Explainable artificial intelligence (XAI) for interpreting the contributing factors feed into the wildfire...

Area of Science:

• Environmental Science
• Artificial Intelligence
• Geospatial Analysis

Background:

• Wildfires pose a significant environmental threat in Australia, necessitating accurate prediction of fire occurrence and susceptibility.
• Machine learning (ML) models are valuable for analyzing complex wildfire hazards but can be limited by data quality and irrelevant input variables.
• Explainable AI (XAI) offers a method to understand and refine ML models, addressing uncertainties in wildfire susceptibility prediction.

Purpose of the Study:

• To interpret the results of a deep learning (DL) model for wildfire susceptibility prediction using Explainable AI (XAI).
• To identify the key environmental and meteorological factors influencing wildfire susceptibility predictions.
• To enhance the reliability and transparency of wildfire risk assessment models.

Main Methods:

• Development of a deep learning (DL) model for wildfire susceptibility prediction.
• Application of Shapley additive explanations (SHAP) to interpret the DL model's predictions.
• Analysis of topographical, landcover/vegetation, and meteorological data, including Normalized Difference Moisture Index (NDMI).

Main Results:

• SHAP plots identified significant contributing factors to wildfire susceptibility, including humidity, wind speed, rainfall, elevation, slope, and NDMI.
• The study demonstrated the relative importance of various parameters in the prediction model.
• Insights were gained into the reasoning behind the DL model's susceptibility predictions.

Conclusions:

• Explainable AI (XAI), specifically SHAP, enhances the understanding of wildfire susceptibility models.
• Identifying high-contributing factors allows for more effective wildfire hazard control and management strategies.
• Developing interpretable models is crucial for improving the accuracy and trustworthiness of wildfire risk mapping.