Multidimensional Kernel Principal Component Analysis of False Alarms of Rapidly Intensifying Atlantic Tropical Cyclones

Abstract Previous work investigating rapid intensification (RI) predictability within Atlantic tropical cyclones (TCs) has revealed numerous challenges in identifying parameters useful in predicting RI. In particular, false alarm RI forecasts are notably high, driving down forecast skill despite higher RI detection rates. RI forecast improvement will be achieved by identifying spatial patterns distinct to RI false alarm cases that help distinguish them from correctly forecast RI events. A previously developed machine learning ensemble was used to hindcast on a database of Atlantic TCs spanning 1999 – 2016 to identify false alarm RI forecasts, as well as correct RI forecasts. Each set of cases (correct and false alarm) was used to develop kernel principal component (KPC) based composites of important meteorological features within RI TCs to facilitate meteorological comparisons against correctly forecast events. Optimality was determined using a simple metric based on a cluster analysis silhouette and compared against a baseline of traditional k-means cluster analysis, hierarchical analysis, and rotated PC analysis. Once optimal cluster configurations were identified for each meteorological feature, events within each cluster were averaged to yield composites, which in turn were used to facilitate meteorological comparisons between false alarm and correctly forecast RI events.