Labeled dataset for training despeckling filters for SAR imagery

Area of Science:

• Remote Sensing
• Artificial Intelligence
• Image Processing

Background:

• Supervised learning models require labeled datasets, typically input-output pairs.
• For image processing tasks like SAR despeckling, a noisy image and its corresponding denoised ground truth are needed.
• Existing SAR despeckling methods often rely on synthetically corrupted images due to the lack of real-world ground truth.

Purpose of the Study:

• To present a new dataset for SAR despeckling using real Sentinel-1 SAR images.
• To overcome the limitation of synthetic noise in current SAR despeckling datasets.
• To facilitate the development and training of AI and Deep Learning models for SAR image enhancement.

Main Methods:

• Utilized Sentinel-1 SAR images from the same geographical region captured at different times.
• Processed and merged multiple SAR images to create a single ground truth image.
• Split all SAR images (noisy and ground truth) into 1600 smaller images of 512x512 pixels each.
• Organized the dataset into 3000 images for training and 200 for validation, with labeled folders.

Main Results:

• A dataset comprising 3200 images (1600 noisy, 1600 ground truth) was generated.
• The dataset is structured for supervised learning, offering real SAR data pairs.
• The dataset is divided into training and validation sets for model development.

Conclusions:

• The proposed dataset provides a realistic foundation for training AI models for SAR despeckling.
• This resource addresses the need for authentic data in developing advanced SAR image processing techniques.
• The availability of this dataset is expected to advance research in AI-driven SAR image analysis.