STAR-ESDM: A Generalizable Approach to Generating High-Resolution
Climate Projections through Signal Decomposition
Abstract
High-resolution climate projections provide crucial insights into
assessing climate risk and developing climate resilience strategies. The
Seasonal Trends and Analysis of Residuals empirical statistical
downscaling model (STAR-ESDM) is a computationally-efficient and
flexible approach to generating high-resolution climate projections that
can be applied globally using a broad range of predictands and
predictors that can be sourced from weather stations, gridded datasets,
satellites, reanalysis, and global or regional climate models. It uses
signal processing combined with Fourier filtering and kernal density
estimation techniques to decompose and smooth any quasi-Gaussian time
series, gridded or point-based, into multi-decadal long-term means
and/or trends; static and dynamic annual cycles; and probability
distributions of high-frequency variability. Long-term predictor trends
are bias-corrected and predictor components are used to map remaining
predictand components to future conditions. Components are then
recombined for each station or grid cell to produce a continuous,
high-resolution bias-corrected and downscaled time series at the spatial
and temporal scale of the original time series. Comparing STAR-ESDM
output with high-resolution daily temperature and precipitation
projections generated by a fully dynamical global model demonstrates
that the method is extremely robust, capable of accurately reproducing
projected changes for all but the most extreme temperature and
precipitation values. For most continental areas, biases in 1-in-1000
hottest and coldest temperatures are less than 0.5°C and biases in the
1-in-1000 wet day precipitation amounts are less than 5 mm/day. As
climate impacts intensify, STAR-ESDM represents a significant advance in
generating consistent high-resolution projections to comprehensively
assess risk and optimize resilience.