Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.
Advertisement
Scientific Reports volume 15, Article number: 38308 (2025) Cite this article
1665 Accesses
Metrics details
Photovoltaic (PV) systems play a vital role in the global transition to renewable energy, yet their efficiency is often compromised by surface defects such as dust accumulation, bird droppings, and cracks. Traditional inspection methods are inefficient, while existing deep learning-based detection models struggle with limited adaptability, large model sizes, and inadequate performance under real-world conditions. To address these challenges, we propose the DCD-YOLOv8s algorithm—an enhanced version of the YOLOv8 architecture that integrates deformable convolutional networks (DCNv3), coordinate attention (CA), and dynamic head (DyHead) modules. These enhancements are designed to strengthen feature extraction, object localization, and detection accuracy while minimizing computational overhead. A custom dataset was constructed by combining a public PV panel defect database with field-collected images, further expanded through data augmentation and self-training strategy. Experimental results demonstrated that DCD-YOLOv8s achieved superior results, with an F1-score of 92.8%, mAP@50 of 95.0%, and mAP@50–95 of 82.3%, while maintaining a high inference speed of 45.9 FPS. Comparative evaluations against YOLOv5s, YOLOv6s, YOLOv7s, YOLOv8s, YOLOv10s, RT-DETR-R18, and YOLOv11s confirm its superior performance of DCD-YOLOv8s in identifying PV surface defects. Ablation studies validated the individual and combined efficacy of the integrated modules. Although real-time UAV-based deployment was not conducted, a mission planning framework was proposed. These results highlight DCD-YOLOv8s’s strong potential for integration into real-time UAV-based inspection systems, contributing to cost-effective and reliable PV system maintenance.
With the increasing awareness of the protection of the social environment and the requirements of sustainable development, reliance on traditional fossil fuels is associated with significant pollution and no longer meet the needs of current development. As a result, there has been a widespread shift toward and develops various new types of clean energy. Photovoltaic (PV) power generation exhibits substantial potential compared to other new energy sources due to its safety characteristics, pollution-free nature, short power plant construction period, and fewer environmental restrictions¹.
According to Solar Power Europe’s “EU Market Outlook for PV Power 2023–2027,” the European Union installed a record 41.4 gigawatts (GW) of new PV capacity in 2022, marking a 47% increase compared to 2021. This growth trajectory is expected to continue, with projections indicating that annual installations will exceed 50 GW in 2023 and more than double to 85 GW by 2026. These developments align with the European Commission’s “EU Solar Energy Strategy,” which aims to achieve over 320 GW of total installed PV capacity by 2025 and nearly 600 GW by 2030². Figure 1. illustrates the top 10 countries producing the most solar energy from 2017 to 2023, which shows the increasing use of solar energy around the world³.
Top 10 countries in solar energy production from 2017 to 2023.
Despite this rapid growth, PV systems still face significant challenges such as high operational costs and reduced efficiency, often caused by environmental and mechanical factors. Surface defects such as dust accumulation, cracks, and bird droppings are among the most common issues that significantly impact the performance and efficiency of PV panels. Dust accumulation on the panel surface obstructs sunlight, reducing energy production based on dust density and panel orientation, and may also generate hot spots, which can accelerate the degradation of solar cells in the long term⁴, cracks, whether caused by thermal stress or physical impacts, allow moisture and contaminants to penetrate the panel’s protective layers, leading to electrical failures, corrosion, and reduced output over time⁵, and bird droppings not only reduce energy production efficiency by shading the cells but also create hot spots by concentrating the shading in small areas, leading to localized temperature increases, and their acidic composition corrodes panel surfaces, further reducing efficiency^6,7. Collectively, these defects on the solar panel accumulate, reducing its production capacity, breaking the thermal balance of the solar panel surface, and causing hot spots on the gray area surface, which may burn the solar panel.
To address these challenges, there is a need to improve the reliability of monitoring systems designed to detect defects and ensure energy production efficiency, avoiding energy production losses. Traditional manual inspection methods are labor-intensive, time-consuming, and prone to human error. Consequently, image-based defect detection using machine vision and deep learning techniques has become a popular approach. These technologies enable the rapid, accurate, and scalable identification of panel defects, even in complex environments.
This study integrates fundamental image-processing techniques with deep-learning approaches to address the challenge of surface defect detection in PV panels. The YOLOv8 architecture a single-stage object detection model known for its accuracy and efficiency in prior applications^8,9,10,11– is enhanced through targeted modifications to optimize its performance for PV defect detection. To the best of our knowledge, this marks the first application of this approach to photovoltaic panel defect detection. The study aims to contribute to real-time automatic detection systems, reducing reliance on human labor by leveraging drone technology, thereby lowering operational costs for photovoltaic power plants. Consequently, the proposed method prioritizes reliability and accuracy, ensuring its practical applicability in real-world environments. The key contributions of this study include developing a defect identification system capable of operating effectively under real-world conditions, with measurable accuracy sufficient for practical applications. The contributions of this study can be listed as follows:
The second and third C2f layers in the base YOLOv8 network were replaced with the DCNv3_C2f module, which provides a deformable convolution to extract features of more complex objects more efficiently¹².
A coordinate Attention (CA) module has been added to address the major problems faced by models in detecting multiple objects, improving feature extraction for the model¹³.
We introduce the Dynamic Head (Dyhead) module to replace the Detect head and improve the model’s detection accuracy¹⁴.
A new, annotated PV defect dataset is constructed and merged with an existing public dataset. Additionally, a comparative evaluation of the performance of YOLOv5s, YOLOv6s, YOLOv7s, YOLOv8s, YOLOv10, RT-DETR-R18, and YOLOv11s is presented in the context of PV defect detection.
Recent advancements in machine vision, computer vision, and image processing have driven significant research into automated detection of surface defects in in PV panels. Numerous studies have focused on deep learning-based methodologies to enhance detection accuracy, automation, and real-time performance in PV inspection systems¹⁵.
In 2019, Yao and Wu developed a Halcon-based system for crack detection in solar cells, contributing to reduced defect rates and labor costs in manufacturing. However, the system was limited in its capacity to identify multiple defect types¹⁶. In the same year, Kurukuru et al.¹⁷ employed Hough transforms and edge detection, achieving better fault identification accuracy, though at the expense of high computational complexity and limited real-time applicability. In 2020, Chen et al. proposed the SACDDS method for detecting crack defects in multicrystalline solar cells. While effective for crack extraction, this approach was restricted to a narrow range of defect categories¹⁸.
