Anthropogenic influence on extreme precipitation over global land areas seen in multiple observational datasets

The study investigates whether anthropogenic forcing has influenced extreme precipitation (annual maximum daily precipitation, Rx1day) over global land areas and whether this signal can be robustly detected in observations despite large internal climate variability, model uncertainty, and observational dataset inconsistencies. Extreme precipitation has substantial societal impacts and is expected to intensify with warming via thermodynamic moisture increases, modulated by circulation changes. Traditional detection and attribution methods rely on long records and often assume the multimodel mean represents the forced fingerprint, a problematic assumption for extreme precipitation given short and uncertain records and large intermodel spread. The authors propose using a machine learning approach that leverages spatial patterns and their time evolution to isolate anthropogenic signals from noise without presupposing a specific model or ensemble-mean fingerprint, aiming to provide multiple lines of evidence across diverse observational datasets.

Prior research has projected robust global and regional increases in extreme precipitation under climate change scenarios, with enhanced variability between wet and dry extremes. Detection and attribution (D&A) studies have identified anthropogenic influence on extreme precipitation in North America, Europe, Asia, and Northern Hemisphere land areas using regression-based fingerprints derived from climate models. However, these approaches face challenges for global extremes due to short observational records, heterogeneous gridding methods, and large model spread in the response of Rx1day. Trend-based detection may miss signals where internal variability masks global-mean changes. Machine learning-based DAI approaches have recently demonstrated the ability to extract forced signals from spatial patterns and even detect anthropogenic influence from single-day weather data. These methods can incorporate internal variability and intermodel differences explicitly and can be interpreted using explainable AI tools, addressing shortcomings in traditional D&A. The study builds on these advancements to assess global land Rx1day using multiple observational and reanalysis datasets.

• Data and preprocessing:
  • Climate models: Daily precipitation from CMIP5 and CMIP6 historical simulations concatenated with high-emissions future projections (CMIP5 RCP8.5; CMIP6 SSP5-8.5), spanning 1920–2099. Data regridded to 2°×2°; Rx1day computed annually at each land grid point (Antarctica and oceans masked), yielding 6082 land grid inputs per year.
  • Internal variability estimate: Pre-industrial control (piControl) simulations from 20 CMIP6 models. Extracted 220 non-overlapping 34-year segments (after removing first three segments to avoid drift) to characterize natural variability, matching the observational analysis period length.
  • Observations/reanalyses (1982–2015, global land coverage): MSWEP v2; GPCC v2018; REGEN_ALL; REGEN_LONG; ERA5; JRA55; MERRA2; CFSR; W5E5 (bias-adjusted ERA5); NCEP2; 20CRv3. All regridded to 2°×2° and Rx1day computed annually.
• ANN detection framework:
  • Task: Supervised regression predicting the year from vectorized Rx1day spatial maps (6082 inputs) so the network learns time-evolving forced patterns amid internal variability and model uncertainty.
  • Architecture: Fully connected ANN with two hidden layers (10 units each), ReLU activation. Loss: mean squared error between actual and predicted year. Optimizer: rmsprop.
  • Regularization and training: L2 regularization (λ=0.001) on weights between inputs and first hidden layer to mitigate spatial autocorrelation overfitting. Trained for up to 1000 epochs with early stopping (patience 50) based on validation loss.
  • Ensemble of ANNs: 51 separate ANNs trained on different random splits of GCMs (∼60% of models, 26, for training; 9 for validation; 9 for testing) to sample model uncertainty.
• Explainable AI and fingerprint extraction:
  • Layer-wise Relevance Propagation (LRP, αβ-rule with α=2, β=1) used to attribute the ANN’s predicted year back to input grid cells, yielding relevance maps (positive values increase predicted year; negative decrease). Relevance is conserved and sums to the output prediction.
  • Time of emergence (departure year) estimated as the first year the predicted year exceeds the 1920–1949 base-period threshold continuously.
  • Signal-to-noise (S:N) diagnostics: Signal defined as multimodel mean change in Rx1day (2070–2099 minus 1920–1949). Noise estimated as (i) internal variability: multimodel mean of base-period Rx1day standard deviation; (ii) intermodel variability: intermodel standard deviation of the signal. Changes in internal variability (dIV) and intermodel variability (dMV) computed between periods (with forced trend removed via regression on 41-year lowess-filtered global-mean surface temperature) to interpret time-varying relevance.
• Detection metrics in observations/reanalyses:
  • For each dataset and ANN, compute correlation (r) between actual and predicted years, and the regression slope of predicted vs actual year (indicator of signal strength). Compare to distributions from piControl segments to assess significance (S:N via z-test; significance when S:N>1.96 corresponds to 95%).

• ANN learns detectable, physically interpretable fingerprints of anthropogenic influence in simulated Rx1day:
  • Predicted-year accuracy improves markedly after mid-to-late 20th century, consistent with signal emergence amid variability. Departure (time of emergence) quartiles across GCMs lie at 1993 (25th percentile) and 2014 (75th percentile), with some departures beginning in the 1970s.
  • Relevance patterns (1982–2015 mean) highlight positive relevance in East Asian and African monsoon regions and along North Pacific and North Atlantic storm-track land regions. Negative relevance appears in arid/semi-arid subtropics (Northern Africa, Middle East, southern South Africa, arid Australia) and some wet regions of central/northwestern South America.
  • Regions with negative relevance exhibit lower S:N due to higher internal variability and larger intermodel spread, consistent with weaker or inconsistent Rx1day increases where dynamic changes offset thermodynamic moistening.
• Time-varying fingerprints:
  • From 1920–1949 to 2070–2099, relevance increases over Africa and Asia (linked to strengthening monsoons), and over North Pacific/Atlantic land regions (consistent with poleward storm-track shifts). Some Mediterranean-climate regions in South Africa and South America show increased relevance (associated with subtropical drying patterns). Decreases in relevance occur over much of South America and the western United States, likely tied to increasing model uncertainty.
  • Grid cells with increasing relevance show similar forced Rx1day changes but much smaller increases in internal and intermodel variability compared to cells with decreasing relevance, aligning relevance changes with the evolving S:N tradeoff.
• Origins of intermodel spread in predicted year:
  • Models with later predicted years during the 1920–1949 baseline have higher baseline Rx1day climatologies in monsoon regions, making their historical patterns more "future-like" relative to the ANN’s learned fingerprint, thus yielding later predictions.
• Detection in observations and reanalyses (1982–2015):
  • Despite only 7 of 11 datasets exhibiting significant positive global-mean Rx1day trends (0.02–0.09 mm/day/decade), all 11 observation/reanalysis datasets show positive correlations between predicted and actual year that exceed those expected from natural variability, indicating detection of an anthropogenic signal via spatial patterns.
  • Slopes (signal strength) vary across datasets but generally align with GCM testing sets. Statistical significance (based on S:N relative to piControl variability):
    • Observations: MSWEP, GPCC, REGEN_ALL at 95% significance; REGEN_LONG at 84%.
    • Reanalyses: ERA5 and CFSR at 90% significance; JRA55, MERRA2, W5E5, NCEP2, and 20CRv3 at 95%.
  • ANN-based detection provides multiple lines of evidence for anthropogenic influence in global terrestrial Rx1day across diverse observational products, even where global-mean trends are not significant.

The machine learning detection framework successfully isolates anthropogenic fingerprints in Rx1day by exploiting the spatial structure and its time evolution, overcoming obstacles posed by short records, large internal variability, and intermodel spread. The relevance patterns emphasize regions where forced changes are robust (monsoon belts, storm-track regions) and de-emphasize areas with low S:N (arid subtropics, parts of South America). Applying the GCM-trained ANNs to observations reveals consistent ordering of years (high correlations with actual year) across all datasets, indicating that the observed spatial evolution resembles the learned forced signal even when absolute climatologies differ and global-mean trends are weak. The results directly address the research question by demonstrating detectable anthropogenic influence on global land extreme precipitation in all observational datasets considered, with signal strengths comparable to those in the models. This underscores the utility of pattern-based detection over trend-only analyses for extremes and provides convergent lines of evidence across independent datasets, strengthening confidence in human influence on extreme precipitation.

The study demonstrates that an ANN trained on CMIP5/6 Rx1day spatial maps, interpreted via LRP, learns physically credible, time-evolving fingerprints of anthropogenic influence and detects this signal across all global land observational and reanalysis datasets examined for 1982–2015. Key contributions include: (1) explicit incorporation of internal variability and intermodel uncertainty without assuming the multimodel mean fingerprint; (2) explainable AI diagnostics linking fingerprints to known physical changes (monsoon intensification, storm-track shifts, subtropical drying); and (3) robust detection in observations despite dataset biases and weak global-mean trends. Future work should aim to disentangle individual forcings within the ANN framework, augment training with paleoclimate data to better sample low-frequency variability, evaluate sensitivity to different explainability methods, and address potential model underestimation of responses to natural forcings and modes such as ENSO. These steps could improve constraints on projections of extreme precipitation.

• Attribution to individual forcings (e.g., aerosols, land-use change, volcanic, solar) is not performed and is challenging within the current ANN regression framework.
• Training datasets may undersample low-frequency internal variability (e.g., AMV, PDO); inclusion of paleoclimate data could help.
• Potential underestimation in GCMs of precipitation responses to natural forcings and variability (e.g., volcanic eruptions, ENSO) may affect learned fingerprints.
• Sensitivity to the choice of ANN visualization method (LRP variant) is not explored; other interpretability techniques could yield different emphasis.
• Observational products exhibit large systematic biases and processing differences; while detection is robust, magnitude estimates vary substantially across datasets, limiting constraints on future projections.