Full-polarimetric scattering characteristics prediction from single/dual-polarimetric SAR data using convolutional neural networks

Abstract

Scattering entropy (H), scattering angle (α) and anti-entropy (A) are useful parameters in synthetic aperture radar (SAR) image classification. Usually, full-polarimetric SAR data are needed to extract these parameters. In this study, the authors firstly try to predict these parameters from single/dual-polarimetric SAR data using convolutional neural network. Experiments are done on GF-3 polarised SAR database, and promising results are obtained, where the parameters H and α, the average relative error reached is <10%, the parameter A, the average relative error reached is around 25%, and the classification performance based on predictive parameters is around 80%. Furthermore, the predicting performance using different single- and dual-polarisation is compared. The results and conclusions provide a new clue for the applications of single/dual-polarimetric SAR.

Keywords

• image classification
• radar polarimetry
• synthetic aperture radar
• radar imaging
• convolutional neural nets
• radar computing
• sar image classification
• scattering entropy
• single-dual-polarimetric sar data
• full-polarimetric scattering characteristic prediction
• antientropy
• full-polarimetric sar data
• synthetic aperture radar image classification
• predictive parameters
• average relative error
• gf-3 polarised sar database
• convolutional neural network