[Paper Review] Region-Based Classification of PolSAR Data Using Radial Basis Kernel Functions With Stochastic Distances

R. G. Negri, A. C. Frery+3 more|arXiv (Cornell University)|May 7, 2018

Synthetic Aperture Radar (SAR) Applications and Techniques4 citations

TL;DR

This paper proposes a novel region-based PolSAR classification method using Support Vector Machines (SVM) with radial basis kernel functions derived from stochastic distances (Kullback-Leibler, Rényi, Bhattacharyya, Hellinger, Chi-Square) between complex Wishart distributions. The approach improves robustness and accuracy over Minimum Distance Classification (MDC), especially under imperfect training data, with Kullback-Leibler and Rényi distances yielding the best performance despite higher computational cost.

ABSTRACT

Region-based classification of PolSAR data can be effectively performed by seeking for the assignment that minimizes a distance between prototypes and segments. Silva et al (2013) used stochastic distances between complex multivariate Wishart models which, differently from other measures, are computationally tractable. In this work we assess the robustness of such approach with respect to errors in the training stage, and propose an extension that alleviates such problems. We introduce robustness in the process by incorporating a combination of radial basis kernel functions and stochastic distances with Support Vector Machines (SVM). We consider several stochastic distances between Wishart: Bhatacharyya, Kullback-Leibler, Chi-Square, Rényi, and Hellinger. We perform two case studies with PolSAR images, both simulated and from actual sensors, and different classification scenarios to compare the performance of Minimum Distance and SVM classification frameworks. With this, we model the situation of imperfect training samples. We show that SVM with the proposed kernel functions achieves better performance with respect to Minimum Distance, at the expense of more computational resources and the need of parameter tuning. Code and data are provided for reproducibility.

Motivation & Objective

• To improve robustness of region-based PolSAR classification under imperfect training samples.
• To investigate the effectiveness of stochastic distances (Bhattacharyya, Kullback-Leibler, Chi-Square, Rényi, Hellinger) as kernel functions in SVM for PolSAR data.
• To compare the performance of SVM with radial basis kernels based on stochastic distances against Minimum Distance Classifier (MDC) using the same distances.
• To evaluate the impact of multiclass strategies (OAA vs. OAO) and class imbalance on classification accuracy.
• To provide open-source code and data for reproducibility and practical application in PolSAR image analysis.

Proposed method

• Formulates radial basis kernel functions using five stochastic distances between complex multivariate Wishart distributions: Bhattacharyya, Kullback-Leibler, Chi-Square, Rényi, and Hellinger.
• Applies these kernel functions within a Support Vector Machine (SVM) framework for region-based classification of PolSAR images.
• Employs both simulated PolSAR data and real data from operational sensors (e.g., ALOS PALSAR) across multiple classification scenarios.
• Compares SVM performance against Minimum Distance Classifier (MDC) using the same stochastic distances, evaluating accuracy and computational cost.
• Uses OAA (One-Against-All) and OAO (One-against-One) multiclass strategies in SVM to assess their impact on classification performance.
• Validates results using quantitative metrics such as overall accuracy and Kappa coefficient, with statistical significance testing.

Experimental results

Research questions

• RQ1 How do different stochastic distances (Bhattacharyya, Kullback-Leibler, Chi-Square, Rényi, Hellinger) perform when used as radial basis kernel functions in SVM for region-based PolSAR classification?
• RQ2 Does SVM with stochastic distance-based kernels outperform Minimum Distance Classifier (MDC) in terms of classification accuracy, especially under imperfect training data?
• RQ3 How do multiclass strategies (OAA vs. OAO) affect the performance of SVM with stochastic distance kernels in PolSAR classification?
• RQ4 What is the influence of class imbalance and intra-class variability on the classification accuracy of MDC and SVM?
• RQ5 Which stochastic distance yields the most robust and accurate classification results when integrated into an SVM kernel for PolSAR data?

Key findings

• SVM with radial basis kernel functions based on Kullback-Leibler and Rényi distances achieved the highest classification accuracy across all scenarios, outperforming MDC.
• The Chi-Square distance performed poorly due to numerical instability and was not recommended for use in either MDC or SVM frameworks.
• MDC performance decreased with increasing intra-class variability (i.e., fewer classes), while SVM performance improved under such conditions.
• SVM with OAA strategy required more computational time than OAO, but provided better separation of classes like AA and PS in complex scenarios.
• In the first scenario (seven classes), both MDC and SVM struggled to distinguish agricultural classes (A1, A2, A3), but SVM showed superior performance in subsequent scenarios.
• SVM was more robust than MDC in handling imperfect training samples, particularly in scenarios with overlapping or ambiguous classes, as demonstrated in actual data experiments.