CN106650765B - High-spectral data based on convolutional neural networks turns the Hyperspectral data classification method of grayscale image - Google Patents
High-spectral data based on convolutional neural networks turns the Hyperspectral data classification method of grayscale image Download PDFInfo
- Publication number
- CN106650765B CN106650765B CN201610805454.5A CN201610805454A CN106650765B CN 106650765 B CN106650765 B CN 106650765B CN 201610805454 A CN201610805454 A CN 201610805454A CN 106650765 B CN106650765 B CN 106650765B
- Authority
- CN
- China
- Prior art keywords
- data
- hyperspectral
- convolutional neural
- classification
- neural network
- 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
Links
- 238000000034 method Methods 0.000 title claims abstract description 80
- 238000013527 convolutional neural network Methods 0.000 title claims abstract description 54
- 239000013598 vector Substances 0.000 claims abstract description 47
- 238000001228 spectrum Methods 0.000 claims abstract description 34
- 230000003595 spectral effect Effects 0.000 claims description 35
- 238000007781 pre-processing Methods 0.000 claims description 16
- 238000010606 normalization Methods 0.000 claims description 11
- 238000013145 classification model Methods 0.000 claims description 8
- 239000011159 matrix material Substances 0.000 claims description 5
- 238000010191 image analysis Methods 0.000 abstract description 3
- 238000013178 mathematical model Methods 0.000 abstract description 3
- VMXUWOKSQNHOCA-UKTHLTGXSA-N ranitidine Chemical compound [O-][N+](=O)\C=C(/NC)NCCSCC1=CC=C(CN(C)C)O1 VMXUWOKSQNHOCA-UKTHLTGXSA-N 0.000 abstract 1
- 238000012549 training Methods 0.000 description 25
- 238000012706 support-vector machine Methods 0.000 description 13
- 238000002474 experimental method Methods 0.000 description 11
- 230000008569 process Effects 0.000 description 8
- 238000012800 visualization Methods 0.000 description 6
- 230000008901 benefit Effects 0.000 description 4
- 230000008859 change Effects 0.000 description 4
- 230000000694 effects Effects 0.000 description 4
- 230000006870 function Effects 0.000 description 4
- 238000003384 imaging method Methods 0.000 description 4
- 230000009286 beneficial effect Effects 0.000 description 3
- 238000013499 data model Methods 0.000 description 3
- 230000007423 decrease Effects 0.000 description 3
- 230000002950 deficient Effects 0.000 description 3
- 238000013461 design Methods 0.000 description 3
- 238000007477 logistic regression Methods 0.000 description 3
- 238000012360 testing method Methods 0.000 description 3
- 230000009466 transformation Effects 0.000 description 3
- 240000004658 Medicago sativa Species 0.000 description 2
- 235000017587 Medicago sativa ssp. sativa Nutrition 0.000 description 2
- 238000004364 calculation method Methods 0.000 description 2
- 238000007635 classification algorithm Methods 0.000 description 2
- 230000009467 reduction Effects 0.000 description 2
- 102100037293 Atrial natriuretic peptide-converting enzyme Human genes 0.000 description 1
- 101710133555 Atrial natriuretic peptide-converting enzyme Proteins 0.000 description 1
- 241001532014 Xanthorrhoea Species 0.000 description 1
- 240000008042 Zea mays Species 0.000 description 1
- 235000005824 Zea mays ssp. parviglumis Nutrition 0.000 description 1
- 235000002017 Zea mays subsp mays Nutrition 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 238000013528 artificial neural network Methods 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 235000005822 corn Nutrition 0.000 description 1
- 238000007405 data analysis Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000003062 neural network model Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000035945 sensitivity Effects 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
- 239000002699 waste material Substances 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Image Analysis (AREA)
Abstract
High-spectral data based on convolutional neural networks turns the Hyperspectral data classification method of grayscale image, and the present invention relates to Hyperspectral data classification methods.The present invention is to solve the image analysis of existing high spectrum dimension and identification accuracy and the demand of practical application mismatches, mathematical model does not meet the problem that the practical atural object regularity of distribution is short of logic etc., and the high-spectral data based on convolutional neural networks proposed turns the Hyperspectral data classification method of grayscale image.This method is by the way that Step 1: pre-processed to obtain the spectrum vector of EO-1 hyperion to EO-1 hyperion initial data, the spectrum vector data of EO-1 hyperion is converted into greyscale image data；Step 2: being classified by the textural characteristics of sample in convolution disaggregated model autonomous learning sample set to greyscale image data sample.And etc. realize.The present invention is applied to Hyperspectral data classification field.
Description
Technical Field
The invention relates to a hyperspectral data classification method, in particular to a hyperspectral data classification method for converting spectral data into a gray scale map based on a convolutional neural network.
Background
The Hyperspectral data classification is an application of Hyperspectral remote sensing, and a spectral Image in which all objects have spectral characteristics and the spectral resolution is within the range of 10l orders of magnitude is called a Hyperspectral Image (Hyperspectral Image). And the hyperspectral remote sensing data are classified by the characteristic that the objects in the same spectral region react in different conditions and the reactions of the same object to different spectrums are obviously different. The traditional hyperspectral data classification method mainly adopts the following two ideas: image classification based on spectral matching, classification based on data statistics. Although a lot of classification methods can be applied to the hyperspectral remote sensing images through the two ideas, in practical application and operation, many defects are still displayed, such as high training cost, high-resolution information waste, mismatching of the existing high-spectral-dimension image analysis and recognition accuracy and the requirement of practical application, incapability of conforming a mathematical model to the actual ground feature distribution rule, logic shortage and the like. Especially, the spectral dimension is continuously increased, so that the existing data analysis capability gradually cannot keep up with the step of high spectral dimension information.
Disclosure of Invention
The invention aims to solve the problems that the existing high-spectral-dimension image analysis and identification accuracy is not matched with the requirements of practical application, a mathematical model does not conform to the actual ground feature distribution rule and lacks logic and the like, and provides a high-spectral data classification method for converting high-spectral data into a gray-scale map based on a convolutional neural network.
The above-mentioned invention purpose is realized through the following technical scheme:
firstly, preprocessing hyperspectral original data to obtain hyperspectral spectral vectors, and converting the hyperspectral spectral vector data into grayscale image data;
step one, performing layer-by-layer normalization on hyperspectral original data to obtain normalized hyperspectral data;
step two, extracting full-wave-band values of each pixel in the normalized hyperspectral data, and enabling the full-wave-band values of each pixel to form one-dimensional vector data
Step three, repeating the step two and the step two to traverse in other pixels in the normalized hyperspectral data, and extracting W × L1 × H one-dimensional pixel spectrum vectors to obtain hyperspectral spectrum vectors;
step four, transposing the spectrum vector of each pixel in the hyperspectral spectrum vector, namely a 1 XH one-dimensional vector into a two-dimensional matrix to obtain WxL two-dimensional matrices, and storing the WxL two-dimensional matrices into a sample set containing WxL gray level pictures, namely gray level image data;
and step two, autonomously learning the texture characteristics of the samples in the sample set through a convolution classification model, and classifying the gray level image data samples.
Effects of the invention
The invention relates to a hyperspectral data classification method based on spectral data to gray-scale map of convolutional neural network, which introduces the theory and model of convolutional neural network in the hyperspectral image classification task, converts the spectral dimension vector data into two-dimensional data in the form of gray-scale map, and tries to understand the abundant information content carried by the hyperspectral data by using imaging language to classify the hyperspectral data. The method is high in classification accuracy and has important significance for fully utilizing hyperspectral information and autonomously learning the abstract characteristics of the hyperspectral information and classifying the hyperspectral information.
The spectrum information is converted into the gray level picture information, the traditional data in a vector form can be converted into a two-dimensional image with texture characteristics, the abundant texture characteristics can well reflect the data change between hyperspectral data spectrum segments, and meanwhile, the vector data is converted into the two-dimensional image data, so that the dimension reduction operation for reducing the data volume can be avoided. On the other hand, the converted texture information is processed by using the convolutional network, and abstract features carried by the data can be independently learned through the multilayer convolutional network. The method is beneficial to realizing the correct classification of the hyperspectral data and simultaneously improving the classification accuracy and the utilization rate of rich information carried by the hyperspectrum.
