Detection of Pomelo in Overlapping Conditions Using Drones

Mahdaniar

and Indrabayu

Department of Informatics Hasanuddin University Makassar, Indonesia

Keywords: Detection, Pomelo, YOLOv11, CLAHE, Overlapping.

Abstract: Detection of Pomelo on trees in overlapping conditions and the similarity of colors between fruits and leaves

are the main challenges in the implementation of smart farming systems. This study aims to develop an

automatic detection system of Pomelo using a drone with a computer vision approach based on the YOLOv11

algorithm combined with the CLAHE (Contrast Limited Adaptive Histogram Equalization) image contrast

enhancement technique. The research methodology includes image data collection, pre-processing, data

labeling, model training, and evaluation using mAP, precision and recall. The initial results showed that

YOLOv11 provide suboptimal performance in the detection process, resulting in the result, precision: 92%,

recall: 81%, mAP50: 90%, mAP50-95: 72%. After YOLOv11 is integrated with CLAHE, the performance

has been improved, achieving precision: 92%, recall: 84%, mAP50: 95%, mAP50-95: 67%.

1 INTRODUCTION

The Industrial Revolution 4.0 encourages the

transformation of automation systems that are

smarter, more efficient, and sustainable, driven by

the development of digital technology, including in

agriculture. One of the leading agricultural

commodities is horticulture (Indrabayu et al., 2019a).

Horticultural crops that have the potential to be

developed in Indonesia are citrus commodities.

Citrus (Citrus Sp) is one of the horticultural

commodities that functions as a source of national

income for the country, the fruit originating from

Asia that can grow in tropical and subtropical areas

(Addi et al., 2021). This is marked by the increasing

consumption of pomelos in Indonesia from year to

year. In general, pomelos are a o vitamin C source

useful for human health. Pomelo juice contains 40-

70 mg of vitamin C per 100 g (Khattak et al., 2021).

In addition to pomelo flesh, the properties and

benefits of pomelos are also contained in pomelo

peels. The content of pomelo peel has benefits

ranging from sedatives, skin smoothers, mosquito

repellent (Adelina et al., 2020) (Sugadev et al.,

2020).

The availability of computer vision technology,

supported by the improvement of computer hardware

capabilities, has become a major supporting factor in

https://orcid.org/0009-0005-3135-5861

https://orcid.org/0000-0003-2026-1809

the development of automated farming systems.

These systems are designed to solve various

problems in the agricultural sector with high level of

flexibility, effectiveness, and efficiency. This

development has also led to an increase in the

number of studies focusing on automatic fruit

detection (Indrabayu et al., 2019b).

Previous research, a system for classifying ripe,

immature and rotten pomelos quickly and efficiently

by 335 Images taken using a 1024x768 pixel

resolution digital camera are stored in JPG format

using Naïve Bayes, Decision Tree and Neural

Network methods (Wajid et al., 2018), automatic

classification of pomelos at the fruit packaging

factory with 120 images and three categories of

pomelos, namely Bam, Payvandi, Thomson using the

Gradient Descent, Stochastic Gradient Descent

(SGD), RMS Prop and Adam methods, the results of

the study obtained the best SGD method with an

optimization of 0.95 (Pathak et al., 2020) detection

lime fruit on a tree for estimation of yield using the

YOLOv5-CS method using a dataset of 3000 images

with a detection accuracy of 96.7% (Lyu

et al.,

2022a

)

, pomelo fruit color detection based on frame

selection detection on 811 pomelo fruit images using

the YOLOv5 method obtained an accuracy rate of

99.4% [9], system for the classification of pomelo

quality based on the characteristics specified on 953

Mahdaniar, and Indrabayu,

Detection of Pomelo in Overlapping Conditions Using Drones.

DOI: 10.5220/0014266900004928

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 1st International Conference on Research and Innovations in Information and Engineering Technology (RITECH 2025), pages 25-31

ISBN: 978-989-758-784-9

images of pomelo leaves using Convolutional Neural

Network (CNN) with a 93.8% accuracy rate (Asriny

et al., 2020) (Xu et al., 2023a), detection of leaf-

covered raspberry fruits, which interfere with crop

estimation and garden efficiency, with the limitations

of traditional technology and challenges in the use of

hyperspectral technology using Logistic Regression

(LR), Random Forest (RF) methods (Chen et al.,

2024), and detection of tomato fruit ripeness under

conditions of changing lighting, obstruction, and

fruit overlap, as well as models that are too large for

limited devices using the GFS-YOLOv11 method

obtained results of Precision (P): 92.0%, Recall (R):

86.8%, mAP50: 93.4%, mAP50-95: 83.6% (Wei et

al., 2024).

