Abstract An appropriate probabilistic description of snow loads is important in reliability analysis. Probability distributions for annual maximum ground snow loads are studied in detail in this paper based on the samples of snow loads on the ground, which are simulated by a snowmelt model. Then, selection of probability distribution function and parameter estimation method for ground snow loads is found to have a great influence on reliability assessment results. The reliability of members of steel roof structures subjected to snow load at four representative sites in China is investigated. The reliability indices of studied roofs are found not to be enough to reach the target value 3.2. The possible reason is that characteristic value of snow loads on the ground specified in Chinese load code is not conservative and the partial factor 1.4 for snow loads on roofs is not enough. To ensure designed roofs subjected to extreme snow loads a target reliability index 3.2, large partial factors for snow load that varies with COV of ground snow loads are expected to be used.