Turbulence strength C n2 estimation from video using physics-based deep learning.

Area of Science:

• Optics and atmospheric physics
• Image processing and computer vision
• Machine learning for scientific applications

Background:

• Long-distance imaging is degraded by atmospheric turbulence, quantified by the refractive-index structure constant (Cn2).
• Accurate Cn2 estimation is vital for applications in weather forecasting, astronomy, aviation safety, and optical communications.
• Existing methods for Cn2 estimation include meteorological data analysis, optical scintillometry, and passive video camera analysis.

Purpose of the Study:

• To conduct a comparative analysis of classical image gradient methods and deep learning approaches for estimating Cn2.
• To introduce and evaluate a novel physics-based deep learning network for improved Cn2 estimation.
• To release a unique dataset of video captures and reference scintillometer measurements for Cn2 estimation research.

Main Methods:

• Comparative analysis of classical image gradient techniques and convolutional neural network (CNN)-based deep learning methods.
• Development and training of deep learning models using a newly collected dataset of video data and ground truth scintillometer measurements.
• Introduction of a novel physics-based network architecture integrating learned convolutional layers with a differentiable image gradient method.

Main Results:

• Deep learning methods demonstrate higher accuracy on training data compared to classical methods.
• Deep learning models exhibit generalization errors when applied to unseen image datasets.
• The proposed physics-based network achieves high accuracy and generalizability across different image datasets.

Conclusions:

• While deep learning offers high accuracy, generalization remains a challenge compared to classical methods.
• The novel physics-based network effectively balances accuracy and generalizability for Cn2 estimation.
• The released dataset will facilitate further research and development in atmospheric turbulence sensing.