Yuan Zhuang: Evaluating Geo-Foundation Model Embeddings for Wildfire Risk Assessment
Authors:
Yuan Zhuang (UNSW School of Risk and Actuarial Studies)
Sanaa Hobeichi (UNSW Climate Change Research Centre, ARC Centre of Excellence for the weather of the 21st Century)
Peng Shi (Wisconsin School of Business, University of Wisconsin-Madison)
Fei Huang (UNSW School of Risk and Actuarial Studies, feihuang@unsw.edu.au)
Abstract:
Recent advances in geo-foundation models have enabled the extraction of rich environmental representations from large-scale Earth observation and geospatial datasets. Trained on diverse multimodal data sources, these models have the potential to capture complex spatial and environmental relationships that may be valuable for downstream Earth system and disaster risk applications. However, their effectiveness for natural hazard assessment remains insufficiently explored, particularly in the context of wildfire risk.
This research investigates whether AlphaEarth embeddings can support wildfire assessment, using susceptibility mapping in Victoria, Australia as a case study. We compare foundation model embeddings with traditional hand-curated environmental covariates using multiple machine learning models. Preliminary results show that AlphaEarth embeddings achieve competitive predictive performance and stronger cross-model consistency, despite requiring substantially less preprocessing effort. Additional analyses suggest that the embeddings effectively capture broad environmental background conditions related to wildfire occurrence, while remaining less sensitive to short-term meteorological extremes.
The study highlights both the opportunities and limitations of geo-foundation model embeddings for Earth system applications. More broadly, it demonstrates how foundation models may lower technical barriers for climate risk mapping and support scalable downstream applications such as insurance exposure analysis and disaster risk assessment.
Keywords: Geo-foundation models; wildfire susceptibility; Earth observation; remote sensing; climate risk assessment