CN108256557B - Hyperspectral image classification method combining deep learning and neighborhood integration - Google Patents

Hyperspectral image classification method combining deep learning and neighborhood integration Download PDF

Info

Publication number
CN108256557B
CN108256557B CN201711415902.1A CN201711415902A CN108256557B CN 108256557 B CN108256557 B CN 108256557B CN 201711415902 A CN201711415902 A CN 201711415902A CN 108256557 B CN108256557 B CN 108256557B
Authority
CN
China
Prior art keywords
layer
training
network
sample set
hyperspectral image
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN201711415902.1A
Other languages
Chinese (zh)
Other versions
CN108256557A (en
Inventor
孟红云
张小华
樊宏渊
田小林
朱虎明
曹向海
侯彪
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Xidian University
Original Assignee
Xidian University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Xidian University filed Critical Xidian University
Priority to CN201711415902.1A priority Critical patent/CN108256557B/en
Publication of CN108256557A publication Critical patent/CN108256557A/en
Application granted granted Critical
Publication of CN108256557B publication Critical patent/CN108256557B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
    • G06F18/2135—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on approximation criteria, e.g. principal component analysis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/10—Terrestrial scenes
    • G06V20/13—Satellite images
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/10—Terrestrial scenes
    • G06V20/194—Terrestrial scenes using hyperspectral data, i.e. more or other wavelengths than RGB

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Physics & Mathematics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • General Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Astronomy & Astrophysics (AREA)
  • Remote Sensing (AREA)
  • Multimedia (AREA)
  • Image Analysis (AREA)

Abstract

The invention provides a hyperspectral image classification method combining deep learning and neighborhood integration, which mainly solves the problems of more training samples and poor classification effect in the prior art, and adopts the technical scheme that: selecting different neighborhood scales in the hyperspectral data to obtain a data set combining different spatial information; respectively inputting data sets of different spatial information into different self-coding networks to obtain classification results under different spatial information; connecting the classification results, and training a new automatic encoder network as training data to serve as a final integrated network; connecting the classification results of the self-encoders to the test samples under different spatial information to form the test sample of the integrated network; and inputting the new test sample into the integrated network to obtain a final classification result of the hyperspectral image. The invention has the advantages of less training samples and high classification precision, and can be used for environmental monitoring, land utilization, target identification and the like.

