Abstract
The importance of perception through all the senses has been recognized in previous studies on landscape preference, but data on aural perception, as opposed to the visual, remains rare. We seek to bridge this gap by analyzing texts that describe more than 3.5 million georeferenced images, created by more than 12000 volunteers in the Geograph project. Our analysis commences by extracting and automatically disambiguating descriptions that potentially contain verbs and nouns of sound (e.g. rustle, bellow, echo, noise) and adjectives of sound intensity (e.g. deafening, quiet, vociferous). Using random forests we classify more than 8000 descriptions based on the type of sound emitter into geophony (e.g. rustling wind, bubbling waterfall), biophony (e.g. gulls calling, bellowing stag), anthrophony (e.g. roaring jets, rumbling traffic) and perceived absence of sound (e.g. not a sound can be heard) with a precision of 0.81. Further, we additionally classify these descriptions as negative, neutral and positive using an Opinion Lexicon and GloVe word embeddings. Our results show that sentiment classification gives an additional level of understanding of descriptions classified into different types of sound emitters. We see that geophony, biophony and anthrophony cannot be uniquely classified as positive or negative. Our results demonstrate how text can provide a valuable, complementary to field-based studies, source of spatially-referenced information about aural landscape perception.
Chesnokova, Olga