An illustration of model agnostic explainability methods applied to environmental data

Area of Science:

• Environmental Science
• Computer Science
• Statistics

Background:

• Machine learning and deep neural models historically lack uncertainty quantification and inference capabilities.
• Explainable AI (XAI) has emerged to address concerns regarding transparency, fairness, and model interpretability.
• Understanding input importance is crucial for reliable environmental data prediction models.

Purpose of the Study:

• To explain which input features are most important in models predicting environmental data.
• To apply and illustrate model-agnostic explainability methods for environmental forecasting.
• To enhance the interpretability of machine learning models used in environmental science.

Main Methods:

• Focus on three general, model-agnostic explainability methods: feature shuffling, interpretable local surrogates, and occlusion analysis.
• Implement and demonstrate specific applications of these XAI techniques.
• Utilize a variety of machine learning models for analysis.

Main Results:

• Successfully applied feature shuffling, interpretable local surrogates, and occlusion analysis to identify important predictors.
• Demonstrated the utility of these methods in the context of long-lead forecasting of soil moisture.
• Quantified the influence of specific inputs, such as Pacific sea surface temperature anomalies, on model predictions.

Conclusions:

• Model-agnostic explainability methods provide valuable insights into the drivers of environmental predictions.
• Feature shuffling, interpretable local surrogates, and occlusion analysis are effective tools for understanding model behavior.
• These XAI techniques improve the trustworthiness and applicability of machine learning in environmental forecasting, particularly for soil moisture.