Exploring the potential of Flickr User–Generated Content for Tourism Research: Insights from Portugal

Flickr is a popular photo-sharing platform recognised as a source of data for research (Thomee et al., 2016) where users can share and organise their photos. It was developed by Ludicorp, a Vancouver, Canada-based company founded in 2004 by Stewart Butterfield and Caterina Fake and is one of the most popular sites worldwide, ranking top 400 globally, top 290 in USA, and top 2 in its category – photography (Broz, 2022). This social network allows users to add geographical references (longitude and latitude) to each uploaded photo. Users can also associate other data such as photo description, hashtags, location, the capture and publication date, among others. Developers, using the Flickr API interface, access this data for further analysis.

In February 2017, Flickr hosted approximately 13 billion photos from 122 million users that come from 72 countries (USA, 31.03%, UK, 9.83% and Germany, 5.26%), getting up to 60 million visits per month. By 2022, around 25 million photos were uploaded to Flickr every day (Broz 2022). Even if not all social groups or travellers use Flickr, Kádár and Gede (2021) highlight that it is a valid research database for measuring tourism demand, delivering comparable data on visitor flows from different parts of the world. For operational purposes, in this article, a user corresponds to a person who is registered on the Flickr social network to host and share photos.

The Flickr social network has stored a huge quantity of georeferenced photographs that have been used by many tourism scholars to study the spatiotemporal behaviour of tourists (Giglio et al., 2020; Höpken et al., 2020; Jing et al., 2020), particularly tourists’ decision-making choices (Bettaieb & Wakabayashi, 2021), destination consumption (Solazzo et al., 2022; Spyrou et al., 2017) or the development of predictive factors (Domènech et al., 2020). The classification and quantification of visitors are also interesting topics explored by several scholars (Kádár, 2014; Wood et al., 2013). Using 16 European historic tourist cities, Kádár (2014) concludes that the correlation coefficient between registered tourist bed nights and both user numbers and number of photographs are very high. Wood et al. (2013) used the locations of photographs in Flickr from 836 recreational sites around the world and found that crowd-sourced information is a reliable proxy for empirical visitation rates.

Some researches on itineraries and attraction recommendations (Lim et al., 2018; Sarkar & Majumder, 2021; Sun et al., 2019) and tourism attractions/destinations rankings (Al-Sultany, 2018; Zhou et al., 2015) have used geotagged photos from Flickr. In addition, it is important to highlight that currently one of the most researched topics is sustainability: not only in studies focused on national parks, coastal areas, and other protected areas, but also on climate change (Alieva et al., 2022; Arkema et al., 2021; Barros et al., 2019, 2020; Sottini et al., 2021).

According to Derdouri and Osaragi (2021), researchers are using more photo-based and geotagged social data in tourism research. These user-generated contents (UGC) can be very useful to better understand tourist behaviour at destinations, especially the popular hotspots and coldspots, tourist mobility patterns, and more. But as these authors note, “distinguishing between tourists and locals from this data is problematic since residence information is often not provided” (Derdouri & Osaragi, 2021, p.575).

As mentioned before, scholars of tourism have used different approaches for distinguishing tourists from locals (Derdouri & Osaragi, 2021): the heuristic methods, supervised machine learning (ML) algorithms, and Shannon entropy (SHEN).

Several authors have used the heuristic approaches, choosing a threshold related to the minimum length of stay to identify visitors (Table 1). A time span of 21 days between the first and last photo taken and at least two POIs visited in the same city identifies the photographer as a tourist, was the range selected by De Choudhury et al. (2010), and a time span of 30 days between the first and last photo taken, was the range chosen to identify tourists by Kádár & Gede (2022), Önder (2017), and Girardin et al. (2007, 2008). Kádár and Gede (2013) considered a threshold of 5 days – i.e., if photos were taken in a period shorter than 5 days, the user is classified as a visitor. These authors state that in an urban context, tourists stay in a destination for a shorter period of time. In the case of Budapest, Kádár and Gede (2013) mention that typical visitors stay on average 3 days, in line with Balińska (2020), who mentions that trips to cities are usually short-term (max 3 nights).

Straumann et al. (2014) employed a semi-automated methodology to classify the user location attribute of Flickr user profiles, extracting the countries of residence. At a global scale, Wood et al. (2013) did the same, considering that the user’s current location was the origin of each trip. For Derdouri and Osaragi (2021), ML-based methods neglect factors related to the distinction of visitors and locals, including weather, population density, and the content of users’ posts. With the SHEN method, the same authors mention that the dispersal of values of a given variable is calculated. Tourists and locals can be identified by temporal entropy: that is, tourists remain for a brief time in destinations, taking photos for a limited time, and by spatial entropy: that is, tourists and locals have different travel trajectories with different spatial distributions of images. Using these parameters, they can only differentiate two groups.

After presenting the main methods of distinguishing tourists from residents, we can conclude that there is no consensus and optimal method. Time-based approaches present weaknesses such as the different thresholds chosen by researchers or the inclusion of residents that do not share photos with frequency in the group of visitors (Straumann et al., 2014). These authors also highlight that time-based approaches rely on large amounts of data which need a long processing time.

This research aims to present and discuss the methods of identifying visitors and residents in studies using geotagged photos available in Flickr social network.

This research uses geotagged Flickr photos and their associated user data, from the period between 1 January 2010 and 31 December 2022 in the area of mainland Portugal. For this 12-year period, a total of 1,434,268 photos were identified on Flickr, uploaded by 31,286 users. The “flickr. photos.search” and “flickr.photos.getInfo” methods of the Flickr API were requested by Python script to store data in the MySql database management system: namely, data of geolocalized photographs (P) in mainland Portugal and data regarding the users (U) who submitted those same photographs (Figure 1).

Figure 1:

The flowchart of data collection

Source: Author’s construction.

This database will be used in this paper as a basis to describe and discuss the sequence of steps that could be followed by scholars using georeferenced Flickr photos in their tourism investigations. In the following section, the process of cleaning the collected data and the method of identifying visitors and their nationalities will be analysed and described.

When uploading photos to Flickr, the metadata from the digital camera is also uploaded (Pereira et al., 2011). But users can also place their photos on a map, and Flickr automatically assigns longitude and latitude values. This whole process introduces noise and redundancy that require pre-processing to clean and remove faulty data (Figure 2). For Chen et al. (2019a), Flickr datasets contain erroneous records resulting from malfunctioning or inaccurate hardware, or photos with missing true data. Pre-processing is therefore necessary to clean and remove faulty or irrelevant data. If the purpose of an investigation is to analyse the trajectory of each user, date-related metadata will be a very important factor to consider in order to ensure that the data has a correct timestamp associated with it. For example, if a photo does not contain the EXIF property, Flickr automatically defines when the photo was captured as being equal to the upload date, and this may generate bias in the database. However, as Girardin et al. (2008) highlight, this seems to have little impact in their study.

Figure 2:

The flowchart of data cleaning

Source: Author’s construction.

The different techniques used to perform the data clean are described below, and are summarised in Figure 2. The database of photos taken in mainland Portugal between 2010 and 2022 and available on Flickr, corresponds to 1,144,981 photos shared by 29,890 users (20.2% of photos were deleted).

The following steps were considered:

The API flickr.photos.search returns the photos associated with a rectangular space that includes the entire territory of mainland Portugal and some Spanish territory. Considering this fact, we chose to detect the country associated with each geographic point, and then remove the photos associated with Spain. This technique implied the elimination of 200,133 photographs (14.1%).

Separating visitors from residents of the study area is another important step of pre-processing the database. There is no method by which all users can be rigorously identified as visitors or locals, therefore there will always be users who may be misclassified as visitors or locals. However, the existence of a large amount of data always results in a generous sample.

