BIAS COMPENSATION IN H/a/Α POLARIMETRIC SAR DECOMPOSITION AND ITS IMPLICATION FOR THE CLASSIFICATION

Yahia, Mohamed; Khalfa, Faten; Chabir, Marwa; Aguili, Taoufik

doi:10.2528/pierb16033005

Cited by 2 publications

(1 citation statement)

References 29 publications

(63 reference statements)

Supporting

Mentioning

Contrasting

Order By: Relevance

“…At present, the feature extraction methods for PolSAR images can be broadly classified into two main categories: simple data transformations and polarimetric target decomposition. The methods based on simple data transformations involve the polarized scattering matrix and its corresponding vector forms [4][5][6][7], or mathematical transformations of the scattering matrix such as the polarimetric coherency matrix and polarimetric covariance matrix [8][9][10][11]. The method based on polarization target decomposition is to decompose the polarization coherence matrix or covariance matrix according to different physical mechanisms, and then use the decomposition parameters representing the target scattering or geometric structure to construct features.…”

Section: Introductionmentioning

confidence: 99%

An Innovative PolSAR Image Classification Method Based on Non-Negative Constraints Stacked Sparse Autoencoder Network with Multi-Features Joint Representation Learning

2023

AJCIS

View full text Add to dashboard Cite

This paper proposed a framework based on joint multi-feature representation learning to reduce the inherent speckle phenomenon in Polarimetric Synthetic Aperture Radar (PolSAR) images interfere with the scattering characteristics of land objects. Firstly, the corresponding 6-dimensional real vector is obtained from the covariance matrix of PolSAR data and combined with the polarized feature vector obtained by the polarization decomposition method to improve the differentiation ability of similar features in images. Secondly, the stacked sparse autoencoder (SSAE) is employed, where the non-negative constraint method is incorporated to make the sparse features in the depth space robust by filtering the weights. Finally, a non-negative constrained SSAE model is proposed to effectively accomplish the classification task of PolSAR images. In the experiments, the proposed classification network is trained layer by layer using unlabeled data, the softmax classifier is trained with a small number of labeled pixels. The parameters obtained from the above steps are used as initial parameters to train the entire classification framework with labeled pixels, the resulting well-trained model is used to predict the labels corresponding to pixels in the datasets. Through experiments using the Flevoland and San Francisco Bay datasets, the results demonstrate that the proposed method yields superior classification results compared with traditional SVM, AE, and Gray Level Co-generation Matrix (GLCM) classification methods.

show abstract