LightTassel-YOLO：一种基于无人机遥感的玉米雄穗实时检测方法

doi:10.12133/j.smartag.SA202505021

摘要/Abstract

摘要：

［目的/意义］ 玉米雄穗的精准识别是制种生产的关键环节。针对现有目标检测模型在复杂大田场景下的研究存在数据维度受限、特征提取不足、计算负载较高、检测效率低下等问题，本研究提出一种基于改进YOLOv11n的大田玉米雄穗实时检测模型LightTassel-YOLO，旨在快速、准确地识别玉米雄穗，以实现去雄无人机的高效作业，减少人工干预的影响。 ［方法］ 利用无人机连续获取2023—2024年玉米抽雄期数据，构建了覆盖玉米抽雄不同阶段、多品种、多高度及多气象条件的大规模高质量玉米雄穗数据集。首先，将EfficientViT应用于主干网中，以增强在多尺度特征中感知信息的能力；其次，设计C2PSA-CPCA模块通过为特征图动态分配通道和空间维度的注意力权重，有效增强网络对目标特征提取能力的同时降低了计算复杂度；最后构建C3k2-SCConv模块，促进代表性特征学习的同时达成低成本空间特征重构，提高模型检测准确率。［结果与讨论］ LightTassel-YOLO为玉米雄穗检测提供了一种可靠方法，最终模型的准确率为92.6%，召回率为89.1%，AP@0.5为94.7%，较基准模型YOLOv11n分别提升2.5、3.8、4.0个百分点，参数量仅为3.23 M，计算量为6.7 GFLOPs。此外，LightTassel-YOLO还与目前主流的目标检测算法Faster R-CNN，SSD和YOLO系列的多个版本进行对比，验证本研究提出方法在综合性能上均优于上述算法，在典型田间场景中，模型亦展现出优异适应性。 ［结论］ 本研究所提出的方法为玉米雄穗精准监测提供了有效的理论基础，对提升田间管理智能化水平具有重要意义。

关键词: 玉米雄穗检测, YOLOv11, EfficientViT, CPCA, SCConv, 无人机

Abstract:

[Objective] The accurate identification of maize tassels is critical for the production of hybrid seed. Existing object detection models in complex farmland scenarios face limitations such as restricted data diversity, insufficient feature extraction, high computational load, and low detection efficiency. To address these challenges, a real-time field maize tassel detection model, LightTassel-YOLO (You Only Look Once) based on an improved YOLOv11n is proposed. The model is designed to quickly and accurately identify maize tassels, enabling efficient operation of detasseling unmanned aerial vehicles (UAVs) and reducing the impact of manual intervention. [Methods] Data was continuously collected during the tasseling stage of maize from 2023 to 2024 using UAVs, establishing a large-scale, high-quality maize tassel dataset that covered different maize tasseling stages, multiple varieties, varying altitudes, and diverse meteorological conditions. First, EfficientViT (Efficient vision transformer) was applied as the backbone network to enhance the ability to perceive information across multi-scale features. Second, the C2PSA-CPCA (Convolutional block with parallel spatial attention with channel prior convolutional attention) module was designed to dynamically assign attention weights to the channel and spatial dimensions of feature maps, effectively enhancing the network's capability to extract target features while reducing computational complexity. Finally, the C3k2-SCConv module was constructed to facilitate representative feature learning and achieve low-cost spatial feature reconstruction, thereby improving the model's detection accuracy. [Results and Discussions] The results demonstrated that LightTassel-YOLO provided a reliable method for maize tassel detection. The final model achieved an accuracy of 92.6%, a recall of 89.1%, and an AP@0.5 of 94.7%, representing improvements of 2.5, 3.8 and 4.0 percentage points over the baseline model YOLOv11n, respectively. The model had only 3.23 M parameters and a computational cost of 6.7 GFLOPs. In addition, LightTassel-YOLO was compared with mainstream object detection algorithms such as Faster R-CNN, SSD, and multiple versions of the YOLO series. The results demonstrated that the proposed method outperformed these algorithms in overall performance and exhibits excellent adaptability in typical field scenarios. [Conclusions] The proposed method provides an effective theoretical framework for precise maize tassel monitoring and holds significant potential for advancing intelligent field management practices.

Key words: maize tassel detection, YOLOv11, EfficientViT, CPCA, SCConv, UAV

中图分类号:

S513

曹玉莹, 刘银川, 高新悦, 贾银江, 董守田. LightTassel-YOLO：一种基于无人机遥感的玉米雄穗实时检测方法[J]. 智慧农业(中英文), 2025, 7(6): 96-110.

CAO Yuying, LIU Yinchuan, GAO Xinyue, JIA Yinjiang, DONG Shoutian. LightTassel-YOLO: A Real-Time Detection Method for Maize Tassels Based on UAV Remote Sensing[J]. Smart Agriculture, 2025, 7(6): 96-110.

