Determination of Low-Intensity Tropical Cyclone Centers in Geostationary Satellite Images Using a Physics-Enhanced Deep-Learning Model
H.G. Wang, Qing Xu, Xiaobin Yin, Yongcun Cheng
Abstract
The incomplete eye structure during the generation and weakening stages of tropical cyclones (TCs) makes it difficult to accurately locate low-intensity TCs in satellite infrared (IR) images. Here, we develop a physics-enhanced deep convolutional neural network (CNN) to determine centers of tropical depressions (TDs) and tropical storms (TSs) with maximum sustained wind speed (MSW) below 63 kt. This is accomplished by integrating consecutive IR images from the Himawari-8 geostationary satellite and historical information of TCs including center position, MSW, and the minimum pressure. Multi-channel images of 196 TCs over the Northwest Pacific from 2015 to 2021 are randomly divided into a 3:1:1 ratio for model training, validation, and testing. Sensitivity experiments are designed to investigate the influence of different inputs on model performance. The best results are achieved by combining 18 hours of images at three IR channels and historical TC information at 3-h intervals as model inputs. The mean distance between the model identified center and that recorded in the Best Track dataset for TD and TS levels are 20.1 km and 19.1 km, respectively. This indicates an accuracy improvement of 63.0% and 54.6%, respectively, over the model which only considers images at the current moment. Compared with some other state-of-the-art models, the center positions of TDs and TSs determined by our model agree better with the Best Track records. The CNN model also performs quite well in determining the center of stronger TCs, with an average error of 14.1 km, indicating it is robust for all-level TCs.