Enhancing Automated Frontal Boundary Detection with Geostationary Satellite Imagery

Frontal boundaries are a primary feature of surface analysis maps created by human forecasters. The identification and classification of these boundaries is time-consuming and often open to interpretation. In previous work, we trained a UNET3+ model to identify frontal boundaries, including drylines, over NOAA's Unified Surface Analysis Domain. This model used thermodynamic variables at five vertical levels sourced from the ERA5 reanalysis dataset as input, and human-drawn fronts from NOAA were treated as ground truth. The model has been in operation at NOAA's Weather Prediction Center since August 2023, with positive feedback from the operational forecasters. One of the shortcomings of this model is the absence of satellite data in the training process; satellite imagery can help locate fronts and other types of boundaries more easily through cloud patterns and water vapor content, in particular over open waters where surface observations are more sparse than on land, where environmental data can also be less reliable. To address this shortcoming, we have developed a new UNET3+ model that will include Geostationary Operational Environmental Satellite (GOES) data, specifically GOES-16 and GOES-17 that provide data over much of North America, and surface analyses from continental observations every three hours. This new model will further expedite and enhance the frontal and boundary analysis process by providing context that may not be reflected in the ERA5 reanalysis data or operational NWP models.