• 42

  KuckrejaK.DanishM. S.NaseerM.DasA.KhanS.KhanF. S. (2024). “Geochat: grounded large vision-language model for remote sensing,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 27831–27840.

  • Google Scholar
  • View reference in article

• 43

  KumarS.MishraS.KhannaP. (2022). Satellite remote sensing based maize yield estimation using NDVI and phenological metrics. Int. J. Remote Sens.43 (15-16), 5721–5740.

  • Google Scholar
  • View reference in article

• 44

  LakraK.SharmaA.SinghR. (2024). Machine learning-based surface soil moisture prediction from Sentinel-1 SAR data in wheat croplands. J. Indian Soc. Remote Sens.52 (4), 879–892.

  • Google Scholar
  • View reference in article

• 45

  LewisP.PerezE.PiktusA.PetroniF.KarpukhinV.GoyalN.et al (2021). Retrieval-augmented generation for knowledge-intensive NLP tasks. Adv. Neural Information Processing Systems33, 9459–9474.

  • Google Scholar
  • View reference in article

• 46

  LiJ.LiD.SavareseS.HoiS. (2023). “BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models,” in Proceedings of ICML 2023.

  • Google Scholar
  • View reference in article

• 47

  LiJ.XuM.XiangL.ChenD.ZhuangW.YinX.et al (2024). Foundation models in smart agriculture: basics, opportunities, and challenges. Comput. Electron. Agric.222, 109032. 10.1016/j.compag.2024.109032

  • CrossRef
  • Google Scholar
  • View reference in article

• 48

  LiX.WenC.HuY.YuanZ.ZhuX. X. (2024). Vision-language models in remote sensing: current progress and future trends. IEEE Geoscience Remote Sens. Mag.12 (2), 28–52. 10.1109/mgrs.2024.3383473

  • CrossRef
  • Google Scholar
  • View reference in article

• 49

  LiuH.LiC.WuQ.LeeY. J. (2023). “Visual instruction tuning,” in Advances in Neural Information Processing Systems, 36.

  • Google Scholar
  • View reference in article

• 50

  LiuF.ChenD.GuanZ.ZhouX.ZhuJ.YeQ.et al (2024). Remote CLIP: a vision language foundation model for remote sensing. IEEE Trans. Geoscience Remote Sens.62, 1–16. 10.1109/tgrs.2024.3390838

  • CrossRef
  • Google Scholar
  • View reference in article

• 51

  LuoJ.LuoX.ChenX.XiaoZ.JuW.ZhangM. (2024). “Semi-supervised fine-tuning for large language models,” in Findings of the Association for Computational Linguistics: NAACL 2025, 2795–2808.

  • Google Scholar
  • View reference in article

• 52

  MaC.LiuM.DingF.LiC.CuiY.ChenW.et al (2022). Wheat growth monitoring and yield estimation based on remote sensing data assimilation into the SAFY crop growth model. Sci. Rep.12 (1), 5473. 10.1038/s41598-022-09535-9

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 53

  MaesW. H.SteppeK. (2019). Perspectives for remote sensing with unmanned aerial vehicles in precision agriculture. Trends Plant Sci.24 (2), 152–164. 10.1016/j.tplants.2018.11.007

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 54

  McNairnH.JiaoX. (2024). Integrating SAR and optical time series for crop condition monitoring. Remote Sens. Environ.305, 114085.

  • Google Scholar
  • View reference in article

• 55

  MetaA. I. (2024). Introducing Meta Llama 3: The Most Capable Openly Available LLM to Date. CA, United States: Ai.meta.

  • Google Scholar
  • View reference in article

• 56

  MinS.KrishnaK.LyuX.LewisM.YihW. T.KohP.et al (2023). “Factscore: fine-grained atomic evaluation of factual precision in long form text generation,” in Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, 12076–12100.

  • Google Scholar
  • View reference in article

• 57

  MomayaM.KhannaA.SadavarteJ.SankheM. S. (2021). Krushi: the farmer chatbot. ICCICT, 1–6.

  • Google Scholar
  • View reference in article

• 58

  MullaD. J. (2013). Twenty five years of remote sensing in precision agriculture. Biosyst. Eng.114 (4), 358–371.

  • Google Scholar
  • View reference in article

• 59

  Odisha (2025). PaddyMitra chatbot launched to empower Odisha’s farmers with AI-driven insights. International Rice Research Institute. Available online at: https://www.irri.org/news-and-events/news/paddymitra-chatbot-launched-empower-odishas-farmers-ai-driven-insights.

  • Google Scholar

• 60

  OuyangL.WuJ.JiangX.AlmeidaD.WainwrightC.MishkinP.et al (2022). Training language models to follow instructions with human feedback. Adv. Neural Information Processing Systems35, 27730–27744.

