CN105320965A - Hyperspectral image classification method based on spectral-spatial cooperation of deep convolutional neural network - Google Patents

Abstract

The present invention relates to a hyperspectral image classification method based on spectral-spatial cooperation of a deep convolutional neural network, which leads the conventional deep convolutional neural network applied to a two-dimensional image into the three-dimensional hyperspectral image classification problem. Firstly, the convolutional neural network is trained by using a small volume of label data, and a spectral-spatial feature of a hyperspectral image is autonomously extracted by using the network without carrying out any compression and dimensionality reduction processing; then, a support vector machine (SVM) classifier is trained by using the extracted spectral-spatial feature so as to classify an image; and finally, the trained neural network is combined with the trained classifier, the neural network extracts a spectral-spatial feature of a to-be-classified target and the classifier determines a specific category of the extracted spectral-spatial feature so as to acquire a structure (DCNN-SVM) that can autonomously extract the spectral-spatial feature of the hyperspectral image and carry out classification to the spectral-spatial feature, thereby forming a set of hyperspectral image classification method.

Description

Based on the hyperspectral image classification method of the sky spectrum associating of degree of depth convolutional neural networks

Technical field

The invention belongs to remote sensing information process technical field, relate to a kind of sorting technique of high spectrum image, particularly relate to the hyperspectral image classification method of a kind of sky based on degree of depth convolutional neural networks spectrum associating.

Background technology

High-spectrum remote sensing spectral resolution is high, imaging band is many, contain much information, and is used widely in remote sensing application field.Classification hyperspectral imagery technology plays an important role in such applications, from former high spectrum image, wherein extract the feature being used for classifying, the nicety of grading impact of this step on high spectrum image is huge, and the strong robustness of characteristic of division, significantly can improve nicety of grading; On the contrary, the characteristic of division that robustness is poor then obviously can reduce classifying quality.

In recent years, degree of depth study is made outstanding achievements in feature extraction, for improving classification hyperspectral imagery precision, SAE, DBN even depth model in degree of depth study is introduced in the classification of high spectrum image, and on the basis of spectrum signature, introduce space characteristics, utilize degree of depth learning model, the empty spectrum signature of autonomous extraction high spectrum image, effectively raises classification hyperspectral imagery precision.

But, existing these utilize depth model to extract these methods of the empty spectrum signature of high spectrum image, the very complicated elder generation of way when extracting empty spectrum signature, often need first to carry out the dimensionality reduction on spectral space to former high spectrum image, then the information after dimensionality reduction is combined with spectrum information obtains sky spectrum signature.Dimension-reduction treatment calculated amount is large, and have lost certain spectrum information, affects precision.

Summary of the invention

The technical matters solved

In order to avoid the deficiencies in the prior art part, the present invention proposes the hyperspectral image classification method of a kind of sky based on degree of depth convolutional neural networks spectrum associating.One is overcome when extracting space characteristics to need to carry out complicated spectral space dimensionality reduction, and two is make full use of spectrum information, and empty spectrum signature obtains by learning by oneself in data.

Technical scheme

Based on a hyperspectral image classification method for the sky spectrum associating of degree of depth convolutional neural networks, it is characterized in that step is as follows:

Step 1: adopt high spectrum image is normalized, wherein: ij denotation coordination position, s represents spectral coverage, x _max, x _minrepresent the maximal value in three-dimensional high-spectral data and minimum value respectively;

Step 2, extract original empty spectrum signature: center pixel and eight neighborhood pixel totally nine pixel vectors of extracting the high spectrum image after normalized as the original empty spectrum signature of center pixel being positioned at (i, j) position;

Step 3: randomly draw the data of data as training convolutional neural networks containing label of more than 6% from the data that step 2 extracts;

Step 4, structure convolutional neural networks: the input of network is the original empty spectrum signature extracted in step 2, the convolution kernel of network is one-dimensional vector, every layer of convolutional layer back connects one deck pond layer, selects 2 ~ 3 layers of convolutional layer, connect softmax layer after the full articulamentum of network according to spectral space dimension;

Step 5: with the training data extracted in step 3, utilizes stochastic gradient descent algorithm training convolutional neural networks, enables convolutional neural networks independently extract the empty spectrum signature of high spectrum image; Remove the softmax layer at the convolutional neural networks end trained, retain full articulamentum and part before, make convolutional neural networks as feature extractor;

Step 6: with feature extractor, feature extraction is carried out to whole high spectrum image and obtain the empty spectrum signature of the degree of depth;

Step 7: operation is normalized to the empty spectrum signature of the degree of depth extracted, the empty spectrum signature of the degree of depth that the training data of convolutional neural networks is obtained by feature extractor is as sorter training data;

Step 8, utilize sorter training data to train a SVM classifier, view data to be sorted is obtained the empty spectrum signature of the degree of depth through feature extractor, by empty for degree of depth spectrum signature through SVM classifier, the result of SVM classifier output category.

Beneficial effect

The hyperspectral image classification method of a kind of spectrum of the sky based on degree of depth convolutional neural networks associating that the present invention proposes, is incorporated into three-dimensional classification hyperspectral imagery problem by traditional application degree of depth convolutional neural networks on 2d.First, utilize a small amount of label data, training convolutional neural networks, and utilize this network independently to extract the empty spectrum signature of high spectrum image, without any need for the process of compression dimensionality reduction; Then, utilize empty spectrum signature Training Support Vector Machines (SVM) sorter extracted, image is classified; Finally, in conjunction with the neural network trained and the sorter trained, neural network extracts the empty spectrum signature of target to be sorted, the specific category of the empty spectrum signature extracted determined by sorter, obtain an empty spectrum signature independently can extracting high spectrum image and to its structure of classifying (DCNN-SVM), thus form the method for a set of classification hyperspectral imagery.

Beneficial effect of the present invention is: solve the problem needing the complex process of carrying out spectral space dimensionality reduction or compression in existing sorting technique, take full advantage of spectrum information, from the empty spectrum signature of principal and subordinate's data learning, improves nicety of grading.Simultaneously introduce SVM classifier at sorting phase, solve the problem that nicety of grading under small sample is not high.

Accompanying drawing explanation

Fig. 1: based on the hyperspectral image classification method process flow diagram of the sky spectrum associating of CNN

Embodiment

Now in conjunction with the embodiments, the invention will be further described for accompanying drawing:

Step 1 inputs high light figure view data, according to formula operation is normalized to data.Wherein ij denotation coordination position s represents spectral coverage, is generally 100-240 spectral coverage, x _max, x _minrepresent the maximal value in three-dimensional high-spectral data and minimum value respectively.

Step 2 extracts original empty spectrum signature, and by high spectrum image, center pixel and eight neighborhood pixel have nine pixel vectors altogether extract the original empty spectrum signature as the center pixel being positioned at (i, j) position.

The data of a small amount of data containing label as training CNN are randomly drawed in the data that step 3 extracts from step 2.

Step 4 builds convolutional neural networks, and with the original empty spectrum signature extracted in step 2 for input, adopt one-dimensional vector as convolution kernel, every layer of convolutional layer back connects one deck pooling layer, suitably selects the convolution number of plies according to spectral space dimension.Connect softmax layer after the full articulamentum of network and calculate output, feedback error.The quantity of convolutional layer is two to three layers, and the convolution kernel quantity of every layer is 10-20, the selection of dimension of convolution kernel 3,5,7 proper.

4a) convolutional layer, convolutional layer forward operation formula is:

$x_j^l=f\left(\underset{i\∈M_j}{\mathrm{\Σ}}{x}_i^{l-1}*k_{ij}^l+\right)b_j^l$

Convolutional layer reversal error propagation formula is:

${\mathrm{\δ}}_j^l={\mathrm{\β}}_j^{l+1}\left(f^{\′}\left(u_j^l\right)oup\left({\mathrm{\δ}}_j^{l+1}\right)\right)$

In above-mentioned formula, l represents the number of plies, and ij represents the mapping numbering of last layer and current layer respectively, the convolution kernel that expression can learn, b represents bigoted, M _jrepresent the selection to input, in the present invention, be adopt the full structure (mapping of current layer is all connected with all mappings of front one deck) connected, * represents convolution operation, and o represents by element multiplication, and up () represents up-sampling operation.In sensitivity after calculating, convolution kernel and biased local derviation can be calculated according to the following equation:

$\frac{\∂E}{\∂b_j}=\underset{u,v}{\Σ}{\left({\mathrm{\δ}}_j^l\right)}_{uv}$

$\frac{\∂E}{\∂k_{lj}^l}=\underset{u,v}{\Σ}{\left({\mathrm{\δ}}_j^l\right)}_{uv}{\left(p_i^{l-1}\right)}_{uv}$

Be that renewable convolution kernel is with biased after calculating convolution kernel and biased local derviation.

