Abstract: As the dam monitoring effect quantity is a random variable, the peaks over threshold (POT) model based on the extreme value theory can be adopted to study the monitoring index of the monitoring effect quantity. However, the threshold in the existing POT model is usually determined by graphic methods, which require human judgment and lead to strong subjectivity and randomness, and automatic computer identification is difficult to realize. In this study, a threshold increment sequence is constructed and the corresponding monitoring indexes under different thresholds are calculated. An improved threshold determination method is then developed based on the "3σ criterion" in probability theory, in which the most reasonable threshold is selected while the difference between the dangerous value and the warning value of the monitoring index approaches to the standard deviation of the measured value sequence. Verification example shows that the improved method is of a clear theoretical foundation and effectively reduces the subjectivity and random errors of the graphical methods. By using computer programs, automatic recognition of the most reasonable threshold can be realized, which enhances the practicability of the POT model in determination of the dam safety monitoring index.