PolSAR image classification using shallow to deep feature fusion network with complex valued attention.

Area of Science:

• Remote Sensing
• Geospatial Analysis
• Machine Learning

Background:

• Polarimetric Synthetic Aperture Radar (PolSAR) data offers rich information for land cover interpretation.
• Extracting features from PolSAR data presents unique challenges compared to optical imagery.
• Deep Learning (DL) methods, particularly Convolutional Neural Networks (CNNs), show promise in addressing these challenges.

Purpose of the Study:

• To propose a novel three-branch fusion Complex-Valued CNN (CV-ASDF2Net) for PolSAR image classification.
• To evaluate the performance of CV-ASDF2Net against existing state-of-the-art methods.
• To assess the model's effectiveness in feature extraction and classification accuracy.

Main Methods:

• Development of a novel three-branch fusion Complex-Valued CNN architecture (CV-ASDF2Net).
• Utilizing kernel capabilities to process local information and the complex-valued nature of PolSAR data.
• Comparative analysis using Airborne Synthetic Aperture Radar (AIRSAR) datasets (Flevoland, San Francisco) and ESAR Oberpfaffenhofen dataset.

Main Results:

• The proposed CV-ASDF2Net achieved notable improvements in Overall Accuracy (OA): 1.30% and 0.80% for AIRSAR datasets, and 0.50% for the ESAR dataset.
• Exceptional performance on the Flevoland dataset, reaching 96.01% OA with only 1% sampling ratio.
• Quantitative and qualitative evaluations confirmed the superior classification performance.

Conclusions:

• The CV-ASDF2Net model demonstrates significant advancements in PolSAR image classification accuracy.
• The model effectively extracts complex features from PolSAR data, outperforming existing methods.
• The proposed approach offers a powerful tool for land cover interpretation using PolSAR imagery, especially with limited training data.