基于无人机影像与YOLOv12n-RCL的向日葵成熟期盘腐病危害程度检测

摘要:

高效、精准地评估向日葵盘腐病危害程度对早期防治、精准施药和产量损失估测至关重要。针对传统病害分级方法效率低、主观性强的问题，该研究提出一种基于改进YOLOv12n的向日葵成熟期盘腐病危害程度检测模型。首先，将主干和颈部网络的C3K2模块替换为融合感受野注意力卷积（receptive-field attention convolution, RFAConv）与坐标注意力机制（coordinate attention, CA）的C3K2-RC模块，以增强模型在复杂环境中对不同等级盘腐病的特征提取能力；其次，在颈部网络集成轻量级的上采样算子CARFAE提升特征重构效果；最后，引入一种分离批量归一化的轻量级共享卷积检测头LSCSBD（lightweight shared convolutional separable batch normalization detection），以提高模型对小尺度病斑的检测精度与速度。试验结果表明，YOLOv12n-RCL模型的精确率、召回率、mAP 0.5 和mAP 0.5-0.95 分别达到了83.4%、81.2%、84.8%和50.2%，较基准模型YOLOv12n分别提高了3.8、3.2、3.4和4.0个百分点；参数量、计算量和模型大小分别压缩至2.11 M、5.7 G和4.5 MB，较原模型YOLOv12n分别降低了17.9%、12.3%和19.6%。基于混淆矩阵的病害等级误检结果表明，改进模型对不同等级盘腐病的错误预测比例均在10%以下。边缘设备部署测试显示，在包含 1298 个不同等级盘腐病样本的70幅测试图像中，YOLOv12n-RCL仅出现10例漏检和8例误检，与基准模型YOLOv12n相比，检测效果更优，且帧率更高，达到了27.5帧/s。综上，YOLOv12n-RCL模型在提升检测精度的同时，实现了轻量化设计与部署效率的有效平衡，可为向日葵成熟期盘腐病智能防治装备的研发提供算法参考。

Abstract:

Sunflower is one of the preferred oil crops in the mildly to moderately saline-alkali soils, due to its strong tolerance to nutrient-poor soils and drought. Among them, sunflower disk rot can be caused by sclerotinia sclerotiorum infection, leading to increasing sterile seed rates. Nutritional components can also decline, such as kernel protein and oil content. Thereby, both the edible and commercial value of the seeds are often required for high yield. Consequently, it is crucial to efficiently and accurately detect disk rot severity for early disease control, precision pesticide application, and yield estimation. However, conventional disease classification has been limited to manual efficiency and subjectivity in recent years. In this study, an improved YOLOv12n-RCL model was proposed to detect the severity of sunflower disk rot at the mature stage. 1) C3K2-RC modules were used to replace C3K2 ones in the backbone and neck networks. Receptive-field attention convolution (RFAConv) and coordinate attention (CA) mechanisms were integrated to enhance the feature extraction from the different severity grades of disk rot in complex environments. 2) A lightweight upsampling operator, CARFAE, was integrated into the neck network for feature reconstruction. 3) A lightweight shared convolutional detection head with separated batch normalization (LSCSBD) was introduced to improve the detection accuracy and speed for small-scale lesions. Experimental results show that the YOLOv12n-RCL model achieved a precision, recall, mAP 0.5, and mAP 0.5~0.95 of 83.4%, 81.2%, 84.8%, and 50.2%, respectively, which was improved by 3.8, 3.2, 3.4, and 4.0 percentage points over the baseline model. The number of parameters, computational complexity, and model size were reduced to 2.11 M, 5.7 GFLOPs, and 4.5 MB, respectively, corresponding to reductions of 17.9%, 12.3%, and 19.6%, compared with the original model. The normalized confusion matrix indicated that the recall for the disease severity grades 0, 1, 2, 3, and 4 were 83.0%, 81.2%, 79.1%, 80.7%, and 82.0%, respectively, indicating a balanced recognition over all five grades. Furthermore, 8.2% of samples with a true label of grade 2 were misclassified as grade 3, and 9.3% of samples with a true label of grade 3 were misclassified as grade 2. Therefore, the confusion between severity grades occurred primarily between grades 2 and 3. The misclassification rate remained below 10% in the rest, indicating its strong performance for the disk rot at different grades. A field test was conducted on 70 images with 1 298 samples of different grades of disk rot. The YOLOv12n-RCL model achieved only 10 missed detections and 8 false detections. Compared with the baseline YOLOv12n model, the superior performance was achieved with a higher frame rate of 27.5 frames per second (FPS). Visualization was also conducted on the sunflower disk rot areas using UAV orthophotography. The improved model maintained stable performance even in complex scenarios with dense target distribution and leaf occlusion, without significantly missing or false detections. In summary, the YOLOv12n-RCL model improved the detection accuracy to effectively balance the lightweight deployment efficiency. The finding can also serve as an algorithmic reference to detect the sunflower disk rot at the mature stage in smart agriculture.

