A deep learning approach for urban building classification in Megacity Tokyo

Title: From human mobility to building functions: A deep learning approach for urban building classification in Megacity Tokyo

deep-learning-building-classification-tokyo

Abstract

Accurately identifying building function is essential for urban management, urban renewal, and promoting sustainable city development. Previous studies on building function classification have primarily focused on extracting external physical characteristics from remote sensing imagery and socio-economic attributes from Points of Interest (POI) data. However, these studies often overlook the patterns of human mobility within buildings, making it challenging to identify building functions accurately. To address these issues, this study mines the latent features embedded in POI and human trajectory data, and constructs a deep learning model, which integrates POI category semantics and human mobility patterns to identify urban building functions. We evaluated the proposed model in the 23 districts of Tokyo, Japan. The results indicate that the proposed model is able to extract features obtained from diverse data sources to identify building functions, achieving a test accuracy of 90.27 % and a Kappa coefficient of 0.8858. The building function mapping results in Tokyo demonstrate that the proposed model can accurately classify building functions in megacities. This study finds that human mobility patterns within buildings significantly improve the identifying accuracy of residential and commercial buildings. The building function mapping results of this study can provide effective data support for urban planning in Tokyo.

Highlights

A deep learning model (STAF-Net) to classify urban building functions is proposed.
STAF-Net leverages both POI semantics and human mobility patterns of a building.
It outperforms other baselines, achieving a classification accuracy of 90.27 %.
Successfully identified the functions of over 93 thousand buildings in Tokyo.