County Galway, Ireland, was selected as the study area because of its heterogeneous landscape, socio-cultural heritage and high visitor numbers. In 2018, visitor numbers were estimated at 1 million domestic and 1.8 million international, contributing a total of €800 million in revenue (Cunningham et al. 2015, Galway County Council 2021). Located on the west coast of Ireland (53°19' N, -9°00' W) (Fig. 1), the county covers 6,150 km² with a population of 175,000 in 2011 (CSO 2011a). The scope of this study focused on spatial trends over a continuous, semi-natural landscape. Therefore, the area of interest was restricted to exclude Galway City as an urban hotspot, Lough Corrib as an inaccessible expanse of freshwater and islands. The remaining area (5,850 km²) contains a diverse landscape of natural features including mountainous areas, grasslands, wetlands, forested areas and coastline. The region also contains areas of biological interest such as Connemara National Park, 19 Special Protection Areas (SPAs) and 77 Special Areas of Conservation (SACs) (NPWS 2022).

Figure 1.

Case-study area of interest, County Galway and its location on the west coast of Ireland. Areas excluded from analysis (Galway City and Lough Corrib) are highlighted.

Data were collected from the Flickr social media platform using the statistical programming language R v.4.1.2 (R Core Team 2021), the Flickr API and the R package photosearcher v.1.0 (Fox et al. 2020). Flickr was selected because of its sizeable userbase estimated to have uploaded 5.67 billion photos between 2004 and 2016 (Ding and Fan 2019), previous work that suggested the platform hosts a more diverse userbase compared with other platforms (Oteros-Rozas et al. 2018) and its use in similar studies on the topic (Wood et al. 2013, Tenerelli et al. 2016, Wood et al. 2020). Flickr has the benefit of complementary data access policies that permit the collection and use of data for academic research purposes (Fox et al. 2020). Photosearcher calls on the Flickr API to retrieve data based on user-provided parameters. All searches were conducted in Feburary 2022 and parameterised to retrieve photo records with GPS coordinates taken between 1 January 2010 and 31 December 2019. This time range was selected to ensure a sufficiently large dataset for modelling purposes, to reflect the widespread use and accuracy of GPS devices and to exclude the disruption of travel restrictions imposed due to COVID-19 legislation in 2020. Two datasets were collected: a validation dataset and a modelling dataset. Firstly, popular locations, based on official national tourism statistics with recorded visitor numbers from Fáilte Ireland (2019a), were used to validate the relationship between social media records and visitation. Validation sites were selected, based on two criteria: 1) sites with visitor number esimates by the national tourism body for at least 4 years between 2010 and 2019 (Fáilte Ireland 2019a) and 2) solely indoor sites (e.g. concert venues, indoor museums) identified using authors' expert knowledge were removed as they were suspected to provide limited potential for in-situ ecosystem service supply. This process produced 38 sites (Suppl. material 1). The second dataset was collected to model visitation across the area of interest. A vector file defining Galway was used to retrieve photo records and the output was cleaned to contain only photo ID, user ID, date taken and GPS coordinates.

The photo-user-days (PUD) metric developed by Wood et al. (2013) has been used as an indicator for visitation, based on geotagged social media data in a number of studies (Sonter et al. 2016, Lee et al. 2019, Wood et al. 2020). The PUD metric is defined as the number of users who upload at least one photograph in a day, at a given area or location and is designed to prevent the inflation of photo-counts based on extremely active users. For the validation dataset, PUD values were calculated at the site level to match official visitor statistics and provide an appropriate comparison to validate PUD as a proxy for visitor numbers (Fáilte Ireland 2019a). The relationship between visitor numbers and PUD counts was checked using Pearson's correlation statistic and the suitability of this test checked using the Shapiro-Wilk test for normality (Crawley 2015). In the modelling dataset, PUD values were calculated per 2 km grid square as users may choose to visit multiple places in a single day and it was desirable to capture these multiple visits. GPS points were assigned a 200 m buffer, based on a conservative estimate of technological accuracy and previous work that found photos uploaded in Western Europe had a mean inaccuracy of 100 m (Zielstra and Hochmair 2013). GPS points were overlayed with a 2 km² grid and assigned a grid ID based on spatial overlap. A variable representing the combination of user ID, grid ID and date was constructed. The final PUD dataset was created by randomly slicing this variable to give one data point per unique user ID, grid ID and date combination. This sampling technique prevents the inflation of data based on individuals contributing many photos from a single visit to one grid cell in one day, but allows users to contribute to multiple cells per day.

