Academic Journal of Science and Technology 
ISSN: 2771-3032 | Vol. 14, No. 1, 2025 

 

150 

Image	Recognition	Method	for	Corn	Northern	Leaf	
Blight	of	Unmanned	Aerial	Vehicles	Based	on	Deep	
Learning	
Xianyuan Liu, Wei Su 

Jilin Institute of Chemical Technology, Jilin 132000, China 

 

Abstract: Aiming at the problems of high complexity, false detection and missing detection of maize big spot based on UAV 
image, an improved algorithm for maize big spot detection was proposed. The algorithm is based on EMA and the improved 
YOLOv8-BiFPN feature pyramid network. By adding the efficient multi-scale EMA attention mechanism module, the capturing 
ability of detail information is improved, thus enhancing the feature extraction ability of the model. The YOLOv8 structure is 
improved into YOLOV8-BIFPN feature pyramid network by integrating BiFPN structure, which can extract context information 
more efficiently. By introducing WloU loss function, low quality samples in training data are filtered effectively, and the 
generalization ability of the model is improved. In this paper, the accuracy P was increased by 2.7%, the recall rate R was 
increased by 4.7%, the average accuracy mAP50 was increased by 3.1%, and the model size was reduced by 2.78M. Compared 
with other YOLO algorithms, the proposed method has significant advantages in the detection of corn big spot disease. 

Keywords: EMA Attention Mechanism; UAV image; YOLOv8-BiFPN; Corn Large Spot Disease. 
 

1. Introduction 
In China's agricultural structure, corn is the largest crop, 

accounting for nearly 40% of the total grain output, with a 
planting area of 650 million mu. It is the foundation of grain 
production, the link between crop cultivation and breeding, 
more than 60% of its output is used for the production of feed, 
and it plays an important role in the economy, sustainable 
agriculture, and the development of rural livelihoods, and is 
also key to ensuring food security and promoting sustainable 
social and economic development. However, in recent years, 
the frequent occurrence of corn diseases, especially the large 
spot disease, poses a major threat to corn leaves, affecting 
photosynthesis and thus threatening the yield and quality of 
corn. 

Due to the inefficiency, high cost, subjectivity, and lack of 
timeliness of traditional manual detection methods, machine 
vision technology is introduced to achieve rapid and effective 
identification of diseases, thereby formulating control 
measures to reduce the negative impact of diseases, ensuring 
food security, and promoting sustainable agricultural 
development. With the development and cost reduction of 
UAV technology, its flexibility and accessibility have been 
widely applied in various fields, and its aerial photography 
images play an important role in target detection. Combined 
with machine vision technology, it can achieve rapid and 
effective disease detection, providing innovative scientific 
and technological means for agricultural production.  

At present, YOLOv8, as a more advanced target detection 
model, has shown great potential in the agricultural field, 
especially in crop disease detection. Lu et al. proposed the 
YOLOv8_Rice model, which significantly improved the 
recognition ability of small-sized and irregularly shaped 
disease types by optimizing the attention mechanism and 
introducing deformable convolution. In addition, He et al. 
proposed the EDS-YOLOv8 algorithm, which significantly 
improved the accuracy of dense weed detection. Sun et al. 
proposed a lightweight network to improve the YOLOv8 

model, effectively reducing the model's parameter volume 
and computational complexity while maintaining accuracy in 
the detection of tobacco pests in complex environments. Deng 
Pengfei et al. used the fusion structure of Inception and VGG 
as the feature extraction model, and compressed the model 
through transfer learning, channel pruning, and knowledge 
distillation, achieving rapid identification of corn diseases. Ye 
et al. proposed an improved YOLOv8 lightweight pest image 
recognition model, which significantly improved the 
performance of small target detection by introducing 
BiFormer attention mechanism and GELAN network. Liu et 
al. proposed a feature-enhanced detection algorithm based on 
SSD, integrating the residual structure of the ResNet50 
network into the backbone network, improving the detection 
ability for small targets. The YOLOv8 model has also been 
applied to agricultural resource monitoring and yield 
estimation. Yu et al. proposed a model based on improved 
YOLOv8, using PConv technology to improve model 
detection accuracy in complex situations. 

