AI Unravels Extreme Precipitation Dynamics: CNN-Based Analysis

By Dr Silpaja Chandrasekar, PhDReviewed by Susha Cheriyedath, M.Sc.Dec 19 2023

In a paper published in the journal Communications Earth & Environment, researchers introduced a groundbreaking approach to analyze the link of extreme precipitation to climate shifts. By employing a convolutional neural network (CNN) trained using data from 10,000 precipitation stations, this method skilfully captured intricate spatial parameters within a generalized extreme value model, showcasing sensitivity to annual mean and global mean surface temperature.

This innovative technique bypasses the limitations of traditional statistical analyses, offering robust insights into the sensitivity of extreme precipitation to climate change. The resulting high-resolution maps for North America, Europe, Australia, and New Zealand unveil remarkable spatial variability, providing a nuanced understanding of this critical relationship.

Background

Extreme precipitation events (EPEs) stand as some of the most impactful climate-related hazards, prompting significant societal concern regarding shifts in their frequency and intensity. A common rule-of-thumb among practitioners, like civil engineers, relies on the Clausius–Clapeyron relationship (CCR), suggesting a 1°C temperature rise corresponds to a 7% surge in atmospheric water capacity and subsequently a 7% escalation in EPE severity. Climate change impacts numerous precipitation-influencing processes, notably affecting local and large-scale dynamics that alter precipitation patterns, significantly diverging from the notion suggested by the Clausius–Clapeyron relationship.

CNN-Based Precipitation Analysis Evaluation

The study utilized daily precipitation data obtained from the Global Historical Climatology Network (GHCN) and supplemented with data from the New Zealand National Climate Database (CLIDB) due to the decline in GHCN sites. This analysis focused on four target regions: North America, Europe, Australia, and New Zealand, covering data from 1950 to 2019, with specific periods for evaluation and training. The GHCN+CLIDB dataset provided annual maximum daily precipitation values for block maxima calculations.

The selection of sites for CNN training involved a systematic process, starting with eight chosen sites representing diverse climates and population centers. Subsequently, 10,000 sites were selected for CNN training, ensuring geographical diversity and at least ten years of block maxima data. Annual mean global mean near-surface temperature data from HadCRUT.5.0.1.055 served as a climate covariate to model non-stationarity in Generalized Extreme Value (GEV) distributions.

The CNN architecture, depicted in Supplementary Fig. S5, integrated features from different spatial resolutions (1/256∘ to 1/4∘) to capture the spatial morphology of precipitation at varying scales. This architecture comprised four parallel branches, employing ResNet blocks to facilitate deep learning while addressing vanishing gradient issues. The CNN training process involved optimization techniques like RAdam, Lookahead, and a cosine decay schedule, executed over multiple epochs, and was tailored separately for each target region.

Evaluation of CNN's performance involved assessing its ability to capture non-stationarity in GEV distributions and spatial generalization. The negative log-likelihood across evaluation sites gauged the CNN's capacity to represent precipitation extremes against traditional direct GEV fits. The evaluation highlighted CNN's superior performance in North America and Europe, although Australia and New Zealand showed mixed results due to unique site-specific anomalies.

Another aspect of the evaluation involved comparing CNN's performance in estimating rare events (One-in-1000-year events) against direct GEV fits. The CNN-derived estimates proved more consistent with the total observations, suggesting the CNN's enhanced ability to model extreme events compared to traditional natural fits, mainly due to its exposure to a more extensive range of training data.

Novel CNN Approach: Extreme Precipitation

The study utilized a CNN approach to analyze extreme precipitation patterns, focusing on 10,000 GHCN+CLIDB sites between 1960 and 2019, training the CNN to simulate extreme precipitation events under varying global mean surface temperatures. However, limitations arose from inhomogeneous data coverage, leading to CNN's training using data only from specific regions (North America, Europe, Australia, and New Zealand). It resulted in reported results particular to these areas.

One primary output was the derivation of 1% annual exceedance probability (AEP) precipitation levels, highlighting that these extremes tend to be around 2.5 times the magnitude of typical maximum yearly 1-day precipitation depths. The CNN effectively learned regional variations, associating higher precipitation extremes with specific geographical features like coastal areas along warm oceans and the windward sides of mountain ranges.

Another key finding was CNN's capability to infer the sensitivity of extreme precipitation to climate change. It diverged from the conventional Clausius–Clapeyron relationship (CCR) expectation of a 7% increase per 1°C of warming due to the impact of local and large-scale dynamics on regional precipitation responses. These dynamics included factors like sea surface temperature variations, changes in large-scale atmospheric circulation, and shifts in storm tracks, influencing precipitation patterns.

While the CNN-based approach demonstrated its ability to learn from observational data, there were caveats. The model's limitations lay in deriving precipitation sensitivities solely from historical observations, potentially ignoring future non-linearities in the climate system. It raised the need to augment observational data with transient model simulations, allowing CNN to update its understanding of precipitation non-stationarity.

Overall, this innovative CNN-based methodology presents an additional avenue for projecting changes in extreme precipitation. Acknowledging its role as one of several necessary approaches, it contributes to robust climate change projections. As AI-based methods evolve and integrate intrinsic knowledge of climatic physics, they hold promise in amalgamating historical observations with scientific understanding for more reliable future climate change assessments.

Conclusion

To sum up, this innovative CNN-based approach enriches understanding of extreme precipitation and its response to climate change. While offering valuable insights, it's evident that this method complements a spectrum of approaches essential for robust climate projections. Continuing advancements in AI-driven methodologies hold immense potential in merging historical observations with scientific understanding, reinforcing the reliability of future climate predictions.