=Paper=
{{Paper
|id=Vol-3736/paper14
|storemode=property
|title=Method of early landfill fire detection using the YOLOv8 neural network
|pdfUrl=https://ceur-ws.org/Vol-3736/paper14.pdf
|volume=Vol-3736
|authors=Olga Pavlova,Tetiana Hovorushchenko,Andrii Kuzmin,Tymur Isayev,Houda EL Bouhissi
|dblpUrl=https://dblp.org/rec/conf/icyberphys/PavlovaHKIB24
}}
==Method of early landfill fire detection using the YOLOv8 neural network==
https://ceur-ws.org/Vol-3736/paper14.pdf
Method of early landfill fire detection using the
YOLOv8 neural network⋆
Olga Pavlova1† *, Tetiana Hovorushchenko†, Andrii Kuzmin†, Tymur Isayev† and
Houda EL Bouhissi2†
1 Khmelnytskyi National University, Instytuts’ka str., 11, Khmelnytskyi, 29016, Ukraine
2 LIMED Laboratory, Faculty of Exact Sciences,University of Bejaia, 06000, Bejaia, Algeria
Abstract
In addressing the challenge of early landfill fire detection using external surveillance cameras, this
study proposes a method leveraging the YOLOv8 neural network model. An experiment was
conducted utilizing the YOLOv8-s variant trained on the Roboflow dataset, with performance
evaluated through precision, recall, and F1 score metrics. The results demonstrated an average
precision of 0.93707, indicating that the model correctly identifies objects 93.7% of the time, thereby
maintaining a low false positive rate. Additionally, the model achieved a high recall of 0.9061,
successfully detecting 90.6% of actual objects and exhibiting a low false negative rate. The F1 score
was 0.9213, reflecting a balanced trade-off between precision and recall. These metrics collectively
suggest that the YOLOv8 model is both accurate and robust, making it a reliable tool for early
detection tasks. The promising results underscore the model's potential for real-world applications
where high accuracy and reliability are essential. Future work will focus on developing an
information system for early landfill fire detection based on the method proposed in this study.
Keywords
Image processing, landfill fire detection, neural networks, YOLOv8
1. Introduction
Fires in solid waste landfills are a significant hazard due to the large impact on the environment,
health and safety and can even cause loss of life. During a fire in Hrybovychi, Lviv region,
Ukraine [1-3] in 2016, three rescuers died, and a significant amount of harmful substances were
released into the air. Another large-scale fire occurred in 2023 on the plain [4]. However, this
time, fortunately, the death of people was avoided.
Fires at solid waste landfills are among the most complex and long-lasting, extinguishing
which requires the involvement of significant resources, efforts, means and time. Forecasting
and prevention of fires at landfills is extremely complicated, as it is difficult to determine
possible centers of temperature increase due to different specific heat capacities of waste. Until
ICyberPhyS-2024: 1st International Workshop on Intelligent & CyberPhysical Systems, June 28, 2024, Khmelnytskyi, Ukraine
∗ Corresponding author.
† These authors contributed equally.
pavlovao@khmnu.edu.ua (O. Pavlova); tat_yana@ukr.net (T.Hovorushchenko); andriy1731@gmail.com(A.Kuzmin);
tymuri@gmail.com (T. Isayev); houda.elbouhissi@gmail.com (H. El Bouhissi)
0000-0003-2905-0215 (O. Pavlova); 0000-0002-7942-1857(T.Hovorushchenko); 0009-0005-6489-225X (A.Kuzmin); 0000-0003-
3239-8255 (H. El Bouhissi)
© 2023 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
CEUR
ceur-ws.org
Workshop ISSN 1613-0073
Proceedings
the fire or smoke has reached the surface, it is almost impossible to detect the source of ignition
visually. Fires mainly occur during the fire-hazardous period in the summer. The main cause of
fires remains human imprudence and carelessness, neglect of fire safety rules, careless handling
of fire, violation of technological regulations for disposal of solid household waste. Landfills are
capable of spontaneous combustion. The process is caused by the biochemical decomposition
of waste, which is accompanied by an increase in temperature to 40-70 °C.
The main causes of landfill fires include:
1) Decomposition heat - organic materials in landfills decompose, producing heat. When this
heat accumulates without adequate ventilation, it can ignite surrounding waste.
