CN106022355A - 3DCNN (three-dimensional convolutional neural network)-based high-spectral image space spectrum combined classification method - Google Patents

Abstract

The invention relates to a 3DCNN (three-dimensional convolutional neural network)-based high-spectral image space spectrum combined classification method. According to the feature that high-spectral image data is of a three-dimensional structure, the three-dimensional convolutional network suitable for a high-spectral image is constructed and high-spectral image space spectrum combined classification is completed. First of all, data blocks within a certain neighborhood scope, taking a pixel to be classified as a center, are extracted from an original high-spectral image as initial space spectrum features, and through combination with a label of the pixel to be classified, the constructed 3DCNN is trained. Then, by use of the trained 3DCNN, the high-spectral image space spectrum combined classification is completed. The method has the following advantages: 1, the problem of need of complex processing of spectrum space dimension reduction or compression in the prior classification arts is solved; 2, the 3DCNN suitable for the high-spectral image data of the three-dimensional structure is constructed, rich information of the high-spectral image is fully utilized, and the trouble of manual set of features in advance is omitted; 3, the 3DCNN-based high-spectral image space spectrum combined classification method enlarges the application scope of depth learning and also provides a new approach for high-spectral image classification; and 4, the high-spectral image classification precision is improved.

Description

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

Technical field

The invention belongs to remote sensing information process technical field, relate to the sorting technique of a kind of high spectrum image, be specifically related to A kind of high spectrum image sky based on 3DCNN spectrum joint classification method.

Background technology

High-spectrum remote sensing spectral resolution is high, imaging band is many, contain much information, and obtains extensively in remote sensing application field General application.Classification hyperspectral imagery technology is highly important content in Hyperspectral imagery processing technology, mainly comprises spy Levying extraction and classification two parts, wherein extraction feature from former high spectrum image, high spectrum image is divided by this step The impact of class precision is huge: the strong robustness of characteristic of division, it is possible to nicety of grading is greatly improved；On the contrary, robustness is poor Characteristic of division then can substantially reduce classifying quality.

In recent years, degree of depth study is made outstanding achievements in terms of feature extraction, for improving classification hyperspectral imagery precision, various Depth model is introduced in the classification of high spectrum image, and on the basis of spectrum signature, introduces space characteristics, profit Use degree of depth learning model, the autonomous empty spectrum signature extracting high spectrum image, effectively raise classification hyperspectral imagery essence Degree.

But, these these methods utilizing depth model to extract high spectrum image sky spectrum signature existing, extracting sky The sufficiently complex elder generation of way during spectrum signature, generally requires and former high spectrum image first carries out the dimensionality reduction on spectral space, then will Information after dimensionality reduction is combined with spectrum information and obtains sky spectrum signature.Dimension-reduction treatment is computationally intensive, and have lost certain spectrum Information, affects precision.

Summary of the invention

Solve the technical problem that

In place of the deficiencies in the prior art, the present invention proposes a kind of high spectrum image sky based on 3DCNN spectrum Joint classification method, overcomes and needs when extracting space characteristics to carry out the pretreatment such as complicated spectral space dimensionality reduction, make full use of Spectrum information and spatial information, in conjunction with the advantage of degree of depth independent study study, autonomous extraction high spectrum image degree of depth sky spectrum spy Levy, improve nicety of grading.

Technical scheme

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

Step 1: input hyperspectral image data is normalized；

Step 2, take original empty spectrum signature: from high spectrum image, extract the n × n centered by pixel to be sorted × L adjacent Data block P in the range of territory_n×n×L, this data block is special as the original empty spectrum of the pixel being positioned at data block center Levy；

Step 3: in the data containing label extracted in step 2, randomly draws half or less than half Data are as the data of training 3DCNN

Step 4: build 3DCNN network, network overall structure be divided into two parts, Part I comprise one layer defeated Entering layer, connect excitation operation layer after two-layer every layer of convolutional layer of Three dimensional convolution layer, excitation operation layer uses unsaturation excitation letter Number ReLU carries out excitation operation；Part II comprises one layer of full articulamentum, one layer of softmax classification layer；Network is whole Body structure comprises seven layers, and the integral operation of network comprises forward direction computing, reverse derivation, and convolution kernel updates three part operations；

Step 5, utilize training data that 3DCNN is trained: to use under stochastic gradient in network training data Fall method training network parameter, after having trained, this 3DCNN can independently extract the empty spectrum signature of high spectrum image complete Constituent class；

Step 6: data to be sorted input the 3DCNN trained, completes high spectrum image sky spectrum joint classification.

