Source-linked AI summary
Super-resolution-based Change Detection Network with Stacked Attention Module for Images with Different Resolutions
Mengxi Liu, Qian Shi, Andrea Marinoni, Da He, Xiaoping Liu, Liangpei Zhang
TL;DR
Change detection must often compare high- and low-resolution images because high-resolution observations are not continuously available, while subpixel methods can accumulate errors in high-resolution imagery. SRCDNet learns a super-resolved image, enhances multi-scale features with stacked CBAM attention, and uses metric learning for change decisions. It outperformed baselines, reaching 87.40% F1 on BCDD and 92.94% on CDD while obtaining the highest accuracies in 4× and 8× resolution-difference experiments.
Problem
High-resolution change detection with different-resolution bi-temporal images is needed, but subpixel methods can accumulate errors because of intraclass heterogeneity and interclass similarity.
Method
SRCDNet restores low-resolution images using adversarial super-resolution, applies five CBAMs to multi-scale features, and uses metric learning to obtain the change map.
Results
87.40% and 92.94% F1 scores were obtained on BCDD and CDD, respectively, with the highest accuracies in 4× and 8× resolution-difference experiments.
Takeaways & Limitations
SRCDNet provides an end-to-end approach for high-resolution change detection with different-resolution images and can also operate without the SR module when resolutions match.
Takeaways & Limitations
The setup assumes HR imagery at one timestamp and LR imagery at another, with the resolution-difference factor specified as N = 4 or 8.
Abstract
from arXiv · showhide
Change detection, which aims to distinguish surface changes based on bi-temporal images, plays a vital role in ecological protection and urban planning. Since high resolution (HR) images cannot be typically acquired continuously over time, bi-temporal images with different resolutions are often adopted for change detection in practical applications. Traditional subpixel-based methods for change detection using images with different resolutions may lead to substantial error accumulation when HR images are employed; this is because of intraclass heterogeneity and interclass similarity. Therefore, it is necessary to develop a novel method for change detection using images with different resolutions, that is more suitable for HR images. To this end, we propose a super-resolution-based change detection network (SRCDNet) with a stacked attention module. The SRCDNet employs a super resolution (SR) module containing a generator and a discriminator to directly learn SR images through adversarial learning and overcome the resolution difference between bi-temporal images. To enhance the useful information in multi-scale features, a stacked attention module consisting of five convolutional block attention modules (CBAMs) is integrated to the feature extractor. The final change map is obtained through a metric learning-based change decision module, wherein a distance map between bi-temporal features is calculated. The experimental results demonstrate the superiority of the proposed method, which not only outperforms all baselines -with the highest F1 scores of 87.40% on the building change detection dataset and 92.94% on the change detection dataset -but also obtains the best accuracies on experiments performed with images having a 4x and 8x resolution difference. The source code of SRCDNet will be available at https://github.com/liumency/SRCDNet.
3. Dept. of Engineering, University of Cambridge, Cambridge, UK
This section identifies an institutional affiliation and lists the paper's main research keywords.
- The listed affiliation is the State Key Laboratory of Information Engineering in Surveying, Mapping and Remote Sensing at Wuhan University.
- The paper concerns change detection in remote sensing images.
- Its stated technical themes include super-resolution, metric learning, and fully convolutional networks.
1. Introduction
The introduction motivates change detection when bi-temporal remote-sensing images have different resolutions and presents SRCDNet to address this problem. The method combines super-resolution, stacked attention, and metric learning, achieving strong results across resolution settings.
- Motivation: High-resolution images improve surface characterization but are not continuously available, so change detection often compares images with different resolutions.
- Motivation: Downsampling loses detailed spatial information, whereas interpolation neglects semantic information needed for detailed change detection.
- Motivation: Subpixel-based methods can accumulate errors from former fine-resolution land-cover maps, especially with high-resolution intraclass heterogeneity and interclass similarity.
- Proposed method: SRCDNet learns a super-resolved image from the low-resolution input, extracts multi-level features with five CBAMs, and predicts changes using metric learning.
- Results: 87.40% and 92.94% F1 scores were achieved on BCDD and CDD, respectively, with the highest accuracies also obtained in 4× and 8× resolution-difference experiments.
2. Related Work
Related work spans fully convolutional change detection, metric-learning-based change decisions, recurrent and attention mechanisms, and deep-learning super-resolution methods.
- Fully convolutional networks and U-Net variants have been widely used for pixel-wise change detection with bi-temporal images.
- Metric-learning approaches replace decoder upsampling by calculating feature distances, minimizing distances for unchanged areas and maximizing them for changed areas.
- Recurrent neural networks can capture spectral or temporal relationships, but their use is limited by infrequent spectral information and deficient temporal information in some high-resolution settings.
- Attention mechanisms have been adopted to emphasize useful spatial, spectral-temporal, or change-related information in extracted features.
- Deep-learning super-resolution research has produced CNN-based methods including SRCNN and the 20-layer VDSR network.
3. Methodology
SRCDNet combines adversarial super-resolution with metric-learning change detection for bi-temporal images at different resolutions. Its stacked attention feature extractor enhances multi-scale information, and the SR module can be removed when resolutions match.
- Architecture: SRCDNet contains a GAN-based super-resolution module and a change detection module that generates change maps from bi-temporal inputs.The SR module reconstructs LR inputs toward HR resolution, while the change detection module extracts features and makes metric-learning-based decisions.