Currently, most research focuses on improving

detection by color and from different fruit targets.

(Indrabayu et al., 2017) (Janowski et al., 2021),

however, studies related to accurate detection of

green pomelos in gardens have received less

attention because they are complicated and not easy,

especially the detection of green pomelos, which are

very important for predicting garden yields (Lyu et

al., 2022b).

Fruit detection with deep learning can be done

using the computer vision method (Xu et al., 2023b),

You Only Look Once (YOLO) is a single-stage

object detector that has demonstrated excellent

performance for detecting with high precision and

accuracy (Liu et al., 2018), and with the most

developed versions. YOLOv11, the latest model

from Ultralytics, demonstrates superior performance

on a wide range of computer vision tasks, including

object detection, instance segmentation, feature

extraction, pose estimation, object tracking, and

classification, to real-time object detection

(Hidayatullah et al., 2025). Studies to address the

problem of low detection accuracy in complex

plantation environments (such as varying lighting

conditions, branch and leaf occlusion, fruit overlap,

and small targets) used YOLOv11 in detecting

occluded pears and the accuracy results showed

precision, recall, mAP50, and mAP50–95 values of

96.3%, 84.2%, 92.1%, and 80.2%, respectively

[Zhang et al., 2025).

Based on the observations in the pomelo field, the

main obstacle is difficult to distinguish the color of

pomelos and leaves, particularly in conditions of

uneven illumination, overlapping fruits, or leaf

occlusion. This makes manual detection difficult,

time-consuming, and lowers the accuracy of fruit

ripeness classification, which impacts harvest

efficiency. To overcome this problem, a pomelo

detection and classification system was developed

using the YOLOv11 algorithm with the addition of

CLAHE, to support smart farming practices and

improve sustainability of agricultural production.

Beyond model development, the proposed

system can be deployed on drones to enable real-time

monitoring in orchards. Running inference on drones

requires lightweight models that can operate under

limited computational resources. Previous work has

shown that compact models such as YOLOv4-tiny

achieve significantly faster inference while

maintaining satisfactory accuracy, making them

suitable for onboard drone applications (Mpouziotas

et al., 2023). This integration ensures practical field

implementation while enhancing the scalability and

efficiency of precision agriculture systems.

2 MATERIALS AND METHODS

The entire research process, from data preparation to

fruit detection, is illustrated in the flowchart in

Figure 1. This flowchart provides a comprehensive

overview of the methodological workflow

implemented in this study. Here is the workflow

process from 1 to 8.

Figure 1: System Design Flow.

The dataset was partitioned into three subsets,

consisting of 1,301 images for training, 124 images

for validation, and 62 images for testing. However,

the relatively small percentage of validation and

RITECH 2025 - The International Conference on Research and Innovations in Information and Engineering Technology

testing data may introduce bias in the evaluation

results, as limited samples can lead the model to

appear to perform well while its performance may

not generalize to more diverse real-world conditions

(Brigato et al., 2020). The picture above illustrates

the design of the proposed system for detecting

pomelos on trees in challenging conditions where the

color and edges of pomelos are similar to those of

leaves, which can be divided into several main stages

as follows:

2.1 Video Input and Extract Video

Frames

The system starts by receiving video as input. This

video likely contains footage that shows pomelos in

various conditions, such as covered in leaves or

stacked. From the given video, the system extracts

the frame (frame-by-frame extraction). It means that

the video is broken down into several static images

so that it can be further processed in the next stage.

2.2 Preprocessing

After the frame is extracted, a preprocessing stage is

carried out to improve the quality of the data before it

is fed into the detection model. One of the methods

mentioned in the diagram is CLAHE (Contrast

Limited Adaptive Histogram Equalization), which is

useful for improving image contrast, especially in

suboptimal lighting conditions.

In this study, CLAHE was used with the

parameters Clip Limit 2.0 and Tile Grid Size (32, 32).

