An Improved Deep Learning Algorithm for Operational Detection of Frontal Boundaries

Weather fronts are a labor-intensive and time-consuming component of surface analyses generated by forecasters. In previous work, we successfully deployed two types of U-Net models trained over the Continental United States for operational use at the Weather Prediction Center to predict the locations and types of fronts over North America and the Unified Surface Analysis domain. One model predicted the locations of cold and warm fronts, while the other predicted the locations of stationary and occluded fronts. While the two-model system was a tool that gave forecasters more confidence in their analyses, we have identified numerous weaknesses in this system, including poor performance over complex terrain, slow runtimes on non-GPU enabled machines, and conflicting predictions between the two models. To combat these weaknesses, we trained a single U-Net model with a simplified architecture to identify cold, warm, stationary, and occluded fronts and drylines. This new “5-class” model significantly outperforms the previous two-model system with cold, warm, stationary, and occluded fronts over the Unified Surface Analysis domain and shows good skill in dryline identification over North America. During a summer internship at The Weather Company (TWC), the 5-class model was used to provide TWC forecasters with first-guess frontal boundary guidance up to five days in advance over the entire globe. The model has been evaluated by TWC forecasters and is currently being integrated into TWC operational workflows. We discuss the strengths of the 5-class model over the two-model system, including performance increases over various regions and more physically consistent predictions. The simplified architecture of the new 5-class model allows for faster predictions with fewer computational resources and has been deployed operationally at the Weather Prediction Center in place of the old two-model system.