3D cloud nowcasting: merging extrapolation with machine learning

Abstract

Short-term prediction of clouds (0-3 hours), or cloud nowcasting, is critical for both civilian and military operations, ranging from solar energy management to intelligence gathering. Despite the capabilities of contemporary numerical weather prediction (NWP) models, nowcasting methods based on near real-time observations (i.e. satellite imagery) hold operational value due to their relative computational efficiency and accuracy for short-term applications. However, successful cloud nowcasting over large- to global-domains must overcome three challenges simultaneously: 1) trajectory accuracy, 2) computational efficiency, and 3) cloud dissipation and formation. In this dissertation we propose a framework that merges a computationally efficient, cloud extrapolation method with NWP fields and ML to better solve all three challenges. First, a commonly used nowcasting approach involves using two or more satellite images to retrieve the apparent cloud motions, or the optical flow (OF), which can be used to extrapolate the future location of those features. However, such approaches generally assume that the OF fields remain static with time, which is a limitation when applied to complex, piecewise cloud fields observed by satellites. This study introduces a new OF-based nowcast method that adapts a computer vision technique for image interpolation, commonly referred to as warping, to account for temporal changes to OF fields derived from infrared satellite imagery. With the use of satellite-based cloud property retrievals, we demonstrate that competitively accurate three-dimensional cloud nowcasting can be efficiently implemented using full-disk geostationary imagery that keeps pace with current-generation image scan cadences. Second, to address cloud dissipation, we propose a method that leverages OF-based nowcasting to generate ML training labels that designates cloud pixels in an initial cloud field for removal with time. Utilizing a dual U-Net approach, one to predict dissipation and the other timing, we demonstrate a proof-of-concept for low-level cloud dissipation that is integrated into an OF-based nowcast framework. Lastly, to address cloud formation, a post-process approach is implemented that combines OF-based nowcasts and NWP relative humidity fields into a U-Net designed to provide probabilistic predictions for cloud. We demonstrate that this approach further improves cloud nowcasting skill where OF-based nowcast approaches fail to predict cloud. Together, these innovations yield a framework that merges OF, NWP, and ML to produce a computationally efficient and skillful approach for three-dimensional cloud nowcasting on a global scale.