Valuing cultural ecosystem services combining deep learning and benefit transfer approach

Abstract

The importance placed by policy-makers on cultural ecosystem services (CES) is on the rise worldwide, inducing a growing interest in the development of innovative approaches for their mapping, characterization, and assessment. Previous studies have shown how the analysis of the content of geotagged images shared on social media can be used to shed a light on CES provision and consumption without having to resort to time-consuming surveys and interviews. However, to make this approach viable over large study areas, the image classification process must be automatized and new developments of deep learning field have the potential to do so. To evaluate this potential, we have gathered over 91,000 images acquired within British Columbia (Canada) provincial parks system and then trained and deployed convolutional neural networks (CNN) to classify them. Images were classified based on: (i) their relevance in the context of CES research; (ii) the type of CES depicted (aesthetic or recreation); and (iii) the type of aesthetic experience or recreational activity depicted; with an estimated accuracy of 85 %. Then we combined the outcomes of the image classification process with images metadata to assess the value of the recreational ecosystem service provided by the most popular parks with an innovative crowdsourced benefit transfer approach, obtaining values that range from the 105.2 M CAN$ of Cypress park to the 0.3 M CAN$ of Lac du Bois park. Our findings show how crowdsourced social media images, coupled with the new developments of deep learning, offer to ecosystem managers across the world the opportunity of gathering cheap, fine scale, and comprehensive information about CES. Finally, we address the limitations and future perspectives of automated image classification for CES research.