Geo-referenced data which are often communicated via maps are inaccessible to the visually impaired population. We summarise existing approaches to improving accessibility of geo-referenced data and present the Atlas.txt project which aims to produce textual summaries of such data which can be read out via a screenreader. We outline issues involved in generating descriptions of geo-referenced data and present initial work on content determination based on knowledge acquisition from both parallel corpus analysis and input from visually impaired people. In our corpus analysis we build an ontology containing abstract representations of expert-written sentences which we associate with macros containing sequences of data analysis methods. This helps us to identify which data analysis methods need to be applied to generate text from data.