Semantic-Aware Remote Sensing Change Detection with Multi-Scale Cross-Attention

Remote sensing image change detection plays a vital role in diverse real-world applications such as urban development monitoring, disaster assessment, and land use analysis. As deep learning strives, Convolutional Neural Networks (CNNs) have shown their effects in image processing applications. Ther...

Full description

Saved in:
Bibliographic Details
Main Authors: Xingjian Zheng, Xin Lin, Linbo Qing, Xianfeng Ou
Format: Article
Language:English
Published: MDPI AG 2025-04-01
Series:Sensors
Subjects:
Online Access:https://www.mdpi.com/1424-8220/25/9/2813
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Remote sensing image change detection plays a vital role in diverse real-world applications such as urban development monitoring, disaster assessment, and land use analysis. As deep learning strives, Convolutional Neural Networks (CNNs) have shown their effects in image processing applications. There are two problems in old-school change detection techniques: First, the techniques do not fully use the effective information of the global and local features, which causes their semantic comprehension to be less accurate. Second, old-school methods usually simply rely on differences and computation at the pixel level without giving enough attention to the information at the semantic level. To address these problems, we propose a multi-scale cross-attention network (MSCANet) based on a CNN in this paper. First, a multi-scale feature extraction strategy is employed to capture and fuse image information across different spatial resolutions. Second, a cross-attention module is introduced to enhance the model’s ability to comprehend semantic-level changes between bitemporal images. Compared to the existing methods, our approach better integrates spatial and semantic features across scales, leading to more accurate and coherent change detection. Experiments on three public datasets (LEVIR-CD, CDD, and SYSU-CD) demonstrate competitive performance. For example, the model achieves an F1-score of 96.19% and an IoU of 92.67% on the CDD dataset. Additionally, robustness tests with Gaussian noise show that the model maintains high accuracy under input degradation, highlighting its potential for real-world applications. These findings suggest that our MSCANet effectively improves semantic awareness and robustness, offering a promising solution for change detection in complex and noisy remote sensing environments.
ISSN:1424-8220