604 / 2019-03-05 00:06:08

Geographical generalisation of convolutional neural networks

convolutional neural network,streetview,urban design

全体主题 > Artificial Intelligence in Urban Planning and Management

Deep convolutional neural networks(CNN) have achieved human-level accuracies in many machine vision tasks such as scene recognition, object detection and image segmentation. This level of computer intelligence has led to advances in intelligent transportation, medical imaging and robotics. Recent research in geography and urban analytics have used these methods in city identification, path learning, and in estimating socio-economic profiles.

A key limitation in using CNNs is the lack of research on how such models generalises across geographies. The aim of this research is to study this research gap. Specifically, we aim to study the extent CNN model generalises for both classification and localisation problems for street frontage case study. The quality of street frontages is an important factor in urban design, as it contributes to the perception of safety and conviviality of the public space.

For this study, we collect street images from Google Streetview API (2017 Google Inc. Google and the Google logo are registered trademarks of Google Inc.) for eight cities around the world namely London, Paris, New York, Barcelona, San Francisco, Hong Kong, Tokyo and Kyoto. We trained a CNN classifier and R-CNN localiser first using the data in London. We then used the train model to make inference for all eight cities to study the extent the model generalises across geographies. We then trained a model using four cities namely London, New York, San Francisco and Tokyo to make inference for the other four cities.

The research finds encouraging results in the use of a single-source image classifier in inferring the urban frontage quality of a different city. The results are expectedly more accurate for cities that are more similar to London either physically, temporally and culturally in the case of NYC and Paris and poorer for cities more distant away such as Tokyo. This finding aligns with Tobler’s first law of geography where nearer things are more similar than distant things. The research also finds that the model trained with using four cities generalises better for the other four cities. These research suggests the inclusion of samples across geographies improves the overall generalisation of CNN models. Further research is needed with the aim of developing a universal geographical model.