Zarei (2026) Deciphering regional hydro-climatic drivers across diverse climatic zones using a causal explainable ensemble with interaction-aware stable framework

Identification

• Journal: Journal of Hydrology Regional Studies
• Year: 2026
• Date: 2026-09-24
• Authors: Abdol Rassoul Zarei
• DOI: 10.1016/j.ejrh.2026.104011

Research Groups

Department of Range and Watershed Management (Nature engineering), College of Agriculture, Fasa University, Fasa, Iran

Short Summary

This study develops the Dynamic Causal Explainable Ensemble with Interaction-aware Stable Index (DCEE-ISI) framework to causally attribute hydro-climatic drivers of actual evapotranspiration (AET) across 40 diverse synoptic stations in Iran. The research reveals that conventional correlational models significantly overestimate precipitation's role, while the proposed causal framework identifies maximum temperature and wind speed as the dominant forcing mechanisms in 26 and 8 strategic stations, respectively, aligning with established hydro-climatic theories.

Objective

• To develop and apply the Dynamic Causal Explainable Ensemble with Interaction-aware Stable Index (DCEE-ISI) framework to causally attribute hydro-climatic drivers of actual evapotranspiration (AET) across diverse climatic zones, overcoming limitations of conventional correlational methods by disentangling mechanistic influences from statistical associations, accounting for non-linear interactions, and temporal non-stationarity.

Study Configuration

• Spatial Scale: 40 diverse synoptic stations across Iran, encompassing hyper-arid central deserts, humid coastal plains, and alpine environments.
• Temporal Scale: Multi-decadal period from 1967 to 2024 (58 years) using monthly datasets.

Methodology and Data

• Models used:
  • Dynamic Causal Explainable Ensemble with Interaction-aware Stable Index (DCEE-ISI) framework.
  • Extreme Gradient Boosting (XGBoost) for predictive modeling.
  • Shapley Additive Explanations (SHAP) for diagnostic interpretability and decoding non-linear interactions.
  • Generalized Causal Forests for extracting mechanistic causal structures (Average Treatment Effect - ATE).
  • Temporal Stability Regularization (TSR) to account for non-stationarity.
  • Shannon Entropy for adaptive weighting of framework components.
  • Benchmark models: Multiple Linear Regression (MLR), Random Forest (RF), Permutation Importance.
• Data sources:
  • Monthly climatic records (1967–2024) for 40 synoptic stations from the Iran Meteorological Organization (IRIMO).
  • Climatic drivers: mean maximum temperature (Tm), mean minimum temperature (Tmi), mean maximum relative humidity (RHm), mean minimum relative humidity (RHmi), mean wind speed at 2 meters (W), mean sunshine duration (S), and total precipitation (P).
  • High-resolution (4 km) gridded Actual Evapotranspiration (AET) data from the TerraClimate dataset (Abatzoglou et al., 2018) as the response variable.

Main Results

• A systematic divergence was found between correlational (SHAP-based) and causally-informed attribution, with conventional models consistently overestimating precipitation's role. For example, at Ahvaz station, precipitation's SHAP value was 15.38, while its causal impact (Ci) was 0.392.
• The DCEE-ISI framework identified maximum temperature as a dominant forcing mechanism in 26 stations and wind speed in 8 strategic stations, aligning with Budyko framework and sensible heat advection theories.
• Non-linear synergistic effects were quantified, showing substantial interaction effects (Ii > 5.0) for maximum temperature and wind speed in several stations (e.g., Hamadan, Kermanshah, Qazvin).
• Temporal Stability Regularization revealed high volatility in traditional feature importance, with precipitation exhibiting severe temporal instability (Ti≈0.000) in multiple stations.
• Shannon Entropy-based adaptive weighting dynamically adjusted component contributions, assigning paramount weights to causality (wC) in regions with severe confounding structures (e.g., 0.434 at Karaj, 0.429 at Qazvin).
• At the national scale, precipitation was a dominant variable in 21 stations, maximum temperature was dominant in 13 stations (and strong in 13 others), and wind speed was strong or dominant in 8 strategic stations.
• Minimum and maximum relative humidity were consistently categorized as weak drivers across most stations, indicating their role as reactive atmospheric states rather than proactive drivers of AET.

Contributions

• Development of an integrated attribution framework (DCEE-ISI) combining machine learning, explainability, and causal inference.
• Introduction of a composite index (I-SCII) that jointly accounts for predictability, causality, interaction strength, and temporal stability.
• Explicit incorporation of temporal non-stationarity through a stability regularization scheme.
• Application and validation across a long-term, multi-climatic observational dataset in Iran.
• Emphasis on consistency across spatial units and station-level distributional analysis to rigorously assess attribution patterns beyond traditional mean-based summaries.

Funding

No funds, grants, or other support was received.

Citation

@article{Zarei2026Deciphering,
  author = {Zarei, Abdol Rassoul},
  title = {Deciphering regional hydro-climatic drivers across diverse climatic zones using a causal explainable ensemble with interaction-aware stable framework},
  journal = {Journal of Hydrology Regional Studies},
  year = {2026},
  doi = {10.1016/j.ejrh.2026.104011},
  url = {https://doi.org/10.1016/j.ejrh.2026.104011}
}