TransY-Net: Learning Fully Transformer Networks for Change Detection of Remote Sensing Images

Tianyu Yan, Zifu Wan, Pingping Zhang, Gong Cheng, Huchuan Lu

Abstract

In the remote sensing field, Change Detection (CD) aims to identify and localize the changed regions from dual-phase images over the same places. Recently, it has achieved great progress with the advances of deep learning. However, current methods generally deliver incomplete CD regions and irregular CD boundaries due to the limited representation ability of the extracted visual features. To relieve these issues, in this work we propose a novel Transformer-based learning framework named TransY-Net for remote sensing image CD, which improves the feature extraction from a global view and combines multi-level visual features in a pyramid manner. More specifically, the proposed framework first utilizes the advantages of Transformers in long-range dependency modeling. It can help to learn more discriminative global-level features and obtain complete CD regions. Then, we introduce a novel pyramid structure to aggregate multi-level visual features from Transformers for feature enhancement. The pyramid structure grafted with a Progressive Attention Module (PAM) can improve the feature representation ability with additional inter-dependencies through spatial and channel attentions. Finally, to better train the whole framework, we utilize the deeply-supervised learning with multiple boundary-aware loss functions. Extensive experiments demonstrate that our proposed method achieves a new state-of-the-art performance on four optical and two SAR image CD benchmarks.

Introduction

Change Detection (CD) is a cornerstone of remote sensing, used for applications ranging from urban planning to disaster monitoring. The goal is to take two images of the same location from different times and highlight significant changes. However, traditional deep learning methods—primarily based on Convolutional Neural Networks (CNNs)—often suffer from two major flaws: incomplete detection regions ("holes") and irregular boundaries. These issues arise because CNNs focus on local textures and often fail to recognize large objects as single units, especially in complex scenes with varying light and shadows.

Methodology: The TransY-Net Architecture

To overcome these limitations, we propose TransY-Net, a fully Transformer-based "Y-shaped" network designed for high-precision change detection.

Overview of the TransY-Net architecture, featuring Siamese Feature Extraction and Progressive Change Prediction.

Siamese Feature Extraction (SFE)

Instead of standard convolutions, TransY-Net utilizes weight-shared Swin Transformers as its backbone. Transformers excel at modeling long-range dependencies, allowing the model to consider the entire image context when deciding if a pixel has changed. This global perspective is crucial for eliminating the "holes" typically found in CNN-based detection maps.

Progressive Change Prediction (PCP)

TransY-Net employs a decoder pyramid to integrate multi-scale features. Unlike models that only subtract features, TransY-Net performs both summation (to highlight commonalities) and difference (to highlight changes). A key innovation is the Progressive Attention Module (PAM), which applies both spatial and channel-level attention to refine the boundaries of detected regions, ensuring sharp and accurate edges.

Hybrid Supervision

The model is trained using a combination of Weighted Binary Cross-Entropy, SSIM, and soft IoU losses. This hybrid approach, combined with deep supervision across multiple side-outputs, ensures the network learns both global overlap and fine structural details.

Experimental Results

TransY-Net was evaluated on six diverse benchmarks, including four optical datasets (LEVIR-CD, WHU-CD, SYSU-CD, Google-CD) and two Synthetic Aperture Radar (SAR) datasets (Ottawa, Sulzberger).

ROC curve comparison on a change detection benchmark, showing TransY-Net (Ours) achieving the highest area under the curve.

Quantitative Performance

The model achieved state-of-the-art (SOTA) performance across the board. On the LEVIR-CD dataset, it reached an F1 score of 91.90%, outperforming previous Transformer-based models like BIT and ChangeFormer. On the WHU-CD dataset, it achieved a staggering 93.38% F1 score. Furthermore, the model demonstrated exceptional robustness on noisy SAR imagery, successfully localizing changes in flood zones and ice shelves.

Qualitative Analysis

Visual results indicate that TransY-Net produces much cleaner, more rectangular building footprints compared to the irregular outputs of earlier models. By leveraging global context, the model effectively distinguishes "pseudo-changes" (like moving shadows) from actual structural shifts.

Conclusion

TransY-Net addresses the critical "holes and edges" problem in remote sensing change detection. By combining the global modeling power of Swin Transformers with a progressive attention-based pyramid, it sets new benchmarks for accuracy and visual quality. While the model's computational complexity and data requirements remain areas for future optimization, its performance establishes it as a leading framework for high-fidelity remote sensing analysis.

Summary updated

Abstract

In the remote sensing field, Change Detection (CD) aims to identify and localize the changed regions from dual-phase images over the same places. Recently, it has achieved great progress with the advances of deep learning. However, current methods generally deliver incomplete CD regions and irregular CD boundaries due to the limited representation ability of the extracted visual features. To relieve these issues, in this work we propose a novel Transformer-based learning framework named TransY-Net for remote sensing image CD, which improves the feature extraction from a global view and combines multi-level visual features in a pyramid manner. More specifically, the proposed framework first utilizes the advantages of Transformers in long-range dependency modeling. It can help to learn more discriminative global-level features and obtain complete CD regions. Then, we introduce a novel pyramid structure to aggregate multi-level visual features from Transformers for feature enhancement. The pyramid structure grafted with a Progressive Attention Module (PAM) can improve the feature representation ability with additional inter-dependencies through spatial and channel attentions. Finally, to better train the whole framework, we utilize the deeply-supervised learning with multiple boundary-aware loss functions. Extensive experiments demonstrate that our proposed method achieves a new state-of-the-art performance on four optical and two SAR image CD benchmarks. The source code is released at https://github.com/Drchip61/TransYNet.
BibTeX
@article{yan2023transy,
  author = {Tianyu Yan and Zifu Wan and Pingping Zhang and Gong Cheng and Huchuan Lu},
  title = {{TransY-Net: Learning Fully Transformer Networks for Change Detection of Remote Sensing Images}},
  journal = {IEEE Trans. Geosci. Remote. Sens.},
  year = {2023},
  doi = {10.1109/tgrs.2023.3327253},
}
Figure