Combining Optical Flow and ML Methods for Satellite-based Cloud Nowcasting

The Optical Variability Evaluation of Regional Cloud Asymmetries in Space and Time (OVERCAST) project, sponsored by the US Navy Office of Naval Research (ONR), aims to create a global, near real-time, atmospheric 3D cloud field analysis. This initiative provides a unique opportunity to enhance cloud nowcasting capabilities through innovative methodologies. Cloud nowcasting, the prediction of short-term (0-3 hour) cloud movements, is critical for various applications ranging from weather forecasting, renewable energy management, and even battlespace awareness for Department of Defense (DoD) missions. This study contributes to OVERCAST by investigating approaches to combine the capability of UNet models with an optical flow warping technique that produces advective nowcasts of satellite imagery. Specific attention is placed on narrowing the focus for ML models to only the prediction of cloud formation and dissipation by eliminating advection as a required task. The CLAVR-x (Clouds from AVHRR Extended) is employed to create an initial state of 3D clouds using cloud top height and cloud geometric thickness which is advected based on dense optical flow retrievals and then updated with ML methods. To assess performance, a validation method is established that compares the resulting 3D cloud fields via the use of traditional forecast metrics such as probability of detection, critical skill index, and false alarm rate as well as a fraction skill score (FSS) to account for spatial variability of predictions. Results indicate that a simple optical flow warping technique with occlusion reasoning outperforms persistence, model wind-based advection methods, and optical flow without warping. Preliminary results from the integration of ML will be presented.