Bridging lab and field: A review and roadmap for unmanned aerial vehicle-based field crop counting with deep learning

• Comprehensive review of UAV-based deep learning for large-scale field crop counting. • Framework links agronomic targets to detection, segmentation, or regression. • Key bottlenecks: semantic gaps, temporal decay, and scaling to field reliability. • Economic constraints and integration with farm workflows limit deployment. • Roadmap outlines four frontiers for robust, scalable, decision-oriented adoption. Field crop counting is crucial for precision agriculture and global food security. The integration of Unmanned Aerial Vehicles (UAVs) and deep learning has become the leading technology for this task. However, existing reviews often overlook the unique challenges of large-scale field crops, such as vast areas and high planting density. This paper provides a comprehensive review of UAV-based deep learning methods for in-field object counting. We analyze 52 key studies and introduce a framework to guide method selection. This framework connects agronomic targets to detection, segmentation, or regression approaches based on their specific characteristics. Our review also examines the entire data workflow, covering data acquisition, annotation strategies, and eight public datasets. We also link data choices directly to the applicability of different models. Finally, we identify the key gaps preventing successful real-world deployment and propose a future research roadmap. This roadmap focuses on four strategic frontiers: large foundational models, unified open-set counting paradigms, dynamic spatiotemporal phenotyping, and autonomous edge-cloud systems. This review aims to support the advancement of intelligent and sustainable agriculture by bridging the gap between research and practical application.