Table 4 classifies KSC and Pavia U data by using a CNN-gray scale, a CNN-oscillogram, a linear kernel support vector machine, a RBF kernel support vector machine, a PCA transformation support vector machine and an automatic coding and logistic regression SAE-LR method, and compares the classification results. The average accuracy and the total accuracy of each method classification are given in the two tables, and the classification accuracy of all the subdivided types in the two data sets is respectively given, so that the results are clearer, and the parts where the best results are obtained by the classification method provided by the invention are displayed in bold.
The accuracy of the spectral information gray level image classification method based on the convolutional neural network on the Pavia U data set exceeds that of a classification method based on a support vector machine; however, the performance on KSC datasets is less than optimal because the number of network layers and parameters chosen are considered based on the most stable balance, rather than the number of network layers and parameters that achieve the best results under each dataset. In addition, the convolutional neural network is a probabilistic model, and the classification accuracy generally decreases as the class of the classification increases.
Advantages of the design
(1) Converting one-dimensional data to two-dimensional data can accommodate larger data volumes and spectral dimensions.
(2) By the method, the classification accuracy of the hyperspectral images is improved.
(3) The time used for the classification process is shortened by using the GPU.
(4) And the abstract characteristics of the data are independently learned by utilizing the deep convolutional network, so that a data model with deficient logic is avoided.
The method utilizes Matlab to preprocess data, converts one-dimensional data into two-dimensional gray image data, realizes the classification of the transformed hyperspectral data by a convolutional neural network model through a Caffe platform and a Linux operating system, and utilizes a GPU to accelerate experiments, so as to reduce the operation time consumption caused by huge calculation amount and shorten the time for classification.
Drawings
Fig. 1 is a schematic diagram of a data preprocessing process based on a spectral information gray-scale map classification method according to an embodiment; wherein W is the width of hyperspectral original data; l is the length of the hyperspectral original data; h is the depth of hyperspectral original data; a is the width of the two-dimensional matrix;
FIG. 2 is a line graph of the overall accuracy on a KSC data set as a function of the number of training samples for various proposed methods of the example;
FIG. 3 is a line graph of the overall accuracy of the various proposed methods on the Pavia U data set as a function of the number of training samples;
FIG. 4(a) is a graph showing the visualization result of Alfalfa (Alfalfa) gray scale image No. 1, which is an example of an Indianpens dataset;
FIG. 4(b) is a graph of the visualization results of a Corin (Corn) grayscale chart No. 4, using the Indianpens dataset as an example, as set forth in the examples;
FIG. 4(c) is a graph of the visualization results of a Grass-schedule gray scale map No. 5, using the Indianpens dataset as an example, proposed by the example;
FIG. 4(d) is a graph of the visualization results of a Grass-trees gray scale map No. 6, using an Indianpens dataset as an example, as set forth in the examples;
FIG. 4(e) is a graph of the visualization results of the Grass-past-mowed grayscale map No. 7, using the Indianpens dataset as an example, proposed by the example;
Detailed Description
The first embodiment is as follows: the hyperspectral data classification method for converting hyperspectral data into a gray-scale map based on the convolutional neural network is specifically prepared according to the following steps:
firstly, preprocessing hyperspectral original data to obtain hyperspectral spectral vectors, and converting the hyperspectral spectral vector data into grayscale image data; (the input of the convolutional neural network needs two-dimensional image data), and the convolutional neural network is applied to a hyperspectral data classification task;
the classification method for converting spectral information based on a convolutional neural network into a grayscale image is described below by interpreting a data preprocessing process. The data preprocessing process is shown in fig. 1:
step one, performing layer-by-layer normalization on hyperspectral original data to obtain normalized hyperspectral data;
step two, extracting full-wave-band values of each pixel in the normalized hyperspectral data, and enabling the full-wave-band values of each pixel to form one-dimensional vector data
Step three, repeating the step two and the step two to traverse in other pixels in the normalized hyperspectral data, and extracting W × L1 × H one-dimensional pixel spectrum vectors to obtain hyperspectral spectrum vectors;
the hyperspectral image data classification is carried out by taking pixels as targets, and the extracted one-dimensional vector represents data information of a certain pixel in a full-spectrum section.
