Kevala et al. (2025) SARCDNet-an enhanced deep learning network for change detection from bi-temporal SAR images

Identification

• Journal: Scientific Reports
• Year: 2025
• Date: 2025-12-31
• Authors: Vibha Damodara Kevala, Vishal Mukundan, Sravya Nedungatt, Shilpa Suresh, Shyam Lal
• DOI: 10.1038/s41598-025-31488-y

Research Groups

• Department of Electronics and Communication Engineering, National Institute of Technology Karnataka, Surathkal, Mangaluru, India
• Department of Mechatronics, Manipal Institute of Technology, Manipal Academy of Higher Education, Udupi, Karnataka, India

Short Summary

This paper introduces SARCDNet, an enhanced deep learning network for change detection from bi-temporal Synthetic Aperture Radar (SAR) images. SARCDNet, featuring an Adaptive Fusion Block, effectively mitigates speckle noise and significantly improves change detection accuracy across various public datasets, particularly for flood detection.

Objective

• To develop an enhanced deep learning network (SARCDNet) for change detection from bi-temporal SAR images, aiming to mitigate the effects of speckle noise and improve prediction accuracy by enhancing the relevance of extracted features through an innovative Adaptive Fusion Block.

Study Configuration

• Spatial Scale: Image sizes ranging from 256 pixels × 256 pixels to 384 pixels × 384 pixels. The model processes image patches of 12 pixels × 12 pixels.
• Temporal Scale: Bi-temporal SAR images with time gaps ranging from months (e.g., May to August 1997 for Ottawa) to approximately one year (e.g., June 2008 to June 2009 for Yellow River and Farmland), or before/after specific events (e.g., Sulzberger Ice Shelf after tsunami, Chao Lake during high water level).

Methodology and Data

• Models used:
  • SARCDNet (proposed deep learning network)
  • Adaptive Fusion Block (AFB) incorporating depthwise separable convolutions, adaptive average pooling, GlobalFilter module (using Fast Fourier Transform and Inverse Fast Fourier Transform), channel attention mechanism, and Multi-Head Attention.
  • Pre-processing: SRAD filter for speckle reduction, log-ratio difference image generation.
  • Pseudo-labeling: Binary and multi-class Fuzzy C-Means (FCM) clustering.
• Data sources:
  • Satellite SAR images from RADARSAT, Sentinel-1, and Envisat sensors.
  • Public datasets: Ottawa, Yellow River, Farmland, Sulzberger, and Chao Lake.

Main Results

• SARCDNet consistently outperformed five state-of-the-art models across all five datasets.
• Chao Lake dataset (flood detection): Achieved significant improvements over the next best model (LANTNet) with a 4.26% increase in F1 score, 0.841% increase in Percentage of Correct Classification (PCC), 4.92% increase in Kappa coefficient (κ), and 4.67% increase in Matthews Correlation Coefficient (MCC).
• Yellow River dataset: Demonstrated substantial improvements over WBANet with a 2.74% increase in F1 score, 0.857% in PCC, 3.35% in κ, and 3.41% in MCC, highlighting its effectiveness in mitigating speckle noise.
• Computational Efficiency: SARCDNet exhibited the lowest computational complexity among compared models, with 47.16 K FLOPs and 6.287 K trainable parameters, indicating lower resource requirements and faster training times.
• Ablation Study: Confirmed the critical role of the Adaptive Fusion Block, channel attention, and global filter operations in enhancing feature representation and discrimination.
• Visual Results: SARCDNet effectively reduced false negatives (Ottawa, Sulzberger), preserved edges well with fewer false positives (Yellow River, Farmland), and showed minimal false positives (Chao Lake).

Contributions

• Introduction of the Adaptive Fusion Block (AFB) which integrates a channel attention mechanism in the spatial domain and an adaptive global filter operation in the frequency domain. Unlike previous methods, AFB incorporates trainable weights for adaptive learning of frequency domain features and performs feature modulation by combining spatial and frequency domain features, thereby effectively mitigating speckle noise.
• Development of SARCDNet, a computationally efficient deep learning model that combines the AFB with a shallow Convolutional Neural Network (CNN). Its use of separable convolutions significantly reduces model complexity, requiring fewer computational resources while achieving superior classification results compared to existing state-of-the-art models.

Funding

No funding was received for conducting this study.

Citation

@article{Kevala2025SARCDNetan,
  author = {Kevala, Vibha Damodara and Mukundan, Vishal and Nedungatt, Sravya and Suresh, Shilpa and Lal, Shyam},
  title = {SARCDNet-an enhanced deep learning network for change detection from bi-temporal SAR images},
  journal = {Scientific Reports},
  year = {2025},
  doi = {10.1038/s41598-025-31488-y},
  url = {https://doi.org/10.1038/s41598-025-31488-y}
}