Deep learning paradigms are widespread approaches for manifold environmental applications. Their strength consists in the possibility of processing massive quantities of data as well as performing complex tasks, with unsupervised procedures. The use of deep learners, based on neural computing, proved its effectiveness for data classification, data generation, time series prediction, etc.. In particular, recurrent neural networks are suitable for time series prediction. Here, recurrent neural networks are used to model groundwater levels of the large shallow porous aquifer of Brindisi. This is an aquifer with a catchment of approximately 1000 km2 located in the upper east Salento peninsula, in southeast Italy. It is hosted by the shallow Quaternary deposits consisting of weakly cemented sands, which diffusively outcrops in its catchment. This aquifer is recharged by local rainfalls, which do not significantly change across the catchment, in terms of volumes and time distribution. This shallow aquifer was monitored by the former National Hydrographic Bureau, whereas piezometric wells were used to acquire groundwater level measurements with a frequency of one measure every three days. However, these time series, available for 6 wells between 1952 and 2002, are not complete, since several gaps, i.e. missing data exist. These time series, together with the time series of rainfall data are used to train a recurrent neural network, which is able to predict average monthly groundwater levels as well as reconstructing missing data, whereas time series are incomplete. Differently from machine learning approaches, multiple well groundwater time series, as well as multiple rain gauge time series will be used together, in order to train the deep learner.