Comparative wildfire susceptibility modelling in heterogeneous terrains

As wildfire frequency and intensity escalate globally due to climate change, the development of robust, scalable predictive models becomes critical for effective disaster risk reduction. This research evaluates the adaptability of the Spatio-Temporal Google Earth Engine (STGEE) framework, originally designed for soil erosion modelling, to generate Wildfire Susceptibility Indices (WSI) across morphologically contrasting environments. The study focuses on two distinct sample areas: the rugged, mountainous terrain of Los Guájares, Spain, and the flat, homogeneous landscape of Hortobágy National Park, Hungary.

The methodology employs a Machine Learning (ML) approach within the cloud-computing environment of Google Earth Engine (GEE). A key innovation of this study is the adaptive selection of mapping units based on geomorphological characteristics. For the mountainous Spanish region, Slope Units (SUs) bounded by drainage and divide lines are utilized to capture topographic effects such as wind patterns and fire acceleration. Conversely, a pixel-based approach (30m * 30m) is applied to the Hungarian plain to address the relative topographic homogeneity.

The modelling process integrates a dual-component database. The inventory dataset comprises historical fire extents derived from Landsat and Sentinel-2 (MSI) products, paired with randomly sampled pseudo-absences. These are correlated with a suite of multi-source environmental conditioning factors, including topographic metrics (elevation, slope, aspect, TWI), vegetation and fuel proxies (NDVI, EVI), hydrological status (MNDWI), climatic variables (LST, precipitation, wind speed), and anthropogenic drivers (distance to roads and settlements).

Predictive modelling is performed using the Random Forest (RF) ensemble algorithm, selected for its capacity to handle non-linear interactions and multi-collinearity. To ensure model robustness and mitigate spatial autocorrelation, performance is validated using Spatial K-fold Cross-Validation. Model accuracy is assessed via the Area Under the Receiver Operating Characteristic Curve (AUROC), while Variable Importance Measurement (VIM) based on Gini Impurity is used to identify dominant fire drivers.

Preliminary hypotheses suggest that susceptibility in Los Guájares is primarily driven by topographic factors, specifically slope and aspect, whereas the Hortobágy model is expected to show higher sensitivity to vegetation moisture content and anthropogenic proximity. By successfully applying a unified methodology to heterogeneous terrains, this research aims to demonstrate the versatility of the STGEE framework in supporting targeted fire prevention strategies across diverse landscape types.

---

This research was funded by the Vicerrectorado de Investigación (University of Granada) with the Plan Propio PP2022.PP-12 on the “Caracterización de propiedades clave en la relación agua-suelo para el estudio de la influencia del fuego en el balance hídrico y el carbono para el planteamiento de estrategias de restauración”. Also, it is based on work funded by COST Action (grant no. FIRElinks CA18135), supported by COST (European Cooperation in Science and Technology).