Towards enhanced fire fuel estimation with satellite-derived predictive models

This report presents a technical note on developing predictive models that link satellite observations, such as vegetation optical depth and solar-induced fluorescence, to key fuel variables, specifically fuel load and fuel moisture content, which are crucial for wildfire forecasting.

The extreme gradient boosting model, a scalable and efficient decision tree-based machine learning algorithm, was implemented as a non-linear observation operator to integrate satellite-derived observations into an offline land data assimilation framework. Data preparation ensured temporal alignment with satellite observations and involved preprocessing steps to improve physical consistency. Feature importance analysis quantified global predictor contributions, while Shapley additive explanations analysis offered detailed insights into predictor impact and directionality at a granular level. Both analyses confirmed the model’s reliance on physically meaningful relationships, consistent with known vegetation-water-energy dynamics.

The model performed well in capturing spatial and seasonal patterns, particularly in regions with clear phenological cycles, such as crops and savannas. Some challenges persist in areas with dense canopies or sparse vegetation, where signal saturation or soil-vegetation decoupling can reduce prediction accuracy. Filtering for orography, snow cover, and steep slopes is included as a preprocessing step in the land data assimilation framework.

The resulting predictive models, along with a sample of data, are provided as open-source code through this GitHub repository https://github.com/selgarroussi/fuelity.