| Wildfire prediction models typically discretize study areas into uniform grids, ignoring the heterogeneous spatial distribution of ignitions. We challenge this paradigm by showing that \emph{how} data is discretized matters more than \emph{which} model is used. We propose an unsupervised fire-zone segmentation algorithm combining watershed detection with K-means clustering to define prediction units directly from historical fire patterns. Experiments across six French departments and six forecasting models show that fire-zone segmentation consistently outperforms grid-based approaches, with mean IoU improvements of +3--6\% depending on spatial scale. The method is computationally lightweight ($<$10s per configuration) and fully parallelizable. Our results demonstrate that optimizing spatial discretization yields significant, reproducible performance gains for short-term wildfire forecasting. Code will be released on GitHub. |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.