An Uncertainty-Audited GeoAI Framework for Sustainable Historic-District Micro-Renewal: Street-View Perception and Network-Based Field-Audit Prioritization

Sustainable regeneration of historic districts requires small-scale interventions to be targeted carefully because conservation constraints, limited resources, and heterogeneous street conditions make blanket renewal inappropriate. We develop an uncertainty-audited GeoAI workflow that supports field-audit prioritization for conservation-sensitive micro-renewal. In a Nanluoguxiang, Beijing case study, 2144 directional street-view images represented 536 panoramas, of which 535 entered final perception analysis. Semantic segmentation and repeated vision-language-model scoring produced streetscape indicators and AI-estimated visual-impression constructs. On a blinded 20-panorama subset, ten independent raters showed excellent average-measures reliability for perceived vitality and oppressiveness (ICC(2,10) = 0.967 and 0.954), and human mean ratings were strongly associated with Qwen scores (Spearman rho = 0.951 and 0.921; both p < 0.001). Leakage-resistant, spatially nested validation explained 30.0% of held-out vitality variance and 40.0% of held-out oppressiveness variance. A 48-specification stochastic priority analysis yielded a median top-15 overlap of 13/15 with the reference specification. Joint propagation of measurement, segmentation-architecture, spatial/model, panorama-to-edge, and decision-weight uncertainty identified E0079, E0076, and E0129 as the strongest inner-core candidates, with conservative conditional selection frequencies of 0.608, 0.605, and 0.577. These frequencies are algorithmic acceptability measures under tested assumptions, not calibrated probabilities, causal effects, or construction recommendations. A separate local Gemma 3 12B check on the same 20 panels showed moderate positive vitality agreement with human means (Spearman rho = 0.752) and with Qwen scores (rho = 0.752), as well as positive but weaker oppressiveness agreement with human means (rho = 0.671) and Qwen scores (rho = 0.654). Gemma underestimated vitality by 0.700 points on average but had near-zero mean oppressiveness bias (+0.010), so cross-model validity remains construct-specific rather than established globally. The workflow supports transparent resource targeting and follow-up field investigation while preserving the need for stakeholder consultation, engineering assessment, cross-model validation, and cross-district replication. Here, conservation sensitivity is implemented as a constraint on subsequent field verification and intervention appraisal; heritage significance is not directly estimated by the ranking model.