Abstract
Remote sensing semantic segmentation is essential for geoscientific analysis. Recently, although multimodal image fusion has achieved notable improvements over single-image approaches, existing methods still suffer from insufficient utilization of the observed information, especially the alignment and fusion of multimodal details information. To address above issues, a novel adaptive frequency-domain fusion network (AFF-Net) is proposed in this letter to achieve effective fusion of visible images and DSM data through a Frequency-Domain Fusion (FDF) module, and enhance the recognition capability of target boundaries via a Semantic-Guided Detail Alignment (SDA) strategy. Specifically, the FDF module conducts an adaptive frequency-wise alignment and fusion scheme between visible images and DSM by decomposing them into low and high frequency subbands, enabling the model to capture more fine-grained complementary information across various frequency domains, rather than coarse-grained fusion in the spatial domain. The SDA strategy further restores high-frequency boundary details while suppressing noise, enabling precise object delineation. AFF-Net was evaluated on the ISPRS Vaihingen and Potsdam datasets, achieving mIoU scores of 84.87%, 87.11%, as well as mF1 scores of 91.61% and 92.93%, respectively. These results show better performances over state-of-the-art methods and verify the effectiveness and superiority of the proposed method. The code is available at https://github.com/lanmao-create/AFF-Net.
Funding Information
  • Program for University Key Laboratory of Guangdong Province (26IRTSTHN003)
  • State Key Laboratory of Drug Research, Chinese Academy of Sciences (SKLSD2025-HNZZ-01)

This publication has 20 references indexed in Scilit: