Explainable machine learning for the prediction and assessment of complex drought impacts

Area of Science:

• Environmental Science
• Data Science
• Disaster Management

Background:

• Droughts cause significant societal, economic, and environmental damage.
• Machine learning (ML) models are powerful predictive tools but often lack transparency, hindering trust in critical applications like disaster assessment.
• Explainability in ML is vital for stakeholders to understand and trust model predictions, especially in high-stakes scenarios.

Purpose of the Study:

• To develop an explainable ML pipeline for predicting multi-dimensional drought impacts in the U.S.
• To enhance the trustworthiness of ML models in drought impact assessment through explainability.
• To interpret the relationships between drought indicators and their impacts at regional scales.

Main Methods:

• Utilized an explainable ML pipeline integrating the XGBoost model and SHAP (SHapley Additive exPlanations) values.
• Trained models on a comprehensive database of drought impacts from the U.S. Drought Impact Reporter.
• Evaluated model performance using the F2 score, comparing against baseline models.

Main Results:

• The XGBoost models significantly outperformed baseline models, achieving an average F2 score of 0.883 nationally and 0.942 at the state level.
• SHAP analysis identified Standardized Precipitation Index (SPI) and Standardized Temperature Index (STI) as key predictors of drought impacts.
• Interpretable relationships were revealed: negative SPI values were positively associated with complex drought impacts, enhancing model trustworthiness.

Conclusions:

• Explainable ML, specifically using XGBoost and SHAP, offers a robust method for accurately predicting complex drought impacts.
• The study demonstrates the importance of SPI and STI in predicting drought impacts, with their influence varying by location and impact type.
• This approach improves stakeholder trust and provides actionable insights for regional drought management and response strategies.