Nowcasting 3D Cloud Fields Using Forward Warping Optical Flow

Abstract Large- to global-domain short-term prediction of clouds (0–3 h), or cloud nowcasting, remains relevant to civilian and military applications ranging from solar energy production to intelligence gathering. Despite the capabilities of contemporary numerical weather prediction 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. A commonly used nowcasting approach involves using two or more images to retrieve the apparent motions of features, or optical flow, which can be used to extrapolate the future location of those features. However, such approaches generally assume that the optical flow field remains unchanged with respect to time, which is challenging to apply to piecewise cloud fields from satellite imagery. Here, we propose a method to nowcast clouds that adapts a computer vision technique for image interpolation, commonly referred to as warping, to account for temporal changes to optical flow fields derived from infrared satellite imagery. We evaluate the proposed method for 991 randomly selected regional cases from 2024 and perform a detailed analysis on three specific cases from 2023. Applying a dense (every image pixel) optical flow retrieval technique to full-disk GOES infrared imagery, we demonstrate that forward warping, when coupled with simple occlusion reasoning, improves nowcasting skill. In addition, we leverage the same optical flow method to predict satellite brightness temperatures and compare the resulting nowcast to the predictions produced from a U-Net architecture implemented autoregressively. Ultimately, the optical flow method outperforms the U-Net at the 3-h prediction time frame when evaluated using root-mean-square error.