Climate–driven attribution of the water level reversal in Lake Qinghai: Non–linear thresholds revealed by interpretable machine learning

The Lake Qinghai Basin is located on the northeastern region of the Tibetan Plateau (TP). As China’s largest inland saltwater lake, Lake Qinghai serves as a critical sentinel for alpine regional water security. Lake Qinghai water level experienced a dramatic structural regime shift around 2004, characterized by an asymmetric “V–shaped” reversal from continuous recession to rapid expansion, where the post–2005 rate of rise being three times greater than the preceding decline. This study utilizes 65 years of multi–source data (1960–2024) and an interpretable machine learning framework coupling eXtreme Gradient Boosting (XGBoost) and SHapley Additive exPlanations (SHAP) to investigate the non–linear physical mechanisms and quantify the threshold effects governing this hydrological recovery. Focusing explicitly on natural climatic drivers, the research isolates these factors from potential anthropogenic influences to determine the primary causes of this structural reversal. This research confirms the structural “V–shaped” reversal of the lake’s hydrology and identifies precipitation (57.2 %) and sunshine duration (20.5 %) as the dominant drivers within the analyzed climatic framework, while the net temperature effect remains weak due to the counterbalancing of meltwater recharge and evaporation. Crucially, the study quantifies non–linear tipping points governing this recovery: the regime shift in runoff generation triggers non–linear amplification when annual precipitation exceeds 513.84 mm, and a “dimming effect” significantly suppresses evaporation when sunshine duration falls 2814.64 h. Additionally, the contribution of minimum temperature (8.3 %) significantly outweighs that of maximum temperature, implicating cryospheric degradation and baseflow regulation as key auxiliary mechanisms. Consequently, these findings elucidate that the lake’s rapid recovery is not a simple linear response to climate change but a threshold–driven transition, providing essential quantitative insights for adaptive water resource management in alpine regions. • Structural “V–shaped” reversal of lake hydrology occurred around 2004. • XGBoost–SHAP quantified non-linear thresholds of driving factors. • Non–linear runoff amplification is triggered above a 513.84 mm precipitation threshold.