Two-Stage Artificial Intelligence Algorithm for Calculating Moisture-Tracking Atmospheric Motion Vectors
AffiliationDepartment of Hydrology and Atmospheric Sciences, University of Arizona
MetadataShow full item record
PublisherAmerican Meteorological Society
CitationOuyed, A., Zeng, X., Wu, L., Posselt, D., & Su, H. (2021). Two-Stage Artificial Intelligence Algorithm for Calculating Moisture-Tracking Atmospheric Motion Vectors. Journal of Applied Meteorology and Climatology, 60(12), 1671–1684.
RightsCopyright © 2021 American Meteorological Society.
Collection InformationThis item from the UA Faculty Publications collection is made available by the University of Arizona with support from the University of Arizona Libraries. If you have questions, please contact us at firstname.lastname@example.org.
AbstractMuch of the errors of atmospheric motion vectors (AMV) may be a consequence of algorithms not incor-porating dynamical information. A physics-informed, artificial intelligence algorithm was developed that corrects errors of moisture tracking AMV (from the movement of water vapor) using numerical weather prediction (NWP) fields. The University of Arizona (UA) algorithm uses a variational method as a first step (fsUA); the second step then filters the first-stage AMVs using a random forest model that learns the error correction from NWP fields. The UA algorithm is compared with a traditional image feature tracking algorithm (JPL) using a global nature run as the “ground truth.” Experiments use global all-sky humidity fields at 500 and 850 hPa for 1–3 January 2006 and 1–3 July 2006. UA outputs AMVs with root-mean-square vector differences (RMSVDs) of 2 m s-1 for the tropics and ∼2–3 ms-1 for midlatitudes and the poles, whereas JPL outputs much higher RMSVDs of ∼3 ms-1 for the tropics and ∼3–9 ms-1 for the midlatitudes and poles. Although the algorithm fsUA produces approximately the same global RMSVDs as the JPL algorithm, fsUA has a higher resolution since it outputs an AMV per pixel, whereas the JPL algorithm uses a target box that effectively smooths the vec-tors. Furthermore, UA’s RMSVDs are lower than the intrinsic error (calculated from the differences between two reanaly-sis datasets). Even for error-prone regions with low moisture gradients and where winds are oriented along moisture isolines, UA’s absolute speed difference with “truth” stays within ∼3ms-1. © 2021 American Meteorological Society.
Note6 month embargo; published online: 04 January 2022
VersionFinal published version