We use convolutional neural networks (CNN), a type of deep-learning model, to achieve this goal. The training data consist of TC-centered 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 channels used consist of 10 infrared (IR) bands and ProxyVis, which is equivalent to the visible channel by day and a synthetic IR-based reconstruction of the visible channel by night. This allows our CNNs to work during both day and night. The images are available every 30 min throughout the lifetime of each TC, over a 2500-by-2500-km TC-centered domain with 2-km spacing. The domain centers used for these images come from the best-track dataset. During CNN-training, we apply random translations to generate a mismatch between the image center and best-track (considered truth) TC center. The job of the CNN is to correct this mismatch and find the TC center. The translation distance is drawn from a quasi-Gaussian distribution with a mean of 45 km, while the translation direction is drawn from a uniform distribution over [0, 360) deg. Thus, the CNN must be able to correct highly inaccurate center fixes, with errors up to 45 km and well beyond in any direction.
Each CNN estimate contains two values: the estimated zonal (Δx) and meridional (Δy) distance from the image center to the true TC center. The CNN performs uncertainty quantification (UQ), outputting an ensemble of 50 (Δx, Δy) pairs. Whereas a deterministic CNN without UQ would be trained to minimize the mean squared error (MSE) loss function, our uncertainty-aware CNN is trained to minimize the continuous ranked probability score (CRPS), ensuring that the ensemble has both a skillful mean and spread. Moreover, we combine several CNNs into a large ensemble, each CNN being trained with a different random seed and a different set of three channels (at five lag times: 0, 30, 60, 90, and 120 min before the present). The large ensemble achieves two objectives. First, it allows us to exploit information from all 11 geostationary channels; training a single CNN with 11 channels is too memory-intensive. Second, the large ensemble improves our UQ, allowing us to resolve both the aleatory and epistemic components of uncertainty. The CRPS approach used for individual CNN captures only aleatory uncertainty, whereas the multi-model approach captures epistemic uncertainty as well. Graphically, the large ensemble of TC-center locations can be presented as a point cloud (one point per ensemble member) or set of probability contours. We call the large ensemble GeoCenter.
Preliminary results – without ProxyVis or western Pacific data – suggest that GeoCenter errors are comparable to errors based on microwave data from other methods. For example, our mean (median) distance error for all validation data is 25 (23) km. Also, GeoCenter provides useful uncertainty estimates, as determined by objective evaluation tools such as the spread-skill plot and discard test. We will present a thorough evaluation of GeoCenter, including comparisons with existing operational methods and case studies.
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.

