6.8 Statistical Modeling of Extreme Precipitation with TRMM Data

Tuesday, 9 January 2018: 3:15 PM
Room 18B (ACC) (Austin, Texas)
Levon Demirdjian, Univ. of California, Los Angeles, CA; and Y. Zhou and G. J. Huffman

Handout (1.0 MB)

This paper improves upon an existing extreme precipitation monitoring system

based on the Tropical Rainfall Measuring Mission (TRMM) daily product

(3B42) using new statistical models. The proposed system utilizes a regional

modeling approach, where data from similar grid locations are pooled to increase

the quality and stability of the resulting model parameter estimates to

compensate for the short data record. The regional frequency analysis is divided

into two stages. In the first stage, the region defined by the TRMM measurements

is partitioned into approximately 28,000 non-overlapping clusters

using a recursive k-means clustering scheme. In the second stage, a statistical

model is used to characterize the extreme precipitation events occurring in

each cluster. Instead of applying the block-maxima approach used in the existing

system, where the Generalized Extreme Value probability distribution is

fit to the annual precipitation maxima at each site separately, the present work

adopts the peak-over-threshold method of classifying points as extreme if they

exceed a pre-specified threshold. Theoretical considerations motivate using

the Point Process framework for modeling extremes. The fitted parameters

can be used to estimate trends and to construct simple and intuitive average recurrence

interval (ARI) maps which reveal how rare a particular precipitation

event is given its location. The new methodology eliminates much of the random

noise that was produced by the existing models due to a short data record,

producing more reasonable ARI maps when compared with NOAA’s longterm

Climate Prediction Center ground-based observations. Furthermore, the

proposed methodology can be applied to other extreme climate records.

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner