Hyperspectral remote sensing image classification method based on the combination of six-layer convolutional neural network and spectral-space information

Abstract

The present invention discloses a hyperspectral remote sensing image classification method based on the combination of six-layer convolutional neural network and spectral-space information. The method comprises: selecting the hyperspectral remote sensing image data of a certain number of bands; performing space mean-filtering on the two-dimensional image data of each selected band and then converting the format of the multi-band data corresponding to each pixel element; converting the one-dimensional vector into a square matrix, meaning that each pixel elements corresponds to a square matrix data; then, designing a six-layer classifier based on the deep learning template with an input layer, a first convolution layer, a largest pooling layer, a second convolution layer, a full connection layer and an output layer; extracting the square matrix data corresponding to several pixel elements as a training set to be inputted into the classifier and training the classifier; extracting the square matrix data corresponding to several pixel elements as a training set to be inputted into the trained classifier; observing the output classification result of the trained classifier; comparing with the real classification information; and verifying the performances of the trainer. With the method of the invention, higher classification accuracy can be obtained than from the currently available 5-CNN method.

Description

technical field [0001] The invention belongs to the field of remote sensing image processing, and relates to a classification method for different object categories in a hyperspectral image, in particular to a hyperspectral remote sensing image classification method based on a six-layer convolutional neural network and spectral-spatial information combination. Background technique [0002] The current hyperspectral remote sensing data can obtain dozens or even hundreds of spectral band information at the same time, and the rich spectral information greatly improves the ability to identify and distinguish various types of ground objects. Moreover, with the improvement of the spatial resolution of hyperspectral sensors, researchers can use hyperspectral remote sensing images to analyze ground objects with small spatial structures. Due to the rich surface information of hyperspectral remote sensing images, wide coverage, and multi-temporal characteristics, its application techn...

AI Technical Summary

Problems solved by technology

Aiming at the problem of insufficient labeled samples caused by the difficulty and high cost of labeling samples in hyperspectral remote sensing images, the existing research proposes a semi-supervised classification model using a stacked denoising autoencoder, which is combined with an edge-preserving denoising filtering algorithm. Combined with hyperspectral remote sensing images for classification [Reference: Wang Qiaoyu. Hyperspectral remote sensing image classification based on deep learning [D]. Huaqiao University, 2016]

Among the methods proposed by existing studies, some use deep learning methods, but only consider the spectral information without combining the spatial information of the image, and some combine spectral-spatial information, but do not use the convolutional neural network structure as a classifier. Without combining the spatial filtering algorithm with the convolutional neural network, the classification effect is limited

Images