Description

Hyperspectral image classification method combining deep learning and neighborhood integration
Technical Field
The invention belongs to the technical field of digital image processing, and particularly relates to a hyperspectral image classification method which can be used for environment monitoring, land utilization and target identification.
Background
By combining an imaging technology and a spectrum technology, the hyperspectral remote sensing can simultaneously obtain data with continuous space and spectrum. Hyperspectral images are an effective tool in earth surface monitoring and are widely used in agriculture, mineralogy, ground detection, physics, astronomy and environmental science. A common technique in these applications is to classify each pixel in the hyperspectral image.
The classification method of the hyperspectral image mainly comprises a classification method based on spectral information, a classification method based on spatial information and a classification method combining the spatial information and the spectral information, wherein:
the classification method based on the spectral information only utilizes the spectral information of the hyperspectral image for classification, and a decision tree algorithm, a neural network algorithm and the like are commonly used. The methods only consider the spectral information of the pixels, but do not consider the neighborhood information of the pixels, and in fact, the pixels of the high-spectrum image and the adjacent pixels are often in the same class, so that the classification effect obtained by the classification method only depending on the spectral information is very limited.
The hyperspectral image classification method based on spatial information only utilizes the spatial information of hyperspectral images for classification, and typical methods are a feature extraction method based on wavelet analysis and a feature extraction method based on gray level co-occurrence matrix. The method is a feature extraction method based on artificial experience, so the method needs better prior knowledge to obtain a better classification result.
A hyperspectral image classification method based on spatial-spectral combination is a method for classifying by combining hyperspectral pixel spectral information and spatial information. Typical methods include a sparse representation classification method based on space-spectrum combination and a hyperspectral image classification method based on deep learning. The sparse representation classification method based on the combination of the space spectrum is a popular classification algorithm at present, and obtains a good classification effect to a certain extent, but only extracts the shallow feature of the hyperspectral image. The hyperspectral image classification method based on deep learning is a research hotspot in recent years, and is more and more applied to actual classification due to the extremely strong deep feature extraction capability, but the hyperspectral image classification method based on deep learning is greatly restricted because a large number of labeled samples are required for network training and the hyperspectral image labeled samples are deficient.
Disclosure of Invention
The invention aims to provide a hyperspectral image classification method combining deep learning and neighborhood integration, and aims to solve the problems that deep-layer spatial spectrum features of hyperspectral images cannot be well extracted and a large number of training samples are required in the prior art.
In order to achieve the purpose, the technical scheme of the invention comprises the following steps:
(1) inputting a hyperspectral image containing C categories: x ═ X1，x2，……，xi，……，xNAnd randomly selecting 10% of samples from each type of hyperspectral pixels as a training sample set S, and using the rest samples as test samples T, wherein x isiRepresents the ith sample in the hyperspectral image, which is B0The dimension spectral vector i is 1, 2, …, N represents the number of samples of the hyperspectral image, C is more than or equal to 2, B0The number of wave bands of the hyperspectral image is different, and the spectral dimensions of pixels of the image obtained by different hyperspectral imagers are different;
(2) inputting a training sample set S into an automatic encoder network connected with a softmax classifier to perform network training to obtain a trained classification network;
(3) inputting the training sample set S and the test sample T into the trained network to respectively obtain the probability classification results of the training sample set and the test sample set
Figure BDA0001521947210000021
And
Figure BDA0001521947210000022
wherein N is1For the number of training samples, N2C is the number of sample types;
(4) performing dimensionality reduction on the hyperspectral image X by using a principal component analysis method to obtain a dimensionality-reduced hyperspectral image: x '═ X'1,x′2,……,x′i,……,x′NX 'therein'iRepresenting the ith sample of the hyperspectral image after dimensionality reduction, and the dimensionality is represented by B0Reducing to B;
(5) on the dimensionality-reduced hyperspectral image X ', by each sample X'iSelecting a window with the space size of 3 multiplied by 3 as a center, and obtaining a new training sample set S 'and a new test sample set T' which are added with neighborhood information correspondingly;
(6) inputting the new training sample set S' into an automatic encoder network connected with a softmax classifier to perform network training to obtain a new trained network;
(7) inputting the new training sample set S 'and the new test sample set T' into a new trained network to respectively obtain the probability classification results of the new training sample set S
Figure BDA0001521947210000023
And the probability classification result of the newly measured sample book T
Figure BDA0001521947210000024
Wherein N is1For the number of training samples, N2C is the number of sample types;
(8) repeating (5) - (7) to obtain probability classification results of training samples obtained when the selection window size is 5 × 5, 7 × 7, 9 × 9, 11 × 11, 13 × 13, 15 × 15
Figure BDA0001521947210000031
And probabilistic classification of test samples
Figure BDA0001521947210000032
Wherein N is1For the number of training samples, N2C is the number of sample types;
(9) will be provided with
Figure BDA0001521947210000035
Probability classification result of cascading into one total training sample
Figure BDA0001521947210000033
And P isSTraining the network as a training sample set of a new automatic encoder network to obtain a trained integrated network;
(10) will be provided with
Figure BDA0001521947210000036
Probabilistic classification of a total test sample
Figure BDA0001521947210000034
And P isTAnd inputting the test sample set into a trained integrated network to obtain a final classification result.
Compared with the prior art, the invention has the following advantages:
1) the method uses a hyperspectral image classification method combining deep learning and neighborhood integration, effectively extracts deep spatial spectral features of the hyperspectral image, and has strong robustness on classification;
2) the invention provides a hyperspectral image classification method combining deep learning and neighborhood integration, which integrates classification information obtained by a plurality of networks, so that the integrated network inherits the information obtained by the plurality of networks and makes up for the defect of insufficient extraction characteristics of a single network, thereby enabling the integrated network to obtain a good classification effect and solving the problems that a large number of training samples are needed and the classification effect is poor in a common deep learning method.
Drawings
FIG. 1 is a general flow chart of an implementation of the present invention;
FIG. 2 is a sub-flow diagram of the present invention for training a single classification network and obtaining classification results;
FIG. 3 is a diagram of a distribution of real terrain features using images in simulation of the present invention;
FIG. 4 is a graph of the classification results of FIG. 3 for a single network at different scales in the present invention;
fig. 5 is a diagram of the classification results of fig. 3 using the final integration network of the present invention.
Detailed Description
Referring to fig. 1, the implementation steps of the invention are as follows:
step 1, selecting an original training sample set S and an original test sample set T from an input hyperspectral image X.
Inputting a hyperspectral image containing C categories: x ═ X1，x2，……,xi,……,xNAnd randomly selecting 10% of pixels from each type of hyperspectral images as an original training sample set S, and using the rest samples as an original test sample set T, wherein the hyperspectral images use Indian Pine images and images of the university of Pavia in a public data set, and x is xiRepresents the ith sample in the hyperspectral image, which is B0A dimensional spectral vector i equal to 1, 2, … …, N representing the number of samples of the hyperspectral image, B0The number of the wave bands of the hyperspectral images is different, the size of the hyperspectral images is different in images obtained by different hyperspectral imagers, C is larger than or equal to 2, the number of the categories C contained in different hyperspectral images is different, for example, Indian Pine images contain 16 categories, and images of university of Pavia contain 9 categories.
And 2, training the network connected with the automatic encoder of the softmax classifier by using the original training sample set S.
(2a) Inputting an original training sample set S into an automatic encoder network, training a first layer of the network, and obtaining parameters of the trained first layer:
(2a1) taking an original training sample set S as a first layer of an m-layer automatic encoder network as input, wherein m is more than or equal to 2, obtaining hidden layer characteristics of the first layer by using initial parameters of the first layer of the network, and obtaining reconstruction data by using the hidden layer characteristics;
(2a2) continuously adjusting the first layer network parameters to minimize the error between the data of the input layer of the first layer network and the reconstructed data, and obtaining trained first layer network parameters;
(2b) converting the training sample into the hidden layer characteristic of the first layer by using the trained parameters of the first layer;
(2c) taking the hidden layer characteristics of the first layer as the input of the second layer of the network, training the second layer of the network by using the hidden layer characteristics of the first layer in a training mode of the first layer to obtain the parameters of the trained second layer, and converting the input of the second layer into the hidden layer characteristics of the second layer by using the parameters; in the same way, the same strategy is adopted for the following layers until the hidden layer characteristic of the last layer is obtained;