With the development of agricultural automation and 
intelligence, the application of YOLOv8 algorithm models in 
agricultural robots and intelligent agricultural equipment is 
also becoming more and more widespread. Guo et al. used the 
improved YOLOv8 algorithm model to achieve early 
detection of harmful plants by combining multi-source image 
data. Zhu Qi et al. proposed a DSCS-YOLO-based apple 
surface defect detection method, designed a DSCS shallow-
deep feature selection module based on the Dense module and 
SE module, improved the original decoupling head with ELU 
activation function, and used the Wise-IoU loss function for 
nonlinear gain, achieving dynamic focused learning of the 
network. Liu et al. proposed a lightweight apple detection 
algorithm Faster-YOLO-AP based on improved YOLOv8, 
which is specifically designed for edge computing devices to 
achieve real-time object detection. Wang et al. proposed a 
YOLOv8+ model for strawberry maturity detection and 
classification, providing technical support for intelligent 
harvesting systems. Zhang et al. improved the YOLOv8 



 

151 

algorithm by introducing Mish activation function and 
DenseBlock structure, improving the accuracy and robustness 
of agricultural pest and disease detection. Zhu et al. proposed 
the Poly-YOLOv8 model and designed a new loss calculation 
algorithm, improving the model's detection ability for 
polygonal targets. 

The aforementioned methods have some issues with real-
time and accuracy in detection speed and model precision. In 
response to these issues, this paper improves the YOLOv8n 
as the benchmark algorithm to achieve UAV image corn large 
spot disease detection, and the main improvements and 
innovations to the YOLOv8 algorithm are as follows: first, a 
new attention mechanism module is introduced to improve 
recognition accuracy, then the new feature pyramid network 
structure adopts cross-scale connection methods and 
weighted feature methods to effectively integrate multi-scale 
features, fully solving the problem of large scale differences 
and feature loss, in addition, the introduction of a new loss 
function effectively reduces the negative impact of low-
quality samples, further enhancing the model's generalization 
performance.  

2. Improved YOLOv8 Algorithm 
The YOLOv8 algorithm is known for its fast speed and 

high accuracy, and has been widely applied. The innovation 
of the YOLOv8 model structure first uses a deep 
convolutional neural network as its feature extractor, which 
greatly helps in capturing richer image features. At the same 
time, multi-scale feature fusion technology is introduced into 
the algorithm, making it more effective in handling targets of 
different sizes and improving the algorithm's ability to detect 
small targets. YOLOv8 also optimizes the loss function by 
introducing new regularization terms and adjusting the 
predicted bounding box, further improving the accuracy of 
detection. YOLOv8 has achieved simultaneous improvement 
in recognition speed and detection accuracy through these 
improvements. The optimization of YOLOv8 has achieved 
good results in target detection tasks. 

2.1. Maize Large Spot Disease Recognition 
Algorithm Model 

To improve the detection accuracy of maize large spot 
disease, this paper proposes an improved optimization plan 
based on the YOLOv8 network architecture. The model is 
based on the YOLOv8 base algorithm and is improved and 
optimized, as shown in Figure 1. 

 

 

Figure 1. Improved and optimized YOLOv8 network structure 
 

The specific improvements of the improved model 
compared to the pre-improved model are as follows. 

Backbone Network: Three multi-scale attention EMA 
modules are introduced to capture detail information and 
improve the model's feature extraction capability, while 
retaining the original SPPF module in the model, which can 
solve the problem of repeated feature extraction of images by 

convolutional neural networks, speed up the generation of 
candidate boxes, and save computational costs. 

Neck Network: A lightweight YOLOv8-BiFPN feature 
pyramid network structure is proposed by fusing BiFPN and 
the original YOLOv8 structure to enhance the model's multi-
scale feature fusion capability and improve the network's 
detection accuracy for small targets. 



 

152 

Head Network: The original network loss function is 
optimized with WIoU Loss, introducing a dynamic non-
monotonic focusing mechanism to improve the model's 
generalization capability. 

In summary, this design method improves the existing 
algorithm, achieving high detection speed and model 
lightweight while improving detection accuracy, thereby 
enhancing the extraction capability of maize large spot 
disease features and enhancing the stability of the detection 
process.  

