Mapping Public Green Space Using Sentinel-2 Imagery and Convolutional Neural Network at a Global Scale

Abstract

Urban green spaces (UGSs) are significant to the urban ecosystem and can have a positive influence on human health, both physically and mentally. Traditional methods of mapping UGS are mainly field surveys, however, the development of (very) high-resolution satellite images, together with machine learning and deep learning techniques has provided a faster way of extracting UGS at a large scale. This paper employed convolutional neural networks (CNNs) to conduct semantic segmentation of UGS from Sentinel-2 imagery for multiple cities around the globe. The main architectures used were U-Net model from base level and U-Net model with ResNet-50 and VGG-16 backbones pretrained on ImageNet. Results show that U-Net with pretrained ResNet-50 backbone achieved the best performance, with the average OA, IoU, F-score, and AUC of 0.9695, 0.9188, 0.9572, and 0.9584. When testing our model on external cities, the metrics above became 0.8767, 0.5570, 0.6448, and 0.6534, respectively, showing a relatively good generalization ability of the model to predict UGSs where benchmark city data is not available, for instance, cities in the global south. Moreover, our model identified some UGSs which are not in the ground truth datasets because of the hard definition of UGS, such as golf courses and cemeteries. However, since these areas indeed provide similar functions and benefits as UGSs, we argue that our model might be a potentially effective way to get more complete datasets for UGS distributions.