Answers are grounded in this paper and cite their sources.
Drought-driven wildfires are intensifying even in humid subtropical forests, yet existing warning systems still rely on static, weather-only maps. Such maps capture neither how susceptibility changes within a season nor how it varies from one hillside to the next. We developed a parallel CNN-LSTM model integrating local landscape context with antecedent moisture and vegetation histories to produce weekly wildfire susceptibility maps across Taiwan. We tested spatial patches from 5 × 5 to 25 × 25 grid cells and antecedent windows from 0 to 24 weeks to identify input configurations that carried skill. On the temporally held-out 2022–2024 test set, the parallel CNN-LSTM with a 25 × 25 spatial patch and a 12-week antecedent window achieved the highest 0.82 AUC-ROC and flagged elevated susceptibility at the location of a major 2023 wildfire in southwestern Taiwan. These scores rank locations and weeks by relative wildfire susceptibility rather than representing calibrated probabilities of fire occurrence. An event-level comparison showed that the parallel model identified wildfires missed by the standalone CNN; these events exhibited persistent antecedent moisture deficits and declining vegetation greenness, consistent with cumulative drying that current-week inputs alone did not adequately capture. Factor importance and effect assessment showed that drought and vegetation stress dominated susceptibility during the fire season, whereas preconditioning factors, terrain and human access, mattered more once severe fire weather eased. By translating antecedent drought memory into weekly, spatially explicit maps, this approach offers a dynamic susceptibility layer to support Taiwan's wildfire warning system and a reference for testing in other humid subtropical regions.
No takes yet. Share an insight, caveat, or question.
Yu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: