GPR systems with a single central frequency suffer limitations due to the unavoidable trade-off between resolution and penetration depth that multi frequency equipments can overcome. We propose a new semi-supervised multi-frequency merging algorithm based on Deep Learning and specifically on Bi-Directional Long-Short Term Memory to automatically merge varying numbers of data sets at different frequencies. The proposed methodology is tested on synthetic and field data, to evaluate performances and robustness. The proposed merging algorithm can manage the complementarity of information at different central frequencies, properly merging different types of data. Results show not only a smooth transition in time, but, even more important, a remarkable broadening of the bandwidth thus increasing the overall resolution. Our approach is not limited to specific frequency components or geological setting but can be potentially exploited to merge any type of dataset with different spectral components.