(2d) and (3) training the softmax classifier by taking the hidden layer characteristics of the last layer obtained in the step (2c) as the input of the softmax classifier, and then finely adjusting the whole network to obtain the classification network trained by the original training sample set S.
Step 3, inputting the original training sample set S and the original test sample T into the classification network obtained in the step 2d to respectively obtain the probability classification results of the original training sample set and the original test sample set
Figure BDA0001521947210000041
Wherein N is1For the number of training samples, N2To test the number of samples, C is the number of sample classes.
And 4, obtaining classification information by using a single network.
Referring to fig. 2, the specific implementation of this step is as follows:
(4a) performing dimensionality reduction on the hyperspectral image X by using a principal component analysis method to obtain a dimensionality-reduced hyperspectral image: x '═ X'1,x′2,……,x′i,……,x′NX 'therein'iRepresenting the ith sample of the hyperspectral image after dimensionality reduction, and the dimensionality is represented by B0Reducing to B;
(4b) on the dimensionality-reduced hyperspectral image X ', by each sample X'iSelecting a 3 × 3 window as the center, and combining the windowsThe inner 9 samples are cascaded into a vector to obtain a hyperspectral image X' containing spatial information;
(4c) correspondingly connecting the samples in the hyperspectral image X with the samples in the hyperspectral image X 'containing spatial information respectively to obtain a hyperspectral image X' containing spatial information;
(4d) acquiring a new training sample set S 'and a new test sample set T' according to the positions of the original training sample set S and the original training sample set T in the hyperspectral image X 'in the step (1) and the corresponding positions in the hyperspectral image X';
(4e) inputting the new training sample set S 'into an automatic encoder network connected with a softmax classifier, and carrying out network training according to the following steps to obtain a network trained by the new training sample set S':
(4e1) taking a training sample as the input of a first layer of an m-layer automatic encoder network, wherein m is more than or equal to 2; training the first layer by using a training sample to obtain the parameters of the trained first layer;
(4e2) converting the training sample into the hidden layer characteristic of the first layer by using the trained parameters of the first layer;
(4e3) taking the hidden layer characteristics of the first layer as the input of the second layer of the network, training the second layer of the network by using the hidden layer characteristics of the first layer in a training mode of the first layer to obtain the parameters of the trained second layer, and converting the input of the second layer into the hidden layer characteristics of the second layer by using the parameters; in the same way, the same strategy is adopted for the following layers until the hidden layer characteristic of the last layer is obtained;
(4e4) and (5) training the softmax classifier by taking the hidden layer features of the last layer obtained in the step (4e3) as the input of the softmax classifier, and then finely adjusting the whole network to obtain the trained network.
(4f) Inputting the new training sample set S 'and the new test sample set T' into the network obtained in the step (4e) to respectively obtain the probability classification result of the new training sample set S
Figure BDA0001521947210000051
And the probability classification result of the newly measured sample book T
Figure BDA0001521947210000052
Wherein N is1For the number of training samples, N2To test the number of samples, C is the number of sample classes.
And 5, respectively selecting windows with different sizes to obtain corresponding probability classification results.
(5a) Selecting the window size of 5 multiplied by 5, repeating (4b) - (4f), and obtaining the probability classification result of the training sample set in the 5 multiplied by 5 window
Figure BDA0001521947210000053
And probabilistic classification of test sample sets
Figure BDA0001521947210000054
(5b) Selecting the window size to be 7 multiplied by 7, repeating (4b) to (4f), and obtaining the probability classification result of the training sample set when the window size is 7 multiplied by 7
Figure BDA0001521947210000055
And probabilistic classification of test sample sets
Figure BDA0001521947210000056
(5c) Selecting the window size of 9 multiplied by 9, repeating (4b) - (4f), and obtaining the probability classification result of the training sample set when the window size is 9 multiplied by 9
Figure BDA0001521947210000061
And probabilistic classification of test sample sets
Figure BDA0001521947210000062
(5d) Selecting the window size to be 11 multiplied by 11, repeating (4b) to (4f), and obtaining the probability classification result of the training sample set when the window size is 11 multiplied by 11
Figure BDA0001521947210000063
And summary of test sample setsRate classification results
Figure BDA0001521947210000064
(5e) Selecting a window size of 13 multiplied by 13, repeating (4b) to (4f), and obtaining the probability classification result of the training sample set when the window size is 13 multiplied by 13
