Answers are grounded in this paper and cite their sources.
• Proposed an enhanced YOLOv5s-BiFormer model that outperforms mainstream lightweight detectors (YOLOv7-tiny, YOLOv8s, YOLOv12s, YOLOv13s, SSD) in classifying citrus SSWF (spring shoots with flowers) and SSWOF (spring shoots without flowers) under complex field conditions; it features a compact parameter size (7.3 M), moderate computational cost (26.1 G Flops), and stable real-time speed (104 FPS), fully meeting the deployment requirements of resource-constrained devices. • Identified the optimal phenological window for precise spring shoot recognition (14 to 4 days before full bloom) via 10 temporal grouping schemes and segmented regression analysis, achieving >90% precision and >75% recall to streamline monitoring efficiency. • Developed a real-time Android application based on Flask and cloud architecture, enabling spring shoot recognition and quantification within 5–8 seconds across device tiers, bridging model development and on-site deployment. • Established a cross-regionally robust monitoring system (SSWF mAP: 95.0%–99.3%; SSWOF mAP: 91.0%–99.1% across six citrus-producing regions), providing data-driven guidance for refined orchard management. Orchard production management is closely tied to key phenological stages of tree development, where accurate monitoring of emerging organs (e.g., citrus spring shoots) is critical for agricultural decision-making and production efficiency. However, real-time precise detection faces challenges from dynamic morphological variations, subtle feature differences (e.g., flower bud presence/absence), and complex field environments (illumination, occlusion, orientation). To tackle this, we propose an enhanced YOLOv5s-BiFormer object detection model by integrating YOLOv5s with the BiFormer attention mechanism, leveraging its dual strengths of CNN-based local feature sensitivity and Transformer-based global context modeling to distinguish spring shoots with (SSWF) and without flowers (SSWOF) under natural conditions. To optimize monitoring efficiency, we systematically evaluated the model’s recognition performance across seven growth stages prior to severe shoot occlusion via 10 temporal grouping schemes and segmented regression analysis. Results revealed a “first-increase-then-decrease” trend in detection accuracy as spring shoots developed, identifying the optimal phenological window for precise image recognition: 14 to 4 days before full bloom (-14d to -4d). Experimental validation showed YOLOv5s-BiFormer outperforms not only the baseline YOLOv5s but also state-of-the-art lightweight models (YOLOv7-tiny, YOLOv13s, SSD) in key metrics: SSWF recall improved by 0.9% (82.8%), SSWOF precision increased by 1.3% (88.8%), with mAP reaching 89.6% (SSWF) and 84.0% (SSWOF) — 17.1% and 19.8% higher than YOLOv7-tiny, 4.9% and 9.6% higher than YOLOv8s, 9.3% and 13.7% higher than YOLOv12s, 4.3% and 10.3% higher than YOLOv13s, and 55.2% and 63.6% higher than SSD, respectively. With a compact parameter size of 7.3 M (only 4.3% larger than YOLOv5s), moderate computational cost (26.1 G Flops, lower than YOLOv8s and SSD), and stable real-time inference speed (104 FPS), the model strikes a balance between performance and efficiency, fully meeting practical deployment needs on resource-constrained devices. A cross-regional model trained on China’s major citrus datasets exhibits robust generalization (precision>90%, recall>85%, mAP50>90%). An Android app, based on Flask and cloud architecture, enables 5-8s real-time identification/quantification of spring shoot development, integrating with flowering prediction tools to form a comprehensive orchard monitoring system. This work provides innovative digital sensing methods for non-destructive citrus spring shoot monitoring, facilitating refined orchard management and data-driven decisions.
No takes yet. Share an insight, caveat, or question.
Liu et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: