By combining detrended fluctuation analysis (DFA) method with surrogate data method, and using the Heuristic segmentation algorithm as well as Chi-Square statistics, we develop a new method to determine the threshold of extreme events, e.g. stochastically re-sorting detrended fluctuation analysis (S-DFA) method. The S-DFA method has a certain phsical background, when the occurrence rate of the data is small, then these data belong to little-probability events and they contain so little information about the dynamic system, the states corresponding to these data are abnormal states or extreme states of the system. When the occurrence rate of the data is large or even in distribution these data do not belong to little-probability events and they contain much information about the system, the states corresponding to these data are normal states of the system. Compared with the Percentile curves method, the S-DFA method gives the critical value between extreme event and non-extreame event, which is definite and unique. We also extensively validate the effectiveness of S-DFA method through extreme event detection.