Finding tourism niche on image-based social media: Integrating computational methods

Destination image marketing and social media

The portrayal of a destination's image has emerged as a paramount concern in marketing (Tasci and Gartner, 2007). The image of a destination wields significant influence over tourists’ decisions to visit, their purchasing behavior, and their overall satisfaction (Chon, 1990; Chi and Qu, 2008). Previous studies have delved into the creation, development, and impact of a destination's image on tourists’ behavior (Gallarza et al., 2002; Govers et al., 2007). Traditional sources, notably media outlets, have long served as the bedrock for shaping destination images (Cohen-Hattab and Kerber, 2004; Chon, 1990; Govers et al., 2007). Despite being influenced by indirect experiences, these traditional media images shape the initial perception of the audience, consequently influencing the selection process among potential destinations (Goodall, 1988; Govers et al., 2007).

In the pre-Internet era, individuals relied on all available organic, induced, and autonomous resources to gather information (Gartner, 1994). However, a common issue during that time was the inconsistency between advertised and received images, necessitating destination image promotion to address this concern in order to generate positive word of mouth (Chen et al., 2020; Ferrer-Rosell and Marine-Roig, 2020; Kumar and Kaushik, 2017). In the Internet era, people now turn to various online sources such as news, communities, blogs, and social media platforms to gather information about destinations (Kim et al., 2017; Pan and Li, 2011). Even before traveling, consumers already form cognitive and affective impressions of a destination, and this trend continues in the modern Internet era (Hunter, 2016; Pan and Li, 2011). The burgeoning popularity of social media, particularly platforms like Instagram, has forced destination marketers to reevaluate their traditional marketing strategies. Instagram, for instance, captures the attention of travelers (Liu et al., 2019) and has become a primary source of tourism information (Kim et al., 2017).

It is worth noting that user-generated visual content often creates a different perception from what destination marketing organizations intend (Lim et al., 2012). Presently, a destination's image is a complex amalgamation of promoted images, user-generated content, and travel experiences (Alcázar et al., 2014; Wang et al., 2020). On-off information loops have become a natural phenomenon, where online data influences offline visits, which, in turn, influences online data (Lee et al., 2017). As a result, developing a destination's image requires a strategic approach that incorporates both online and offline marketing tools (Baker and Cameron, 2008; Leung et al., 2013; McCartney et al., 2008). Another factor contributing to the complexity of tourism information distribution is the paradigm shift in how tourism information is delivered (Chung and Koom, 2015). Visual photos and videos, predominantly found on social media, have become the most sought-after type of tourism information, especially among younger generations (Fatanti and Suyadnya, 2015; Kladou and Mavragani, 2015). The use of social media by travelers helps reduce uncertainty while enhancing the utility of tourism experiences (Gretzel et al., 2006).

In fact, Instagram has emerged as a pivotal platform for tourism marketing, particularly in the context of South Korea. With an impressive monthly user count of 18,912,201 as of August 2022, this photo-sharing platform has claimed the top spot among social networking services and community apps in the country (Choi, 2022). In light of this, Instagram presents an incredibly potent tool for destination marketing specialists who aim to gain a profound understanding of tourist preferences and interests. By leveraging Instagram's vast reach, marketers can create targeted campaigns and customize travel packages that cater to specific niche markets.

While previous studies exploring Instagram's potential for visual image analysis have not specifically focused on the Korean context, they offer valuable insights into how such analyses can illuminate broader aspects of tourist behavior and experiences. For instance, Mukhina et al. (2017) conducted a data crawling analysis of images from Saint Petersburg, Russia, utilizing geotags in Instagram profiles to identify variations in attraction points favored by tourists and locals. Similarly, Liu et al. (2021) examined the night market sightseeing experience in Taiwan through the Instagram profiles of market consumers. Their findings indicated that perceived usefulness and ease of use positively influenced behavioral intentions, attitudes, and trust. Furthermore, Yu and Egger (2021) employed an ML approach to evaluate tourist images on Instagram, uncovering the significance of the color blue in driving user engagement in photos featuring natural scenery, high-end gastronomy, and sacral architecture.