Environmental variables were selected, based on natural, human and built capital attributes identified as factors contributing to cultural service flows (particularly recreation) in the UN System of Environmental-Economic Accounting framework (United Nations 2021) and previous studies using geotagged social media records (Tenerelli et al. 2016, Byczek et al. 2018, Tieskens et al. 2018, Chang and Olafsson 2022). Natural capital attributes included biophysical variables describing ecosystems at a given location linked to potential ecosystem service supply, built capital attributes included infrastructure associated with the accessibility and attraction of a given place and human capital was included using population density as a proxy for service demand (Table 1). While this does not capture the complexity of all factors linked to cultural service flows, this schema was designed to cover the variety of capital stocks identified in literature, given the available data for the study area. Spatial data related to these variables were collected from existing databases and clipped to the area of interest. The raw spatial data were saved and a second set of maps were created by calculating an indicator value per 2 km grid cell using zonal statistics tools in ArcMap v.10.7.

Four modelling exercises were applied in this study: 1) global regression (using both presence and count data), 2) local GWR (presence data), 3) MaxENT and 4) InVEST (Table 2).

All 38 validation sites returned PUD counts that were plotted against official visitor numbers reported by Fáilte Ireland (Fig. 2) (Fáilte Ireland 2019a). The Pearson correlation coefficient was 0.7 indicating a positive correlation and a significant relationship was observed using OLS regression (p < 0.001 at the 0.05 level, R² = 0.3). This result supports the use of social media-derived PUD as a visitation proxy in Ireland.

Figure 2.

a) Location of 38 tourist sites, b) scatterplot showing significant positive correlation between official visitor counts and photo-user-days visitation metric (log scale).

A total of 25,170 geotagged photographic records were retrieved (Suppl. material 3) and converted into 6,762 PUDs (Fig. 3a-b). The number of PUD per grid cell ranged from 0 to 413 with a mean of 5.65 and standard deviation of 20.8 (Fig. 3c). Spatial analysis using the Getis-Ord Gi* tool in ArcMap v.10.7 showed significant clustering of PUD, and therefore, a non-random distribution (Fig. 3d). Hot spots were found in western, coastal areas and the area south-east of Galway City. Cold spots of low PUD count were concentrated in the east of the region.

Figure 3.

a) GPS points of 25,170 photographs retrieved, b) GPS points of 6,762 photo-user-days (PUD), c) aggregation of PUD points into 2 km grid resolution (log scale), d) Getis-Ord Gi* statistic hot and cold spots.

The global logistic model displayed evidence of spatial autocorrelation that may mask local relationships. GWR was used to investigate this and the resulting coefficients were mapped to show variation in magnitude, direction and significance (Fig. 6). Distance to the coast was found to produce collinearity problems and dropped from this model. Monte Carlo simulation produced a significiant result (p = 0) for all predictors, further supporting the presence of spatially-varying relationships. Several variables showed changes in significance levels across the area of interest, most notably town distance that showed a negative coefficient in eastern areas and a positive coefficient in the west. In other words, areas close to towns in the east of the region are more likely to have a PUD record, but remote areas far from towns are more likely to record a PUD in the west. The presence of recreational sites was found to be significant in three hotspot areas across the landscape. The ROC for the GWR local model was plotted with an AUC of 0.909 (Fig. 4). Compared to the non-spatial global logistic model, the GWR local model showed an improved model fit as indicated by its AUC value (0.909 compared to 0.839) and AIC value (1469 compared to 1629).

Figure 6.

GWR coefficient surfaces for 12 environmental predictor variables used to model photo-user-day occurrence. Only significant coefficients at the 95% level are shaded. Warm colours indicate a positive coefficient, while cold colours indicate a negative coefficient.

The MaxEnt model used 100 m rasters of environmental predictors and PUD occurrence coordinates to predict areas of visitation suitability (Fig. 7). The mean AUC value was 0.8 indicating an improved discrimination compared to a random prediction model. The results show hot spots of high suitability for PUD visits in clusters in the east of the county, along road networks and coastal areas and surrounding Galway City. Dark blue areas indicate low likelihood of suitability for PUD occurrence and span the predominantly agricultural areas in the east and wetlands with poor connectivity and low population. Variable response curves are included in Suppl. material 5.

Figure 7.

Average suitability for Flickr-derived photo-user-day occurrence (100 replicates) at 100 m resolution generated by the maximum entropy model.

MaxEnt’s jackknife analysis, based on average AUC values, is shown in Fig. 8. Distance to the coast, elevation, presence of major roads and presence of recreational sites contributed the most information to the model in isolation (shown in dark blue bars). Similarly, the model performance showed the greatest decrease with the removal of elevation and major roads (shown by the light blue bars) suggesting that these variables contain information that is not captured by other variables. The variables with the lowest contribution to the model include geological heritage sites, river length and slope that suggests they have limited predictive power in isolation.

Figure 8.

Jackknife tests for environmental predictor contributions to MaxEnt prediction models. Dark blue shows models with the variable in isolation, light blue shows the full model minus the variable and red shows maximal model for comparison.

The output of the InVEST recreation model reports an OLS regression of log transformed annual PUD values (retrieved from an archived dataset) on environmental predictors (Table 4). Direct statistical comparison to other models is not appropriate given differing input data; however, we provide some observations regarding outputs and computation. The InVEST model uses log annual PUD as the response variable that is characterised by low values and produces correspondingly small absolute coefficient values. This model output included land-cover variables manually dropped from previous models due to collinearity concerns. The InVEST tool does not compute or report collinearity check. Other differences compared to previous analyses include river length and geological heritage being insignificant. All land-cover variables are reported with a negative sign suggesting that any cell dominated by one land-cover type has fewer PUD counts. This is consistent with a positive and significant coefficient for habitat diversity.

This is the first study to use a social media-derived PUD indicator in the Irish context and the positive correlation between PUD and official visitor counts is consistent with other validation studies (Sonter et al. 2016, Fisher et al. 2019, Kim et al. 2019). Application of the PUD metric successfully revealed global trends, local relationships and predictive suitability maps. These results demonstrate the applicability of social media-derived analysis for providing spatially explicit information for ecosystem service assessments. This information is useful for environmental management by providing a mechanism for decision-makers to account for these services and corresponding benefits in policy and planning. For example, this information could contribute to designing future tourism plans and local development strategies given the projected increase in visitor numbers to the area and understanding how and where those people interact with nature (Galway County Council 2021). In the long-term, spatially explicit assessment of the contributions ecosystems make to peoples' lives can support the preservation of these contributions for future generations (Díaz et al. 2018, Dasgupta 2021).

A core set of environmental attributes representing a blend of natural, human and built capital were identified as correlated with PUD across all models: coastal distance, presence of major roads, population density, habitat diversity, elevation and presence of recreational sites. The finding that a blend of capital stocks was correlated with PUD mirrors results in other contexts (Tenerelli et al. 2016, Byczek et al. 2018, Tieskens et al. 2018), and demonstrates that natural and biophysical characteristics, socio-cultural context and accessibility are all implicated in the potential flow of cultural ecosystem services (Byczek et al. 2018). This is aligned with the natural capital approach that considers the underpinning stocks that give rise to ecosystem services in terms of their unique combination, configuration and condition at a given place (Chan et al. 2012a, Jones et al. 2016, Costanza et al. 2017, Mace 2019). Another shared result between all models used was that protected status was not correlated with the PUD indicator, unlike in other studies, for example, USA (Figueroa-Alfaro and Tang 2017) and Japan (Yoshimura and Hiura 2017), which may suggest that protected areas serve a different role in people-environment relationships compared to other contexts. The majority of protected areas in Co. Galway fall under SAC and SPA designations that are targeted at nature conservation under EU-wide nature directives (NPWS 2022). This may be a reason why the protected status does not appear to be correlated with in-person visitation in Co. Galway, as the primary purpose for their designation lies in biodiversity conservation rather than providing a societal utility. It may be that characteristics that allow for sites of high biodiversity value prevent high visitor numbers, for example, inaccessibility and lack of man-made infrastructure or sites unsuitable for such development.

