Abstract:High-dimensional environmental systems often contain variables that are strongly predictive, structurally informative, and physically coupled, but these roles are not equivalent. In precipitation dynamics, this distinction is particularly important because rainfall emerges from multiscale thermodynamic, kinematic, microphysical, and land--atmosphere interactions, while observations are sparse, spatially redundant, and strongly imbalanced. In this work, we study the relationship between structural causal discovery and predictive sufficiency in a high-dimensional precipitation system. Using HRRR atmospheric fields and MRMS precipitation observations over the Southwestern United States, we apply a projection-based formulation of entropic regression to identify candidate causal parents of next-hour precipitation. The method evaluates variables through their incremental conditional information contribution under a one-hour temporal delay, using spatially aggregated superpixel representations to assess structural consistency across the domain. Compared with transfer entropy and causation entropy, entropic regression produces a more concentrated and stable selection profile, revealing a compact set of six physically interpretable variables associated with moisture availability, reflectivity, vertical motion, wind organization, and convective instability. Predictive experiments using only these structurally selected variables show strong short-horizon discrimination of precipitation occurrence, but limited calibration, intensity prediction, and fixed-threshold event-detection skill. These results demonstrate that structural relevance does not imply predictive closure. The selected variables form a stable informational backbone of the precipitation process, but they do not constitute a complete predictive state.
From: Abd AlRahman AlMomani [view email]
[v1]
Sat, 30 May 2026 12:38:02 UTC (1,821 KB)
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。