Data-Driven Extraction of Spatiotemporal Causal Networks of Extreme Snowfall Based on Long-Term Observations in the Japanese Archipelago

hiroshi matsuda , Shin'ichi HOMMA

Understanding the spatiotemporal dynamics and long-term trends of extreme weather events is crucial for climate risk assessment. However, extracting causal network structures and underlying trends from observational data remains challenging due to large-scale common influences, spurious correlations, and historical missing values.

In this study, we propose a generalized data-driven framework for extracting spatiotemporal causal network structures and underlying trends by systematically integrating statistical and machine learning techniques. The framework consists of two main stages. First, it employs Graphical Lasso and Markov blanket extraction to reduce the influence of large-scale common factors, followed by the identification of causal structures using a Linear Non-Gaussian Acyclic Model (LiNGAM). Second, it applies robust regression and state-space modeling (Kalman filtering) to mitigate the effects of extreme anomalies and missing values, enabling the extraction of smooth non-linear trends.

As a representative case study, the framework is applied to over a century of annual maximum snow depth records (135 years) in the Japanese archipelago. The results reveal distinct network structures and provide a visualization of meso-scale causal propagation patterns associated with the Japan-Sea Polar-airmass Convergence Zone (JPCZ), as well as localized seesaw relationships influenced by topography. Furthermore, the state-space model identifies a decadal regime shift in the late 1980s, indicating a non-linear decline in extreme snowfall associated with temperature thresholds between snow and rain. These findings demonstrate the applicability of the proposed framework for analyzing complex spatiotemporal dynamics of extreme weather.