Park-NET: identifying Urban parks using multi-source spatial data and Geo-AI

Abstract

Public urban green spaces (PUGSs) are a vital part of cities. They have a positive influence on the ecosystem and the well-being of local citizens. This is widely recognizable now with the United Nations (UN) mentioning public urban green spaces accessibility in their Sustainable Development Goals (SDG). To make the most of PUGSs, create them, and manage them sustainably we need to understand them and their spatial and temporal characteristics better. There is also a need for more accessible and cheaper PUGSs datasets so that officials and locals can use that in designing more sustainable neighbourhoods and cities. While spatial PUGSs data for some cities are available in land use dataset, these are often inaccessible to outside researchers, and not updated frequently. Hence there are methodological limitations on how PUGS can be created more efficiently for a wider usage. Deep learning methods with the usage of satellite images are a great way to achieve that because they can capture geometric patterns of land cover types using freely available satellite images of moderate spatial and temporal resolution. This study evaluates two convolutional network architectures, U-Net and U-Net with a ResNet-34 encoder, for semantic segmentation of PUGSs at a metropolitan level from multiple cities worldwide. Open-source data, mainly Sentinel and ground truth data about PUGSs from various open data portals are used as an input data. The chosen best model had an average test Intersection over Union (IoU) of 0,5610, and average test F1 score of 0,64515 across two external cities, which is a moderate to good performance of the deep learning model in detecting PUGSs. It shows that this approach is promising to create reliable, new PUGSs datasets that can help officials in urban planning.