Objective Traditional agricultural production systems exhibit evident limitations in production efficiency, resource utilization and environmental sustainability, and are undergoing a gradual transition toward informatization and intelligent modernization. As a core enabling technology of smart agriculture, agricultural vision plays a crucial role in key applications such as crop production monitoring and livestock and poultry breeding management, and has significant practical value for improving agricultural productivity. However, most existing vision models rely on large-scale labeled datasets. In agricultural scenarios, complex environments and highly variable data acquisition conditions lead to high annotation costs and long data preparation cycles, which substantially limit large-scale deployment.

Method This paper focuses on three representative learning paradigms with limited annotations in agricultural vision, namely semi-supervised, weakly supervised and self-supervised learning. Their fundamental principles and commonly adopted frameworks are reviewed.

Result The performance characteristics and applicability of these three learning paradigms are summarized in the context of typical agricultural vision tasks. Key challenges, including limited cross-domain generalization and interference from noisy annotations, are further analyzed.

Conclusion Future research directions are discussed, such as the construction of datasets and evaluation benchmarks, the development of agriculture-specific pre-trained models, and active learning–driven low-cost iterative strategies, providing references for the advancement and application of agricultural vision technologies.