2.2. Backbone Network Improvement.  
By introducing the EMA (Efficient Multi-Scale Attention) 

attention module, the ability of the attention mechanism to 
obtain local important information is improved. EMA, as a 
multi-scale efficient attention mechanism, reshapes the 
channel dimension while maintaining the original channel 

dimension, improving computational ability while 
minimizing the loss of channel information. EMA is mainly 
responsible for global information encoding, can manage the 
channel weights of multiple parallel subnetworks, and 
promotes cross-dimensional information integration between 
these subnetworks, thereby producing output features. Its 
structure is shown in Figure 2. When the input features 

C H WR  X   are divided into N sub-features 

  C//C H
0 1, , , R W

i N i
 

 X X X X X   according to the EMA 

channel number, let N<<C, and assume that the learned 
weights can enhance the feature representation of key areas in 
all sub-features. EMA uses a 3x3 branch and two 1x1 
branches as parallel paths to obtain feature weights, which can 
obtain multi-scale features while avoiding dimensionality 
reduction. This study replaces the three C2f in the backbone 
network with EMA. 

 

h

h h h

h

c
w

∙ ∙ ∙ ∙

n

c//n c//n c//n
w w w

X
c//n

w

n

n*batch size

    
Figure 2. EMA structure diagram "n" represents grouping, "X Avg Pool" represents one-dimensional horizontal global pool, 

"Y Avg Pool" represents one-dimensional vertical global pool 
 

2.3. Neck Improvement 
To improve the detection accuracy of small targets such as 

large spot disease lesions, YOLOv8 has successfully 
integrated the advantages of PAN (Path Aggregation Network) 
and FPN (Feature Pyramid Network) structures, and this 

paper further introduces the concept of BiFPN (Bidirectional 
Feature Pyramid Network), proposing an innovative 
YOLOv8-BiFPN feature fusion network through new cross-
scale connection methods and feature fusion methods, the 
detailed network structure is shown in Figure 3. 

 



 

153 

P3

P4

P5

P6

P7

           
(a)BiFPN                           (b)YOLOv8-BiFPN 

Figure 3. BiFPN, YOLOv8-BiFPN feature network structure 
 

The construction method of the FPN feature pyramid is 
top-down, which can capture detailed information of images 
at different scales, thereby enhancing the performance and 
semantic information robustness of the algorithm. 
Nevertheless, the issue of semantic information differences at 
different levels in FPN still exists in YOLOv8. The 
construction method of PAN adds a bottom-up feature 
transmission method to FPN, effectively enhancing the 
transmission of low-level positioning features to the high 
level, and constructing a fused feature pyramid. YOLOv8 
further optimizes on the basis of PAN, by removing nodes that 
do not participate in feature fusion, both simplifying the 
network and improving efficiency. However, these 
improvements still have certain limitations in small target 
detection. Small targets are easily disturbed by large targets 
during feature extraction, and inconspicuous feature 
information is also easily ignored by the network, leading to 
the gradual loss of key information of small targets, thereby 
affecting the accuracy of detection. 

The BiFPN facilitates the connection of nodes at the input 
and output ends within the same level via an edge, thereby 
eliminating nodes incapable of feature integration. Moreover, 
it incorporates weighted feature fusion techniques, which 
supplant the conventional feature concatenation approach, 
thus more effectively preserving the hierarchical structure of 
features. In this study, an enhancement to the BiFPN is 
proposed, introducing a novel cross-layer connection 
methodology. Specifically, for the second and third layers of 
the network, 3×3 and 1×1 convolutional operations are 
respectively integrated to adjust the dimensions of the feature 
maps to 80×80, and subsequently merge these with the 
outputs of the corresponding layers. Additionally, a 1×1 
convolutional layer is inserted between the input and 
intermediate nodes, and a weighted fusion strategy is 
employed, which not only reduces the number of model 
parameters but also more effectively retains a greater amount 
of original feature information. 

2.4. Loss Function 
The WIoU loss function is used to replace the original loss 

function to improve the prediction accuracy of anchor frames. 
The WIoU v3 loss function dynamically allocates gradient 
gains to suppress harmful gradients generated by low-quality 
images, overcoming the insufficient localization performance 
of the YOLOv8n network, allowing the network to more 
accurately locate lesions in joint misalignment detection, 
thereby reducing errors and improving the model's 
recognition accuracy and generalization capability. 

