Mapping urban spatial perception using deep learning approaches

This study aims to demonstrate that urban perception can be quantitatively and spatially analyzed by integrating deep learning methods with Geographic Information Systems (GIS). Within this scope, a total of 4,520 street-level images collected from the city of Eskişehir were utilized to train three distinct Convolutional Neural Network (CNN) architectures—EfficientNet-B1, ResNet18, and VGG19—using a transfer learning approach. Predictive models were developed for six perceptual dimensions: liveliness, depressing, safer, beautiful, boring, and wealthier. The performance of these models was evaluated using regression metrics including MAE, RMSE, and R². Among the tested architectures, EfficientNet-B1 exhibited the highest predictive accuracy, particularly in themes associated with tangible visual cues such as safer, livelier, and beautiful. In contrast, all models demonstrated limited predictive performance in more abstract and subjective themes such as boring and depressing. The model outputs were visualized through heat maps, revealing spatial patterns that largely correspond with the known urban morphology of the study area. A significant contribution of this study is the provision of a digitized and mappable data infrastructure pertaining to urban perception. Nevertheless, the limited explanatory capacity of the models indicates that visual data alone may be insufficient to comprehensively capture the complexity of urban perception. Future research should explore the integration of multimodal data sources (e.g., audio, text, user comments, and location check-ins) and adopt explainable artificial intelligence (XAI) frameworks to enhance both the interpretability of the models and their applicability within urban planning contexts.