Unsupervised SAR Imagery Feature Learning with Median Filter-Based Loss Value.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Limited availability of labeled Synthetic Aperture Radar (SAR) imagery databases.
• Scarcity of pre-trained neural networks for SAR image analysis.
• Challenges posed by speckle noise in SAR imagery.

Purpose of the Study:

• To address the limitations of SAR data scarcity and noise for neural network training.
• To present comprehensive data augmentation techniques for SAR imagery.
• To propose a novel filter-based method for enhancing SAR datasets.

Main Methods:

• Detailed description of SAR imagery characteristics and limitations.
• Presentation of various data augmentation strategies.
• Introduction of a novel filter-based method to mitigate speckle noise.

Main Results:

• The proposed filter-based method effectively reduces speckle noise in SAR imagery.
• Demonstrable improvement in dataset quality, evidenced by loss value functions.
• Enhanced feature detection capabilities in neural networks trained with the improved dataset, confirmed by layer-wise analysis.

Conclusions:

• The developed filter-based augmentation method significantly improves SAR datasets for neural network training.
• The method leads to more robust feature detectors, enhancing model performance.
• Trained neural networks are provided for open use, facilitating quicker implementation of CNN-based solutions.