The global, non-spatial logistic model was found to display spatial autocorrelation (violating the assumption of independence) and a higher AIC value compared to both spatial model alternatives (SAM and GWR). Therefore, models that account for the spatial nature of geo-tagged social media data should be used in such cases. Spatially varying local relationships were found using GWR. This is consistent with studies that found evidence of local relationships when modelling cultural ecosystem service flows (Tenerelli et al. 2016, Schirpke et al. 2018, Kim et al. 2019). In the most extreme case of town distance, the relationship was reversed across the study area with remote areas correlated with visitation in the west and places close to towns correlated with visitation in the east. By definition, this phenomenon was obscured by models that produce one single relationship for the entire region: in both the global logistic regression and the SAM town, distance was not significant, while the InVEST model suggested a positive relationship overall. Previous European studies identified natural areas close to urban sites as potential hotspots for providing cultural ecosystem services (Ridding et al. 2018, Long et al. 2020) and, in Galway, this appears to be the case for some areas, but not everywhere. The rugged, mountainous and wetland landscapes in western areas may be perceived as more attractive to visit because of their remoteness and attract visitors away from urban areas, whereas the predominantly agricultural landscapes surrounding towns in the east may not. In less pronounced examples of local-level relationships, some variables were found to be significant, but only in limited areas rather than across the entire region, for example, habitat diversity, recreational sites and geological heritage. The identification of local relationships using GWR shows that caution should be applied when downscaling global average trends (non-spatial GLM, SAM, InVEST) to local areas (Fotheringham 2020, Comber et al. 2022).

The MaxEnt model was the only method used for prediction due to its presence-only approach (as opposed to testing for correlations in GWR and SAM models). Jackknife analysis showed that elevation, coastal distance, recreational sites and town distance were the variables of greatest predictive influence in the model. Other variables were found to have limited contributions to model prediction, such as river length, water cover and geological diversity, despite showing significant correlations in regression analysis. This result can support the prioritisation of data collection when designing management interventions as some variables appear to be more informative. Results in this case-study suggest rural areas, close to the coast, of moderate elevation and with a major road should be prioritised for targeted management interventions. These areas have the potential to experience high visitor volumes through in-situ cultural ecosystem service supply and the associated anthropogenic disturbance could contribute to negative ecological consequences and compromise long-term service flows.

The InVEST model presented some differences compared to the other models, such as the inclusion of land-cover variables and different significance levels for water cover and geological diversity. This is not unexpected given that InVEST is premised on a different response variable dataset (archived Flickr database), but it does provide a comparator to triangulate with other methods. Overall, the variables of greatest significance in user-led regression techniques (coast distance, elevation, recreational sites, major roads) were also identified as significant using InVEST with less intensive processing of data required. Stepwise model optimisation and diagnostics are not provided by the default InVEST tool and so any changes must be led by the user manually inspecting model outputs, making desired changes and re-running the model in its entirety, which can be time-consuming. These characteristics may be limitations to the InVEST model depending on the research context and similar remarks have been stated in literature (Byczek et al. 2018).

This is the first social media-based spatial modelling study in the Irish context and so comparison is limited. Previous research used a stated-preference methodology to elicit aesthetic preferences of rural landscapes in Ireland, based on a nationally-representative survey of 430 individuals (O'Donoghue et al. 2020). The results showed some overlap between highly valued aesthetic characteristics and characteristics correlated with the social media-derived PUD variable, for example, freshwater (lakes, ponds, rivers), marine areas and beaches, heritage monuments and geological features (mountains and cliffs). On the other hand, built attributes (roads, fences, buildings) were assigned a low value in the stated preference study, but urban areas and roads were found to be correlated with the PUD indicator. These differences may be due to the different phenomena investigated (visitation versus aesthetics), the different cohorts sampled, differences between revealed behaviour and stated preferences and the reality that not all areas in the landscape may be equally accessible and, therefore, visitor occurrence may not reflect a true “choice” amongst all possible options. This emphasises the value of assembling a diverse suite of tools to investigate cultural ecosystem services and the contributions they make to people’s lives (Zhang et al. 2020).

