Polsar Image Classification Method Based on Fusion of FCN and Sparse-Low Rank Subspace Representation

Abstract

The invention discloses a PolSAR image classification method based on fusion of FCN and sparse-low-rank subspace representation, which fully utilizes the powerful ability of FCN to automatically learn nonlinear deep multi-scale spatial features from PolSAR data itself, and based on sparse-low rank subspace representation. The linear dimensionality reduction algorithm for low-rank graph embedding can simultaneously capture the advantages of local and global structural information of PolSAR data in a low-dimensional subspace, and effectively fuse FCN and sparse-low-rank subspace representations to solve the problem of PolSAR images. Non-linear deep spatial feature extraction problems, dimensionality reduction problems, and effective fusion of polarization information and spatial information, and effectively solve the classification problem of PolSAR images. The fused multi-level subspace features obtained by the present invention contain various types of information, including linear and nonlinear, shallow and deep, local and global, and polarization and multi-scale space, so It is highly discriminative and can greatly improve the classification accuracy of PolSAR images.

Description

technical field [0001] The invention belongs to the technical field of image processing, and in particular relates to a PolSAR image classification method represented by a fully convolutional neural network and a sparse-low rank subspace. Background technique [0002] Synthetic Aperture Radar (SAR), with its almost all-day and all-weather imaging capability, plays an important role in various ground observation applications such as agriculture and forestry management, urban planning, disaster monitoring and environmental protection. Fully polarimetric SAR (Polarimetric SAR, PolSAR) is an advanced SAR system that can measure the complex scattering matrix of a medium. The scattering matrix integrates the amplitude, phase and polarization characteristics of the target, so it can greatly improve the ability to distinguish different ground objects when used in image classification tasks. PolSAR image classification is an important part of remote sensing image interpretation, and...

AI Technical Summary

Problems solved by technology

To sum up, there are four main challenges: (1) the lack of high-quality, well-labeled remote sensing data; (2) the complexity of remote sensing data hinders the learning of robust discriminative features; (3) the How to migrate between data; (4) How to choose the appropriate model depth in the case of limited computing time and training data

Therefore, it may not be realistic to rely entirely on deep learning to solve PolSAR image classification problems at present

Images