A New Deep Learning Algorithm for SAR Scene Classification Based on Spatial Statistical Modeling and Features...

Area of Science:

• Remote Sensing
• Computer Vision
• Artificial Intelligence

Background:

• Synthetic Aperture Radar (SAR) scene classification is crucial for various applications.
• Deep learning offers powerful hierarchical feature learning for complex SAR data.
• Existing methods face challenges in achieving high accuracy and generality.

Purpose of the Study:

• To propose a novel deep learning framework, FRN-MSF, for accurate SAR scene classification.
• To enhance feature extraction by integrating multi-scale spatial information.
• To validate the proposed framework's performance and generalizability across different SAR datasets.

Main Methods:

• Constructed a Multi-Scale Omnidirectional Gaussian Derivative Filter (MSOGDF).
• Generated Multi-scale Spatial Features (MSF) using MSOGDF, GLGCM, and Gabor transformation.
• Developed a Feature Recalibration Network (FRN) integrating Depthwise Separable Convolution (DSC), SE Block, and CNN for high-level feature learning.

Main Results:

• Achieved a mean accuracy of 98.18% on eleven types of SAR scenes from four systems.
• Demonstrated strong generalizability with a 94.56% mean accuracy on Gaofen-3 and TerraSAR-X datasets.
• FRN-MSF outperformed previous classifications on the same datasets.

Conclusions:

• The FRN-MSF framework significantly advances SAR scene classification accuracy.
• The integration of multi-scale spatial features and recalibration network enhances performance.
• FRN-MSF shows great potential for robust and generalizable SAR scene classification.