Data-Driven Fire-Zone Segmentation Improves Wildfire Prediction

Traditional wildfire prediction models typically divide geographical areas into uniform grids, which often overlooks the non-uniform distribution of fire ignitions. This new research proposes that the method of data discretization is more impactful than the specific forecasting model used. The study introduces an unsupervised fire-zone segmentation algorithm. This algorithm combines watershed detection with K-means clustering to create prediction units that are directly derived from historical wildfire patterns, rather than arbitrary grids. This data-driven approach allows the model to better capture the heterogeneous spatial characteristics of fire occurrences. Experiments conducted across six French departments using six different forecasting models consistently showed that the fire-zone segmentation method delivered superior performance compared to grid-based approaches. It achieved mean IoU (Intersection over Union) improvements of 3-6%, depending on the spatial scale. The method is also computationally efficient, demonstrating that optimizing spatial discretization can lead to significant and reproducible gains in short-term wildfire forecasting accuracy.