Hyperspectral image classification method based on three-dimensional densely connected convolutional neural network

Abstract

The invention provides a hyperspectral image classification method based on a three-dimensional densely connected convolutional neural network. The method is characterized by comprising the followingsteps that step one, the three-dimensional cube data block of a hyperspectral image is inputted; step two, the three-dimensional cube data block is processed by using the three-dimensional dense spectral block so as to obtain an interspectral characteristic graph; step three, the interspectral characteristic graph is processed by using a three-dimensional transition layer so as to obtain a compression characteristic graph; step four, the compression characteristic graph is processed by using a three-dimensional dense space spectral block so as to obtain a space characteristic graph; step five,the predictive tag vector is obtained according to the space characteristic graph through a pooling layer, a compression layer, a dropout layer and a full connection layer; step six, the objective function is determined; step seven, the predictive tag vector is substituted in the objective function so as to obtain the loss of iterative training; step eight, the parameter to be optimized is optimized according to the loss; and step nine, the step one to five, seven and eight are repeated and the parameter to be optimized is optimized for multiple times so as to obtain the predictive tag vectorunder the condition of minimum loss, i.e. the classification result of the hyperspectral image.

Description

technical field [0001] The invention relates to a hyperspectral image classification method, in particular to a hyperspectral image classification method based on a three-dimensional densely connected convolutional neural network. Background technique [0002] Traditional ordinary images with only a small amount of information have only a narrow visible light band. Hyperspectral sensors generally have hundreds of bands, and each band independently absorbs signals within the band range, and generates corresponding two-dimensional images according to different feedback signals of different substances on the spectrum of each band, and the data of all bands is finally Together form a multi-channel 3D data. Therefore, hyperspectral images contain a large amount of information and have many typical applications, such as target detection in hyperspectral images in civil and military fields. Among them, hyperspectral image classification plays a very important role in the field of...

AI Technical Summary

Problems solved by technology

[0005] To sum up, in the existing hyperspectral image classification methods, traditional machine learning methods have complex feature extraction projects, and the classification accuracy cannot reach the most advanced level

However, deep convolutional networks and their extended networks mostly have a long model training time.

Images