Climate-driven changes in the predictability of seasonal precipitation

The study investigates how anthropogenic climate change will alter the predictability of seasonal precipitation over global land. Predictability at seasonal timescales largely arises from teleconnections between slowly varying sea surface temperature (SST) patterns (e.g., ENSO, PDO, AMO) and precipitation. While future changes in modes of climate variability and their teleconnections have been explored, their implications for practical predictability of land precipitation are not well known. The authors aim to quantify historical spatial patterns and drivers of seasonal precipitation predictability and to project how predictability will change by late 21st century under SSP2-4.5 and SSP3-7.0, with implications for water, agriculture, and ecosystem management.

Prior work shows seasonal precipitation over many regions is influenced by SST anomalies and indices across basins: Indian Ocean SSTs affect Australia, Africa, and parts of Asia; Pacific SSTs (ENSO) impact North and South America including Amazonia and the southern U.S.; and other modes such as PDO, AMO, and NPGO contribute to predictability. Multiple studies project changes in ENSO-related precipitation variability and teleconnections under warming, including potential amplification of ENSO-driven rainfall variability and altered teleconnections over Pacific-North America, Australia, central Africa, and western South America. However, the consequences of these evolving teleconnections for seasonal precipitation predictability have not been comprehensively assessed.

Observations and preprocessing: Global seasonal SSTs (1°×1°) from COBE-SST2 and land precipitation from GPCP (2.5°×2.5°, interpolated to 1°×1°) and GPCC (1°×1°) were combined for 1964–2014. Using the 1979–2014 overlap, GPCC (1964–1979) was linearly adjusted to match GPCP mean and variability before combining. Seasonal anomalies (relative to each season’s climatology) were computed and detrended via an 11-year centered moving average. Seasons: MAM, JJA, SON, DJF. Arid land grid cells with long-term mean seasonal precipitation <50 mm were excluded. SST modes via EOF: Empirical orthogonal function (EOF) analysis was applied to global seasonal SST anomalies to extract leading spatial modes and principal component (PC) time series. The first four PCs, collectively explaining ≥40% of global seasonal SST variance, were retained. EOF/PCs were obtained separately for observations and for each CMIP6 model and experiment (historical, SSP2-4.5, SSP3-7.0). Predictive modeling and predictability score: For each land 1°×1° grid cell and season s, a multiple linear regression predicted the seasonal precipitation anomaly y^s using previous-season (s−1) SST PCs x^(s−1). The model form y_l^s = β_0 + Σ_j β_j x_j^(s−1), with j being the two best PCs selected from the first four PCs for that grid. Model selection compared two classes: (1) using the same first n-leading PCs everywhere; (2) best-n-PCs allowing any subset of the first four PCs per grid. Five-fold cross-validation (out-of-sample) assessed skill, using the Nash–Sutcliffe Efficiency (NSE) as predictability metric. Monte Carlo shuffling established the 95% significance threshold (NSE ≈ −0.013). The best-2-PCs model provided the optimal global performance across seasons and was adopted for all analyses. Historical benchmark and model selection: Predictability maps from observations (1964–2014) established seasonal baselines and identified dominant PC sources by grid. For CMIP6, 26 models (154 ensembles) with required experiments were evaluated. Pattern correlation coefficients (PCC_NSE) between observed and simulated historical predictability maps were computed per season and ensemble. The 10 best-performing models (32 ensembles total) that most closely reproduced observed predictability patterns and leading SST EOFs were selected for future projections. Future projections and aggregation: Future predictability changes (ΔNSE) were computed as the difference between future (2049–2099) and historical (1964–2014), using the best-2-PCs model and cross-validated NSE. For each model m, ensemble means of ΔNSE_m were calculated, using climatology if historical skill < 0. Multi-model ensemble (MME) means (equal weight per model) produced ΔNSE maps for SSP3-7.0 and SSP2-4.5; stippling indicates ≥80% model agreement on the sign of change. Additional diagnostics assessed changes in SST variability (variance explained by PC1, Niño3.4 variability, ΔEOF1 patterns) and precipitation climatology and interannual variance. Robustness: All fields were bilinearly remapped to 1°×1°. Sensitivity tests with linear detrending yielded similar results to moving-average detrending, with slightly smaller predictability changes.

