Center Fixing Tropical Depressions and Tropical Storms Using Machine Learning - Nighttime Visible Imagery

The first step in most tropical cyclone (TC) forecasting techniques is determining the storm's central position. In mature TCs, the center is highlighted by a distinct eye and curved band pattern; however, in intensifying and decaying storms, the central position is often obscured by thick clouds or overlying cirrus. Generally, microwave imagery addresses this challenge; however, microwave sensors aboard polar orbiting satellites provide discontinuous coverage throughout the TC life cycle; therefore, analysts rely on geostationary satellite imagery for extended forecast periods. This study assesses the benefits of incorporating machine learning-derived nighttime visual imagery to improve analysis of center fix in intensifying and decaying tropical depression (TD)- and tropical storm (TS)-strength TCs during periods of darkness and when polar orbiting satellites are unavailable. The machine learning model used for this study, referred to as Machine Learning – Nighttime Visual Imagery (NVI), incorporates the Advanced Himawari Imager's 3.9µm, 8.6µm, 10.4µm, and 12.4µm wavelength bands, as well as their brightness temperature differences, to predict lunar reflectance comparable to that of Lunar Reflectance Values derived from measured radiances using the VIIRS Day/Night Band (DNB) sensor.

The study is divided into two parts: the first, a subjective imagery analysis with the Joint Typhoon Warning Center (JTWC), and the second, an objective analysis using the Automated Rotational Center Hurricane Eye Retrieval (ARCHER-2) algorithm. The study with the JTWC assesses subjective center fix accuracy using shortwave infrared (SWIR) and NVI imagery. Eleven operational analysts center fixed 16 Western Pacific TC image sets - results pending. For the objective analysis, 914 NVI, SWIR, and LWIR image sets, comprised of TD- and TS-strength cyclones, were ingested into ARCHER-2. Results showed that NVI improves center fix forecasts by 26km (28%) and 51km (44%) over SWIR and LWIR, respectively.