Xuchen Ying, Binbin Wang
Abstract Bubble plumes play an important role in both natural and engineered aquatic systems. Optical imaging has been widely used to study bubble size distributions, meandering, and plume growth. Machine learning (ML) offers new opportunities in analyzing optical images of bubble plumes. However, its application is limited by the lack of large, labeled image data sets that realistically represent plume conditions. Here, we introduce PlumeDEBuG, a synthetic image generator designed to reproduce key physical characteristics of bubble plumes, including bubble size distributions and spatial organization, using images derived from laboratory experiments. We first constructed an experimental database of more than 86,000 isolated bubbles of 1–15 mm with associated geometric labels. PlumeDEBuG enables users to generate synthetic plume images by specifying bubble size distributions (uniform, Gaussian, log‐normal, bimodal, or Weibull) and arranging bubbles with either Gaussian or random spatial patterns to resemble void fraction profiles. Validation against laboratory plume images confirms that the generated data sets replicate plume statistics. Finally, we trained deep learning (DL) models (YOLOv8 and SAM) on PlumeDEBuG images and show improved detection accuracy from their pretrained models. Across the images used in this study, the result showed a mean precision of 0.99 in IoU = 0.5 (mAP@50) and 0.91 averaged over IoU thresholds from 0.50 to 0.95 (mAP@50–95) using YOLOv8, and along with an average instance detection rate of 82.6% based on SAM segmentation. Based on our evaluation, we found that the apparent void fraction has a strong influence on the performance of DL models in detecting bubbles.