Interpretable deep learning method to quantify the impact of extreme temperatures on vegetation productivity in China

As a key ecological parameter, NPP measures the photosynthetic efficiency of plants in capturing atmospheric carbon. With the warming of the climate, extreme temperature events are frequent, which has exerted a profound influence on NPP. Previous studies on the drivers of NPP have predominantly relied on linear approaches. This study innovatively employs deep neural networks (DNN) integrated with SHAP method, integrating nationwide daily meteorological data (2001-2020) and MODIS annual NPP products, to establish the first quantitative causal linkages between extreme temperature events and NPP dynamics across China. The findings indicate that: (1) The study period witnessed a pronounced increase in both occurrence rates and severity of extreme heat events across China, contrasted by a distinct decrease in extreme cold episodes, with notable regional variations in these trends. (2) NPP exhibits a distinct southeast-to-northwest decreasing gradient across China, and from 2001 to 2020, NPP in most parts of China showed an upward trend, but NPP in some mountainous and hilly areas in the south decreased slightly. (3) Regional analyses reveal contrasting NPP responses to temperature extremes. In southern region and Tibetan Plateau region, NPP shows a positive association with elevated temperature extremes, while moderate low temperature is more conducive to vegetation growth in northern region. (4) The DNN model constructed in this study performs better (R²≈0.89) than other models in simulating the spatiotemporal distribution of NPP in China. The SHAP analysis identifies frost days and annual precipitation as the primary drivers of NPP variation across China, with most factors exhibiting threshold-dependent effects on NPP.