Research press release


Scientific Reports

Inequalities in UK can be detected using deep-learning image analysis



今回、Esra Suelたちの研究グループは、ディープラーニングの手法により、一般公開されている街路画像と政府統計を使って、英国の4つの主要都市(ロンドン、バーミンガム、マンチェスター、リーズ)における不平等を検出するため、コンピュータープログラムを訓練した。コンピュータープログラムは、まず15万6581種の郵便番号に対応するロンドンの画像(合計52万5860点)を使って訓練され、次に、そのわずか1%に当たる(ウェストミッドランド、グレーターマンチェスター、ウェストヨークシャーで収集された)追加データを使って微調整を行った上で、バーミンガム、マンチェスター、リーズにおける不平等の検出を行ったところ、ロンドンの場合とほぼ同じ予測精度が得られた。



Social, economic, environmental and health inequalities within UK cities can be detected using a deep-learning computer method that analyses publicly available street imagery. The findings are published in Scientific Reports this week.

Detailed measurements of the substantial inequalities that exist within large cities like London are crucial for informing and evaluating policies that aim to reduce them. However, only a small number of countries have fully linked statistical datasets that allow for real-time measurements.

Esra Suel and colleagues used a deep-learning approach to train a computer programme to detect inequalities in four major UK cities - London, Birmingham, Manchester and Leeds - using publicly available street level images and government statistics. Trained on 525,860 images from London corresponding to 156,581 postcodes, the programme predicted outcomes with similar accuracy in the other three cities, after it had been fine-tuned with only 1% of additional images collected in the West Midlands, Greater Manchester and West Yorkshire.

The authors hypothesized that some features of cities and urban life, such as quality of housing and the living environment, have direct visual signals that a computer could recognize. These visual signals include building materials and disrepair, cars, or local shops. Combined with government statistics on outcomes such as housing conditions, mean income, or mortality and morbidity rates for one city, images may be used to train a computer programme to detect inequalities in other cities that lack statistical data.

The authors found that their computer programme was most successful at recognizing differences in quality of the living environment and mean income, and least successful for crime and self-reported health.

doi: 10.1038/s41598-019-42036-w


メールマガジンリストの「Nature 関連誌今週のハイライト」にチェックをいれていただきますと、毎週各ジャーナルからの最新の「注目のハイライト」をまとめて皆様にお届けいたします。