Map building and environment modelling is a main task in mobile robot navigation. Obtaining a lightweight and robust model of the environment is crucial when processes are going to be run in a robot with low computing power and memory, as in the case of most climbing robots. This article proposes the use of different neural network architectures to identify from the data captured with a LiDAR sensor those points contained in planes belonging to reticular structures. Our purpose is to remove irrelevant information such as trees or soil in order to reduce the computation and memory requirements for mapping or localization tasks. For training these neural networks, an automatic dataset generation and labelling process has been developed through simulated environments. The experiments evidence the capacity of neural networks to segment elements of the structure contained in a plane.