Mapping Urban Mixed Land Use via Multimodal Fusion and Large Language Models

Abstract

The precise delineation of urban land use, particularly mixed land use, is essential for sustainable urban planning and resource allocation. However, traditional pixel-based mapping methods face a profound "semantic gap" and struggle to decode the complex, overlapping functional dynamics of modern cities. This study proposes a novel object-based GeoAI framework that integrates morphological segmentation, multimodal feature fusion, and Large Language Model (LLM) reasoning to identify urban mixed functions at the parcel level. Using the core districts of Beijing and Changsha as study areas. We first generat high-fidelity, seamless urban parcels by applying dynamic semantic buffering and topological repair to crowdsourced road networks. To capture deep socio-economic semantics, we utilize a contrastive learning architecture to align spatial locations with textual Point of Interest (POI) embeddings, fusing them with multispectral imagery, 3D building morphology, and space syntax metrics. Following Graph Neural Network (GNN) spatial contextual smoothing and HDBSCAN unsupervised clustering, we develop a zero-shot Chain-of-Thought (CoT) inference mechanism using an LLM. This approach accurately deduces primary land use categories and quantifies mixed land use proportions based on soft-clustering probabilities. Our results demonstrate exceptional capability in identifying intricate micro-level spatial patterns, such as transit-oriented developments and commercial-residential integrations. Validation against the 2022 EULUC-China 2.0 baseline yielded area-weighted Overall Accuracies of 70.91% for Beijing and 76.90% for Changsha. In-depth analysis reveals that this statistical attenuation is not a model failure, but rather the manifestation of a systematic mismatch in time—capturing rapid urban renewal between 2022 and 2024—and ontological dissonance between classification taxonomies. By transcending deterministic single-label paradigms, this framework provides a highly authentic, interpretable representation of urban complexity, establishing a new methodological framework for LLM integration in Geospatial AI (GeoAI).