Figure BDA0001521947210000065
And probabilistic classification of test sample sets
Figure BDA0001521947210000066
(5f) Selecting a window with the size of 15 multiplied by 15, repeating (4b) to (4f), and obtaining the probability classification result of the training sample set when the window is 15 multiplied by 15
Figure BDA0001521947210000067
And probabilistic classification of test sample sets
Figure BDA0001521947210000068
Figure BDA0001521947210000069
Wherein N is1For the number of training samples, N2To test the number of samples, C is the number of sample classes.
And 6, training the integrated network according to the result of the step 5.
(6a) Classifying the probability of the training sample set obtained in the step 5
Figure BDA00015219472100000610
Cascading to obtain a new sample set:
Figure BDA00015219472100000611
(6b) will PSAnd (3) as a training sample set of a new automatic encoder network, carrying out network training according to the same training mode as the step (2) to obtain a trained integrated network.
And 7, classifying all the test samples by using the trained integrated network to obtain a final result.
(7a) Probability classification result of the test sample set obtained in the step 5
Figure BDA00015219472100000612
Cascading to obtain a probability classification result of a test sample set:
Figure BDA00015219472100000613
(7b) will PTInputting the classification result into the integrated network obtained in the step 6 to obtain a final classification result.
The effect of the present invention can be further illustrated by the following simulation results.
1. Simulation conditions
The hardware platform is as follows: intel (R) core (TM) i5-3210M, 8GB RAM, software platform: MATLAB R2014 a.
The simulation experiment adopts an Indian Pine image obtained by an AVIRIS of a NASA jet propulsion laboratory in 1992 at 6 months in North Indiana as shown in FIG. 3, wherein the image has a real ground object distribution diagram, the image size is 145 multiplied by 145, the total number of the wave bands is 220, the wave bands for removing noise and absorbing the air and water areas are 200, and the total number of the 16 types of ground object information is shown in Table 1.
TABLE 1 Ind Pin image information of 16 kinds of ground objects
Figure BDA0001521947210000071
2. Emulated content and analysis
Simulation 1, selecting neighborhood windows with different sizes, and obtaining classification results of the test sample under the neighborhood windows with different sizes by using the method shown in FIG. 2, wherein the classification results are shown in FIG. 4; wherein:
FIG. 4(a) shows a classification result diagram when the size of the neighborhood window is selected to be 1X 1,
figure 4(b) shows a classification result diagram when the size of the neighborhood window is selected to be 3 x 3,
figure 4(c) shows a diagram of the classification results when the size of the neighborhood window is chosen to be 5 x 5,
figure 4(d) shows a classification result diagram when the size of the neighborhood window is selected to be 7 x 7,
figure 4(e) shows a classification result diagram when the size of the neighborhood window is chosen to be 9 x 9,
figure 4(f) shows a classification result diagram when the size of the neighborhood window is chosen to be 11 x 11,
figure 4(g) shows a graph of the classification results when the size of the neighborhood window is chosen to be 13 x 13,
fig. 4(h) is a diagram showing the classification result when the size of the neighborhood window is selected to be 15 × 15.
Simulation 2, the method of the present invention, i.e., the method shown in fig. 1, is used to classify the test samples, and the result is shown in fig. 5.
As can be seen from fig. 3, 4 and 5, the classification result obtained by using the integrated network in the present invention is better than the classification result of a single network in different scales shown in fig. 4, and has stronger similarity with the Indian Pine reference diagram shown in fig. 3, thereby demonstrating that the present invention effectively improves the classification accuracy of the hyperspectral images.
The classification accuracy is an index for evaluating classification performance, namely the ratio of the number of correctly classified samples to the total number of samples. The higher the classification accuracy, the better the performance of the classification method. The accuracy of the classification result of the single network under different scales shown in fig. 4 and the accuracy of the classification result of the final integrated network shown in fig. 5 are respectively counted, and the results are shown in table 2.
TABLE 2 Classification correctness for Single networks and for Integrated networks under different scales
Figure BDA0001521947210000081
As can be seen from Table 2, the classification accuracy of the final integrated network is far higher than that of a single network under different scales, which shows that the invention has better classification performance.
In summary, compared with a common hyperspectral image classification method, the hyperspectral image classification method has a better classification effect, and compared with other deep learning methods, the hyperspectral image classification method overcomes the defects that a large number of training samples are needed and the classification result is poor, and is suitable for the hyperspectral image classification problem in reality.

Claims (5)

1. A hyperspectral image classification method combining deep learning and neighborhood integration comprises the following steps:
(1) inputting a hyperspectral image containing C categories: x ═ X1，x2，……，xi，……，xNAnd randomly selecting 10% of samples from each type of hyperspectral pixels as a training sample set S, and using the rest samples as a test sample set T, wherein xiRepresents the ith sample in the hyperspectral image, which is B0The dimension spectral vector i is 1, 2, …, N represents the number of samples of the hyperspectral image, C is more than or equal to 2, B0The number of wave bands of the hyperspectral image is different, and the spectral dimensions of pixels of the image obtained by different hyperspectral imagers are different;
(2) inputting a training sample set S into an automatic encoder network connected with a softmax classifier to perform network training to obtain a trained classification network;
(3) inputting the training sample set S and the test sample set T into the trained network to respectively obtain the probability classification results of the training sample set and the test sample set
Figure FDA0001521947200000011
And
Figure FDA0001521947200000012
wherein N is1For the number of training samples, N2C is the number of sample types;
(4) performing dimensionality reduction on the hyperspectral image X by using a principal component analysis method to obtain a dimensionality-reduced hyperspectral image: x' ═ X1',x'2,……,xi',……,x'NIn which xi' represents the ith sample of the hyperspectral image after dimensionality reduction, and the dimensionality is represented by B0Reducing to B;
(5) on the hyperspectral image X' after dimensionality reduction, each sample XiSelecting a window with the space size of 3 multiplied by 3 as the center to obtain a corresponding new training sample set S 'and a new test sample set T' added with neighborhood information;
(6) inputting the new training sample set S' into an automatic encoder network connected with a softmax classifier to perform network training to obtain a new trained network;
(7) inputting the new training sample set S 'and the new test sample set T' into a new trained network to respectively obtain the probability classification results of the new training sample set S
Figure FDA0001521947200000013
And the probability classification result of the newly measured sample book T
Figure FDA0001521947200000014
Wherein N is1For the number of training samples, N2C is the number of sample types;
(8) repeating (5) - (7) to obtain probability classification results of the training sample sets obtained when the selection window sizes are 5 × 5, 7 × 7, 9 × 9, 11 × 11, 13 × 13 and 15 × 15 respectively
Figure FDA0001521947200000015
And probabilistic classification of test sample sets
Figure FDA0001521947200000021
Wherein N is1For the number of training samples, N2C is the number of sample types;
(9) will be provided with
Figure FDA0001521947200000022
Probability classification result of cascading into one total training sample
Figure FDA0001521947200000023
And P isSTraining the network as a training sample set of a new automatic encoder network to obtain a trained integrated network;
(10) will be provided with
Figure FDA0001521947200000024
Probabilistic classification of a total test sample
Figure FDA0001521947200000025
And P isTAnd inputting the test sample set into a trained integrated network to obtain a final classification result.
2. The method of claim 1, wherein the training sample set S is input into the softmax classifier-connected autoencoder network for network training in step (2), and the steps are as follows:
(2a) taking a training sample as the input of a first layer of an m-layer automatic encoder network, wherein m is more than or equal to 2; training the first layer by using a training sample to obtain the parameters of the trained first layer;
(2b) converting the training sample into the hidden layer characteristic of the first layer by using the trained parameters of the first layer;
(2c) taking the hidden layer characteristics of the first layer as the input of the second layer of the network, training the second layer of the network by using the hidden layer characteristics of the first layer in a training mode of the first layer to obtain the parameters of the trained second layer, and converting the input of the second layer into the hidden layer characteristics of the second layer by using the parameters; in the same way, the same strategy is adopted for the following layers until the hidden layer characteristic of the last layer is obtained;
(2d) and (3) training the softmax classifier by taking the hidden layer characteristics of the last layer obtained in the step (2c) as the input of the softmax classifier to obtain a trained classifier, and then finely adjusting the whole network to obtain the trained network.
3. The method of claim 2, wherein the training of the first layer with the training sample in step (2a) is performed by:
(2a1) taking a training sample as an input layer of a first layer network, and obtaining hidden layer characteristics and reconstruction data of the first layer by using initial parameters of the first layer of the network;
(2a2) and continuously adjusting the parameters of the first layer network to minimize the error of the data of the input layer and the reconstructed data of the first layer network, thereby obtaining the trained parameters of the first layer.
4. The method according to claim 1, wherein in step (5), a 3 × 3 window is selected from the hyperspectral image X ' after dimensionality reduction with each sample as a center, and a corresponding new training sample set S ' and a new test sample set T ' to which neighborhood information is added are obtained, and the method comprises the following steps:
(5a) on the hyperspectral image X' after dimensionality reduction, a window with the space size of 3 multiplied by 3 is selected by taking each sample as the center, and 9 samples in the window are cascaded into a vector to obtain a hyperspectral image X containing space information;
(5b) correspondingly connecting samples in the hyperspectral image X with samples in the hyperspectral image X 'containing spatial information respectively to obtain a hyperspectral image X' containing spatial information;
(5c) and (2) acquiring a new training sample set S ' and a new test sample set T ' according to the positions of the training sample set S and the test sample set T in the hyperspectral image X in the step (1) and the corresponding positions in the hyperspectral image X ' ″ containing the space spectrum information.
5. The method of claim 1, wherein the training sample set S' is input into the softmax classifier-connected autoencoder network for network training in step (6), and the following steps are performed:
(6a) taking a training sample as the input of a first layer of an m-layer automatic encoder network, wherein m is more than or equal to 2; training the first layer by using a training sample to obtain the parameters of the trained first layer;
(6b) converting the training sample into the hidden layer characteristic of the first layer by using the trained parameters of the first layer;
(6c) taking the hidden layer characteristics of the first layer as the input of the second layer of the network, training the second layer of the network by using the hidden layer characteristics of the first layer in a training mode of the first layer to obtain the parameters of the trained second layer, and converting the input of the second layer into the hidden layer characteristics of the second layer by using the parameters; in the same way, the same strategy is adopted for the following layers until the hidden layer characteristic of the last layer is obtained;
(6d) and (5) training the softmax classifier by taking the hidden layer characteristics of the last layer obtained in the step (6c) as the input of the softmax classifier to obtain a trained classifier, and then finely adjusting the whole network to obtain the trained network.
CN201711415902.1A 2017-12-25 2017-12-25 Hyperspectral image classification method combining deep learning and neighborhood integration Active CN108256557B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201711415902.1A CN108256557B (en) 2017-12-25 2017-12-25 Hyperspectral image classification method combining deep learning and neighborhood integration