In 2021, Zubair et al.¹⁹ introduced a classification neural network for defect detection in photoluminescence (PL) images; however, the method lacked precise spatial localization capabilities. Concurrently, Wang et al.²⁰ used an unsupervised recurrent neural network with electroluminescence (EL) imaging, achieving high optimization accuracy. Nevertheless, the performance of their approach was hindered by the inherent unpredictability of unsupervised learning when applied to rare or complex defects.
In 2022, Alaa et al.²¹ enhanced anomaly detection for PV cell defects using Faster R-CNN with lightweight attention modules, improving crack detection in complex scenes. Similarly, improvements to YOLOv5 significantly boosted accuracy for detecting defects like cracks, black cores, and finger interruptions²², However, these enhancements were evaluated only on a limited set of defect types, restricting generalizability.
In 2023, a series of YOLO-based models further advanced detection capabilities. Li et al.²³ proposed GBH-YOLOv5 with Ghost convolution for small target detection, but faced constraints in lightweight design. Prabhakaran et al.²⁴ developed the RMVDM model, achieving high accuracy with low time complexity; however, its reliance on small datasets raised concerns about overfitting. Chen et al.²⁵ Improved Faster R-CNN with spatial attention, achieving better performance on small targets, though it was limited to cracks and broken grids. Lu et al.²⁶ integrated coordinate attention and a decoupled head into YOLOv5, yielding notable gains in both classification and localization tasks. Kshetrimayum et al.^27,28 introduced a UAV-based system for detecting and cleaning bird droppings on PV panels using YOLOv7 and image mosaicking. Their approach calculated GPS positions of defects and optimized UAV cleaning paths, achieving high detection accuracy and low localization error. a UAV-based system for detecting and cleaning bird droppings on PV panels using YOLOv7 and image mosaicking. Their approach calculated GPS positions of defects and optimized UAV cleaning paths, achieving high detection accuracy and low localization error.
In 2024, Sun T et al.²⁹ presented a method for detecting dust accumulation on photovoltaic panels using the PP-YOLO deep learning algorithm. The authors developed a detection model using the YOLOv5 algorithm, enhanced with a lightweight backbone network, activation function, and attention mechanisms, which improves the detection accuracy and speed. The proposed model demonstrates better recognition accuracy and speed compared to existing models, achieving precision and recall rates of 89.71% and 90.23%, respectively.
Despite these advancements, existing methods exhibit persistent limitations, including a narrow focus on specific defect types, insufficient validation under real-time conditions, computational inefficiency, and constrained generalization due to limited or unrepresentative datasets. To address these shortcomings, this study introduces the DCD-YOLOv8s model, an enhanced YOLOv8s architecture incorporating DCNv3, CA, and DyHead modules. The proposed model enables comprehensive multi-defect detection, achieves a favorable balance between accuracy and speed, and demonstrates robust performance across diverse environmental conditions. Table 1 summarizes the key limitations of prior approaches and highlights how this study contributes to bridging these gaps.
Section 3.1 provides an overview of the methodology architecture, followed by Sect. 3.2, which discusses the mechanisms employed in the model development process. Section 3.3 describes the augmentation techniques utilized in the research. Next, Sect. 3.4 presents the self-training method, and Sect. 3.5 concludes with an analysis of drone mission planning.
This study is organized into three main phases. The first phase entails a literature review of relevant studies to establish a comprehensive understanding of the current research landscape, including the challenges and limitations of existing solar panel defect detection approaches based on advanced deep-learning frameworks.
The second phase, which is the primary focus of this work, is the development of a model called DCD-YOLOv8s. This model has been improved based on the original YOLOv8 architecture through the integration of three modules and mechanisms: DCNv3, CA, and Dyhead. The DCNv3 module is utilized in this research to optimize the base network by replacing the second and third C2f modules. To address the major challenges faced by models in multi-object recognition, a CA module is integrated to detect objects in images with complex distances, enhancing the model’s ability to extract features. Additionally, the Dyhead module is introduced to replace the detection head, further improving the model’s detection accuracy. These modules were selected based on a comprehensive review of recent literature^{12,13,14,30,31} and their alignment with the specific demands of photovoltaic surface defect detection. Although alternative modules such as SE, CBAM, or FPN etc., were not implemented in this study, the chosen combination was deemed more suitable for the accurate detection of fine-grained, spatially localized anomalies, including cracks, dust, and bird droppings.
The third phase involves the development of a flight path plan for the drone to demonstrate the potential for real-time deployment of the proposed system. Despite the absence of real-time testing owing to current equipment limitations, Fig. 2 provides an illustration of the UAV-based detection and analysis framework. This framework is tailored for the surveillance and inspection of power plants by leveraging a deep learning-based architecture, complemented by drone mission planning utilizing the DJI Mavic Air 2 Pro. Information pertaining to UAV mission planning, as outlined in Sect. 3.5, originates from the manufacturer’s specifications, as referenced in Reference³², and serve as a valuable addition to this study in the absence of practical deployment. The improved YOLOv8s network architecture proposed in this study is presented in Fig. 3.
Framework of the proposed method.
The scheme of the DCD-YOLOv8s model’s network structure.
The CA mechanism operates through two sequential stages: feature embedding and attention generation. In the first stage, the module encodes spatially distributed features while retaining location-specific information. In the second stage, it generates attention weights that guide the network to focus on informative regions. This dual-stage design enables the encoding of both channel relationships and long-range spatial dependencies without sacrificing positional precision¹³.
Conventional channel attention mechanisms typically rely on global average pooling to aggregate spatial information. However, such pooling methods tend to eliminate spatial detail, making it difficult to preserve accurate location information, especially in tasks that require fine-grained localization. To address this limitation, the CA block introduces a decomposition strategy that replaces traditional 2D global pooling with two separate one-dimensional pooling operations applied along the horizontal and vertical, respectively.
This decomposition produces two direction-aware feature maps: direction-aware attention, which captures contextual dependencies along the horizontal direction to highlight structural patterns, and location-aware attention, which retains fine-grained positional information within each channel along the vertical direction. These feature maps preserve long-range information along one spatial axis while maintaining positional detail along the other. As a result, the network is able to encode both global context and precise spatial cues, improving its ability to detect small or irregular objects. Figure 4. Shows the structural diagram of the CA attention mechanism.
Structural diagram of CA attention mechanism.
