Architecture for classifying facets on a planet as one of four surface types. Several periodic convolutions recognize patterns in the curves similar to the rotation axis retrieval network (Fig. 12). To maintain a high level of resolution of rotation phases, no down-sampling is used in the convolutions. Four dense layers with 500 to 1000 nodes estimate the probability of each surface type for all visible facets. We applied five spherical convolutions in series and add the outputs to achieve the final result. We used four filters in the spherical convolutions to match the dimensions.