Anthropogenic climate change has reduced drought recovery probabilities across the western US

The study addresses when and how droughts recover in the western United States, a critical question for water resource managers who must plan allocations and restrictions. The region’s hydroclimate features cool, wet winters and hot, dry summers, with high interannual variability. Recent decades have brought persistent and severe drought conditions, driven by natural variability and amplified by anthropogenic climate change, which increases atmospheric evaporative demand and the co-occurrence of hot and dry conditions. While forecasting tools and historical analog methods can estimate drought recovery probabilities (PDR), they assume stationarity that may no longer hold. The authors hypothesize that climate change has reduced the probability and speed of drought recovery once drought conditions are established. The study’s purpose is to quantify changes in PDR across five major western US basins and attribute observed changes to anthropogenic climate change.

Prior work documents multiple drought types and varied definitions of recovery, emphasizing the need for surplus moisture beyond current demand to resolve accumulated deficits. Studies show that warming has increased evaporative demand, intensified droughts (including flash droughts), and raised the likelihood of compound hot-dry extremes in the western US. Hydrologic and seasonal forecasts, as well as analog approaches (e.g., ESP and soil moisture analogs), can provide skill up to 12 months, but their reliance on historical stationarity is problematic under climate change. Modeling studies project increased drought frequency and duration under warming, and observational attributions indicate substantial anthropogenic contributions to recent southwestern megadrought conditions. However, prior to this work, no study explicitly examined how climate change has altered the probability of recovery once drought is underway.

Data: The analysis used monthly gridded (0.25°) observational climate data (precipitation, Tmin, Tmax, wind speed, actual vapor pressure, downward shortwave radiation; 1901–2023) to compute specific humidity and reference evapotranspiration (ETo) via the ASCE Penman–Monteith method. Self-calibrating Palmer Drought Severity Index (scPDSI) was then calculated using 1950–2016 as the reference period. Model-based data came from 23 CMIP6 GCMs (130 ensemble members) for historical (1850–2014) and SSP2-4.5 scenarios. Five western US HUC2 basins were analyzed: Upper Colorado (UCB), Lower Colorado (LCB), Great Basin (GB), Pacific Northwest (PNW), and California (CA). Model data were bilinearly interpolated to 0.1° and spatially averaged by basin. Counterfactual construction: To estimate a world without anthropogenic climate change, the multi-model mean trends in precipitation, near-surface air temperature, wind speed, specific humidity, and shortwave radiation from CMIP6 were subtracted from the observed fields (first-order monthly mean trends only), preserving observed variability. Counterfactual scPDSI was calibrated to observations. Experimental dataset: To control for initial drought severity and increase sample size, the authors initialized each calendar month with severe drought (scPDSI = −4; soil moisture at the 2nd percentile) and drove 24-month scPDSI evolutions using basin-mean observed or counterfactual precipitation and ETo. Drought/recovery definitions and PDR: Drought was defined as scPDSI below the 10th percentile and recovery as scPDSI above the 30th percentile for each month based on a 1901–1980 baseline (aligned with USDM thresholds: <10th ≈ D2–D4; >30th = no drought). For each drought start date, the probability of drought recovery (PDR) was computed for each of the following 24 months as the fraction of cases with scPDSI ≥ recovery threshold in the month of interest. PDR was computed separately for observations, counterfactual observations, each model ensemble, and their multi-model mean. Time periods included historical (1901–1980), recent (2000–2021), and contemporary (model-based 2000–2040). Statistical analysis: Differences between observed and counterfactual traces used McNemar’s test (paired binaries) with p<0.05. Historical vs recent observational traces used t-tests on monthly scPDSI distributions. For models, significance was assessed by bootstrapping across 23 models to form 95% CIs; recent (or contemporary) mean PDR falling outside historical CI indicated significance. Moving-window PDR (21-year windows every 5 years) was used to examine variability and estimate Time of Emergence (ToE) per model as the first time PDR falls and remains below the 1900–1980 average through 2050; the multi-model ToE is the median across models. Seasonal attribution used differences in PDR between historical and contemporary periods by month and season (JFM, AMJ, JAS, OND), with Welch one-way ANOVA testing for warm (Apr–Sep) vs cool (Oct–Mar) differences. Contemporary minus historical anomalies in precipitation and ETo were also assessed. Months-to-recovery were tabulated over 24 months after drought onset (non-recovery within 24 months assigned 25); medians were computed per simulation and aggregated to model medians. Model screening: Climatological biases and trends in tas, pr, and ETo were reviewed; no CMIP6 models were culled.

