Luo, Zhewen2025-01-062025-01-062025-01-06http://hdl.handle.net/10393/50035https://doi.org/10.20381/ruor-30809This thesis investigates the degree to which advanced GeoAI techniques, particularly deep learning models, and how these models can reproduce older adults perceptions of urban green spaces and to map the quality of green spaces across the city using street-level imagery. The study compares three popular deep-learning feature extractors to predict and map seniors' perceptual responses to green spaces in Ottawa, utilizing data derived from Mapillary street-level images. In these images, senior citizen volunteers were seated in front of a computer and selected between images based on perceptual dimensions such as safety and aesthetics. A few AI ranking models are constructed for the purpose of identifying how well street-level photographs are valued by seniors within Ottawa’s green spaces. Our focus is mainly on finding the best-performing model and backbone training dataset via experimentation with various generations of deep learning architectures and using statistics to determine how such differences affect human perceptual modelling. Experimentation included the determination of the best ranking loss function for our data, comparing different loss functions, effects of hidden layers and how different pre-trained weights from different domains affect model performance. The RankNet algorithm demonstrated superior performance compared to RankSVM in a range of architectural configurations, including ViT-B16, VGG16, and ResNet50, with an average difference of 8.45% in accuracy for each perceptual attribute. The model deemed most effective overall adopted ViT-B16 as its foundation, providing the highest absolute accuracy despite the accuracy not being statistically different from VGG19 or ResNet50. Contrary to previous research we did not find the addition of Softmax to be of any significance in model preformance. Finally, we found that there is no statistically significant differences in using different-object-domain training weights in the transfer learning. Detailed prediction maps were predicted on park photographs from Mapillary that indicated how seniors perceive public spaces across the city. By combining deep learning, image segmentation, and spatial analysis, the research offers an empirical contribution of how AI can contribute insights into how urban planning can be enhanced to meet the needs of an ageing population and creates a foundation for further investigation into age-friendly environments.enElderly Urban PerceptionOttawaMachine-learning ApproachThesis