Predicting Filtered Image Quality Using Transfer Learning on Sentinel-1 Speckle Noise with DenseNet-121

• Date: 2025
• Main Authors: Al-Senaikh, Raed, Rubel, Oleksii
• Format: Article
• Language: English
• Published: Scientific Centre for Aerospace Research of the Earth Institute of Geological Science National Academy of Sciences of Ukraine, Kyiv, Ukraine 2025
• Subjects: Speckle noise; Sentinel-1; image quality metrics; DenseNet-121; transfer learning
• Online Access: https://ujrs.org.ua/ujrs/article/view/293
• Journal Title: Ukrainian Journal of Remote Sensing of the Earth

Speckle noise inherent to synthetic aperture radar (SAR) imagery degrades image quality and complicates automated analysis in Earth observation applications. Quantitative assessment of despeckling results requires computing quality metrics against reference images, which are unavailable in operational SAR scenarios. This paper presents a method for a priori prediction of filtered Sentinel-1 SAR image quality metrics before applying speckle noise filters. Unlike existing approaches predicting relative quality improvement, the proposed method predicts absolute values of five metrics (PSNR, WSNR, SSIM, MS-SSIM, FSIM) for a specific filter, enabling direct comparison and rational filter selection. The methodology employs transfer learning of DenseNet-121 convolutional neural network, pre-trained on ImageNet, adapted for single-channel SAR inputs through architectural modifications including input layer transformation, pooling optimization, and regression head replacement. A novel synthetic data generation pipeline utilizes histogram matching of Sentinel-2 optical images with Sentinel-1 SAR references to create training samples preserving ground truth. Dynamic gamma-distributed speckle noise addition with variable ENL ∈ [2, 6] enhances data variability and model robustness. Experiments with six classical filters (Gamma MAP, Lee, Enhanced Lee, Frost, SRAD, Kuan) demonstrate high prediction accuracy across all filter-metric combinations. The coefficient of determination R² reaches 0.997 for best combinations and exceeds 0.97 for most of the 30 trained models. Mean absolute prediction errors remain below 0.29 dB for PSNR and 0.014 for SSIM across all tested configurations. The approach enables a priori quality prediction without reference images, allowing optimization of SAR processing workflows and resource planning before resource-intensive despeckling. Contributions of Authors: Conceptualization, Raed A. and Oleksii R.; methodology, Raed A.; formal analysis, Raed A.; investigation, Raed A. and Oleksii R.; data curation, Raed A. and Oleksii R.; writing—original draft preparation, Raed A.; writing—review and editing, Raed A. and Oleksii R.; visualization, Raed A. All authors have read and agreed to the published version of the manuscript. Funding: This research received no external funding. Data Availability Statement: The processed dataset is available in the Zenodo repository (https://zenodo.org/uploads/17253925). Python implementation code is available in the GitHub repository (https://github.com/rsenaikh/Predicting_Quality_after_Noise_Removal). Acknowledgments: The authors would like to express their sincere gratitude to the Copernicus Data Space Ecosystem for providing open access to Sentinel-1 and Sentinel-2 data. We are also grateful to reviewers and editors for their valuable comments, recommendations, and attention to the work. Conflicts of Interest: The authors declare no conflict of interest.