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

Original paper by Gavin D. Madakumbura, Chad W. Thackeray, Jesse Norris, Naomi Goldenson & Alex Hall and published by Nature Connections

Abstract

The intensification of extreme precipitation under anthropogenic forcing is robustly projected by global climate models, but highly challenging to detect in the observational record. Large internal variability distorts this anthropogenic signal. Models produce diverse magnitudes of precipitation response to anthropogenic forcing, largely due to differing schemes for parameterizing subgrid-scale processes. Meanwhile, multiple global observational datasets of daily precipitation exist, developed using varying techniques and inhomogeneously sampled data in space and time. Previous attempts to detect human influence on extreme precipitation have not incorporated model uncertainty, and have been limited to specific regions and observational datasets. Using machine learning methods that can account for these uncertainties and capable of identifying the time evolution of the spatial patterns, we find a physically interpretable anthropogenic signal that is detectable in all global observational datasets. Machine learning efficiently generates multiple lines of evidence supporting detection of an anthropogenic signal in global extreme precipitation.

Introduction

Extreme precipitation can have devastating direct societal impacts such as flooding, soil erosion, and agricultural damage¹, as well as causing indirect health risks and impacts². Anthropogenic warming acts to intensify Earth’s hydrologic cycle³. This intensification is manifested in part through increased extreme precipitation as a result of greater atmospheric moisture with warming following the Clausius–Clapeyron relationship. However, circulation changes can act to enhance or reduce this increase^4,5,6,7. Future projections by climate models following climate change scenarios show a robust increase in extreme precipitation, globally and on regional scales^8,9,10,11. Moreover, increased variation between wet and dry extremes is projected, which could have devastating societal impacts^12,13. These changes in extreme precipitation may have already become apparent on a regional basis^14,15,16.

Recent studies have detected anthropogenic influence in historical changes to extreme precipitation across the domains of North America^17,18, Europe^18,19, Asia^18,19,20, and Northern Hemisphere land areas as a whole²¹. These attempts are part of a larger category of studies known as Detection and Attribution (D&A)^22,23,24. Often, they initially extract the spatial or spatiotemporal patterns of climate-system response to anthropogenic forcing (so-called fingerprints) from an ensemble of global climate models (GCMs). Projection of observations onto these fingerprints allows for detection of the signal^24,25. The presence of a signal that can be statistically distinguished from internal variability confirms the influence of external forcing. Thus, traditional D&A methods rely on long-term observations^24,26. In the case of extreme precipitation, traditional methods may be difficult to apply globally due to inordinately short records and large observational uncertainty, reflected in multiple global datasets produced with very different assumptions^27,28,29,30. Another key difficulty with traditional methods is that the models produce a large spread in the extreme precipitation response to historical anthropogenic forcing³¹. This spread, the model uncertainty, occurs alongside large internal variability in the models’ simulations of the historical period. These two effects create significant uncertainty in the character of the true anthropogenic signal. In past research, spread in the response has been suppressed by assuming the anthropogenic fingerprint can be derived from the ensemble-mean change in extreme precipitation³². Here, we aim to take these uncertainties fully into account, by making no assumptions about how to derive the anthropogenic signal from GCM data.

Machine learning-based methods for the detection of anthropogenic influence (DAI) have been shown to overcome the reliance on trends^33,34 and are even capable of detecting the human influence from weather data on a single day³⁵. An artificial neural network (ANN) is trained to predict a proxy of external forcing (e.g., the year of the data) based on the spatial maps of the target variable from an ensemble of GCM simulations. Under this supervised learning approach, the ANN learns the spatial patterns that best represent the external forcing from the background noise arising from the internal variability and model uncertainty^33,34. Observations can then be fed to this trained ANN to assess the presence of an anthropogenic signal in observations^33,34,35. This ANN DAI method can identify the nonlinear combinations of the forced signal, internal climate variability, and intermodel variability³⁴. This method also has the advantage of being able to explicitly include internal variability and model uncertainty. It does not assume that any model or any model-derived quantity, such as the ensemble mean of the models, is the true anthropogenic signal. It uses the raw GCM data, with GCM internal variability included. In addition, novel visualization techniques allow for the interpretability of the ANNs formerly considered as black boxes, making them explainable^36,37, or interpretable in terms of physical processes or system behavior. The use of these visualization techniques alongside the ANN DAI method allows one to capture the time-varying dynamic fingerprints of each input and evaluate their physical credibility^34,38.

In this study, we apply the ANN DAI method and the ANN visualization technique known as Layer-wise Relevance Propagation (LRP)^39,40 to global maps of annual maximum daily precipitation (Rx1day) over land. Using Coupled Model Intercomparison Project, phase 5 (CMIP5)⁴¹ and phase 6 (CMIP6)⁴² model ensembles, we first aim to understand how the ANN is detecting the anthropogenic signal and interpret it physically. Then we use the ANN to detect the anthropogenic influence on Rx1day in several land-only observational and reanalysis datasets. Thus, we are agnostic about which GCM is correct, and which gridded dataset is a true representation of the observed record. In this way, we efficiently generate multiple lines of evidence as to the presence of an anthropogenic signal in the various instantiations of the observed record.

Results

ANN-identified fingerprints of anthropogenic influence

We first discuss the ability of the ANN to predict the year of occurrence for a series of simulated annual Rx1day maps. Predictions of the simulated Rx1day year (Fig. 1a, b) show that the ANN struggles during roughly the 1920–1970 period. But prediction accuracy gradually increases, noticeably starting from the late twentieth century. This characteristic, a near-constant predicted year followed by a positive trend, is consistent with the emergence of the anthropogenic signal from the noise of natural variability⁴³. Compared to when this technique is applied to global-mean temperature (ref. ³³), there is a lag in the emergence of the anthropogenic signal in extreme precipitation. This delay is likely due to larger internal and intermodel variability in extreme precipitation. We estimate this time of emergence (departure year) as the year after which the ANN prediction continuously exceeds a selected base period (1920–1949)^33,43. In GCMs, the predicted year departs from the base period in the 1970s, but the departures mostly occur later, with lower and upper quartiles of 1993 and 2014, respectively (Fig. 1c). The ANN suggests that there is a detectable anthropogenic signal in the GCM’s Rx1day during the historical period, consistent with traditional statistical methods⁴⁴.

Read more.