CN106022355B

CN106022355B - High spectrum image sky based on 3DCNN composes joint classification method

Info

Publication number: CN106022355B
Application number: CN201610301687.1A
Authority: CN
Inventors: 李映; 张号逵; 曹莹
Original assignee: Northwestern Polytechnical University
Current assignee: Northwestern Polytechnical University
Priority date: 2016-05-09
Filing date: 2016-05-09
Publication date: 2019-03-29
Anticipated expiration: 2036-05-09
Also published as: CN106022355A

Abstract

The invention relates to a 3DCNN-based hyperspectral image space-spectrum joint classification method. Aiming at the characteristic that hyperspectral image data is a three-dimensional structure, a three-dimensional convolution network suitable for hyperspectral images is constructed to complete the hyperspectral image space-spectrum joint classification. First, from the original hyperspectral image, a data block in a certain neighborhood range centered on the pixel to be classified is extracted as the initial spatial spectral feature, and the constructed 3DCNN network is trained in combination with the label of the pixel to be classified. Then, the trained 3DCNN is used to complete the space-spectrum joint classification of hyperspectral images. The beneficial effects are: 1) the problem of complex processing of spectral space dimensionality reduction or compression in the existing classification technology is solved; 2) a 3DCNN suitable for hyperspectral image data of three-dimensional structure is constructed, and the hyperspectral image is fully utilized Rich information and saves the trouble of artificially presetting features; 3) 3DCNN-based hyperspectral image space-spectrum joint classification method. 4) The classification accuracy of hyperspectral images is improved.

Description

High spectrum image sky based on 3DCNN composes joint classification method

Technical field

The invention belongs to remote sensing information process technical fields, are related to a kind of classification method of high spectrum image, and in particular to A kind of high spectrum image sky spectrum joint classification method based on 3DCNN.

Background technique

High-spectrum remote sensing spectral resolution is high, imaging band is more, contains much information, and obtains extensively in remote sensing application field Using.Classification hyperspectral imagery technology is highly important content in Hyperspectral imagery processing technology, mainly includes feature extraction And classification two parts, wherein the extraction feature from former high spectrum image, which influences the nicety of grading of high spectrum image Huge: the strong robustness of characteristic of division can greatly improve nicety of grading；On the contrary, the poor characteristic of division of robustness then can be bright It is aobvious to reduce classifying quality.

In recent years, deep learning was made outstanding achievements in terms of feature extraction, to improve classification hyperspectral imagery precision, various depths Degree model is introduced in the classification of high spectrum image, and on the basis of spectrum signature, introduces space characteristics, utilize depth Model is practised, the autonomous empty spectrum signature for extracting high spectrum image effectively raises classification hyperspectral imagery precision.

However, existing these extract these methods of high spectrum image sky spectrum signature using depth model, it is empty extracting The sufficiently complex elder generation of way when spectrum signature generally requires first to carry out former high spectrum image the dimensionality reduction on spectral space, then by dimensionality reduction Information later obtains sky spectrum signature in conjunction with spectrum information.Dimension-reduction treatment is computationally intensive, and has lost certain spectrum information, influences Precision.

Summary of the invention

Technical problems to be solved

In order to avoid the shortcomings of the prior art, the present invention proposes a kind of high spectrum image sky spectrum connection based on 3DCNN Classification method is closed, overcomes and needs to carry out the pretreatment such as complicated spectral space dimensionality reduction when extracting space characteristics, make full use of spectrum information And spatial information, it is autonomous to extract high spectrum image depth sky spectrum signature in conjunction with the advantage of deep learning autonomous learning, improve classification Precision.

