Fire-season vapour pressure deficit as a spatial indicator of satellite-detected vegetation fire in subtropical China: predictive value, geographical confounding and transferability

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.

Authors

Institutions

Publication Details

Journal
Ecological Indicators
Published
2026-11-18
DOI
https://doi.org/10.1016/j.ecolind.2026.115872
Primary Topic
Fire effects on ecosystems
Type
article
Field-Weighted Citation Impact
0.00

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Fire-season vapour pressure deficit as a spatial indicator of satellite-detected vegetation fire in subtropical China: predictive value, geographical confounding and transferability

Pujie Wei, Byron B. Lamont, Lin Chen, Weixing Xue et al.
3 links
Ecological Indicators
Fire effects on ecosystems
article

Fire-season vapour pressure deficit as a spatial indicator of satellite-detected vegetation fire in subtropical China: predictive value, geographical confounding and transferability

Pujie Wei, Byron B. Lamont, Lin Chen, Weixing Xue, Zhaogui Yan, Yu Liu
article en

Abstract

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.

Ecological IndicatorsVol. 193
University of Vienna (AT), Huazhong Agricultural University (CN), Curtin University (AU), Research Institute of Tropical Forestry (CN), Chinese Academy of Forestry (CN), University of Göttingen (DE)
Fundamental Research Funds for the Central Universities
Life on land, Climate action
Fire effects on ecosystems
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon