Deep learning-based detection of power transmission lines using YOLOv4 and YOLOv8 Authors: Tugce Nur Karadeniz, Sami Ekici, Engin Avci Published: 2025 ABSTRACT AbstractReliable energy distribution is considered a cornerstone of modern society, and the maintenance of power transmission lines (PTLs) plays a vital role in achieving this goal. The accurate and rapid detection of PTLs not only enhances the efficiency of energy distribution networks but also is crucial in preventing potential failures and accidents. Traditional inspection methods often rely on the human eye, leading to time-consuming, costly, and risky processes. In this context, the emergence of unmanned aerial vehicle (UAV) technology has brought about a revolutionary change in power transmission line inspections. UAVs offer unprecedented efficiency, safety, and cost-effectiveness in monitoring energy infrastructure. The integration of deep learning-based object detection methods with UAV technology enables safe and effective monitoring of power transmission lines. Particularly, the You Only Look Once (YOLO) algorithm facilitates faster and more accurate results by combining object detection and classification processes within a single network. This study comprehensively examines the object detection applications of the YOLO algorithm. The results indicate that all versions of YOLOv8 outperform YOLOv4 in terms of object detection, demonstrating high performance. The precision, recall, and F1 score performance metrics for the tested YOLOv8 versions were found to be over 99%. The average mAP50 and mAP50-95 values were found to be 0.995 and 0.919, respectively.Keywords: Power transmission lines, Unmanned aerial vehicles, Deep learning, YOLOv8Subject terms: Energy science and technology, Electrical and electronic engineering Introduction Power transmission lines (PTLs) are critical for ensuring the maintenance and operation of reliable energy distribution systems, particularly as societies become increasingly dependent on electricity1. These lines are the infrastructure systems that enable the transmission of electrical energy from production sources to transformer stations and final consumers (homes, workplaces, etc.). Energy transmission lines ensure the safe and efficient transport of electrical energy at various voltage levels. Electricity is transported from power generation stations to consumption points via high and ultra-high voltage transmission lines, which stretch thousands of kilometers across mountains, forests, and hard-to-reach areas. Given the crucial role of the transmission system, system operators place great importance on maintenance activities to ensure grid availability1. Traditional methods for inspecting power transmission lines have shown little change over the years. These methods rely on visual inspections performed by human crews from the ground or via helicopters2. Visual inspections of transmission lines are often conducted by teams of two inspectors on foot. Inspectors use binoculars and sometimes infrared cameras to check the condition of the lines. This method is slow and physically demanding. In rugged terrain, under poor weather conditions, or during extreme weather events such as hurricanes and snowstorms, this approach becomes even more difficult3. While helicopter-assisted inspections offer advantages in terms of reaching hard-to-access areas and increasing the speed of inspection, they also have numerous drawbacks, including high costs, reduced accuracy due to speed, and the risk of contact with live lines. Furthermore, the detection rates of these methods are limited by the visual observation capabilities of the inspectors2,3. Recently, as shown in Fig. 1, alternative technologies such as UAVs and climbing robots have been tested and implemented for the inspection of power transmission lines4,5. UAVs are defined as technology that can fly without a human pilot onboard and remain airborne autonomously. Smart inspection systems offer 2.5 times greater efficiency compared to traditional methods, saving significant time for grid maintenance personnel and reducing their workload, thus increasing operational efficiency. Moreover, these systems stand out due to their ability to perform critical tasks efficiently without risking human lives while offering lower operational costs compared to manned equivalents2. UAVs, with their environmentally friendly features, present promising solutions in this sector. These systems are easily and quickly deployable, scalable, and flexible with their innovative designs. Additionally, they feature capabilities such as self-organization, cost-effectiveness, and high maneuverability. UAVs are diverse in terms of the payloads they can carry, which may include communication devices, navigation equipment, sensors, cameras, and artificial intelligence (AI) technologies. They also vary widely in terms of configuration, size, range, weight, motor type, and performance characteristics. UAVs can carry various equipment such as communication devices, navigation systems, sensors, and cameras6,7. Power transmission line monitoring scheme. UAVs can be produced much more cheaply compared to manned aircraft due to their smaller and lighter frames. Unmanned flights make UAV accidents more financially and socially acceptable, encouraging the execution of tasks that were previously deemed impossible. Additionally, the fact that UAVs have a lifespan of over 10 years and can be operated at low costs with minimal maintenance required provides significant advantages4. Visual inspections primarily focus on examining the condition of structural components such as transmission lines, conductors, and towers. Additionally, detecting contamination or physical damage to power insulators is a key application. Visual inspection methods are also widely used to detect violations near tower structures or on the ground. In offline applications, UAVs can record a large number of images of the transmission grid. These images can then be analyzed to detect patterns that indicate fatigue, cracks, rust, corrosion, or other damage. UAVs can also be used for surveillance against potential terrorist attacks on transmission lines8. This new approach aims to offer faster, safer, and more efficient solutions in the inspection of power transmission lines, while also having the potential to create a significant transformation in the energy sector by meeting the increasing demand for energy and ensuring the security of transmission line systems2. In the literature, studies exist that explore the use of UAVs for inspecting power transmission lines. For example, Oswaldo Menéndez et al.8, addressed this challenge by developing a monocular camera-based system for the autonomous monitoring of transmission lines. The system can accurately position a commercial UAV over the transmission grid. A key feature of the system is its ability to determine the UAV’s position accurately even in the absence of a GNSS signal, thus enabling real-time reporting of the transmission lines’ condition8. Sharma et al.9 provide a comprehensive review of object detection approaches and challenges specifically in power infrastructure, identifying the critical need for robust detection under cluttered backgrounds and low visibility conditions. Similarly, Aboalia et al.10 proposed a method leveraging feature-level fusion of infrared and visible light images to enhance detection accuracy in complex environments. These works highlight the importance of improving detection algorithms under diverse operational scenarios. European energy companies have also begun adopting UAV technology. For example, the Italy-based energy company Enel utilizes UAV technology for the monitoring and maintenance of power transmission lines. Thanks to regular UAV inspections, Enel can detect faults early and optimize its maintenance operations. Additionally, by using UAVs, Enel improves workplace safety and reduces inspection costs, effectively employing UAV technology to achieve efficiency and cost savings in its energy sector operations11. Energy companies in the United States also widely use UAV technology for the inspection and maintenance of power transmission lines. Southern Company employs UAVs to monitor electrical transmission lines and optimize maintenance processes. The company has achieved reductions in maintenance costs and significant improvements in fault detection time through the use of UAVs. Xcel Energy, one of the first companies to receive approval from the Federal Aviation Administration (FAA), uses UAVs for monitoring and inspecting power transmission lines. With UAVs, Xcel Energy can conduct faster and safer inspections in hard-to-reach areas, leading to significant improvements in maintenance and repair processes4. As China’s economy grows, the demand for electricity also increases, requiring higher standards for the security of PTLs and other electrical infrastructure12. Liu et al. proposed a ”Safe Flight Approach” for UAVs. Existing UAVs are vulnerable to strong winds, posing a risk of colliding with electrical towers and transmission lines. In their study, they applied a wind formula based on wind force, vehicle weight, and other factors13. Faced with the challenges of complex backgrounds in UAV-captured images and determining the optimal positioning of the UAV relative to the target, Wang et al.14 conducted a study. In this study, they performed two main tasks to facilitate the detection of power towers using images captured by UAVs. Using a deep learning-based object detector, they automatically detected normal and abnormal power towers in aerial images. By combining target positioning with data recorded by the UAVs—such as position, altitude, and angle—they were able to approximate the geographic coordinates of the detected abnormal power towers. Thus, through the combination of UAV-based imaging and deep learning technologies, they were able to quickly and automatically detect and locate abnormal power towers in transmission lines14. In their study, the YOLO-Spp3 algorithm was trained over 52 epochs, with a batch size of 60,000. With YOLO-Spp3, they achieved a mean average precision (mAP) of 68%. This method can accurately classify normal and abnormal power towers and generate appropriate bounding boxes for each detected tower14. Since 2013, China’s State Grid and Southern Power Grid have been performing smart maintenance of transmission lines using UAVs. They obtain over 80 million images annually and detect approximately 75,000 faults. These faults include transmission line errors, protection clamp issues, grounding device faults, and environmental problems. The large volume of image data and variety of fault types necessitate the effective