for better or for worse,
IDF CURVES UNDERPIN RISK ASSESSMENT
EXISTING GUIDANCE LEAVES GAPS
THE CLIMATE IS CHANGING BUT SAMPLING VARIABILITY IMPEDES TREND ESTIMATION
Fagnant et al. (2020): Rolling window estimates of the 100 year return level. Each line shows a different gauge from the same \(5^\circ \times 3^\circ\) region.
(EVEN WITH SOPHISTICATED TOOLS)
do we capture extremes and their trends?
Yuhao Liu, Guha Balakrishnan, Ashok Veeraraghavan, & James Doss-Gollin (in prep.): diffusion models for downscaling. (T) low-resolution input, (M) high-resolution model output, and (B) ground truth for 8 time steps.
NONSTATIONARY MODELS
INCREASE ESTIMATION UNCERTAINTY
More parameters, same data more uncertainty (Serinaldi & Kilsby, 2015)
SPATIALLY VARYING COVARIATES
We model statistical parameters as latent spatial fields in a hierarchical Bayesian framework, leading to improved estimates.
Histogram showing the quantile of each observed annual maximum given the posterior predictive distribution for that location and year. For a perfect model, this will be uniformly distributed. Preliminary results.
HAZARD HAS ALREADY INCREASED
Change in 100 year return level in 2022 minus 1980 for (L) 1- and (R) 24-hour precipitation. Preliminary results.
HAZARD WILL FURTHER INCREASE
Posterior expected 100 year return level for 1-hour rainfall over time. Projections use RCP6 CO\(_2\) concentrations. We compare to NOAA Atlas 14, a widely used stationary analyis (Perica et al., 2018). Preliminary results.