We study the tomography of spatial qudits encoded on structured light photons. While direct position measurements with cameras do not provide an informationally complete Positive Operator Valued Measure (POVM) in the space of fixed order modes, we complement this POVM with an astigmatic transformation. The enlarged POVM is informationally complete, allowing full characterization of the spatial quantum state from simple intensity measurements in both the intense and in the low photocount regimes. For intense light, the standard technique of linear inversion is used. For the low photocount regime, we employ Bayesian mean inference, and study how the quality of the tomographic reconstruction behaves as we increase the photocounts. In both cases, we also perform the tomography using a convolutional neural network, which displays an increased flexibility in exchange for a slightly lower quality reconstruction in some of the cases. These methods will be useful for classical and quantum communication with structured light.