A Data-Driven Framework for Characterizing Nonlinear Responses of River Ice Growth and Decay in the Shisifenzi Reach of the Upper Yellow River

Abstract

The Shisifenzi Reach, with its sharp curvature, serves as a prime observational target for river ice research in the Upper Yellow River. Using six winter seasons (2020–2026) of monitoring data, this study developed a Random Forest with SHAP and Accumulated Local Effects (ALE) to characterize nonlinear responses river ice growth and decay to 26 hydro-thermal features. Distinctive contributions are: (1) The 30-day cumulative freezing degree-hours is the dominant driver, and a threshold near 4000 °C·h separates slow growth from rapid thickening. (2) Water level elevation, though indirect, shows an obvious nonlinear link: ice is thickest when it stays in the 988.5–989 m range; above this, the ALE effect drops—often because flow velocity disturbance is slight in the range of 988.5–989 m and high water level mainly coincides with warmer temperature. (3) Short-term thermal features show little effects. Other thermal variables and feature interactions also show saturation. (4) Combining the ALE marginal effect thresholds with observed freeze-up data clarifies the two-stage nonlinear mechanism. In the final decay stage, ice disappears faster than temperature variations alone explain. The data-driven framework provides interpretable, site-specific information useful for river ice management.