 

1WIoUv WIoU IoUL R L                 (1) 

 
2 2

2 2 *

( ) ( )
exp( )

( )
gt gt

WIoU
g g

x x y y
R

W H

  



        (2) 

 

WIoU loss function is a loss function used for bounding 
box regression in target detection, which optimizes the quality 
assessment of anchor frames during training by dynamically 
allocating gradient gains. Compared with the CIoU loss 
function, the WIoU v3 loss function introduces a dynamic 
non-monotonic mechanism, which uses "outlier degree" 
instead of IoU for anchor frame quality assessment and 
provides a wise gradient gain allocation strategy. This strategy 
reduces the competitiveness of high-quality anchor frames 
while also reducing the harmful gradients generated by low-
quality examples, allowing WIoU to focus more on ordinary 
quality anchor frames and improve the overall performance 
of the detector. 

3. Experimental Data and Evaluation 
Metrics 

3.1.  Dataset 
The corn large spot disease image acquisition platform is 

composed of a UAV-mounted camera. Table 1 shows the 
corresponding equipment models and parameter 
configurations. 

 

Table 1. UAV Platform Parameters 
Numble Name Scheme 2 

1 UAV DJI Mavic Air 2s 
2 Image Resolution 6000×4000 
3 Maximum Flight Time 34 minutes 
4 Camera Sensor 1/2 inch CMOS, 48mp 



 

154 

 

To construct the corn large spot disease disease dataset, this 
study used the LabelImg tool to accurately label 4000 corn 
leaf disease images. To improve the generalization capability 
of the model during evaluation, the dataset was divided in a 
ratio of 8:1:1 to form the training set, test set, and validation 

set. A total of 3200, 400, and 400 images were obtained, 
meeting the needs for comprehensive training and model 
evaluation. This annotation and division process aims to 
provide the model with accurate enough training data to 
improve its performance in practical applications. 

 

 
Figure 4. Dataset Annotation File Statistics 

 

3.2. Experimental Environment and 
Parameter Configuration 

 
Table 2. YOLOv8 Model Training Environment 

Configuration Parameter 
CPU Intel(R) Core(TM) i7-12700 

Memory 32G 
GPU 

Video Memory 
Training Environment 

Operating System 
Development Environment 

NVIDIA GeForce RTX4070 
12G 

CUDA 12.4 
Windows10 (64-bit) 

Python3.11 Pytorch2.3 
 

Parameter Settings: The training period is set to 150 epochs, 
under which the model can reach a converged state; the 
learning rate is set to 0.01, and the number of images 
processed per batch is 16, ensuring that the GPU can operate 
at full power; the input image size is 640×640 pixels, and the 
images are scaled to a uniform size to improve processing 
speed. 

3.3. Evaluation Metrics 
In this paper, the following evaluation metrics are mainly 

used to measure the performance of the model: precision (P), 
recall rate (R), average precision (mAP), F1 score, model 
parameter quantity (Params), total floating-point operations 
(FLOPs). These metrics together form a comprehensive 
evaluation system, and the related formulas are as follows: 

 

TP
P=

TP+FP
                   (3) 

 
TP

R=
TP+FN

                   (4) 

 
1

0
AP= ( )P R dR                  (5) 

 

0mAP=

n

i
i

AP

n



                  (6) 



 

155 

2PR
F1=

P+R
                    (7) 

 
Precision and recall rate can be measured using true 

positives (TP), true negatives (TN), false positives (FP), and 
false negatives (FN). AP represents the average precision on 
each category, which can measure the model's performance 
on different categories, and mAP is the average value of all 
APs, which can measure the model's performance on all 
categories. The definition of mAP is shown in formula (6). 
mAP50 is the total category average precision when the IoU 

threshold is set to 0.5. 

4. Experimental Results and Analysis 

4.1. Ablation Study 
To ensure the rationality of the ablation study comparison, 

all participating models are trained based on the same dataset 
and parameter environment. In this study, YOLOv8n is used 
as the benchmark algorithm to implement a series of 
improvement measures, and the corresponding ablation 
experiments are completed. The results of the experiments are 
shown in Table 1. 

