This paper explores how the dynamics and structure of land ownership can be studied by automating the vectorisation of historical cadastral sources. Our analysis focuses mainly on three cadastres of Lausanne (1722, 1831, 1883). The plans are first georeferenced, before being semantically segmented using a neural model, and vectorised. We study the dynamics of persistence using a spatial matching methodology, detecting parcel fusions and divisions through time. We also investigate the social structure of ownership by matching representative owners to their socioprofessional status, based on historical population censuses.