[Paper Review] Classification of polarimetric SAR images using compact convolutional neural networks

Mete Ahishali, Serkan Kıranyaz+2 more|arXiv (Cornell University)|Apr 26, 2022

Synthetic Aperture Radar (SAR) Applications and Techniques47 references24 citations

TL;DR

This paper proposes a compact and adaptive convolutional neural network (CNN) for polarimetric SAR (PolSAR) image classification, using a sliding-window approach that bypasses manual feature extraction. The method achieves high accuracy (up to 99.39%) with minimal training data (<0.1% of data) and low computational cost, enabling real-time processing on standard hardware, even with small window sizes (7×7 to 19×19).

ABSTRACT

Classification of polarimetric synthetic aperture radar (PolSAR) images is an active research area with a major role in environmental applications. The traditional Machine Learning (ML) methods proposed in this domain generally focus on utilizing highly discriminative features to improve the classification performance, but this task is complicated by the well-known ?curse of dimensionality? phenomena. Other approaches based on deep Convolutional Neural Networks (CNNs) have certain limitations and drawbacks, such as high computational complexity, an unfeasibly large training set with ground-truth labels, and special hardware requirements. In this work, to address the limitations of traditional ML and deep CNN-based methods, a novel and systematic classification framework is proposed for the classification of PolSAR images, based on a compact and adaptive implementation of CNNs using a sliding-window classification approach. The proposed approach has three advantages. First, there is no requirement for an extensive feature extraction process. Second, it is computationally efficient due to utilized compact configurations. In particular, the proposed compact and adaptive CNN model is designed to achieve the maximum classification accuracy with minimum training and computational complexity. This is of considerable importance considering the high costs involved in labeling in PolSAR classification. Finally, the proposed approach can perform classification using smaller window sizes than deep CNNs. Experimental evaluations have been performed over the most commonly used four benchmark PolSAR images: AIRSAR L-Band and RADARSAT-2-C-Band data of San Francisco Bay and Flevoland areas. Accordingly, the best obtained overall accuracies range between 92.33-99.39% for these benchmark study sites.

Motivation & Objective

• Address the high computational complexity and data requirements of deep CNNs in PolSAR classification.
• Overcome the limitations of traditional machine learning methods that rely on manual, high-dimensional feature engineering.
• Enable accurate PolSAR classification with minimal labeled training data and low computational overhead.
• Demonstrate robustness and generalization across different PolSAR datasets and window sizes.
• Enable real-time processing by eliminating the need for specialized hardware and extensive pre-processing.

Proposed method

• Proposes a compact, adaptive 4-layer CNN architecture with shared weights and reduced parameters to minimize computational load.
• Uses a sliding-window classification approach to process small image patches (7×7 to 19×19 pixels) directly from second-order PolSAR descriptors.
• Processes raw complex backscattering matrices without manual feature extraction, leveraging end-to-end learning for joint feature extraction and classification.
• Employs multi-look coherency matrices and polarimetric descriptors (e.g., SPAN, Pauli decomposition) as input channels (3–6 channels) to preserve physical scattering information.
• Applies backpropagation with adaptive learning to train the network efficiently on limited labeled data.
• Uses hyper-parameter tuning (multipliers for neurons and layers) to assess model robustness and avoid overfitting.

Experimental results

Research questions

• RQ1 Can a compact CNN achieve high classification accuracy in PolSAR images with minimal labeled training data?
• RQ2 How does the proposed method compare to traditional machine learning and deep CNNs in terms of computational efficiency and data requirements?
• RQ3 What is the optimal window size for accurate PolSAR classification using compact CNNs?
• RQ4 Can the model generalize across different PolSAR datasets (e.g., San Francisco Bay, Flevoland) with cross-site validation?
• RQ5 How sensitive is the model’s performance to variations in network depth and width (hyper-parameters)?

Key findings

• The proposed compact CNN achieved overall accuracies between 92.33% and 99.39% on four benchmark PolSAR datasets (AIRSAR L-Band and RADARSAT-2 C-Band for San Francisco Bay and Flevoland).
• Best performance was achieved with 99.39% OA on SFBay_L using a 4-channel input and 21×21 window size.
• The model generalized well in cross-site validation, achieving 88.11% OA on Flevo_L when trained on Flevo_C, even with combined cropland classes.
• The method required less than 0.1% of the total data for training, significantly reducing labeling costs and data dependency.
• Increasing network depth or width (e.g., m=8, n=1 or n=2) led to overfitting and non-convergence, indicating that compactness is crucial for stability.
• The model maintained high accuracy across window sizes from 7×7 to 19×19, demonstrating robustness to small patch sizes.