Atmospheric turbulence and jitter degrade satellite AI detection
A new study reveals that atmospheric turbulence and satellite pointing jitter significantly impair AI-based object detection in Earth Observation imagery, yet these distortions are rarely included in training datasets. Researchers developed an enhanced image simulator incorporating vertical-path turbulence and jitter from platform vibrations to generate physically realistic distorted images. Using YOLOv8 and RetinaNet for vessel detection, they found YOLOv8 recall dropped from 91% under ideal conditions to 60% with weak turbulence and below 40% under strong turbulence or jitter. RetinaNet proved more robust, maintaining about 75% recall across degraded conditions. The findings highlight the need to include such distortions in training data for reliable EO AI models.
Key facts
- Atmospheric turbulence and pointing jitter degrade EO imagery but are rarely in training datasets.
- An enhanced image simulator incorporates vertical-path turbulence and satellite jitter.
- Vessel detection tested with YOLOv8 and RetinaNet on simulated distorted images.
- YOLOv8 recall fell from 91% to 60% under weak turbulence and below 40% under strong conditions.
- RetinaNet maintained approximately 75% recall across degraded conditions.
- The study underscores the importance of realistic distortions in training data.
- Source is arXiv paper 2605.22268.
- Research focuses on Earth Observation AI robustness.
Entities
Institutions
- arXiv