2) Chemical reactions - certain chemicals and materials can react exothermically, generating
heat and potentially leading to combustion.
3) Human activities - careless disposal of lit cigarettes, fireworks, or other flammable items
by individuals can start fires.
4) Arson - intentional setting of fires, whether for malicious purposes or to reduce waste
volume.
5) Faulty equipment - malfunctioning machinery or exposed electrical wiring can create
sparks that ignite flammable materials.
6) Overheating - overloaded electrical systems in or near the landfill can overheat and cause
fires.
7) Gas build-up - methane, produced by the anaerobic decomposition of organic waste, is
highly flammable. Without proper gas collection systems, methane can accumulate and
ignite.
The consequences of the fires that happen on landfills are the following:
1) Air pollution - burning waste releases toxic smoke and harmful pollutants such as dioxins,
furans, and particulate matter, contributing to air quality degradation and respiratory
health issues.
2) Greenhouse gas emissions - fires release large amounts of CO2 and other greenhouse
gases, exacerbating climate change.
3) Soil and water contamination - ash and residues from fires can seep into the soil and
groundwater, contaminating local water supplies and harming ecosystems.
4) Health risks - respiratory problems: exposure to smoke and toxic fumes can cause acute
respiratory issues, exacerbate chronic conditions like asthma, and increase the risk of
long-term health problems such as lung cancer.
5) Chemical exposure - communities near landfills are at risk of exposure to hazardous
chemicals released during fires, which can lead to various health complications.
6) Firefighting and cleanup -extinguishing landfill fires and managing the aftermath require
significant financial resources, manpower, and time.
7)Property damage - fires can spread to nearby areas, damaging infrastructure, properties,
and agricultural land, resulting in economic losses.
8)Landfill closure - ongoing fires can force landfills to temporarily close, disrupting waste
management services and creating additional public health and environmental
challenges.
9)Rehabilitation costs - post-fire site rehabilitation to ensure the area is safe for future use
can be expensive and time-consuming.
10) Public safety concerns - evacuations: severe landfill fires may necessitate the evacuation
of nearby communities to protect residents from harmful smoke and fumes.
11) Long-term exposure risks - prolonged exposure to pollutants from recurring landfill fires
can lead to chronic health issues for local populations.
This research focuses on the Sustainable Development Goals (SDGs)[5] established by the
United Nations (UN) and adopted by all UN Member States in 2015. Specifically, it addresses
SDG 12: Responsible Consumption and Production, and SDG 15: Life on Land. Landfill fires
pose significant and undeniable harm to all living beings on Earth [6], therefore the problem of
landfill fires early detection is quite relevant for faster disposal and prevention of emissions of
harmful substances into the air and human casualties.
2. Related works
During the study, an analysis of the most recent scientific publications on early landfill fire
detection was conducted. The research [7] helps to detect the gases emitted from garbage, with
the project's importance increasing due to the rise in solid waste. The purpose of [8] is to
highlight the issue of landfill fires and their effects on air, soil, and water, drawing from a review
of documented fires and fire indicators in the areas where the authors conducted their research.
The goal of the study [9] was to determine the exact temperature at which each waste kind
ignites and maintains smoldering, which would aid in proactive fire prevention and effective
waste management. The study [4] compares the output and underlying assumptions of each
model and suggests the FODM and LandGEM SP simulations can be suitable for estimating
methane emissions in the conditions of Khulna. The study [5] concentrates on developing a
methodology to create 3D thermal models by projecting TIR image data onto a 3D model
generated from RGB images and identifying thermal anomalies using photointerpretation and
GIS analysis. This source [6] proposed a comprehensive method for PFAS screening in leachate
samples using suspect and nontarget analysis. The study [7] presents a developed methane
emission model which can be replicated globally. The research [8] focuses on managing leachate
by recirculating the nutrient-rich fluid back into the landfills, transforming them into
bioreactors. This approach maximizes landfill performance parameters, making them more
efficient for electricity production in waste-to-energy plants. The paper [9] presents an
experimental investigation into the physio-chemical properties of landfills by recirculating
leachate to achieve sustainable performance characteristics in landfill models. The study [10]
aims to explore various aspects of mine fire data using CatBoost and LightGBM methods to
reduce human fatalities and material losses during the construction of deep underground
engineering projects. The study [11] highlights effective solid waste management practices and
discusses ways to manage it sustainably through resource recovery. The work [12] aims to
predict the fire danger rating of underground mining production processes by applying
advanced unsupervised and supervised machine learning techniques. The study [13] introduces
a transparent decision-making framework for landfill site selection that combines multi-criteria
decision making, fuzzy set theory, GIS and eXplainable Artificial Intelligence (XAI). This study
[14] includes an analysis of various machine learning (ML) algorithms for municipal solid waste
management to improve procedures and mitigate adverse environmental impacts. The paper
[15] discusses an innovative application of dioxin-like persistent organic pollutants (dl-POPs)
emission trends as a measure of environmental performance for designing effective municipal
solid waste management (MSWM) schemes.