The results demonstrate the applicability of social media to developing spatial models of visitation. The differences in model outputs identified have the potential to create differing interpretations when a single approach is deployed in isolation. Therefore, we find merit in investigating multiple modelling tools—even if, ultimately, one is favoured for final reporting. While this study presented four workflows for clarity and comprehension, there is flexibility to customise these approaches to user needs beyond what is presented here. Some observations regarding data availability, data processing, expertise, research question and spatial extent required are presented below to provide information for future cultural ecosystem service assessment using these methodologies.

The focus of this study was the application of social media-derived data in different spatial models, not social media-derived data itself. The caveats of social media as a data source have been well-described in literature, but its use continues to increase, especially in studies related to cultural ecosystem services (Cheng et al. 2019, Ghermandi and Sinclair 2019, Zhang et al. 2020). We make some general comments here to contextualise the use of social media data and, despite these caveats, the results presented remain relevant as the first such application in Ireland and they provide transferable insight regarding spatial model selection for other contexts.

Potential inaccuracy in the geo-tagged coordinates was accounted for by using a 200 m buffer when calculating the PUD variable, given the suspected technological accuracy in northern Europe (Zielstra and Hochmair 2013, Tieskens et al. 2018). MaxENT requires precise GPS points and InVEST assumes GPS-tags are accurate. If there are suspected errors in GPS accuracy, a user can choose to manually validate GPS tags by inspecting photographic content (although this is costly for large samples) (Walden-Schreiner et al. 2018) or select a resolution coarse enough to render the suspected error marginal.

Social media user-groups are a self-selecting, unrepresentative sample of the general population and so data collected carries with it the bias of the userbase (Ghermandi and Sinclair 2019, Sinclair et al. 2020). Typically, this bias skews towards wealthier individuals, younger people and contains a gender-bias depending on the platform (Zhang et al. 2020), although some studies suggest Flickr is more diverse than other leading platforms (Fox et al. 2020). We acknowledge that the results of this study represent a self-selecting group of Flickr contributors and do not intend to generalise the overall population interacting with the landscape of County Galway. While skewed towards a specific userbase, the results of this pilot study provide a first-look into social media data applications in Ireland and future work can complement these findings by deploying a range of social media platforms and other data sources (visitor statistics, choice-based experiments, participatory data collection) (Wood et al. 2020).

The social media content used in this study was not screened or filtered beyond the removal of areas outside the area of interest. The volume of data collected (25,000 data points contributed by 1,866 users) and the widespread distribution of those points across the landscape suggest that the sampled data contain a diversity of information and were successfully applied as per other similar studies exploring visitation (Sonter et al. 2016, Sinclair et al. 2020). We emphasise that this study considered in-situ cultural ecosystem services broadly without disaggregating specific services or assuming user intentions and, therefore, analysis is based on the location of the photographer, not the content captured in the photograph itself (which would require inspection) (Tenerelli et al. 2016, Langemeyer et al. 2018). There is also an assumption that the physical presence of an individual in an ecosytem represents a potential cultural ecosystem service flow. Investigators may choose to filter content to varying degrees depending on their available resources, data needs and research questions (e.g. based on sentiment, content, contributor or machine-learning) (Langemeyer et al. 2018, Oteros-Rozas et al. 2018, Lee et al. 2019, Fox et al. 2021). Even with these efforts, it is not possible to know with certainty how the intent and values behind the location content was created without the input of the individual contributor. We also do not suggest that the findings of this study reveal true “preferences” as this implies a degree of choice amongst all options that was not tested in this study and there is an inherent bias towards more “picturesque” places as an artefact of using a photo-sharing platform (Clemente et al. 2019). All social media-based studies must grapple with these problems and, despite these caveats, the popularity and breadth of social media applications continues to grow (Ghermandi and Sinclair 2019, Zhang et al. 2020).