Identification of Precipitation Propagation Factors in the North American Monsoon with Self-Organizing Maps and Hierarchical Clustering

Abstract:
Understanding the spatial structure and propagation mechanisms of precipitation during the North American Monsoon (NAM) remains critical for improving short-term forecasting in southeastern Arizona. This study applies unsupervised machine learning techniques to identify dominant precipitation regimes and their associated synoptic-scale drivers. For this study, the NAM season in Arizona is truncated to July 1st through August 31st; this is done to highlight and ensure capture of the mature phase of the NAM. Daily precipitation from the PRISM dataset for these days is analyzed over a 1° × 1° domain around Tucson, Arizona; capturing both the city itself and the surrounding mountainous terrain.

A Self-Organizing Map (SOM) is first used to classify spatial precipitation patterns based solely on the precipitation fields. To further refine regime classes and reduce pattern redundancy, weighted agglomerative hierarchical clustering is then applied to the SOM node vectors. This secondary clustering is weighted based on the number of Best-Matching Units (BMUs). The subsequent regimes are grouped into 4-6 major precipitation patterns representing the different propagation states.

These propagation states are used to generate composite fields using ERA5 reanalysis data. The composite fields show the synoptic-scale atmospheric conditions by using mid-tropospheric geopotential height, moisture transport, and wind structure. By linking surface precipitation morphology to synoptic and mesoscale atmospheric structure, this approach provides a pathway towards improved situational awareness and event extent prediction for forecasters in southeastern Arizona during the yearly monsoon.

Ongoing work focuses on refining secondary clustering and the analysis of vertical profiles and field composites.