Performance evaluation of ST-DBSCAN and spatio-temporal spatially constrained K-means (ST-SCKM) for wildfire risk zoning and resilience analysis

Wildfire risk assessment requires methodologies capable of capturing both spatial contiguity and temporal variability. This study proposes a Spatio-Temporal Spatially Constrained K-Means (ST-SCKM) framework for delineating cohesive wildfire risk zones, and compares its performance with the density-based ST-DBSCAN algorithm and a spatially constrained Ward’s approach. Using 131,049 active fire detections from the MODIS FIRMS archive combined with ERA5 meteorological variables for California (2019–2024), spatial coordinates were projected into a planar UTM system and Fire Radiative Power (FRP) was log-transformed. To ensure operational interpretability, the clustering objective function strictly utilizes spatial and temporal components, while environmental features are utilized post hoc for risk profiling. By incorporating a dynamic K-Nearest Neighbors (KNN) soft-penalty constraint, ST-SCKM ( w s = 0.5 , w t = 1.5 ) reduces spatial fragmentation and large-cluster effects commonly observed in density-based methods. Under the evaluated parameter settings, the proposed approach achieves higher clustering performance than the tuned baselines, with a Silhouette Score of 0.4728 (vs. 0.3492) and a Calinski-Harabasz index of 7113.98 (vs. 108.17), while retaining all observations without classifying events as noise. The resulting macro-zones are subsequently characterized into Low, Moderate, High, and Extreme risk categories. Temporal analysis indicates persistent spatio-temporal patterns, particularly for Extreme-risk zones during the 2020–2021 drought period. Spatial comparison with historical wildfire footprints demonstrates an overwhelming 99.81% quantitative overlap (Intersection over Union proxy) between high-risk zones and major fire-affected areas. Furthermore, a forward validation test successfully predicted severe hotspot expansions for the unseen 2023–2024 period, confirming its predictive utility in unseen data. Overall, the ST-SCKM framework provides a data-driven approach for transforming spatio-temporal point data into contiguous risk zones, with potential applications in supporting wildfire risk assessment and resilience-oriented decision-making.