Geostationary satellite-derived positioning of a tropical cyclone center using artificial intelligence algorithms over the western North Pacific

Abstract

Abstract Artificial intelligence (AI) models were developed to determine the center of tropical cyclones (TCs) in the western North Pacific. These models integrated information from six channels of geostationary satellite imagery: the brightness temperature of four infrared (IR) and one shortwave IR channels, as well as the reflectivity of one visible channel. The first model is a convolutional neural network designed for spatial data processing, and the second is a convolutional long short-term memory model that effectively captures spatiotemporal information. For training, verification, and testing purposes, spatial images from six channels were obtained from the Japanese Himawari-8 satellite from 2016–2021. The position of the European Center for Medium-range Weather Forecast 6- or 12-h prediction was assigned as an initial value to the AI models. Errors in the initial value were 20–50 km compared to the Joint Typhoon Warning Center best track, depending on TC intensity. Weak (strong) TCs exhibited large (small) errors. This error dependency was found in Automated Rotational Center Hurricane Eye Retrieval (ARCHER) product, which is currently used by several operational organizations. ARCHER errors were typically small when observations from both geostationary and polar orbiting satellites were included. Significant errors remained in the absence of microwave channel information from polar orbiting satellites. This study successfully developed two AI models that consistently determined the location of the TC center using only six-channel images from geostationary satellites. These models exhibited comparable or better performance than the ARCHER products. The newly developed AI models can potentially be implemented for operational use.