• Observed decreases in PDR: Across California and the three southwestern basins (UCB, LCB, GB), PDR 18 months after severe drought declined from 56–65% (1901–1980) to 30–47% (2000–2021), i.e., roughly 25–50% lower. PNW shows limited observed change due to strong internal variability and historically low PDR in the 1920s–1930s.
• Attribution: Counterfactual analyses indicate that in southwestern basins, about one-third to all of the observed PDR decrease at month 18 is attributable to anthropogenic climate change. The primary mechanism is increased ETo causing larger P − ETo deficits in non-winter months.
• Model-based evidence: CMIP6 multi-model means show significant PDR reductions in recent decades for several regions. At month 18, modeled PDR is lower by about 17% (±6%) in PNW and 13% (±5%) in CA, and ~0–10% lower in southwest regions. When defining the contemporary period as 2000–2040, all regions show significant reductions.
• Time of Emergence (model-based medians): CA ~1985 (1975–1995), PNW ~1980 (1970–1990), LCB ~1995 (1985–2005), UCB ~2010 (2000–2020), GB ~2015 (2005–2025). All regions exhibit low PDR after 2000.
• Seasonal mechanisms: PDR recovery rates have slowed in non-winter months (AMJ/JAS) across most regions due to significant increases in ETo, reducing P − ETo. In CA and PNW, modest winter improvements in PDR (0.1–1.4 percentage points per month increase) are simulated due to slight increases in winter precipitation, but these gains do not compensate for non-winter losses.
• Increased time to recovery: Under contemporary climate (2000–2040), models indicate droughts take ~1–4 months longer to recover (median ~13–16 months vs ~12 months historically). This lengthening is significant (p<0.05) for all regions except GB in models; experimental results suggest GB also experiences lengthening due to climate change.

The study demonstrates that the probability of drought recovery has declined across the western US, addressing the central hypothesis that anthropogenic climate change has slowed drought recovery. Multiple lines of evidence—observations, controlled experiments initialized at equal severity, and CMIP6 simulations—converge to show reduced PDR, particularly 18 months after onset, with a substantial anthropogenic component via enhanced evaporative demand. Seasonal analysis clarifies that warming-driven increases in ETo during the warm season are the dominant driver of slower recovery, while any winter precipitation gains, notably in coastal basins, are insufficient to offset non-winter losses. These results imply longer drought durations and greater persistence once droughts begin, complicating water management that relies on historical analogs and timing of recovery. Management implications include increased risk of water supply shortfalls (e.g., ‘day zero’ scenarios) and ecological stress, underscoring the need to integrate non-stationarity into planning, enhance conservation and adaptive agricultural practices, and expand storage and recharge to bridge extended recovery times. The findings also highlight regional contrasts: clearer observational attribution in the Colorado basins versus stronger modeled signals in coastal regions, likely reflecting higher internal precipitation variability and anthropogenic shifts in variability along the coast.

This work quantifies and attributes a significant, climate change–driven decline in the probability and pace of drought recovery across five western US basins. PDR is ~25–50% lower in recent decades than in early 20th century baselines, with at least one-third of this reduction attributable to anthropogenic climate change, primarily via increased evaporative demand outside winter. Models indicate recovery now takes ~1–4 months longer, a trend expected to continue, emphasizing the need for drought planning that accounts for longer durations and reduced recovery likelihood. Practical responses include demand reduction during droughts, adaptive agricultural systems, and managed aquifer recharge to buffer extended deficits. Future research should evaluate drought recovery using additional indicators (e.g., soil moisture, streamflow, snowpack) with differing sensitivities and memory, assess anthropogenic effects on higher-frequency variability and extreme precipitation, and refine land-surface and arid-region feedbacks in models to improve PDR projections and attribution.

• Metric sensitivity: scPDSI captures precipitation and ETo but omits critical hydroclimatic processes such as snowpack dynamics and snow–rain partitioning; results may therefore be conservative regarding reduced recovery.
• Counterfactual scope: The counterfactual removes only first-order monthly mean anthropogenic trends and does not alter higher-frequency variability or extremes, potentially understating anthropogenic influences on variability.
• Internal variability and sample size: High multidecadal variability in observations and limited drought samples in some windows constrain detection of trends and attribution, especially in PNW and coastal basins.
• Model uncertainties: CMIP6 biases and uncertainties in precipitation trends and land-surface feedbacks in arid regions/seasons may affect modeled PDR changes and regional differences (e.g., GB).
• Spatial aggregation: Basin-mean analysis may mask important sub-basin heterogeneity in drought and recovery processes.