图 1 不同等级盘腐病样本实例

Figure 1. Examples of disk rot samples at different grades

图 2 YOLOv12n-RCL模型结构图

注：Conv为卷积块，Conv 2d为二维卷积块，BatchNorm（BN）为批量归一化层，SiLU为激活函数，C3K为自定义卷积核大小的特征提取模块，C3K2为残差连接特征提取模块，C3K2-RC为融合了感受野卷积和坐标注意力机制的C3K2模块，Split为特征分割操作，n 为模块出现的次数，Concat为特征拼接层，A2为区域注意力机制，A2C2f为结合了区域注意力机制和残差连接的特征提取模块，Scaling为特征缩放操作，ABlock为区域注意力多层感知机模块，CARFAE为上采样算子，LSCSBD为批量归一化的轻量级共享卷积检测头。下同。

Figure 2. Structure diagram of the YOLOv12n-RCL model

Note: Conv refers to the convolutional block, Conv 2d refers to the 2D convolutional block, BatchNorm (BN) refers to the batch normalization layer, SiLU is the activation function, C3K is the custom convolution kernel size feature extraction module, C3K2 is the residual connection feature extraction module, C3K2-RC is the C3K2 module that integrates receptive field convolution and coordinate attention mechanism, Split refers to the feature splitting operation, n is the number of times the module appears, Concat is the feature concatenation layer, A2 is the region attention mechanism, A2C2f is the feature extraction module combining region attention mechanism and residual connections, Scaling refers to the feature scaling operation, ABlock is the region attention multilayer perceptron module, CARFAE is the upsampling operator, and LSCSBD refers to the lightweight shared convolutional detection head with batch normalization. The same below.

图 3 RFAConv模块结构图

Figure 3. Structure diagram of the RFAConv module

图 4 CA注意力机制工作流程图

注：Residual为残差连接，X Avg Pool为宽度方向的平均池化，Y Avg Pool为高度方向的平均池化，r 为通道缩减比例系数，BatchNorm + Non-linear为标准化和非线性处理单元，S为Sigmoid激活函数，下同。

Figure 4. Workflow diagram of CA attention mechanism

Note: Residual refers to the residual connection, X Avg Pool refers to average pooling along the width dimension, Y Avg Pool refers to average pooling along the height dimension, r denotes the channel reduction ratio coefficient, BatchNorm + Non-linear refers to the batch normalization and non-linear processing unit, and S refers to the Sigmoid activation function. The same below.

图 5 C3K2- RC网络结构图

注：RFAConv为感受野注意力卷积，RFAConv-Bottleneck为添加感受野注意力卷积的瓶颈层，CA为坐标注意力机制，下同。

Figure 5. C3k2-RC network structure diagram

Note: RFAConv refers to the receptive field attention convolution, RFAConv-Bottleneck refers to the bottleneck layer incorporating receptive field attention convolution, and CA indicates the coordinate attention mechanism. The same below.

图 6 CARAFE上采样网络结构图

注：X 表示输入特征图，$ {C}{m} $为通道压缩后的特征图通道数，$ N({\chi }{l},{k}{up}) $表示以$ {\chi }{l} $为中心、$ {k}{up} $为大小的局部邻域，$ \sigma $为上采样倍数，$ W{l}^{{'}} $表示上采样核预测模块生成的权重，$ {X}^{{'}} $表示输出特征图。

Figure 6. Network structure diagram of CARAFE upsampling

Note: X refers to the input feature map, $ {C}{m} $ is the number of channels of the feature map after channel compression, $ N({\chi }{l},{k}{up}) $ indicates the local neighborhood centered at $ {\chi }{l} $ with a size of $ {k}{up} $, σ is the upsampling factor, $ W{1}^{{'}} $ refers to the weights generated by the upsampling kernel prediction module, and $ {X}^{{'}} $ refers to the output feature map.

图 7 LSCSBD检测头网络结构图

注：P3、P4、P5为第3、4和5层输入特征图，Conv_BN 1×1为批量归一化卷积层，Conv_GN 3×3为组归一化共享卷积层，BnAct为批量归一化和激活函数的组合模块，Conv-Reg为回归卷积层，Conv-Cls为分类卷积层，Scale为缩放因子。

Figure 7. LSCSBD detection head network structure diagram

Note: P3, P4, and P5 refer to the feature maps at the 3 rd, 4 th, and 5 th layers. Conv_BN 1×1 is the batch normalization convolutional layer, Conv_GN 3×3 is the group normalization shared convolutional layer, BnAct is the combination module of batch normalization and activation function, Conv-Reg is the regression convolutional layer, Conv-Cls is the classification convolutional layer, Scale is the scaling factor.

图 8 模型优化前后的性能评估结果

注：mAP 0.5 表示IoU阈值为0.5时的平均精度均值，mAP 0.5-0.95 表示IoU阈值为0.95时的平均精度均值，下同。

Figure 8. Performance comparison between the baseline and the optimized model

Note: mAP 0.5 represents the mean average precision at an IoU threshold of 0.5, and mAP 0.5-0.95 represents the mean average precision across IoU thresholds from 0.5 to 0.95. The same below.

图 9 C3K2引入不同模块后的输出感受野

Figure 9. Output receptive field of C3K2 with the integration of different modules

图 10 添加CARAFE模块前后输入输出特征对比图

注：图中红色框代表向日葵目标。

Figure 10. Comparison of input and output features before and after adding the CARAFE module

Note: The red frame in the image represents the sunflower target.

图 11 消融试验的可视化效果

Figure 11. Visualization effect of the ablation study

图 12 不同目标检测模型识别效果对比

注：图中蓝色圆表示漏检，红色三角表示错检。

Figure 12. Performance comparison of different object detection models

Note: In the figure, The blue circle indicates missed detections, while the red triangle indicates false detections.

图 13 不同等级病害的混淆矩阵

Figure 13. Confusion matrix for different disease severity grades

图 14 向日葵盘腐病检测系统界面

Figure 14. Sunflower disk rot detection system interface

图 15 向日葵盘腐病区域监测与分级结果可视化

Figure 15. Visualization of monitoring and grading results for sunflower disk rot regions

表 1 向日葵盘腐病数据集的样本信息

Table 1 Sample information of the sunflower disk rot dataset

病害等级

Disease grade 训练集

Training set 测试集

Test set 验证集

Validation set 增强训练集

Augmented training set
0 7108 900 1012 28432
1 6517 872 855 26068
2 6633 767 881 26532
3 6400 847 860 25600
4 6817 873 810 27364

表 2 试验的软硬件配置参数

Table 2 Hardware and software configuration parameters for experiments

配置Configuration 参数Parameters
中央处理器 Intel(R)Core(TM)i7-11700k
图形处理器 NVIDIA GeForce RTX 3060 Ti
优化算法 SGD
操作系统 Ubuntu 18.04
配置环境 Pytorch 2.6.0+Python 3.10.16
训练轮次 200
批次大小 16
输入图像尺寸 1024×1024
初始学习率 0.01
动量系数 0.935
权重衰减次数 0.0005

表 3 C3K2-RC模块不同引入位置的对比试验

Table 3 Comparative experiment on different positions for the introduction of the C3K2-RC module %