These studies underscore the immense potential of Instagram as a research tool for tourism and provide valuable insights into the analysis of visual images to understand tourist behavior and experiences. By applying these insights to the Korean context, destination marketing specialists can capitalize on Instagram's widespread popularity in the country to enhance engagement and boost tourism revenues. As such, the potential of Instagram as an indispensable tool for tourism research and marketing cannot be overstated, warranting further attention and exploration from scholars and practitioners alike.

Social media visual images & machine-learning method

Recent research has established a variety of techniques for analyzing social media data for advertising and consumer research purposes, including spatio-temporal visualizations (Gomez et al., 2019), social and geographic aspects (Boy and Uitermark, 2016), picture attributes (Aramendia-Muneta et al., 2021), color composition (Yu et al., 2020), and AI prediction of popularity and user engagement (Gupta et al., 2020). The significance of social media visual image has expanded beyond the creation of destination photos (Filier et al., 2021) to include the provision of travel tips, culinary, and cultural activities (Fatanti and Suyadnya, 2015; Rossi et al., 2018; Salleh et al., 2017). For instance, Instagram visual images depict contemporary tourism concepts (Smith, 2018) for cultural tourism marketing (Mele et al., 2021).

Previous research had used social media metrics such as “likes” and “comments” to interpret the content of visual images (e.g. Yılmazdoğan et al., 2021). The distribution pattern of “likes” and “comments” on Instagram varies across image features such as people, animals, water, and multi-images (Aramendia-Muneta et al., 2021). The color scheme of tourism-related Instagram photos is also used to understand the post's response and popularity (Yu et al., 2020). The time of day that Instagram posts are also used to indicate the level of user engagement (Gupta et al., 2020). The recent increased use of machine learning (ML) has increased the automated visual analysis of destination photos. Computational methodologies aided in the processing of enormous amounts of data that humans are incapable of handling. As the algorithms used in these methods produce consistent results, the results are more reliable.

As a result, previous ML-based image categorization research has provided fruitful insights into tourism marketing. For instance, He et al. (2021) classified UGC images based on cognitive and affective elements and filtered them into “core” networks of words describing the images. Afrefieva et al. (2021) used k-means principal component analysis and topic models to cluster photos and investigate the relationship between tourist site location and common images associated with them. In other words, prior studies are effective at identifying current popularity patterns and classifying them.

Despite their insights, they paid less attention to satisfying marketers’ desire to segment consumers into specific niches and to execute focused marketing activities. However, marketers need to help consumers finding relevant information in the Internet era of information overload. The increase in the number of fully independent and air-and-hotel-only travelers implies that marketers need to provide all the options that travelers can take (Lew, 2008). According to Lew (2008), this is why social media is an excellent venue for marketers to promote niche tourism. Marketers can discover and promote specific tourism demands and use social media to create travel products. Robinson and Novelli (2005) distinguish three different approaches of niche tourism: a geographical and demographic approach that location and population of tourism consumption play a key role; product-related approaches such as activities, attractions, food, and others are the main parts of tourism; a customer-related approach that represents a specific lifestyle or holiday experiences.

In contrast to mass tourism, niche tourism focuses on the customized needs of tourists. Niche tourism can be highly encouraged by the use of social media. For example, if a large number of images on social media are related to outdoor adventure or nature tourism, it may be an indication that there is a demand for these types of tourism experiences. On the other hand, if there are fewer images related to urban tourism or cultural experiences, it may suggest that there is less demand for these types of tourism offerings. As a result, social media is used in niche tourism to create content that highlights the unique experiences and destinations provided by niche tourism providers (Cillo et al., 2021). Niche tourism providers can attract potential customers’ attention by creating and sharing engaging content. Tourism photography on social media can help to create a virtuous cycle of tourism activities that connect online and offline, as well as post-tourism activities (Lee et al., 2017; Jansson, 2018).

