Abstract
Change detection in Synthetic Aperture Radar (SAR) images is a critical task for environmental monitoring and disaster management, yet it is frequently hindered by inherent speckle noise and the scarcity of high-quality labeled data. This paper presents a novel framework utilizing a "Siamese samples" Convolutional Neural Network (CNN) architecture. By leveraging a Joint-Similarity Classifier (JSC) to generate initial pseudo-labels, the proposed method learns robust local descriptors that effectively suppress noise. Experimental results on real-world datasets demonstrate that this approach achieves state-of-the-art accuracy with a superior speed-to-performance trade-off compared to traditional dual-branch architectures.
Introduction
Synthetic Aperture Radar (SAR) is an invaluable tool for remote sensing due to its ability to operate in all weather conditions and at any time of day. However, the coherent nature of SAR imaging introduces "speckle noise," a granular interference that complicates the identification of actual ground changes. Traditional change detection methods often struggle to distinguish between these noise artifacts and meaningful temporal variations.
While deep learning has shown promise, its application to SAR change detection is limited by the difficulty of obtaining ground-truth labels. Researchers often resort to "pseudo-labels" generated by unsupervised algorithms, but these labels are frequently coarse and inaccurate. There is a pressing need for network architectures that are inherently robust to such noisy training data.
Proposed Methodology
The authors propose a comprehensive three-stage framework designed to extract clean change information from noisy SAR pairs.
1. Pre-classification via Joint-Similarity Classifier (JSC)
To address the lack of labels, a Joint-Similarity Classifier (JSC) is employed. Derived from fuzzy c-means clustering, the JSC generates a coarse initial change map. These "pseudo-labels" serve as the starting point for training the neural network. The logic involves comparing pixel values and similarity indices against defined thresholds to refine the initial data.
2. "Siamese Samples" CNN Architecture
The core innovation is the "Siamese samples" architecture. Unlike standard Siamese networks that use two parallel branches, this model treats patch pairs as indiscriminate samples fed into a single-branch network.
- Feature Extraction: The lower layers learn to extract high-level, noise-invariant descriptors for each patch.
- Metric Learning: The top fully connected layers join these descriptors to learn a distance metric, classifying the relationship between patches as "changed" or "unchanged."
3. Iterative Fine-tuning
The network is fine-tuned using the coarse labels from the JSC. As the CNN learns to identify discriminative spatial features, it effectively acts as a sophisticated filter, "seeing through" the initial label noise to produce a significantly cleaner final result.
Experimental Results
The model was rigorously tested on several real-world SAR datasets, including the Yellow River Estuary (Coastline and Farmland), Ottawa, and Bern.
- Accuracy and Robustness: On the Coastline dataset, the proposed method achieved a Kappa coefficient of 90.43% and a Percentage of Correct Classification (PCC) of 99.80%. These results surpassed standard Siamese, Pseudo-Siamese, and 2-Channel architectures.
- Efficiency: The "Siamese samples" approach offers a compelling balance between computational speed and detection accuracy. While 2-Channel networks may be slightly faster, they often lack the precision of the proposed single-branch descriptor learning method.
- Qualitative Analysis: Visual inspections of the resulting change maps show a marked reduction in "false alarm" spots in unchanged regions, demonstrating the network's superior speckle noise suppression capabilities.
Conclusion
This research demonstrates that a single-branch CNN architecture, when combined with an effective pseudo-labeling strategy, can overcome the significant challenges posed by speckle noise in SAR change detection. By learning robust local descriptors, the model achieves high accuracy without requiring perfect training data. Future work will explore the application of this framework to heterogeneous data sources and the development of fully unsupervised learning paradigms.
Summary updated