The Tarim River, the largest inland river in China, sits in the Tarim River Basin (TRB), which is an arid area with the ecosystem primarily sustained by water from melting snow and glaciers in the headstream area. To evaluate the pressures of natural disasters in this climate-change-sensitive basin, this study projected flash droughts in the headstream area of the TRB. We used the variable infiltration capacity (VIC) model to describe the hydrological processes of the study area, Markov chain Monte Carlo to quantify the parameter uncertainty of the VIC model. Ten downscaled general circulation models from the Coupled Model Intercomparison Project Phase 6 (CMIP6) were used to drive the VIC model, and the standardized evaporative stress ratio was applied to identify flash droughts. The results demonstrated that the VIC model after Bayesian parameters uncertainty analysis can efficiently describe the hydrological processes of the study area. In the future (2021-2100), compared with the plain region, the alpine region has higher flash drought frequency and intensity. Compared with the historical period (1961-2014), the frequency, duration, and intensity of flash droughts tend to increase throughout the study area, especially for the alpine area. Moreover, based on variance decomposition, CMIP6 model is the most important uncertainty source for flash drought projection, followed by the shared socioeconomic pathway of climate change scenario and VIC model parameters.

We develop a Bayesian model to predict the maximum thickness of seasonally frozen ground (MTSFG) using historical air temperature and precipitation observations. We use the Stefan solution and meteorological data from 11 stations to estimate the MTSFG changes from 1961 to 2016 in the Yellow River source region of northwestern China. We employ an antecedent precipitation index model to estimate changes in the liquid soil water content. The marginal posterior probability distributions of the antecedent precipitation index parameters are estimated using Markov chain Monte Carlo sampling methods. We compare the results of our stochastic method with those obtained from the traditional deterministic method and find that they are consistent in general. The stochastic approach is effective for estimating the historical changes in the frozen ground depth (root-mean-square errors = 0.13-0.35 m), and it provides more information on model uncertainty regarding soil moisture variations. Additionally, simulation shows that the MTSFG has decreased by 0.31 cm per year over the last 56 years on the northeastern Qinghai-Tibet Plateau. This decrease in frost depth accelerated in the 1990s and 2000s. Considering the lack of data on seasonally frozen soil monitoring, the Bayesian method provides a pragmatic approach to statistically model frozen ground changes using available meteorological data.