A Digital Lens on Urban Biodiversity: Exploring Urban Nature Perceptions with Large Language Models

Urban landscapes are increasingly recognized as complex socio-ecological systems where biodiversity, built form, and human experience intersect. Beyond the intricate interconnections among biodiversity and provisioning and regulating ecosystem functions, these environments deliver Cultural Ecosystem Services (CES) such as aesthetic enjoyment, recreation, social relationships, and sense of place. Yet these intangible and experiential benefits remain underrepresented in biodiversity research, largely due to methodological constraints in capturing and interpreting public perceptions at scale. Recent advances in digital data availability, particularly user-generated content on social media, combined with the rapid evolution of large language models (LLMs) now provide promising avenues for uncovering how urban biodiversity and landscape features are encountered, interpreted, and valued in everyday life. By revealing how individuals articulate value, meaning, and emotional connection to urban nature in situ, such approaches help illuminate the perceptual and experiential pathways through which urban environments foster pro-environmental intentions, stewardship attitudes, and forms of place-based ecological care. 

This study examines the potential of LLMs to analyze public perceptions of urban nature using textual and visual content from the Flickr platform. Employing a prompt-based, zero-shot learning framework, several LLM architectures were adapted to identify nature-related elements and CES associated with urban landscapes. Model outputs were systematically compared across prompt formulations and validated against classifications made by human researchers to assess accuracy, interpretive depth, and sensitivity to linguistic variation. The findings demonstrate that LLMs can effectively detect subtle, context-dependent expressions of how people perceive and describe urban nature, capturing collective patterns of appreciation and engagement that reflect cultural, aesthetic, and ecological dimensions of urban environments. However, model performance varied across prompt structures and model types, underscoring the need for methodological transparency, careful prompt engineering, and iterative model refinement to ensure reliable outcomes in AI-assisted environmental research. 

By integrating LLMs, social-media analytics, and landscape research, this study presents a scalable and replicable approach for interpreting the human dimensions of biodiversity in cities. It advances digital methodologies for assessing the social and cultural foundations of urban ecosystems and offers insights that support inclusive, evidence-based urban planning aligned with the goals of the Global Biodiversity Framework (Target 12).