Historical predictability patterns: Boreal winter (DJF) precipitation is most predictable globally, with about 58% of land area showing NSE > 0; boreal spring (MAM) is least predictable, with about 49% of land area having NSE > 0 and most regions with NSE < 0.25. Predictability is generally higher in the tropics than extra-tropics. Regions with high DJF predictability include much of Amazonia (especially northeastern Amazonia), the northwestern and southern United States, southern Africa, and the Maritime Continent; northern Australia peaks in SON. ENSO-like PC1 is the dominant source in SON, DJF, and MAM, explaining 18.6% (JJA) to 31.8% (SON) of global SST variance; in DJF, PC1 contributes as a predictor over >42% of global land area with skill above climatology. PC2 resembles PDO; PC3 and PC4 resemble AMO and NPGO, though higher-order PCs can be mixed. Future projections (SSP3-7.0, 2049–2099 vs 1964–2014; similar but smaller magnitude in SSP2-4.5): - Tropics (23°N–23°S): Predictability generally increases year-round, with strongest, most consistent increases over the Maritime Continent in JJA and central Africa in DJF. Notable exceptions include declines in northern South America (northern Amazonia) during boreal winter and decreases over parts of eastern India. - Extra-tropics: Significant increases projected in central Asia (including Iran, Afghanistan, Turkmenistan) during DJF and MAM. Elsewhere, signals exist but inter-model variability limits robustness. Drivers and related changes: CMIP6 models project increased dominance and variability of ENSO: increases in variance explained by PC1 (MME increase largest in SON: 2.3 ± 4.8%) and in Niño3.4 SST variability; ΔEOF1 indicates weakened equatorial central/eastern Pacific loading in MAM but enhancement in SON and DJF. Many regions show increased interannual precipitation variability in all seasons (with exceptions: modest decreases in southern Africa and eastern South America in SON). These changes in SST variability and teleconnections, and modified atmospheric propagation efficiency to extratropics, likely underlie predictability changes. Implications and regional risks: Northern South America faces compounding risks: projected dry-season (DJF) precipitation declines, increased interannual variability, and decreased seasonal predictability, elevating wildfire and drought risk and challenging seasonal risk management. Conversely, increased predictability over parts of Africa and the Maritime Continent may provide opportunities to improve seasonal water and ecosystem management.

The work demonstrates that anthropogenic warming alters SST variability and land–atmosphere teleconnections in ways that reshape practical predictability of seasonal precipitation. By quantifying historical predictability constrained by previous-season SSTs and projecting changes with a CMIP6 model subset that reproduces observed patterns, the study links increased ENSO dominance and widespread increases in interannual variability to spatially heterogeneous predictability changes. Increased predictability in many tropical regions (except northern South America) and in parts of central Asia (DJF, MAM) suggests potential to enhance seasonal forecasting utility, whereas declines in key hotspots (e.g., northern Amazonia DJF) will hinder anticipatory water, agriculture, and fire management. These findings directly address the research question by identifying where and when predictability will likely increase or decrease, connecting changes to evolving SST modes, and highlighting management-relevant consequences.

This study provides a global, seasonally resolved assessment of how climate change is projected to alter the predictability of seasonal precipitation derived from previous-season SSTs. Using an EOF-based reduction of SST variability and cross-validated best-2-PC linear prediction, validated against observations and a selected subset of CMIP6 models, the authors find robust increases in predictability across much of the tropics (notably the Maritime Continent and central Africa) and central Asia in boreal winter and spring, but declines in northern South America during boreal winter. Enhanced ENSO variability and increased interannual variability of both SSTs and precipitation are key contributors. The results identify regions where seasonal water and ecosystem management can leverage improved predictability, and others where diminished predictability will pose challenges. Future research should: (i) use idealized experiments and causal network approaches to isolate teleconnection mechanisms under warming; (ii) assess the role of SST persistence at seasonal lead times; (iii) explore nonlinear and nonstationary predictors and multi-source predictors (e.g., soil moisture, stratospheric signals); and (iv) evaluate predictability at different lead times and subseasonal windows.

Predictability is estimated via linear models using only previous-season global SST PCs; true intrinsic predictability cannot be directly measured. The linear framework may miss nonlinear and nonstationary SST–precipitation relationships. Higher-order PCs can mix physical modes, potentially reducing physical interpretability; while rotation could aid interpretation, it likely would not change linear predictive skill. Analyses exclude arid regions and rely on a single predictand (seasonal totals) and a single lead (one season). Observational uncertainties (especially in dry regions) and dataset blending/adjustment may affect historical benchmarking. Model selection focuses on reproducing historical predictability patterns; projections remain uncertain where model spread is high. Changes in SST persistence (ocean memory) at seasonal lead times and altered atmospheric teleconnection efficiency introduce additional uncertainties. Results are not directly comparable to potential precipitation predictability (PPP) metrics based solely on precipitation dynamics.