The probability distributions of soil parameters can be updated with limited site-specific information via probabilistic back analyses. The updated probability distributions can be further used for a more realistic slope reliability assessment. However, few attempts have been made to conduct probabilistic back analyses accounting for the inherent spatial variability of soil properties. The main challenge is the so-called curse of dimensionality encountered when thousands of random variables are used to model spatial variability. The BUS approach (Bayesian Updating with Structural reliability methods) can tackle the high-dimensional back analysis problem by transforming it into an equivalent structural reliability problem, but it requires an evaluation of the likelihood multiplier that is tedious and time-consuming. This paper proposes a modified BUS approach for the probabilistic back analysis of soil parameters and reliability updating of slopes in spatially variable soils. With this approach, the curse of dimensionality and evaluation of the likelihood multiplier can be effectively avoided, and the computational accuracy is significantly improved. Two slope examples are investigated to illustrate the effectiveness of the proposed approach. The soil parameters and their probability distributions for a slope section can be well determined through a probabilistic back analysis, which facilitates an accurate identification of the causes of slope failures and a deep understanding of the actual performance of in-service slopes.