Change Guiding Network: Incorporating Change Prior to Guide Change Detection in Remote Sensing Imagery

Chengxi Han, Chen Wu, Haonan Guo, Meiqi Hu, Jiepan Li, Hongruixuan Chen

TL;DR

This paper introduces CGNet, a novel deep learning architecture for remote sensing change detection that addresses the common issues of inaccurate edge detection and internal holes in change maps. By leveraging a "change guiding map" derived from deep semantic features, CGNet effectively guides multi-scale feature fusion and utilizes a self-attention module (CGM) to capture long-range dependencies, leading to improved detection accuracy.

What Problem Does This Paper Address?

Change detection (CD) in remote sensing imagery is crucial for monitoring urban development, land use, and environmental changes. While deep learning has significantly advanced CD, existing methods, particularly those based on the U-Net architecture, often struggle with precisely delineating the boundaries of changed areas and can produce "holes" within detected change regions. This is attributed to the limited receptive field of traditional convolutional layers and insufficient expression of change features. The paper aims to overcome these limitations by incorporating prior knowledge of change to guide the feature extraction and fusion process.

Methods

The core of the proposed solution is the Change Guiding Network (CGNet), which builds upon the U-Net architecture. The key innovations are:

  • Change Guiding Map (CGM): Deep semantic features are used to generate a "change guiding map." This map acts as prior information, guiding the fusion of multi-scale features. The idea is that by understanding where changes are likely to occur at a semantic level, the network can better focus its feature extraction.
  • Change Guide Module (CGM): A novel self-attention module designed to enhance the network's ability to capture long-distance dependencies between pixels. This addresses the limited receptive field of standard convolutional layers, which is a major cause of inaccurate edges and internal holes. The CGM uses a "change guiding map" to focus its attention on relevant change information.
  • Hierarchical Feature Fusion: The network employs a hierarchical approach, integrating features from different levels of a VGG-16 backbone encoder. This allows for the capture of both coarse and fine-grained change information.

Datasets

The proposed CGNet was evaluated on four publicly available, large-scale remote sensing change detection datasets:

  • LEVIR-CD [24]: High-resolution (0.5m/pixel) building change detection dataset.
  • WHU-CD [37]: Very high-resolution (0.075m/pixel) building change detection dataset.
  • SYSU-CD [38]: A challenging dataset with various types of modifications.
  • S2Looking-CD [39]: Features side-looking satellite images with different off-nadir angles, presenting challenges like small targets and dense arrangements.

The paper also mentions the use of standard data splits for training, validation, and testing, with images cropped to 256x256 for fair comparison.

Key Insights and Ideas

  • Prior knowledge is powerful: The paper demonstrates that explicitly incorporating a "change guiding map" derived from deep semantic features significantly improves the precision of change detection, especially at object boundaries.
  • Self-attention for context: The CGM module effectively leverages self-attention to overcome the inherent limitations of convolutional receptive fields, leading to better handling of long-range spatial dependencies crucial for accurate change localization.
  • Addressing edge and hole issues: The combination of the change guiding map and the CGM module directly tackles the persistent problems of inaccurate edge detection and internal holes in change maps, which are common in previous methods.
  • Hierarchical approach benefits: Integrating features from multiple scales and using them in a guided fusion process enhances the overall representation of change features.

Key Results

CGNet consistently achieved state-of-the-art performance across the four evaluated datasets, outperforming 13 other methods.

  • Quantitative Performance: CGNet achieved the best F1-score, Precision, Recall, and IoU metrics on the WHU-CD dataset. It also secured first place in F1 and IoU on LEVIR-CD and SYSU-CD, and first place in F1 and IoU on S2Looking-CD, with strong performance in other metrics as well.
  • Qualitative Results: Visualizations (Figures 4-8) clearly show CGNet's ability to produce cleaner change maps with more accurate edges and fewer internal holes compared to other methods. The ablation studies further confirm the effectiveness of the CGM modules.
  • Efficiency: While CGNet is more complex than some pure CNN methods, it offers a good balance between accuracy and training/inference time, outperforming many attention-based and transformer-based methods in terms of efficiency.

Limitations and Future Work

  • Computational Cost: While CGNet offers improved efficiency compared to some complex models, it is still a deep learning model with associated computational demands. The paper notes that transformer-based methods can be large in parameter count and require significant computing resources.
  • Future Exploration: The authors suggest that the concept of using deep change information to guide multi-scale feature fusion is a promising avenue for future research and could be explored in other state-of-the-art CD methods. The paper also mentions the intention to open-source their code to facilitate further research in the field.

Summary updated

Abstract

The rapid advancement of automated artificial intelligence algorithms and remote sensing instruments has benefited change detection (CD) tasks. However, there is still a lot of space to study for precise detection, especially the edge integrity and internal holes phenomenon of change features. In order to solve these problems, we design the Change Guiding Network (CGNet), to tackle the insufficient expression problem of change features in the conventional U-Net structure adopted in previous methods, which causes inaccurate edge detection and internal holes. Change maps from deep features with rich semantic information are generated and used as prior information to guide multi-scale feature fusion, which can improve the expression ability of change features. Meanwhile, we propose a self-attention module named Change Guide Module (CGM), which can effectively capture the long-distance dependency among pixels and effectively overcomes the problem of the insufficient receptive field of traditional convolutional neural networks. On four major CD datasets, we verify the usefulness and efficiency of the CGNet, and a large number of experiments and ablation studies demonstrate the effectiveness of CGNet. We're going to open-source our code at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/ChengxiHAN/CGNet-CD</uri> .
BibTeX
@article{han2023change,
  author = {Chengxi Han and Chen Wu and Haonan Guo and Meiqi Hu and Jiepan Li and Hongruixuan Chen},
  title = {{Change Guiding Network: Incorporating Change Prior to Guide Change Detection in Remote Sensing Imagery}},
  journal = {IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing},
  year = {2023},
  doi = {10.1109/jstars.2023.3310208},
}
Figure