 

Table 3. Ablation Study Results 
Experiment Improvement 

Backbone 
Improvement 

Neck 
Improvement 

Loss 
P/% R/% mAP

50/% 
Params

/M 
FLOP

s/G 
F1/% 

1 × × × 80.1 73.4 77.2 9.31 20.4 76.6 
2 √ × × 81.1 74.2 78.1 10.35 21.5 77.5 
3 
4 
5 
6 
7 
8 

× 
× 
√ 
√ 
× 
√ 

√ 
× 
√ 
× 
√ 
√ 

× 
√ 
× 
√ 
√ 
√ 

80.1 
79.8 
81.4 
80.3 
81.7 
82.8 

75.4 
73.8 
74.7 
73.6 
76.9 
78.1 

77.7 
77.4 
78.8 
77.5 
79.2 
80.3 

7.51 
11.13 
6.72 
10.72 
7.48 
6.53 

16.7 
21.7 
16.3 
21.5 
16.7 
16.1 

77.8 
76.7 
77.9 
76.8 
79.2 
80.4 

 

In-depth analysis of the experimental results: Experiment 2 
integrated the EMA attention mechanism into the backbone 
network, and most performance indicators have improved, 
indicating that the optimization of the backbone network has 
enhanced the model's feature extraction capability and 
improved detection accuracy. Experiment 3 integrated the 
YOLOv8-BiFPN structure into the neck of YOLOv8n, and 
compared with the original algorithm, all performance 
indicators have significantly improved except for precision, 

ensuring that the detection accuracy is improved while the 
model is simplified. Considering that precision is not a core 
indicator of model evaluation, it can be considered that the 
improvement of the network neck is very successful. Finally, 
experiment 4 used WIoU as a new method for bounding box 
regression, and compared with the benchmark algorithm, it 
successfully improved the network's detection accuracy 
without increasing the model's complexity. 

 

 
Figure 5. Ablation Study mAP50 Comparison 

 

In the YOLOv8n and the improved backbone network 
structures, after introducing the WIoU loss function, the 
mAP50 increased by 0.2% and 0.3% respectively, the recall 
rate increased by 0.4% and 0.2% respectively, but the 
precision decreased, indicating that the improvement effect is 
not significant. However, when WIoU was applied to the 
improved neck network structure, the mAP50 significantly 
increased by 2%, the recall rate increased by 3.5%, and the 
precision also improved, reaching 1.6%. These comparative 
results indicate that the WIoU loss function performed better 

in the network optimized with the neck structure, implying 
that the WIoU loss function is more suitable for networks that 
have undergone multi-scale feature fusion. Experiment No. 8 
shows that after integrating all the improvements proposed in 
this paper into YOLOv8n, compared with the original 
algorithm, the method proposed in this paper increased the 
mAP50 by 3.1%, precision by 2.7%, recall rate by 4.7%, 
reduced the parameter volume by 2.78 million, reduced the 
total floating-point operations by 4.3G, and increased the F1 
score by 3.8%. These improvements make the algorithm 



 

156 

superior to YOLOv8n in all performance indicators. As 
shown in Figure 5, which compares the changes in mAP50, it 
not only shows the improvement in detection accuracy but 
also the simplification of the model, meeting the needs for 
real-time and accuracy. 

4.2. EMA Quantity Comparison Experiment 
There are 4 C2f modules in the YOLOv8 backbone 

network. Adding different quantities of EMA will have 
different effects on the model. This experiment compares the 
model by adding different quantities of EMA modules after 
the C2f modules in the YOLOv8 backbone network. The 
results are shown in Table 2. 

 

Table 4. EMA Quantity Comparison Experiment Results 

Method mAP50% Params/M FLOPs/G 
YOLOv8n 77.2 9.31 20.4 

1EMA 77.4 9.34 20.9 
2EMA 
3EMA 
4EMA 

77.6 
77.8 
77.8 

9.36 
9.38 
9.41 

21.2 
21.5 
21.9 

 
From the experimental data in Table 2, it is known that 

adding the EMA attention mechanism can improve the 
model's feature extraction capability and thereby improve 
detection accuracy, but at the same time, it causes an increase 
in the number of parameters, which will slow down the 
model's computation speed. Therefore, as the number of EMA 
introductions increases, a balance point needs to be sought 
between the number of parameters and the improvement of 
accuracy. When adding up to 3 EMA mechanisms, accuracy 
is improved and the increase in the number of parameters is 
within an acceptable range. However, after adding 4 EMA 
mechanisms, the accuracy did not significantly improve, and 
the number of parameters continued to increase. Therefore, 

adding 3 EMA mechanisms in the backbone network can 
achieve an ideal balance between speed and accuracy 
improvement. 

