PMC-PhaseNet: ODConv-EMA enhanced YOLOv11n with focal loss for fine-grained phenological staging of Pinus tabuliformis male pine cones

Read full text Summarise with VIRAT
Authors: Changhao Zhou, Xiaodong Li, You Fu, Wentao Chen, Yan Cheng, Wenliang Wu
Cited by: 0
References: 33

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Accurate, automated identification of phenological stages in Pinus tabuliformis male pine cones is critical for intelligent harvesting systems, yet fine-grained classification in natural field conditions remains challenging due to small target sizes, scale variation, background occlusion, and high morphological similarity within the same phenological stage. To tackle these application-specific challenges, we enhance YOLOv11n by integrating Omni-dimensional Dynamic Convolution (ODConv) into the backbone to suppress background interference, incorporating an Efficient Multi-scale Attention (EMA) module into the neck network to capture multi-scale contextual information, and applying a focal loss function to focus training on hard-to-classify samples, resulting in the proposed PMC-PhaseNet. This configuration aims to provide a practical solution for multi-scale variation, background occlusion, and hard sample learning in similar fine-grained phenological stage classification tasks. To support this research, we curate a specialized dataset, PMC-Phase, comprising images meticulously annotated with three distinct phenological stages (elongation stage, ripening stage, decline stage). Experiment results demonstrate that PMC-PhaseNet achieves improvement compared to the baseline model YOLOv11n and other mainstream detectors. Compared to the more recent YOLOv12n, PMC-PhaseNet shows improved metrics (e.g., +1.32% in mAP 50) when evaluated on this specific fine-grained phenological staging task. It achieves a precision of 91.88%, a recall of 83.32%, mAP 50 of 88.61%, and mAP 50-95 of 70.72%. Notably, the resulting PMC-PhaseNet model features 2.61M parameters and 5.9 GFLOPs, achieving these improvements while balancing lightweight compactness and performance. These findings not only provide a robust and efficient vision solution for the automated phenology monitoring of Pinus tabuliformis male pine cones but also establish a solid foundation for future on-device deployment, highlighting its considerable practical potential.

Fields of Study & Subjects

Smart Agriculture and AI Advanced Neural Network Applications Industrial Vision Systems and Defect Detection

Publication details

Journal / source
Smart Agricultural Technology
Volume
14
Pages
102053–102053
Publication date
28 March 2026
Language
EN