GloSoFarID: Global multispectral dataset for Solar Farm IDentification in satellite imagery

Solar Photovoltaic (PV) technology has emerged as a key solution for meeting the global demand for clean and renewable energy. By harnessing solar power for electricity generation, solar PV plays a significant role in mitigating climate change and reducing pollution. Its ability to replace traditional power generation methods reduces global carbon emissions and lessens dependence on finite resources, paving the way for a sustainable future [1].

A critical aspect of the global transition to clean energy is the precise monitoring of solar panel farms, which are integral to renewable energy infrastructure worldwide. Understanding the growth and distribution of these farms is essential for assessing progress towards a sustainable energy landscape [2]. Therefore, monitoring these farms is a pivotal step in comprehensively understanding the global adoption of clean energy resources.

In line with this goal, our project aims to make a significant contribution by creating a comprehensive dataset of multispectral satellite images of solar panel farms. This dataset will be a foundational resource for developing effective machine learning models for monitoring solar panel farms globally. The focus on using these models highlights the importance of accurate and advanced monitoring techniques in promoting a sustainable and cleaner energy future worldwide. This dataset is a pioneering effort, representing the first global compilation of solar panel farm data, and includes comprehensive multispectral satellite information. Our dataset enhances existing resources in three ways:

• •

  It employs mid-resolution data and includes rural areas globally, addressing previous limitations in geographic coverage.

• •

  It provides multispectral satellite images with $13$ multispectral bands, offering more comprehensive information that comes in an easy-to-use format with a complementary how-to-use guide and script.

• •

  The dataset is updated with recent data from $2021$ to $2023$, doubling the size of the existing data sources and ensuring current and relevant insights into the global solar energy landscape.

To construct this dataset, we devised a methodical three-step process that incorporates modern techniques. As illustrated in Fig. 2, our approach begins with the creation of a robust initial training set, proceeds with the training of state-of-the-art (SOTA) models, and concludes by processing the combined outputs of the top-performing models to generate a novel dataset.

This section presents a detailed analysis of the proposed dataset, including comprehensive statistical information. It aims to give users a clear insight into the dataset’s characteristics and strengths. Additionally, our Github repository contains an easy-to-use guide for working with this dataset regardless of the machine learning framework.

Running benchmark segmentation models on our dataset is essential to establish a performance baseline and enable a fair comparative evaluation. This not only enhances our understanding of the dataset but also provides a standardized reference for the broader research community, fostering transparency, consistency, and credibility in future studies. To analyze and benchmark the dataset, we employed two widely recognized semantic segmentation models and one simplified model, namely the Fully Convolutional Network (FCN), U-Net, and Half-UNet depicted in Figure 4.

The FCN, introduced by Long et al. [7], is renowned for its ability to perform end-to-end pixel-wise prediction through the extensive utilization of convolutional layers, allowing it to maintain spatial information throughout the network. This makes FCN particularly effective for semantic segmentation tasks.On the other hand, the U-Net architecture, proposed by Ronneberger et al. [9], features a U-shaped structure that incorporates both encoder and decoder components. The encoder captures contextual information by progressively downsampling the input, while the decoder facilitates precise localization by upsampling and merging features from different resolutions. The Half-UNet [8], a simplified model, offers a more lightweight alternative. While retaining the essence of U-Net, Half-UNet reduces computational complexity by utilizing only half of the U-shaped architecture. This streamlining allows for faster processing and lower resource requirements, making it an attractive choice for scenarios with limited computational resources.

For the training of three models, a batch size of $16$ was selected to balance computational efficiency and model convergence, and training was conducted over $100$ epochs. To mitigate the risk of overfitting, we implemented early stopping with a patience value set to $10$ based on validation loss. These hyperparameter choices were made to strike a balance between model complexity and training efficiency. The summary of the performance metrics and model size are presented in TABLE III. Additionally, Fig. 5 illustrates the qualitative results of the three models.

This study presented a comprehensive global dataset of multispectral satellite images of solar panel farms. Our dataset enhances existing resources in three ways: by incorporating mid-resolution data from rural areas globally, using multispectral satellite images for richer detail, and updating the timeframe of data to $2021$-$2023$ for a current view of the global solar energy landscape. The evaluation of this dataset using benchmark models has established its robustness and set a baseline for future improvements. It has proven effective in accurately identifying solar panel areas on a global scale. This dataset is a significant asset for the solar energy research community, potentially driving advancements in renewable energy development.