Zhou et al. (2026) An impact-based drought classification method using real-world agricultural drought records and explainable automated machine learning

Identification

• Journal: Journal of Hydrology
• Year: 2026
• Date: 2026-02-05
• Authors: Keke Zhou, Jianzhu Li, Ting Zhang, Xiaogang Shi, Ping Feng
• DOI: 10.1016/j.jhydrol.2026.135078

Research Groups

• State Key Laboratory of Hydraulic Engineering Intelligent Construction and Operation, Tianjin University, Tianjin 300350, China
• School of Social and Environmental Sustainability, University of Glasgow, Dumfries, UK

Short Summary

This study introduces a novel impact-based framework combining causal inference with explainable Automated Machine Learning (AutoML) to classify drought severity and identify its primary drivers in China. The framework, leveraging real-world impact records, outperforms conventional methods, revealing that non-climatic factors (latitude, geopotential height) and climatic factors (soil moisture, evaporation) are key drivers, and indicating a significant intensification of drought severity across China from 1980 to 2024.

Objective

• To develop a novel impact-based framework that synergizes causal inference with explainable Automated Machine Learning (AutoML) to classify drought severity and uncover its primary causal drivers in China.
• To enhance climate resilience, inform adaptive water management strategies, and mitigate drought risks by providing a robust assessment framework.

Study Configuration

• Spatial Scale: National scale for China, with specific regional analysis in South China, Yunnan, Henan, and Hebei.
• Temporal Scale: Trend analysis conducted for the period between 1980 and 2024. Classification models incorporate antecedent climatic information.

Methodology and Data

• Models used: Automated Machine Learning (AutoML), Shapley Additive Explanations (SHAP) for explainability, PCMCI+ (Peter and Clark Momentary Conditional Independence plus) algorithm for causal inference.
• Data sources: Observed real-world agricultural drought impact records, linked with a comprehensive set of drought-related variables.

Main Results

• The AutoML model significantly outperforms conventional machine learning approaches and widely used standardized drought indices across multiple evaluation metrics, accurately depicting both in-sample and out-of-sample drought events and capturing drought impacts on vegetation and ecosystems.
• Explainability (SHAP) and causal analysis (PCMCI+) reveal that two non-climatic variables, latitude and geopotential height, exert the strongest influence on the model’s drought classifications.
• Soil moisture and evaporation from bare soil are identified as the most influential climatic drivers of drought severity, with soil moisture exhibiting particularly strong influence across South China.
• Trend analysis indicates a significant intensification of drought severity across China between 1980 and 2024, with marked increases observed in Yunnan, Henan, and Hebei.
• Incorporating antecedent climatic information consistently improves classification performance, with AutoML models exhibiting the greatest gains.

Contributions

• Presents a novel impact-based drought classification framework that integrates causal inference and explainable Automated Machine Learning (AutoML), enhancing accuracy and practical relevance by utilizing real-world drought impact records.
• Advances the transparency and interpretability of drought assessment through the integration of SHAP and PCMCI+ methodologies.
• Demonstrates superior performance of the AutoML-based model compared to conventional machine learning approaches and standardized drought indices in depicting drought events and their impacts.
• Identifies key non-climatic (latitude, geopotential height) and climatic (soil moisture, evaporation from bare soil) drivers of drought severity in China.
• Provides quantitative evidence of a significant intensification of drought severity across China between 1980 and 2024, highlighting specific vulnerable regions.

Funding

Not specified in the provided text.

Citation

@article{Zhou2026impactbased,
  author = {Zhou, Keke and Li, Jianzhu and Zhang, Ting and Shi, Xiaogang and Feng, Ping},
  title = {An impact-based drought classification method using real-world agricultural drought records and explainable automated machine learning},
  journal = {Journal of Hydrology},
  year = {2026},
  doi = {10.1016/j.jhydrol.2026.135078},
  url = {https://doi.org/10.1016/j.jhydrol.2026.135078}
}