Quantifying uncertainty in deep learning models for urban safety perception: An MC Dropout approach

Deep learning models have revolutionized the automated analysis of urban perception. However, these models typically provide point predictions without associated uncertainty estimates, limiting their application in critical urban planning and policy decisions. This research explores the implementation of Monte Carlo Dropout (MC Dropout) as a method for quantifying uncertainty in deep learning models for urban safety perception. Using ResNet-based convolutional neural networks, we evaluated this approach on both global urban datasets (PlacePulse 2.0) and a specific case study of Santiago, Chile, focusing specifically on the safety dimension of urban perception. Our findings reveal that MC Dropout yields per-image uncertainty estimates without requiring significant architectural modifications, alongside a modest improvement in predictive accuracy (from 66.81% to 67.64%). We show that these estimates vary systematically with urban characteristics, with distinct patterns observed between different urban districts, and that the MC Dropout spread is systematically narrower than the discrepancy between the stochastic and deterministic predictions. We are explicit about what the study does not establish: none of our analyses tests the uncertainty estimates against predictive error, so we do not claim that they are calibrated or that they rank predictions by reliability, and we specify in detail the risk–coverage protocol and pairwise proper scoring rule that such a validation would require. The potential implications for urban analytics—triage of cases for human verification, mapping the spatial distribution of predictive dispersion within cities, and improved interpretability of model decision-making—are presented as prospective applications conditional on that validation. The methodological framework established here can be adopted and extended for future studies and practical applications in urban environments.