On the Development of a Dense Optical Flow Benchmark Dataset for Satellite Meteorology Applications
New and advanced techniques in deriving the cloud brightness motion in passive imagery sequences, or so-called Optical Flow, are enabled by rapid scanning from current generation satellites. These techniques can retrieve accurate estimates for cloud motions at every image pixel, or “Dense” optical flow (DOF), including in regions with multi-layer motions, strong deformations, and near texture-less features. While DOF techniques are gaining popularity to fill in data gaps left by Atmospheric Motion Vector (AMV) techniques and to add information for decision making and machine-learning algorithms, little is known on DOF performance, strengths, and weaknesses specifically for satellite meteorology applications. This is because traditional open-source DOF techniques were not designed to track the complex fluid motions of clouds and utilize ancillary data from highly advanced imagers such as the 16-channel Geostationary Operational Environmental Satellite (GOES)-R series Advanced Baseline Imager (ABI).
At the Cooperative Institute for Research in the Atmosphere (CIRA), DOF retrieval algorithms have been developed for satellite imagery applications following the recent research advances and utilizing ancillary information available from the ABI. These DOF techniques are validated using a novel benchmark dataset to be described within this presentation. The dataset contains cloud-top winds retrieved from an aircraft wind lidar in April 2019 in collaboration with the NASA Langley Research Center, 30 sec imagery from overlapping meso-sectors within the GOES-16 ABI, and a 1.5 hr 6-sec refresh rate geostationary imagery dataset from the pre-operational GOES-17 ABI tests. Preliminary results to be shown demonstrate that the CIRA DOF techniques provide similar performance to operational AMVs while also accurately filling in data gaps typically quality controlled away with current operational approaches. Interpolation validation demonstrates success where clouds steadily vary from frame to frame, and problems in regions with rapid image evolution, such as within bubbling deep convection. Performance for both winds and interpolation are found to be dependent on rapid (≤ 5 min) scanning for cloud features. Future modifications to DOF techniques based on the benchmarks, and future planned benchmarks for new DOF applications, will be discussed here.
