Anthropogenic Influence on Extreme Precipitation Over Global Land Areas Seen in Multiple Observational Datasets
Detection of the anthropogenic influence on extreme precipitation is challenging due to the large natural variability associated with these extremes, and the large uncertainty in climate models that are used for this task. Observations of extreme precipitation also have large uncertainties. Previous attempts have not taken all these uncertainties into account, especially the observational uncertainty. Scientists at the University of California, Los Angeles used explainable artificial intelligence to address the above uncertainties.
Past attempts to detect the anthropogenic influence on extreme precipitation were based on long-term trends and mainly limited to a single quality-controlled dataset with limited spatial coverage. Using the time evolution of the spatial pattern of extreme precipitation, this study detects the human influence in extreme precipitation over the global land area in multiple disparate observational records. Different techniques used to create these observational datasets lead to different representations of historical precipitation. However, by observing the anthropogenic signal in all of them, we provide multiple lines of evidence of the human influence on extreme precipitation.
Although it is well established that anthropogenic climate change is influencing extreme precipitation, detecting this signal in historical records has until now been challenging. And the uncertainty that arises from observational datasets constructed using different techniques has not previously been addressed. This is mainly because traditional techniques that are used to extract the signal from noise in observations require long-spanning high-quality datasets. In this study, scientists trained an artificial neural network (ANN) to identify the changes in spatial patterns of global maps of annual maximum daily precipitation using CMIP5 and CMIP6 global climate model simulations. Using ANN visualization techniques, they assessed the physical interpretability of the ANN and obtained a physically interpretable fingerprint. Finally, they applied the trained model to historical records and found that the human influence can be seen in all the datasets they analyzed for the period 1982-2015.