A Combined CNN-LSTM Network for Ship Classification on SAR Images.

Area of Science:

• Remote Sensing and Earth Observation
• Machine Learning and Artificial Intelligence
• Signal Processing and Imaging

Background:

• Satellite synthetic aperture radar (SAR) imagery provides global, all-weather data crucial for remote sensing and maritime surveillance.
• Machine learning, particularly convolutional neural networks (CNNs), faces challenges in SAR target classification due to limited training data and SAR image feature constraints.
• Existing CNNs struggle with SAR data due to distinct imaging mechanisms, complicating transfer learning from natural image datasets.

Purpose of the Study:

• To develop a specialized shallow CNN architecture optimized for synthetic aperture radar (SAR) datasets.
• To investigate the impact of CNN component variations (filter number, size) on SAR image classification performance.
• To enhance SAR image classification by combining CNNs with Long short-term memory (LSTM) networks.

Main Methods:

• Proposed a novel shallow CNN architecture tailored for SAR image characteristics.
• Systematically analyzed and optimized CNN layer parameters (filters, sizes) to improve discrimination and reduce complexity.
• Integrated the optimized CNN with LSTM networks for sequential data processing in SAR classification tasks.
• Conducted comparative experiments against six state-of-the-art CNN models (VGG16, ResNet50, Xception, DenseNet121, EfficientNetB0, MobileNetV2) on FUSAR-Ship, OpenSARShip, and MSTAR datasets.

Main Results:

• The proposed shallow CNN architecture achieved competitive accuracy on SAR datasets, outperforming standard models in efficiency.
• Optimizing CNN components led to reduced redundancy, improved discrimination capabilities, and decreased network size and learning time.
• The CNN-LSTM combination demonstrated superior performance for SAR image classification compared to standalone CNNs.
• Evaluations on challenging datasets (FUSAR-Ship, OpenSARShip) with limited and imbalanced data highlighted the model's robustness.

Conclusions:

• Customized CNN architectures are effective in overcoming the specific challenges of SAR target classification.
• The proposed shallow CNN combined with LSTM offers a computationally efficient and accurate solution for SAR image analysis.
• This research demonstrates a promising direction for improving automated target recognition using SAR imagery.