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201711415902.1A CN108256557B (en) 2017-12-25 2017-12-25 Hyperspectral image classification method combining deep learning and neighborhood integration

Publications (2)

Publication Number Publication Date
CN108256557A CN108256557A (en) 2018-07-06
CN108256557B true CN108256557B (en) 2021-09-28

Family

ID=62724013

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201711415902.1A Active CN108256557B (en) 2017-12-25 2017-12-25 Hyperspectral image classification method combining deep learning and neighborhood integration

Country Status (1)

Country Link
CN (1) CN108256557B (en)

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109344698B (en) * 2018-08-17 2021-09-03 西安电子科技大学 Hyperspectral band selection method based on separable convolution and hard threshold function
CN109522918B (en) * 2018-09-14 2020-05-05 广东工业大学 Hyperspectral image feature extraction method based on improved local singular spectrum analysis
CN111079850B (en) * 2019-12-20 2023-09-05 烟台大学 Depth-space spectrum combined hyperspectral image classification method of band significance
CN112348049A (en) * 2020-09-28 2021-02-09 北京师范大学 Image recognition model training method and device based on automatic coding

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105320965A (en) * 2015-10-23 2016-02-10 西北工业大学 Hyperspectral image classification method based on spectral-spatial cooperation of deep convolutional neural network
CN106845418A (en) * 2017-01-24 2017-06-13 北京航空航天大学 A kind of hyperspectral image classification method based on deep learning
CN106897737A (en) * 2017-01-24 2017-06-27 北京理工大学 A kind of high-spectrum remote sensing terrain classification method based on the learning machine that transfinites
CN107194423A (en) * 2017-05-19 2017-09-22 杭州电子科技大学 The hyperspectral image classification method of the integrated learning machine that transfinites of feature based random sampling

