The landscape of computer vision datasets in smart agriculture: trends, taxonomies, challenges, and future directions

Abstract Modern agriculture is increasingly driven by data-centric and artificial intelligence (AI)-enabled approaches to improve productivity, resource efficiency, and precision in decision-making. In this context, computer vision datasets have emerged as critical resources for enabling automated agricultural analysis. This study presents a large-scale, data-centric review of agricultural computer vision datasets using a PRISMA-guided systematic review methodology. A systematic search of the peer-reviewed data journals Scientific Data and Data in Brief, conducted through 31 January 2026, identified 2,308 records, from which 253 eligible dataset publications were retained after duplicate removal and eligibility screening based on predefined inclusion and exclusion criteria. We systematically analyze the dataset landscape across multiple dimensions, including vision tasks, annotation strategies, acquisition environments, sensing modalities, and geographic distribution. Furthermore, we propose a unified multi-dimensional taxonomy that organizes datasets based on task orientation, semantic content, and design characteristics, enabling structured comparison and informed dataset selection. Our analysis identifies key limitations, including sampling bias, underrepresentation of crops and regions, inconsistent annotation practices, and limited interoperability, which collectively hinder reproducibility and model generalization. Additionally, challenges in dataset accessibility and the absence of standardized benchmarking frameworks restrict effective knowledge transfer. Finally, we outline critical directions for future research, emphasizing standardized dataset design, scalable annotation frameworks, and robust evaluation protocols to support the development of reliable and generalizable agricultural AI systems.