<p>Wildfires are increasingly dangerous in Maule (Chile), where they start in several places and in different manner in each places. Most traditional prediction models assume that space is homogeneous, making it hard identifying localized ignition patterns and adjusting to different types of land. This article presents a lightweight geospatial framework that uses supervised classification models (Random Forest, Naive Bayes, and Artificial Networks), which are trained on either spatially segmented or global data in combination with density-based clustering algorithms (DBSCAN and HDBSCAN). The method employs only geographic coordinates (latitude and longitude), to explore whether territory segmentation by itself makes binary wildfire categorization more accurate and easier to understand. The method was used to evaluate wildfire risk in the Maule region. Initial results show that geographical segmentation can help with wildfire modeling and geospatial risk assessment of areas for which there is little data.</p>