Sparse Manifold-Regularized Neural Networks for Polarimetric SAR Terrain Classification

Area of Science:

• Remote Sensing
• Artificial Intelligence
• Signal Processing

Background:

• Polarimetric Synthetic Aperture Radar (PolSAR) data offers rich information for Earth observation.
• Accurate classification of PolSAR data is crucial for various applications.
• Conventional deep neural networks (DNNs) face challenges in feature extraction from complex SAR data.

Purpose of the Study:

• To propose a novel deep neural network, DSMR, for improved feature extraction and classification of PolSAR data.
• To leverage sparse filtering and manifold regularization for enhanced PolSAR analysis.
• To demonstrate the effectiveness of DSMR across different SAR systems and incidence angles.

Main Methods:

• Developed a deep neural network (DSMR) integrating sparse filtering and manifold regularization.
• Employed dual sparsity (population and lifetime) for global feature learning.
• Utilized neighborhood-based manifold regularization to preserve local data structures.
• Incorporated spatial information during preprocessing to weight data samples.

Main Results:

• DSMR automatically learns relevant features from raw SAR data.
• The dual sparsity mechanism requires minimal parameter tuning.
• Manifold regularization reduces the need for extensive training samples.
• Experimental results show improved classification accuracy compared to conventional DNNs, particularly for large incidence angles.

Conclusions:

• DSMR offers a robust approach for PolSAR data feature extraction and classification.
• The proposed method achieves superior performance over existing DNNs.
• DSMR demonstrates significant potential for advancing SAR image analysis and interpretation.