Landslide with stepwise deformation often occurs under water level fluctuations in reservoir areas. Its displacement curve is multi-step

with intermittent

sudden

(quasi) periodic characteristics. This study selects a typical reservoir landslide case with the stepwise feature. Based on years of on-site monitoring

the feature of stepwise deformation is revealed. A theory of stick-slip is introduced to reasonably explain the mechanism of stepwise deformation. Secondly

considering the limitations of existing methods in the early warning for this landslide type

a new method of threshold definition and warning is developed based on machine learning. The results show that the seepage effect controls the process of stick-slip of the sliding zone soils. It is the likely mechanism of stepwise deformation of the landslide. Moreover

based on artificial neural network (ANN) and C5.0 algorithms

a two-dimensional threshold scheme for water level factor is defined

and the inverse proportional relationship between water level and rate of water level change is founded. The accuracy rate of early warning reaches 86.7%. This case study provides references for the introduction of the concept of stick-slip into landslide mechanism studies and machine learning techniques into threshold definition.