Explainable Frontal Boundary Predictions for Applications in Operational Environments

Abstract

Frontal boundaries drive many high-impact weather events around the globe. Identifying fronts through various thermodynamic fields increases predictability of hazardous weather phenomena. Frontal analysis is still primarily done by human forecasters, often implementing their own subjectivity rules and criteria for determining frontal positions and types. Subjective placements of fronts can result in various solutions by different forecasters when given identical sets of data. Numerous studies have attempted to make frontal analysis more consistent through numerical frontal analysis, using sets of rules and thresholds with thermodynamic fields to locate and classify fronts. In recent years, machine learning algorithms have gained more popularity in meteorology due to their ability to learn complex relationships within large quantities of atmospheric data. We present a novel machine learning algorithm that predicts five different types of frontal boundaries - cold, warm, stationary, and occluded fronts and drylines. The algorithm was able to locate 76-86\% and 70-81\% of fronts over CONUS and NOAA's Unified Surface Analysis domain, respectively, on an independent testing dataset. We applied two Explainable Artificial Intelligence methods to the model - permutation studies and saliency maps. Permutation studies allowed us to determine variable importance for each frontal type. Saliency maps for the selected case study gave us insight as to how the model output can change as the ambient environment is modified. While more work needs to be done to improve the algorithm, we have demonstrated that machine learning can be used to develop an accurate and efficient model for detecting frontal boundaries.