Wishart Deep Stacking Network for Fast POLSAR Image Classification.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Polarimetric Synthetic Aperture Radar (POLSAR) imagery offers rich information for Earth observation.
• Traditional classification methods struggle to fully exploit complex polarimetric features.
• Deep learning architectures show promise but require specialized adaptations for POLSAR data.

Purpose of the Study:

• To propose a novel deep learning architecture, Wishart Deep Stacking Network (W-DSN), specifically designed for POLSAR image classification.
• To develop an efficient implementation of the Wishart distance for integration into neural networks.
• To improve the accuracy and efficiency of POLSAR image classification using deep learning.

Main Methods:

• A fast implementation of Wishart distance using a linear transformation was developed.
• A single-hidden-layer neural network (NN), the Wishart Network (WN), was created using the fast Wishart distance.
• A multi-layer deep learning architecture, W-DSN, was formed by stacking WNs.

Main Results:

• The fast Wishart distance implementation significantly speeds up POLSAR image classification (e.g., 768,000 pixels classified in 0.53 seconds).
• Both the single-layer WN and the deep W-DSN architectures demonstrated improved classification accuracy for POLSAR images.
• The W-DSN architecture showed further accuracy gains compared to the single-layer WN.

Conclusions:

• The proposed W-DSN offers an efficient and accurate deep learning solution for POLSAR image classification.
• The integration of fast Wishart distance provides a robust method for utilizing polarimetric information in NNs.
• This work establishes a foundational connection between POLSAR interpretation and specialized deep learning architectures.