Abstract

Visual landscape assessment informs highway corridor planning decisions, yet conventional surveys scale poorly with corridor length. Multimodal large language models (MLLMs) offer a scalable alternative, but their alignment with road-user preferences remains poorly understood. This study aimed to quantify the extent to which MLLM-derived visual preference rankings align with those of road users. The comparison used 80 images across 16 landscape character groups along 418 km of Malaysia’s North–South Expressway. Five MLLMs (ChatGPT, Claude, Gemini, Kimi, and Qwen) were queried under three prompt formulations using complete pairwise comparison with AB/BA reversal, yielding 94,800 judgements. Bradley–Terry rankings were then compared against rating-scale responses from 400 road users. The five models converged strongly (Kendall’s W = 0.926), whereas baseline AI–human agreement was moderate (image-level ρ = 0.622; group-level ρ = 0.697). Divergences are concentrated in two opposing categories. Paddy landscapes, ranked first by humans, fell to thirteenth in the AI ranking, whereas advertisement-dominated scenes were overvalued. Excluding the paddy group raised correlations to 0.772 and 0.911. A theory-directed prompt achieved comparable gains (ρ = 0.775 and 0.929) and restored paddy to third rank. A hybrid AI-screening, human-targeted protocol is proposed for corridor-scale visual planning.

IPC Classification

Keywords