Unraveling deep-learning detection of atmospheric Kármán vortex streets: dataset, baseline benchmarking, and optimization
Atmospheric Kármán vortex streets, which frequently emerge on the leeward side of isolated islands under specific atmospheric conditions, play a critical role in large-scale atmospheric dynamics. However, their automated detection and the influence of atmospheric variables on their formation remain...
Saved in:
| Main Authors: | , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Taylor & Francis Group
2025-07-01
|
| Series: | Big Earth Data |
| Subjects: | |
| Online Access: | https://www.tandfonline.com/doi/10.1080/20964471.2025.2522492 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Atmospheric Kármán vortex streets, which frequently emerge on the leeward side of isolated islands under specific atmospheric conditions, play a critical role in large-scale atmospheric dynamics. However, their automated detection and the influence of atmospheric variables on their formation remain insufficiently explored. Here, we firstly developed an object detection dataset utilizing 20 years of cloud imagery from the MODIS aboard the Terra and Aqua satellites, facilitating the application of artificial intelligence algorithms for the automated detection of Kármán vortex streets. Given that these vortex streets can manifest under both cloudy and cloud-free conditions, we further assessed the contribution of various spectral bands to vortex detection by implying a multi-channel classification network. This network, built upon a custom-designed mini-ResNet dual-level classification network and trained on MODIS HDF-format sensor data, demonstrates that the dataset is particularly suited for detecting Kármán vortex streets in cloudy conditions. Different types of advanced algorithms were applied on the dataset, yielding baseline performance metrics and potential detection methods. Additionally, we proposed two optimization strategies tailored to Kármán vortex streets detection. First, leveraging the intrinsic properties of vortex structures, we introduced a Gamma loss optimization for bounding box aspect ratios, which significantly enhanced performance across six YOLO-based models. Second, we investigated the role of atmospheric variables in vortex formation and incorporated an auxiliary-task-enhanced learning approach to further refine detection accuracy, drawing upon findings from our recently published work. In summary, this study provides a novel dataset that supports meteorological and climatological research, addressing a critical gap in atmospheric data. Furthermore, the proposed auxiliary optimization strategies highlight the distinct challenges of Kármán vortex streets detection compared to conventional object detection, underscoring the unique structural and atmospheric characteristics of these phenomena. |
|---|---|
| ISSN: | 2096-4471 2574-5417 |