and accurate application of deep learning-based object detection on these images. This not only saves labor and resources but also increases maintenance efficiency15. Huaxi Wu et al.15 worked on detecting rusty, loose, and broken power transmission lines, as well as foreign objects that could cause fires or short circuits on the lines. They enhanced the sample images in their database using a simulated augmentation of foreign object faults. Simulated samples were added to the training dataset, and the YOLOv5 algorithm was trained over 200 epochs, achieving an object detection rate of 81.22%15. The environmental interactions and applicability of local planning methodologies for UAVs are critical when advancing toward a target. In this context, the integration of innovative features such as collaborative planning and global optimization allows existing strategies to be integrated with other global planning methods. The proposed approach aims to enable online navigation, allowing UAVs to operate in a more dynamic and interactive manner5. The distinction between types of power transmission lines is similarly important; their varying physical structures and designs present different challenges for UAV visual perception and navigation systems. In particular, double-circuit lines, which contain two parallel electrical circuits, offer more complex structures, while single-circuit lines represent simpler designs with just one electrical circuit. Differences between real-world conditions and simulation environments play a decisive role in realistically evaluating UAV performance. While real-circuit images are used to analyze UAV performance, double-circuit and single-circuit images are modeled in simulation scenarios. Additionally, algorithms for detecting and tracking power transmission lines may perform differently depending on the type of line. In this context, the distinction between double-circuit, single-circuit, and real-circuit types is critical for assessing and improving the effectiveness of algorithms when necessary. Ultimately, this distinction contributes to understanding and adapting the technical requirements needed for UAVs to perform visual-based inspection and landing tasks more effectively. This, in turn, increases UAV operational efficiency and enables the development of more reliable and effective aerial vehicles. Power line detection presents several challenges due to the thin and elongated structure of wires, their low contrast against complex backgrounds such as vegetation or sky, and varying lighting and weather conditions. Additionally, occlusions caused by trees or towers, as well as noise in the aerial imagery, can further complicate accurate detection. These factors necessitate robust models capable of generalizing well across different scenarios and environments. In conclusion, this study presents innovative approaches aimed at improving the security of energy infrastructure and making maintenance processes more efficient. Automatic inspection methods carried out using UAVs significantly reduce costs while also improving workforce safety. Future research will provide important opportunities to further develop and enhance the applicability of these technologies2. Material and method Figure 2 shows a process consisting of three stages. First, labeling was performed using a pre-existing dataset. In this phase, PTLs were marked. Then, YOLOv8 models were trained via Google Colab, and certain parameters were applied for the learning process. The accuracy and detection capabilities of the model were tested, and the data obtained during training were analyzed. Finally, based on this information, the model’s performance was evaluated. The details of these tasks are presented under the following headings. Overview of methodology. Deep learning The foundations of deep learning can be traced back to the development of artificial neural networks in the 1940s and 1950s. However, modern deep learning emerged in the early 2000s, led by researchers like Hinton and Salakhutdinov16. In the 2010s, significant advancements were made in deep learning with the use of large datasets and powerful GPUs17. Specifically, the victory of AlexNet in the ImageNet competition in 2012 marked the widespread recognition of deep learning18. As shown in Fig. 3, deep learning is a subset of machine learning, capable of learning from complex data and making predictions19. Upon reviewing the literature, it has been proven that deep learning is effective in various fields, including image and video recognition, audio processing, and natural language processing20. Hierarchy of learning methods. Deep learning-based remote sensing object detection has a wide range of applications, and YOLO has become a prominent model due to its fast performance and high accuracy21. YOLO is a highly popular object detection algorithm in the field of deep learning. It is a single-stage object detection algorithm known for its fast inference speed and high accuracy. It enables the detection of multiple objects in an image and CNNs to distinguish between them22. As shown in the timeline in Fig. 4, YOLOv1, developed by Joseph Redmon and Ali Farhadi in 2015, attracted attention with its fast processing speed and high accuracy23. Over time, YOLO has undergone many versions, each improving performance and addressing specific issues. For