Updated reference cartographic datasets are one of the main input data sources in several application domains requiring spatial modelling. The aim of this work is to semi-automatically extract building footprints from very high-resolution satellite images to update already existing cartographic datasets or generate them if missing over the target area. The proposed approach for building footprint detection takes advantage of a convolutional neural network for segmenting VHR satellite images. The images from which the building footprints are extracted are different from the images used to train the network. The adopted architecture enables to generalize the learning features from images that have different statistics from the segmented image. One application domain where this feature is critical is emergency mapping since it enables to rapidly extract reference information over the area of interest. Such reference data allows possible damages to infrastructures to be assessed and delineated. The results indicate that the procedure can be adopted for purposes that range from urban sprawl monitoring to rapid mapping activities using a different approach in comparison to classic image segmentation methods. For example, in emergency mapping activities, if no reference data is available (either from authoritative or open data sources) building footprints must be manually extracted by means of visual interpretation. The proposed procedure enables a faster reference data extraction, limiting the human intervention. © The Authors.