Source-linked AI summary
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
Precise remote-sensing change detection remains limited by inaccurate edges and internal holes associated with insufficient change-feature expression in conventional U-Net structures. The paper proposes CGNet, which uses deep semantic change maps to guide multiscale fusion and introduces the self-attention CGM for long-distance pixel dependencies. On four change-detection datasets, experiments and ablations demonstrate CGNet’s usefulness and effectiveness.
Problem
Conventional U-Net-based change-detection methods have insufficient change-feature expression, producing inaccurate edge detection and internal holes.
Method
CGNet generates deep semantic change maps as prior information for multiscale fusion and uses the self-attention CGM to capture long-distance pixel dependencies.
Results
Experiments and ablation studies on four change-detection datasets demonstrate the usefulness, effectiveness, and practical value of CGNet and CGM.
Takeaways & Limitations
CGNet provides a change-guided architecture for improving change-feature expression, including edge integrity and internal-hole handling.
Abstract
from arXiv · showhide
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 overcome 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 https://github.com/ChengxiHAN/CGNet-CD.
I. INTRODUCTION
Remote-sensing change detection has advanced through machine learning, deep learning, CNNs, and transformer-based methods, but precise feature-detail extraction remains challenging. Existing approaches include threshold-based, classification-based, CNN-based, attention-enhanced, and transformer-based methods.
- Change detection identifies variations in conditions across remote-sensing images acquired at different periods.
- Traditional change-detection methods include change vector analysis, multivariate analysis, principal component analysis, slow feature analysis, and decision-function-based classification.
- Deep learning has been widely applied to remote-sensing imagery classification, semantic segmentation, object detection, and change detection.
- CNN-based change-detection models use convolutional feature extraction, with representative designs including early fusion, Siamese concatenation, and Siamese difference networks.
- Prior change-detection studies use attention mechanisms, deep supervision, recurrent models, multiscale Siamese networks, and nested connections to improve detection performance.
- Existing methods still require improvement in extracting detailed change information, while HANet addresses sample imbalance and low model accuracy.
- Transformer-based approaches have also been introduced for change detection, including SwinSUNet, BIT, Change Former, RSP-BIT, and TransUNetCD.
T1 T2 GT
The paper motivates CGNet by identifying broken edges and internal holes in prior change maps, then proposes deep-feature guidance and self-attention to improve change-feature expression. Experiments and ablations on four change-detection datasets support the approach’s usefulness and practicability.
- Motivation: Previous methods produce edge breakage and internal holes in both feature maps and binary change maps.
- Motivation: These defects are associated with insufficient context extraction, small receptive fields, and inadequate multiscale fusion.
- CGNet: CGNet uses deep change-guiding features as prior knowledge to direct multiscale feature fusion.
- CGM: The Change Guide Module is a self-attention module designed to capture long-distance pixel dependencies and enlarge the effective receptive field.
- Results: CGNet with CGM is reported to address inaccurate edges and internal holes in change features.
- Evaluation: Experiments and ablation studies on four change-detection datasets evaluate the usefulness and practicability of the proposed approach.
II. CGNET
CGNet uses a hierarchical encoder–decoder architecture in which deep semantic change maps guide multiscale fusion. Its CGM self-attention modules model long-distance pixel dependencies to address conventional convolutional receptive-field limitations.
- A. Overall architecture: CGNet is presented as a hierarchical feature network built on a conventional U-Net architecture.
- A. Overall architecture: The network extracts bi-temporal image features from coarse to fine using VGG-16 as its encoder foundation.
- A. Overall architecture: Its decoder concatenates features from different depths and refines them with designed convolutional blocks.
- A. Overall architecture: Deep semantic features generate change maps that serve as prior information for guiding multiscale feature fusion.
- B. Change Guide Module: CGM is a self-attention module that compensates for long-distance pixel dependencies and insufficient receptive fields in traditional CNNs.
- B. Change Guide Module: Three CGMs are placed at different feature-extraction stages, guiding fusion from deep to shallow features.
B. Change Guide Module
The Change Guide Module (CGM) uses deep semantic features to generate change-guide maps that guide multi-scale fusion, while self-attention captures long-distance pixel dependencies. This design focuses feature expression on likely changes and broadens the receptive field for change detection.
- Attention weights encode the importance of change information, with larger weights assigning greater influence during attention calculation.The corresponding values represent the information being weighted.
- Self-attention captures internal pixel correlations and provides a larger receptive field than conventional convolutional processing.The module is introduced to address limited receptive-field coverage and capture long-distance dependencies.
- CGM generates change-guide maps from deep semantic feature maps to provide prior information for multi-scale feature fusion.The guide maps are intended to improve the expression ability of change features.
- The original self-attention formulation is computationally intensive, motivating the change-guided design described for CGM.The passage identifies computational cost as a limitation of the original formulation.