In this research, the authors have decided to use the information provided in each user’s profile, as done by Yan et al. (2017), considering that the Flickr user describes the country of residence and the time zone information (see Table 2). Regarding the time zone, it was considered that if the values are different from Portugal (Madeira and Azores), the Flickr user is classified as a visitor. Portugal has two different time zones. In mainland Portugal and the Madeira archipelago, Western European Time (GMT +0) applies. In the Azores archipelago, located west in the Atlantic Ocean, the time zone is GMT-1.

It was also pertinent to analyse the activity of users who uploaded photos of Portugal to their profiles. As we can see in Table 3, a significant percentage of users (26%) uploaded only one photo taken in Portugal, with a reduction in the percentage of users as the number of photos increases. Only 1% of visitors uploaded 12 photos in their profiles. These figures deserve further reflection in future work to understand whether the extreme values, that is, users with many photos and users with only 1 or 2 photos, have an influence on the quality of the collected data.

Although the number of users with nationality is already quite large, there are still 50% of users without identification of the area of residence. This can be improved by applying other complementary methods, such as the option followed by Girardin et al. (2008) or Önder (2017): that is, the presence in the area over time as the discriminating factor. These authors explain that this strict threshold was selected in order to capture the real one-time tourists. Both options are correct and if used together could contribute to reducing errors in visitor selection, improving samples quantitatively.

It is important to note that in this second method, the choice of an appropriate threshold will depend on the research objectives and the size of the geographical area under study. However, there is still no consensus in the existing literature (Önder et al., 2016). It is usual, at a city scale, to use a 5-day threshold to consider the user to be a visitor. As explained by Kádár and Gede (2013), this short period correlates better with the stays of typical visitors in urban destinations. In studies on a national scale, it is advisable to choose a longer time span to include international visitors who made a multi-destination visit to the national territory. And in the case of Portugal, there are also large numbers of Portuguese emigrants who usually return to Portugal during the summer months for their holidays, usually staying for up to 30 days. As this is an experimental work, the authors chose not to apply this second method to distinguish resident and visiting users in this research.

For operationalisation purposes, a visitor is defined as “any person travelling to a place other than that of his/her usual environment for less than 12 months and whose main purpose of visit is other than the exercise of an activity remunerated from within the place visited” (UNWTO, 2007, p.444).

Through users’ profiles, from a population of 29,890 users (photographers), 12,144 visitors were identified (Table 2).

The graph in Figure 3 shows the monthly distribution of the percentage of photos uploaded by users classified as visitors between 2010 and 2022. It can be observed that the concentration of photos in the summer months (June to September) coincides with the high season of tourist activity in Portugal. The average seasonality rate in Portugal for the period 2016–2022 is 39.49% (TravelBI, 2023). This value measures tourism demand in the three months of highest demand (July, August and September) and as can be seen from Figure 3, the highest percentage of photographs (33%) is also taken during these three months.

Figure 3:

Users’ activity by month (2010-2022).

Source: Author’s construction

Previous studies have used various methods to detect users’ place of residence, such as spatial, spatio-temporal and content-based methods (Heikinheimo et al., 2022). The spatial approach uses the maximum number of posts per user (max posts) as the place of residence, as done by Hawelka et al. (2014) with data from Twitter. According to Heikinheimo et al. (2022), centrality measures based on the spatial distribution of geotagged posts are also used in several researches to estimate the country of residence. In this case, the mean centre or centre of mass correspond to the average (or a weighted average) of the latitude and longitude. If the median centre is selected it minimises the distance to all points and smooths the effect of outlying points compared to the mean (Heikinheimo et al., 2022). Other studies used clustering methods such as the density-based spatial clustering of applications with noise algorithm (DBSCAN) to detect the activity centres of users, as done by Luo et al. (2016) in Chicago.

Through a spatio-temporal approach, scholars can use the Max timedelta method: the place with the longest difference between the date of first and last post; the Max period (a selected period of social media activity differentiates between locals and visitors in the region) as done by Girardin et al. (2008), among others. The content approach uses the self-reported home locations in user profile; or in images or text content.

In this article, the content approach was selected, allowing identification of 7,574 users who specifically indicate the country or city of residence. The time zone displayed by users was also used to detect users’ place of residence. This new method, which can be classified as a spatial approach, allows the identification of more than 4,253 user countries. The following table shows the number and percentage of users, by country, who shared photos in Portugal in the timeframe of the study.

According to Table 4, the most representative nationalities of the users who visited Portugal, on average, in the period between 2010 and 2022, are: Spain (21%), United Kingdom (15%), France, United States and Italy with 8% respectively, and Germany (7%). In order to verify the existence of a relationship between the average proportion of users by nationality (2010–2022) and the proportion of average guests for the period 2018–2022, a linear regression was calculated (Figure 4) obtaining a coefficient of determination of 0.9128 [P(F<=f)=0,0135]. If the average proportion of users by nationality for the period 2018–2022 is selected, the linear regression has a coefficient of determination of 0.9024 [P(F<=f)=0,006836].

Figure 4:

Guests (2018-2021) versus Users (2010-2022) in Portugal

Source: Author’s construction

Although there is no information available on the website of the National Institute of Statistics of Portugal on the number of guests by nationality between 2010 and 2017, it is interesting to note that the main foreign markets according to the number of guests correspond to the most important nationalities in terms of number of foreign Flickr users who visited Portugal. This means that despite the number of Flickr users decreasing worldwide, it is still an excellent and valid research database for measuring tourism demand (Kádár & Gede, 2021).

In addition, to detect the country of residence, it was decided to test two more methods included in the Spatial approach: the Mean centre – the place where the geographic mean of user posts is located; and the Median centre – the place where the geographic median of user posts is located (Heikinheimo et al., 2022). All photographs with public access were extracted from each user (more than 33 million photos) and through the geographical coordinates provided in the photographs, centrographic measures mean centre and median centre were calculated. Subsequently, the validation of the information obtained was carried out, assuming that the users entered their location correctly. Through the mean and median centre, the degree of correctness of the country of residence was also evaluated for 10 countries, including Portugal (Table 5).

It is concluded that the median method obtains better results, in line with the results obtained by Heikinheimo et al. (2022), where the hierarchical median centres method corresponded the most to the official visitor statistics. It can also be observed that in general, the greater the distance between Portugal and the user’s country of origin, the greater the error. Spain, UK, France or Italy register the smallest errors and distant countries, such as the USA, Brazil or Canada register lowest accuracy. This is a topic that will deserve further analysis in future studies.

For users classified as visitors who do not have their country of residence registered on Flickr, it is possible to use the median method together with the time zone data available in the user’s profile. After the algorithm calculates the median of the geographic coordinates (latitude and longitude), the data obtained is validated: that is, for each user, it is confirmed whether the country resulting from the median coincides with the time zone information entered by the user. If so, the country of residence is considered valid. Subsequently, the name of the country that resulted from the median calculation is compared with the user self-reported home location (Content approach). In this way, the degree of accuracy of the algorithm is validated. These results are presented in Table 6.

In fact, the combination of methods considerably improves the accurate identification of the users’ country of origin, with very significant levels of accuracy for countries such as Portugal, Brazil, United States or Spain.

It is also possible to analyse the number of visits made by each user/visitor, identifying the first-time and repeat visitors. For this purpose, a spatio-temporal approach was selected using the Max period, that is, a selected period of social media activity differentiates between locals and visitors in the region. Taking into account that Max days corresponds to the maximum time each user stays in Portugal, a threshold of 30 days was selected. Then the number of visits of each user was counted.

Figure 5 shows the number of visits made by each visitor between 2010 and 2022, and it can be seen that 79% of users have visited Portugal only once, which means that for the period under analysis, they are considered first-time visitors.

Figure 5:

Number of visits by user in Portugal (2010-2022)

Source: Author’s construction.

As can be observed in Table 7, the number of repeaters is relatively low (21%). A more detailed analysis of the number of visits according to the country of residence shows that the Spanish, the main inbound market, are the ones who most repeat their visits to Portugal (31%), which may be justified by geographic proximity.