A spectral-spatial dimension reduction method for tensor hyperspectral images based on deep convolutional neural networks

Abstract

The invention discloses a tensor hyperspectral image spectrum-space dimensionality reduction method based on a deep convolutional neural network. The method comprises steps of: in view that it may significantly increase the parameter space of the deep convolutional neural network to directly use high-band tensor data, performing dimensionality reduction on the waveband of a normalized hyperspectral image by introducing a maximum likelihood intrinsic dimensionality estimation algorithm and principal component analysis to obtain a low-band hyperspectral image; converting the low-band hyperspectral image into a tensor low-band hyperspectral image by means of a window field, and keeping the spectrum and space information of each pixel; and performing spectrum-space dimensionality reduction on the tensor low-band hyperspectral image by means of the deep convolutional neural network in order that a characteristic subjected to the dimensionality reduction includes spectrum information and space information. The tensor hyperspectral image spectrum-space dimensionality reduction method may acquire a high overall classification precision and Kappa coefficient by using the spectrum characteristic and space field characteristic of the hyperspectral data.

Description

technical field [0001] The invention belongs to the technical field of hyperspectral remote sensing image processing, and in particular relates to a tensor-type hyperspectral image spectral-space dimensionality reduction method based on a deep convolutional neural network. Background technique [0002] In recent years, with the rapid development of remote sensing technology, it has become very easy to obtain high-precision and high-resolution remote sensing images. Hyperspectral images collected by hyperspectral image sensors can provide rich band information and spatial information, and have strong Therefore, it is widely used in environmental monitoring, vegetation classification, crop growth monitoring and other fields. Classifying each pixel in hyperspectral remote sensing images is one of the common methods in these applications. In recent years, some discriminant-based methods in the field of machine learning have been successfully applied to hyperspectral image class...

AI Technical Summary

Problems solved by technology

Typical vector-based dimensionality reduction algorithms include principal component analysis (PCA), linear discriminant analysis, and local-preserving projection. Although hyperspectral images have rich band information and can represent hyperspectral images to a certain extent, this idea There are still inherent defects: 1) This idea is based on the assumption that "adjacent pixels are independent of each other", but there is a strong correlation between adjacent pixels in hyperspectral images, so this assumption is not valid; 2 ) Converting a three-dimensional hyperspectral image into one-dimensional vector data will destroy the internal structure of the image, resulting in loss of information; 3) vectorizing the hyperspectral image will lead to the problem of "dimension disaster", etc.

Hu et al. introduced DCNN into hyperspectral image classification task for the first time, but their work is based on vector data and adopts 1D convolution mode. This idea not only does not consider the spatial information of hyperspectral image, but also cannot give full play to DCNN. feature extraction capability

Images