The publications analyzed above consider the probability of occurance fires on the landfills
and estimate the emissions of harmful substances into the air and the amount of air and land
pollution caused by the degradation of household waste on the landfills. These studies have
merely theoretical background and do not propose early fire detecting methods.
The works [22], [23] propose the application of machine vision, namely pattern recognition
using artificial neural networks for finding free parking spaces. In [24], the application of the
YOLOv8 model is proposed for automating the security of warehouses using images from
outdoor surveillance cameras.
Therefore, taking into account the abovementioned analysis and based on the comparative
analysis of machine learning methods and technologies in [23] and [24], it was decided to apply
the YOLOv8 artificial neural network model for learning fire and smoke recognition for the
purpose of early detection of fires in landfills.
3. Dataset preparation and model selection for the implementation
For the early detection of landfills fires, a method of recognizing smoke and fire images from
outdoor surveillance cameras using machine vision was chosen. Since the task before us is the
task of image classification, that is, trying to predict whether there is smoke or fire in the image
or not. For the task of classification, the supervised learning method is best suited. Supervised
learning is an approach to machine learning defined by using labeled datasets to train
algorithms for data classification and outcome prediction.
A labeled data set has an output labeled with tags corresponding to the input data so that
the machine can understand what to look for in the unseen data. The working principle of the
supervised learning method is presented in Figure 1.
Data preparation is one of the most important steps in the neural network training process
[27]. The accuracy and efficiency of the model depends on training data quality. Especially, at
the stage of data collection, attention should be paid to:
- Volume of data. Typically, the more data used, the better the model can be trained;
- Data quality. The data must be accurate, complete and representative of the problem to be
solved;
- Variety of data. It is important that the data set contains examples from all possible
categories or classes that the model will need to recognize.
Neural network developers often list their guidelines for dataset preparation. In particular,
the developers of the YOLO neural network provide the following recommendations [28]
regarding data preparation:
- number of images per class: >=1500.
- number of copies (objects with labels) per class: >=10000.
- background images: 0-10% of the total amount.
Figure 1: The working principle of the supervised learning method [25].
In addition to quantitative recommendations, there are others, in particular:
- Variety of images. For cases of real use, it is recommended to use images obtained at
different times of the day, different seasons, different weather, different lighting,
different angles, from different sources (cameras), etc.;
- Marking sequence. Every occurrence of every class in all images needs to be labeled.
Incomplete labeling is not sufficient;
- Marking accuracy. Labels should tightly cover each object. There should be no gaps
between the object and its bounding box.
The prepared data should be divided into 3 groups:
1) Training set - the dataset used to train the model.
2) Verification (validation) set - the dataset is utilized to assess the model's performance
throughout the training process.
3) Test set - the data set used to finalize the model's performance after training is complete.
The ratio between these sets is determined as follows: as a rule, 10-20% of the total volume of
prepared data is allocated to the validation and test sets, and 80-60% to the training set,
respectively. These parameters can vary, even go beyond the mentioned limits, depending on
the total size of the prepared data, the complexity of the model, the used training
hyperparameters, as well as the subject area of use of the trained model.
Since this study pursues a specific goal, namely the detection of ignition in landfills, the
Roboflow dataset [26] was chosen for training the neural network. It contains 9686 images with
two classes - smoke and fire, and split for training, validation and testing.
An example of an image from the Roboflow dataset is presented in Figure 2.
Other images from this dataset that were used for the neural network model training are
presented in Figure 3.
Figure 2: An example of an image with “smoke” and “fire” classed from the Roboflow dataset
[26].
