UrbanMoE: A Sparse Multi-Modal Mixture-of-Experts Framework for Multi-Task Urban Region Profiling

Urban region profiling, the task of characterizing geographical areas, is crucial for urban planning and resource allocation. However, existing research in this domain faces two significant limitations. First, most methods are confined to single-task prediction, failing to capture the interconnected, multi-faceted nature of urban environments where numerous indicators are deeply correlated. Second, the field lacks a standardized experimental benchmark, which severely impedes fair comparison and reproducible progress. To address these challenges, we first establish a comprehensive benchmark for multi-task urban region profiling, featuring multi-modal features and a diverse set of strong baselines to ensure a fair and rigorous evaluation environment. Concurrently, we propose UrbanMoE, the first sparse multi-modal, multi-expert framework specifically architected to solve the multi-task challenge. Leveraging a sparse Mixture-of-Experts architecture, it dynamically routes multi-modal features to specialized sub-networks, enabling the simultaneous prediction of diverse urban indicators. We conduct extensive experiments on three real-world datasets within our benchmark, where UrbanMoE consistently demonstrates superior performance over all baselines. Further in-depth analysis validates the efficacy and efficiency of our approach, setting a new state-of-the-art and providing the community with a valuable tool for future research in urban analytics

💡 Research Summary

Urban region profiling aims to characterize fine‑grained attributes of city districts such as carbon emissions, population density, and nighttime light intensity. Existing work largely treats each indicator as an isolated prediction problem and relies on a single data modality, which limits the ability to capture the complex inter‑dependencies among urban metrics. Moreover, the field lacks a publicly available, well‑aligned multi‑modal benchmark, making reproducible comparison difficult.

This paper addresses both gaps. First, the authors construct a comprehensive benchmark covering three diverse cities. For every region they collect (i) high‑resolution satellite imagery, (ii) POI statistics (counts per category), and (iii) a textual summary of POI information generated by a large language model. The three modalities are aligned and paired with three target tasks: carbon emission, population, and night‑light intensity. A set of strong baselines—including single‑modal CNNs, multimodal Transformers, and existing Mixture‑of‑Experts (MoE) models—is provided to enable fair evaluation.

Second, the paper proposes UrbanMoE, a sparse multi‑modal Mixture‑of‑Experts architecture specifically designed for multi‑task urban profiling. The pipeline consists of three stages.

1. Multimodal Feature Extraction – Satellite images and textual summaries are encoded with RemoteCLIP, yielding visual embeddings e_i and textual embeddings e_t. Two learnable context vectors are added: a region‑specific embedding r and a POI‑distribution embedding p. Concatenating these yields two modality‑specific inputs z_i =