4b) pooling layer, the forward operation of this layer is the operation of simple down-sampling, and reversal error propagation formula is:

${\mathrm{\δ}}_i^l={\mathrm{\Σ}}_{j=1}^M{\mathrm{\δ}}^{l+1}*k_{ij}$

If lower one deck is not convolutional layer, so error propagation mode is identical with BP network error circulation way.

4c) excitation function, in order to improve training speed, have employed undersaturated excitation function in the present invention:

f(x)＝max(0，x)

Compared with saturated arctan function, sigmoid function etc., this unsaturated function convergence faster, training gets up more to save time.

4d) softmax layer, forward computing formula and the local derviation computing formula of this layer are as follows:

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

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

C in above-mentioned formula represents the numbering of the true classification of current sample data, a total D classification.Step 5 adopts stochastic gradient descent method training network parameter in network training data, gets 20-25 sample at random at every turn, and after having trained, this degree of depth convolutional network independently can extract the empty spectrum signature of high spectrum image.

Step 6 removes the softmax layer of the convolutional neural networks trained, retain full articulamentum and former part, the part remained can, as empty spectrum signature extraction apparatus, allow all hyperspectral image data through this network, and the output of full articulamentum is the empty spectrum signature of the degree of depth learning to arrive.

Step 7 is normalized operation to the empty spectrum signature of the degree of depth extracted.

The empty spectrum signature of the degree of depth that the network training data chosen in step 3 are obtained by convolutional network by step 8 is as sorter training data, utilize these data to train RBF-SVM sorter, the RBF-SVM sorter of having trained and the convolutional network removing softmax layer together form one independently can extract the empty spectrum signature of high spectrum image and the structure of classifying to it.

Claims (1)

1., based on a hyperspectral image classification method for the sky spectrum associating of degree of depth convolutional neural networks, it is characterized in that step is as follows:

Step 1: adopt high spectrum image is normalized, wherein: ij denotation coordination position, s represents spectral coverage, x _max, x _minrepresent the maximal value in three-dimensional high-spectral data and minimum value respectively;

Step 2, extract original empty spectrum signature: center pixel and eight neighborhood pixel totally nine pixel vectors of extracting the high spectrum image after normalized as the original empty spectrum signature of center pixel being positioned at (i, j) position;

Step 3: randomly draw the data of data as training convolutional neural networks containing label of more than 6% from the data that step 2 extracts;

Step 4, structure convolutional neural networks: the input of network is the original empty spectrum signature extracted in step 2, the convolution kernel of network is one-dimensional vector, every layer of convolutional layer back connects one deck pond layer, selects 2 ~ 3 layers of convolutional layer, connect softmax layer after the full articulamentum of network according to spectral space dimension;

Step 5: with the training data extracted in step 3, utilizes stochastic gradient descent algorithm training convolutional neural networks, enables convolutional neural networks independently extract the empty spectrum signature of high spectrum image; Remove the softmax layer at the convolutional neural networks end trained, retain full articulamentum and part before, make convolutional neural networks as feature extractor;

Step 6: with feature extractor, feature extraction is carried out to whole high spectrum image and obtain the empty spectrum signature of the degree of depth;

Step 7: operation is normalized to the empty spectrum signature of the degree of depth extracted, the empty spectrum signature of the degree of depth that the training data of convolutional neural networks is obtained by feature extractor is as sorter training data;

Step 8, utilize sorter training data to train a SVM classifier, view data to be sorted is obtained the empty spectrum signature of the degree of depth through feature extractor, by empty for degree of depth spectrum signature through SVM classifier, the result of SVM classifier output category.