Detecting Weather Fronts from Satellite Observations: What Does Reanalysis Add?

Weather fronts are transition regions between air masses with contrasting thermodynamic properties, associated with mid-latitude precipitation and severe weather. We investigate machine learning (ML) approaches for detecting weather fronts and quantify the impact of reanalysis data on performance. We compare two configurations of the same ML model for pixel-wise frontal classification over the North Atlantic and Europe: (1) only using geostationary infrared satellite observations, (2) incorporating 28 ERA5 reanalysis variables in addition to the satellite observations. We find that our satellite-only model achieves a neighbourhood critical success index at a 150km matching radius (CSI@150) of 0.402, within the range of agreement among independent meteorological agencies, while the satellite+ERA5 model achieves 0.575 (Δ = +0.173, 95% CI [0.163, 0.183]). Our analysis shows that near-surface winds and moisture are least predictable from the satellite representation (R^2 = 0.23 for winds compared with 0.88 for temperature) and that their removal produces the largest decreases in frontal detection performance.