In this regard, this study will benefit tourism marketers by integrating previous research findings and constructing them into a new market niche identified by social media. The combination of an ML-based method and social media data will identify market segment niches by revealing data knowledge from social media information. It should be noted that the use of ML-based methods does not specify the type of market segment in advance. It is a data-driven approach to understanding and searching for market niches.

Research question

The study's research question is as follows: What insights can be gained about niche tourism sites through the use of machine learning data analysis methods on Instagram photos? In the process of answering the question, we also want to know which image factor is more popular with users than others in terms of the hashtag, image color, and photo content

Research design

In order to understand and explore a new market niche tourism through social media, the study combined traditional and recent development in ML research methods. For example, ML methods such as deep learning are used to extract variables for Poisson regression. It is common that deep learning method is used for object (e.g. building and person), and image context (e.g. outdoor vs indoor) detection (Zhao et al., 2019). Thus, the object and context of Instagram photos are identified by deep learning technology and those identified objects and contexts are transformed into categorical variables. In fact, categorical variables are also made by applying a text mining method that can cluster words into, what is called, topics (Blei et al., 2003). Digital image processing is another realm used for color detection in the photos, making another variables for statistical analysis. Finally, word embedding analysis is used to identify niche tourist attraction photos. Figure 1 depicted the structure of the research design. The detailed procedure will be explained in this section after the research question is asked below.

OPEN IN VIEWER

Figure 1.

Study structure.

Data collection

Instagram photos with the hashtag “#Seoul” were collected using the Python module “Instaloader.” The collected photos were posted between March 2018 and November 2019. The data was collected from 10 December 2019 to 12 December 2019. During the photo collection stage, the compiled data included the unique ID, title, upload date, likes, the number of comments, hashtags, and the location associated with each picture at the time of posting. The final dataset for analysis consisted of 30,013 photos with the hashtag #Seoul that were posted in English.

Data analysis

The analysis methodology is two-folded. First, to determine the popularity features of visual images, a traditional statistical analysis was used which incorporated variables derived from deep learning and text mining. Deep learning is used to recognize the dominant colors and visual-content labels in photographs. These features are incorporated into the statistical analysis as analysis variables. Additionally, there are variables that describe the visual representations generated by the text analysis method, Latent Dirichlet Allocation (LDA) topic model. The topic model algorithm was utilized to classify the visual images into four content categories, based on the user-generated hashtags and visual-content labels extracted from the deep learning study. Second, word embedding analysis was used to show how the marketer's input may be applied to find a specific photo that fits within the advertising scheme. Further details about the variables and the analysis are explained next.

Text mining analysis

The top 50 terms based on TF and TF-IDF are shown in Table 1. According to the analysis, posture hashtags are frequently used but do not help identify the posted features. For example, words such as “standing,” “sitting,” and “holding” are in the top five in TF, but fall to lower levels in TF-IDF, indicating that these are not thematic words but rather broad ones.

OPEN IN VIEWER

Table 1.

Top 50 words frequency result: TF and TF-IDF.

TF (term frequency)				TF-IDF (TF-inversed document frequency)
Rank 1–25		Rank 26–50		Rank 1–25		Rank 26–50
Standing	19,916	Plate	5333	Young	15914.2	Girl	12518.5
Man	18,488	Girl	4662	Person	15911.7	Riding	12459.8
White	16,495	Riding	4615	Large	15881.9	Sign	12406.7
Sitting	16,426	Sign	4573	Outdoor	15736.9	Walking	12305.3
Holding	15,756	Walking	4494	People	15719.7	Small	11882.3
Table	14,565	Small	4180	Indoor	15499.3	Standing	11756.7
Woman	14,071	Posing	3792	Woman	15358.7	Posing	11312.3
Person	11,225	Photo	3667	Wearing	15236.3	Photo	11116.8
Young	11,015	Shirt	3641	Group	15188.1	Shirt	11075.3
Large	10,336	Filled	3487	Table	15172.9	Filled	10824.3
Outdoor	9426	Board	3287	Food	15116.5	Board	10483.5
People	9352	Wooden	3197	Street	15048.8	Wooden	10324.6
Indoor	8610	Cake	3173	Black	14991.8	Cake	10281.5
Wearing	7972	Bench	3089	Building	14986.2	Bench	10128.9
Group	7871	Snow	3057	Room	14750.6	Snow	10069.9
Food	7728	Covered	2905	Red	14695.0	Covered	9783.0
Street	7599	Train	2855	Front	14639.9	Train	9686.1
Black	7495	Computer	2766	Holding	14626.9	Computer	9510.5
Building	7485	Kitchen	2706	Sitting	14262.0	Kitchen	9389.8
Room	7091	Boy	2693	White	14222.2	Boy	9363.4
Red	7005	Blue	2687	City	13778.1	Blue	9351.2
Front	6922	Parked	2645	Phone	13657.4	Parked	9265.2
City	5834	Playing	2627	Water	13483.1	Playing	9228.0
Phone	5705	Umbrella	2625	Plate	13285.7	Umbrella	9223.9
Water	5526	Field	2590	Man	12898.2	Field	9151.0

However, there are words that are related to the theme. First, it appears that the description of people is centered on the words “young,” “woman,” “person,” (and “people” collectively). These are the terms with the highest TF and TF-IDF scores. This indicates that these hashtags are common and reveal a theme composition in the dataset. Second, it appears that dining is one of the most popular Instagram topics in Seoul. Higher TF and TF-IDF scores were found for the words “indoor,” “table,” “food,” and “plate.” Third, as shown in Table 1, “outside” is associated with “building,” “city,” “street,” and “walking.” This is understandable given that the dataset's environment is the city rather than the countryside, where visitors can enjoy a scenic view of valleys and mountains.

Although the TF and TF-IDF analyses shed light on the dataset's likely three themes (woman-people, food-dining, and building-outdoor), there is insufficient evidence to determine whether the themes or words identify photo postings. These three themes are, in fact, analogous to the three classification categories revealed by deep learning technology. In this sense, hashtags are similar to photo content features in that they contain reliable information. However, it is possible that the scheme's number will differ when it is recognized automatically via a computational model. Simple counting measures, such as TF or TF-IDF, as well as pre-defined category classification (deep learning), do not take multiple classification entries into account. To investigate this issue, the topic model method was employed.

However, when compared to human inspection, automatic computer systems do not always produce the best-optimized figure (Arun et al., 2010; Chang et al., 2009). As a result, each issue and clustered subjects were investigated using the LDA visualization technique, varying the number of topics (Sievert and Shirley, 2014). After reviewing automatic computational indexes and topic contents, it was determined that four topics are the best fit for human evaluation. The number differs from the initial assessment based on TF and TF-IDF. Following the topic modeling analysis, it became clear that the “people” category needed to be divided into two distinct topics. Note that LDA topic modeling is more or less descriptive in our analysis because it is used to identify themes and words for finding a niche, not the niche itself.

Figure 2 depicts the four identified subjects as well as the top 20 terms that are most likely to belong to those topics. The dataset's topics can be distinguished according to who posted them and where they were posted. Because the words in each topic can overlap, the topics with the most relevant keywords in the topic were labeled. The relevance of a word is determined by the likelihood that it belongs to a specific topic category. For example, the first topic in Figure 2 shows that “woman” has the highest probability of belonging to the topic “indoor-people-group-woman,” while “woman” also belongs to the topic “table-plate-food,” but not as strongly to the topic “indoor-people-group-woman.” Only word descriptions of agents and objects, such as man, woman, and table were chosen for statistical analysis rather than behaviors or status, such as sitting, holding, and so on. It only leaves us with the words “indoor,” “outdoor,” “man,” “woman,” “person,” “people,” “young,” “street,” “food,” “room,” “table,” and “plate.” These words will be used as variables in the Poisson regression analysis to determine what types of Instagram photos get the most “likes.”

OPEN IN VIEWER

Figure 2.

Latent Dirichlet allocation (LDA) topic model result.

Poisson regression analysis and top-pick postings

The results of the stepwise Poisson regression analysis are shown in Table 2. Due to the collinearity issue between words, the words “people” and “room” are excluded from the analysis. The results of the regression analysis are summarized below.

OPEN IN VIEWER

Table 2.

Poisson regression results: stepwise modeling with posting time, picture color, and hashtag referenced by deep learning result.

		Model 1			Model 2			Model 3			Model 4			Model 5
		IRR		SE	IRR		SE	IRR		SE	IRR		SE	IRR		SE
Const.		19.33		0.11	20.22	***	0.12	18.25	***	0.12	18.69	***	0.13	19.16	***	0.13
Posting time	Afternoon (12:00 – 18:00)	0.61	***	0.00	0.62	***	0.00	0.61	***	0.00	0.60	***	0.00	0.60	***	0.00
	Evening (18:00 – 22:00)	2.68	***	0.02	2.69	***	0.02	2.61	***	0.02	2.57	***	0.02	2.57	***	0.02
	Night & early morning (22:00 – 9:00)	1.26	***	0.01	1.27	***	0.01	1.26	***	0.01	1.27	***	0.01	1.27	***	0.01
Number of hashtags		0.99	***	0.00	0.99	***	0.00	0.99	***	0.00	0.99	***	0.00	0.99	***	0.00
Image content Color (image detection API)
	Brown				0.86	***	0.00	0.90	***	0.00	0.90	***	0.00	0.90	***	0.00
	White				0.98	***	0.00	0.99	***	0.00	0.97	***	0.00	0.97	***	0.00
	Blue				0.70	***	0.01	0.69	***	0.01	0.71	***	0.01	0.71	***	0.01
	Green				1.07	***	0.01	1.12	***	0.01	1.11	***	0.01	1.10	***	0.01
	Red				1.08	***	0.01	1.09	***	0.01	1.07	***	0.01	1.07	***	0.01
	Yellow				1.07	***	0.01	1.13	***	0.01	1.13	***	0.01	1.13	***	0.01
	Purple				1.06	***	0.02	1.06	***	0.02	1.07	***	0.02	1.07	***	0.02
	Pink				0.91	***	0.01	0.93	***	0.01	0.90	***	0.01	0.89	***	0.01
	Orange				1.77	***	0.02	1.93	***	0.02	1.92	***	0.02	1.92	***	0.02
Teal					0.56	***	0.02	0.56	***	0.02	0.58	***	0.02	0.58	***	0.02
Image content analysis with text mining
	Person							1.24	***	0.00
	Man										0.97	***	0.00	0.98	***	0.00
	Woman										1.15	***	0.01	1.15	***	0.01
	Young										1.28	***	0.01	1.27	***	0.01
	Group										0.92	***	0.00	0.93	***	0.00
	Food							0.86	***	0.00
	Plate										0.99	***	0.00	0.98	***	0.00
	Table										0.89	**	0.00	0.88	*	0.00
	Outdoor													1.01	*	0.00
Street														0.92	***	0.00
LR X2		92158.6***			99,643***			106514.2***			115062.9***			115560.5***
Pseudo R2		0.059			0.063			0.068			0.073			0.074
N		23,123

Note: IRR: Incidence Rate Ratios: SE: Standard Error, Reference for color variation—either gray or black.

First, evening (18:00–22:00) posts tend to have more “likes” than those uploaded at any other time. The odd ratio is 2.57, meaning that evening postings get about 2.6 times more “likes” than those in the morning (9:00–12:00). It is not sure if the number of “likes” is synchronized with posting times, later in the evening or later the next day, but the number of “likes” is higher than others when the upload time is later in the evening. Second, warm color tones in Instagram posts receive more “likes” than other colors. Color schemes such as “the colors green, yellow, and orange indicate that more than one odd ratio for the number of likes is statistically significant.” In terms of effects, the color orange is the most potent. In model 2, another warm color, “red,” was significant, but adding the “hashtags” variable nullified the effect. Third, when compared to building features, people—particularly “young” and “women”—receive more “likes” on Instagram. The feature related to “food” is identical to the feature related to “building.” Furthermore, the types of visual images that achieve popularity on Instagram based on five features were identified according to the results of the Poisson regression analysis: (1) uploaded in the “evening,” (2) prevalence of color “orange” and hashtags containing, (3) “young,” (4) “woman,” and (5) “outdoor.”

Figure 3 indicates the top three most selected photos in these categories. Images belonging to the same series were excluded. For example, there are postings in the same places within a few seconds or minutes later. Those were treated as the same image content. Interestingly, the locations of these postings are somewhat scattered, but they are top-rated tourist destinations in Seoul. For example, Ikseondong is a place that reveals the phenomenon of so-called “over-tourism” (Dodds and Butler, 2019).

OPEN IN VIEWER

Figure 3.

Top selected photos according to Poisson regression analysis. (Note: the map is the official city tour guide, and the arrows indicate the location of the photo.)

Figure 3 shows a photo of Ikseongdong taken in front of a café, which is an example of a niche activity within the well-known traditional Korean house “hanok” village. The photo showcases an unexpected visit to a modern and Western café, which deviates from the typical tourist experience of touring around the Hanok village. This finding has implications for the development of niche tourism spots that cater to the changing interests and preferences of tourists. Similarly, the photo taken at N Seoul Tower captures a “natural green image” with postures, showcasing a unique way of experiencing the attraction beyond the typical skyline views. This finding suggests that tourists are interested in exploring alternative ways of experiencing popular tourist attractions, which could lead to the development of new niche tourism activities. Additionally, the photo taken at Dongdaemun Plaza depicts stylish photo-taking, which is a niche activity compared to the typical shopping experience in the near area. This finding highlights the potential for niche tourism activities that cater to tourists’ interests in photography and self-expression. Overall, these findings provide insights into the development of niche tourism spots that cater to tourists’ evolving interests and preferences, contributing to the promotion of diversified and sustainable tourism.

Word embedding analysis for top picks with word combinations

This study also focused on more specific features of popular Instagram postings. While previous examples indicated the most “likes” postings, they do not represent the whole series of Instagram posts under the diverse domains. In order to find other popular postings in a relatively small and niche domain, Instagram postings that belong to a topic domain whose numbers are relatively low in terms of the total number of “likes” but highest in the sectoral domain we also tracked. Table 3 presents this tracking result in two domains: the most popular domain and the domain that received less attention among the statistically significant variables in Table 2. As word embedding can combine words in the distance calculation, the word combination was used. Table 3 left side indicates the top similarity order for the equation “woman”  +  “young”  +  “outdoor,” whereas the right side shows the order for the equation “table”  +  “sitting”  +  “indoor.”

OPEN IN VIEWER

Table 3.

Glove word embedding result.

Women, young, outdoor		Table, sitting, indoor
Man	0.754	Food	0.688
Outdoor	0.738	Woman	0.594
Wearing	0.707	Black	0.560
Holding	0.705	Small	0.556
Standing	0.682	Person	0.545
Girl	0.665	Plate	0.537
Person	0.658	Window	0.524
Building	0.618	Photo	0.524
Boy	0.615	Holding	0.511
People	0.604	Front	0.502
Front	0.583	Building	0.499
Posing	0.583	Man	0.497
Street	0.555	Cup	0.469
Yellow	0.546	Wooden	0.464
Black	0.529	Room	0.456

After reviewing the words from the Glove word embedding analysis, keywords from the previous topic modeling analysis such as “girl,” “building,” “boy,” “street,” “food,” “woman,” “plate,” and “window” were selected from a marketer's perspective. Figure 4 depicts the most popular pick postings based on these keywords. The image on the left accurately depicts tourist activities. Visitors wearing traditional Korean clothing, known as “Hanbok,” are granted free admission to old Seoul palaces by the Korean Cultural Heritage Administration. This policy is clearly demonstrated in Instagram photos of a tourist wearing a Hanbok and touring the nearby palace area. Although not all of the words correspond to the exact image content, they do provide insight into tourist activities and how they were shared on Instagram. The possible reason for this word mismatch may be from the way photos were uploaded. Often users upload a series of photos at the same time in one posting with hashtags covering all photos uploaded. However, in general, the results of the word embedding method, which used insights from topic modeling and Poisson regression, provided useful information for communicating more strategically with potential tourists.

OPEN IN VIEWER

Figure 4.

Top selected photos with themes based on word-embedding analysis.

Theoretical contributions

While previous studies have examined the benefits of using unique approaches to analyze images of destinations, it remains a challenge to determine which aspects of images shared on social media are linked to better attractiveness in the context of different types of tourism themes. To address this gap, this study aimed to integrate standard statistical analysis with text mining and deep learning analysis methods. By doing so, this study demonstrates a more comprehensive approach to analyzing social media data that takes into account both the visual and textual information available in these posts. The findings of this study suggest that cutting-edge analysis methods, such as machine learning and text mining, can be used alongside traditional variable-based statistical analysis to identify the visual attributes of destinations that are most strongly associated with niche tourism themes. This analysis approach can help destination marketers and policymakers develop more targeted and effective strategies for promoting niche tourism products and services that cater to the changing interests and preferences of tourists.

Overall, this study's innovative approach and findings provide valuable insights for tourism researchers, marketers, and policymakers. The integration of cutting-edge analysis methods with traditional statistical approaches provides a more complete understanding of the visual and textual attributes of destinations that are most strongly associated with niche tourism themes. This study underscores the importance of data-driven approaches for developing evidence-based tourism policies and strategies that can enhance the competitiveness and sustainability of destinations.

Managerial contributions

In the era of social media, tourists form cognitive and emotional impressions of a place even before visiting it (Hunter, 2016; Pan and Li, 2011). As a result, destination marketers are required to adjust their traditional marketing strategies to meet the changing landscape. The current study provides valuable insights for marketers seeking to employ social media as a marketing tool for tourism. Our methodology, which combines popularity and visual image analysis components, can assist in identifying specialized tourism sectors and appropriate promotional regions, particularly in the post-COVID-19 era when marketing resources are limited, and tourist demands are diverse. Furthermore, our methodology includes a way for destination marketers to actively pursue promotional interests through word combination.

By examining which Instagram posts receive the most likes, our study provides insights into the tourist attractions that are most popular and what visitors find most appealing. This information can help visitors create personalized itineraries tailored to their unique interests and preferences, leading to a more fulfilling travel experience in Seoul. Overall, the present study contributes to the body of knowledge on tourism destination visual image analysis in social media and provides practical implications for destination marketers looking to leverage social media for tourism marketing.

Limitations and future research

Despite the study's practical and theoretical implications, it has some limitations. Because this study examined 30,013 Instagram posts in one city, first, it is difficult to say that the findings will apply equally to other destinations’ tourism marketing contexts. Second, because the postings in this study were obtained in English using the search phrase #Seoul, the analysis omitted social media posts in other languages like Chinese and Japanese. As a result, the findings of this study should be interpreted with caution. Third, our research is based on Microsoft Azure, which was not specially trained for tourist visual images. For example, the “animal” category was not useful to analyze because the algorithm could not distinguish between live animals and dead ones for cooking. In the future, the creation of tourism-specific trained algorithms might considerably improve our research outcomes. Video clips could be incorporated into the existing technique in future research. Although video for tourist information has been considered, it has rarely been studied. While the current research focuses on visual image processing, future research could integrate video tourist footage.

Finally, it should be noted that the present study is exploratory in nature and focuses on a specific set of techniques for identifying niche tourism themes in Instagram posts. While our findings offer valuable insights into the potential of using these techniques to identify destination image attributes associated with niche tourism themes, further theoretical framework development is needed to explore the underlying motivations and behaviors of tourists in relation to these themes. Additionally, future research could apply similar techniques to other social media platforms to gain a more comprehensive understanding of the visual and textual attributes associated with niche tourism themes, and to identify effective strategies for promoting and marketing niche tourism experiences.