  • Google Scholar
  • View reference in article

• 61

  PaulA.MachavaramR. (2025a). Design, kinematic modeling, and hybrid AI-based optimization of a vision-guided 6-DOF robotic arm for capsicum harvesting in greenhouse environments. Frankl. Open12, 100388. 10.1016/j.fraope.2025.100388

  • CrossRef
  • Google Scholar
  • View reference in article

• 62

  PaulA.MachavaramR. (2025b). Generative AI in agriculture 4.0: applications, challenges, and integration in the Indian context. Food Humanity5, 100889. 10.1016/j.foohum.2025.100889

  • CrossRef
  • Google Scholar
  • View reference in article

• 63

  PaulA.MachavaramR.KumarD.NagarH. (2024). Smart solutions for capsicum harvesting: unleashing the power of YOLO for detection, segmentation, growth stage classification, counting, and real-time mobile identification. Comput. Electron. Agric.219, 108832. 10.1016/j.compag.2024.108832

  • CrossRef
  • Google Scholar
  • View reference in article

• 64

  Pon ArasanA.RadhamaniS.PazhanivelanS.KavithaR.RajaR.KumaraperumalR. (2025). Mapping and monitoring of weeds using unmanned aircraft systems and remote sensing. Plant Prot. Sci.61 (Issue 1), 44–55. 10.17221/74/2024-PPS

  • CrossRef
  • Google Scholar
  • View reference in article

• 65

  RadfordA.KimJ. W.HallacyC.RameshA.GohG.AgarwalS.et al (2021). Learning transferable visual models from natural language supervision. arxiv.org/abs/2103.00020.

  • Google Scholar
  • View reference in article

• 66

  RaffelC.ShazeerN.RobertsA.LeeK.NarangS.MatenaM.et al (2020). Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res.21 (140), 1–67.

  • Pubmed Abstract
  • Google Scholar
  • View reference in article

• 67

  ReichsteinM.Camps-VallsG.StevensB.JungM.DenzlerJ.CarvalhaisN. (2019). Deep learning and process understanding for data-driven Earth system science. Nature566 (7743), 195–204. 10.1038/s41586-019-0912-1

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 68

  RunningS. W.NemaniR. R.HeinschF. A.ZhaoM.ReevesM.HashimotoH. (2004). A continuous satellite-derived measure of global terrestrial primary production. BioScience54 (6), 547–560. 10.1641/0006-3568(2004)054

  • CrossRef
  • Google Scholar

• 69

  ShepherdK. D.MillerM. A.KisituB.MilesB. G.GbedeviK.ChungaP. Z.et al (2025). Virtual Agronomist-an AI-assisted chatbot for guiding crop management decisions of smallholder farmers in Africa. agriRxiv.20250269995.

  • Google Scholar

• 70

  SinghR. K.RizviJ.BeheraM. D.BiradarC. (2021). Automated crop type mapping using time-weighted dynamic time warping-A basis to derive inputs for enhanced food and nutritional security. Curr. Res. Environ. Sustain.3, 100032. 10.1016/j.crsust.2021.100032

  • CrossRef
  • Google Scholar

• 71

  SinghN.Wang'ombeJ.OkangaN.ZelenskaT.RepishtiJ.MishraS.et al (2024). Farmer. Chat: scaling AI-Powered agricultural services for smallholder farmers. arXiv preprint arXiv:2409.08916.

  • Google Scholar

• 72

  SonobeR.YamayaY.TaniH.WangX.KobayashiN.MochizukiK. (2018). Crop classification from Sentinel-2-derived vegetation indices using ensemble learning. J. Appl. Remote Sens.12 (02), 1. 10.1117/1.jrs.12.026019

  • CrossRef
  • Google Scholar

• 73

  StutselB.JohansenK.MalbéteauY. M.McCabeM. F. (2021). Detecting plant stress using thermal and optical imagery from an unoccupied aerial vehicle. Front. Plant Sci.12, 734944. 10.3389/fpls.2021.734944

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 74

  Suaza-MedinaT.RamirezG. M.MarinF. (2024). NDVI and agro-climatic indicators for smallholder maize yield prediction in Colombia. Comput. Electron. Agric.218, 108672.

  • Google Scholar

• 75

  SunZ.SunL.LiuY.LiY.CrusiolL. G. T.ChenR.et al (2024). Estimating fraction of absorbed photosynthetically active radiation of winter wheat based on simulated Sentinel-2 data under different varieties and water stress. Remote Sens.16 (2), 362. 10.3390/rs16020362

  • CrossRef
  • Google Scholar

• 76

  TamásJ.LénártC.BuraiP. (2023). Time-integrated NDVI as a proxy for grain yield: sentinel-2 analysis in Hungarian maize systems. Remote Sens.15 (8), 2117.

  • Google Scholar

• 77

  ThenkabailP.LyonJ. G.TurralH.BiradarC. (2009). Remote Sensing of Global Croplands for Food Security. Oxfordshire, United Kingdom: Taylor and Francis Group.

