SDF2Net boosts PolSAR classification accuracy by up to 1.3%
A novel three-branch CNN fusion achieves 96.01% accuracy with only 1% training data.
A team from the University of Sharjah and other institutions introduced SDF2Net (Shallow to Deep Feature Fusion Network), a specialized deep learning architecture for polarimetric synthetic aperture radar (PolSAR) image classification. Unlike conventional CNNs used in optical imagery, SDF2Net employs a three-branch fusion of complex-valued convolutional neural networks designed specifically to handle the complex-valued nature and local spatial information of PolSAR data. The network fuses features from shallow layers (fine details) with deeper layers (semantic context), enabling richer representations for land cover classification.
Tested on three benchmark datasets—AIRSAR Flevoland, AIRSAR San Francisco, and ESAR Oberpfaffenhofen—SDF2Net outperformed multiple state-of-the-art approaches. It achieved accuracy improvements of 1.3% and 0.8% on the AIRSAR datasets and 0.5% on ESAR. Notably, on Flevoland, SDF2Net reached 96.01% overall accuracy using only 1% of the labeled pixels for training, demonstrating exceptional data efficiency. This makes the model highly practical for real-world remote sensing applications where labeled data is scarce. The paper is available on arXiv (2402.17672) and was updated in May 2026.
- Three-branch complex-valued CNN fusion (SDF2Net) specifically designed for PolSAR data challenges.
- Outperforms state-of-the-art by 1.3% (Flevoland), 0.8% (San Francisco), and 0.5% (Oberpfaffenhofen).
- Achieves 96.01% accuracy with only 1% training samples, showing strong data efficiency.
Why It Matters
Enables more accurate land cover mapping from radar imagery with minimal labeled data, critical for environmental monitoring.