- Architecture: When bi-temporal images have matching resolutions, SRCDNet can omit the SR module and operate as a conventional change detection network.This configuration is described as improving the model’s generality.
- Super-resolution module: The generator learns SR images from LR inputs, while the discriminator distinguishes generated SR images from HR images during adversarial training.The generator uses convolutional and residual blocks followed by sub-pixel convolution, and the discriminator outputs a binary classification probability.
- Feature extraction and attention: Five CBAMs enhance hierarchical features by emphasizing useful channel-wise and spatial-wise information before multi-scale fusion and metric-based change decision.Four CBAMs process residual-block outputs, and a fifth CBAM further refines the fused feature pairs.
- Change decision: Metric learning computes Euclidean distances between CBAM-refined bi-temporal features and uses contrastive loss to separate changed from unchanged areas.The resulting distance map is used to generate the final change map through threshold segmentation.
4. Settings
The experiments evaluate SRCDNet on two real high-resolution change detection datasets against five established baselines. They vary the resolution gap from none to 4× and 8× while reporting precision, recall, F1-score, and IoU.
- Datasets: The evaluation uses BCDD building-change imagery and CDD season-varying imagery containing changes in buildings, roads, forests, cars, and tanks.BCDD provides 0.2-m earthquake-related image pairs, while CDD contains 16,000 Google Earth image pairs with 3–100 cm spatial resolution.
- Baselines: Five baselines are compared: FC-EF, FC-Siam-diff, FC-Siam-conc, BiDateNet, and STANet.The baselines represent fusion, Siamese, recurrent, and spatial-temporal attention approaches to change detection.
- Resolution settings: Three experiments test no resolution difference, a 4× difference, and an 8× difference between bi-temporal images.For X4 and X8, T2 images are downsampled and bicubically interpolated for comparison with same-resolution methods.
- Implementation: Training uses PyTorch for 100 epochs with Adam optimization, an initial learning rate of 0.0001, and batch size eight.The generator loss factors are set to α=0.006, β=0.001, and λ=0.001.
- Evaluation metrics: Performance is measured using precision, recall, F1-score, and IoU.F1 combines precision and recall, while IoU measures intersection over union between detections and ground truth.
5. Experiments and Analysis
Across X1, X4, and X8 experiments, SRCDNet achieves the strongest quantitative and visual change-detection performance, including when bi-temporal images have substantial resolution differences. Its advantages are especially apparent in preserving object boundaries, detecting small changes, and reducing missed or pseudo-changes.
- X1 experiments: 87.40% F1 and 77.63% IoU are the highest BCDD scores, with SRCDNet also achieving 90.13% recall and 84.84% precision.STANet ranks second with 84.96% F1 and 73.86% IoU.
- X1 experiments: 92.94% F1 and 86.81% IoU are the highest CDD scores, while STANet reaches 91.44% F1 and 84.23% IoU.Metric-learning methods capture small changes, whereas STANet shows spillover associated with higher recall and lower precision.
- X1 experiments: SRCDNet preserves precise building footprints and detects a building reduction that the other methods miss in X1 visual comparisons.On CDD visualizations, it also extracts small changes while maintaining their boundaries and shapes.
6. Discussion
Ablation studies show that SAM and SRM each improve change detection, while their combination produces the strongest results across 4× and 8× resolution differences. SRM also improves restored-image quality, although larger resolution differences make image restoration more difficult.
- Ablation results: 84.07% and 88.77% are the SAM-integrated model’s F1 values on BCDD and CDD, compared with base-model values of 77.53% and 86.56%.The increased precision indicates more accurate change extraction with SAM.
- Visual analysis: SRM produces more precise change boundaries and more comprehensive results, while SRCDNet combines these gains with SAM’s improved change extraction.The visual comparisons report fewer missed alarms and more detailed change results with SR images.
- 8× resolution difference: 81.69% F1 and 69.05% IoU are achieved by SRCDNet on BCDD with 8× resolution difference, surpassing the component baselines.The base, SAM-integrated, and SRM-integrated models achieve F1 values of 72.26%, 77.22%, and 75.17%, respectively.
- 8× resolution difference: 83.24% is SRCDNet’s highest F1 on CDD with 8× resolution difference, while SAM and SRM baselines reach 79.91% and 80.59%.The SRM baseline is more effective than the SAM baseline in this CDD setting.
- Restored-image quality: SRCDNet’s restored images achieve the highest PSNR and SSIM in all experiments, whereas bicubic interpolation obtains the lowest values.The SRCDNet metrics are slightly higher than those of the Base+SRM restored images.
- Restored-image quality: X4 experiments show larger PSNR and SSIM gains from SRM than X8 experiments, indicating greater restoration difficulty as resolution differences increase.The paper attributes this to greater image-information loss at larger resolution differences.
7. Conclusion
The paper proposes SRCDNet, an end-to-end network for change detection from bi-temporal images with different resolutions. It combines super-resolution, multi-scale feature extraction with stacked attention, and metric learning, and reports the best results among compared methods across same- and different-resolution settings.
- Method: SRCDNet combines a GAN-based super-resolution module, a Siamese multi-scale feature extractor with stacked attention, and deep metric learning for change-map generation.The super-resolution module reconstructs LR images to HR size before feature comparison.
- Results: The method achieves the best results on BCDD and CDD with same-resolution images and outperforms comparative methods under different resolutions.The conclusion describes this as a more general solution for multi-scene change detection.