Variable Imaging Projection Cloud Scattering Tomography.

Area of Science:

• Atmospheric science and remote sensing.
• Computational imaging and inverse problems.

Background:

• Scattering-based computed tomography (CT) reconstructs volumetric scattering media from multi-directional images.
• Traditional physics-based optimization methods for CT are slow, difficult to scale, and lack flexibility for variable imaging conditions.

Purpose of the Study:

• To develop a fast and flexible deep-neural network (DNN) solution for variable imaging projection scattering tomography of clouds (VIP-CT).
• To enable efficient and high-quality volumetric recovery of large cloud fields for climate studies.

Main Methods:

• A deep-neural network (DNN) was trained on a large, physics-based labeled volumetric dataset.
• The DNN architecture is scale-oblivious, allowing application to arbitrarily large domains.
• The method was validated by demonstrating CT of a real cloud using empirical data.

Main Results:

• VIP-CT significantly outperforms state-of-the-art methods in image quality.
• The DNN achieves much faster inference speeds, enabling near real-time processing for spaceborne observations.
• The approach demonstrates flexibility in handling variable viewing geometries and resolutions.

Conclusions:

• VIP-CT provides a high-quality, fast, and flexible solution for scattering-based CT of clouds.
• This learning-based approach offers a potential model for nonlinear CT problems in other scientific domains.
• The developed code is publicly available for further research and application.