Ship Classification in High-Resolution SAR Images via Transfer Learning with Small Training Dataset

Area of Science:

• Remote Sensing
• Artificial Intelligence
• Marine Science

Background:

• Synthetic Aperture Radar (SAR) is a crucial all-weather remote sensing technology for oceanographic observations and object tracking.
• Deep learning (DL) and neural networks (NN) show promise for SAR ship classification, but limited labeled data hinders model training.
• Over-fitting is a significant challenge when training neural networks on small datasets for SAR ship classification.

Purpose of the Study:

• To address the bottleneck of limited labeled SAR ship data for deep learning models.
• To improve the accuracy and effectiveness of SAR ship classification using convolutional neural networks (CNNs).
• To mitigate over-fitting issues inherent in training on small datasets.

Main Methods:

• Application of convolutional neural networks (CNNs) for SAR ship classification on small datasets.
• Development and implementation of a novel data augmentation technique.
• Integration of the proposed data augmentation method with transfer learning.

Main Results:

• The proposed method, combining data augmentation and transfer learning, effectively addresses the challenge of limited SAR ship data.
• Experiments demonstrate high classification accuracies for different ship types.
• The study validates the effectiveness of the developed approach for SAR ship classification.

Conclusions:

• The novel data augmentation and transfer learning strategy significantly enhances deep learning-based SAR ship classification performance.
• This approach offers a viable solution for overcoming data limitations in training accurate ship classification models.
• The findings highlight the potential of advanced machine learning techniques for maritime surveillance and analysis using SAR imagery.