Semantic enhancement and change consistency network for semantic change detection in remote sensing images

Semantic Change Detection (SCD) identifies binary change information and determines the ‘from-to’ types, revealing not only ‘where’ changes occurred but also ‘what’ the changes are. It has become a popular research topic in remote sensing, with many models proposed, yet challenges remain. Semantic e...

Full description

Saved in:
Bibliographic Details
Main Authors: Zhenghao Jiang, Biao Wang, Peng Zhang, Yanlan Wu, Zhiyuan Ye, Hui Yang
Format: Article
Language:English
Published: Taylor & Francis Group 2025-08-01
Series:International Journal of Digital Earth
Subjects:
Online Access:https://www.tandfonline.com/doi/10.1080/17538947.2025.2496790
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Semantic Change Detection (SCD) identifies binary change information and determines the ‘from-to’ types, revealing not only ‘where’ changes occurred but also ‘what’ the changes are. It has become a popular research topic in remote sensing, with many models proposed, yet challenges remain. Semantic extraction remains insufficient, and inconsistencies in change features hinder SCD progress. Segment Anything Model 2 (SAM2) has proven effective in extracting global features. However, when processing remote sensing images, SAM2's performance declines. This is particularly evident in images containing diverse ground objects, where interclass similarities are pronounced and intraclass variations are substantial. The model struggles because it lacks semantic information from the local features of adjacent objects. This paper proposes a semantic enhancement and change consistency network(SCNet). By enhancing the model's ability to extract semantic information from complex land-cover classes, we incorporate multi-scale adaptive modules, leveraging the efficient zero-shot capability of SAM2. Specifically, we have designed a semantic alignment module (SA) to enhance change information consistency. This module continuously supervises segmentation and final detection results using interactive self-attention. SCNet achieves state-of-the-art performance on the SECOND, Hi-UCD-mini, and SJH-SCD (our custom) datasets, demonstrating an impressive balance between efficiency and accuracy. The code will be available at https://github.com/XiaoJ058/RS-SCD.
ISSN:1753-8947
1753-8955