An additional advantage of this design is its efficiency; the CA block introduces minimal computational overhead and is highly modular, making it easy to integrate into existing lightweight architectures such as, MobileNeXt³³, EfficientNet³⁴ and MobileNetV2³⁵. This makes it particularly well-suited for real-time and resource-constrained applications.
In the YOLOv8 backbone network, the module of C2f is essential for leveraging multi-scale features and integrating contextual information to enhance detection precision. However, the stacking of these modules often leads to redundant channels and a limited receptive field, which hinders the model’s ability to detect small, occluded, or multi-scale objects. This limitation increases the likelihood of missed detections or false positives, especially for small objects. Although visual layer information offers rich spatial information, it frequently lacks sufficient semantic detail. To solve this issue, it has been proposed to improve feature representation by integrating Deformable Convolutional Networks (DCN) into the visual layer³⁶, nevertheless, excessive use of DCN layers can reduce computational efficiency without yielding significant performance gains³⁷.
To overcome these challenges, we performed experiments to assess inference speed and detection accuracy. The results demonstrate that replacing the C2f module (P3 and P4 layers) in the backbone of the YOLOv8 network with the DCNv3_C2f module achieves optimal performance. This modification enhances alignment with object shapes and sizes by enabling the network to dynamically adjust its receptive field during sampling, particularly for small objects. Furthermore, the technique improves the bounding box regression parameters during prediction regression, enhancing the model’s capacity to depict small, occluded, and multi-scale objects.
DCNs extend traditional CNNs by enabling convolutional kernels to adapt their receptive fields to object geometry. DCNv1¹² introduced learnable 2D offsets for flexible spatial sampling, and DCNv2³⁰ further incorporated modulation scalars to reweight sampled features; however, both employed separate mechanisms for offsets and weights, limiting spatial generalization and adding complexity. DCNv3³¹ advances this design with a refined offset learning mechanism, employing a lightweight sub-network to predict more precise deformations. As illustrated in Fig. 5, this allows the sampling grid to dynamically align with object shape and scale, improving detection of irregular and partially occluded forms. Moreover, optimized implementation and reduced parameterization make DCNv3 more efficient than earlier versions or standard convolutions^38,39, which often struggle with subtle or partially visible objects due to limited spatial adaptability.
Structure of deformable convolution (3 × 3 Kernel).
In the context of PV defect detection, the architectural enhancements of DCNv3 make it particularly well-suited to addressing the unique challenges posed by such defects. PV surface anomalies, including cracks, dust deposits, and bird droppings, often exhibit irregular geometries and non-uniform distributions; DCNv3’s capacity to deform receptive fields allows the network to adapt its sampling locations with greater precision, thereby improving sensitivity to subtle or fine-grained defects that standard convolutions or earlier DCN variants may overlook. Moreover, since defects occur across a wide range of scales, from hairline cracks to large dust patches, the adaptive receptive fields of DCNv3 enable dynamic adjustment to object size, ensuring robust detection across different scales within a single inspection task. The model’s efficiency is further enhanced by its reduced parameterization in the offset prediction module, which supports real-time inference without compromising accuracy, a critical requirement for UAV-based outdoor inspections. Additionally, PV images are often affected by occlusion and noise due to dirt, reflections, or shadows; under such conditions, DCNv3 can redirect its receptive fields toward visible regions while leveraging contextual information to reconstruct the full structure, thereby reducing false negatives and improving overall robustness. Collectively, these task-specific advantages highlight DCNv3’s suitability for PV defect detection and provide a strong rationale for its integration into the proposed DCD-YOLOv8s model.
The detection head plays a critical role in object detection models by refining the extracted features and predicting object classes, locations, and confidence scores. In the baseline YOLOv8 architecture, the default detection head often exhibits limitations in accurately detecting small, occluded, or scale-variant targets. To overcome these challenges, the original detection head was replaced with the DyHead, an attention-based module designed to enhance feature representation adaptability without adding significant computational overhead. DyHead improves detection performance by sequentially integrating three types of attentions:
Scale-aware attention: This module facilitates effective fusion of features across multiple scales by dynamically weighting them based on semantic relevance. It enables the model to attend to fine-grained features from small objects while maintaining robustness to large contextual variations. This is essential for photovoltaic defect detection, where defect sizes vary significantly (e.g., cracks and bird droppings).
Spatial-aware attention: Spatial attention identifies and enhances discriminative regions in the feature maps. It operates across the spatial hierarchy to emphasize relevant spatial locations while suppressing background noise. DyHead employs deformable convolutions in this step to sparsely focus on informative spatial regions, further refining spatial sensitivity.
Task-aware attention: This component adapts features for specific prediction tasks—such as classification and localization by applying task-specific channel recalibration. It enhances the model’s ability to generalize across multiple object categories by enabling dynamic switching of functional branches, effectively supporting multi-task learning.
These three attention modules are applied sequentially and can be stacked to form a deeper head architecture. Each stage builds upon the refined output of the previous attention mechanism, resulting in a unified representation that simultaneously captures scale variation, spatial structure, and task relevance¹⁴. Figure 6 illustrates this modular stacking and the flow of feature transformation through each attention block.
Dynamic head structure.
In deep learning, data augmentation is essential, especially when dealing with small dataset sizes. It substantially increases sample diversity, thereby enhancing the model’s ability to generalize across various scenarios. There are numerous different augmentation strategies available for object detection tasks, ranging from basic strategies such as rotation and random cropping to advanced approaches like gridmask⁴⁰ and neural style transfer⁴¹. Researchers can select appropriate augmentation strategies based on the unique characteristics of their dataset to optimize model performance.
To simulate this data preparation process, we restricted the augmentation range to only the training image sets, where the horizontal flip fusion technique^42,43 was used to enhance the model’s ability to identify defects in PV panels from different angles. This technique horizontally flips the original images with minor scaling adjustments, enabling the model to detect PV panel defects from various angles. In addition to considering the influence of lighting conditions on the detection accuracy, we modify the image brightness as part of our augmentation technique to diversify the dataset. The brightness is varied within a reasonable range to simulate different lighting scenarios. In addition, zoom-in and zoom-out techniques, which involve randomly increasing and decreasing the image size.
Optimizing limited resources to create models with improved generalization capabilities has been a major research focus in the field of object identification. Because semi-supervised object detection (SSOD) approaches generate anchor boxes efficiently, many studies have investigated their integration with two-stage detection frameworks, including Mask R-CNN and Faster R-CNN. However, using SSOD techniques on one-stage models presents difficulties, as these models generate significantly more anchor boxes, which often leads to lower-quality pseudo-labels.
A simple yet effective semi-supervised object detection (SSOD) strategy based on self-training was adopted in this study. Initially, a preliminary model is trained using the available annotated data. This model is then utilized to create pseudo-labels for the unannotated portion of the dataset, with only those instances surpassing a confidence threshold of 0.7 being retained for further use. The selection of this threshold is grounded in its successful application in prior SSOD research^44,45,46, where it was established as a reliable criterion for identifying high-confidence pseudo-labels. Consistent with these findings, our methodology also employs this threshold to ensure the quality of the pseudo-labeled data. Finally, the model is retrained using a combination of the original annotated data and the newly generated pseudo-labeled data, culminating in the development of the final model.
UAV missions can be planned for a single inspection by calculating how long it will take to fly over a given area while considering spare battery availability. The UAV mission planning and related assumptions presented here are derived from manufacturer-provided specifications rather than direct experimental data. The DJI Mavic Air 2 Pro, equipped with three spare batteries, achieves a total flight time of 93 min, including two breaks in the middle. However, estimates indicate that the additional time required to reposition the drone during a battery change, due to increased power consumption caused by changing wind resistance, is approximately six minutes per break. The area covered by the drone depends on the field of view (FOV) of the onboard camera and its altitude⁴⁷. Under windless conditions, the DJI Mavic Air 2 Pro can travel at a speed of 25 km/h.
When implementing real-time inspection of photovoltaic panels, the drone can operate at an altitude of 30 m, taking approximately 73.8 min to cover a 15-acre area in a single mission. Operation at lower altitudes could facilitate the identification of smaller defects; however, this approach would decrease the coverage area per battery cycle. Comprehensive flight analysis and computation for UAVs as an inspection approach require extensive experimental data and validation to substantiate the assumptions, which could be addressed in future research³².
Section 4.1 introduces the dataset preparation and preprocessing. Section 4.2 describes the experimental evaluation Indicators.
Our study utilizes the Panel Solar Computer Vision Project⁴⁸ as the experimental dataset, which was re-processed and integrated with our own database collected from Yemen and prepared for this project to suit the development and evaluation of machine learning models for detecting surface defects in solar panels. The publicly available database contains 6,493 images classified into four basic defect types: cracks, bird droppings, dust accumulation, and general (clean) panel images. These categories represent common issues that affect solar panels’ efficiency and longevity. The dataset is divided into subsets for training (4,546 images), validation (1,299 images), and testing (648 images).
To simulate scenarios with limited data, since the public database contains augmentation techniques, we reprocessed the data by removing those images and preserving only the original images. From these, 20% were manually selected as training, validation, and test samples for our project, totaling 1,142 images. These images are split in a ratio of 7:2:1, resulting in 797 images for the training set, 230 images for the validation set, and 115 images for the test set, as shown in Fig. 7, Box B. Next, we applied augmentation techniques described in Sect. 3.3 to double the database 5 times, resulting in a total of 5,710 images. These images are then divided into 3,997 images for the training set, 1,142 images for the validation set, and 571 images for the test set, as shown in Fig. 7, Box C. Finally, using the SSOD (Self-training) technique, our pre-processed database (own dataset) using augmentation techniques and consisting of 3427 images is integrated into the training set, leading to a total of 7424 images for the training set while keeping the size of the validation and test sets constant at 1142 and 571 images, respectively, as shown in Fig. 7, Box D. By addressing real-world challenges in solar panel maintenance, the final dataset supports applications in automated defect detection, predictive maintenance, and energy optimization. Figure 8 shows four examples from each category.
Schematic diagram of dataset preprocessing.
Example of images in different image sets using in the experimentation of this work and presented as a general dataset.
Accuracy indicators include mAP, precision, recall, and F1-Score. Generally, the higher the value of these parameters, the higher the recognition accuracy rate, that is, the better the detection accuracy of the model. Use the formula in Eqs. (1–3) to calculate recall, precision, and F1-Score:
where (:phi:) =1 and the TP represents true positive cases of correct classification (correctly identified samples); FP represents false positive cases (incorrectly identified samples); TN represents true negative cases of correct classification; and FN represents false negative cases (the number of missing targets among correct targets). The average accuracy (AP) and the paramount average accuracy (mAP) can be calculated using the Eqs. (4) and (5):
where (:pleft(rright)) = curve formed by the precision of the detected object with the change of the recall; and r=recall. V= total number of detected object categories; the value of V in this search is 4, including clean, bird droppings, cracks and dust.
Section 5.1 presents the results, analysis, and discussion resulting from training and testing the model. Section 5.2 examines the ablation experiments and their outcomes. Section 5.3 presents a comparative analysis of the performance of one-stage algorithms. Finally, Sect. 5.4 addresses the limitations of the study and outlines potential directions for future research.
The experimental setup consisted of a server running 64-bit Windows Server 2023, featuring an AMD R9-7945HX processor and a GeForce RTX 4060 GPU. The experiments were conducted using Python 3.10 and the PyTorch framework. Model training was accelerated with a GPU, while testing employed both GPU and CPU resources. For hyperparameter parameters, the learning rate of 0.01 was setting, and the momentum parameter on 0.937 preventing the model from getting stuck in local minima while avoiding overshooting the optimal solution. The Stochastic Gradient Descent (SGD) optimizer was used to update the parameters of the designed model efficiently; The batch size was selected at 8 and weight decay was set at 0.0005. The specific configurations and detailed specifications of the experimental environment parameters and hyperparameters used throughout the training process are provided in Table 2. Figure 9 illustrates the training and testing mechanism for the DCD-YOLOv8s model.
Model training and testing for DCD-YOLO8s.
This section provides a detailed analysis of the efficiency of the DCD-YOLOv8s model, which is based on the YOLOv8s framework. When the model is compared using the suggested method, we discover that the proposed DCD-YOLOv8s model produced excellent results in performance measures, the model’s F1-score, mAP50, and mAP50–95 values were 0.928, 0.95, and 0.823, respectively. These results demonstrate the effectiveness of the proposed model in accurately detecting surface defects in PV panels with high confidence.
A comparative analysis of the loss function between the DCD-YOLOv8s model and the original YOLOv8s model was then performed on the training and validation datasets to evaluate the effectiveness of the loss function in enhancing convergence, as shown in Fig. 10.