example, YOLOv2 made enhancements to detect objects of varying sizes, while YOLOv3 improved performance with features like a backbone network and spatial pyramid pooling. YOLOv4 introduced enhancements such as CSPDarknet53 as a backbone and PANet for path aggregation, resulting in improved accuracy and training stability. YOLOv5, implemented in PyTorch, offered lightweight models with better deployment flexibility. YOLOv6 introduced improvements for industrial applications, including better anchor-free detection and advanced label assignment strategies. YOLOv7 expanded the detection framework to support additional tasks such as instance segmentation and pose estimation. YOLOv8, the latest release, integrates an anchor-free design, dynamic label assignment, and auto-learning of hyperparameters for even better performance23. YOLOv8 offers the best performance in terms of both speed and accuracy compared to previous models24. YOLO frames object detection as a regression problem, simultaneously predicting bounding boxes and associated class probabilities in an image25. By passing the image through neural networks just once, YOLO delivers faster performance compared to other object detection algorithms26. In this study, the YOLOv8 algorithm was used to detect power transmission lines from UAV images. Timeline of YOLO versions28. YOLOv8 algorithm YOLOv8 was released by Ultralytics, the company that developed YOLOv5, in 202323. YOLOv8 has become dominant in target detection due to its strong generalization capabilities and increased accuracy. In YOLO, predictions are made by passing the input image through the neural network only once. This method enables real-time object detection with a delay of less than 25 ms, providing rapid predictions. The input image is divided into grid cells of S × S dimensions, with each cell associated with bounding box predictions, confidence scores, and class probabilities2. As shown in Fig. 5, the model’s architecture consists of the backbone, neck, and head sections27. In the input part, the image serves as the input data for the model. In the backbone section, convolutional layers are present, gradually reducing the dimensions of the image. YOLOv8 uses the C2f module instead of the C3 module to enhance gradient flow and offer a lightweight design. The C2f (CSP Module) is designed to process deep feature maps and minimize information loss. This module works in integration with shortcut connections, contributing to the processing flow by directly transferring some of the features from previous layers. The Spatial Pyramid Pooling Fast (SPPF) layer pools pixels at different scales to summarize a larger area of the feature maps from the backbone. This improves the model’s ability to recognize objects of various sizes and helps it detect objects in a wider context by combining information at different scales. The SPPF layer is located near the end of the backbone, just before the neck begins. YOLOv8 architecture. In the neck section, there are concatenation and upsampling layers. The features from the neck are processed in these layers. These operations integrate information from previous layers and help align the dimensions23. In the head section, detection layers perform convolutional operations on feature maps at each scale in the final stage, and these layers are used for object detection. Detecting objects at various scales increases the model’s ability to recognize objects of different sizes. Conv applies various convolutional operations and the CSP module in the final stages of detection layers. These operations play a critical role in extracting object classes and coordinates. The model presents an anchor-free architecture equipped with a head structure capable of handling object detection, classification, and regression tasks independently. This design allows each component to focus on its specific task, thus improving the model’s overall performance and accuracy23. In the output layer of YOLOv8, the sigmoid activation function is used to determine the object score, which indicates the likelihood that the bounding box contains an object. Additionally, the softmax function is applied to determine class probabilities, representing the likelihood of the objects belonging to each potential class. It uses the CIoU loss function for calculating bounding box loss and the binary cross-entropy loss function for classification loss21. These loss functions significantly enhance the model’s performance, especially in detecting small objects. Additionally, YOLOv8 can be run via a command-line interface (CLI) or installed as a PIP package23. The model also includes Soft-NMS, a variant of the non-maximum suppression (NMS) technique found in YOLOv5. Soft-NMS processes overlapping bounding boxes by applying a soft threshold instead of completely eliminating them22. YOLOv8 supports various computer vision tasks, including object detection, instance segmentation, pose/keypoint detection, oriented object detection, and classification. Each model type is optimized to perform a specific task in the best possible way. Moreover, it provides flexibility by working in different modes such as inference, validation, training, and exporting23. Experimental studies Data set In this study, comparisons will be made using the results obtained from YOLOv4, based on the Power Transmission Line dataset prepared