Up to the present, the snow load for design of buildings in Japan has been based on the snow depth because of the lack of data of the snow load itself. Concerning the snow depth, series of the annual maximum snow depth (AMD) have been considered as one of the most important factors. Most of researchers have investigated the statistical properties of the snow depth. However, most of the research has discussed what probability distribution function is appropriate to explain these properties and none has discussed the behaviour of AMD on a probability paper, which depends on observation points. In this paper, not only the behaviour of AMD on the probability paper but also the properties of the population of AMD are investigated in order to propose a reasonable method to calculate more appropriate return period values of AMD. Five types of probability distribution functions, which are types I, II and III extreme value distributions, normal distribution and lognormal distribution, are applied with Thomas plot and Hazen plot techniques to AMD of 423 points in Japan. Although the behaviour of AMD of these points can not be explained by any sole distribution or sole plotting techniques, the proposed method is available to almost all points. In the method, the linear regression analysis is applied only to the 1/3 larger data of AMD on the type I extreme value distribution probability paper. Moreover, this method may be more reasonable than the others to calculate the values of AMD for a long return period through considering snowfall mechanism.