TL;DR
This paper introduces deSpeckNet, a deep learning model for SAR image despeckling that can adapt to unknown noise statistics without requiring clean reference images for every new dataset. This generalization capability is achieved through a two-phase training approach, making it more robust and applicable to diverse SAR acquisitions.
What Problem Does This Paper Address?
Synthetic Aperture Radar (SAR) imagery is invaluable for Earth observation due to its all-weather, day/night capabilities. However, a significant challenge in SAR imaging is the presence of "speckle," a multiplicative noise that degrades image quality and hinders analysis. Traditional speckle reduction methods often rely on specific noise models or require paired noisy and clean images for training deep learning models. This limits their applicability to new SAR datasets with different noise characteristics (e.g., from different sensors, acquisition modes, or processing levels). The core problem this paper addresses is the lack of a generalizable SAR despeckling method that can effectively reduce speckle without prior knowledge of the noise distribution or the availability of clean reference images for every new scenario.
Methods
The authors propose deSpeckNet, a deep learning framework that tackles SAR despeckling in two phases:
- Phase 1: Supervised Training: The model is initially trained on a dataset where a clean reference image can be synthesized, typically by taking a temporal average of multi-temporal SAR images. This phase uses a Siamese architecture with two branches: one to estimate the clean image ($\hat{X}$) and another to estimate the noise ($\hat{N}$). Three loss functions are employed:
- $L_{clean}$: Mean Squared Error (MSE) between the reconstructed clean image and the reference clean image.
- $L_{Noisy}$: MSE between the reconstructed noisy image (derived from $\hat{X}$ and $\hat{N}$ using the multiplicative model) and the input noisy image. This loss is crucial for providing a learning signal in the second phase.
- $L_{TV}$: Total Variation loss, which enforces smoothness on the reconstructed clean image by minimizing differences between neighboring pixels, helping to preserve edges.
- Phase 2: Unsupervised Fine-tuning: This phase allows deSpeckNet to adapt to new SAR images with unknown noise statistics without needing clean reference labels. The $L_{clean}$ loss is down-weighted, effectively using the input noisy image as a pseudo-reference. This encourages the model to preserve spatial structures while still leveraging the $L_{Noisy}$ and $L_{TV}$ losses to reduce speckle and enforce smoothness.
The architecture uses convolutional neural network (CNN) blocks, but importantly, it avoids pooling layers to maintain feature map sizes and increase network depth, which is beneficial for reconstructing images to their original resolution.
Datasets
The paper utilizes several SAR datasets for training and evaluation:
- Training Data: Sentinel-1 GRD (Ground Range Detected) images from Indonesia (Pegunungan Barisan area) were used for the initial supervised training. A stack of 23 multi-temporal images from July 2018 to April 2019 was used to synthesize the reference label.
- Testing Data:
- Sentinel-1 GRD: Images from the Democratic Republic of Congo (DRC), Netherlands (Utrecht and Flevoland), acquired in various modes and locations.
- ALOS-2/PALSAR-2: An L-band SAR image acquired over Japan.
- Iceye X2: An X-band SAR image acquired over Germany.
The paper highlights the use of multi-temporal images to create temporally averaged "clean" reference labels for evaluation where possible. For areas where temporal stability is low or multi-temporal data is unavailable (like Flevoland, Japan, and Germany), qualitative assessments are performed, and metrics are derived from homogeneous regions.
Key Insights and Ideas
- Generalization through Unsupervised Fine-tuning: The most significant insight is the development of a two-phase approach that allows a deep learning model to generalize to new SAR datasets without requiring specific clean reference data for each. This is achieved by fine-tuning the model using the noisy input image itself as a pseudo-reference.
- Importance of the $L_{Noisy}$ Loss: The inclusion of an MSE loss on the reconstructed noisy image ($L_{Noisy}$) in Phase 1 is critical. It provides a learning signal that is transferable to Phase 2, where clean references are absent, enabling the model to adapt effectively.
- Implicit Noise Model Estimation: deSpeckNet implicitly learns and adapts to the speckle noise distribution of the input image during fine-tuning, rather than relying on explicit, pre-defined noise models. This makes it robust to variations in sensor characteristics and processing.
- Siamese Architecture for Flexibility: The Siamese architecture, with separate branches for clean image and noise estimation, allows for independent tuning and adaptation of these components during the fine-tuning phase.
Key Results
- Quantitative Performance: deSpeckNet consistently achieved state-of-the-art or comparable results across various quantitative metrics (PSNR, SSIM, DG, EPI, ENL) compared to traditional filters (Lee Sigma) and other deep learning methods (SAR-BM3D, SAR-CNN). For example, in the Indonesia test area, deSpeckNet achieved a PSNR of 45.70 ± 0.05, SSIM of 0.97 ± 0.00, and ENL of 101.83 ± 0.09, often outperforming or matching SAR-CNN which was fine-tuned on a temporally averaged image.
- Qualitative Performance: Visually, deSpeckNet effectively reduced speckle noise while preserving image features and edges across diverse landscapes and sensors, including high-resolution Iceye X2 data. It demonstrated superior performance in maintaining subtle features compared to other methods.
- Adaptability: The model showed remarkable adaptability to different sensors (Sentinel-1, ALOS-2, Iceye) and geographic regions, even when trained on a different sensor's data.
- Noise Distribution Estimation: The paper demonstrates that deSpeckNet can accurately estimate the speckle noise distribution (often Gamma or exponential) for various test areas, which is a key component of its adaptability.
Limitations and Future Work
- Computational Cost: The initial training phase of deSpeckNet is computationally intensive (31.8 hours), though the fine-tuning phase is significantly faster. This is a trade-off for its generalization capabilities.
- Artefacts in Ratio Images: While deSpeckNet effectively despeckles images, the ratio images (Y/X) show some residual speckle and artefacts, particularly in scenes with strong deterministic targets. This is attributed to using the noisy input image as a reference in Phase 2, which can lead to imperfectly uncorrelated speckle.
- Future Work: The authors suggest improving loss functions to better handle mixed urban and rural scenes with both strong deterministic and distributed targets. Further research could focus on optimizing the computational efficiency and exploring more advanced techniques for handling complex scattering scenarios.
Summary updated