Development of Geostationary Satellite Atmospheric Motion Vectors Forecasting Algorithm by CNN Model

Atmospheric motion vector (AMV) is an important factor that affects most meteorological phenomena in numerical weather prediction. Despite of its significance, the conventional algorithm of moisture tracking for AMV calculated with most of remote sensing data uses the cross-correlation coefficient (CCC) method, resulting in low-resolution (target-based) output and much of errors. In addition, forecasting AMVs is impossible in conventional method because it requires water vapor data 10 minutes from the current time to calculate current winds. For better moisture flow tracking, convolutional neural network (CNN) frames were used that track motion, which is called optical flow estimation in computer vision. The pixel-based high-resolution AMVs are calculated by using the water vapor channel images into the PWC-Net (CNNs for optical flow using pyramid, warping, and cost volume). For each pixel, linear regression is used to forecast AMVs. The performance of the AMVs calculated by CNN was validated by comparing those results and the Korean geostationary satellite GEO-KOMPSAT-2A (GK2A) AMVs with wind fields of ERA5 data at 100-1000 hPa. Experiments used infrared brightness temperature images of three water vapor channels at 6.2 µm, 7.0 µm, and 7.3 µm over Korean Peninsula for 2022. As to root-mean-square vector differences (RMSVDs), the tracking performance of this study was found to be more accurate than the GK2A AMVs ­— 1.3 to 21.93 m/s more accurate for the cloudy sky and 0.32 to 14.9 m/s more accurate for the clear sky above 400 hPa. The results using the CNN model showed better moisture tracking performance than the conventional method, especially for low altitudes. It also enables to obtain higher resolution AMVs with pixel-based tracking rather than conventional target-based tracking. Furthermore, the mean RMSVDs of forecasted AMVs are 1.97 m/s, 2.66 m/s, 3.32 m/s, and 5.28 m/s when the forecast lead time is 10 min, 20 min, 30 min, and 1 hr, respectively. Consequently, high-resolution AMV forecasts with accuracy, which could not be calculated by the conventional method, were obtained by CNN model, and can be used to advance the accuracy of weather forecasting.

 

KEYWORDS: Moisture Tracking; Optical Flow; Atmospheric Motion Vectors; Wind Forecasting; Remote Sensing