Source-linked AI summary
SAR image despeckling through convolutional neural networks
G. Chierchia, D. Cozzolino, G. Poggi, L. Verdoliva
TL;DR
SAR despeckling is difficult because speckle statistics vary across acquisition conditions and clean references are generally unavailable. The paper trains a residual-learning CNN to estimate speckle using multiplicative-noise-aware processing and multitemporal SAR data, achieving promising objective and visual results on synthetic and real data, including improved performance over reference techniques on simulated images.
Problem
SAR speckle can impair automatic analysis, while its statistical properties vary across sensors, acquisition modalities, multilooking conditions, and land cover.
Method
A residual-learning CNN estimates the speckle component using a multiplicative-noise-aware loss and training pairs built from multitemporal SAR data with multilook clean approximations.
Results
The method shows promising objective and visual results on synthetic and real SAR data; on simulated images, SAR-CNN has about 1 dB, 2 dB, and 2.5 dB average PSNR gains over three reference techniques.
Takeaways & Limitations
CNN-based residual learning is a promising approach for SAR despeckling across synthetic and real data, including challenging high-resolution imagery.
Takeaways & Limitations
Real-data assessment is challenging because the original noiseless signal is unavailable, requiring separate synthetic and real-data validation procedures.
Abstract
from arXiv · showhide
In this paper we investigate the use of discriminative model learning through Convolutional Neural Networks (CNNs) for SAR image despeckling. The network uses a residual learning strategy, hence it does not recover the filtered image, but the speckle component, which is then subtracted from the noisy one. Training is carried out by considering a large multitemporal SAR image and its multilook version, in order to approximate a clean image. Experimental results, both on synthetic and real SAR data, show the method to achieve better performance with respect to state-of-the-art techniques.
1. INTRODUCTION
SAR speckle is multiplicative noise that can impair automatic image analysis, while its statistics vary substantially across sensors, acquisition modalities, multilooking conditions, and land cover. The paper therefore proposes CNN-based learning to avoid explicit statistical modeling and reports experiments on synthetic and real SAR data.
- Speckle can severely impair automatic SAR-image operations such as classification and segmentation, making despeckling a central issue.
- Existing despeckling approaches commonly rely on detailed statistical models of the signal and speckle.
- Speckle statistics may vary significantly with the sensor, acquisition modality, multilooking, and land cover, especially for high-resolution data.
- The proposed approach uses a convolutional neural network to avoid the statistical-modeling problem and learn an implicit model from suitable images.
- The method adapts residual-learning CNN denoising to multiplicative SAR noise through a specialized training-set construction.
- The study evaluates the proposed method on both synthetic and real SAR data before drawing conclusions.
2. PROPOSED METHOD
The proposed SAR-CNN learns the speckle component rather than the clean image, using a multiplicative-noise-aware loss and residual learning. Training pairs are generated from multitemporal SAR data by approximating clean references with multilooking and stable regions.
- The 17-layer fully convolutional network recovers the speckle component, which is subtracted from the noisy image to obtain the despeckled result.All layers use 64 feature maps with 3×3×64 filters except the single-band input and output layers.
- The training criterion combines homomorphic log and exp transforms with a similarity measure designed for multiplicative speckle noise.
- The loss function uses clean-noisy amplitude patch pairs, network output R_Θ, trainable parameters Θ, and the nonzero mean c of log-speckle.
- Residual learning speeds training and improves performance by making the network reproduce noise instead of an output very similar to the input.This is particularly useful when SAR training data are scarce.
- Clean references are approximated by averaging multitemporal SAR components and retaining regions without significant temporal changes.More temporal instances make the clean reference more reliable before noisy patches and clean versions are extracted for training.
3. EXPERIMENTAL RESULTS
The experiments evaluate SAR-CNN on synthetic and real SAR data using objective metrics, visual inspection, and comparisons with PPB, NL-SAR, and SAR-BM3D. SAR-CNN generally performs best on synthetic data and shows promising speckle suppression and detail preservation on real high-resolution imagery.
- Experimental setup: Validation is divided into synthetic experiments with noiseless references and real-image experiments because original noiseless SAR signals are unavailable.Synthetic data use PSNR and SSIM, while real-data assessment relies mainly on visual inspection and no-reference metrics.
- Experimental setup: SAR-CNN is compared with PPB, NL-SAR, and SAR-BM3D using reference implementations and parameters from their papers.The compared methods were selected for competitive performance and software availability.
- Results on simulated SAR images: In all but one synthetic test case, SAR-CNN provides the best PSNR, with average gains of about 1 dB, 2 dB, and 2.5 dB over the reference techniques.SSIM results show similar behavior, and visual inspection indicates improved detail preservation.
- Results on high-resolution SAR images: On COSMO-SkyMed data, SAR-CNN shows speckle suppression comparable to NL-SAR and detail preservation comparable to SAR-BM3D.ENL and αβ results identify SAR-CNN and NL-SAR as achieving the best scores for speckle suppression and detail preservation.
- Results on high-resolution SAR images: The real-data improvement over competitors is less striking because training used well-despeckled patches rather than clean patches.The real experiment uses a 16000×16000 COSMO-SkyMed image with 25 co-registered temporal components.
4. CONCLUSION
The paper investigates CNN-based SAR image despeckling using residual learning and a training procedure based on multitemporal SAR data and its multilook version. Results on synthetic and real SAR data are promising under both objective and visual assessment.
- 4. CONCLUSION: CNN-based SAR despeckling uses residual learning and multitemporal SAR data with a multilook version for training.The training data approximate a clean reference from the same scene.
- 4. CONCLUSION: Synthetic and real SAR experiments show promising results under both objective metrics and visual assessment.The conclusion summarizes the method's performance across both data settings.