Reproduce “A machine learning approach for image classification in synthetic aperture RADAR” on OpenResearch

About this paper

Romina Gaburro, Patrick Healy, Shraddha Naidu, Clifford Nolan

Researchers at the University of Limerick developed a machine learning approach for object classification in Synthetic Aperture Radar (SAR) data, demonstrating that Convolutional Neural Networks (CNNs) trained on raw SAR data consistently yield higher accuracy than those trained on reconstructed images. The method achieved 100% accuracy in simulated geometric shape detection and 75% accuracy in classifying eight real-world sea ice types from Sentinel-1 imagery.

Key insights

• Raw Synthetic Aperture Radar (SAR) data consistently yields superior classification accuracy for CNNs compared to backprojected SAR images, indicating that crucial information is better preserved in the raw signal for machine learning.
• Convolutional Neural Networks effectively process SAR data to classify both geometric shapes and complex environmental features, addressing challenges posed by speckle noise and blurriness.
• The dynamic nature of SAR allows for better distinction of multiple objects compared to static imaging modalities like Electrical Impedance Tomography (EIT) in certain classification scenarios.

Summary by alphaXiv.