Promises and pitfalls of using computer vision to make inferences about landscape preferences: Evidence from an urban-proximate park system (data and code)

We compare preferences for landscape features derived through a computer vision algorithm (Google Cloud Vision) used to analyze social media photographs with preferences derived through a traditional on-site intercept survey. We surveyed visitors in Boulder Open Space and Mountain Parks lands in Colorado (USA) in May and June, 2018. We downloaded all Flickr photographs within Boulder Open Space and Mountain Parks lands from 2004 - 2018, and ran the photographs through Google Cloud Vision to get up to 10 labels for each image. We compare the content in Flickr photographs to the features that visitors say positively impacted their experience on surveys.<br><br>This paper is currently under review.<br><br><b>Contents of this repository:</b><br><br><b>GoogleVision: </b><br>Contains raw data exported from Google Vision, as well as a codebook for how we coded each label to match the landscape categories we asked about in the survey. Also contains a full database connecting the Google Vision labels and presence/absence of each feature to the Flickr data.<br><br><b>Code:</b><br>Contains one R script that makes the maps and runs spatial cluster analysis, and one R script that does all the data cleaning and analysis for the Flickr and survey data. You will need to download the contents in the GoogleVision, shapefiles, and survey folders to run this code. This folder also contains a Python script that we used to download Flickr data within Boulder through the Flickr API, and another R script used only to generate table E.1 in the supplementary material.<br><br><b>Shapefiles: </b><br>Contains all the spatial data needed to reproduce maps and run R code. This includes OSMP trails, trailheads, lands, landscape character areas, survey locations, and coordinates of Flickr points.<br><br><b>Survey:</b><br>Contains the data from a visitor survey in Boulder OSMP lands from May and June 2018 (in a CSV), as well as a codebook to interpret the data, and the survey instrument.<br><br>