Wildfire forecasts move to a 12-hour clock: FireGraphNet's directed graph lifts near-front ranking 1.7–1.9 times a distance rule

The scored object changes from which cells burn on the next day to which unburned cells ahead of the front newly burn in roughly the next 12 hours, shortening the prediction unit from a day to half a day. Most learned models predict daily on targets that include burned area, and the closest graph node-classification work is daily, uses undirected edges without features, and scores cells drawn from the fire's final perimeter. On five California test fires used for no model selection, the five FireGraphNet variants reach average precision 0.0624–0.0710 against 0.0369 for the strongest distance baseline, with no-skill prevalence at 0.0165.

The predictive skill comes from passing neighborhood information along edges rather than from each cell's own features. A multilayer perceptron on cell features without message passing scores 0.0394, at the level of the ring-distance baseline at 0.0369. The edge-blind GraphSAGE variant has the lowest average precision of the five and the worst Brier score and calibration error, while the edge-conditioned GINE variant has a Brier score of 0.035 against 0.050–0.094 for the others.

The same graph model keeps ranking skill outside its training region, while the no-message-passing control has no skill out of state. The out-of-state set was declared evaluation-only at freeze time and run once under pre-registration, requiring pooled lift above 1 and an established-track stratum also above 1. Across five Oregon, Washington and Colorado fires the pooled lift is 1.62 and the established-track stratum is 1.73, with per-fire lifts from 1.12 on East Troublesome to 3.22 on Santiam Complex, while the no-message-passing control sits at 1.06.

Trained on the same candidate objective, a raster convolutional network becomes statistically indistinguishable from the graph model on pooled ranking, but the two separate on operating behavior. The same Res18-U-Net trained on a full-mask target that includes the burned footprint scores 0.0230, below the distance baseline of 0.0369. Under the candidate objective that network reaches 0.0769 against 0.0710 for GINE, but its Brier score and calibration error are 6.9 and 3.6 times worse, it loses 31% of attainable CSI when the validation-frozen threshold is transferred against 2% for GINE, and it carries about 2.4 times the seed-to-seed standard deviation.

The claims cover 2018–2021 California fires plus the 2020–2021 out-of-state confirmation set, and are limited to retrospective roughly 12-hour prediction within five H3 rings of an active front; that region holds 88.8% of newly burned cells in the California test set and 71.9% out of state. The intended readers are researchers and engineers working on wildfire spread modeling and evaluation, and every model input is public, with no forecast product required.

The authors state that the probabilities remain uncalibrated and require calibration on separate data and evaluation under deployment-specific thresholds and costs; the five-ring candidate horizon omits some rapid runs, satellite labels carry detection uncertainty and variable observation intervals, and the nominal FEDS timestamps sit about 8.5 hours from observation times. Fuel and barrier edits measure counterfactual sensitivities of a learned model, and on the fuel-dominated CZU fire random barrier edits even increase predicted spread, so they do not establish management effectiveness.