Solar granulation is the visible signature of convective cells emerging from the upper convective zone towards the solar surface. High-resolution images have revealed the complexity of the granulation, evidencing special phenomena such as exploding granules or granular lanes, which are known to be directly related to the emergence of small-scale magnetic flux. Unveiling the nature of magnetic emergence in granules requires extensive statistical studies. The development of new automatic tools has become crucial to perform statistics on a large amount of data expected by new/upcoming instrumentation: DKIST or Sunrise III.

In this contribution, we present the current advances of our classification algorithm of solar granulation based on neural networks including the exploration of recurrent modules for covering the temporal dimension. An initial model was tested using U-net architecture in a supervised approach using the continuum intensity of the IMaX instrument onboard the Sunrise I and their corresponding segmented maps. We study the performance of this approach to assess the versatility of the U-Net architecture for single-frame segmentation. We found an interesting potential of the U-Net to identify granules reaching matching in pixels greater than 80%, achieving high levels of accuracy in the identification of the intergranular network and allowing the effective separation of granular morphologies. We identify per-class accuracy levels of around 60% in single snapshots which are substantially reduced when temporal sequences are included as extra channels. Recurrent modules added within the deep layers seem to improve the prediction accuracy compared with the previous case.