Figure 3: An example of the the images for neural network model training from the Roboflow
dataset [26].
4. Method of early landfill fire detection using the YOLOv8 neural
network
Method of early landfill fire detection using the YOLOv8 neural network consists of the
following steps:
1. Preparation and training of a neural network model on a dataset with "smoke" and "fire"
classification.
2. Validation of the neural network model on real images of fires at landfills.
3. Automatic detection of fires at landfills in videos from outdoor surveillance cameras and
saving images using the existing YOLOv8 model using Python program.
4. Manual verification of recognition results, separation of cases of false detection of fire or
smoke.
5. Evaluation of the automated detection quality.
6. Manual labeling of incorrectly processed images.
The method is visually represented in Figure 4.
Figure 4: Visual representation of Method of early landfill fire detection using the YOLOv8
neural network.
5. Experiments and Results
To conduct the experiments, the YOLOv8 neural network model was trained on the Roboflow
dataset for 125 epochs.
The training results are presented in Table 1.
The table shows that already at the 125th epoch, the train/box_loss and train/dfl_loss
indicators begin to decrease, so further training is not effective.
We have reached the maximum indicators.
Evaluation of training results was carried out in 2 ways: using the obtained metrics which is
presented in Figures 5-8, and manually (saving data from the video using the newly created
model, manually searching for erroneous results).
Table 1
Results of YOLOv8 training on Roboflow “smoke” and “fire” dataset
Epoch train trai metrics metrics/ metrics metrics val/b val/cl val/dfl
s /box n/dfl /precisi recall(B) /mAP5 /mAP5 ox_lo s_los _loss lr/pg0 lr/pg1 lr/pg2
_loss _los on(B) 0(B) 0-95(B) ss s
s
1.55 2.15 0.7260 0.610 1.037 0.964 0.0033
1 35 9 1.617 0.77306 5 0.7791 3 7 71 1.3708 0.003326 26
1.39 1.46 0.7507 0.8322 0.663 1.004 0.997 0.0066
2 99 12 1.4798 0.84439 5 6 28 1 59 1.3348 0.006607 07
1.43 1.52 0.7758 0.8288 0.654 1.029 1.043 0.0098
3 72 41 1.512 0.8035 3 5 68 5 3 1.3554 0.009835 35
1.50 1.61 0.7705 0.8286 0.648 1.029 1.053 0.0097
4 73 59 1.5615 0.778 8 9 1 8 1 1.3595 0.009762 62
1.46 1.53 0.8079 0.8598 0.691 0.948 0.871 0.0096
5 77 33 1.5346 0.86732 8 5 35 68 23 1.307 0.009683 83
1.44 1.47 0.8606 0.696 0.977 0.942 0.0096
6 3 98 1.5262 0.83168 0.8007 9 06 22 86 1.3246 0.009604 04
1.41 1.41 0.8409 0.8663 0.717 0.900 0.803 0.0095
7 4 46 1.5109 0.87346 3 7 64 3 14 1.2951 0.009525 25
1.40 1.38 0.8805 0.730 0.888 0.801 0.0094
8 14 7 1.4966 0.89735 0.8526 5 37 59 01 1.2676 0.009446 46
1.37 1.34 0.8545 0.8972 0.749 0.850 0.732 0.0093
9 41 01 1.4791 0.90382 4 4 6 62 99 1.2328 0.009366 66
1.33 1.30 0.8388 0.8876 0.753 0.852 0.705 0.0092
10 15 2 1.459 0.89978 3 1 38 26 44 1.2372 0.009287 87
… … … ... … … … … … … … …. …
0.69 0.38 0.8893 0.9210 0.841 0.575 0.404 0.0003
123 156 405 1.005 0.90633 6 4 59 5 56 1.0956 0.000338 38
0.68 0.37 0.8979 0.9211 0.842 0.576 0.404 0.0002
124 591 955 1.0044 0.89673 8 5 08 13 7 1.0963 0.000258 58
0.68 0.37 0.8980 0.9211 0.841 0.575 0.405 0.0001
125 454 691 1.0016 0.89841 2 7 77 85 65 1.096 0.000179 79
Figure 5: Metrics for training results evaluation.
Figure 6: Precision-Confidence Curve.
Figure 7: Recall-Confidence Curve.
Figure 8: Precision-Recall Curve.