- CGM combines weighted input features, multi-head self-attention, and a convolution operation to produce its output feature map.The input feature map is weighted using the change guide map before attention and convolutional processing.
- The change guide map helps CGM focus attention on significant change information while disregarding abundant background pixels.This weighting strategy reflects the highly unbalanced nature of binary change detection, where changed pixels are relatively sparse.
C. Model Details
The model details describe cross-entropy training with deep supervision of the change-guiding map and evaluation on four remote-sensing change-detection datasets. The datasets cover building and land-use changes across varied imaging conditions and are assessed using precision, recall, IoU, and F1-related comparisons.
- Model details: The model uses cross-entropy loss for the highly unbalanced binary change-detection task.Deep supervision applies cross-entropy between the change-guiding map and ground truth, and this loss is added to the overall objective.
- Model details: The change-guiding map provides prior information to the decoder to enhance change-feature representation.This supervision and guidance connect the change map to decoder-side feature processing.
- Datasets: Experiments use LEVIR-CD, WHU-CD, SYSU-CD, and S2Looking-CD, covering building changes, earthquake-related rebuilding, varied land-use changes, and side-looking rural imagery.S2Looking-CD additionally includes off-nadir angles, lighting variation, and rural-scene complexity.
- Evaluation: Precision, recall, and IoU are used to statistically assess prediction maps against ground truth, with higher F1 and IoU indicating better change-detection performance.The evaluation compares predicted maps with ground-truth labels using TP, TN, FP, and FN totals.
B. Experimental environment
The experiments use PyTorch on an NVIDIA RTX 3090 GPU with AdamW optimization, fixed training settings, data augmentation, and validation-based model selection.
- Training setup: Each model is trained in PyTorch using one NVIDIA RTX 3090 GPU and the AdamW optimizer.The reported optimizer settings use weight decay 0.0025 and learning rate 0.0005.
- Data processing: Data augmentation includes random Gaussian noise, salt-and-pepper noise, and random cropping.These augmentations are part of the reported experimental preprocessing.
- Training setup: Training uses a batch size of eight and a maximum of fifty epochs, with the model having the best validation F1 and IoU saved for testing.The selected checkpoint is based on validation performance.
C. Comparison with state-of-the-art methods
CGNet is compared with a broad set of CNN-, attention-, and transformer-based change-detection methods across four remote-sensing datasets. It achieves leading quantitative results and produces more complete change regions with fewer false detections and internal holes.
- Across four datasets, CGNet performs best in F1 and OA.
- LEVIR-CD: CGNet achieves first place in F1 and IoU on LEVIR-CD, while HCGMNet is nearly second.
- WHU-CD: On WHU-CD, the hierarchical change-guidance framework ranks first in F1, Pre., Rec., and IoU.
- SYSU-CD: On challenging SYSU-CD, CGNet ranks first in F1 and IoU and second in Precision.
- S2Looking-CD: On extremely imbalanced S2Looking-CD, CGNet ranks first in F1 and IoU and second in Recall.
- Qualitative comparison: Qualitative results report fewer false detections and improved incomplete edges and internal holes, whereas some baselines retain these errors.
C. Ablation Study
The ablation study evaluates the Change Guide Module at different network stages. Models with more CGMs achieve progressively stronger accuracy, with the full three-CGM CGNet performing best.
- The study compares a VGG 16_BN base model with single, double, and triple CGM configurations.
- Across datasets, the base model without CGM has the lowest accuracy, while CGNet with three CGMs has the highest.
- Visualization results agree with quantitative results, and the experiments support improved change-information extraction by CGM.
COMPONENT (SINGLE CGM AND DOUBLE CGM) ON FOUR
The study compares one, two, and three CGM modules across four datasets using F1-score. Performance generally increases as CGMs are added, alongside improved edge completeness and fewer internal holes.
- CGNet with three CGMs has the highest accuracy, while double-CGM models outperform single-CGM models.
- Adding CGMs progressively improves change-information extraction, with three modules addressing incomplete edges and internal holes most effectively.
- Visualization and quantitative results are consistent across the ablation comparisons.
ABLATION STUDY ON THE TRAINING AND INFERENCING TIME
The comparison examines training and inference efficiency alongside accuracy across change-detection methods. CGNet is reported to balance these dimensions, retaining competitive inference performance as dataset size increases.
- The study compares training speed per epoch on WHU-CD with test-set inference time across the evaluated methods.
- Pure CNN models such as FC-EF, FC-Siam-conc, and FC-Siam-diff run faster, but their accuracy is low.
- Attention-based methods improve accuracy compared with pure CNN methods, but their greater complexity substantially increases running time.
- CGNet can ensure both model accuracy and training efficiency, providing competitiveness across the two dimensions of time and accuracy.
- CGNet maintains good inference-time performance, with its advantage more apparent on SYSU-CD and S2Looking-CD, which contain relatively large amounts of data.