图/表 16

参考文献 31

[1]	WANG X W, LI X Y, LOU Y S, et al. Refined evaluation of climate suitability of maize at various growth stages in major maize-producing areas in the north of China[J]. Agronomy, 2024, 14(2): ID 344.
[2]	YANG J X, ZHANG R R, DING C C, et al. YOLO-detassel: Efficient object detection for Omitted Pre-Tassel in detasseling operation for maize seed production[J]. Computers and electronics in agriculture, 2025, 231: ID 109951.
[3]	THOMPSON A L, THORP K R, CONLEY M M, et al. Comparing nadir and multi-angle view sensor technologies for measuring in-field plant height of upland cotton[J]. Remote sensing, 2019, 11(6): ID 700.
[4]	CHIVASA W, MUTANGA O, BIRADAR C. UAV-based multispectral phenotyping for disease resistance to accelerate crop improvement under changing climate conditions[J]. Remote sensing, 2020, 12(15): ID 2445.
[5]	KUMAR C, MUBVUMBA P, HUANG Y B, et al. Multi-stage corn yield prediction using high-resolution UAV multispectral data and machine learning models[J]. Agronomy, 2023, 13(5): ID 1277.
[6]	LU H, CAO Z G, XIAO Y, et al. Region-based colour modelling for joint crop and maize tassel segmentation[J]. Biosystems engineering, 2016, 147: 139-150.
[7]	KURTULMUŞ F, KAVDIR İ. Detecting corn tassels using computer vision and support vector machines[J]. Expert systems with applications, 2014, 41(16): 7390-7397.
[8]	ZHANG W Q, WU S, WEN W L, et al. Three-dimensional branch segmentation and phenotype extraction of maize tassel based on deep learning[J]. Plant methods, 2023, 19(1): ID 76.
[9]	GAO R, JIN Y S, TIAN X, et al. YOLOv5-T: A precise real-time detection method for maize tassels based on UAV low altitude remote sensing images[J]. Computers and electronics in agriculture, 2024, 221: ID 108991.
[10]	YU X, YIN D M, NIE C W, et al. Maize tassel area dynamic monitoring based on near-ground and UAV RGB images by U-Net model[J]. Computers and electronics in agriculture, 2022, 203: ID 107477.
[11]	WAN T Y, RAO Y, JIN X, et al. Improved U-Net for growth stage recognition of in-field maize[J]. Agronomy, 2023, 13(6): ID 1523.
[12]	LIU L B, YU L J, WU D, et al. PocketMaize: An Android-smartphone application for maize plant phenotyping[J]. Frontiers in plant science, 2021, 12: ID 770217.
[13]	LI Y H, ZHAO Z X, LUO Y F, et al. Real-time pattern-recognition of GPR images with YOLOv3 implemented by tensorflow[J]. Sensors, 2020, 20(22): ID 6476.
[14]	PU H L, CHEN X, YANG Y Y, et al. Tassel-YOLO: A new high-precision and real-time method for maize tassel detection and counting based on UAV aerial images[J]. Drones, 2023, 7(8): ID 492.
[15]	NIU S W, NIE Z G, LI G, et al. Multi-altitude corn tassel detection and counting based on UAV RGB imagery and deep learning[J]. Drones, 2024, 8(5): ID 198.
[16]	JIA Y J, FU K, LAN H, et al. Maize tassel detection with CA-YOLO for UAV images in complex field environments[J]. Computers and electronics in agriculture, 2024, 217: ID 108562.
[17]	LIU Y L, CEN C J, CHE Y P, et al. Detection of maize tassels from UAV RGB imagery with faster R-CNN[J]. Remote sensing, 2020, 12(2): ID 338.
[18]	FALAHAT S, KARAMI A. Maize tassel detection and counting using a YOLOv5-based model[J]. Multimedia tools and applications, 2023, 82(13): 19521-19538.
[19]	REDMON J, DIVVALA S, GIRSHICK R, et al. You only look once: Unified, real-time object detection[C]// 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway, New Jersey, USA: IEEE, 2016: 779-788.
[20]	KHANAM R, HUSSAIN M. Yolov11: An overview of the key architectural enhancements[EB/OL]. arXiv: 2410.17725, 2024.
[21]	CAI H, LI J, HU M, et al. Efficientvit: Multi-scale linear attention for high-resolution dense prediction[EB/OL]. arXiv: 2205.14756, 2022.
[22]	VOITA E, TALBOT D, MOISEEV F, et al. Analyzing multi-head self-attention: Specialized heads do the heavy lifting, the rest can be pruned[EB/OL]. arXiv:1905.09418, 2019.
[23]	NASCIMENTO M G D, PRISACARIU V, FAWCETT R. DSConv: Efficient convolution operator[C]// 2019 IEEE/CVF International Conference on Computer Vision (ICCV). Piscataway, New Jersey, USA: IEEE, 2019: 5147-5156.
[24]	HUANG H, CHEN Z, ZOU Y, et al. Channel prior convolutional attention for medical image segmentation[J]. Computers in biology and medicine, 2024, 178: ID 108784.
[25]	WOO S, PARK J, LEE J Y, et al. CBAM: Convolutional Block attention module[C]// Computer Vision-ECCV 2018. Cham, Germany: Springer, 2018: 3-19.
[26]	LI J F, WEN Y, HE L H. SCConv: Spatial and channel reconstruction convolution for feature redundancy[C]// 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway, New Jersey, USA: IEEE, 2023: 6153-6162.
[27]	ZHANG X L, WANG Z S, WANG X S, et al. STARNet: An efficient spatiotemporal feature sharing reconstructing network for automatic modulation classification[J]. IEEE transactions on wireless communications, 2024, 23(10): 13300-13312.
[28]	CHEN H, WANG Y, GUO J, et al. Vanillanet: The power of minimalism in deep learning[J]. Advances in neural information processing systems, 2023, 36: 7050-7064.
[29]	QIN D F, LEICHNER C, DELAKIS M, et al. MobileNetV4: Universal models for the mobile ecosystem[C]// Computer Vision – ECCV 2024. Cham, Germany: Springer Nature Switzerland, 2024: 78-96.
[30]	MA N N, ZHANG X Y, ZHENG H T, et al. ShuffleNet V2: Practical guidelines for efficient CNN architecture design[C]// Computer Vision-ECCV 2018. Cham, Germany: Springer, 2018: 122-138.
[31]	WANG A, CHEN H, LIN Z J, et al. Rep ViT: Revisiting mobile CNN from ViT perspective[C]// 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway, New Jersey, USA: IEEE, 2024: 15909-15920.

