salzburg bild stadt

Project Profile

Title: Data-driven Tourism for Sustainability

Type: Research project funded by FFG (the Austrian Research Promotion Agency)

Collaborating Organizations: DatenVorsprung GmbH, TU Graz, FH Salzburg, Donau Uni Krems, TSG Tourismus Salzburg GmbH, Fuscher Freges GmbH (TVB Bruck/Fusch), Nexyo GmbH

Location: Salzburg, Austria

ObjectiveDevelop deep-learning models for predicting and managing future tourism flows in the city of Salzburg, to prevent overcrowding and optimize the living, as well as the tourism environment

Problem Description and Motivation

  • Overcrowded streets: due to its topography, Salzburg’s historic city centre is characterised by multiple narrow streets that are easily filled with people (both tourists and residents)

  • Seasonal fluctuation: visitors from all around the world visit this iconic city throughout the entire year, in very high numbers

  • Overloaded sightseeing hotspots: the high number of tourists visiting the main attractions create overcrowding nearby specific points of interest

  • This poses a challenge to the city authorities, that have to manage a very delicate balance between different key players, on different levels:

    • From an economic perspective – guaranteeing there is profit for all parties involved

    • From a sustainability perspective – ensuring that the environment does not suffer from mass tourism (with high levels of pollution and waste)

    • From a social perspective – providing an optimized experience to visitors and a well balanced relationship between residents and tourists

Technologies and Solutions

3 years
Tourism data gathered from the "Salzburg Card" (more details below), data on weather and global holidays, mobile phone location data (covering 1-2% of tourists), OpenStreetMap
Models (AI-based and others)
RNNs, LSTMs, NeuralODE, Transformer, CT-GRN, ARIMA (statistics-based)
Using DatenVorsprung's technology to ensure the prediction models we use are robust throughout unexpected anomalies (such as global pandemics or big scale events)

Our technology is especially made for regions which suffer from overcrowding and overwhelming tourism flows. It should be used to prevent it, to guide tourists to places and distribute them over the city.

Innovation and Insights

We applied a wide range of recurrent neural network (RNN) models to tourist flow predictions and for this used the stems of the dataset of the “Salzburg Card”, which was provided to us. 

  • The “Salzburg Card” grants the visitors access to several touristic attractions and the dataset we used consists of the time-stamps of entries to each point of interest (POI), the normalised visitor count for a given hour, and the location of the POI
  • We then compared the accuracy of the different models, especially deep-learning (DL) models and ARIMA (traditional method on forecasting tourist flow time-series)
  • We found out that the transformer-based DL models, which we created, outperformed traditional ARIMA
  • Creation of first high frequency tourism dataset, which results of real life data
  • We discovered that graph neural networks (GNNs) are more suitable for incorporating spatial structure using sparse geolocation data

© FFG dTS; Agent-based simulation of tourism flow through the city of Salzburg

In the simulation above you can see several agents, here the tourists, represented as yellow dots, freely moving throughout Salzburg’s historic city centre. The green circles correspond to the main touristic attractions, the so called points of interest (POIs), and the blue figures represent accommodations. This simulation allows us to visualise how agents move between different hotspots and the density of the agents in different streets. The thicker and darker the line gets, the higher the density of tourists in a specific street.

Publication List

  • Lemmel, J., et al., “Deep-Learning vs Regression: Prediction of Tourism Flow with Limited Data”, arXiv e-prints, 2022, doi: 10.48550/arXiv.2206.13274
    Accepted for publication at the IJCAI’22 Workshop AI for Time Series Analysis (AI4TS-22)
  • Lemmel, J., et al., “Prediction of Tourism Flow with Sparse Geolocation Data”, arXiv e-prints, 2023, doi: 10.48550/arXiv.2308.14516
    Accepted for publication at the 5th International Data Science Conference – iDSC2023

Further questions? Please do not hesitate to contact us.