Step four, transposing the spectrum vector of each pixel in the hyperspectral spectrum vector, namely a 1 × H one-dimensional vector, into a two-dimensional matrix to obtain W × L two-dimensional matrices, and storing the W × L two-dimensional matrices, namely, imwrite (the imwrite is a function in matlab) into a sample set containing W × L gray-scale pictures, namely, gray-scale image data, wherein each picture in the sample set represents full-spectrum information of a certain pixel, and a certain pixel point of a single picture represents data value representation of the pixel under a specific waveband. The texture characteristics of the whole picture are very obvious and are enough to express the data information of the target pixel, and the texture characteristics also reflect the change of the information between data layers;
and step two, autonomously learning the texture characteristics of the samples in the sample set through a convolution classification model, and classifying the gray level image data samples.
The effect of the embodiment is as follows:
the hyperspectral data classification method based on the spectral data to gray map of the convolutional neural network is characterized in that the theory and the model of the convolutional neural network are introduced into a hyperspectral image classification task, spectral dimension vector data are converted into two-dimensional data in a gray map form, rich information carried by the hyperspectral data is tried to be understood by an imaging language, and the hyperspectral data are classified. The method is high in classification accuracy and has important significance for fully utilizing hyperspectral information and autonomously learning the abstract characteristics of the hyperspectral information and classifying the hyperspectral information.
The spectrum information is converted into the gray level picture information, the traditional data in a vector form can be converted into a two-dimensional image with texture characteristics, the abundant texture characteristics can well reflect the data change between hyperspectral data spectrum segments, and meanwhile, the vector data is converted into the two-dimensional image data, so that the dimension reduction operation for reducing the data volume can be avoided. On the other hand, the converted texture information is processed by using the convolutional network, and abstract features carried by the data can be independently learned through the multilayer convolutional network. The method is beneficial to realizing the correct classification of the hyperspectral data and simultaneously improving the classification accuracy and the utilization rate of rich information carried by the hyperspectrum.
Table 4 classifies KSC and Pavia U data by using a CNN-gray scale, a CNN-oscillogram, a linear kernel support vector machine, a RBF kernel support vector machine, a PCA transformation support vector machine and an automatic coding and logistic regression SAE-LR method, and compares the classification results. The average accuracy and the total accuracy of each method classification are given in the two tables, and the classification accuracy of all the subdivided types in the two data sets is respectively given, so that the results are clearer, and the parts where the best results are obtained by the classification method provided by the invention are displayed in bold.
The accuracy of the spectral information gray level image classification method based on the convolutional neural network on the Pavia U data set exceeds that of a classification method based on a support vector machine; however, the performance on KSC datasets is less than optimal because the number of network layers and parameters chosen are considered based on the most stable balance, rather than the number of network layers and parameters that achieve the best results under each dataset. In addition, the convolutional neural network is a probabilistic model, and the classification accuracy generally decreases as the class of the classification increases.
Advantages of the design
(1) Converting one-dimensional data to two-dimensional data can accommodate larger data volumes and spectral dimensions.
(2) By the method, the classification accuracy of the hyperspectral images is improved.
(3) The time used for the classification process is shortened by using the GPU.
(4) And the abstract characteristics of the data are independently learned by utilizing the deep convolutional network, so that a data model with deficient logic is avoided.
In the embodiment, Matlab is used for preprocessing data, one-dimensional data is converted into two-dimensional gray image data, a convolution neural network model is used for classifying the transformed hyperspectral data through a Caffe platform and a Linux operating system, and a GPU is used for accelerating the experiment, so that the operation time consumption caused by huge calculation amount is reduced, and the time for classification is shortened.
