Researchers at the Data Science Lab at Warwick Business School took more than 200,000 images of places in the UK that had been rated for their beauty on the website
Scenic-or-Not and showed them to a deep learning model in order to find out what makes a scenic location beautiful.
The deep learning model processed all 200,000 images and labelled them with information on what was in the picture, such as “valley”, “grass”, “no horizon” or “open space”. Using these labels, the researchers were able to investigate which attributes of a scene led to higher scenic scores.
The scientists then trained a new deep learning model to look at pictures and rate them itself.
Chanuki Seresinhe, of the Data Science Lab at Warwick Business School, said: “We tested our model in London and it not only identified parks like Hampstead Heath as beautiful, but also built-up areas such as Big Ben and the Tower of London.”
Ms Seresinhe, along with WBS Data Science Lab directors Suzy Moat and Tobias Preis, used the
MIT Places Convolutional Neural Network – a deep learning model – to analyse the images from
Scenic-or-Not, which were rated by 1.5 million people, and find what attributes, such as “trees”, “mountain”, “hospital” and “highway”, corresponded to high and low scenic ratings.
Deep learning models are a particular kind of “neural network” – simulated networks of neurons, like those in the human brain – and have driven recent dramatic advances in artificial intelligence tasks, such as facial recognition and speech recognition.
Using the MIT Places deep learning model, the researchers found that features such as "valley", "coast", "mountain" and "trees" were associated with higher scenicness.
However, some man-made elements also tended to improve scores, including historical architecture such as "church", "castle", "tower" and "cottage", as well as bridge-like structures such as "viaduct" and "aqueduct". Interestingly, large areas of greenspace such as "grass" and "athletic field" led to lower ratings of scenicness rather than boosting scores.
Ms Seresinhe added: “It appears that the old adage ‘natural is beautiful’ seems to be incomplete: flat and uninteresting green spaces are not necessarily beautiful, while characterful buildings and stunning architectural features can improve the beauty of a scene.
“I am fascinated by how deep learning can help us to develop a deeper insight into what human beings collectively might understand to be beautiful.”
Dr Moat, Associate Professor of Behavioural Science at Warwick Business School and co-director of the Data Science Lab, said: ”These findings are of particular interest in the context of our previous
research, which showed that people who live in areas rated as more scenic report their health to be better, even when we take data on greenspace into account.
“Our new results shine light on why a location being green might not be enough for it to be considered attractive. This distinction has clear relevance for planning decisions which aim to improve the wellbeing of local inhabitants.”
The scientists then adapted the deep learning model to rate the scenicness of new locations, and tested it on more than 200,000 photographs of London that the model hadn’t seen before.
Professor Preis, Professor of Behavioural Science and Finance at Warwick Business School and co-director of the Data Science Lab, said: “It was fascinating to see that the model understood that bridges and historical architecture increase the perceived beauty of a scene, while grass and greenery is not necessarily scenic.
“Our previous results make it clear that scientists and policymakers alike need measurements of environmental beauty, not just measurements of how green places are. Games like Scenic-or-Not can help us collect millions of ratings from humans, but having a model which can automatically tell us whether a place is beautiful or not opens up completely new horizons.”