Beneficial effect

A kind of based on 3DCNN high spectrum image sky spectrum joint classification method that the present invention proposes, for high-spectrum As data are the feature of three dimensional structure, structure is applicable to the Three dimensional convolution network of high spectrum image and completes high spectrum image sky Spectrum joint classification.First, from original high spectrum image, extract the certain contiguous range centered by pixel to be sorted Interior data block is as initial empty spectrum signature, and the 3DCNN network that the label training combining pixel to be sorted builds. Then, trained 3DCNN is utilized to complete high spectrum image sky spectrum joint classification.

The beneficial effects of the present invention is: 1) solve and existing sorting technique needs carry out spectral space dimensionality reduction or pressure The problem of the complex process of contracting；2) 3DCNN of the hyperspectral image data being applicable to three dimensional structure, a side are constructed Face has given full play to the ability of degree of depth independent study extraction feature, on the other hand independently extracts depth characteristic, makes full use of Information that high spectrum image is abundant also eliminates the trouble artificially presetting feature；3) Gao Guang based on 3DCNN Spectrogram picture sky spectrum joint classification method, i.e. extends the range of application of degree of depth study, also provides for classification hyperspectral imagery New approaches.4) improve classification hyperspectral imagery precision.

Accompanying drawing explanation

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

Detailed description of the invention

In conjunction with embodiment, accompanying drawing, the invention will be further described:

Step 1 inputs hyperspectral image data, according to formula pairData are normalized behaviour Make.Wherein x_ijsRepresenting a pixel in high spectrum image, i, j represent that this pixel is positioned in high spectrum image respectively Coordinate position, s represents that the spectral coverage of high spectrum image, existing high spectrum image generally comprise 100-240 spectral coverage, x_··smax、x_··sminRepresent that three-dimensional high spectrum image is in the maximum of s wave band and minima respectively.

Step 2 extracts original empty spectrum signature, extracts the certain neighborhood centered by pixel to be sorted from high spectrum image In the range of data block P_n×n×L, n represents the size of neighborhood block, typically takes 5 or 7, and L represents spectral coverage sum, data block P_n×n×LBeing the three dimensional structure of n × n × L, this data block is in the original empty spectrum signature of the pixel of data block center.

The data that step 3 extracts from step 2 are randomly drawed a certain amount of data containing label as training The data of 3DCNN, typically choose and have the data of the half of label data total amount or less than half as training data.

Step 4 builds 3DCNN and utilizes training data to be trained 3DCNN.Network overall structure is divided into two Part, Part I comprises one layer of input layer, and (ground floor convolutional layer comprises 2 Three dimensional convolution to two-layer Three dimensional convolution layer Core, second layer convolutional layer comprises four three dimensional convolution kernel, and the Spatial Dimension of every layer of convolution kernel is set to 3, and spectrum dimension sets For 2-9), connect excitation operation layer after every layer of convolutional layer, excitation operation layer uses unsaturated excitation function ReLU to carry out Excitation operation, this part comprises altogether five layer networks.Part II comprises one layer of full articulamentum, and one layer of softmax divides Class layer, network overall structure comprises seven layers.The integral operation of network mainly comprises forward direction computing, reverse derivation, convolution Core updates three part operations:

4a) forward direction computing is broadly divided into convolutional layer forward direction computing, excitation function forward direction computing, grader forward direction computing three Point, wherein the formula of middle convolutional layer forward direction computing is:

$u_{ij}^{xyz}=b_{ij}+\underset{p=1}{\overset{P_i}{\Σ}}\underset{q=1}{\overset{Q_i}{\Σ}}\underset{r=1}{\overset{R_i}{\Σ}}{k}_{ij}^{pqr}{x}_i^{(x+p)(y+q)(z+r)}$

After representing convolution algorithm, i-th layer of network, jth opens (x, y, z) value of position on characteristic pattern.P_i、Q_i、R_i Representing convolution kernel space dimension and the size of spectrum dimension, k represents convolution kernel,Represent i-th layer of network without convolution operation it Front data.

