Point processes are often used as spatial or spatio-temporal distribution models of occurrences. In this PhD dissertation, we focus first on Cox processes driven by a hidden process associated with a Dirichlet process. This model corresponds to hidden occurrences influencing the stochastic intensity of observed occurrences. We generalize the notion of "Shot noise Cox process" introduced by M{\o}ller and develop its bayesian analysis by a Gibbs sampler combined with a Metropolis-Hastings algorithm. We show that our MCMC method is a reversible jump one. The model takes into account a random number of hidden contributions producing effects on the observed point process intensity. Therefore the parameter space has a variable dimension. We focus the statistical inference on the estimation of the hidden contribution expected value, the hidden contribution expected number, the spatial influence and correlation parameters. The contribution equality test and contribution independence test are proposed. Applications in epidemiology and ecology are shown from \textit{Rubus fruticosa} data, \textit{Ibicella lutea} data and death number data in counties of Georgia, USA. Two situations are considered with respect to available data~: firstly, the spatial positions of occurrences are observed between several pairs of consecutive dates; secondly, counts are carried out over a fixed time interval in several spatial sampling units. Secondly, we focus on point processes with memory introduced by Kagan, Ogata and Vere-Jones, pioneers in statistical seismology. In fact, spatio-temporal point processes play an important role in the studies of earthquake catalogs since they consist of seismic events with their dates and spatial locations. We studied an ETAS (Epidemic Type Aftershock Sequence) model with time independent background intensity and several triggering functions taking into account previous events. We illustrate our approach with a seismicity study of the Lesser Antilles arc. A comparaison study of Gamma, Weibull, Log-Normal and modified Omori law triggering function models is also carried out. We show that the modified Omori law does not fit the Lesser Antilles seismic data and the best adjusted triggering function is the Weibull model. Consequently, the waiting time between aftershocks is weaker in the Lesser Antilles zone compared to the one in regions with seismicity described by the modified Omori law. In other words, aftershock aggregativity is higher in the Lesser Antilles region. The stochastic background intensity following a Dirichlet process centered on a spatial log-normal process is discussed.
