Comparative Analysis of Optical Flow and Traditional Feature Tracking for Atmospheric Motion Vector Retrieval: Implications for Future Satellite Missions

Understanding atmospheric winds is crucial for climate research and weather forecasting. Atmospheric motion vectors (AMVs), retrieved from spaceborne measurements, aid research on convective and extreme weather phenomena, either directly or through data assimilation. The current global satellite constellation routinely produces AMVs by tracking features in sequences of geostationary (GEO) or high latitude low Earth orbit (LEO) images.

Future 3D global wind observation systems may combine both active and passive measurement techniques. Assessing the robustness of passive AMVs is essential for understanding their contributions to this system. However, accurately estimating AMVs from satellite image sequences remains challenging.

Given the complexities of accurately estimating AMVs from satellite images, we adopt an optical flow approach. This method produces a dense vector field for each pixel in image pairs. This study compares the performance of optical flow with traditional feature tracking techniques, examining variations based on weather conditions, image temporal separation, and pressure level. Our optical flow method derives AMVs from just two images without needing initial guesses for configurations or pressure levels, enhancing mission planning flexibility. This comprehensive research encompasses diverse water vapor datasets, multiple time snapshots, different altitudes, high spatial resolution, and broad area coverage.

To evaluate the optical flow algorithm's performance, we used reference datasets from simulations representing various weather phenomena, containing water vapor variables and wind vector fields across pressure levels. We compared the AMVs derived from water vapor field sequences using both traditional feature tracking and optical flow algorithms against simulated wind vectors. We assessed the uncertainties in wind vectors for both methods by contrasting wind estimates with winds from the reference simulations.

Unlike traditional feature tracking, optical flow consistently produces dense AMVs for each pixel in image pairs, resulting in smaller errors. The wind speeds and AMVs from the optical flow closely mirror the reference simulations, while traditional feature tracking often shows over-smoothing and erratic behavior, producing less detailed, piecewise constant speed maps. Our findings are important in shaping the architecture and instrument performance projections for future satellite missions targeting atmospheric wind retrievals.

Optical flow outperforms traditional feature tracking, enhancing AMV accuracy by 30-50% across all weather regimes, pressure levels, and time intervals. AMV algorithm performance is influenced by atmospheric state. We have provided a detailed quantification of uncertainties in both optical flow and traditional feature tracking results, based on the state of the atmosphere. Optical flow demonstrates relative invariance across time intervals. In contrast, traditional feature tracking results are highly sensitive to the chosen time intervals, showing larger errors over shorter time spans. This consistency in optical flow results supports greater flexibility in planning future missions.