The training results during 200 epochs: (a) A comparison of the validation loss curves between the DCD-YOLO8s and YOLOv8s models; (b) A comparison of the training loss curves between the DCD-YOLO8s and YOLOv8s models; (c) the training dataset curves for the box-loss curve, cls-loss curve, and dfl-loss curve; (d) the validation dataset curves for the box-loss curve, cls-loss curve, and dfl-loss curve.
Figure 10a,b displays the loss function curve (box loss curve) of DCD-YOLOv8s, represented by the blue curve, and the loss function curve (box loss curve) of the original YOLOv8s, represented by the orange curve, across both the training and validation sets. It is evident that with the increase in iterations, the mean values of each loss function decrease significantly, indicating that the model learns and improves its predictions over time. Thanks to the optimized loss function, predictions become more accurate and converge more quickly, aligning closer to the ground truth. The mean values of the loss functions tend to converge as the number of training epochs approaches 200.
Figure 10c,d shows the loss function curves for the values ​​of box loss, classification loss, and distribution focal loss on the training and validation sets of the DCD-YOLOv8s model plotted against the training iteration epochs, where the classification loss determines the probability that a detected object belongs to a specific class, whereas the box loss determines the difference between expected and actual bounding box coordinates, and the distribution focal loss determines the network’s discovery of potential coordinates around the goal values. The horizontal axis represents the number of training epochs, while the vertical axis denotes the loss function. All of the created model’s loss functions are jointly minimized through training and validation, demonstrating the model’s increasing improvement in identifying PV panel defects in the images.
To provide a thorough visual analysis of the performance of the proposed model, we calculated new evaluation metrics, supplementing the metrics that had previously been computed. These additional measures include the precision-recall curve, confusion matrices, recall-confidence curve, and F1-confidence curve, as shown in Fig. 11. These additional assessment metrics provide a more nuanced understanding of the model’s ability to detect and classify the surface defects on PV panels in the images correctly. The balance between recall and precision through various decision thresholds is shown by the F1-score and precision-recall curves. The precision-confidence curve illustrates that higher confidence levels correspond to increased precision. The proposed model attained a peak F1-score of 0.93, as demonstrated by the F1-confidence curve. A large area under the recall-confidence curve signifies the model’s high recall and minimal false negatives. The positioning of the proposed model at the top-right corner of the precision-recall curve signifies a substantial area beneath it, highlighting its efficiency.
Performance diagnostic of: (a) F1-confidence curve; (b) precision-recall curve; (c) precision-confidence curve; and (d) recall-confidence Curve.
Confusion matrices provide a detailed understanding of the model’s classification accuracy by revealing the numbers of true positives, true negatives, false positives, and false negatives for every category. The diagonal values of the matrix represent the proportion of correctly predicted categories. Rows correspond to actual categories, while columns indicate predicted categories, and the diagonal values show the proportion of each class that was accurately predicted. The confusion matrix is shown in Fig. 12.
Confusion matrix of the developed DCD-YOLO8s model.
Figure 13 presents visual results demonstrating the effectiveness of the proposed DCD-YOLOv8s algorithm in identifying and localizing key photovoltaic defect classes, including dust, clean panels, cracks, and bird droppings. Through the use of simplified and representative examples accompanied by clear annotations, the figure underscores the model’s robustness in detecting diverse defect types under realistic inspection conditions. The accurate and high-confidence detections across all cases highlight the practical applicability of the model in real-world scenarios. Collectively, these visual results reinforce the quantitative findings by illustrating the model’s strengths in generalization, detection precision, and visual clarity across major defect categories.
Visual detection results of the propsed model: (A) dust, (B) Clean panel, (C) Crack, and (D) Bird droppings.
Based on the accurate experimental evaluation and detailed analysis of the outcomes, the effectiveness and superiority of the proposed method in detecting photovoltaic panel defects are verified. Figure 14; Table 3 presents the results of the ablation study and the training dynamics to compare the results of proposed model with the baseline YOLOv8s and its optimized variants using CA, DCNv3, and DyHead across four key performance indicators: mAP50, mAP50–95, precision, and recall for evaluate their individual and collective contributions to the model’s detection accuracy.
Curves of mAP, mAP50-95, precision, and recall for comparison of the stability and robustness of the ablation experiments over 200 epochs of training.
The initial addition of the CA module to the baseline YOLOv8s model resulted in a noticeable decline in performance across all major metrics. This finding indicates that while CA enhances positional and channel awareness, its standalone integration may disrupt the balance of feature extraction when not paired with modules that complement its spatial encoding capabilities. Next, the subsequent inclusion of DCNv3 alongside CA yielded marginal improvements in certain metrics, yet the overall mAP remained unchanged compared to the YOLOv8s + CA variant. This indicates that the DCNv3 module contributes to enhancing the model’s effectiveness to some degree, albeit insufficient to surpass the performance of the original YOLOv8s.
The introduction of the DyHead module to the original YOLOv8s yielded a more pronounced improvement, with the mAP increasing to 94%, representing a 0.6% enhancement over the baseline model, though accompanied by a slight 0.3% reduction in F1-score. This outcome underscores the effectiveness of the DyHead module in dynamically adapting to multi-scale features, which is particularly valuable in detecting PV defects of varied sizes. Further, when DyHead was combined with DCNv3, a slight increase of 0.1% in mAP was observed, though a slight decline in F1-score persisted.
The most significant performance gains were achieved by integrating the CA, DCNv3, and DyHead modules in a triple-fusion configuration. This approach outperformed all other ablation settings, yielding substantial improvements: a 2.4% increase in the F1-score, a 1.6% increase in mAP, a 3.5% increase in precision, and a 1.4% increase in recall. These results indicate that the triple fusion approach is highly effective in capturing shape variations and local features of PV panel defect objects, thereby significantly improving detection accuracy.
In the mAP50 and mAP50–95 plots, the red curve shows both faster convergence and higher final values. This indicates that the proposed model not only learns more efficiently but also achieves more accurate object localization across a wider range of intersection-over-union thresholds, particularly for fine-grained photovoltaic surface defects such as cracks, dust, and bird droppings. Notably, while other variants (e.g., YOLOv8s + CA + DCNv3 or YOLOv8s + DyHead) also exhibit improved performance over the baseline, their curves plateau earlier and at lower performance levels than DCD-YOLOv8s, confirming the complementary benefit of integrating all three modules. In the precision and recall curves, DCD-YOLOv8s consistently maintains the highest and most stable values, indicating a lower false positive rate and strong sensitivity to true positives, both critical for robust and generalizable real-time deployment. These results confirm that DCD-YOLOv8s outperforms baseline and partially modified models in both accuracy and training stability, benefiting from the combined strengths of DCNv3, CA, and DyHead. A visual representation of these key performance metrics is presented in Fig. 15 to facilitate direct comparison.
Graphical representation comparing the performance of the ablation experiments with the proposed DCD-YOLOv8s model.
Table 4 presents a detailed comparison of several state-of-the-art object detection models, all evaluated under consistent training conditions, 200 epochs with an input resolution of 640 × 640 pixels. The key evaluation metrics include F1-score, mAP@50, mAP@50–95, and frames per second (FPS), providing a comprehensive assessment of both accuracy and real-time efficiency in defect detection scenarios.
The proposed DCD-YOLOv8s model achieved the highest F1-score (92.8%), indicating an excellent balance between precision and recall. It outperformed recent models such as YOLOv10s (92.7%) and RT-DETR-R18 (92.6%) and showed notable improvements over earlier variants like YOLOv6s (87.3%) and YOLOv7s (88.7%). These improvements indicate that DCD-YOLOv8s effectively reduces both false positives and false negatives, which is crucial for real-world object detection tasks.
In terms of mAP@50, DCD-YOLOv8s reached 95.0%, closely aligning with YOLOv10s (94.9%) and YOLOv11s (95.4%). Compared to the baseline YOLOv8s (93.4%), this represents a meaningful improvement, primarily due to the integration of DCNv3, Coordinate Attention (CA), and DyHead modules, which collectively improve both classification and localization precision.
Under the more stringent mAP@50–95 metric, DCD-YOLOv8s achieved 82.3%, matching YOLOv10s and surpassing all other models, including YOLOv11s (82.2%), RT-DETR-R18 (82.0%), and YOLOv8s (81.6%). This demonstrates the model’s robustness and consistent performance across a broader range of intersection-over-union thresholds, which is crucial for detecting objects of varying sizes and shapes.
Despite its enhanced architecture, DCD-YOLOv8s maintained high inference efficiency, achieving 45.9 FPS comparable to YOLOv10s (46.5 FPS) and YOLOv11s (46.8 FPS), and outperforming YOLOv8s (45.1 FPS). These results confirm that the added modules do not significantly impact runtime performance while contributing to measurable accuracy gains.
Although Table 4 shows only marginal gains, approximately 0.1% improvements in F1-score and mAP@50 compared to YOLOv10s, alongside a slight reduction of 0.6 FPS in inference speed, these numerical metrics do not fully capture the practical benefits of the proposed modifications. The incorporation of DCNv3, CA, and DyHead modules strengthens the model’s ability to detect complex, small, irregular, and occluded defects commonly encountered in photovoltaic surface inspection. Notably, DCD-YOLOv8s exhibits more stable convergence behavior, fewer false positives, and greater consistency across varied testing conditions. Figure 16 provides a visual comparison of defect detection between the proposed DCD-YOLOv8s model and YOLOv10s, demonstrating a slight improvement in our model’s ability to detect and predict defects, particularly under challenging lighting conditions, partial shading, occlusion, and irregular defect patterns.
Comparison of defect detection between DCD-YOLOv8s and YOLOv10s under challenging conditions.
These qualitative advantages, while not fully reflected in standard benchmarks, translate into meaningful improvements in robustness, generalization, and real-world reliability. Furthermore, the minor reduction in inference speed remains well within the acceptable range for real-time UAV-based applications, where detection accuracy and consistency often take precedence over minimal differences in processing speed. Overall, the proposed model demonstrates a favorable balance between accuracy and efficiency, making it a practical and reliable solution for real-world photovoltaic defect detection tasks. Figure 17 graphically compares the performance of DCD-YOLOv8s with the other models across performance metrics.
Graphical representation comparing the performance of the various models with the proposed DCD-YOLOv8s model.
This study introduces DCD-YOLOv8s, an enhanced object detection model designed to address the challenge of defect detection in photovoltaic (PV) panels. By integrating Deformable Convolutional Networks (DCNv3), Coordinate Attention (CA), and Dynamic Head (DyHead) into the YOLOv8s architecture, the proposed model significantly improves detection accuracy, feature representation, and computational efficiency. Experimental evaluations validate the model’s superior performance, achieving an F1-score of 92.8%, mAP@50 of 95.0%, and mAP@50–95 of 82.3%, while maintaining a high inference speed of 45.9 FPS. Comparative and ablation analyses further confirm the model’s robustness and suitability for real-time applications in automated PV inspection systems. Thus, DCD-YOLOv8s provides a scalable and practical solution that reduces reliance on manual inspection and enables more efficient, cost-effective PV system maintenance.
Despite these promising results, several limitations remain. The model’s evaluation was confined to a specific dataset, which may limit its generalizability to other environmental conditions or rare defect types such as snail trails, partial shading, or cement residue. Additionally, dynamic factors such as lighting variation, weather conditions, and panel orientation were not fully addressed, which may affect real-world performance. Furthermore, while a UAV mission planning framework was proposed, practical deployment was not conducted due to hardware limitations. Future research should focus on expanding and diversifying the dataset, implementing advanced augmentation strategies, and conducting real-time UAV field trials to validate the model’s effectiveness under operational conditions. Addressing these challenges will further enhance the applicability of DCD-YOLOv8s and support its integration into intelligent inspection systems for sustainable solar energy infrastructure.
The data that support the findings of this study are available from the corresponding author upon reasonable request.
Li, J. & Huang, J. The expansion of china’s solar energy: challenges and policy options. Renew. Sustainable Energy Rev. 132, 110002. https://doi.org/10.1016/j.rser.2020.110002 (2020). https://doi.org:.
Article  Google Scholar 
Reuters EU solar growth slows, raising fears for energy transition. https://www.reuters.com/business/energy/eu-solar-growth-slows-raising-fears-energy-transition-2024-12-17/?utm (2024).
Review, W. P. Solar Power by Country 2025. https://worldpopulationreview.com/country-rankings/solar-power-by-country (2025).
Nezamisavojbolaghi, M. et al. The impact of dust deposition on PV panels’ efficiency and mitigation solutions: review Article. Energies 16, 8022 (2023).
Article  Google Scholar 
DroneQuote. Watch out for these 11 common solar panel defects!. https://blog.dronequote.net/watch-out-for-these-11-common-solar-panel-defects-dronequote/ (2023).
Birds, P. Discover the negative impact of dirt and bird droppings on your solar panels. https://prosolarbirds.com/en/impact-of-dirt-and-bird-droppings/ (2024).
Squad, S. The Impact of bird droppings on solar panel efficiency and how to address it. https://squeegeesquad.com/solar-panel-cleaning/the-impact-of-bird-droppings-on-solar-panel-efficiency-and-how-to-address-it/ (2024).
Debsi, A., Ling, G., Al-Mahbashi, M., Al-Soswa, M. & Abdullah, A. Driver distraction and fatigue detection in images using ME-YOLOv8 algorithm. IET Intel Transp. Syst. 18, 1910–1930. https://doi.org/10.1049/itr2.12560 (2024).
Article  Google Scholar 
Xiong, C., Zayed, T. & Abdelkader, E. M. A novel YOLOv8-GAM-Wise-IoU model for automated detection of Bridge surface cracks. Constr. Build. Mater. 414, 135025. https://doi.org/10.1016/j.conbuildmat.2024.135025 (2024).
Article  Google Scholar 
Al-Mahbashi, M., Li, G., Peng, Y., Al-Soswa, M. & Debsi, A. Real-Time distracted driving detection based on GM-YOLOv8 on embedded systems. J. Transp. Eng. Part. Syst. 151, 04024126. https://doi.org/10.1061/JTEPBS.TEENG-8681 (2025).
Article  Google Scholar 
Zhao, S., Peng, Y., Wang, Y., Li, G. & Al-Mahbashi, M. Lightweight YOLOM-Net for automatic identification and Real-Time detection of fatigue driving. Computers Mater. & Continua. 82, 4995–5017 (2025).
Article  Google Scholar 
Dai, J. et al. Deformable convolutional networks. In IEEE International Conference on Computer Vision (ICCV). 764–773 (IEEE, 2017).
Hou, Q., Zhou, D. & Feng, J. Coordinate attention for efficient mobile network design. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 13708–13717 (IEEE, 2021).
Dai, X. et al. Dynamic head: Unifying object detection heads with attentions. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 7369–7378 (IEEE, 2021).
Dung, C. V. & Anh, L. D. Autonomous concrete crack detection using deep fully convolutional neural network. Autom. Constr. 99, 52–58. https://doi.org/10.1016/j.autcon.2018.11.028 (2019). https://doi.org:.
Article  Google Scholar 
Yao, G. & Wu, X. Halcon-based solar panel crack detection. In 2nd World Conference on Mechanical Engineering and Intelligent Manufacturing (WCMEIM). 733–736 (IEEE, 2019).
Kurukuru, V. S. B., Haque, A., Khan, M. A. & Tripathy, A. K. Fault classification for photovoltaic modules using thermography and machine learning techniques. In International Conference on Computer and Information Sciences (ICCIS). 1–6 (IEEE, 2019).
Chen, H. et al. Structure-aware-based crack defect detection for multicrystalline solar cells. Measurement 151, 107170. https://doi.org/10.1016/j.measurement.2019.107170 (2020). https://doi.org:.
Article  Google Scholar 
Abdullah-Vetter, Z. et al. Localization of defects in solar cells using luminescence images and deep learning. In IEEE 48th Photovoltaic Specialists Conference (PVSC). 0745–0749 (IEEE, 2021).
Wang, Y. et al. Adaptive automatic solar cell defect detection and classification based on absolute electroluminescence imaging. Energy 229, 120606. https://doi.org/10.1016/j.energy.2021.120606 (2021).
Article  Google Scholar 
Al-Waisy, A. S. et al. Identifying defective solar cells in electroluminescence images using deep feature representations. PeerJ Comput. Sci. 8, e992. https://doi.org/10.7717/peerj-cs.992 (2022).
Article  PubMed  PubMed Central  Google Scholar 
Zhang, M. & Yin, L. Solar cell surface defect detection based on improved YOLO v5. IEEE Access. 10, 80804–80815. https://doi.org/10.1109/ACCESS.2022.3195901 (2022).
Article  Google Scholar 
Li, L., Wang, Z. & Zhang, T. GBH-YOLOv5: ghost Convolution with bottleneckcsp and tiny target prediction head incorporating YOLOv5 for PV panel defect detection. Electronics 12, 561 (2023).
Article  Google Scholar 
Prabhakaran, S., Uthra, R. A. & Preetharoselyn, J. Deep Learning-Based model for defect detection and localization on photovoltaic panels. Comput. Syst. Sci. Eng. 44 https://doi.org/10.32604/csse.2023.028898 (2023).
Chen, A., Li, X., Jing, H., Hong, C. & Li, M. Anomaly detection algorithm for photovoltaic cells based on lightweight Multi-Channel Spatial attention mechanism. Energies 16, 1619 (2023).
Article  Google Scholar 
Lu, S., Wu, K. & Chen, J. Solar cell surface defect detection based on optimized YOLOv5. IEEE Access. 11, 71026–71036. https://doi.org/10.1109/ACCESS.2023.3294344 (2023).
Article  Google Scholar 
Kshetrimayum, S., Liou, J. J. H. & Huang, Y. P. A deep learning based detection of bird droppings and cleaning method for photovoltaic solar panels. In IEEE International Conference on Systems, Man, and Cybernetics (SMC). 3398–3403 (IEEE, 2023).
Kshetrimayum, S., Liu, C. Y. & Huang, Y. P. An intelligent system for automatic detection and cleaning of bird droppings on solar panels. In International Automatic Control Conference (CACS). 1–6 (IEEE, 2023).
Sun, T., Gao, H., Fan, S. & Hu, X. Enhancing dust detection on photovoltaic panels with PP-YOLO: A deep learning approach. In 4th International Conference on Machine Learning and Intelligent Systems Engineering (MLISE). 24–28 (IEEE, 2024).
Zhu, X., Hu, H., Lin, S. & Dai, J. Deformable ConvNets V2: More deformable, better results. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 9300–9308.
Xiong, Y. et al. Efficient deformable convnets: Rethinking dynamic and sparse operator for vision applications. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition 5652–5661.
Shihavuddin, A. S. M. et al. Image based surface damage detection of renewable energy installations using a unified deep learning approach. Energy Rep. 7, 4566–4576. https://doi.org/10.1016/j.egyr.2021.07.045 (2021).
Article  Google Scholar 
Zhou, D., Hou, Q., Chen, Y., Feng, J. & Yan, S. Rethinking bottleneck structure for efficient mobile network design. In European Conference on Computer Vision. 680–697 (Springer).
Tan, M., Le, Q. & Efficientnet rethinking model scaling for convolutional neural networks. in International Conference on Machine Learning. 6105–6114 (PMLR).
Sandler, M., Howard, A., Zhu, M., Zhmoginov, A. & Chen, L. C. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern recognition. 4510–4520.
Deng, L. et al. Focus-enhanced scene text recognition with deformable convolutions. In IEEE 5th International Conference on Computer and Communications (ICCC). 1685–1689 (IEEE, 2019).
Xi, W., Sun, L. & Sun, J. Upgrade your network in-place with deformable convolution. In 19th International Symposium on Distributed Computing and Applications for Business Engineering and Science (DCABES). 239–242 (IEEE, 2020).
Li, L., Wang, Z. & Zhang, T. Photovoltaic panel defect detection based on ghost convolution with BottleneckCSP and tiny target prediction head incorporating YOLOv5. arXiv (2023).
Naeem, U. et al. Aerial Imaging-Based soiling detection system for solar photovoltaic panel cleanliness inspection. Sensors 25, 738 (2025).
Article  PubMed  PubMed Central  ADS  Google Scholar 
Yang, J. GridMask based data augmentation for Bengali handwritten grapheme classification. In Proceedings of the 2020 2nd International Conference on Intelligent Medicine and Image Processing. 98–102 (Association for Computing Machinery, 2020).
Bagheri, A., Simeone, O. & Rajendran, B. Adversarial training for probabilistic spiking neural networks. In 2018 IEEE 19th International Workshop on Signal Processing Advances in Wireless Communications (SPAWC). 1–5 (IEEE, 2018).
Mujeeb, A., Dai, W., Erdt, M. & Sourin, A. One class based feature learning approach for defect detection using deep autoencoders. Adv. Eng. Inf. 42, 100933. https://doi.org/10.1016/j.aei.2019.100933 (2019).
Article  Google Scholar 
Tang, S., Zhu, Y. & Yuan, S. A novel adaptive convolutional neural network for fault diagnosis of hydraulic piston pump with acoustic images. Adv. Eng. Inf. 52, 101554. https://doi.org/10.1016/j.aei.2022.101554 (2022).
Article  Google Scholar 
Li, Y., Huang, D., Qin, D., Wang, L. & Gong, B. Improving object detection with selective self-supervised self-training. In European Conference on Computer Vision. 589–607 (Springer, 2020).
Liu, Y. C. et al. Unbiased teacher for semi-supervised object detection. arXiv (2021).
Trappey, A. J. C. & Shen, O. T. C. A universal traffic sign detection system using a novel self-training neural network modeling approach. Adv. Eng. Inf. 62, 102674. https://doi.org/10.1016/j.aei.2024.102674 (2024).
Article  Google Scholar 
Álvarez-Tey, G., Jiménez-Castañeda, R. & Carpio, J. Analysis of the configuration and the location of thermographic equipment for the inspection in photovoltaic systems. Infrared Phys. Technol. 87, 40–46. https://doi.org/10.1016/j.infrared.2017.09.022 (2017). https://doi.org:
Article  ADS  Google Scholar 
Panel Solar Dataset. https://universe.roboflow.com/susan-ifblr/panel-solar-bw945
Download references
This paper was funded by Princess Nourah bint Abdulrahman University Researchers Supporting Project, under grant No. (PNURSP2025R755), Nourah bint Abdulrahman University, Riyadh, Saudi Arabia. The authors, therefore, gratefully acknowledge and thank Nourah bint Abdulrahman University for its technical and financial support.
School of Electronic and Control Engineering, Chang’an University, Xi’an, 710064, Shaanxi, People’s Republic of China
Mohammed Al-Mahbashi
Department of Mechatronics Engineering, Faculty of Engineering, Sana’a University, Sana’a, Yemen
Mohammed Al-Mahbashi
Department of Computer Sciences, College of Computer and Information Sciences, Princess Nourah Bint Abdulrahman University, P.O. Box 84428, Riyadh, 11671, Saudi Arabia
Sara Abdelwahab Ghorashi & Hafiza Elbadie Ahmed Elsrej
School of Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing, 210094, People’s Republic of China
Abdolraheem Khader
Department of Information Technology, Faculty of Computing and Information Technology, King Abdulaziz University, P.O. Box 344, Rabigh, 21911, Saudi Arabia
Sharaf J. Malebary
School of Energy and Electrical Engineering, Chang’an University, Xi’an, 710064, Shaanxi, People’s Republic of China
Gang Li
School of Information Engineering, Chang’an University, Xi’an, 710064, Shaanxi, People’s Republic of China
Mohammed Al-Soswa & Akram AL-Radaei
Search author on:PubMed Google Scholar
Search author on:PubMed Google Scholar
Search author on:PubMed Google Scholar
Search author on:PubMed Google Scholar
Search author on:PubMed Google Scholar
Search author on:PubMed Google Scholar
Search author on:PubMed Google Scholar
Search author on:PubMed Google Scholar
Mohammed Al-Mahbashi: Conception of the study, methodology design, manuscript drafting, and analysis and interpretation of data. Gang Li: Project Manager and Study Supervision. Abdolraheem Khader: Contributed to the study design and performed the statistical analysis. Mohammed Alsoswa and Akram AL-Radaei: Assisted in interpreting the results, revising the article and data collection. Sharaf J. Malebary: Data curation, writing—review & editing. Sara A. Ghorashi and Hafiza E. A. Elsreja: Funding acquisition and contributed to the interpretation of results and critically reviewed and revised the article. All authors have reviewed and approved the final version of the manuscript.
Correspondence to Abdolraheem Khader or Gang Li.
The authors declare no competing interests.
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.
Reprints and permissions
Al-Mahbashi, M., Ghorashi, S.A., Khader, A. et al. An effective approach to improving photovoltaic defect detection using the new DCD-YOLOv8s model. Sci Rep 15, 38308 (2025). https://doi.org/10.1038/s41598-025-22307-5
Download citation
Received: 23 June 2025
Accepted: 26 September 2025
Published: 03 November 2025
Version of record: 03 November 2025
DOI: https://doi.org/10.1038/s41598-025-22307-5
Anyone you share the following link with will be able to read this content:
Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative
Advertisement
Scientific Reports (Sci Rep)
ISSN 2045-2322 (online)
© 2025 Springer Nature Limited
Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.

source