The advent of new sources of spatial data and associated information (e.g. Volunteered Geographic Information (VGI)) allows us to explore non-expert conceptualisations of space, where the number of participants and spatial extent coverage encompassed can be much greater than is available through traditional empirical approaches. In this paper we explore such data through the prism of landscape preference or scenicness. VGI in the form of photographs is particularly suited to this task, and the volume of images has been suggested as a simple proxy for landscape preference. We propose another approach, which models landscape aesthetics based on the descriptions of some 220000 images collected in a large VGI project in the UK, and more than 1.5 million votes related to the perceived scenicness of these images collected in a crowdsourcing project. We use image descriptions to build features for a supervised machine learning algorithm. Features include the most frequent uni- and bigrams, adjectives, presence of verbs of perception and adjectives from the "Landscape Adjective Checklist". Our results include not only qualitative information relating terms to scenicness in the UK, but a model based on our features which can predict some 52% of the variation in scenicness, comparable to typical models using more traditional approaches. The most useful features are the 800 most frequent unigrams, presence of adjectives from the "Landscape Adjective Checklist" and a spatial weighting term.