Related Experiment Video
Updated: May 17, 2025
03:31
Author Spotlight: Enhancement of Salient Object Detection for Smart Grid Applications
Published on: December 15, 2023
438
Semantic-Aware Remote Sensing Change Detection with Multi-Scale Cross-Attention
Xingjian Zheng1, Xin Lin2, Linbo Qing3
1College of Design and Engineering, National University of Singapore, Singapore 119077, Singapore.
Sensors (Basel, Switzerland)
|May 14, 2025
Summary
This study introduces a new deep learning model, the multi-scale cross-attention network (MSCANet), for remote sensing image change detection. MSCANet improves accuracy by better integrating spatial and semantic features across different scales, enhancing change detection in complex environments.
Area of Science:
- Remote Sensing
- Computer Vision
- Deep Learning
Background:
- Change detection in remote sensing is crucial for urban planning, disaster management, and land use analysis.
- Traditional methods struggle with accurately comprehending global and local features, leading to semantic inaccuracies.
- Pixel-level analysis in older techniques often overlooks critical semantic information.
Purpose of the Study:
- To propose a novel deep learning model, the multi-scale cross-attention network (MSCANet), for enhanced remote sensing image change detection.
- To address limitations in existing methods by improving the integration of multi-scale spatial and semantic features.
- To develop a more robust and accurate change detection solution for complex and noisy remote sensing data.
Main Methods:
- Implemented a multi-scale feature extraction strategy to capture and fuse information at various spatial resolutions.
- Introduced a cross-attention module to improve the model's comprehension of semantic-level changes between bitemporal images.
- Utilized Convolutional Neural Networks (CNNs) as the foundational architecture for the proposed MSCANet.
Main Results:
- MSCANet demonstrated competitive performance on public datasets (LEVIR-CD, CDD, SYSU-CD), achieving high F1-scores (e.g., 96.19% on CDD).
- The model achieved a high Intersection over Union (IoU) of 92.67% on the CDD dataset.
- Robustness tests confirmed the model's ability to maintain high accuracy even with input degradation, such as Gaussian noise.
Conclusions:
- The proposed MSCANet effectively integrates spatial and semantic features across multiple scales for more accurate and coherent change detection.
- The model exhibits improved semantic awareness and robustness, making it a promising solution for real-world remote sensing applications.
- MSCANet offers a significant advancement in change detection, particularly in challenging environments with noise and complexity.