Clip Limit 2.0 controls the extent to which contrast is

enhanced. This value was chosen because it provides

a significant contrast enhancement without causing

excessive effects such as the appearance of bright

spots (noise amplification). Meanwhile, Tile Grid

Size (32, 32) determines the size of the image division

grid. With this size, the image is divided into small

blocks with sufficient detail so that local contrast in

certain areas can be enhanced. If the grid is too small,

the result can increase noise, while a grid that is too

large actually makes important details less visible.

The combination of these two parameters allows

CLAHE to produce images with clearer details, thus

helping object detection models like YOLO to

recognize oranges more accurately.

(a) (b)

Figure 2: Sample dataset (a) before applying CLAHE and

(b) after applying CLAHE

2.3 Annotations

Data annotation refers to the process of labelling or

adding information to different types of data,

including text, images, and videos. At this stage, the

processed frames are manually or semi-automatically

annotated to mark the pomelo object. This annotation

process is important to properly train the pomelo

detection model.

2.4 Data Splitting

Data splitting is a method to divide a dataset into three

parts: data training, validation, and testing. It is one

of several factors that affects the performance of

model. The proportions used were 70% for the

training set, 20% for the test set and 10% for the

validation set.

2.5 Training Model

The pomelo detection model was trained using

YOLO (You Only Look Once), precisely the version

of YOLOv11 mentioned in the diagram. This

algorithm works by processing an image only once

and immediately generating predictions in the form of

bounding box locations, confidence scores, and

object classes. YOLOv11 introduces significant

improvements to its architecture, such as a more

efficient backbone for feature extraction, a spatial

attention module for capturing details of both small

and large objects, and an anchor-free and decoupled

detection head, resulting in more stable, faster, and

more accurate predictions. These advantages make

YOLOv11 highly effective for detecting oranges

under various conditions, such as when the fruit is

close together, covered by leaves, or in less than ideal

lighting, making it highly suitable for automated fruit

detection and counting tasks.

Detection of Pomelo in Overlapping Conditions Using Drones

Figure 3: YOLOv11 detection architecture.

The YOLOv11 architecture consists of three

main parts: the backbone, the neck, and the head. The

backbone extracts feature from the input image

through convolutional layers and C3K2 blocks,

which efficiently capture visual patterns. Feature

from the backbone are then processed in the neck,

which combines information from multiple scales

with unsampled and concat operations, as well as

SPPF and C2PSA modules to expand the spatial

context and emphasize important areas. The results of

this processing are forwarded to the anchor-free,

decoupled head, resulting in more stable class and

bounding box location predictions. The head feature

three detection branches at different scales, allowing

YOLOv11 to detect small, medium, and large objects

with high accuracy.

Table 1: Comparison of this study with previous studies.

Study / Method Dataset / Target Approach Reported Performance

Anis Ilyana et al., 2025 Coffee YOLO11

Ripe fruit: mAP50 = 77.4%,

Precision = 64.5%, Recall = 81.2%

Half-ripe fruit: mAP50 = 69.5%,

Precision = 62.4%, Recall = 67.9%

Dhungana, P et al., 2025 subsea pipeline

YOLOv8n and

YOLOv11n+CLAHE

YOLOv11n without enhancement:

mIoU = 70.98%, Dice = 81.29%

YOLOv11n with CLAHE: mIoU =

70.48%, Dice = 80.77%

Sapkota et al., 2025 Apple

YOLOv11 and

CBAM

Result Trunk: 83% precision (with

CBAM) vs 80% (without CBAM).

Result Branch: 75% precision (with

CBAM) vs 73% (tanpa CBAM)

Ours Pomelo

YOLOv11 and

CLAHE

Best performance YOLOv11

without CLAHE: Precision = 85%,

Recall = 56%, mAP50 = 72%.

Best performance YOLOv11 with

CLAHE: Precision = 83,8%, Recall

= 78,2%, mAP50 = 85,1%.

Table 2: Configuration of the training model.

