Integrating Remote Sensing and machine learning for dynamic burn probability mapping in data-limited contexts

Effective Burn probability mapping is crucial for proactive fire management and enhancing firefighting efficiency. Typically, these maps rely on static variables like topography, vegetation density, and fuel availability. Dynamic data sources such as remote sensing data offer precise, easy-access information for structuring dynamic Burn probability assessment tools. This study introduces a remote sensing-based Burn probability prediction model tailored for the State of Jalisco, Mexico, leveraging satellite data and machine learning algorithms (Logistic regression, Random Forest, XGBoost) to support public policy development. The model utilizes multispectral datasets, local geographic information, and algorithms such as logistic regression and random forest to identify high-risk wildfire areas. All evaluated parameters presented significant differences between the Fire-Affected and Non-Fire-Affected groups. Both NDVI and NDWI presented strong correlations to the presence of fire events, with smaller dispersion values for Fire-Affected entries within the dataset compared to Non-Fire-Affected entries, indicating high potential for its use as predictor of Burn probability. The model delivers a robust decision support system by integrating climatic, topographical, and anthropogenic factors. The XGBoost model incorporating nine parameters, identified as the best-performing by a recursive feature elimination analysis, presented an AUC value of 0.96 and a Sensitivity of 0.9333. Our findings highlight that this approach effectively identifies high-risk areas, aiding in targeted policy interventions and resource allocation to mitigate wildfire impacts, and offering a low-cost alternative for Burn probability monitoring in developing countries and resource-restricted areas. • Remote sensing & local data combined to build accurate burn probability model for wildfire risk assessments. • XGBoost model outperformed other tested models with an AUC value of 0.96. • Nine key variables were integrated for accurate Burn probability prediction. • Model aids targeted policies & efficient resource allocation to reduce wildfire impacts