Detection and Localization of Stationary Waves on Venus Using a Self‐Supervised Anomaly Detection Model

Abstract The atmosphere of Venus features stationary waves of planetary scale, which are believed to be caused by wind flowing over elevated terrains. Detecting these waves can provide valuable insight into atmospheric dynamics. However, manual detection is time‐consuming, especially with the increasing volume of new images. To address this, we propose a self‐supervised anomaly detection model based on a vector‐quantized generative adversarial network to identify and localize stationary atmospheric waves in spacecraft images of Venus. Our model is developed over Akatsuki images of Venus and is trained on smaller cropped images containing features other than stationary waves, allowing it to learn the underlying patterns of normal cloud features. We enhance the model with spatial attention modules, skip connections, and latent space transformation to improve image reconstruction and avoid mixing stationary wave features with artifacts. For inference, we use an overlapping sliding window approach, considering the reconstruction loss of each grid slice to identify anomalies. The model records bounding boxes with high reconstruction loss, localizes stationary waves, and draws the final bounding box. Experiments demonstrate that our model outperforms benchmark models using ultraviolet imaging (UVI) and longwave infrared imaging (LIR) acquired by the Akatsuki Venus orbiter, achieving an area under the curve (AUC) score of 94.78% for UVI and 92.84% for LIR images, and an intersection over union score of 66.53% for UVI and 61.49% for LIR images. The ablation study shows consecutive performance improvements with each added module.