With the deployment of high quality and dense permanent seismic networks over the last 15 years comes a dramatic increase of data to process. In order to lower the threshold value of magnitudes in a catalogue as much as possible, the issue of discrimination between natural and anthropogenic events is becoming increasingly important. To achieve this discrimination, we propose the use of a convolutional neural network (CNN) trained from spectrograms. We built a database of labeled events detected in metropolitan France between 2020 and 2021 and trained a CNN with three-component 60 s spectrograms ranging frequencies from 1 to 50 Hz. By applying our trained model on independent French data, we reach an accuracy of 98.2 per cent. In order to show the versatility of the approach, this trained model is also applied on different geographical areas, a post-seismic campaign from NW France and data from Utah, and reaches an accuracy of 100.0 per cent and 96.7 per cent respectively. These tests tend to hypothesise that some features due to explosions compared to earthquakes are widely shared in different geographical places. In a first approach, we propose that it can be due to a contrast in the energy balance between natural and anthopogenic events. Earthquake seismic energies seem to be more continuous as a function of frequency (vertical bands features in a spectrogram) and conversely for explosions (horizontal strips).