by Lucas Diniz and colleagues5. The dataset contains three different classes: double-circuit, simple-circuit, and ”real-circuit. Real-world data were collected via a vision-based system integrated into a UAV flying over power transmission lines. The dimensions of the 290 collected images were set at 1920 pixels in width and 1088 pixels in height. Images for the double-circuit and simple-circuit classes were generated in the Gazebo simulator by creating different scenarios. To increase the robustness of the model, some synthetic images included elements such as complex backgrounds, trees, mountains, and buildings that resemble real-world data. The dimensions of the synthetic images were defined as 320 pixels in width and 240 pixels in height. For each synthetic class, 290 different images were produced. Data augmentation techniques such as random rotation, flipping, and contrast adjustments were applied to generate additional data. Through the data augmentation process, a total of 174 new images were created, with 58 images for each class. This ensured that the sample count for each class was balanced, eliminating the issue of class imbalance. In total, the original dataset consists of 1,044 images, evenly distributed among three classes (348 per class): real-circuit, simple-circuit, and double-circuit. In this study the original dataset images were annotated using the RoboFlow platform in YOLO format, which encodes object classes and normalized bounding box coordinates. The hyperparameters used by Lucas Diniz et al.5 in the YOLOv4 algorithm are shown in Table 1. They conducted the training using a Tesla K80 GPU in the Google Colab environment. The training took approximately 35 hours, resulting 98.47% mAP in the validation image set. Table 1Hyperparameter values for YOLOv4 model.HyperparameterValueInitial Learning Rate0.001Momentum0.9Weight Decay0.0005OptimizerSGD Hyperparameter values for YOLOv4 model. Data labeling and preprocessing To prepare the dataset for YOLO training, a series of preprocessing and augmentation steps were applied to improve data quality and model generalization. RoboFlow is a library developed for object detection and segmentation tasks, containing open-source artificial intelligence algorithms28. Each image in the dataset was labeled using the RoboFlow platform. The images labeled with RoboFlow were divided into training, validation, and test datasets for the YOLOv8 models. The training dataset was used for the model’s learning process, the validation dataset was used for model improvement, and the test dataset was used for predictions and evaluations. In this study, a series of preprocessing steps were applied to the dataset created for transmission line detection to prepare the images for model training. Firstly, the Auto-Orient operation was applied to automatically align the images. Then, to focus on a specific region of the images and enhance the detection performance of the model, a static crop operation was performed. During this process, the 25–75% horizontal region and 30–75% vertical region of the images were cropped, and only this selected region was used as input to the model. After the cropping process, the images were resized to the model’s input resolution of 640x640 pixels and converted to grayscale to reduce computational complexity and speed up the training process. To increase the variety of the dataset and improve the model’s capability to detect objects from different perspectives, horizontal flip and 90° rotation (both clockwise and counter-clockwise) data augmentation techniques were applied to each image, generating three additional augmented samples per training example. This method aims to enhance the model’s overall performance by increasing the diversity of the dataset28. YOLOv8 offers five different variations depending on the depth and width of the layers: Nano (YOLOv8n), Small (YOLOv8s), Medium (YOLOv8m), Large (YOLOv8l), and Xlarge (YOLOv8x)23. The total number of images used for each YOLOv8 model after preprocessing and augmentation is shown in Table 2. As seen in Fig. 6, the object from the real-circuit class is marked. Table 2Image distribution for YOLOv8 models after preprocessing and augmentation.VersionTotal imagesTrainingValidationTestingYOLOv8n23742065172137YOLOv8s23782069172137YOLOv8m23842075172137YOLOv8l23982089172137YOLOv8x23932084172137 Image distribution for YOLOv8 models after preprocessing and augmentation. Roboflow sample labeling. The training process was carried out in the Google Colab environment using an NVIDIA Tesla T4 GPU. During this process, all YOLOv8 models were trained for 80 epochs using the hyperparameter values given in Table 3 and the AdamW optimizer. Table 3Hyperparameter values for YOLOv8 models.HyperparameterValueHyperparameterValueInitial Learning Rate (lr0)0.01hsv_v0.4Final Learning Rate (lrf)0.01Translate0.1Momentum0.937Scale0.5Weight Decay0.0005Flipud0.0hsv_h0.015Mosaic1.0hsv_s0.7Copy paste0.0 Hyperparameter values for YOLOv8 models. Results and discussion Processed image examples for each model’s training results, along with the confusion matrix, and loss and training graphs, are presented separately. The processed and bounding-box annotated images of real-circuit, simple-circuit, and double-circuit results from the YOLOv8n model training are shown in Fig. 7 with their class probabilities. Note that the