The understanding of subsurface information on the Earth is crucial in numerous fields such as economics of oil and gas, geophysical exploration, archaeology and hydro-geophysics, particularly in a context of climate change. The methodology consists in reconstructing the seismic velocity model of the near surface, that contains information about the basement structure, by solving the inverse problem and resolving the related complex nonlinear systems with the data collected from seismic experiments and measurements. In the last few years, many deep neural networks have been proposed to simplify the seismic inversion problem based, for instance, on automatic differentiation of the adjoint operator, or on automatic arrival time picking. However, such approaches require a large amount of labeled training data, which are hardly available in real applications. We present here a deep learning approach for arrival time picking, aimed to deal with unlabeled data. The main building blocks are transfer learning, as well as a semi-supervised learning strategy where the pseudo-labels are greedily computed with robust regression, and classification algorithms. The hybrid method showcases very high scores when evaluating on synthetic data, and its application to a real dataset containing a limited amount of labeled data shows the computational efficiency and very accurate results.