The forward direction of convolutional layer has operated and has utilized undersaturated excitation function ReLU to complete to encourage forward direction computing, formula afterwards For:

$y_{ij}^{xyz}=max(0,u_{ij}^{xyz})$

After excitation operation (x, y, z) value of position, the formula of final step grader forward direction computing is:

$y=-\underset{ij}{\Σ}(x_{ijc}-log\underset{d=1}{\overset{D}{\Σ}}{e}^{x_{ijd}})$

In formula, c represents the numbering of the true classification of current sample data, and one has D represents classification sum

4b) reversely derivation, corresponding to forward direction computing, also comprises the derivation to convolutional layer, the derivation to excitation function, to dividing The derivation of class device.Public according to the mathematics on basis to convolutional layer forward direction operational formula and excitation function forward direction operational formula derivation Formula can be derived, and reality typically takes the form of convolution to complete the derivation process of convolutional layer during writing code, The derivation formula of grader is by Part III:

$\frac{dz}{{\mathrm{dx}}_{ijd}}=-\frac{dz}{dy}(O_{d=c}-y_{ijc})$

O_D=cRepresent teacher signal be dimension be the one-dimensional vector of D, be at d, i.e. d=c current true class number Value be 1 remaining everywhere value be 0.

4c) convolution kernel updates is to complete in reverse derivative operation, carries out convolution kernel more after calculating the local derviation of convolution kernel New single stepping, more new formula are as follows:

$v_{l+1}=0.9\·v_l-0.0005\·\mathrm{\ϵ}\·k_l-\mathrm{\ϵ}\·\frac{\∂J}{\∂k}{|}_{k_l}$

k_l+1=k_l+v_l+1

L represents that iterations, ε represent learning rate, and learning rate typically chooses 0.01.

Step 5 uses stochastic gradient descent method training network parameter in network training data, takes 20-100 the most at random Individual sample, randomly selects depending on the quantity of sample is the classification number according to data to be sorted every time, typically chooses to be sorted The integral multiple of data category number, after having trained, this 3DCNN can independently extract the empty spectrum signature of high spectrum image also Complete classification.

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

Claims (1)

1. high spectrum image sky based on a 3DCNN spectrum joint classification method, it is characterised in that step is as follows:

Step 1: input hyperspectral image data is normalized；

Step 2, take original empty spectrum signature: from high spectrum image, extract the n × n centered by pixel to be sorted × L adjacent Data block P in the range of territory_n×n×L, this data block is special as the original empty spectrum of the pixel being positioned at data block center Levy；

Step 3: in the data containing label extracted in step 2, randomly draws half or less than half Data are as the data of training 3DCNN

Step 4: build 3DCNN network, network overall structure be divided into two parts, Part I comprise one layer defeated Entering layer, connect excitation operation layer after two-layer every layer of convolutional layer of Three dimensional convolution layer, excitation operation layer uses unsaturation excitation letter Number ReLU carries out excitation operation；Part II comprises one layer of full articulamentum, one layer of softmax classification layer；Network is whole Body structure comprises seven layers, and the integral operation of network comprises forward direction computing, reverse derivation, and convolution kernel updates three part operations；

Step 5, utilize training data that 3DCNN is trained: to use under stochastic gradient in network training data Fall method training network parameter, after having trained, this 3DCNN can independently extract the empty spectrum signature of high spectrum image complete Constituent class；

Step 6: data to be sorted input the 3DCNN trained, completes high spectrum image sky spectrum joint classification.