Dataset Epoch Precision Recall Loss Train

Loss

Validation

mAP 50 mAP 50-95

YOLOV8

100 0.7200 0.5400 0.8607 0.9366 0.6955 0.5014

200 0.8300 0.5133 0.8755 0.9122 0.7144 0.5288

300 08009 0.4703 0.8679 0.9324 0.6700 0.5000

Baseline

(YOLOV11)

100 0.7353 0.5603 0.9435 0.9558 0.7019 0.5185

200 0.8502 0.5377 0.9072 0.9393 0.7203 0.5387

300 0.8124 0.4889 0.9021 0.9545 0.6740 0.5087

CLAHE +

YOLOV11

100 0.7851 0.7795 1.0578 0.9916 0.8379 0.5626

200 0.8233 0.7821 0.9950 0.9707 0.8471 0.5579

300 0.8385 0.7665 0.9906 0.9677 0.8506 0.5741

RITECH 2025 - The International Conference on Research and Innovations in Information and Engineering Technology

2.6 Evaluation

After the model is trained, the evaluation stage aims

to assess its performance in detecting pomelos. The

assessment is conducted using metrics such as mAP

(mean Average Precision) to measure how well the

model can recognize objects in the image. In addition,

a confusion matrix is employed to calculate precision

and recall, providing a more detailed overview of the

model’s ability to distinguish between correct and

incorrect detections.

2.7 Build Model Testing

The trained models were tested with new videos to

see how well the models could detect pomelos in

more real conditions. These trials aim to measure the

reliability of the model before it is implemented in a

real environment.

3 RESULTS AND DISCUSSION

3.1 Training Perfomance of Yolov11

and (Yolov11 + Clahe)

In this study, the YOLOv11 model was trained with

parameter configurations designed to balance

convergence speed and generalization capabilities.

The initial learning rate was set at 0.01 with a gradual

decrease scheme to maintain weight update stability

during the training process. A momentum parameter

of 0.937 was applied to accelerate convergence by

utilizing the previous gradient direction, while a

weight decay of 0.0005 was used as a regularization

mechanism to reduce the risk of overfitting.

Furthermore, a batch size of 16 was selected

considering the limitations of CPU-based computing,

while the training process was carried out for 300

epochs so that the model had the optimal opportunity

to achieve the best performance. These parameters

were used consistently in both the baseline YOLOv11

model training and the YOLOv11 model with the

CLAHE method to ensure a fair comparison in

performance evaluation.

Table 1 Based on the comparison table, it can be

concluded that each study produced different

performance outcomes due to variations in research

objects, dataset conditions, and applied methods. Anis

Ilyana et al. (2025) demonstrated that the ripeness

level of coffee significantly affects detection accuracy,

with ripe fruits yielding better results than half-ripe

fruits. Dhungana et al. (2025) showed that the

application of CLAHE in subsea pipeline detection

does not always provide benefits, and in fact, slightly

reduced performance compared to results without

CLAHE. Meanwhile, Sapkota et al. (2025) proved that

the addition of the CBAM module improved detection

precision for apples, particularly in more complex

areas such as the trunk and branches.

In this study, the use of CLAHE with YOLOv11

showed a significant improvement in recall and

mAP50 for pomelo detection. This indicates that

CLAHE helps the model identify more objects that

might otherwise be difficult to detect, albeit with a

slight compromise in prediction precision. Overall,

this comparison highlights that the effectiveness of

additional methods such as CLAHE or CBAM largely

depends on the type of object, image quality, and

dataset characteristics.

The table 2 shows the results of the evaluation of

the YOLOv11 model on two types of datasets: original

and those that have been processed using CLAHE.

The model with the original dataset yielded the highest

precision off 0.8502 and the 50th mAP of 0.72032 at

the 200th epoch, but the recall was low (0.5377),

indicating many undetected objects.

.In contrast, the use of CLAHE significantly

improves performance. Precision and recall were

more stable, reaching 0.8385 and 0.7665, respectively,

on the 300th epoch. The values of mAP 50 and mAP

50-95 were also higher, at 0.8506 and 0.5741. These

results prove that CLAHE preprocessing helps

improve the detection of low- contrast objects, making

the model more accurate and reliable

Figure 4: Yolov11 training results graph (mAP 50, mAP 50-

95, precision, and recall) with 300 epochs.

Detection of Pomelo in Overlapping Conditions Using Drones

Figure 5: Graph of Yolov11+ Clahe training results (mAP

50, mAP 50-95, precision, and recall) with 300 epochs.

3.2 Testing Performance of Yolov11

dan (Yolov11 + Clahe)

The table 3 shows the results of the evaluation of the

YOLOv11 model on two types of datasets: original

and those that have been processed using CLAHE.

The model with the original dataset yielded the

highest precision of 0.8502 and the 50th mAP of

0.72032 at the 200th epoch, but the recall was low

(0.5377), indicating many undetected objects.

