Work

Quantifying hydrothermal-driven dynamics of gross primary productivity in the Ta-Pieh mountains based on explainable artificial intelligence

Year: 2025

Type: article

Abstract: In the context of global warming and intensified climate extremes, accurately quantifying regional responses of gross primary productivity (GPP) to hydro-thermal drivers is crucial for assessing carbon balance and guiding ecosystem management. Taking the Ta-pieh Mountains as a case study, we integrated multi-source remote-sensing and meteorological data (2000–2022) and applied Theil-Sen trend analysis, Mann-Kendall tests, and least-squares cross-wavelet analysis to characterize the spatio-temporal variations of GPP and climate variables. We then implemented a per-pixel sliding-window modeling framework using four tree ensemble algorithms—Random Forest (RF), eXtreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM) and Category Boosting (CatBoost)—and employed TreeSHAP to quantify contributions of temperature, precipitation, root-zone soil moisture (RZSM), and vapor-pressure deficit (VPD) to GPP. Results reveal that temperature consistently dominates GPP variability across most years in the Ta-pieh Mountains, with VPD contributions closely tracking temperature fluctuations. Precipitation shows a one-month lagged effect on GPP. Moreover, reduced precipitation and lower RZSM impose stronger limitations on carbon uptake during drought events, exemplified by the autumn drought in 2019, than under typical conditions. The per-pixel sliding-window approach yielded high accuracy in predicting GPP and clearly delineated each driver’s influence across spatial units and climate scenarios via SHAP analysis. These methods provide a transferable toolkit for analyzing vegetation dynamics and climate risks, offering robust, science-based guidance for ecosystem management in mountainous and transition-zone regions.

Source: Journal of Hydrology Regional Studies

DOI: 10.1016/j.ejrh.2025.102797