Technical solution

A kind of high spectrum image sky spectrum joint classification method based on 3DCNN, it is characterised in that steps are as follows:

Step 1: input hyperspectral image data is normalized；

Step 2 takes original empty spectrum signature: adjacent from the n × n × L extracted centered on pixel to be sorted in high spectrum image Data block P within the scope of domain_n×n×L, using the data block as the original empty spectrum signature for the pixel for being located at data block center；

Step 3: in the data containing label extracted in step 2, randomly selecting the number of half or less than half According to the data as training 3DCNN

Step 4: the network of 3DCNN is constructed, network overall structure is divided into two parts, and first part includes one layer of input layer, Connection excitation operation layer, excitation operation layer are carried out using unsaturation excitation function ReLU after every layer of convolutional layer of two layers of Three dimensional convolution layer Excitation operation；Second part includes one layer of full articulamentum, one layer of softmax classification layer；Network overall structure includes seven layers, network Integral operation include preceding to operation, reversed derivation, convolution kernel updates three parts operation；

Step 5 is trained 3DCNN using training data: stochastic gradient descent method is used in network training data Training network parameter, the 3DCNN can independently extract the empty spectrum signature of high spectrum image and complete to classify after the completion of training；

Step 6: data to be sorted being inputted into trained 3DCNN, high spectrum image sky is completed and composes joint classification.

Beneficial effect

A kind of high spectrum image sky based on 3DCNN proposed by the present invention composes joint classification method, for high spectrum image The characteristics of data are three-dimensional structure, the Three dimensional convolution network that building is suitable for high spectrum image complete high spectrum image sky spectrum joint Classification.Firstly, from original high spectrum image, extract in certain contiguous range using centered on pixel to be sorted data block as Initial sky spectrum signature, and the 3DCNN network for combining the label training of pixel to be sorted to build.Then, utilization is trained 3DCNN completes high spectrum image sky and composes joint classification.

The beneficial effects of the present invention are: 1) it solves and needs to carry out spectral space dimensionality reduction or pressure in existing sorting technique The problem of complex process of contracting；2) 3DCNN of the hyperspectral image data suitable for three-dimensional structure is constructed, is on the one hand sufficiently sent out The ability of the autonomous extraction feature of deep learning has been waved, it is on the other hand autonomous to extract depth characteristic, take full advantage of high spectrum image Information abundant simultaneously eliminates the trouble for artificially presetting feature；3) the high spectrum image sky based on 3DCNN composes joint classification Method extends the application range of deep learning, also provides new approaches for classification hyperspectral imagery.4) EO-1 hyperion is improved Image classification accuracy.

Detailed description of the invention

Fig. 1: the empty spectrum joint hyperspectral image classification method flow chart based on 3DCNN

Specific embodiment

Now in conjunction with embodiment, attached drawing, the invention will be further described:

Step 1 inputs hyperspectral image data, according to formula pairBehaviour is normalized to data Make.Wherein x_ijsIndicate a pixel in high spectrum image, i, j respectively indicate the coordinate that the pixel is located in high spectrum image Position, s indicate that the spectral coverage of high spectrum image, existing high spectrum image generally comprise 100-240 spectral coverage, x_··smax、x_··smin Respectively indicating indicates three-dimensional high spectrum image in the maximum value and minimum value of s wave band.

Step 2 extracts original empty spectrum signature, from the certain neighborhood extracted centered on pixel to be sorted in high spectrum image Data block P in range_n×n×L, n indicates the size of neighborhood block, generally takes 5 or 7, and L indicates spectral coverage sum, data block P_n×n×LIt is n The three-dimensional structure of × n × L, the data block are the original empty spectrum signatures of the pixel positioned at data block center.

A certain amount of data for containing label are randomly selected in the data that step 3 is extracted from step 2 as training The data of 3DCNN, the general data for choosing the half or less than half that have label data total amount are as training data.

