Motivated by segmentation issues in marine studies, a novel hidden Markov model is proposed for the analysis of cylindrical space-time series, that is, bivariate space-time series of intensities and angles. The model is a multilevel mixture of cylindrical densities, where the parameters of the mixture vary across space according to a latent Markov field, while the parameters of this hidden Markov random field evolve over time according to the states of a hidden Markov chain. It segments the data within a finite number of latent classes that represent the conditional distributions of the data under environmental conditions that vary across space and time, simultaneously accounting for unobserved heterogeneity and space-time autocorrelation. It parsimoniously accommodates specific features of environmental cylindrical data, such as circular-linear correlation, multimodality and skewness. Due to the numerical intractability of the likelihood function, parameters are estimated by a computationally efficient EM algorithm based on the maximization of a weighted composite likelihood. The effectiveness of the proposal is tested in a case study that involves speeds and directions of marine currents in the Gulf of Naples, where the model was capable to cluster cylindrical data according to a finite number of intuitively appealing latent classes.