label ”GD” represents the real-circuit class, ”BD” represents the simple-circuit class, and ”CD” represents the double-circuit class. The processed image samples, obtained from YOLOv8n. PTL detection results are displayed with bounding boxes and associated confidence scores. The confusion matrix illustrates how well the model predictions align with the categories of real-circuit, double-circuit, and simple-circuit in the data set. The confusion matrix, presented in Table 4, clearly shows which classes the model correctly detected and where it made errors. It can be seen that all of the simple-circuit, double-circuit, and real-circuit objects in the validation data set were correctly detected by the algorithms. Table 4Confusion matrix comparison of YOLOv8 models. True positive (TP) values show correctly detected objects. False positive (FP) represents incorrect detections (model detecting an object where there is none). False negative (FN) shows the number of missed detections (model failing to detect an existing object).ModelSimple Circuit (BD)Double Circuit (CD)Real Circuit (GD)YOLOv8nTP: 35, FP: 0, FN: 0TP: 71, FP: 0, FN: 0TP: 66, FP: 0, FN: 0YOLOv8sTP: 35, FP: 0, FN: 0TP: 71, FP: 0, FN: 0TP: 66, FP: 0, FN: 0YOLOv8mTP: 35, FP: 0, FN: 0TP: 71, FP: 0, FN: 0TP: 66, FP: 0, FN: 0YOLOv8lTP: 35, FP: 0, FN: 0TP: 71, FP: 0, FN: 0TP: 66, FP: 0, FN: 0YOLOv8xTP: 35, FP: 0, FN: 0TP: 71, FP: 0, FN: 0TP: 66, FP: 0, FN: 0 Confusion matrix comparison of YOLOv8 models. True positive (TP) values show correctly detected objects. False positive (FP) represents incorrect detections (model detecting an object where there is none). False negative (FN) shows the number of missed detections (model failing to detect an existing object). During the training process, it is crucial to reduce the loss in the validation set together with the training set; otherwise, it may indicate that the model is overfitting. The graph in Fig. 8 details the training and loss behavior of the YOLOv8n model. The training loss graph shows how much each model learned from the training data over time. A decreasing loss trend indicates that the model is improving its predictive abilities and performing the task for which it was trained more effectively. Loss and training graphs of the YOLOv8n model. In the loss curves presented in Fig. 8, the initial fluctuations and spikes observed in early epochs represent a typical phase of the training process, reflecting the model’s adjustment to the learning task. These anomalies gradually diminish as training progresses, with both curves stabilizing by epoch 20, indicating successful convergence without overfitting. The graphs show that the losses in both the training and validation sets have decreased and the performance metrics have improved. Throughout the training process, the model made increasingly accurate predictions and reduced its errors. High precision and recall values indicate that all detected objects were correct. The high mAP values further demonstrate the overall good performance of the model. Performance evaluation metrics The following evaluation metrics were used to analyze the performance of the model in this study: mean Average Precision (mAP), Precision, Recall, and F1 score. Intersection over Union (IoU): IoU is a metric that assesses the overlap between a predicted bounding box and the actual bounding box. It plays a significant role in evaluating the accuracy of object localization. Intersection over Union (IoU): IoU is a metric that assesses the overlap between a predicted bounding box and the actual bounding box. It plays a significant role in evaluating the accuracy of object localization. True Positive (TP): If the Intersection over Union (IoU) value obtained during testing is greater than 0.5, it is considered a TP. True Positive (TP): If the Intersection over Union (IoU) value obtained during testing is greater than 0.5, it is considered a TP. False Positive (FP): If the IoU value obtained during testing is less than 0.5 or if multiple bounding boxes are generated, it is considered an FP. False Positive (FP): If the IoU value obtained during testing is less than 0.5 or if multiple bounding boxes are generated, it is considered an FP. False Negative (FN): If an object should have been detected but no object was detected, or if the IoU value is greater than 0.5 but the class is incorrect, it is considered an FN. False Negative (FN): If an object should have been detected but no object was detected, or if the IoU value is greater than 0.5 but the class is incorrect, it is considered an FN. These metrics are used to comprehensively evaluate the object detection performance of the model. True Positive, False Positive, and False Negative values help determine the correct and incorrect detections made by the model. This provides a clearer understanding of the model’s overall accuracy, precision, and recall capabilities. Using the TP, FP, and FN parameters, the precision and recall and F1-score metrics are calculated29. The formulas for these parameters are shown in Eqs. 1, 2, and 3. Precision: Precision is a measure calculated using the TP, FP, and FN parameters. It represents the ratio of true positive detections to all positive detections, indicating how accurate the model’s predictions are. 1 Precision: Precision is a measure calculated using the TP, FP, and FN parameters. It represents the ratio of true positive detections to all positive detections, indicating how accurate the model’s predictions are. 1 Recall: Recall is another metric calculated using the TP, FP, and FN parameters. Recall represents the ratio of true positive detections to all actual positive instances. This shows how well the model captures the actual positive instances. 