4.3. Comparative Experiment 
To fully evaluate the advantages of the improved YOLOv8 

algorithm proposed in this paper in the detection of corn large 
spot disease, this paper compares the improved algorithm 
with the YOLOv8n benchmark algorithm and other 
representative network models in the field, including Faster-
RCNN, SSD, RetinaNet, YOLOv5, YOLOv6, YOLOv7-tiny, 
and YOLOv8, as shown in Table 3. 

 

Table 5. Comparative Experiment Results of Different Models 
Model P/% R/% mAP50/% F1/% Params/M FLOPs/G 

Faster R-CNN 74.1 78.5 75.6 76.2 76.73 95.7 
SSD 80.4 77.4 79.4 78.9 24.71 57.2 

RetinaNet 74.6 73.1 74.7 73.8 18.9 60.4 
YOLOv5 75.3 76.4 76.3 75.8 6.51 15.8 
YOLOv6 77.1 73.7 77.3 75.4 9.50 17.5 

YOLOv7-tiny 78.5 76.9 77.7 78.2 6.14 12.1 
YOLOv8 80.1 73.4 77.2 76.6 9.31 20.4 

Improved YOLOv8n 82.8 78.1 80.3 80.4 6.53 16.1 
 

The results of the comparative experiment show that the 
improved algorithm proposed in this paper has achieved the 
highest level in terms of precision and mAP50 values. The 
algorithm not only has excellent performance but also the 
fastest detection speed, and the parameter volume is not much 
different from other models. Compared with the benchmark 
algorithm YOLOv8n, the improved algorithm shows 
significant advantages in all indicators. Considering all 
performance indicators, the comprehensive performance of 
the improved algorithm proposed in this paper is the best, 
with good detection ability for corn large spot disease in UAV 
images, and good performance in both speed and accuracy. To 
intuitively show the superiority of this algorithm, the mAP50 
change curve is compared, and the experimental results are 
shown in Figure 6. From Figure 6, it can be seen that all 
algorithms can converge through continuous iteration, but the 
curve of the improved YOLOv8n is significantly higher than 
other algorithms, proving that the improved YOLOv8 
algorithm has effectively improved detection accuracy. 

 
Figure 6. Comparison of mAP50 for Different Models 

 

5. Summary 
In this study, considering the high complexity and low 



 

157 

distinguishability of the morphological and color features of 
corn large spot disease lesions, an improved YOLOv8 corn 
large spot disease detection model was proposed, specifically 
for detecting corn large spot disease. The model's training was 
based on a mixed dataset, which was composed of some 
public dataset images and our own collected corn large spot 
disease UAV images. The following are the main conclusions 
of this study: First, a corn large spot disease image dataset 
under natural conditions was successfully constructed, and its 
data integrity and reliability were ensured, which is crucial for 
model training and evaluation. Second, the EMA attention 
mechanism was added to the backbone network, the 
YOLOv8-BiFPN was designed to optimize the model's neck 
structure, and a new WIoU loss function was introduced to 
improve the YOLOv8 algorithm, making it specifically 
applicable to the detection of corn large spot disease in UAV 
images. Finally, compared with the benchmark algorithm 
YOLOv8n, the proposed algorithm shows superiority in 
many aspects, in comparison with other advanced methods, it 
has the best comprehensive performance, meeting the 
requirements for real-time and accuracy. In future research, 
we plan to increase the types of diseases and the number of 
samples to improve the model's accuracy and recognition 
ability. In addition, we will continue to explore the 
combination and optimization improvement of different 
network models to further improve the model's training speed 
and accuracy. 

References 
[1] Han Dongfeng, Li Feng, Qin Quan, et al. A refined 

identification method for summer corn planting area based on 
GEE and Sentinel-1/2 data[J]. Journal of Marine Meteorology, 
2024, 44(03): 122-132. 

[2] Xu Xingbin. Characteristics of corn disease occurrence and 
control techniques[J]. Modern Agriculture, 2024, (08): 47-49. 

[3] Dai Linhua, Li Yuansong, Shi Rui. Improved YOLOv8n 
algorithm for rice leaf disease detection[J]. Journal of Hubei 
University for Nationalities (Natural Science Edition), 2024, 
42(03): 382-388. 

[4] Meng Wei, Wang Jinghui. Progress in the application of UAV 
technology in forest pest and disease monitoring[J/OL]. Forest 
Science and Communication, 1-6 [2024-10-10]. 

