Modeling Uncertainties in Spatiotemporal Dynamic of Vegetation Above- and Belowground Production on the Tibetan Plateau

Net primary production (NPP), comprising both aboveground (ANPP) and belowground (BNPP) components, serves as a critical indicator for evaluating terrestrial ecosystem responses to climate change. However, poorly characterized BNPP spatiotemporal dynamics limit carbon allocation understanding on the Tibetan Plateau (TP), and the impact of modeling uncertainties (on different models and their parameter settings) is not well constrained. Thus, we developed 97 data-driven models (including linear, machine learning, and deep learning), integrating field observations, and remote sensing datasets to produce 1-km-resolution ANPP and BNPP across the TP (1981–2018). Our results demonstrated strong model performance for both ANPP ( R 2 = 0.80 to 0.88) and BNPP ( R 2 = 0.89 to 0.95), with the Rborist (ANPP) and xgbLinear (BNPP) models showing optimal performance. Spatially, linear and machine learning models exhibited strong consistency in ANPP and BNPP estimates, while deep learning outputs showed marked divergence, particularly in eastern TP. Temporally, we observed a significant decline in total ANPP (–0.003 Pg C·a –1 ) but an increase in BNPP (0.001 to 0.003 Pg C·a –1 ), with intermodel variability in annual totals ranging from 0.13 to 0.32 Pg C·a –1 . The discrepancies stemmed from varying factor influences across models, with evapotranspiration, Bio3 (isothermality), and leaf area index primarily driving ANPP, while BNPP was mainly governed by Bio3, Bio14 (driest-month precipitation), and leaf area index. Notably, the increasing contribution of BNPP to carbon sequestration on the TP highlights the necessity of its independent quantification in global carbon models. This study also reveals uncertainties in vegetation carbon allocation and provides key benchmarks for developing biogeochemical models in climate-sensitive regions.