Urban Heat-Aware Walkability Index Based on LLMs Analysis: A Case Study of Sofia

Abstract

A novel Artificial Intelligence (AI)-driven urban heat-aware walkability index (UHAWI) is proposed to assess pedestrian wellbeing and heat exposure. The approach integrates a conventional walkability index with the Wet Bulb Globe Temperature (WBGT) to assess how thermal stress affects the suitability of urban streets and areas for walking. AI methods are used to analyse spatial patterns and identify thermally comfortable and low-risk pedestrian zones. A Large Language Model (LLM) is used to analyse the calculated index and suggest an intervention. The pipeline is applied to street segments in Sofia, Bulgaria, across 5 consecutive July dates, producing UHAWI scores ranging from 23.1 to 66.5 (mean 43.9), which are then normalised for visualisation, indicating that summer thermal stress binds pedestrian conditions city-wide. Street segments with critical heat-walkability deficits are identified and targeted interventions are suggested, each grouped by empirically documented cooling and walkability effects, with street tree planting yielding the largest mean improvement (+13.3 points). Rubric-based evaluation of the LLM recommendations by a human (overall mean 3.25/5) and an independent LLM-as-judge (2.75/5) found strong spatial coherence and actionability but limited numerical faithfulness, identifying output grounding as the principal limitation of small locally deployed models. The proposed approach offers municipalities a reproducible, privacy-preserving decision support tool for prioritising climate adaptation investment at the street level, and the compute-then-reason architecture generalises to other urban analytics domains where AI assistance must remain locally executable and fully auditable.