GeoCenter: Real-Time Deep Learning with Uncertainty Quantification for Center-Fixing of Tropical Cyclones

This presentation will discuss GeoCenter, a new deep-learning method for objective center-fixing of tropical cyclones (TC). Two features distinguish GeoCenter from existing methods: (1) GeoCenter uses only geostationary satellite data; (2) GeoCenter has been designed from the beginning with real-time implementation in mind and fits within the constraints of an operational environment. Thus, we provide an algorithm that should be straightforward to transition to operations, providing objective global TC center fixes every 10 minutes. Furthermore, based on our testing data, the accuracy of GeoCenter – using only geostationary data – is similar to the accuracy of existing methods that use microwave satellite data, which are less routinely available.

Center-fixing – i.e., determining the location of a tropical cyclone’s (TC) circulation center – is an important first step in the forecasting process. Small errors in the initial center fix can lead to large errors in the current intensity estimate, especially for weak systems, as well as in multi-day forecasts of both TC track and intensity. For operational center-fixing, forecasters rely heavily on microwave (MW) satellite imagery and scatterometer data to complement geostationary satellite imagery, especially for TCs that are weak, poorly organized, or in a remote area. Automated center-fixing methods, such as the Automated Rotational Center Hurricane Eye Retrieval (ARCHER), also rely heavily on MW imagery and are less accurate when using only geostationary imagery. However, the number of operational high-resolution microwave sensors has steadily decreased since 2014, leading to decreased accuracy in center-fixing. Thus, there is an urgent need to develop more accurate methods that use only geostationary data.

To this end, we have developed GeoCenter, an ensemble of convolutional neural networks (CNN). The development data for GeoCenter include satellite images from GOES-16/17/18 and Himawari-9/10, covering TCs in three basins (the Atlantic, eastern Pacific, and western Pacific) over seven years (2016-2022). The inputs for an individual CNN are a multi-channel time series (“video”) at lag times of 0, 30, 60, 90, 120, and 150 min before present (t0); the task is to determine the TC-center location at t0. The GeoCenter ensemble includes four CNNs, where each CNN uses a different set of infrared (IR) channels in the multi-channel time series. Specifically, the sets are {cloud-top phase, ozone, dirty IR}; {shortwave IR, low-level WV, CO2}; {shortwave IR, upper-level WV, mid-level WV}; and {mid-level WV, clean IR, IR window}. Thus, each of the ten IR channels is used at least once in the GeoCenter ensemble. Each CNN also uses GeoProxyVis imagery in the multi-channel time series. During the day, GeoProxyVis is the visible channel normalized by solar zenith angle; at night, GeoProxyVis is a synthetic IR-based reconstruction of the visible channel. Thus, GeoCenter is a day/night algorithm. The ground truth (correct TC center) for training CNNs comes from the operational (Automated Tropical Cyclone Forecast System; ATCF) best-track data. Before training CNNs, we pre-process the satellite data to create 600-by-600-km TC-centered images. During training, we apply random translations to generate a mismatch between the image center and best-track center; the CNNs’ task is to correct this mismatch and find the best-track center. The direction of the random translations is uniformly distributed over [0, 360)°; the distance is drawn from a quasi-Gaussian distribution with a mean of 45 km. Some translations exceed 100 km, so the CNNs are trained to correct both typical and extreme center-fixing errors.

Each CNN in the GeoCenter ensemble produces its own 50-member ensemble, i.e., 50 estimates of the correct TC-center location. Thus, GeoCenter performs uncertainty quantification (UQ) with a 200-member ensemble. Within each CNN, the ensemble is produced by training to minimize the continuous ranked probability score (CRPS) loss function, which encourages a well calibrated ensemble. We assess ensemble calibration with objective tools such as the spread-skill plot, rank histogram, and discard test, which we will feature in our presentation.

Preliminary results suggest that GeoCenter errors – using geostationary data only – are comparable to errors from existing methods that use MW data. We will present results on independent testing data (from different years than the training/validation data) from all three basins. For each storm type – TC, subtropical, extratropical, and disturbance – we will present both objective evaluation scores and case studies. The case studies will involve real-time runs, using only operationally available data from newer systems (2023 or 2024). We will also present the comparison of GeoCenter trained with and without GeoProxyVis. Finally, we will discuss operational implications for GeoCenter.

Disclaimer: The scientific results and conclusions, as well as any views or opinions expressed herein, are those of the author(s) and do not necessarily reflect those of NOAA or the Department of Commerce.