[5] Lu, Y., Yu, J. H., Zhu, X. F., Zhang, B. F., & Sun, Z. F. (2024). 
YOLOv8-Rice: a rice leaf disease detection model based on 
YOLOv8. PADDY AND WATER ENVIRONMENT. 

[6] He, C. C., Wan, F. X., Ma, G. J., Mou, X. B., Zhang, K. K., 
Wu, X. F., & Huang, X. P. (2024). Analysis of the Impact of 
Different Improvement Methods Based on YOLOV8 for Weed 
Detection. AGRICULTURE-BASEL. 

[7] Sun, D. Z., Zhang, K., Zhong, H. S., Xie, J. X., Xue, X. Y., Yan, 
M. L., Wu, W. B., & Li, J. H. (2024). Efficient Tobacco Pest 

Detection in Complex Environments Using an Enhanced 
YOLOv8 Model. AGRICULTURE-BASEL. 

[8] Deng Pengfei, Guan Zheng, Wang Yuyang, et al. A method for 
corn disease recognition based on transfer learning and model 
compression[J]. Computer Science, 2022, 49(S2): 444-449. 

[9] Ye, R., Gao, Q., Qian, Y., Sun, J. H., & Li, T. (2024). Improved 
YOLOv8 and SAHI Model for the Collaborative Detection of 
Small Targets at the Micro Scale: A Case Study of Pest 
Detection in Tea. AGRONOMY-BASEL. 

[10] LIU W J, QIANG J, LI X X, et al. UAV Image Small Object 
Detection Based on Composite Backbone Network[J]. Mobile 
Information Systems, 2022(15):1-11. 

[11] Yu, X., Yin, D. M., Xu, H. G., Espinosa, F. P., Schmidhalter, 
U., Nie, C. W., Bai, Y., Sankaran, S., Ming, B., Cui, N. B., Wu, 
W. B., & Jin, X. L. (2024). Maize tassel number and tasseling 
stage monitoring based on near-ground and UAV RGB images 
by improved YoloV8. PRECISION AGRICULTURE. 

[12] Guo, B. L., Ling, S. K., Tan, H. Y., Wang, S., Wu, C. L., & 
Yang, D. S. (2024). Detection of the Grassland Weed 
Phlomoides umbrosa Using Multi-Source Imagery and an 
Improved YOLOv8 Network. AGRONOMY-BASEL. 

[13] Zhu Qi, Zhou Deqiang, Sheng Weifeng, et al. A DSCS-YOLO-
based apple surface defect detection method [J]. Journal of 
Nanjing Agricultural University, 2024, 47(03): 592-601. 

[14] Liu, Z. F., Abeyrathna, R. M. R. D., Sampurno, R. M., 
Nakaguchi, V. M., & Ahamed, T. (2024). Faster-YOLO-AP: A 
lightweight apple detection algorithm based on improved 
YOLOv8 with a new efficient PDWConv in orchard. 
COMPUTERS AND ELECTRONICS IN AGRICULTURE. 

[15] Wang, C. L., Wang, H. M., Han, Q. Y., Zhang, Z. G., Kong, D. 
D., & Zou, X. J. (2024). Strawberry Detection and Ripeness 
Classification Using YOLOv8+ Model and Image Processing 
Method. AGRICULTURE-BASEL. 

[16] Zhang, L. J., Ding, G. C., Li, C. R., & Li, D. M. (2024). DCF-
Yolov8: An Improved Algorithm for Aggregating Low-Level 
Features to Detect Agricultural Pests and Diseases. 
AGRONOMY-BASEL. 

[17] Zhu, R. X., Hao, F. Q., & Ma, D. X. (2024). Research on 
Polygon Pest-Infected Leaf Region Detection Based on 
YOLOv8. AGRICULTURE-BASEL. 

[18] OUYANG D L, HE S, ZHANG G Z, et al. Efficient Multi Scale 
Attention Module with Cross-Spatial Learning[C] // 
Proceedings of the IEEE International Conference on 
Acoustics   Vancouver : IEEE,2023:12021-12031. 

[19] Tan M, Pang R, Le Q V. Efficientdet: Scalable and efficient 
object detection[C]//Proceedings of the IEEE/CVF conference 
on computer vision and pattern recognition. 2020: 10781-
10790 

[20] Tong Z, Chen Y, Xu Z, et al. Wise-IoU: bounding box 
regression loss with dynamic focusing mechanism[J]. arXiv 
preprint arXiv:2301.10051, 2023.