Step 4 building 3DCNN is simultaneously trained 3DCNN using training data.Network overall structure is divided into two parts, the A part includes one layer of input layer, and (first layer convolutional layer includes 2 three dimensional convolution kernels, second layer convolution to two layers of Three dimensional convolution layer Layer includes four three dimensional convolution kernels, and the Spatial Dimension of every layer of convolution kernel is set as 3, and spectral Dimensions are set as 2-9), after every layer of convolutional layer Connection excitation operation layer, excitation operation layer carry out excitation operation using unsaturation excitation function ReLU, and it includes five which, which has altogether, Layer network.Second part includes one layer of full articulamentum, and one layer of softmax classification layer, network overall structure includes seven layers.Network Integral operation mainly include before to operation, reversed derivation, convolution kernel updates three parts operation:

4a) forward direction operation is broadly divided into before convolutional layer to operation, to operation before excitation function, to operation three before classifier Point, wherein to the formula of operation before middle convolutional layer are as follows:

After indicating convolution algorithm, i-th layer of network, jth opens the value of position (x, y, z) on characteristic pattern.P_i、Q_i、R_iTable Showing the size of convolution kernel space dimension and spectrum dimension, k indicates convolution kernel,I-th layer of network of expression is without the number before convolution operation According to.

The forward direction operation of convolutional layer is completed before utilizing unsaturated excitation function ReLU to complete excitation later to operation, formula Are as follows:

The value of the position (x, y, z) after excitation operation, to the formula of operation before final step classifier are as follows:

C indicates the number of the true classification of current sample data in formula, and a shared D indicates classification sum

4b) reversed derivation corresponds to preceding to operation, also comprising the derivation to convolutional layer, the derivation to excitation function, to point The derivation of class device.It is according to the mathematical formulae on basis to operational formula derivation to before operational formula and excitation function to before convolutional layer It can derive, the practical form that convolution is generally taken during writing code completes the derivation process of convolutional layer, third portion Divide the derivation formula to classifier are as follows:

O_D=cIt indicates that teacher signal is the one-dimensional vector that dimension is D, is d in current true class number, is i.e. is taken at d=c Value be 1 remaining value is 0. everywhere

4c) convolution kernel update is completed in reversed derivative operation, and the local derviation for calculating convolution kernel later carries out convolution kernel The single stepping of update, more new formula are as follows:

k_l+1=k_l+v_l+1

L indicates the number of iterations, and ε indicates that learning rate, learning rate generally choose 0.01.

Step 5, using stochastic gradient descent method training network parameter, takes 20-100 in network training data at random every time A sample, the quantity for randomly selecting sample every time are generally to choose data to be sorted depending on the classification number of data to be sorted The integral multiple of classification number, the 3DCNN can independently extract the empty spectrum signature of high spectrum image and complete to classify after training is completed.

Data to be sorted are inputted trained 3DCNN by step 6, are completed high spectrum image sky and are composed joint classification.

Claims

1. a 3DCNN-based hyperspectral image space-spectrum joint classification method is characterized in that the steps are as follows:

Step 1: Normalize the input hyperspectral image data;

Step 2. Take the original empty spectrum feature: extract the data block P _n×n×L in the neighborhood of n×n×L centered on the pixel to be classified from the hyperspectral image, and use the data block as the data block located in the data block. The original empty spectral feature of the pixel at the center position; L represents the total number of spectral segments, and n represents the size of the neighborhood block;

Step 3: In the data containing labels extracted in step 2, randomly select half or less of the data as the data for training 3DCNN;

Step 4: Build a 3DCNN network. The overall structure of the network is divided into two parts. The first part contains an input layer, two three-dimensional convolution layers, and each convolution layer is connected to the excitation operation layer. The excitation operation layer adopts the unsaturated excitation function ReLU Perform the excitation operation; the second part includes a fully connected layer and a softmax classification layer; the overall structure of the network includes seven layers, and the overall operation of the network includes forward operation, reverse derivation, and convolution kernel update three-part operations;

Step 5. Use the training data to train the 3DCNN: the stochastic gradient descent method is used to train the network parameters on the network training data. After the training is completed, the 3DCNN can autonomously extract the empty spectral features of the hyperspectral image and complete the classification;

Step 6: Input the data to be classified into the trained 3DCNN to complete the joint classification of hyperspectral images.