Answers are grounded in this paper and cite their sources.
In regional machine-learning studies, the predictive value of climatic fire indicators is rarely separated from geographical gradients, observation bias and spatial autocorrelation. We evaluated fire-season (January–April) vapour pressure deficit (VPD) as a spatial indicator of satellite-detected vegetation fire across Sichuan, Chongqing and Hubei, China, using MODIS detections (2001−2020) in burnable 1 km cells, XGBoost with spatial block cross-validation, ablation, geography-only baselines, controls for observation opportunity, and ecoregion and temporal validation. The full model reached an out-of-block AUC of 0.842, and VPD was the strongest of 13 retained predictors (by univariate AUC, SHAP and geographical detector). Removing the climate group reduced AUC by 0.076 (95% CI 0.059–0.092), against ≤0.012 for the other groups, and the highest-ranked 10% of burnable land contained 57% of fire-positive cells. The strongest support came from temporal validation: models calibrated with 2001–2010 data located 2011–2020 fires, including newly burning cells in held-out blocks (AUC 0.834 and 0.814), and climate alone outperformed coordinates alone (ΔAUC 0.022, 95% CI 0.012–0.032). In space, however, climatic information overlapped largely with location (climate added 0.003 to a model with coordinates) and partly with observation opportunity, removing VPD alone did not reduce AUC, and transfer to unsampled ecoregions was limited (AUC 0.48–0.76). The VPD transition point (∼0.49 kPa) refers to 20-year means and is not a daily warning threshold (false-alarm ratio > 99%). Fire-season VPD is thus a low-cost, temporally stable indicator for regional spatial screening, and our diagnostic sequence offers a template for testing climatic indicators in machine-learning fire studies.
No takes yet. Share an insight, caveat, or question.
Chen et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: