Land cover classification of PolSAR image using tensor representation and learning

Mingliang Tao,Jia Su,Ling Wang

doi:10.1117/1.jrs.13.016516

Abstract

We propose a tensor representation for polarimetric synthetic aperture radar data and extend the usage of tensor learning technique for feature dimension reduction (DR) in image classification. Under the tensor algebra framework, each pixel is modeled as a third-order tensor object by combining multiple polarimetric features and incorporating neighborhood spatial information together. A set of training tensors are determined according to the prior knowledge of the ground truth. Then a tensor learning technique, i.e., multilinear principal component analysis, is applied on the training tensors set to find a tensor subspace that captures most of the variation in the original tensor objects. This process serves as a feature DR step, which is critical for improving the subsequent classification accuracy. Further, the projected tensor samples after DR are fed to the k-nearest neighbor classifier for supervised classification. The performance is verified in both simulated and real datasets. The extracted features are more discriminative in the feature space, and the classification accuracy is significantly improved by at least 10% compared with other existing matrix-based methods.

Highlights

Polarimetric synthetic aperture radar (PolSAR) has been an important instrument for active remote sensing since it can provide scattering information under different combinations of wave polarizations.[1]
Our goal is to investigate multifeatures combination and incorporate spatial information together within the tensor algebra framework and to develop a pixel-based feature dimension reduction (DR) method based on the tensor learning techniques for improving the accuracy of PolSAR land cover classification
Since our work mainly focuses on the DR process, the rationality behind the choice of k-nearest neighbor (KNN) is to show that the extracted features by the proposed method are separable so that even a simple classifier can achieve a satisfactory result

Summary

Introduction

Polarimetric synthetic aperture radar (PolSAR) has been an important instrument for active remote sensing since it can provide scattering information under different combinations of wave polarizations.[1]. The first class is based on appropriate statistical modeling of PolSAR data. The most well-known method is the Wishart classifier proposed by Lee et al.,[2] which derives an optimal Bayesian classifier based on the assumption that scattering vectors from a homogeneous region follow a complex joint Gaussian distribution. The performance deteriorates in heterogeneous regions because of inaccurate statistical model description, and more refined statistical models are needed.[3,4] Some advanced non-Gaussian models are investigated for characterizing heterogeneity of the scattering medium by incorporating the texture parameter,[5,6] in which the classification accuracy is improved using more representative statistical model of the data

Objectives

Methods

Results

Discussion

Conclusion