Airborne lidar data is commonly used to generate point clouds over large areas. These points can be classified into different categories such as ground, building, vegetation, etc. The first step for this is to separate ground points from non-ground points. Existing methods rely mainly on TIN densification but there performance varies with the type of terrain and relies on the user’s experience who adjusts parameters accordingly. An alternative may be on the use of a deep learning approach that would limit user’s intervention. Hence, in this paper, we assess a deep learning architecture, PointNet, that applies directly to point clouds. Our preliminary results show mitigating classification rates and further investigation is required to properly train the system and improve the robustness, showing issues with the choices we made in the preprocessing. Nonetheless, our analysis suggests that it is necessary to enrich the architecture of the network to integrate the notion of neighbourhood at different scales in order to increase the accuracy and the robustness of the treatment as well as its capacity to treat data from different geographical contexts.