Detecting Human-Scale Built Environment Changes With Multimodal Large Language Models

Source abstract

Problem, research strategy, and findings Detecting built environment change at the human scale can help planners monitor neighborhood dynamics, target interventions, and provide evidence to support the evaluation of policies and design programs. The growing availability of time-series street-view imagery has made such change detection increasingly feasible. However, existing methods often provide only coarse indications of change rather than precise identification of what has changed and might lack the flexibility to capture diverse forms of built environment transformation, limiting their applicability in planning research and practice. In this study we addressed these limitations by examining the use of ChatGPT to detect fine-grained built environment changes in Guangzhou (China), evaluating its performance, and demonstrating its applications in planning contexts. The results show that ChatGPT identifies a broader range of changes than trained human coders, achieving higher recall (0.813–0.862) but lower precision (0.622–0.792), while demonstrating broadly comparable overall performance. We further applied the approach to identify crosswalk improvements, pedestrian facility enhancements, tree canopy expansion, and building façade renewal, with ChatGPT demonstrating relatively high performance across these applications (F1 scores: 0.810–0.851).Takeaway for practice This study shows how multimodal large language models (MLLMs) can support planning practice by leveraging time-series street-view imagery to detect fine-grained changes in the built environment. By developing and evaluating a prompt-chaining framework, this study has clarified the capabilities and limitations of MLLMs for planning applications. Importantly, planners should treat MLLM outputs as supporting evidence rather than ground truth. Category-guided open-ended prompts can reveal changes beyond predefined event classes, and prompt chaining can decompose complex tasks into transparent and adjustable steps. When applied carefully, MLLMs enable planners to identify and analyze built environment changes at a level of detail that was previously difficult to achieve at scale.