Review of the studies on remote sensing classification based on polarimetric Synthetic Aperture Radar

Yu Li, Jingfei Yang, Hongsheng Zhang, Gang Li, Jie ChenNational 2024 · DOI

Abstract

Polarimetric Synthetic Aperture Radar (PolSAR) has become a cornerstone of modern remote sensing due to its ability to provide all-weather, day-and-night observations with rich physical information. This review synthesizes the development of PolSAR classification methods, ranging from traditional polarimetric decomposition to state-of-the-art deep learning and multi-source fusion. We highlight the shift toward Compact Polarimetry (CP) for large-scale applications and the critical role of preserving phase information through complex-valued neural networks.

Introduction

While traditional optical remote sensing is often hindered by cloud cover and daylight, Synthetic Aperture Radar (SAR) provides reliable all-weather imaging. Polarimetric SAR (PolSAR) enhances this capability by capturing the "fingerprint" of how surfaces reflect microwave energy across different polarizations. The core challenge in the field is effectively translating these complex polarimetric signatures into accurate land-cover maps, ship detections, or environmental assessments. This paper tracks the evolution of classification methodologies from manual feature engineering to automated, AI-driven systems.

Overall framework of the research review

Classification Methodologies

The evolution of PolSAR classification can be categorized into four primary pillars, each offering unique advantages and facing specific challenges:

  1. Polarimetric Decomposition: Techniques such as H/A/$\alpha$ (Cloude-Pottier) and Freeman-Durden decomposition break down radar returns into physical components like surface scattering, double-bounce (from buildings), or volume scattering (from vegetation). These methods provide clear physical meaning but can struggle with complex, mixed-scattering environments.
  2. Classical Machine Learning: Support Vector Machines (SVM), Random Forests, and Bayesian classifiers have been workhorses in the field. They are computationally efficient but require significant manual effort in feature extraction and engineering.
  3. Deep Learning: The adoption of Convolutional Neural Networks (CNNs) and Vision Transformers has revolutionized the field, particularly in complex urban scenes. A key advancement is the use of Complex-valued Neural Networks, which preserve the phase information inherent in radar data—a feature often discarded by standard image-processing models.
  4. Compact Polarimetry (CP): CP represents a strategic middle ground, providing more information than single-polarization SAR while maintaining the wide-area coverage capabilities that Full Polarimetry lacks.
Comparison of different land-cover classification algorithm categories

Data Sources and Applications

Research in this field relies on a diverse array of satellite platforms. High-resolution full-polarimetric data is provided by missions like ALOS-2 (L-band), Radarsat-2 (C-band), and China’s Gaofen-3. Meanwhile, open-access data from Sentinel-1 has democratized the field, enabling large-scale studies despite its dual-pol limitations. Emerging missions like the Radarsat Constellation Mission (RCM) are increasingly focused on Compact Polarimetry to scale applications globally.

These data sources are applied across various domains, including:

  • Marine Monitoring: Detecting oil spills and ships using the unique scattering properties of the sea surface.
  • Urban Mapping: Differentiating between various types of man-made structures and impervious surfaces.
  • Cryosphere Studies: Classifying sea ice and ice sheets, where Compact Polarimetry has shown accuracy comparable to Full Pol systems.

Key Insights and Future Directions

A major trend in modern research is the Physical-AI Hybrid approach, where polarimetric decomposition results are used as input features for deep learning models. This combines the interpretability of physics with the pattern-recognition power of AI. Furthermore, the "gold standard" for high-precision mapping is shifting toward Multi-Source Synergy, fusing SAR's structural data with the spectral information of optical sensors and the height data from LiDAR.

Despite these advances, several hurdles remain. There is a critical shortage of large-scale, high-quality labeled PolSAR benchmark datasets. Additionally, the geometric differences between SAR and optical sensors make pixel-perfect registration a significant technical challenge. Future work is expected to focus on Time-Series Analysis, moving beyond static snapshots to track dynamic environmental changes like deforestation and urban expansion over time.

Conclusion

PolSAR classification is transitioning toward a more integrated and operationally scalable future. By bridging the gap between microwave physics and advanced machine learning, the remote sensing community is developing more robust tools for global Earth observation and environmental monitoring.

Summary updated

Abstract

100124; 2. , 999077; 3. /
BibTeX
@inproceedings{li2024review,
  author = {Yu Li and Jingfei YANG and Hongsheng Zhang and Gang Li and Jie Chen},
  title = {{Review of the studies on remote sensing classification based on polarimetric Synthetic Aperture Radar}},
  booktitle = {National},
  year = {2024},
  doi = {10.11834/jrs.20242346},
}
Figure