Abstract: [Purposes] To address the limitations of traditional urban functional zone (UFZ)identification methods，such as insufficient multi-source data fusion and limited feature representation under complex terrain conditions. [Methods] This study proposes a deep learning-based multimodal fusion framework.The model extracts heterogeneous features from Sentinel-2 imagery，POI data，and DEM information.An auxiliary weighting layer (AWL)is designed to adaptively evaluate the contribution of each auxiliary data source.A dual-branch attention module (DBAM)dynamically balances the influence of remote sensing and auxiliary data through a cross-modal attention mechanism.Additionally，a multi-modal feature fusion module (MFFM)is constructed by integrating depthwise separable convolutions with Transformer encoders to enhance cross-modal semantic representation and improve classification of urban areas. [Findings] Experiments conducted in Nan'an district，Chongqing，using 2019 labeled samples across nine UFZ categories demonstrate that the proposed method achieves an overall accuracy of 88.02% and a Kappa coefficient of 0.847. [Conclusions] These results confirm the effectiveness of the framework in UFZ classification for complex urban environments．

Key words: multi-modal data, urban functional zones, feature extraction, deep learning, urban planning