  • Google Scholar

• 78

  ThenkabailP. S.TeluguntlaP. G.XiongO.CongaltonR. G.OzdoganG.GiriM.et al (2021). PP 1868: Global Cropland-Extent Product at 30-m Resolution (GCEP30) Derived from Landsat Satellite Time-Series Data for the Year 2015 Using Multiple Machine-Learning Algorithms on Google Earth Engine Cloud.

  • Google Scholar

• 79

  ThorntonP. K.EricksenP. J.HerreroM.ChallinorA. J. (2014). Climate variability and vulnerability to climate change: a review. Glob. Change Biol.20 (11), 3313–3328. 10.1111/gcb.12581

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 80

  TorresR.SnoeijP.GeudtnerD.BibbyD.DavidsonM.AttemaE.et al (2012). GMES Sentinel-1 mission. Remote Sens. Environ.120, 9–24. 10.1016/j.rse.2011.05.028

  • CrossRef
  • Google Scholar

• 81

  TouvronH.LavrilT.IzacardG.MartinetX.LachauxM.-A.LacroixT.et al (2023). LLaMA: open and efficient foundation language models. arxiv.org/abs/2302.13971.

  • Google Scholar

• 82

  van HaterenT. C.ChiniM.MatgenP.TeferleF. N. (2023). High-resolution soil moisture retrieval from Sentinel-1 at sub-field scale. Remote Sens. Environ.295, 113664.

  • Google Scholar

• 83

  Van TrichtK.GilliamsS.ZanagaD.BattudeM.GrosuA.BrombacherJ. A.et al (2023). WorldCereal: a dynamic open-source system for global-scale, seasonal, and reproducible crop and irrigation mapping. Earth Syst. Sci. Data15, 5491–5515. 10.5194/essd-15-5491-2023

  • CrossRef
  • Google Scholar

• 84

  VaswaniA.ShazeerN.ParmarN.UszkoreitJ.JonesL.GomezA. N.et al (2017). Attention is all you need. arxiv.org/abs/1706.03762.

  • Google Scholar

• 85

  VerdouwC.TekinerdoganB.BeulensA.WolfertS. (2021). Digital twins in smart farming. Agric. Syst.189, 103046. 10.1016/j.agsy.2020.103046

  • CrossRef
  • Google Scholar

• 86

  VermaP.PaulA.MachavaramR.BhattacharyaM. (2024). Cotton Growth Stages Detection Using fine-tuned YOLOv8 Deep Learning Model.

  • Google Scholar

• 87

  VijayvargiaA.NagpalA.PundalikK.SavarkarA.GautamS.SinghP.et al (2025). Intent aware context retrieval for multi-turn agricultural question answering. arXiv Preprint arXiv:2508.03719.

  • Google Scholar

• 88

  WangJ.ZhangS.LizagaI.ZhangY.GeX.ZhangZ.et al (2024). UAS-based remote sensing for agricultural monitoring: current status and perspectives. Comput. Electron. Agric.227, 109501. 10.1016/j.compag.2024.109501

  • CrossRef
  • Google Scholar

• 89

  WeiJ.WangX.SchuurmansD.BosmaM.XiaF.ChiE.et al (2022). Chain-of-thought prompting elicits reasoning in large language models. Adv. Neural Information Processing Systems35, 24824–24837.

  • Google Scholar

• 90

  WeissM.JacobF.DuveillerG. (2020). Remote sensing for agricultural applications: a meta-review. Remote Sens. Environ.236, 111402. 10.1016/j.rse.2019.111402

  • CrossRef
  • Google Scholar

• 91

  WengX.PangC.XiaG.-S. (2025). Vision-language modeling meets remote sensing. IEEE Geoscience Remote Sens. Mag.13 (3), 276–323.

  • Google Scholar

• 92

  WisemanL.SandersonJ.ZhangA.JakkuE. (2019). Farmers and their data: an examination of farmers’ reluctance to share their data through the lens of the laws impacting smart farming. NJAS Wageningen J. Life Sci.90-91, 100301–100310. 10.1016/j.njas.2019.04.007

  • CrossRef
  • Google Scholar

• 93

  XiaoA.XuanW.WangJ.HuangJ.TaoD.LuS.et al (2025). Foundation models for remote sensing and Earth observation: a survey. IEEE Geoscience Remote Sens. Mag.

  • Google Scholar

• 94

  ZhaoR.LiY.MaM. (2021). Mapping paddy rice with satellite remote sensing: a review. Sustainability13 (2), 503. 10.3390/su13020503

  • CrossRef
  • Google Scholar

• 95

  ZhaoX.ChenB.JiM.WangX.YanY.ZhangJ.et al (2024). Implementation of large language models and agricultural knowledge graphs for efficient plant disease detection. Agriculture14 (8), 1359. 10.3390/agriculture14081359

  • CrossRef
  • Google Scholar

• 96

  ZhouG.LihuangQ.GambaP. (2025). Advances on multimodal remote sensing foundation models for Earth observation downstream tasks: a survey. Remote Sens.17 (21), 3532. 10.3390/rs17213532

  • CrossRef
  • Google Scholar