替换位置 Replace position P R mAP 0.5 mAP 0.5-0.95
主干 80.0 76.9 81.9 48.1
颈部 78.0 75.8 81.7 47.0
主干+颈部 81.1 77.8 82.2 48.3

表 4 消融试验结果

Table 4 Ablation test results

C3K2-RC CARAFE LSCSBD P/%R/%mAP 0.5/%mAP 0.5-0.95/%参数量

Parameters/M 计算量

FLOPs/G 模型大小

Model size/MB
□□□79.6 78.0 81.4 46.2 2.57 6.5 5.6
■□□81.1 77.8 82.2 48.3 2.59 6.6 5.7
□■□79.6 78.4 81.6 47.0 2.23 5.8 4.9
□□■80.2 78.7 82.7 48.4 2.44 6.4 5.3
■■□82.7 80.1 83.6 49.7 2.33 6.3 5.1
■□■83.7 81.2 84.5 50.1 2.46 6.5 5.4
□■■79.9 78.0 82.1 48.4 2.09 5.6 4.5
■■■83.4 81.2 84.8 50.2 2.11 5.7 4.5
注：“■”表示采用了该模块，“□”表示未采用该模块，FLOPs为浮点数运算次数。下同。
Note: "■" indicates that the module is adopted, "□" indicates that the module is not adopted, and FLOPs stands for the number of floating-point operations. The same below.

表 5 不同检测网络的综合性能对比

Table 5 Comparative analysis of the overall performance of different detection networks

模型

Model 类型

Type P/%R/%mAP 0.5/%mAP 0.5-0.95/%参数量

Parameters/M 计算量

FLOPs/G 模型大小

Model size/MB
Faster R-CNN Two-Stage 70.1 65.6 68.8 40.2 137.1 370.2 114.2
SSD One-Stage 73.1 66.4 72.1 42.3 24.0 61.1 105.2
CenterNet One-Stage 72.1 67.1 75.6 43.7 32.7 70.2 90.6
RT-DETR(resnet50)DETR 71.9 68.2 78.7 44.1 42.8 130.5 96.7
YOLOv5s One-Stage 78.9 77.3 81.2 46.4 7.26 16.4 13.7
YOLOv7-tiny One-Stage 80.1 81.0 80.5 45.5 6.02 13.2 11.7
YOLOv8n One-Stage 79.1 78.0 81.4 46.7 3.01 8.2 6.3
YOLOv9 t One-Stage 76.5 74.2 79.5 44.7 1.73 6.4 4.3
YOLOv10n One-Stage 79.2 78.7 80.4 45.1 2.69 8.2 5.8
YOLOv11n One-Stage 78.3 78.8 80.6.46.0 2.61 6.6 5.7
YOLOv12n One-Stage 79.6 78.0 81.4 46.2 2.57 6.5 5.6
YOLOv12n-RCL One-Stage 83.4 81.2 84.8 50.2 2.11 5.7 4.5

