The spatial and temporal distribution of forest fires displays a complex pattern which strongly influences the forest landscape and the neighbouring anthropogenic development. Statistical methods developed for spatio-temporal stochastic point processes can be employed to find a structure, detect over-densities and trends in forest fire risk and address towards prevention and forecasting measures. The present study considers the Portuguese mapped burnt areas official geodatabase resulting from interpreted satellite measurements, covering the period 1990-2013. The main goal is to detect whether space and time act independently or whether, conversely, neighbouring events are also closer in time, interacting to generate clusters. To this purpose, the following statistical methods were applied: (1) the geographically weighted summary statistics, to explore how the average burned area vary locally through the investigated region; (2) the bivariate K-function, to test the space-time interaction and the spatial attraction/independency between fires of different size; and (3) the space-time kernel density, allowing elaborating smoothed density surfaces and representing over-densities of large versus medium versus small fires and on north versus south region. The proposed approach successfully allowed finding and mapping spatio-temporal patterns within this large data series. Specifically, medium fires tend to aggregate around small fires, while large fires aggregate at a larger distance and longer times, indicating that the return time following these events is longer than for small and medium fires. The density maps shows that hot spots are present almost each year in the northern region, with a higher concentration in the northern areas, while the southern half of the country counts lower surface densities of fires, which are mainly concentrated in the central period (2000-2007).