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9594983B2 (en) * 2013-08-02 2017-03-14 Digimarc Corporation Learning systems and methods

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105320965A (en) * 2015-10-23 2016-02-10 西北工业大学 Hyperspectral image classification method based on spectral-spatial cooperation of deep convolutional neural network
CN106845418A (en) * 2017-01-24 2017-06-13 北京航空航天大学 A kind of hyperspectral image classification method based on deep learning
CN106897737A (en) * 2017-01-24 2017-06-27 北京理工大学 A kind of high-spectrum remote sensing terrain classification method based on the learning machine that transfinites
CN107194423A (en) * 2017-05-19 2017-09-22 杭州电子科技大学 The hyperspectral image classification method of the integrated learning machine that transfinites of feature based random sampling

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
Semisupervised Hyperspectral Image Classification via Neighborhood Graph Learning;Daniel Jiwoong Im etc.;《IEEE Geoscience and Remote Sensing Letters》;20150611;第12卷(第09期);第1913- 1917页 *
基于空谱信息挖掘和稀疏表示学习的高光谱图像分类;张二磊;《中国优秀博硕士学位论文全文数据库(博士)信息科技辑》;20161215(第12期);第I140-19页 *
基于集成学习的高分遥感图像玉米区高精度提取算法研究;李大威;《中国优秀博硕士学位论文全文数据库(博士)》;20170715(第07期);第D047-17页 *

Also Published As

Publication number Publication date
CN108256557A (en) 2018-07-06

Similar Documents

Publication Publication Date Title
Zhang et al. Hyperspectral unmixing via deep convolutional neural networks
CN110084159B (en) Hyperspectral image classification method based on combined multistage spatial spectrum information CNN
CN108985238B (en) Impervious surface extraction method and system combining deep learning and semantic probability
CN109993220B (en) Multi-source remote sensing image classification method based on double-path attention fusion neural network
CN110443143B (en) Multi-branch convolutional neural network fused remote sensing image scene classification method
CN107292343B (en) Hyperspectral remote sensing image classification method based on six-layer convolutional neural network and spectrum-space information combination
CN108256557B (en) Hyperspectral image classification method combining deep learning and neighborhood integration
Charles et al. Learning sparse codes for hyperspectral imagery
CN108830330B (en) Multispectral image classification method based on self-adaptive feature fusion residual error network
CN109840560B (en) Image classification method based on clustering in capsule network
CN110110596B (en) Hyperspectral image feature extraction, classification model construction and classification method
CN106845418A (en) A kind of hyperspectral image classification method based on deep learning
CN109410184B (en) Live broadcast pornographic image detection method based on dense confrontation network semi-supervised learning
Fan et al. Superpixel guided deep-sparse-representation learning for hyperspectral image classification
CN107358203B (en) A kind of High Resolution SAR image classification method based on depth convolution ladder network
CN105139028A (en) SAR image classification method based on hierarchical sparse filtering convolutional neural network
CN107832797B (en) Multispectral image classification method based on depth fusion residual error network
CN107944483B (en) Multispectral image classification method based on dual-channel DCGAN and feature fusion
CN105117736B (en) Classification of Polarimetric SAR Image method based on sparse depth heap stack network
CN104239902A (en) Hyper-spectral image classification method based on non-local similarity and sparse coding
CN111783884B (en) Unsupervised hyperspectral image classification method based on deep learning
CN111222442A (en) Electromagnetic signal classification method and device
CN111639587A (en) Hyperspectral image classification method based on multi-scale spectrum space convolution neural network
CN113673556A (en) Hyperspectral image classification method based on multi-scale dense convolution network
CN111222545B (en) Image classification method based on linear programming incremental learning

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant