The label "coast" was one of the characteristics associated with higher scores of "scenicness" in the study.

Researchers from the Data Science Lab at Warwick Business School have applied a quantitative approach to something traditionally thought of as subjective by asking: what qualifies as “beautiful scenery?”

The team, led by Chanuki Seresinhe, used deep learning methods to train a computer to recognize and categorize characteristics of geo-tagged images.

They used more than 200,000 images of locations throughout the UK that had been rated based on their scenic beauty via the website Scenic-or-Not. About 1.5 million people rated the photos, and only images that had been rated at least three times were used for the research. The rating scale ranged from 1 to 10 – with 10 being considered “very scenic” and 1 equating to “not scenic.”

Seresinhe and colleagues used the MIT Places Convolutional Neural Network – a deep learning model – to analyze and label certain aspects of the images to determine if the overall image was considered scenic or not. The model simulates networks of neurons just like those in the human brain. These types of models have generated great interest and breakthroughs for both facial and speech recognition, among other AI-related tasks.

Examples of labels included whether there was a body of water present, trees and mountains, or manmade structures like highways and buildings. These labels allowed the team to investigate which specific elements of a scene led to higher “scenic” scores.

The next step was to have the model analyze another group of images it hadn't seen before and rate them on their "scenicness." The model proved to be successful with both natural beauty as well as manmade areas.

“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,” said Seresinhe, of the Data Science Lab.

Some of the top keywords that correlated to higher ranks of beauty were “valley,” “coast,” “mountain,” and “trees.” Labels like “viaduct,” “tower,” and “castle” were also considered beautiful when looking at images of manmade environments.

But a more unexpected result of the study was that flat, open spaces or simple greenery did not necessary correlate to higher rankings. Labels like “grass,” “no horizon,” and “athletic field” were associated with some of the lowest ratings of “scenicness.”

The authors correlate the lack of appeal of flat, open spaces with Jay Appleton’s theory of “prospect and refuge,” which suggests that humans evolved to prefer outdoors spaces that can easily be surveyed for “prospects” or contain “refuge” to easily hide and escape potential dangers.

“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,” said Seresinhe.

Other more obvious negative terms included “construction site,” “hospital,” and “parking lot.”

The urban areas that ranked well had elements reminiscent of countryside scenery – such as appealing canals or tree-lined walking paths. But the model also learned that images of historical architecture and bridge-like structures can also be deemed as beautiful or scenic.

The latest study complements and provides additional insight into previous research from the Warwick Business School, which found that people living in areas considered more scenic also reported having better health compared.

The team believes that information gathered from the deep learning model could help policy makers and planners make better-informed decisions when examining where to build highways, parks or new housing complexes, and taking into consideration how those construction projects may influence human health.

Using AI and the deep learning model also allowed the researchers to examine and characterize a much larger dataset than a human encoder could handle, and prevented them from having to make conclusions based on a smaller sample.

“The ability to crowdsource large amounts of data, coupled with recent advances in computer vision methods, is opening up a new avenue for research, allowing us to investigate visual perceptions of our environment,” wrote the study authors.

The authors noted that the model can be improved for future research by incorporating more features and data. For example, historical buildings earned higher scores in this study, but that could be in part because categories for modern types of architecture were absent from this data.

Despite being a manmade structure in an urban setting, this image of Big Ben scored high in the study.