U-Net-based Objective Detection of Atmospheric Blocking

Atmospheric blocking is a quasi-stationary high-pressure circulation pattern that disrupts the midlatitude westerlies and is closely linked to high-impact weather extremes. Blocking detection, however, is highly method-dependent, often producing divergent blocking climatologies. This uncertainty also affects future projections, because climate-models frequently underestimate blocking relative to observations, limiting reliable assessments of blocking-related extremes. To address these challenges, we propose an objective deep learning–based framework for blocking detection that can be applied consistently across reanalysis datasets and climate model simulations.

We frame blocking detection as identifying spatial patterns in 2D atmospheric fields, analogous to semantic image segmentation, and employ a U-Net architecture to produce daily blocking masks. A two-stage training strategy is adopted: the network is first pre-trained using labels from the standard Hybrid Index (HYB; Dunn-Sigouin et al. 2013) across all seasons and then fine-tuned with a regionally modified variant, the Regional Hybrid Index (RHYB), using boreal-winter data. This strategy allows the model to incorporate regional dependence in background variability while retraining the broad blocking characteristics learned from HYB.

Although fine-tuning is restricted to boreal winter, the trained model generalizes to boreal summer and detect additional blocking events relative to HYB. When applied to the CESM2 Large Ensemble (LESN2), the framework mitigates the tendency of traditional indices to under-detect blocking frequency. Overall, this approach offers a more objective and transferable detection method that may improve the consistency of blocking diagnostics and support more reliable evaluations of blocking-related extremes in climate-model simulations.