Accurate identification and monitoring of sunflower capitula are crucial for field phenotypic analysis, cultivation management, phenological monitoring, and yield prediction. Manual observation, however, faces significant challenges due to the complexity of field environments and the morphological diversity of sunflower capitula. Unmanned Aerial Vehicles (UAVs) have emerged as an ideal platform for monitoring sunflower capitula due to their low cost and high spatiotemporal resolution. This study introduces Sunflower-YOLO, an enhanced model based on YOLOv7-tiny, designed for detecting sunflower capitula in UAV remote sensing images. The model effectively identifies sunflower capitula and distinguishes between three specific states: open, half-open, and bud. Sunflower-YOLO incorporates several key improvements: the SiLU activation function replaces the original LeakyReLU, enhancing the model’s nonlinear expression capability; a shallow high-resolution feature map and an additional detection head for small targets are introduced during the feature fusion stage to improve the detection performance of small capitula; and the integration of deformable convolution and the SimAM attention mechanism enhances the ELAN structure in the backbone, creating a new DeformAtt-ELAN structure that improves the model’s ability to capture morphological variations and reduces noise interference. Experimental results demonstrate that Sunflower-YOLO achieves precision, recall, and mAP@0.5 of 92.3 %, 89.7 %, and 93 %, respectively, marking improvements of 4.2 %, 4.2 %, and 3.7 % over the original YOLOv7-tiny model. The average precision (AP) for the three growth states is 98.7 %, 93.4 %, and 87 %, with AP for the half-open and bud states improving by 6.5 % and 4.7 %, respectively. The model’s FLOPs is 17.7 G, its size is 13.8MB, and it achieves an FPS of 188.52. Compared to current mainstream state-of-the-art (SOTA) models for object detection, Sunflower-YOLO achieves the highest mAP@0.5 in detecting multiple types of sunflower capitula. The constructed capitulum density map offers a practical view for observing sunflower growth status. This study highlights the immense potential of combining UAV remote sensing technology with YOLO object detection algorithms in monitoring sunflower capitula and their growth processes, providing an innovative and effective approach for precision agriculture practices.

中文翻译：

---

Sunflower-YOLO：无人机遥感图像中向日葵头状花序的检测


向日葵头状花序的准确识别和监测对于田间表型分析、栽培管理、物候监测和产量预测至关重要。然而，由于田间环境的复杂性和向日葵头状花序形态的多样性，人工观察面临着巨大的挑战。无人机（UAV）因其低成本和高时空分辨率而成为监测向日葵头状花序的理想平台。本研究介绍了Sunflower-YOLO，这是一种基于YOLOv7-tiny的增强模型，旨在检测无人机遥感图像中的向日葵头状花序。该模型有效识别向日葵头状花序，并区分三种特定状态：开放、半开放和花蕾。 Sunflower-YOLO融合了几个关键改进：SiLU激活函数替代了原来的LeakyReLU，增强了模型的非线性表达能力；在特征融合阶段引入浅层高分辨率特征图和针对小目标的附加检测头，以提高小头状体的检测性能；可变形卷积和SimAM注意力机制的集成增强了主干中的ELAN结构，创建了新的DeformAtt-ELAN结构，提高了模型捕获形态变化的能力并减少噪声干扰。实验结果表明，Sunflower-YOLO 的精度、召回率和 mAP@0.5 分别达到 92.3%、89.7% 和 93%，比原始 YOLOv7-tiny 模型提高了 4.2%、4.2% 和 3.7%。三种生长状态的平均精度 (AP) 分别为 98.7%、93.4% 和 87%，其中半开状态和芽状态的 AP 分别提高了 6.5% 和 4.7%。 该模型的 FLOP 为 17.7 G，大小为 13.8MB，FPS 为 188.52。与当前主流的物体检测最先进（SOTA）模型相比，Sunflower-YOLO 在检测多种类型的向日葵头状花序方面实现了最高的 mAP@0.5。构建的头状花序密度图为观察向日葵生长状况提供了实用的视图。这项研究凸显了无人机遥感技术与 YOLO 目标检测算法相结合在监测向日葵头状花序及其生长过程中的巨大潜力，为精准农业实践提供了一种创新且有效的方法。