لخّصلي

Abstract

Guano are an important factor affecting the cleanliness of photovoltaic modules on floating solar power plants at sea.It can lead to a decrease in photoelectric conversion efficiency, power loss, and even the occurrence of "hot spots", thereby causing damage to the components. Therefore, the segmentation and detection of guano are crucial for visual automation in the cleaning and inspection processes. However, the composition, density, and thickness of guano naturally vary, leading to inconsistent levels of transparency and color. The uneven intensity of guano images greatly reduces the accuracy of segmentation and detection. Addressing this issue, this study proposes a segmentation algorithm based on combining different color channels to segment guano on the surface of photovoltaic modules. The mean shift method is used for adaptive segmentation to facilitate the detection of guano. Furthermore, the segmentation results obtained by this method are introduced into the input space to improve the traditional Mask RCNN. In addition, this study successfully produced a dataset of guano on the surface of photovoltaic modules using on-site data collection and with the help of a platform built in-house in the laboratory. The experimental results on the self-constructed dataset demonstrate that the enhanced Mask R-CNN model has shown an approximate increase of 5.9% and 6.0% in mAP values for object recognition and segmentation compared to the traditional Mask R-CNN model. This indicates the effectiveness of the methodology proposed in this study. Introduction

Achieving the "dual carbon" goal not only represents China's solemn commitment to the world but also is an inevitable requirement for the country's socio-economic development.We extend our sincere gratitude to the editor and the anonymous reviewers for their professional comments and corrections.However, when utilized for guano identification on the surface of photovoltaic modules in offshore floating photovoltaic power plants, Mask RCNN faces unique challenges such as complex marine environmental backgrounds, extreme variations in lighting conditions, and the diversity of guano shapes and sizes, all of which greatly increase the difficulty of the detection task [21].Furthermore, an innovative separation algorithm is devised, incorporating aspects such as color space, optical image processing models,

The devices and strategies employed in the training of models

The experiments s based on a unified hardware and software environment configuration, the specific configuration is shown in Table.Countries worldwide are realizing the immense potential of offshore photovoltaics, as developing photovoltaic power plants at sea allows for more efficient utilization of sunlight resources while overcoming challenges posed by limited land resources [7].Moreover, by inputting the segmented grayscale images and original RGB images into the Mask RCNN network model, clear color and texture features are introduced to accurately identify and segment guano of various sizes and shapes as well as clean photovoltaic modules.The color image segmentation method used by Navon et al. [16] integrates edge and region-based techniques, while considering local factors to adaptively derive local thresholds, ensuring that any threshold is associated with a specific region, thereby improving the quality of segmentation.This transformation will inject strong momentum into the adjustment of China's energy structure and the promotion of green and low-carbon development, driving economic transformation and upgrading, promoting ecological civilization construction, and achieving sustainable development goals [2].By the end of 2023, the global installed capacity of renewable energy had reached 3870 GW, with solar energy occupying the largest share at 1419 GW. The offshore floating photovoltaic (FPV) industry has shown significant growth globally in recent years [5].In this context, China's 18,000 km coastline and approximately 710,000 km2 of available offshore photovoltaic area present a theoretical capacity to develop nearly 700 million kilowatts of offshore photovoltaic power, showcasing significant development potential [8].Zhang et al. [14] improved the extraction speed significantly by utilizing a fast texture feature extraction method based on the similarity between adjacent pixels while keeping distortion within a reasonable range.Additionally, Li et al. [23] proposed a photovoltaic modules guano detection method based on transfer learning from visible light images collected by drones, achieving intelligent detection of guano.Firstly, by combining different color components to form the feature space, and using the mean shift method to adaptively separate the guano, so that the

CRediT authorship contribution statement

Xifeng Gao: Conceptualization, Writing - original draft.In order to enhance the efficiency and reduce the operation and maintenance costs of offshore photovoltaic power generation systems, it is necessary to develop an efficient and accurate method for guano detection.The combination of these two models will provide diversified solutions for optimizing and upgrading China's energy supply system, advancing the improvement of the clean energy industry chain, and promoting sustainable economic development [4].With the advancement of computer vision, researchers have utilized technologies such as unmanned aerial vehicles and sensors to perform target detection using traditional image segmentation methods.As one of the leading deep learning network, Mask Region-based Convolutional Neural Network (Mask RCNN) has demonstrated outstanding performance in object detection and instance segmentation tasks in numerous fields.For example, Han et al. [22] discussed the challenges of deep learning models in handling high-resolution images in their research and proposed the Feature Pyramid Network to improve the detection performance of small objects.Through the development of offshore floating photovoltaic power plants, China can better harness sunlight resources at sea while addressing challenges associated with limited land resources [9].Mengmeng Liu: Conceptualization, Investigation, Supervision, Validation, Funding acquisition, Writing - review & editing.Lina Yu: Visualization.1.2.

Abstract

Guano are an important factor affecting the cleanliness of photovoltaic modules on floating solar power plants at sea. It can lead to a decrease in photoelectric conversion efficiency, power loss, and even the occurrence of “hot spots”, thereby causing damage to the components. Therefore, the segmentation and detection of guano are crucial for visual automation in the cleaning and inspection processes. However, the composition, density, and thickness of guano naturally vary, leading to inconsistent levels of transparency and color. The uneven intensity of guano images greatly reduces the accuracy of segmentation and detection. Addressing this issue, this study proposes a segmentation algorithm based on combining different color channels to segment guano on the surface of photovoltaic modules. The mean shift method is used for adaptive segmentation to facilitate the detection of guano. Furthermore, the segmentation results obtained by this method are introduced into the input space to improve the traditional Mask RCNN. In addition, this study successfully produced a dataset of guano on the surface of photovoltaic modules using on-site data collection and with the help of a platform built in-house in the laboratory. The experimental results on the self-constructed dataset demonstrate that the enhanced Mask R-CNN model has shown an approximate increase of 5.9% and 6.0% in mAP values for object recognition and segmentation compared to the traditional Mask R-CNN model. This indicates the effectiveness of the methodology proposed in this study.

Introduction

Achieving the “dual carbon” goal not only represents China's solemn commitment to the world but also is an inevitable requirement for the country's socio-economic development. With the comprehensive implementation of the carbon neutrality strategy, the development of clean energy, primarily wind and solar energy, will experience a leap forward [1]. This transformation will inject strong momentum into the adjustment of China's energy structure and the promotion of green and low-carbon development, driving economic transformation and upgrading, promoting ecological civilization construction, and achieving sustainable development goals [2]. In the large-scale development of photovoltaic energy, there are two main models: firstly, large-scale ground-based photovoltaic energy development in the western region with long-distance transmission from west to east; secondly, the development of floating photovoltaic energy systems in the coastal areas of the eastern region [3]. The combination of these two models will provide diversified solutions for optimizing and upgrading China's energy supply system, advancing the improvement of the clean energy industry chain, and promoting sustainable economic development [4].

By the end of 2023, the global installed capacity of renewable energy had reached 3870 GW, with solar energy occupying the largest share at 1419 GW. The offshore floating photovoltaic (FPV) industry has shown significant growth globally in recent years [5]. According to data from the International Energy Agency (IEA), the installed capacity of global FPV has multiplied several times over the past five years, and the compound annual growth rate of FPV is expected to reach 15 % in the next decade. By 2031, the global FPV market is projected to exceed 6 GW, with a cumulative installed capacity surpassing 58 GW [6].

Solar power generation, as one of the primary renewable energy sources globally, continues to exhibit strong momentum. However, the increasing scarcity of land resources in various countries is becoming more apparent, highlighting offshore photovoltaics as an emerging development direction of great interest. Countries worldwide are realizing the immense potential of offshore photovoltaics, as developing photovoltaic power plants at sea allows for more efficient utilization of sunlight resources while overcoming challenges posed by limited land resources [7]. In this context, China's 18,000 km coastline and approximately 710,000 km2 of available offshore photovoltaic area present a theoretical capacity to develop nearly 700 million kilowatts of offshore photovoltaic power, showcasing significant development potential [8]. Through the development of offshore floating photovoltaic power plants, China can better harness sunlight resources at sea while addressing challenges associated with limited land resources [9]. This presents a new direction and opportunity for China's development in the clean energy sector, potentially becoming a key strategic initiative for the country's energy sector transformation and upgrade.

Guano on the surface of photovoltaic modules in floating solar power plants are a common and challenging issue. Research has shown that the obstruction caused by guano can result in a maximum power loss of up to 2 % [10]. Therefore, it is necessary to identify and detect guano on the surface of photovoltaic modules to ensure the normal operation and efficiency of offshore solar power systems. However, traditional methods for detecting guano typically rely on manual inspections, which are labor-intensive, costly, and inefficient [11].

In order to enhance the efficiency and reduce the operation and maintenance costs of offshore photovoltaic power generation systems, it is necessary to develop an efficient and accurate method for guano detection. With the advancement of computer vision, researchers have utilized technologies such as unmanned aerial vehicles and sensors to perform target detection using traditional image segmentation methods. Yang et al. [12] defined color rules and used Gray Level Co-occurrence Matrix to calculate texture features for flame detection. Asatryan et al. [13] achieved smoke detection through a fully segmented and simplified image method in the RGB color space. Zhang et al. [14] improved the extraction speed significantly by utilizing a fast texture feature extraction method based on the similarity between adjacent pixels while keeping distortion within a reasonable range. Song et al. [15] proposed an arc spline model for real-time lane marking detection. This model utilizes a curve-fitting approach to compare the contours of the region of interest (ROI) with standard prototypes, enabling the detection and classification of lane markings. The color image segmentation method used by Navon et al. [16] integrates edge and region-based techniques, while considering local factors to adaptively derive local thresholds, ensuring that any threshold is associated with a specific region, thereby improving the quality of segmentation. Ge et al. [17] choose multiple structural elements based on the geometric features of the target to match image details, preserving fine details while suppressing noise as much as possible. Thresholding the image using a grayscale level-weighted average provides good results for enhancing image edges.

Traditional computer vision methods are limited by the quality of image sampling and require manual design of feature engineering [18]. Additionally, the expressive power of shallow features such as grayscale information, gradients, or texture structures is limited, especially for complex images with rich features [19]. However, deep learning can provide accurate results by learning the features of training data. As one of the leading deep learning network, Mask Region-based Convolutional Neural Network (Mask RCNN) has demonstrated outstanding performance in object detection and instance segmentation tasks in numerous fields. Wang et al. [20] showcased the excellent performance of Mask RCNN on multiple standard datasets, particularly in the application of instance segmentation tasks. However, when utilized for guano identification on the surface of photovoltaic modules in offshore floating photovoltaic power plants, Mask RCNN faces unique challenges such as complex marine environmental backgrounds, extreme variations in lighting conditions, and the diversity of guano shapes and sizes, all of which greatly increase the difficulty of the detection task [21].

Although deep learning models perform well in general environments, there is still room for improvement in specific and challenging application scenarios. For example, Han et al. [22] discussed the challenges of deep learning models in handling high-resolution images in their research and proposed the Feature Pyramid Network to improve the detection performance of small objects. Additionally, Li et al. [23] proposed a photovoltaic modules guano detection method based on transfer learning from visible light images collected by drones, achieving intelligent detection of guano. The algorithm showed an accuracy improvement rate of 1.6 %, a false negative rate of 2.60 %, and a false positive rate of 0.65 %. Yu [24] improved the YOLOv4 algorithm by incorporating the ELAN_Block structure into the backbone network. This enhancement aimed to enhance the detection of guano on photovoltaic modules, leading to a 1.8 % increase in accuracy.

This study aims to enhance the performance of the Mask RCNN model in identifying guano on the surface of floating photovoltaic modules at sea, leveraging the characteristics of deep learning. To address the diversity in the shape and size of guano, a YCrS-based guano separation algorithm is proposed, coupled with an adaptive segmentation using mean shift method. Moreover, by inputting the segmented grayscale images and original RGB images into the Mask RCNN network model, clear color and texture features are introduced to accurately identify and segment guano of various sizes and shapes as well as clean photovoltaic modules. Finally, various evaluation metrics are employed to further enhance identification efficiency, reduce false positive rates, and guide the model training process.

Section snippets

A general framework

The framework proposed in this study encompasses several steps, as illustrated in Fig. 1. Firstly, guano images are collected on-site during project operations, followed by simulated experiments in the laboratory to facilitate data acquisition. Subsequently, the dataset undergoes processing through the utilization of data enhancement and data labeling tools. Furthermore, an innovative separation algorithm is devised, incorporating aspects such as color space, optical image processing models,

The devices and strategies employed in the training of models

The experiments s based on a unified hardware and software environment configuration, the specific configuration is shown in Table. 2. All experiments are performed under the same hardware conditions and utilized the same dataset.

Improvements in the hyperparameter settings of the Mask RCNN are made. The learning rate is set to 0.009, the momentum to 0.9, and the decay factor to 0.0005. After every 70 iterations, the learning rate is reduced by a factor of 0.1. The batch size is set to 2, and

Conclusion

In this study, a deep learning-based guano detection and segmentation framework for photovoltaic module surfaces is proposed. Since research has focused on offshore floating PV plants, accurately detecting, locating, and tracking guano is an important task. To solve this problem, a deep learning model based on Mask RCNN framework is adopted. Firstly, by combining different color components to form the feature space, and using the mean shift method to adaptively separate the guano, so that the

CRediT authorship contribution statement

Xifeng Gao: Conceptualization, Writing – original draft. Ting Wang: Methodology, Software, Writing – review & editing. Mengmeng Liu: Conceptualization, Investigation, Supervision, Validation, Funding acquisition, Writing – review & editing. Jijian Lian: Resources. Ye Yao: Supervision. Lina Yu: Visualization. Yichu Li: Investigation. Yiming Cui: Validation. Runze Xue: Investigation.

Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgements
This study is supported by the National Key R&D Program of China (grant number 2022YFB4200704), Zhejiang Provincial Science and Technology Plan Project (grant number 2022C01004), Tianjin Transportation Technology Development Plan Project (grant number 2022-47). We extend our sincere gratitude to the editor and the anonymous reviewers for their professional comments and corrections.