Figure 9 presents confusion matrix in numerical (a) and in percentage (b) ways. The results
of images from Roboflow dataset verification in which the neural network detected smoke and
fire are presented in Figure 10. The results of real-life landfill fire images verification are
presented in Figure 11.
a)
b)
Figure 9: Confusion matrix in numerical (a) and in percentage (b) representations.
Figure 10: The results of images from Roboflow dataset verification in which the neural
network detected smoke and fire.
The errors of computer vision algorithms are characterized by such parameters as Precision,
Accuracy, Recall, F1 score.
Precision shows how many of the positive predictions turned out to be correct. The average
precision value for our case was calculated by the formula 1.
Precision = TP / (TP + FP) = 0.93707, (1)
TP refers to the count of correctly classified positive examples, while TN denotes the count
of correctly classified negative examples.
Figure 11: The results of real-life landfill fire images verification.
Recall represents the ratio of positive cases correctly identified by the model. The average
recall value for our case was calculated by the formula 2.
Recall = TP / (TP + FN) = 0.9061, (2)
where TP is the number of correctly classified positive examples; TN is the number of correctly
classified negative examples.
F1 score is a metric for measuring model performance in classification tasks. It combines
precision and response into one metric to provide a balanced assessment of model accuracy. F1
score for our case was calculated by the formula 3:
F1 = 2 * (Precision * Recall) / (Precision + Recall) = 0,9213. (3)
6. Conclusions
Therefore, in the course of work on the problem of early landfill fire detection using external
surveillance cameras, it was decided to develop a method of early detection of landfill fires using
YOLOv8 neural network model.
Also, an experiment on YOLOv8-s training was conducted using Roboflow dataset. The
results of the experiment were measured by such metrics as precision, recall and F1 score. The
average precision value that was obtained is 0,93707 is quite high, meaning that when the model
predicts an object, it is correct 93.7% of the time. This indicates that the model has a low false
positive rate and is good at avoiding incorrect predictions. The model is effective at identifying
true positives among the positive predictions it makes. The model has a high recall (0,9061),
meaning it correctly identifies 90.6% of all actual objects. This indicates that the model has a
relatively low false negative rate and is good at finding most of the actual objects present in the
dataset. The model is effective at capturing the majority of true objects in the data. The F1 score,
which is the harmonic mean of precision and recall, is 0.9213. This value is high, indicating a
good balance between precision and recall. The model maintains a good trade-off between
precision and recall, making it robust for the task at hand. The high precision and recall values,
along with a high F1 score, suggest that the neural network model is performing well and is
balanced in terms of making accurate predictions (high precision) and capturing most of the
relevant objects (high recall). The metrics indicate that the model has been trained effectively
on the given dataset. It is capable of making reliable and comprehensive detections or
classifications. Overall, these metrics suggest that the model is highly accurate and has been
trained successfully, making it a strong candidate for deployment in applications where high
precision and recall are critical.
The future efforts of the authors will focus on the development of the information system
for early landfill fire detection based on the proposed in this work method.
References
[1] T. Kovalenko, V. Mykhaylenko, Anthropogenic impact on the environment caused by fire at
Lviv’s municipal solid waste landfill. Bulletin of Taras Shevchenko National University of
Kyiv. Geography. (2019) 78-86. 10.17721/1728-2721.2019.75.13.
[2] The ecological crisis in Lviv. URL: https://www.osw.waw.pl/en/publikacje/analyses/2016-06-
15/ecological-crisis-lviv (Last accessed May 17).
[3] V. Mykhaylenko, et al. Acquiring practice in environmental and social impact assessment:
case study of Lviv city dumpsite, Ukraine. 3 (2021) 154-167.
https://doi.org/10.23939/ep2021.03.154.
[4] There was a fire at garbage dump near Rivne. URL: https://ecopolitic.com.ua/en/news/bilya-
rivnogo-stalasya-pozhezha-na-smittiezvalishhi-2/.
[5] THE 17 GOALS. Sustainable Development. URL: https://sdgs.un.org/goals.
[6] United Nations. Transforming our world: the 2030 Agenda for Sustainable Development.
URL: https://sdgs.un.org/2030agenda (Last accessed March 29, 2024).
[7] S. GayatrI, et al. An Intelligent IoT Based Landfill Fire Prediction and Prevention System
(2023).