Year	Number of images
Year	Train	Validation	Total
2023	2 397	1 026	3 423

Parameters	Setup
Epochs	100
Batch size	16
Learning rate	0.01
Optimizer	AdamW
Number of workers	8
Image size	640 × 640

Model	P/%	R/%	AP@0.5/%	AP@0.5：0.95/%	Params/M	FPS	GFLOPs
YOLOv11n+ StarNet	91.1	86.8	93.5	54.3	2.64	179.5	5.2
YOLOv11n+ VanillaNet	90.9	85.7	93.1	53.3	3.69	101.2	6.2
YOLOv11n+MoblieNetV4	91.2	85.2	92.7	53.2	3.84	112.4	7.2
YOLOv11n+ ShuffleNetv2	91.3	85.8	93.4	54.6	2.48	123.3	5.9
YOLOv11n+ RepViT	91.1	87.6	93.9	55.5	6.16	131.2	17.7
YOLOv11n+ EfficientViT	91.5	87.9	93.8	56.2	3.56	163.9	6.9

Model	P/%	AP@0.5/%	AP@0.5：0.95/%	Params/M	FPS	GFLOPs
YOLOv11n+ EfficientViT_M0	91.5	93.8	56.2	3.56	163.9	6.9
YOLOv11n+ EfficientViT_M1	92.0	94.3	55.8	4.37	152.8	12.6
YOLOv11n+ EfficientViT_M2	91.8	94.1	55.5	5.58	143.2	14.7
YOLOv11n+ EfficientViT_M3	90.8	94.2	55.7	8.27	139.8	18.4
YOLOv11n+ EfficientViT_M4	92.3	94.2	56.0	10.17	137.6	20.5
YOLOv11n+ EfficientViT_M5	91.4	94.3	56.5	13.85	122.4	33.0

Baseline model	Models	EfficientViT	C2PSA-CPCA	C3k2-SCConv	P/%	R/%	AP@0.5/%	Params/M	GFLOPs	FPS
YOLOv11n	Model 1	×	×	×	90.1	85.3	90.7	2.43	6.3	124.9
	Model 2	√	×	×	91.5	87.9	93.8	3.56	6.9	163.9
	Model 3	×	√	×	90.8	86.7	93.3	2.43	6.3	239.2
	Model 4	×	×	√	90.7	86.8	93.2	2.11	5.4	226.3
	Model 5	√	√	×	92.2	87.9	94.1	3.56	6.9	238.5
	Model 6	√	×	√	92.1	88.4	94.4	3.15	6.7	190.8
	Model 7	×	√	√	91.8	87.6	93.9	2.26	5.4	285.6
	LightTassel-YOLO	√	√	√	92.6	89.1	94.7	3.23	6.7	226.9