2 Recall: Recall is another metric calculated using the TP, FP, and FN parameters. Recall represents the ratio of true positive detections to all actual positive instances. This shows how well the model captures the actual positive instances. 2 F1-score: F1-score is a performance metric that combines Precision and Recall. It is defined as the harmonic mean of Precision and Recall, showing the balance between them. 3 F1-score: F1-score is a performance metric that combines Precision and Recall. It is defined as the harmonic mean of Precision and Recall, showing the balance between them. 3 mAP50: Mean average precision is a metric used to evaluate the overall accuracy of the model. In mAP50 metric, the IoU threshold is fixed to 0.50. mAP50: Mean average precision is a metric used to evaluate the overall accuracy of the model. In mAP50 metric, the IoU threshold is fixed to 0.50. mAP50-95: The mAP value is calculated by averaging multiple IoU thresholds from 0.50 to 0.95 at various levels in steps of 0.05. This approach provides a stricter evaluation. mAP50-95: The mAP value is calculated by averaging multiple IoU thresholds from 0.50 to 0.95 at various levels in steps of 0.05. This approach provides a stricter evaluation. These metrics are used to assess the model’s performance by understanding the ratio of correctly detected positive samples and how well the model captures all positive samples. The results provided in Table 5 indicate that all versions of YOLOv8 demonstrate high performance in object detection and achieve good results in most classes. Table 5Performance metrics obtained from the validation dataset using the YOLOv8 Model.VersionClassPrecisionRecallF1-ScoremAP50mAP50-95YOLOv8nBD0.98910.9940.9950.977CD0.99310.9960.9950.935GD1110.9950.834YOLOv8sBD0.99510.9970.9950.981CD0.99610.9980.9950.954GD0.9850.9920.9880.9940.841YOLOv8mBD0.99810.9990.9950.979CD0.99810.9990.9950.938GD0.99810.9990.9950.830YOLOv8lBD0.99710.9980.9950.986CD0.99710.9980.9950.929GD0.98410.9920.9950.827YOLOv8xBD0.99210.9960.9950.989CD0.99810.9990.9950.936GD1110.9950.837 Performance metrics obtained from the validation dataset using the YOLOv8 Model. It is observed that the average precision, recall, and F1 score metrics are above 99%. The mean average precision is calculated by averaging the mAP values of all classes, and the mAP50 value is found to be 0.995. The mAP50-95 values show a decline in some classes, indicating that the model’s performance slightly decreases at stricter thresholds as shown in Table 6. The results in Table 6 belong to the YOLOv4 model extracted from the referenced article5, and mAP50-95 metrics were not presented in the study. Table 6Performance metrics obtained of the YOLOv4 model.ModelClassPrecisionRecallF1-ScoremAP50YOLOv4BD0.9540.8600.9050.985CD0.9240.9260.9200.973GD0.9900.8360.9070.996 Performance metrics obtained of the YOLOv4 model. The overall mAP50-95 value is calculated as 0.919 as shown in Table 7. In general, the model is quite successful in object detection tasks; however, there is potential for performance improvements for certain classes at higher IoU thresholds. Table 7Comparison of average performance metrics across YOLO model variants.ClassPrecisionRecallF1-ScoremAP50mAP50-95Training time (min.)YOLOv8n0.99410.9970.9950.91586YOLOv8s0.9920.9970.9940.9950.92699YOLOv8m0.99810.9990.9950.915148YOLOv8l0.99210.9960.9950.921230YOLOv8x0.99610.9980.9950.921229YOLOv40.9560.8740.9100.984-About 2100 Comparison of average performance metrics across YOLO model variants. Training durations vary according to the model’s capacity and complexity; lighter models (YOLOv8n, YOLOv8s) are trained faster, while heavier models (YOLOv8l, YOLOv8x) have longer training times. YOLOv8m is positioned as a balanced option in terms of training time and performance, making it suitable for larger datasets and complex tasks. Each model has exhibited impressive performance with high IoU thresholds as indicated by the mAP50-95 values. However, the lower mAP50-95 values for the GD class are noteworthy, suggesting the existence of areas for improvement to enhance model performance in more complex object detection scenarios. YOLOv8 versions provide flexibility for different needs and application areas. In the reference study that served as the data source, it is observed that only the mAP metric was used for YOLOv4 performance evaluation, with a calculated value of 0.984. The most notable point in the study is that the training time took approximately 35 hours. These results clearly demonstrate that all YOLOv8 variants significantly outperform YOLOv4 in terms of both training time and performance. Conclusion The advancements in technology have led to the development of YOLOv8, a newer model trained with a larger dataset in a shorter training time. This model has exhibited superior performance not only in overall mAP but also in class-wise mAP values, demonstrating its robustness across various object categories. The ability to efficiently process larger datasets