Advances in Optical Flow Retrieval Methods for Inferring Atmospheric Winds and Motions
Atmospheric Motion Vectors (AMVs) are retrievals of winds found by navigating the quantified brightness motions, or optical flow, from cloud- and water-vapor drifts observed in satellite image sequences. Data processing centers around the world quantify optical flow using least-squares or cross correlation methods which make several assumptions that are often invalid in natural image scenes. For example, least squares tracking breaks down in instances of low texture, strong deformations, illumination changes, transparent/multi-layer motions, and along motion discontinuities, prohibiting retrieval of fine-scale motion patterns relevant for many forecasting applications. Many of these issues can be addressed by using more advanced optical flow retrieval techniques developed over the last four decades by the computer vision community which are now enabled by the new generation of satellite imagers.
The Cooperative Institute for Research in the Atmosphere (CIRA) has developed the Optical flow Code for Tracking, AMV, and Nowcasting Experiments (OCTANE), an open-source toolkit which includes novel methods that relax typical retrieval assumptions and even enable “dense” (every image pixel) motion retrievals. Within this presentation, we will overview novel techniques that have been explored in the last few years, and how they compare to operational methods for retrieving winds within specially designed benchmarks. Further, we will demonstrate examples of how dense optical flow-derived winds can be used for operational purposes other than numerical weather prediction model data assimilation, such as highlighting early signals of deep convection growing in strong wind shear and using drifting estimates to enhance the temporal resolution of imagery. Finally, we will show early results on using machine learning for optical flow estimation, and where artificial intelligence can be used to further push the boundaries of satellite imagery-based winds retrieval.
