Khosravi et al. (2026) A geographically weighted XGBoost framework for Pixel-Level modeling of vegetation responses using Multi-Source Earth Observation data

Identification

• Journal: ISPRS Journal of Photogrammetry and Remote Sensing
• Year: 2026
• Date: 2026-03-10
• Authors: Younes Khosravi, Taha B.M.J. Ouarda
• DOI: 10.1016/j.isprsjprs.2026.03.006

Research Groups

Canada Research Chair in Statistical Hydro-Climatology, Institut National de la Recherche Scientifique, Centre Eau Terre Environnement, INRS-´ET´E, Qu´ebec, QC, Canada.

Short Summary

This study introduces Geographically Weighted XGBoost (GW-XGBoost), a hybrid and interpretable framework, to model pixel-level vegetation responses to climate extremes in the Middle East. The model, calibrated with 30 years of multi-source Earth Observation data, outperforms baseline models and reveals a significant ecological transition where vegetation sensitivity has shifted from cold/precipitation constraints to warm temperatures and episodic moisture pulses.

Objective

• To develop a hybrid geospatial learning framework (GW-XGBoost) that effectively accounts for both spatial heterogeneity and nonlinear responses of vegetation to extreme climate events using multi-source Earth Observation data.
• To enhance Normalized Difference Vegetation Index (NDVI) predictive accuracy and interpretability by integrating local coefficients from geographically weighted regression (GWR) into an Extreme Gradient Boosting (XGBoost) ensemble, with SHapley Additive exPlanations (SHAP) for attribution.

Study Configuration

• Spatial Scale: Middle East (over 7 million square kilometers), with a common spatial resolution of 5 km × 5 km.
• Temporal Scale: 1995–2024 (30 years), analyzed seasonally (Winter, Spring, Summer, Autumn) and across three decadal windows (1995–2004, 2005–2014, 2015–2024) for SHAP analysis.

Methodology and Data

• Models used: Geographically Weighted XGBoost (GW-XGBoost), Ordinary Least Squares (OLS), Geographically Weighted Regression (GWR), Extreme Gradient Boosting (XGBoost), XGBoost + coordinates. SHapley Additive exPlanations (SHAP) for model interpretability.
• Data sources:
  • Vegetation: Normalized Difference Vegetation Index (NDVI) time series derived from Landsat 5 Thematic Mapper (TM), Landsat 7 Enhanced Thematic Mapper Plus (ETM+), and Landsat 8 Operational Land Imager (OLI) data (Google Earth Engine).
  • Climate: Daily temperature and precipitation from ERA5 reanalysis (European Centre for Medium-Range Weather Forecasts - ECMWF), used to compute ETCCDI extreme climate indices.
  • Topography: SRTM15+ Digital Elevation Model (DEM) (NASA Shuttle Radar Topography Mission), used to derive mean elevation, slope, and aspect.

Main Results

• The GW-XGBoost model consistently achieved the highest seasonal performance, with R² values ranging from 0.799 (autumn) to 0.854 (spring), and the lowest RMSE (0.017–0.02) and MAE (0.011–0.013). It improved explained variance by 8–15% over benchmark models.
• SHAP analysis revealed a significant region-wide ecological transition in vegetation sensitivity:
  • Winter and early-spring greenness, previously constrained by cold extremes and multi-day rainfall, is now increasingly controlled by warm-percentile temperature thresholds (e.g., TX90p, TN90p).
  • Summer and autumn vegetation has become more dependent on short-lived moisture pulses (RX5D) under intensifying heat.
• Spatially, compound-sensitive zones (e.g., "Heat + Cold + Precipitation") expanded markedly in spring and summer, particularly along the Zagros and Taurus ranges, coastal corridors, and hyper-arid interiors.
• Comparatively resilient areas persisted along the Persian Gulf margins, the southern Arabian Peninsula, and high-altitude belts.
• A decade-scale temporal holdout (training: 1995–2014; testing: 2015–2024) showed GW-XGBoost improved withheld-decade test performance in spring, summer, and autumn, remaining comparable to XGBoost in winter.

Contributions

• Development of GW-XGBoost, a novel hybrid and interpretable framework that integrates GWR's spatial adaptivity with XGBoost's nonlinear learning capabilities, enhanced by SHAP for transparent attribution of climate drivers.
• Provision of high-resolution sensitivity layers for vegetation responses to extreme climate events, leveraging multi-source Earth Observation data for the Middle East.
• Identification of a profound spatiotemporal ecological transition in the Middle East, detailing how vegetation sensitivity to climate extremes has evolved over three decades (1995-2024).
• Pinpointing emerging hotspots of compound climate stress and areas of comparative resilience, offering practical decision-support for drought early-warning, water-resource planning, and ecosystem adaptation in a climate-vulnerable region.
• Formulation of a coherent set of "rules" (Winter, Spring, Summer, Autumn Rules) that generalize the observed shifts in NDVI-climate dynamics, reflecting a transition from single-stressor to compound-stressor control.

Funding

National Sciences and Engineering Research Council of Canada (NSERC) (funding number: RGPIN-2024–06736).

Citation

@article{Khosravi2026geographically,
  author = {Khosravi, Younes and Ouarda, Taha B.M.J.},
  title = {A geographically weighted XGBoost framework for Pixel-Level modeling of vegetation responses using Multi-Source Earth Observation data},
  journal = {ISPRS Journal of Photogrammetry and Remote Sensing},
  year = {2026},
  doi = {10.1016/j.isprsjprs.2026.03.006},
  url = {https://doi.org/10.1016/j.isprsjprs.2026.03.006}
}