UFRI-MMBench: A Benchmark for Multi-Granularity and Multi-Level Urban Functional Region Identification

Accurately identifying urban functional regions is vital for understanding urban structure, improving emergency response, and promoting sustainable urban development. However, achieving high precision identifying remains challenging due to the limited availability of annotated data. In this paper, we create a novel dataset named Urban Functional Region Identification (UFRI), which includes a substantial collection of very high resolution (VHR) remote sensing (RS) images, annotated with 6 coarse-grained and 4 fine-grained categories. In contrast to RS semantic segmentation datasets, UFRI provides more detailed labels and comprehensive resources for urban scene analysis. Based on UFRI dataset, we propose a Multi-granularity and Multi-level Benchmark (UFRI-MMBench) method for identifying urban functional regions motivated by encoder-decoder structure in semantic segmentation. It combines a transformer-based encoder and a Mamba-based decoder. Additionally, considering Mamba's sequential selective scanning for sequential data, it is difficult to capture the global context information that is not sequentially dependent on the image, resulting in performance limitations. Thus, we also design a Multi-level Feature to Sequence (MF2S) module to convert feature maps from encoder into sequential patches. Our framework facilitates the multi-level fusion of VHR images, leading to accurate identification of urban functional regions. Experimental results demonstrate the effectiveness of our approach, achieving 71.83% mF1 and 57.67% mIoU on coarse-grained dataset, along with 49.25% mF1 and 36.69% mIoU on fine-grained dataset, outperforming the latest approaches.