is critical for real-world applications dominated by variable conditions. The newer hardware used in training YOLOv8 may have contributed to this performance improvement. For example, the NVIDIA Tesla T4 graphics card not only shortens training times by providing higher computing power and efficiency but also enhances the quality of results through better resource management and faster processing speeds. YOLOv8 offers an effective solution for detecting energy transmission lines, and better results can be achieved by modifying the default parameters. This study highlights the effectiveness of deep learning methods in the detection of energy transmission lines and underscores the potential of these technologies to transform the energy sector. Further development of this potential requires research with larger datasets and various deep learning algorithms. The integration of different datasets can enhance the model’s generalization capabilities, allowing it to perform better under diverse environmental conditions. Additionally, the development of real-time detection systems can provide significant benefits in the management and maintenance of energy transmission lines. Such systems enable faster identification of problems, thereby reducing response times and minimizing the risk of interruptions or accidents. Integration with existing infrastructure leads to more proactive maintenance strategies, reducing operating costs and increasing safety. The use of deep learning methods in the detection of energy transmission lines can significantly enhance the efficiency and safety of energy distribution networks. The high accuracy rates achieved with YOLOv8 demonstrate the potential for broader adoption of these methods. With the increasing global energy demand, the implementation of advanced detection technologies is critical to ensuring the reliability of energy supply systems. Research and development in this area can strengthen the resilience of energy infrastructure by improving reliability and efficiency of energy networks and supporting the transition to smart energy management solutions. In this context, advancements in the technologies used for detecting energy transmission lines hold the potential to enhance the sustainability of the energy sector, enabling the achievement of smarter and more efficient energy management solutions. Despite the promising results, this study has certain limitations. The dataset is primarily composed of images collected in clear weather conditions, which may restrict the model’s generalizability to foggy, rainy, or nighttime environments. Furthermore, while the current work focuses solely on power transmission lines, extending the model to detect other infrastructure components such as insulators, towers, or vegetation interference could provide a more holistic monitoring framework. Future work may explore domain adaptation techniques or multi-sensor data fusion to enhance robustness across varied conditions. Additionally, comprehensive comparisons with other state-of-the-art detectors such as YOLOv7, YOLO-NAS, and EfficientDet should be conducted to provide a more thorough evaluation of YOLOv8’s performance. Furthermore, ablation studies to systematically assess how individual preprocessing techniques (cropping, resizing, grayscale conversion) and data augmentation strategies (flipping, rotation) affect detection performance would provide valuable insights into the contribution of each component to the overall model accuracy. Future work may explore domain adaptation techniques or multi-sensor data fusion to enhance robustness across varied conditions. Additionally, comprehensive comparisons with other state-of-the-art detectors such as YOLOv7, YOLO-NAS, and EfficientDet should be conducted to provide a more thorough evaluation of YOLOv8’s performance. Furthermore, ablation studies to systematically assess how individual preprocessing techniques (cropping, resizing, grayscale conversion) and data augmentation strategies (flipping, rotation) affect detection performance would provide valuable insights into the contribution of each component to the overall model accuracy. Acknowledgements This work is supported by the Scientific Research Projects Coordination Unit of Fırat University (FUBAP). We would also like to extend our sincere gratitude to Lucas Diniz and his team for providing the dataset used in this study. Additionally, we acknowledge the support and contributions of the project numbered ADEP.24.23, which provided valuable resources and insights that enriched the scope and quality of this research. Author contributions T.K. conceived the experiment(s), S.E. and E.A. analysed the results. All authors wrote the main manuscript text and reviewed the manuscript. Funding This research did not receive any specific funding. Data availability The datasets generated and/or analysed during the current study are available in the IEEE Dataport repository, https://ieee-dataport.org/documents/power-transmission-line-dataset. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. T. Nur and S. Ekici: These authors contributed equally to this work. References Associated Data Data Availability Statement The datasets generated and/or analysed during the current study are available in the IEEE Dataport repository, https://ieee-dataport.org/documents/power-transmission-line-dataset.