Answers are grounded in this paper and cite their sources.
Net ecosystem exchange (NEE), a critical indicator of terrestrial carbon source‐sink dynamics, remains difficult to estimate across global flux sites due to spatial heterogeneity and sparse flux observations. Existing models overlook local variability, limiting accuracy in site‐level NEE estimation. Geographically Weighted Regression (GWR) captures spatial non‐stationarity in environmental variables on NEE by enabling spatially varying model parameters. This study develops a Geographically Weighted eXtreme Gradient Boosting (GWXGBoost) framework coupling XGBoost's strength in modeling complex nonlinear relationships with GWR's spatial explicitness. Using 8‐day data from 387 global flux sites and remote sensing predictor variables, we optimize model parameters locally per site to account for spatial heterogeneity in NEE drivers. Our results show water‐related variables dominate explanatory power in GWXGBoost at 36% of sites, particularly in arid regions. The GWXGBoost framework outperforms traditional global models, achieving R 2 = 0.63 and RMSE = 1.21 g C m −2 day −1 at the 8‐day scale—with ∼40% higher R 2 and ∼20% lower RMSE than global modeling. Model accuracy is enhanced at the site scale ( R 2 = 0.84) and remains robust across most plant functional types, with the highest value in deciduous broadleaf forests ( R 2 = 0.80) and the lowest in grasslands ( R 2 = 0.37). Independent validation across 140 flux sites ( R 2 = 0.61) demonstrated the model's robustness and multi‐scale generalizability. Moreover, local modeling produces NEE estimates more consistent with observations and better captures site‐specific NEE controls than global models. Our study demonstrates that modeling spatial heterogeneity with localized models improves NEE prediction and supports carbon budget assessments.
No takes yet. Share an insight, caveat, or question.
Shi et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: