Abstract: In the context of China’s “dual carbon” goals, the technology of solar panels is rapidly developing. The installation characteristics of PV are large area, irregular, and multi-scene. Therefore, an efficient method for obtaining PV is needed to provide technical support for real-time detection of its type, location, quantity, and scope. This paper proposes an improved U-Net-based approach for identifying and extracting solar panels from high-resolution remote sensing images, aiming at high-precision intelligent detection and assessment of PV.
1. Introduction
1.1 Research Background and Significance
Against the backdrop of global carbon reduction, solar PV technology is undergoing rapid development. China, with its vast territory and abundant solar energy resources, is one of the countries that can make good use of solar energy. The distribution of solar energy resources in China. In addition, China also has many national-level PV industrial bases, which have made great contributions to the country’s solar energy industry.
In the 75th United Nations General Assembly, President Xi Jinping put forward the “dual carbon” goals, namely carbon peaking and carbon neutrality. The Chinese government has decided to increase the country’s self-contribution efforts and adopt more powerful policies and measures to cope with the increasingly severe situation of national resources and climate.
1.2 Research Status at Home and Abroad
Many researchers have conducted studies on the extraction of solar panels. For example, MS Karoui et al. [30] proposed a partial linear NMF-based unmixing method for the detection and area estimation of solar panels in urban hyperspectral remote sensing data. AMM Sizkouhi et al. [31] focused on autonomous path planning by unmanned aerial vehicles (UAVs) for efficient and convenient detection and maintenance of large PV areas.
However, the existing research on PV extraction faces challenges such as few high-quality PV images and datasets in different scenarios, significant multi-scale features of PV, large scene differences, and diverse interference factors.
1.3 Research Content and Technical Route
This paper proposes an improved U-Net model for extracting solar panels from high-resolution remote sensing images. The main work includes PV dataset selection and processing, model improvement and optimization, experimental design and implementation, and result analysis and discussion.
1.4 Structure of This Paper
The structure of this paper is as follows:
- Section 2: Research Theories and Related Technologies
- Section 3: Overview of Classic Convolutional Neural Network Models
- Section 4: PV Detection Network Model Based on Improved U-Net
- Section 5: Experimental Results and Analysis
- Section 6: Conclusion and Future Work
2. Research Theories and Related Technologies
2.1 Supervised Classification Basic Theory
Supervised classification is a method of classifying unknown samples by learning from known samples. It requires a training dataset with labeled data to train a classifier, which can then classify new data.
2.2 Basic Theory of Deep Learning Methods Based on Convolutional Neural Networks
Convolutional neural networks (CNNs) are a type of deep learning model that have achieved great success in image classification, object detection, and semantic segmentation.
3. Overview of Classic Convolutional Neural Network Models
This section introduces five classic CNN models: U-Net, DeepLabV3+, PSPNet, SegNet, and HRNet, and presents their structural frameworks and principles.
3.1 U-Net
U-Net is named for its U-shaped structure. It was proposed in 2015 and is initially used for medical image segmentation. It is a classic semantic segmentation CNN model based on encoder and decoder.
The main features of U-Net are skip connections and full convolution. It is a combination of ResNet and FPN, suitable for large-scale image classification. Its model depth and computational cost are moderate, and its network structure and theory are relatively mature.
4. PV Detection Network Model Based on Improved U-Net
4.1 Improved U-Net Network Structure
The original U-Net has limitations in extracting PV due to the complexity of PV installation scenarios and interference factors. Therefore, this paper proposes an improved U-Net model.
In the improved model, the CBAM module is fused into the upsampling and downsampling of U-Net, and residual modules are embedded. This enhances the model’s ability to focus on PV pixels and reduce the impact of non-target features, improving extraction accuracy.
5. Experimental Results and Analysis
5.1 Dataset Selection and Processing
The dataset used in this paper is the public high-resolution PV remote sensing images released by Hou Jiang et al. [29]. The images are manually screened to determine the final experimental data.
The dataset includes four scenarios: WaterSurface, Grassland, SalineAlkali, and Cropland. The validation dataset is Shrubwood.
5.2 Experimental Design and Implementation
This paper compares the improved U-Net model with other classic CNN models, including PSPNet, SegNet, DeepLabV3+, and HRNet. The evaluation metrics include precision (mP), recall (mR), F1 score (mF1), and Intersection over Union (mIoU).
5.3 Experimental Results
Table 5.1 shows the experimental results of different models on the Cropland dataset.
| Model | mP | mR | mF1 | mIoU |
|---|---|---|---|---|
| U-Net | 91.23% | 92.59% | 91.90% | 91.37% |
| PSPNet | 89.23% | 92.11% | 90.65% | 88.76% |
| SegNet | 90.84% | 93.26% | 91.91% | 90.81% |
| DeepLabV3+ | 91.93% | 93.69% | 92.80% | 90.38% |
| HRNet | 91.15% | 91.82% | 91.48% | 91.24% |
| Ours | 92.51% | 96.72% | 94.57% | 92.75% |
Table 5.1 Comparison of Experimental Results on Cropland Dataset
The improved U-Net model achieves the highest mIoU of 92.75%, demonstrating its superiority in extracting PV.
Similar comparisons and results are presented for other datasets, including WaterSurface, SalineAlkali, Grassland, and Shrubwood. The improved U-Net model consistently outperforms other models in terms of mIoU.
5.4 Analysis and Discussion
The classic CNN models, such as PSPNet, SegNet, DeepLabV3+, and HRNet, have been relatively mature in researching traditional targets. However, they face challenges in accurately identifying and extracting PV due to PV’s multi-scale features, scene differences, and interference factors.
The improved U-Net model, with the integration of the CBAM module and residual modules, enhances the model’s ability to focus on PV pixels and reduce the impact of non-target features. This leads to improved extraction accuracy and robustness.
6. Conclusion
This paper proposes an improved U-Net model for extracting solar panels from high-resolution remote sensing images. Experimental results demonstrate the superiority of the improved model in terms of extraction accuracy and robustness.