RWKV-CD: A Vision-RWKV Framework for Building Change Detection From Remotely Sensed Imagery
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Buildings serve as a key indicator of human land use, and building change detection plays a crucial role in urban growth and shrinkage monitoring, land use planning, and urban development assessment and management. Current mainstream change detection methods mainly rely on Convolutional Neural Networks (CNN), Transformer, and their variants. Although CNN excel at extracting local image features, their fixed and limited receptive fields hinder effective capture of long-range dependencies along complex building boundaries. Transformer, on the other hand, leverage self-attention mechanisms to effectively capture global information but suffer from quadratic computational complexity, leading to high demand of computing resources and limiting their practical efficiency. Meanwhile, since the spectra of edge pixels often represent a mixture of building and surrounding surfaces (e.g., roads or vegetation), existing methods typically overlook the variations in spectral similarity between the main body and its edges. To address this challenge, we leverage the Receptance Weighted Key Value (RWKV) architecture for its ability to capture and propagate global dependencies with linear complexity. Building upon this foundation, we propose the RWKV-CD, a novel change detection method that integrates RWKV architecture with specialized feature extraction and noise suppression strategies. RWKV-CD employs a dual-branch encoder separately capture global building information and fine-grained edge details, while a cross-attention module adaptively fuse these features. In addition, a noise removal module combining gating mechanism is introduced to reduce the spectral noise and irregular patterns. Comparative experiments on three publicly available datasets (e.g., LEVIR-CD, WHU-CD, and S2Looking) demonstrate that RWKV-CD significantly outperforms existing methods in terms of F1 score and Intersection over Union (IoU), underscoring the strong potential of the RWKV architecture for CD tasks.
Fields of Study & Subjects
Remote-Sensing Image Classification
Automated Road and Building Extraction
Advanced Neural Network Applications
Publication details
- Journal / source
- IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
- Volume
- 19
- Pages
- 13986–14002
- Publication date
- 1 January 2026
- Language
- EN