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

Extreme precipitation events cause devastating societal impacts, including flooding, erosion, and agricultural damage, along with indirect health risks. Anthropogenic warming intensifies the Earth's hydrologic cycle, leading to increased extreme precipitation due to higher atmospheric moisture content (Clausius-Clapeyron relationship). However, circulation changes can either enhance or mitigate this increase. Climate models project a robust increase in extreme precipitation globally and regionally under various climate change scenarios, with increased variation between wet and dry extremes. While regional evidence of these changes has emerged, detecting human influence on extreme precipitation globally has proven difficult. Previous studies focusing on specific regions (North America, Europe, Asia, Northern Hemisphere) have detected anthropogenic influence using Detection and Attribution (D&A) methods. These methods often extract spatial or spatiotemporal patterns (fingerprints) from global climate model (GCM) ensembles. Observations are projected onto these fingerprints for signal detection, relying on long-term observations to distinguish the signal from internal variability. This approach faces challenges with extreme precipitation data due to short records, large observational uncertainty across different global datasets, and large spread in model responses to anthropogenic forcing. Traditional methods often suppress response spread by using ensemble mean changes. This study aims to fully account for uncertainties, making no assumptions about deriving the anthropogenic signal from GCM data.

Existing research on the detection and attribution of anthropogenic influence on extreme precipitation has largely relied on traditional statistical methods, such as trend analysis and fingerprint techniques. These studies, while providing valuable insights for specific regions, often face challenges in handling the large internal variability and uncertainty inherent in both climate model simulations and observational datasets. The reliance on long-term trends may also mask the presence of more complex spatiotemporal patterns of anthropogenic influence that are not captured by simple linear trends. Furthermore, the existing research often simplifies the model uncertainty by using ensemble means, potentially overlooking valuable information regarding the range of possible responses to anthropogenic forcing. This study seeks to overcome the limitations of previous approaches by employing advanced machine learning methods that can explicitly account for both internal variability and model uncertainty, allowing for a more comprehensive and robust detection of anthropogenic influence on extreme precipitation.

This study utilizes a machine learning-based method, specifically an artificial neural network (ANN), to detect anthropogenic influence on annual maximum daily precipitation (Rx1day) over global land areas. The ANN is trained to predict the year of occurrence of simulated Rx1day maps from Coupled Model Intercomparison Project phase 5 (CMIP5) and phase 6 (CMIP6) model ensembles. This supervised learning approach allows the ANN to learn the spatial patterns that best represent the external forcing (anthropogenic influence) while accounting for internal variability and model uncertainty. The ANN does not assume any specific model or model-derived quantity (e.g., ensemble mean) to represent the true anthropogenic signal; instead, it utilizes the raw GCM data, including internal variability. Layer-wise Relevance Propagation (LRP) is used to visualize and interpret the ANN's decision-making process, making the ‘black box’ nature of ANNs more transparent and physically interpretable. The relevance patterns obtained through LRP represent the ANN-identified fingerprints of anthropogenic influence on Rx1day. The trained ANN is then applied to various global land-only observational and reanalysis datasets (MSWEP, GPCC, REGEN_ALL, REGEN_LONG, ERA5, JRA55, MERRA2, CFSR, W5E5, NCEP2, 20CRv3) to detect the presence of an anthropogenic signal in these independent datasets. The study uses 51 different ANNs trained with different random combinations of GCMs to improve robustness. Pre-industrial control (piControl) simulations are used to assess the influence of natural variability and determine the statistical significance of the detected signal in observations. The signal-to-noise ratio (S:N) is calculated to quantify the strength of the detected signal.

The ANN successfully predicted the year of Rx1day maps from GCM simulations, with accuracy gradually increasing from the late 20th century, indicating the emergence of the anthropogenic signal. The ANN-identified fingerprints (relevance patterns) show positive relevance (increased Rx1day) in regions like East Asian and African monsoon regions, and North Pacific and Atlantic storm tracks. Negative relevance (decreased Rx1day) was observed in arid/semi-arid subtropical zones and some wet regions, likely due to negative dynamical components offsetting thermodynamic contributions. The signal-to-noise ratio (S:N) was higher in regions with positive relevance. The time-varying fingerprints revealed increasing relevance over time in Africa, Asia, and some regions associated with storm track shifts, indicating increasing signal strength or decreasing noise. Decreasing relevance in some areas might be due to increased model uncertainty. Analysis of the intermodel spread in predicted years showed that models predicting later years during the baseline period exhibited more future-like Rx1day patterns in their baseline climatologies. Applying the GCM-trained ANN to observational and reanalysis datasets revealed that all datasets showed a positive correlation between actual and predicted years, indicating the presence of an anthropogenic signal. While some datasets did not show a significant trend in globally averaged Rx1day, the spatial pattern analysis revealed the signal. The signal was statistically significant (95% level or higher) in most observational datasets. The study showed that the absence of a significant trend in globally averaged Rx1day does not negate the existence of an anthropogenic signal.

This study demonstrates that the detection of anthropogenic influence on extreme precipitation requires considering spatial patterns rather than relying solely on global mean trends. The use of machine learning, specifically the ANN DAI method with LRP for interpretation, successfully accounts for internal variability and model uncertainty, leading to a robust detection of anthropogenic signals in multiple global datasets, despite large systematic biases among the datasets. The results confirm and extend previous research by providing strong evidence of human influence on historical changes in extreme precipitation, even in datasets showing no significant trend in global-average Rx1day. The variation in signal magnitude across observational datasets underscores challenges in precisely constraining future projections of extreme precipitation.

This study provides robust evidence of anthropogenic influence on global extreme precipitation using a novel machine learning approach. The method successfully accounts for significant internal variability and model uncertainty, revealing a detectable signal across multiple observational datasets, even those lacking significant global mean trends. Future research could focus on decomposing the influence of individual forcings (e.g., aerosols, land use changes) and incorporating paleoclimate data to improve sampling of low-frequency natural variability.

The study acknowledges several limitations. The assessment of individual forcing influences (aerosols, land-use change, etc.) is challenging with the current framework and requires methodological modifications. The training GCMs may undersample low-frequency natural variability, potentially affecting the results. Different ANN visualization techniques could yield varying results. Despite these limitations, the ANN DAI method proves highly valuable for detecting human influence in uncertain variables like extreme precipitation.