Deep Learning and Multi Source 2D and 3D Geospatial Data for Urban Quality of Life Assessment

Urban Quality of Life (UQoL) assessment is essential for improving well-being and guiding urban planning. While most studies focus on developed countries using household surveys and satellite imagery, this study addresses the gaps pertaining to the quantification of UQoL at a very microscale by utilizing multi-modal 2D and 3D data, viz. elevation-stacked Unmanned Aerial Vehicle (UAV) imagery, Google Street View (GSV) imagery, and amenities information derived from crowdsourced OpenStreetMap (OSM). An Unsupervised Domain Adaptation (UDA) based Deep Learning (DL) pipeline is implemented to segment UAV imagery, extracting indicators like built-up and green cover from the segmentation maps. Additionally, DeepLabV3 is used to segment GSV imagery to compute the sky-view factor, while OSM is employed to extract the location information of amenities. The UDA-based ResiDualGAN, equipped with a convolutional resizer model and integrated with the OSA method, trained on the RGB-nDSM dataset, achieved an IoU of 60.48%. Principal Component Analysis (PCA) is applied to create a weighted UQoL index. The results reveal disparities in UQoL across the study area, with higher UQoL in green, amenity-rich regions, emphasizing the importance of utility access. Notably, informal settlements located near essential services exhibited high UQoL despite having limited green cover and higher built-up density, which clearly emphasizes the importance of proximity as a key indicator of UQoL.