Automatic mapping of urban green spaces using a geospatial neural network

Detailed and precise urban green spaces (UGS) maps provide essential data for the sustainable urban development and related studies (e.g. heatwave events, heat related health risk, urban flooding, urban biodiversity and ecosystem services). However, remote sensing of mapping UGS is challenging due to the existence of mixed pixels and the cost and difficulty of collecting quality training data. This study presents a neural network-based automatic mapping method of UGS that integrates the use of Sentinel-2A satellite images and crowdsourced geospatial big data. The proposed neural network consists of three parts: (i) a multi-scale feature extraction module; (ii) a multi-modal information fuse module; and (iii) and a boundary enhancement module. The results showed that the proposed method achieved a high overall classification accuracy of 94.6%, which presents a clear UGS structure of a large scale. This study provides a fresh insight into how remote-sensing and crowdsourced geospatial big data can be integrated to improve urban mapping of green spaces through neural network.