Drought assessment has been outpaced by climate change: empirical arguments for a paradigm shift

Abstract

Despite the acceleration of climate change, erroneous assumptions of climate stationarity are still inculcated in the management of water resources in the United States (US). The US system for drought detection, which triggers billions of dollars in emergency resources, adheres to this assumption with preference towards 60-year (or longer) record lengths for drought characterization. Using observed data from 1,934 Global Historical Climate Network (GHCN) sites across the US, we show that conclusions based on long climate records can substantially bias assessment of drought severity. Bias emerges by assuming that conditions from the early and mid 20th century are as likely to occur in today's climate. Numerical simulations reveal that drought assessment error is relatively low with limited climatology lengths (~30 year) and that error increases with longer record lengths where climate is changing rapidly. We assert that non-stationarity in climate must be accounted for in contemporary assessments to more accurately portray present drought risk.

Fig. 1. Conceptual model describing the drought metric bias associated with a non-stationary climate scenario.

A theoretical accumulated precipitation dataset is presented on the horizontal axis, while the associated probability density function (PDF) is on the vertical axis. [left] Conceptual model showing one way that probability distributions can shift in time when conditions transition from a wetter, less variable state to a drier, more variable state. [right] Demonstration of how this shift can produce both a dry bias during dry times and a wet bias during wet times for a theoretical distribution.

Fig. 2. Probability distribution shift for Global Historical Climatology Network (GHCN) site USC00381770 located at Clemson University in South Carolina.

[left] Subplots show 30-year moving window values of the gamma distribution rate and shape parameters, mean precipitation and coefficient of variation (CV) of precipitation for a 30-day timescale on August 1st. Horizontal lines represent values computed using the temporally integrated distribution. [right] Probability density functions (PDFs) for each of the 30-year moving windows. The color scale represents the 30-year moving window’s final year and the black and white dashed line represents the temporally integrated PDF.

Fig. 3. Summary of simulations that fit gamma probability distributions to a known stationary distribution.

These simulations use random samples of varying climatology lengths (from 2 to 100) and the L-moment method for estimating gamma parameters. A Subplots showing the detailed simulation results for a single gamma distribution pair derived from 1000 simulations; rate and shape parameters as well as the cumulative distribution function (CDF) absolute error and Standardized Precipitation Index (SPI) absolute error are presented. Black lines and gray ribbons represent the median [interquartile] value of the 1000 simulations. B Replication of the aforementioned simulation for 100 randomly selected gamma distribution parameter pairs, each simulated 1000 times, focused on the absolute SPI error.

Fig. 4. Summary of simulation results for a non-stationary distribution using the observed 30-year moving window gamma parameters from 11 Global Historical Climatology Network (GHCN) sites.

The Standardized Precipitation Index (SPI) was estimated by fitting a gamma distribution to random samples of differing lengths, with each random sample being generated by the observed moving window generative gamma distribution for each site. The absolute SPI error was then computed using the known SPI value from the known (observed 30-year moving window) distribution for the most recent observation (e.g. 2020, consistent with operational SPI calculations). A Detailed simulation results for GHCN site USC00381770 located at Clemson University in South Carolina for the 30-day SPI for August 1st. Black lines and gray ribbons represent the median [interquartile] value of the 1000 simulations. B Results for 10 additional GHCN sites for various geographical locations and 30-, 60- and 90-day timescales for August 1st. Gray shading represents the interquartile range for simulations at all sites for each timescale.

Fig. 5. Standardized Precipitation Index (SPI) bias for Global Historical Climatology Network (GHCN) sites across the United States during periods with SPI < −2 (very dry conditions) defined using the period-of-record SPI timeseries.

Bias was computed as the median daily difference between the period-of-record SPI and the 30-year (“contemporary”) SPI from June 1 to August 31, 1991–2020. Dry bias (represented by red) denotes locations where the period-of-record reports conditions that are drier than the most recent 30 years for [left] 30-day, [middle] 60-day and [right] 90-day timescales.