Urban Functional Zone Mapping with a Bibranch Neural Network via Fusing Remote Sensing and Social Sensing Data

Abstract

Urban functional zones (UFZs) are the urban spaces divided by various functional activities and are the basic units of daily human activities. UFZ mapping, which identifies the UFZ categories in different spatial areas of a city, is of considerable significance to urban management, design, and sustainable development. Various deep learning-based (DL-based) methods, which achieved remarkable results in an end-to-end supervised process, were proposed for UFZ mapping. However, the excellent performance of DL-based models relies heavily on a large number of well-annotated samples, which is impossible to obtain in practical UFZ mapping scenarios. Obtaining these well-annotated samples requires a lot of manual costs, which greatly limits the outcome of these methods in practical UFZ mapping tasks. In this article, we proposed a UFZ mapping method using OpenStreetMap-based (OSM-based) sample generation and the bi-branch neural network (BibNet). By adopting the idea of OSM-based sample generation, the proposed method utilized large-scale crowdsourcing labeled data (source domain) in OSM to generate a UFZ dataset (target domain) from OSM using remote sensing and social sensing data. Considering the inconsistent response of UFZ to various data observations, it is difficult to fully reflect the characteristics of UFZs using only remote sensing or social sensing data. We further proposed the BibNet, which utilizes two different deep neural network branches to comprehensively harness remote sensing images and social sensing data to map the UFZ. Experiments were conducted in Shenzhen City and Hong Kong City (Yau Tsim Mong District, Sham Shui Po District and Kowloon City District). The proposed method achieved an overall accuracy (OA) of 94.46% in the testing set of Shenzhen City and OA of 91.90% in the testing set of Hong Kong City.