Agricultural intelligence and LLMs

Li, H., Wu, H., Li, Q., & Zhao, C. (2025). A review on enhancing agricultural intelligence with large language models. Artificial Intelligence in Agriculture, 15(4), 671-685. https://doi.org/10.1016/j.aiia.2025.05.006

Study Type: Literature review

Geographical Context: China-focused with global perspective

Agri-Food Subsector: Cross-sectoral (crop cultivation, animal husbandry, food processing, supply chain management)

Implementation Scale: Conceptual framework spanning laboratory to full-scale applications

GenAI Characterisation

GenAI Technology: Large Language Models (primarily Transformer-based architectures including encoder-only, encoder-decoder, and decoder-only configurations)

Specific Models: GPT series, BERT, XLNet, Claude, specialized agricultural models (AgLLMs, Shennong, SkyEyeGPT)

Industrial Applications:

• Agricultural knowledge question-answering systems

• Crop management and optimization

• Market analysis and decision-making

• Supply chain and production management

• Pest and disease identification

• Multimodal agricultural data processing

Systems Integration: Multiple integration approaches including vector databases, knowledge graphs, RAG (Retrieval-Augmented Generation), and multimodal frameworks

Key Technical Components

Architecture Design: The paper systematically categorizes LLM architectures into three types: encoder-only (for understanding tasks), encoder-decoder (for translation/summarisation), and decoder-only (for text generation). Each architecture serves different agricultural applications.

Pre-training Strategies: Four main approaches identified - Masked Language Modeling (MLM), Next Sentence Prediction (NSP), Causal Language Modeling (CLM), and Permutation Language Modeling (PLM), each with distinct advantages for agricultural knowledge processing.

Fine-tuning Methods:

• Full parameter fine-tuning for comprehensive agricultural knowledge integration

• Parameter-Efficient Fine-Tuning (PEFT) including LoRA, adapters, and prompt tuning

• Prompt engineering for lightweight adaptation without parameter modification4

Agricultural Knowledge Integration

Vector Databases: Implementation of high-dimensional vector storage for agricultural text, images, and multimodal data using technologies like Milvus and FAISS for semantic similarity retrieval.

Knowledge Graphs: Construction of structured agricultural knowledge representations using Neo4j and ArangoDB, enabling entity recognition, relationship extraction, and graph-based reasoning for agricultural Q&A systems.

Multimodal Learning: Integration of text, images, videos, audio, and sensor data using models like CLIP and BLIP-2 for comprehensive agricultural decision-making frameworks.

Practical Applications

Agricultural Question-Answering Systems: Development of hybrid retrieval-based and generation-based systems combining vector databases with knowledge graphs for accurate, contextually-relevant responses to agricultural queries.

Specialized Agricultural LLMs: Examples include AgroNT for crop genetics, FoodS for cuisine applications, and various Chinese agricultural models (Houji, Tiangong Kaifeng, Shennong series) with domain-specific training.

Multimodal Agricultural Intelligence: Applications spanning crop identification, pest detection, yield estimation, livestock monitoring, and precision agriculture through vision-language models.

Technical Challenges and Solutions

Knowledge Integration: Addressing fragmented agricultural knowledge through unified representation frameworks and causal inference enhancement modules.

Multimodal Data Fusion: Developing denoising algorithms and attention mechanisms for cross-modal semantic alignment in noisy agricultural environments.

Agent Communication: Establishing unified standards and conflict resolution mechanisms for multi-agent agricultural systems.

Dynamic Knowledge Updates: Implementing parameter-efficient updating techniques to maintain current agricultural knowledge without catastrophic forgetting.

Future Directions

Crop Management: Precision agriculture applications including health management, planting optimization, and resource allocation through integrated multi-source data analysis.

Market Analysis: Real-time processing of diverse agricultural market data for trend prediction and decision support.

Supply Chain Optimization: IoT integration for automated greenhouse management, inventory optimization, and supplier relationship management.

Research Infrastructure: Development of specialized transformer architectures, domain-specific pre-training tasks, and standardized agricultural AI benchmarks.

Limitations and Considerations

The review acknowledges several critical challenges: limited agricultural domain training data, complex causal relationships in agricultural systems, multimodal data processing difficulties, and the need for human oversight in AI-driven agricultural decisions. The authors emphasise the importance of human-machine collaboration rather than full automation.