In contrast, the use of CLAHE significantly

improves performance. Precision and recall were

more stable, reaching 0.8385 and 0.7665 respectively

on the 300th epoch. The values of mAP 50 and mAP

50-95 were also higher, at 0.8506 and 0.5741. These

results prove that CLAHE preprocessing helps

improve the detection of low-contrast objects, making

the model more accurate and reliable.

𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 

𝑇𝑃



𝐹𝑃

(1)

𝑅𝑒𝑐𝑎𝑙𝑙 

𝑇𝑃



𝐹𝑁

(2)

Explanation:

FN = The model fails to detect an object even though

it exists.

TP = The model successfully

detects an object and the object does indeed exist.

TN = The model does not detect objects, and they do

not exist.

FP = The model detects an object when it doesn't

exist.

Table 3: Testing Configuration.

Dataset

Baseline

(

Yolov8

)

Baseline

(

Yolov11

)

Clahe +

Yolov11

TP 315 328 412

FP 16 11 113

FN 145 118 34

Precision 0.921 0.929 0.921

Recall 0.803 0.815 0.841

mAP 50 0.894 0.907 0.936

mAP 50-

0.717

0.721 0.67

4 CONCLUSIONS

This research successfully developed a pomelo

detection system on trees using the YOLOv11

algorithm combined with the CLAHE contrast

enhancement method. The results of the experiment

showed that preprocessing using CLAHE was able to

improve the accuracy of fruit detection, especially in

conditions of overlap and color similarity between

fruits and leaves. The increased mAP, precision, and

recall values after the use of CLAHE prove the

effectiveness of this approach in dealing with

challenges in complex plantation environments. This

system can be a supporting solution in the

implementation of smart farming, especially to

improve the efficiency and accuracy of the automatic

pomelo harvesting process.

ACKNOWLEDGEMENTS

The authors would like to thank the Master’s Program

in Informatics Engineering, Hasanuddin University,

and the Artificial Intelligent and Multimedia

Processing (AIMP) Thematic Research Group for

their support and facilities during this research. The

authors also acknowledge colleagues who provided

constructive feedback and assistance throughout the

research process. In addition, the authors

acknowledge the use of a generative AI tool in

improving the clarity and grammar of this

manuscript. All contents, analyses, and conclusions

remain the sole responsibility of the authors.

RITECH 2025 - The International Conference on Research and Innovations in Information and Engineering Technology

REFERENCES

Addi, A. Elbouzidi, M. Abid, D. Tungmunnithum, A.

Elamrani, and C. Hano, “An Overview of Bioactive

Flavonoids from Pomelos,” Appl. Sci., vol. 12, no. 1, p.

29, Dec. 2021, doi: 10.3390/app12010029.

Adelina and E. Adelina, “Identification of Morphology and

Anatomy of Local Pomelos (Citrus sp) in Doda Village

and Lempe Village, Central Lore District, Poso

Regency.”

Asriny, S. Rani, and A. F. Hidayatullah, “Pomelo Fruit

Images Classification using Convolutional Neural

Networks,” IOP Conf. Ser. Mater. Sci. Eng., vol. 803,

no. 1, p. 012020, Apr. 2020, doi: 10.1088/1757-

899x/803/1/012020.

Brigato, L., & Iocchi, L. (2021, January). A close look at

deep learning with small data. In 2020 25th

international conference on pattern recognition

(ICPR) (pp. 2490-2497). IEEE.

Chen, J. Wang, R. Xi, and Z. Ren, “Analysis of Leaf cover

on Raspberry Fruits Based on Hyperspectral

Techniques Combined with Machine Learning

Models,” Jul. 15, 2024, Springer Science and Business

Media LLC. doi: 10.21203/rs.3.rs-4607290/v1.

Dhungana, P., Fresta, M., Tamrakar, N., & Dhungana, H.

(2025, Juni 30). YOLO-Based Pipeline Monitoring in

Challenging Visual Environments

(arXiv:2507.02967v1).

https://doi.org/10.48550/arXiv.2507.02967

Hidayatullah, N. Syakrani, M. R. Sholahuddin, T. Gelar,

and R. Tubagus, “YOLOv8 to YOLO11: A

Comprehensive Architecture In-depth Comparative

Review.”

Ilyana, A., Nurdin, N., & Maryana, M. (2025). Real-Time

Detection of Coffee Cherry Ripeness Using YOLOv11.

Journal of Applied Informatics and Computing, 9(4).

https://doi.org/10.30871/jaic.v9i4.9735.

Indrabayu, Mar’atuttahirah, and I. S. Areni, “Automatic

Counting of Chili Ripeness on Computer Vision for

Industri 4.0,” in 2019 IEEE International Conference

on Industry 4.0, Artificial Intelligence, and

Communications Technology (IAICT), BALI,

Indonesia: IEEE, Jul. 2019, pp. 14–18, doi:

10.1109/iciaict.2019.8784858.

Indrabayu, A. R. Fatmasari, and I. Nurtanio, “A Colour

Space Based Detection for Cervical Cancer Using

Fuzzy C-Means Clustering,” in Proceedings of the 6th

International Conference on Bioinformatics and

Biomedical Science, Singapore: ACM, Jun. 2017, pp.

137–141, doi: 10.1145/3121138.3121196.

Janowski, R. Kaźmierczak, C. Kowalczyk, and J. Szulwic,

“Detecting Apples in the Wild: Potential for Harvest

Quantity Estimation,” Sustainability, vol. 13, no. 14, p.

8054, Jul. 2021, doi: 10.3390/su13148054.

Khattak et al., “Automatic Detection of Pomelo and Leaves

Diseases Using Deep Neural Network Model,” IEEE

Access, vol. 9, pp. 112942–112954, 2021, doi:

10.1109/access.2021.3096895.

Liu, Y. Tao, J. Liang, K. Li, and Y. Chen, “Object Detection

Based on YOLO Network.”

Lyu, R. Li, Y. Zhao, Z. Li, R. Fan, and S. Liu, “Green Citrus

Detection and Counting in Orchards Based on

YOLOv5-CS and AI Edge System,” Sensors, vol. 22,

no. 2, p. 576, Jan. 2022, doi: 10.3390/s22020576.

Mpouziotas, D., Karvelis, P., Tsoulos, I., & Stylios, C.

(2023). Automated wildlife bird detection from drone

footage using computer vision techniques. Applied

Sciences, 13(13), 7787.

Pathak, H. Gangwar, and A. S. Jalal, “Performance

Analysis of Gradient Descent Methods for

Classification of Pomelos using Deep Neural

Network,” in 2020 7th International Conference on

Computing for Sustainable Global Development

(INDIACom), New Delhi, India: IEEE, Mar. 2020, pp.

68–72, doi: 10.23919/indiacom49435.2020.9083723.

Sapkota, R., & Karkee, M. (2025, Januari 26). Comparing

YOLOv11 and YOLOv8 for instance segmentation of

occluded and non-occluded immature green fruits in

complex orchard environment (arXiv:2410.19869v3).

https://doi.org/10.48550/arXiv.2410.19869

Sugadev, K. Sucharitha, I. R. Sheeba, and B. Velan,

“Computer vision based automated billing system for

fruit stores,” in 2020 Third International Conference on

Smart Systems and Inventive Technology (ICSSIT),

Tirunelveli, India: IEEE, Aug. 2020, pp. 1337–1342,

doi: 10.1109/icssit48917.2020.9214101.

Wajid, N. K. Singh, P. Junjun, and M. A. Mughal,

“Recognition of Ripe, Unripe and Scaled Condition of

Pomelo Citrus Based on Decision Tree Classification.”

Wei et al., “GFS-YOLO11: A Maturity Detection Model

for Multi-Variety Tomato,” Agronomy, vol. 14, no. 11,

p. 2644, Nov. 2024, doi: 10.3390/agronomy14112644.

Xu, H. Zhao, O. M. Lawal, X. Lu, R. Ren, and S. Zhang,

“An Automatic Jujube Fruit Detection and Ripeness

Inspection Method in the Natural Environment,”

Agronomy, vol. 13, no. 2, p. 451, Feb. 2023, doi:

10.3390/agronomy13020451.

Zhang, S. Ye, S. Zhao, W. Wang, and C. Xie, “Pear Object

Detection in Complex sOrchard Environment Based on

Improved YOLO11,” Symmetry, vol. 17, no. 2, p. 255,

Feb. 2025, doi: 10.3390/sym17020255.

Detection of Pomelo in Overlapping Conditions Using Drones