Ship Classification in High-Resolution SAR Images Using Deep Learning of Small Datasets.

Area of Science:

• Remote Sensing
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning excels in image recognition but faces challenges with small datasets in Synthetic Aperture Radar (SAR) ship classification.
• High-resolution SAR imagery presents unique difficulties for automated feature learning due to data limitations.

Purpose of the Study:

• To apply deep learning, specifically convolutional neural networks (CNNs), for effective ship classification using high-resolution SAR images.
• To address the training bottleneck caused by small datasets in SAR ship classification tasks.

Main Methods:

• Constructed ship chips from high-resolution SAR images and divided them into training and validation sets.
• Developed a ship classification model using Very Deep Convolutional Networks (VGG), pre-trained on ImageNet, and fine-tuned for SAR data.
• Evaluated the model using six scenes of COSMO-SkyMed SAR images.

Main Results:

• The proposed fine-tuned CNN model achieved over 95% average classification accuracy, even with 5-fold cross-validation.
• The VGG16-based ship classification model demonstrated at least a 2% higher accuracy compared to other evaluated models.
• The method proved effective in classifying ships within high-resolution SAR imagery despite dataset size constraints.

Conclusions:

• Fine-tuning pre-trained deep learning models, like VGG16, is a highly effective strategy for ship classification in high-resolution SAR images.
• The developed approach successfully overcomes the limitations of small datasets, paving the way for improved automated analysis of SAR imagery.
• The study highlights the potential of deep learning for enhancing remote sensing applications, particularly in maritime surveillance and object detection.