[8] D. Dabrowska, R. Wojciech, V. Nourani, Causes, types and consequences of municipal waste
landfill fires—Literature review. Sustainability 15.7 (2023).
[9] M. Islam, M. Saptarshi, S. Salahuddin, Determination of smoldering and ignition
Temperature of MSW and Understanding the Landfill Fire in an Open Dump Site in Khulna
Region of Bangladesh (2024).
[10] S. Salahuddin, et al, Prediction of methane emissions and electrical energy generation
potential from MSW landfill in Khulna city of Bangladesh: a model-based approach. Clean
Technologies and Environmental Policy (2023) 1-19.
[11] J. Sedano-Cibrián et al., 4D Models Generated with UAV Photogrammetry for Landfill
Monitoring Thermal Control of Municipal Solid Waste (MSW) Landfills. Applied Sciences
13.24. 2023. 13164.
[12] C. Feng et al, Suspect, Nontarget Screening, and Toxicity Prediction of Per-and
Polyfluoroalkyl Substances in the Landfill Leachate. Environmental Science & Technology
(2024).
[13] A. Rafey, Z. S. Faisal, Modelling and simulation of landfill methane model. Cleaner Energy
Systems 5. 2023. 100076.
[14] O. Khan et al. Experimental investigation and multi-performance optimization of the
leachate recirculation based sustainable landfills using Taguchi approach and an integrated
MCDM method. Scientific Reports 13.1. 2023. 19102.
[15] O. Khan et al. Waste-to-Energy Power Plants: Multi-objective Analysis and Optimization of
Landfill Heat and Methane Gas by Recirculation of Leachate. Process Safety and
Environmental Protection (2024).
[16] M. Kamran et al. Intelligent based decision-making strategy to predict fire intensity in
subsurface engineering environments. Process Safety and Environmental Protection 171
(2023) pp.374-384.
[17] S. Mor and R.Khaiwal, Municipal solid waste landfills in lower-and middle-income countries:
Environmental impacts, challenges and sustainable management practices. Process Safety
and Environmental Protection (2023).
[18] M. Kamran et al. A multi-criteria decision intelligence framework to predict fire danger
ratings in underground engineering structures. Fire 6.11 (2023) 412.
[19] M. Al Awadh, and M. Javed, A decision-making framework for landfill site selection in Saudi
Arabia using explainable artificial intelligence and multi-criteria analysis. Environmental
Technology & Innovation 33 (2024) 103464.
[20] K. S. Pillai et al, Municipal Solid Waste Management: A Review of Machine Learning
Applications. E3S Web of Conferences. Vol. 455. EDP Sciences (2023).
[21] S. V. Ajay et al, Dioxin-like POPs emission trends as a decision support tool for developing
sustainable MSW management scheme–an exploratory study. Journal of Environmental
Management 328 (2023) 117004.
[22] P. Radiuk, O. Pavlova, H. El Bouhissi, V. Avsiyevych, V. Kovalenko, Convolutional Neural
Network for Parking Slots Detection. CEUR-WS, 3156 (2022) pp. 284–293.
[23] O. Pavlova, V. Kovalenko, T. Hovorushchenko, Neural network-based image recognition
method for smart parking. Comput. Syst. Inf. Technol. J. 1 (2021) pp. 49–55.
[24] O. Pavlova, I. Rudyk, H. Bouhissi, Post-processing of video surveillance systems alarm signals
using the YOLOv8 neural network. CEUR Workshop Proceedings. 3675 (2024) 196–207.
[25] Supervised and Unsupervised Learning [Differences & Examples] URL:
https://www.v7labs.com/blog/supervised-vs-unsupervised-learning.
[26] Roboflow. Fire2 Object Detection Dataset by TAEIN. URL:
https://universe.roboflow.com/taein-ibves/fire2-kiqel.
[27] E. M. Cherrat, R. Alaoui, H. Bouzahir, Score fusion of finger vein and face for human
recognition based on convolutional neural network model. International Journal of
Computing, 19(1) (2020) 11-19. https://doi.org/10.47839/ijc.19.1.1688.
[28] Tips for Best Training Result URL:
https://docs.ultralytics.com/yolov5/tutorials/tips_for_best_training_results/.
[29] Model Training with Ultralytics YOLO. URL:https://docs.ultralytics.com/modes/train/.