Advancing site-specific disease and pest management in precision agriculture: From reasoning-driven foundation models to adaptive, feedback-based learning
- 1Department of Horticultural Sciences, Gulf Coast Research and Education Center (GCREC), University of Florida, Wimauma, FL, USA
- 2Department of Soil, Water, and Ecosystem Sciences, Citrus Research and Education Center (CREC), University of Florida, Lake Alfred, FL, USA
Abstract
Site-specific disease management (SSDM) in agricultural crops has witnessed tremendous advancements over the past few decades using conventional machine and deep learning (ML & DL) approaches for real-time computer vision applications. This research has evolved from handcrafted feature extraction to large-scale automated feature learning in single-modality datasets. However, with the rise of foundation models (FMs), the way large-scale crop disease datasets are processed is being fundamentally transformed. Unlike convolutional neural networks (CNNs), which are inherently single-modality architectures and require post-fusion strategies to integrate heterogenous data streams, vision-language FMs provide unified embedding spaces where visual and textual representations are jointly learned from the same training dataset. This enables richer, more coherent cross-modal reasoning, allowing models to interpret disease symptoms described in text, infer relationships between symptoms and management factors, and support interactive Q&A platforms for growers and extension educators. Beyond visual and textual reasoning, FMs are also serving as backbone architectures for robotic perception and decision-making, bridging the gap between scene understanding and physical action in field environments. In this context, the integration of adaptive, imitation, and reinforcement learning (AL, IL, & RL) in robotics is enabling novel applications for field-based disease management. Therefore, this study reviewed ≈40 research articles to highlight the application of FMs for SSDM in crops, focusing on two primary themes: large-language models (LLMs), and vision language models (VLMs). Additionally, the extended role of FMs in enabling AL, IL, RL, and digital twin (DT) frameworks for robotics-based targeted spraying is also discussed. Based on the results and discussion, several key conclusions emerge from this review: (a) FMs are gaining traction, with a notable increase in reported technical literature during 2023-2024, (b) VLMs are being leveraged more than LLMs, with a five-to-tenfold increase in published articles from 2023 to 2024, (c) RL and AL remain in their infancy for developing smart spraying technologies that learn from experience, (d) integrating DTs with RL in cyber-physical systems offer a transformative approach for simulating targeted spraying in virtual environments, (e) addressing sim-to-real gap, the performance drop when models trained in simulated environments are deployed in real-world disease scenarios, will be critical for robust and scalable management systems, (f) while perception models for disease detection are advancing, human-robot collaboration and human-in-the-loop (HiTL) approaches remain underexplored in field-scale crop disease management systems where robots autonomously detect early symptoms and humans validate uncertain cases, and (g) continued advancements in FMs. Multi-modal integration, and real-time feedback are expected to drive the next generation of SSDM technologies.
Related
- Cell-Lens: Advancing LLM Reasoning at Single-Cell Resolution
- Hybrid local–global vision transformer framework for robust plant disease detection in precision agriculture
- A global–local momentum contrastive framework with critical feature calibration for field plant disease identification in precision agriculture
- Advancing sustainable agriculture through multi-parameter fuzzy soft set–based plant disease classification
- Amino Acids as Multifunctional Molecules in Plants: From Fundamental Metabolism to Precision Agriculture
BibTeX
@article{rai2026advancing,
title = {{Advancing site-specific disease and pest management in precision agriculture: From reasoning-driven foundation models to adaptive, feedback-based learning}},
author = {Nitin Rai and Daeun Dana Choi and Nathan S. Boyd and Arnold W. Schumann},
journal = {Computers and Electronics in Agriculture},
year = {2026},
month = may,
doi = {10.1016/j.compag.2026.111859},
url = {https://doi.org/10.1016/j.compag.2026.111859},
note = {CC BY-NC-ND},
}
RIS
TY - JOUR
TI - Advancing site-specific disease and pest management in precision agriculture: From reasoning-driven foundation models to adaptive, feedback-based learning
AU - Nitin Rai
AU - Daeun Dana Choi
AU - Nathan S. Boyd
AU - Arnold W. Schumann
PY - 2026
DA - 2026/05/16
JO - Computers and Electronics in Agriculture
DO - 10.1016/j.compag.2026.111859
UR - https://doi.org/10.1016/j.compag.2026.111859
AB - Site-specific disease management (SSDM) in agricultural crops has witnessed tremendous advancements over the past few decades using conventional machine and deep learning (ML & DL) approaches for real-time computer vision applications. This research has evolved from handcrafted feature extraction to large-scale automated feature learning in single-modality datasets. However, with the rise of foundation models (FMs), the way large-scale crop disease datasets are processed is being fundamentally transformed. Unlike convolutional neural networks (CNNs), which are inherently single-modality architectures and require post-fusion strategies to integrate heterogenous data streams, vision-language FMs provide unified embedding spaces where visual and textual representations are jointly learned from the same training dataset. This enables richer, more coherent cross-modal reasoning, allowing models to interpret disease symptoms described in text, infer relationships between symptoms and management factors, and support interactive Q&A platforms for growers and extension educators. Beyond visual and textual reasoning, FMs are also serving as backbone architectures for robotic perception and decision-making, bridging the gap between scene understanding and physical action in field environments. In this context, the integration of adaptive, imitation, and reinforcement learning (AL, IL, & RL) in robotics is enabling novel applications for field-based disease management. Therefore, this study reviewed ≈40 research articles to highlight the application of FMs for SSDM in crops, focusing on two primary themes: large-language models (LLMs), and vision language models (VLMs). Additionally, the extended role of FMs in enabling AL, IL, RL, and digital twin (DT) frameworks for robotics-based targeted spraying is also discussed. Based on the results and discussion, several key conclusions emerge from this review: (a) FMs are gaining traction, with a notable increase in reported technical literature during 2023-2024, (b) VLMs are being leveraged more than LLMs, with a five-to-tenfold increase in published articles from 2023 to 2024, (c) RL and AL remain in their infancy for developing smart spraying technologies that learn from experience, (d) integrating DTs with RL in cyber-physical systems offer a transformative approach for simulating targeted spraying in virtual environments, (e) addressing sim-to-real gap, the performance drop when models trained in simulated environments are deployed in real-world disease scenarios, will be critical for robust and scalable management systems, (f) while perception models for disease detection are advancing, human-robot collaboration and human-in-the-loop (HiTL) approaches remain underexplored in field-scale crop disease management systems where robots autonomously detect early symptoms and humans validate uncertain cases, and (g) continued advancements in FMs. Multi-modal integration, and real-time feedback are expected to drive the next generation of SSDM technologies.
SN - 0168-1699
SN - 1872-7107
ER -