Improving the Prediction and Discernment of Frontal Boundaries with Deep Learning

Frontal boundaries and other airmass boundaries are important drivers of convective initiation, precipitation and form the backbone of surface analyses created by forecasters at NOAA. Surface analyses are labor-intensive and require forecasters to sift through layers of data in order to properly locate and classify fronts. This often results in surface analyses with frontal boundaries whose types and locations are subjective in nature. We are developing a single deep-learning model with the U-Net 3+ architecture to improve the consistency in the classification of cold, warm, stationary, and occluded fronts, along with drylines. Through previous work we demonstrated that two U-Net 3+ models, one predicting cold and warm fronts and the other predicting stationary and occluded fronts, that the U-Net 3+ architecture is skillful at detecting frontal boundaries. However, our “two-model” approach struggled to properly discriminate between some of the front types, and resulted in lower overall probabilities. By using one model, and through the addition of drylines, we hope to better discern the front types and improve consistency in the surface analysis process. Despite the variations in the horizontal scales relevant to different frontal and air mass boundaries, we find that the U-Net 3+ approach is resilient in its detection of these features, and the developed model is also transferable between datasets. In this presentation, we will compare the model’s performance for frontal identification between ERA5 data and real-time GDAS/GFS data.