Multiple Classifiers Based Semi-Supervised Polarimetric SAR Image Classification Method.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Polarimetric Synthetic Aperture Radar (PolSAR) image classification is crucial for data applications.
• Deep learning methods show promise but struggle with insufficient labeled training data.
• Existing deep learning approaches do not fully utilize PolSAR scattering characteristics.

Purpose of the Study:

• To develop a novel semi-supervised classification method for PolSAR images.
• To enhance classification accuracy using limited labeled samples by leveraging scattering characteristics.
• To combine deep learning with traditional scattering trait-based classifiers.

Main Methods:

• A semi-supervised approach combining a complex-valued convolutional neural network (CV-CNN) with traditional classifiers (Wishart, SVM).
• Majority voting of base classifier outputs on limited samples to generate strong and weak datasets.
• Utilizing the strong training set as pseudo-labels for CV-CNN reclassification of the weak dataset.

Main Results:

• The proposed method achieved significant improvements (3-5%) over base classifiers in most cases.
• Superior performance was observed when the number of labeled samples was small.
• Improvements were more pronounced in agricultural and forest areas compared to built-up areas.

Conclusions:

• The novel semi-supervised method effectively enhances PolSAR image classification accuracy with limited labeled data.
• Combining deep learning with scattering characteristics offers a robust solution for data-scarce scenarios.
• The method demonstrates particular efficacy in classifying natural landscapes like forests and agricultural regions.