[1]李丹, 赵存鹏, 刘素恩, 等. 39份油用向日葵杂交种主要农艺性状与产量的多元分析[J]. 华北农学报, 2023, 38(S1): 154-160.
[2]王彦尊, 王文浩, 何丽芬, 等. 不同向日葵品系对菌核病抗性的综合鉴定[J]. 中国种业, 2023(12): 130-135.
[3]史明. 向日葵病害综合防治技术[J]. 农业开发与装备, 2015(10): 146.
[4]魏伟业. 试论向日葵几种常见病及防治方法[J]. 农民致富之友, 2019(2): 13.
[5]FERENTINOS K P. Deep learning models for plant disease detection and diagnosis[J]. Computers and Electronics in Agriculture, 2018, 145: 311-318. doi: 10.1016/j.compag.2018.01.009
[6]UPADHYAY A, CHANDEL N S, SINGH K P, et al. Deep learning and computer vision in plant disease detection: a comprehensive review of techniques, models, and trends in precision agriculture[J]. Artificial Intelligence Review, 2025, 58(3): 92. doi: 10.1007/s10462-024-11100-x
[7]朱锐, 张家瑜, 黄继超, 等. 基于卷积神经网络的农作物叶片病害检测研究进展[J]. 农业工程学报, 2025, 41(17): 15-28.
ZHU Rui, ZHANG Jiayu, HUANG Jichao, et al. Research progress of crop leaf disease detection based on convolutional neural network[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2025, 41(17): 15-28. (in Chinese with English abstract)
[8]GUAN B, WU Y, ZHU J, et al. GC-Faster RCNN: The object detection algorithm for agricultural pests based on improved hybrid attention mechanism[J]. Plants, 2025, 14(7): 1106. doi: 10.3390/plants14071106
[9]兰玉彬, 孙斌书, 张乐春, 等. 基于改进YOLOv5s的自然场景下生姜叶片病虫害识别[J]. 农业工程学报, 2024, 40(1): 210-216.
LAN Yubin, SUN Binshu, ZHANG Lechun, et al. Identifying diseases and pests in ginger leaf under natural scenes using improved YOLOv5s[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2024, 40(1): 210-216. (in Chinese with English abstract)
[10]GONG X, ZHANG S. A high-precision detection method of apple leaf diseases using improved Faster R-CNN[J]. Agriculture, 2023, 13(2): 240. doi: 10.3390/agriculture13020240
[11]吴世琛, 毛玉明, 胡慧中, 等. 基于改进YOLO11n的葡萄果叶病害检测方法[J]. 农业工程学报, 2025, 41(14): 140-147.
WU Shichen, MAO Yuming, HU Huizhong, et al. Detecting grape leaf diseases using improved YOLO11n[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2025, 41(14): 140-147. (in Chinese with English abstract)
[12]CHEN S, LV F, HUO P. Improved detection of YOLOv4 sunflower leaf diseases[C]//2021 2nd International Symposium on Computer Engineering and Intelligent Communications (ISCEIC). Nanjing, China: IEEE, 2021: 56-59.
[13]于海琳, 兰玉彬, 李京谦, 等. 基于无人机遥感数据和机器学习的向日葵LAI反演[J]. 农业机械学报, 2025, 56(1): 356-365.
YU Hailin, LAN Yubin, LI Jingqian, et al. Sunflower LAI inversion based on unmanned aerial vehicle remote sensing data and machine learning[J]. Transactions of the Chinese Society for Agricultural Machinery, 2025, 56(1): 356-365. (in Chinese with English abstract)
[14]胡根生, 谢一帆, 鲍文霞, 等. 基于轻量型网络的无人机遥感图像中茶叶枯病检测方法[J]. 农业机械学报, 2024, 55(4): 165-175.
HU Gensheng, XIE Yifan, BAO Wenxia, et al. Lightweight network for tea leaf blight detection in UAV remote sensing images[J]. Transactions of the Chinese Society for Agricultural Machinery, 2024, 55(4): 165-175. (in Chinese with English abstract)
[15]PAN P, GUO W, ZHENG X, et al. Xoo-YOLO: A detection method for wild rice bacterial blight in the field from the perspective of unmanned aerial vehicles[J]. Frontiers in Plant Science, 2023, 14: 1256545. doi: 10.3389/fpls.2023.1256545
[16]陈自立, 郭燕, 王明鑫, 等. 基于RT-YOLOv10和无人机遥感影像的烟草病害轻量化检测研究[J]. 山东农业科学, 2025, 57(9): 149-163.
[17]YANG Z, KHAN Z, SHEN Y, et al. GTDR-YOLOv12: Optimizing YOLO for efficient and accurate weed detection in agriculture[J]. Agronomy, 2025, 15(8): 1824. doi: 10.3390/agronomy15081824
[18]李易初, 孟庆林, 石凤梅, 等. 田间防治向日葵菌核盘腐病的药剂筛选[J]. 宁夏农林科技, 2015, 56(7): 53-54.
[19]赵静, 李京谦, 杨蕾, 等. 基于PConv-CGLU与重参数检测头的轻量化膜下棉苗实时检测算法[J]. 农业工程学报, 2025, 41(18): 151-162.
ZHAO Jing, LI Jingqian, YANG Lei, et al. Real-time detecting cotton seedlings under a lightweight film using PConv-CGLU with a heavy parameter detection head[J]. Transactions of the Chinese Society of A