Differential Effects of Soil Moisture and Air Temperature on Vegetation Dynamics in Northwest China's Warming and...

Under the pronounced warming-wetting trend in Northwest China, understanding vegetation responses to the redistribution of hydrothermal resources is essential for interpreting regional ecohydrological processes. Here, we developed a bivariate Long Short-Term Memory (LSTM) model to simulate leaf area index (LAI) dynamics for four representative vegetation types (cold temperate forest, shrubland, grassland, and cropland), using air temperature and soil moisture as predictors. The model reproduces seasonal vegetation phenology well across vegetation types (R² > 0.9), indicating that LSTM effectively captures the cumulative and lagged effects of hydrothermal drivers. However, its performance diverges at the interannual scale. Interannual variability in grasslands in water-limited environments is reasonably represented (R² = 0.31), consistent with their sensitivity to short-term hydroclimatic variability under warming-wetting conditions. In contrast, the model fails to reproduce the observed long-term greening trend in forests when driven solely by hydrothermal variables. This contrast suggests distinct underlying mechanisms across ecosystem types. Grassland dynamics are closely linked to high-frequency hydroclimatic variability, whereas forest growth appears to be governed by slower processes and low-frequency drivers, including CO₂ fertilization, nitrogen deposition, and ecological inertia. As a result, hydrothermal variables alone are insufficient to explain long-term forest dynamics. Overall, these findings highlight a transition from water-limited to energy- and process-limited controls across vegetation types and underscore the limitations of purely climate-driven models. Integrating biogeochemical processes or process-based constraints into machine learning frameworks may therefore be necessary to improve predictions of long-term vegetation change under climate change.