Unified Generative Data Augmentation for Efficient Solar Panel Soiling Localization

As the usage of solar power generation increases, it has become essential to predict power generation accurately. Among the various factors that affect solar power generation, soiling on the panel surface drastically reduces solar power generation. Therefore, accurately identifying the area of soiling on the panel surface helps predict solar power generation. However, most existing studies classify the presence or absence of soiling on the panel or the type of soiling. Additionally, current datasets used for training these models, such as the Solar Panel Soiling Image (SPSI) dataset, suffer from limitations, including a lack of diversity in panel types and a small number of unique soiling shapes. To address these issues, we propose three novel data augmentation techniques—Naïve, Realistic, and Translucent—that generate diverse solar panel images with various soiling patterns. Using Pix2Pix and Copy-Paste methods, we created three corresponding datasets to address the imbalances in the existing SPSI dataset. We trained the DeepLabV3+ model for soiling localization using both the original SPSI dataset and our augmented datasets. Experimental evaluations on real-world solar panels installed at Chungbuk National University demonstrated that models trained on our proposed datasets significantly outperform those trained on SPSI data, with improvements in the Jaccard Index of 3.3%, 2.4%, and 14.6% for the Naïve, Realistic, and Translucent datasets, respectively. These results highlight the effectiveness of our data augmentation techniques for improving soiling localization in solar panels.