An interpretable machine learning approach based on DNN, SVR, Extra Tree, and XGBoost models for predicting daily pan...

Area of Science:

• Hydrology
• Environmental Science
• Data Science

Background:

• Evaporation is a critical hydrological process impacting water resources.
• Machine learning (ML) models offer high accuracy in predicting pan evaporation.
• The 'black-box' nature of ML hinders practical application and physical process understanding.

Purpose of the Study:

• To develop an interpretable ML framework for predicting daily pan evaporation.
• To assess the physical consistency of ML models using interpretability techniques.
• To enhance the transparency and reliability of ML models in hydrological studies.

Main Methods:

• Utilized Extra Tree, XGBoost, SVR, and Deep Neural Network (DNN) models.
• Employed Shapely Additive explanations (SHAP), Sobol sensitivity analysis, and Local Interpretable Model-agnostic Explanations (LIME) for interpretability.
• Analyzed hourly climate datasets from the Sidi Mohammed Ben Abdellah (SMBA) weather station in Morocco.

Main Results:

• ML models demonstrated high accuracy in predicting daily pan evaporation (NSE 0.76-0.83).
• Air temperature (Ta) and solar radiation (Rs) were identified as the most influential climate variables.
• Interpretability analysis revealed consistency between ML predictions and hydro-climatic processes.

Conclusions:

• The interpretable ML framework enhances the reliability and transparency of evaporation prediction models.
• The study provides insights into reducing the 'black-box' nature of ML in hydrological applications.
• The findings support the use of interpretable ML for understanding and managing water resources in semi-arid environments.