Harnessing large vision and language models in agriculture: a review

摘要

Introduction: Agriculture is a cornerstone of human society but faces significant challenges, including pests, diseases, and the need for increased production efficiency. Large models, encompassing large language models, large vision models, and multimodal large language models, have shown transformative potential in various domains. This review aims to explore the potential applications of these models in agriculture to address existing problems and improve production. Methods: We conduct a systematic review of the development trajectories and key capabilities of large models. A bibliometric analysis of literature from Web of Science and arXiv is performed to quantify the current research focus and identify the gap between the potential and the application of large models in the agricultural sector. Results: Our analysis confirms that agriculture is an emerging but currently underrepresented field for large model research. Nevertheless, we identify and categorize promising applications, including tailored models for agricultural question-answering, robotic automation, and advanced image analysis from remote sensing and spectral data. These applications demonstrate significant potential to solve complex, nuanced agricultural tasks. Discussion: This review culminates in a pragmatic framework to guide the choice between large and traditional models, balancing data availability against deployment constraints. We also highlight critical challenges, including data acquisition, infrastructure barriers, and the significant ethical considerations for responsible deployment. We conclude that while tailored large models are poised to greatly enhance agricultural efficiency and yield, realizing this future requires a concerted effort to overcome the existing technical, infrastructural, and ethical hurdles.