UrbanMMCL: Urban region representations via multi-modal and multi-graph self-supervised contrastive learning

Urban region representation learning has emerged as a fundamental approach for diverse urban analytics tasks, where each neighborhood is encoded as a dense embedding vector for effective downstream applications. However, existing approaches suffer from insufficient multi-modal alignment and inadequate spatial relationship modeling, limiting their representation quality and generalizability. To address these challenges, we propose UrbanMMCL, a novel self-supervised framework that integrates multi-modal multi-view contrastive pre-training with unified fine-tuning for comprehensive urban representation learning. UrbanMMCL employs a dual-stage architecture. First, cross-modal contrastive learning aligns diverse data modalities including remote sensing imagery, street view imagery, location encodings, and Vision–Language Model (VLM)-generated textual descriptions. Second, multi-view adaptive graph contrastive learning captures complex spatial relationships across human mobility, functional similarity, and geographic distance perspectives. The framework then fine-tunes all parameters with the learned representations for effective adaptation to downstream tasks. Comprehensive experiments demonstrate that UrbanMMCL consistently outperforms state-of-the-art methods across pollutant emission prediction, population density estimation, and land use classification with minimal fine-tuning requirements, thereby advancing foundation model development for diverse Geo-AI applications.