XAI4Extremes: An interpretable machine learning framework for understanding extreme-weather precursors under climate change

TL;DR

The paper addresses understanding extreme-weather precursors under climate change by proposing XAI4Extremes, a framework that links machine-identified precursors to human knowledge. It introduces a new Indochina heatwave dataset and trains a Transformer-based binary classifier on 23 atmospheric variables over a 7-day window. Using post-hoc interpretability methods, particularly Integrated Gradients, it generates relevance maps (the machine view) and compares them with domain expertise (the human view) to identify key precursors and potential climate-change fingerprints. Preliminary results show the upper-troposphere temperature at 200 hPa emerging as a growingly relevant precursor for Indochina heatwaves, consistent with composite anomalies, and the framework is extensible to other regions and extremes with potential use in adversarial data augmentation.

Abstract

Extreme weather events are increasing in frequency and intensity due to climate change. This, in turn, is exacting a significant toll in communities worldwide. While prediction skills are increasing with advances in numerical weather prediction and artificial intelligence tools, extreme weather still present challenges. More specifically, identifying the precursors of such extreme weather events and how these precursors may evolve under climate change remain unclear. In this paper, we propose to use post-hoc interpretability methods to construct relevance weather maps that show the key extreme-weather precursors identified by deep learning models. We then compare this machine view with existing domain knowledge to understand whether deep learning models identified patterns in data that may enrich our understanding of extreme-weather precursors. We finally bin these relevant maps into different multi-year time periods to understand the role that climate change is having on these precursors. The experiments are carried out on Indochina heatwaves, but the methodology can be readily extended to other extreme weather events worldwide.

Figures (12)

• Figure 1: The XAI4Extremes framework proposed, composed of a novel extreme weather dataset (a), a DL predictive model (b), an interpretability block along with its evaluation (c), that produces relevance maps, or what we called the "machine view" (d). The latter (d) is then compared with existing human expert knowledge (e) for knowledge discovery or for augmenting the dataset with e.g., adversarial samples that can shape and improve model behavior.
• Figure 2: Mean relevance of temperature at 200 hPa for the 5 historical time periods considered and on the 7 days prior to heatwaves in Indochina, for region 1 (a, top), region 2 (a, middle), region 1+2 (a, bottom), along with the corresponding relevance maps -- i.e., "machine view" -- associated to 7 days prior to heatwave, and composite anomalies -- i.e., "human view" -- (b).
• Figure 3: Indochina region used to define heatwaves (dark red).
• Figure 4: Mean relevance, positive and negative, on region 1, for all 23 variables considered, on the 7 days prior to heatwaves in Indochina and for the 5 historical time periods considered.
• Figure 5: Mean relevance, only positive (negative is set to zero) on region 1, for all 23 variables considered, on the 7 days prior to heatwaves in Indochina and for the 5 historical time periods considered. Variable t_200hPa (temperature at 200hPa) is highlighted with red colors across 5 historical time periods, while the other variables are displayed in gray.