Pixel-Wise PolSAR Image Classification via a Novel Complex-Valued Deep Fully Convolutional Network

Yice Cao, Yan Wu, Peng Zhang, Wenkai Liang, Ming Li

TL;DR

This paper introduces CV-FCN, a novel deep learning model for classifying PolSAR (Polarimetric Synthetic Aperture Radar) images. By directly processing complex-valued PolSAR data and incorporating phase information, CV-FCN achieves superior classification accuracy and robustness to speckle noise compared to existing methods, including real-valued deep learning approaches.

What Problem Does This Paper Address?

PolSAR images contain rich information about the Earth's surface, making them valuable for applications like land use/land cover (LULC) classification. While complex-valued (CV) neural networks have shown promise for PolSAR classification, most existing deep learning approaches for this task still rely on converting the complex-valued PolSAR data into real-valued (RV) features. This conversion can lead to a loss of crucial phase information, potentially limiting classification accuracy and making it unclear which projection method is optimal. Furthermore, existing CV deep learning models are often shallow and may not fully leverage spatial context, making them susceptible to speckle noise inherent in SAR imagery. This paper aims to bridge this gap by developing a deep, pixel-wise CV neural network that directly utilizes PolSAR data's complex nature and phase information for improved classification.

Methods

The core of this work is the Complex-Valued Fully Convolutional Network (CV-FCN). Key methodological contributions include:

  • Complex-Valued Architecture: The entire network, including weights, biases, inputs, and outputs, is designed to operate in the complex domain, directly processing raw PolSAR CV data without projection.
  • Complex Downsampling-Then-Upsampling Scheme: This forms the backbone of the CV-FCN, enabling pixel-level labeling.
  • Multi-Level CV Feature Extraction: A complex downsampling section extracts rich, discriminative CV features at various levels, preserving polarization information.
  • Complex Upsampling with Max-Unpooling: A novel complex upsampling scheme, featuring complex max-unpooling layers, is introduced to recover spatial information and enhance robustness to speckle noise. These layers use "max locations maps" to guide the upsampling of real and imaginary parts of feature maps.
  • Complex-Valued Weight Initialization: A new scheme is proposed that considers the statistical distribution of PolSAR data to initialize CV-FCN weights, aiming for faster training and better convergence.
  • Average Cross-Entropy Loss Function: A novel loss function is derived for CV-FCN optimization. It operates on both the real and imaginary parts of the predicted labels, effectively incorporating phase information for more precise classification.

Datasets

The proposed CV-FCN was evaluated on three benchmark PolSAR datasets:

  • Flevoland Benchmark Dataset: A well-known dataset with 14 land cover classes, acquired by NASA/JPL AIRSAR.
  • San Francisco Dataset: A dataset covering four target types: water, vegetation, low-density urban, and high-density urban.
  • Oberpfaffenhofen Dataset: An ESAR dataset from Germany with three classes: built-up areas, woodland, and open areas.

The paper does not introduce new benchmark datasets.

Key Insights and Ideas

  • Phase Matters: Directly using complex-valued data and phase information in deep networks significantly improves PolSAR classification accuracy and robustness compared to real-valued counterparts. This is a recurring theme across all experiments.
  • Deep FCN for Pixel-wise PolSAR: Adapting the Fully Convolutional Network (FCN) architecture to the complex domain is highly effective for pixel-level PolSAR classification, outperforming shallower CV networks.
  • Max-Unpooling for Speckle Resilience: The proposed complex max-unpooling layers are crucial for recovering spatial information, leading to smoother and more coherent classification maps that are less affected by speckle noise.
  • Data-Driven Initialization: A CV weight initialization scheme tailored to PolSAR data statistics can accelerate training and improve performance, suggesting that understanding data characteristics is key for effective deep learning in this domain.

Key Results

  • Superior Performance: CV-FCN consistently outperformed all compared methods (including non-deep methods, RV-NNs, and other CV-NNs) across all three datasets in terms of Overall Accuracy (OA), Average Accuracy (AA), and Kappa coefficient.
  • Significant Accuracy Gains: On the Flevoland dataset, CV-FCN achieved OA of 99.72% and Kappa of 0.9967. For the San Francisco dataset, OA reached 99.69% and Kappa 0.9957. The Oberpfaffenhofen dataset saw an OA of 97.26% and Kappa of 0.9531.
  • CV vs. RV: CV-NNs consistently outperformed their RV counterparts, with CV-FCN showing notable improvements, especially in Average Accuracy (AA).
  • Faster Convergence: The proposed complex-valued weight initialization scheme led to faster convergence and earlier optimal value attainment compared to existing CV initialization methods.
  • Loss Function Impact: The proposed average cross-entropy loss function demonstrated faster convergence and better classification results than CMSE and CMAE, while also achieving better boundary localization and speckle noise robustness.

Limitations and Future Work

  • Visualization of Initialization: While the new CV weight initialization scheme shows promise, the authors suggest further visualization to better understand its superiority.
  • Dataset Limitations: The availability of high-quality, large-scale PolSAR datasets and training data remains a challenge. Training deeper complex-valued networks is difficult due to the risk of overfitting and model collapse.
  • Data Augmentation: The authors note that data augmentation strategies for natural images are not directly suitable for PolSAR due to differing imaging mechanisms. Developing effective PolSAR-specific data augmentation is identified as an urgent need.
  • Further Exploration: Future work could involve exploring more advanced data augmentation strategies and investigating deeper CV networks for PolSAR classification, while addressing overfitting and model collapse.

Summary updated

Abstract

Although complex-valued (CV) neural networks have shown better classification results compared to their real-valued (RV) counterparts for polarimetric synthetic aperture radar (PolSAR) classification, the extension of pixel-level RV networks to the complex domain has not yet thoroughly examined. This paper presents a novel complex-valued deep fully convolutional neural network (CV-FCN) designed for PolSAR image classification. Specifically, CV-FCN uses PolSAR CV data that includes the phase information and uses the deep FCN architecture that performs pixel-level labeling. The CV-FCN architecture is trained in an end-to-end scheme to extract discriminative polarimetric features, and then the entire PolSAR image is classified by the trained CV-FCN. Technically, for the particularity of PolSAR data, a dedicated complex-valued weight initialization scheme is proposed to initialize CV-FCN. It considers the distribution of polarization data to conduct CV-FCN training from scratch in an efficient and fast manner. CV-FCN employs a complex downsampling-then-upsampling scheme to extract dense features. To enrich discriminative information, multi-level CV features that retain more polarization information are extracted via the complex downsampling scheme. Then, a complex upsampling scheme is proposed to predict dense CV labeling. It employs the complex max-unpooling layers to greatly capture more spatial information for better robustness to speckle noise. The complex max-unpooling layers upsample the real and the imaginary parts of complex feature maps based on the max locations maps retained from the complex downsampling scheme. In addition, to achieve faster convergence and obtain more precise classification results, a novel average cross-entropy loss function is derived for CV-FCN optimization. Experiments on real PolSAR datasets demonstrate that CV-FCN achieves better classification performance than other state-of-art methods.
BibTeX
@inproceedings{cao2019pixel,
  author = {Yice Cao and Yan Wu and Peng Zhang and Wenkai Liang and Ming Li},
  title = {{Pixel-Wise PolSAR Image Classification via a Novel Complex-Valued Deep Fully Convolutional Network}},
  booktitle = {Remote},
  year = {2019},
  doi = {10.3390/rs11222653},
}
Figure