Accurate simulation of terrestrial gross primary productivity (GPP) is crucial for understanding global carbon cycles and climate change impacts. While light use efficiency (LUE) models, particularly two-leaf (TL) approaches, outperform big-leaf models, their parameterizations for water stress and meteorological responses remain limited. To address this, we developed an improved water-constrained TL-LUE model (WTL-LUE) based on the revised TL-LUE (RTL-LUE). Using observations from 201 sites in FLUXNET2015 dataset covering ten ecosystems, we optimized WTL-LUE by determining the parameters for temperature and vapor pressure deficit constraint functions through high-quantile regression and introducing a nonlinear photosynthesis response function to light. The optimized model demonstrated significant improvements in GPP estimation, achieving R2 values of 0.71 (RMSE = 2.23 gC m-2 d-1) and 0.74 (RMSE = 2.03 gC m-2 d-1) for daily and 8-day scales, respectively. WTL-LUE outperformed existing LUE models (MOD17, VPM, TL-LUE, RTL-LUE), particularly in dryland ecosystems (savannas, shrublands) and specific vegetation types (croplands, deciduous broadleaf forests, wetlands), underscoring the critical integration of meteorological data with remote sensing for accurate water stress representation. In comparative analyses across environmental gradients, WTL-LUE also demonstrated relative stability advantages over the benchmarked XGBoost machine learning approach. This study provides a robust tool for analyzing global ecosystem dynamics and their responses to climate change.