Evaluating Variations in Tropical Cyclone Precipitation in Eastern Mexico Using Machine Learning Techniques
Laiyin Zhu, Pascual Aguilera
Abstract
Abstract Tropical Cyclone Precipitation (TCP) is one of the major triggers of flash flooding and landslide in eastern Mexico. We apply different statistical and machine learning techniques to study a 99 years TCP climatology in high resolution. Strong correlations exist between location variables and annual mean TCP, as well as between dynamic variables and event TCP. Topographic variables observe mixed signals with the elevation variances positively correlated with TCP. The Random Forest (RF) model is a powerful tool with excellent fitting and predicting skills for TCP variations. It has a very small out‐of‐sample cross‐validation error and well captures the spatial variations of historical TCP events. Only three location variables are needed to construct the best model for the annual mean TCP while the best model needs 18 variables to explain the complex variations in the event TCP. The distance to the track is the most important variable for the event TCP model and many other factors contribute to the TCP collectively and nonlinearly, which can't be captured fully by the previous correlation analysis. They include translation characteristics of the storms, locations of the precipitation grid, and topography. Event TCP is generally larger in storms with slower translation speed and more variance in their tracks. While the lower coastal area generally has a higher probability of TCP, the higher inland has elevation variances that enhance less frequent but extreme TCP events. The RF algorithm is an efficient machine learning approach showing potentials for future Quantitative Precipitation Forecasting (QPF).