Destination positioning: do DMOs promote their destination distinctly in their visual marketing?
Destination branding: does tourist photography align with how DMOs promote the destination?
How distinct and aligned with UGC is European capitals’ DMO branding on Instagram?
1. Lyndon JB Nixon
MODUL University Vienna
School of Applied Data Science
Vienna, Austria
How distinct and aligned with UGC is European
capitals’ DMO branding on Instagram?
24 January 2024 ENTER2024 – Research Track Page 1
2. Goal of this paper
Destination positioning: do
DMOs promote their
destination distinctly in
their visual marketing?
Destination branding: does
tourist photography align
with how DMOs promote
the destination?
24 January 2024 ENTER2024 – Research Track Page 2
Bing Image Creator.
(left) a destination marketing poster for Izmir showing the major sights
(right) tourists in Izmir taking photos of the major sights
3. Background
• Destination image research has been done in e-tourism since decades
• The “beliefs, ideas and impressions that a person has of a destination”
(Crompton, 1979) can influence travel choice, what they do at the
destination and if they return
• Measuring Touristic Destination Image (TDI) requires a descriptive model,
a data source and a methodology to extract from the data in terms of the
model.
• Images are increasingly important in TDI formation (Kim et al., 2014)
• While text mining has been well studied, extracting TDI from images or
videos was either manual or hampered by inaccurate tools…
24 January 2024 ENTER2024 – Research Track Page 3
4. Computer Vision and Destination Image
• Recent advances (2012-) in AI have improved significantly the accuracy of
computer vision models, incl. image classification (label an image with
the most relevant class)
• Computer vision is a valid method to extract TDI from images (Picaso &
Moreno-Gil. 2017)
• Pre-trained models are available (via API or code), with high accuracy
scores, but their training is focused on a large, broad set of labels
• E-tourism researchers have annotated photos with those labels, using
clustering techniques to determine “destination image” specific
characteristics, but what if the original labels were not ideal?
• We have found that – for TDI – accuracy is lower (Nixon, 2018)
24 January 2024 ENTER2024 – Research Track Page 4
5. TDI classes // CV model accuracy
24 January 2024 ENTER2024 – Research Track Page 5
I have used transfer learning to fine-tune state of the art CV
models for 18 specific classes for TDI measurement…
6. TDI classes // CV model accuracy
24 January 2024 ENTER2024 – Research Track Page 6
Different pre-trained models were compared, with Vision
Transformers the state of the art best model architecture…
7. Fine-tuned CV model for TDI
24 January 2024 ENTER2024 – Research Track Page 7
The final model for destination image
measurement is on HuggingFace.
It is fine-tuned on BEiT-L (best accuracy
results on the ground truth dataset)
https://huggingface.co/
lyndonnixon/destination-image-classifier
8. Measuring TDI for a destination
• To get a measurement of TDI, we can use a representative sample of
photography and get a label for each photo in the sample
• A count of the frequency of each label in the set can form a TDI
measure (Stepchenkova & Zhan, 2013)… but how to use this measure
for analysis?
• Our intuition is to represent the counts as 18-dimensional vectors
(like AI “embeddings” – vectors are well understood in mathematics)
• The absolute frequency is divided by the sample size to ensure
comparable measures (across samples of different sizes)
• The sum of the relative frequencies of all labels will be 1
24 January 2024 ENTER2024 – Research Track Page 8
9. Compositional Data Analysis (CODA)
• Our TDIs (vectors) are compositional data as the sum of all values is
fixed. If one value should increase, another must decrease.
• Compositional data analysis (CODA) is the field of applying data
analytics on compositional data (Coenders & Ferrer-Rosell, 2020)
• The vectors (compositions) are transformed into “centred logarithms
of ratios” (ensures linearity in the differences between values)
• CODA statistical approaches can now be used to quantify differences
between compositions, e.g. perceived and projected TDI (Marine-Roig
& Ferrer-Rosell, 2018).
24 January 2024 ENTER2024 – Research Track Page 9
10. Data collection
Instagram is chosen as a leading source of imagery co-creating destination image for consumers
(esp. Gen Z/millennials).
We use the top 10 visited European capitals (Eurostat, 2019) but had to remove Prague due to a lack
of data (sorry!)
We collected DMO and UGC photos posted within 2023.
24 January 2024 ENTER2024 – Research Track Page 10
11. Comparing DMO TDIs
Do DMOs promote their destination distinctly in their visual marketing?
We cluster TDIs by cosine distance and use k-means:
24 January 2024 ENTER2024 – Research Track Page 11
Cluster 0 Cluster 1 Cluster 2 Cluster 3 Cluster 4 Cluster 5
Budapest
Rome
Madrid Athens
Paris
Berlin
Stockholm
Lisbon Vienna
12. Comparing DMO TDIs
Do DMOs promote their destination distinctly in their visual marketing?
We compare clusters by the log ratios of geometric means of the TDI classes
(overall vs cluster). The bar plot highlights the relative differences by cluster:
24 January 2024 ENTER2024 – Research Track Page 12
13. Comparing DMO and UGC TDIs
Does tourist photography align with how DMOs promote the destination?
UGC photos materialise what visitors deem important at a destination (Pan et
al., 2014) & are a valid source for TDI (Hunter, 2016)
DMO marketers want to see their promoted characteristics reflected in
consumer’s TDI (Ji & Wall, 2015)
Aitchinson distance is used to measure similarity in distribution of classes. A
higher value indicates less similarity between UGC and DMO TDI.
Our vectors are centred log-ratios: positive values indicate classes relatively
more present in the TDI. A simple mathematical difference between DMO and
UGC values indicates who focuses more on a characteristic of TDI.
24 January 2024 ENTER2024 – Research Track Page 13
14. Comparing DMO and UGC TDIs
Does tourist photography align with how DMOs promote the destination?
24 January 2024 ENTER2024 – Research Track Page 14
City Vector Distance Sign. Diff. in Class DMO log-ratio value UGC log-ratio value
Athens 0.78 Shops & Markets 1.9 0.77
Paris 2.82 Accommodation
Entertainment
-7.39
2.35
0.08
-0.2
Rome 2.72 Gastronomy
Museum
0.84
1.76
2.64
3.1
Vienna 1.34 Museum
Roads & traffic
3.19
0.6
1.07
1.48
15. Findings
Destination positioning: do DMOs promote their destination distinctly in
their visual marketing?
From the 9 capitals, Lisbon, Madrid and Vienna show the most distinct
positioning, highlighting a different mix of TDI characteristics than their
competitors, e.g. Vienna highlights museums and historical buildings.
Destination branding: does tourist photography align with how DMOs
promote the destination?
UGC photos from Athens are closest to the DMO’s visual marketing. Paris
DMO focuses more on entertainment whereas UGC suggests traveller interest
in accommodation. Rome could promote more its gastronomy and museums.
24 January 2024 ENTER2024 – Research Track Page 15
16. Future work / an invitation
• We showed how computer vision and CODA can be used to extract a TDI
from a set of photos and represent it mathematically for data analysis
and visualisation
However!
- the set of visual classes used is always open to discussion
- the assumption is that extracted TDIs do generalise to the larger
population’s TDI (at least e.g. the Instagram user demographic)
- we lack correlations to other DM metrics such as: would a projected TDI
(DMO) closer to the perceived TDI (UGC) actually lead to more visitors?
24 January 2024 ENTER2024 – Research Track Page 16
17. Thank You / Tesekkürler
lyndon.nixon@modul.ac.at
24 January 2024 ENTER2024 – Research Track Page 17
Image generated by Bing Image Creator.
Prompt: “A researcher in Izmir presenting
tourism research”
Editor's Notes
15 + 5 min
Cluster 0 cities (Budapest, Rome) show relatively more entertainment and monument content. Cluster 1 (Madrid) has relatively much more shops & markets content and much less water content (i.e. lakes, seas) than the others. Cluster 2 (Athens, Paris) proves to be the brands which are closest to the overall mean, suggesting they may not stick out from the other destination marketing. Cluster 3 (Berlin, Stockholm) are brands which show relatively more modern buildings and trees. Cluster 4 (Lisbon) and cluster 5 (Vienna) show the most relative variation in attributes in their marketing. Lisbon highlights more than other destinations animals, beach, landscape and water; Vienna is highlighting more its historical buildings, gastronomy and museums.