Using social media data and machine learning to map recreational ecosystem services

Crowdsourced geotagged social media data and machine learning approaches have emerged as promising tools for mapping ecosystem services, especially cultural ecosystem services that are difficult to assess. Here, we use recreation to show how social media data, machine learning, and spatial analysis techniques can improve our understanding of human-nature interactions and the mapping of recreational ecosystem services. We extracted 80,500 photographs taken in non-urban areas of the Tahoe Central Sierra Initiative project area in California between 2005 and 2019 that were posted to the photo sharing application Flickr and used these as a proxy for recreational visits to the area. Automated image content analysis was used to identify the objects and concepts in the photographs and uncover the types of nature experiences that are important to visitors. Additionally, variable importance, a Random Forest machine learning technique, was used to examine the environmental and landscape variables that drive recreation in the area and to create a classification model that predicts the recreation potential of the entire area based on important variables. The automated image content analysis identified 1,239 unique labels linked to recreation, with mountains, hills, and rocks being the most prominent features (22%). Our Random Forest model indicates that vegetation cover, land cover, elevation, smoke days, and landscape features are major drivers of recreation in the area and are of interest to visitors in the area. The model predicted that 25.9% of the area has the potential to support recreational visits. Most of these recreation potential areas are in protected areas (77.8%), predominantly in conifer forests (66%) and within national forest boundaries, especially the Tahoe National Forest area (37.6%). These results show that recreational ecosystem services vary across landscapes and illustrate the need for improved mapping approaches to determine the provision of ecosystem services in different places. The analysis provides novel insights into the various ways social media data and machine learning techniques can be powerful components of ecosystem service research and how they hold great potential for monitoring and informing management interventions on ecosystem service provision, especially in places with limited traditional onsite visitation data