The second embodiment is as follows: the first difference between the present embodiment and the specific embodiment is: the step one, the preprocessing of the hyperspectral raw data specifically comprises the following steps: performing layer-by-layer normalization on hyperspectral original data:
in the formula,the normalized hyperspectral data is obtained;hyperspectral raw data at the k layer (i, j) position; w is the width of hyperspectral original data; l is the length of the hyperspectral original data; h is the depth of hyperspectral original data; w, L, H is a positive integer;
the normalization mode of the invention selects layer-by-layer internal linear normalization for two reasons: firstly, normalizing the data within the range of 0-1, so as to facilitate the subsequent imwrite of the data into a picture; secondly, due to the imaging principle of hyperspectrum, the data ranges of different spectral bands are too different, and partial spectral band information is ignored after pixel wave spectrums are extracted and converted into images. Other steps and parameters are the same as those in the first embodiment.
The third concrete implementation mode: the present embodiment differs from the first or second embodiment in that: one-dimensional vector data is formed in the first and second stepsThe method specifically comprises the following steps:
wherein,the hyperspectral raw data of the depth H at the position (i, j) are obtained;and (4) obtaining a spectrum vector of a pixel at the (i, j) position in the normalized hyperspectral data. Other steps and parameters are the same as those in the first or second embodiment.
The fourth concrete implementation mode: the difference between this embodiment mode and one of the first to third embodiment modes is: in the second step, the classification of the gray level image data samples specifically comprises the following steps:
classifying the small sample data of the gray image using a Convolutional Neural Network (CNN); analyzing the classification effect, and finally finishing comparison between the spectral image classification based on the CNN and other methods; the sample data includes small sample data and large sample data. Other steps and parameters are the same as those in one of the first to third embodiments.
The following examples were used to demonstrate the beneficial effects of the present invention:
the first embodiment is as follows:
the hyperspectral data classification method based on the convolution neural network for converting the hyperspectral data into the gray-scale map is specifically prepared according to the following steps:
firstly, preprocessing hyperspectral original data to obtain hyperspectral spectral vectors, and converting the hyperspectral spectral vector data into grayscale image data; (the input of the convolutional neural network needs two-dimensional image data), and the convolutional neural network is applied to a hyperspectral data classification task;
the classification method for converting spectral information based on a convolutional neural network into a grayscale image is described below by interpreting a data preprocessing process. The data preprocessing process is shown in fig. 1:
step one, performing layer-by-layer normalization on hyperspectral original data to obtain normalized hyperspectral data;
the preprocessing of the hyperspectral raw data specifically comprises the following steps: performing layer-by-layer normalization on hyperspectral original data:
in the formula,the normalized hyperspectral data is obtained;hyperspectral raw data at the k layer (i, j) position; w is the width of hyperspectral original data; l is the length of the hyperspectral original data; h is the depth of hyperspectral original data; w, L, H is a positive integer;
the normalization mode of the invention selects layer-by-layer internal linear normalization for two reasons: firstly, normalizing the data within the range of 0-1, so as to facilitate the subsequent imwrite of the data into a picture; secondly, due to the imaging principle of hyperspectrum, the data ranges of different spectral bands are too different, and partial spectral band information is ignored after pixel wave spectrums are extracted and converted into images.
Step two, extracting full-wave-band values of each pixel in the normalized hyperspectral data, and enabling the full-wave-band values of each pixel to form one-dimensional vector dataThe method specifically comprises the following steps:
wherein,the hyperspectral raw data of the depth H at the position (i, j) are obtained;and (4) obtaining a spectrum vector of a pixel at the (i, j) position in the normalized hyperspectral data.
Step three, repeating the step two and the step two to traverse in other pixels in the normalized hyperspectral data, and extracting W × L1 × H one-dimensional pixel spectrum vectors to obtain hyperspectral spectrum vectors;
the hyperspectral image data classification is carried out by taking pixels as targets, and the extracted one-dimensional vector represents data information of a certain pixel in a full-spectrum section.
Step four, transposing the spectrum vector of each pixel in the hyperspectral spectrum vector, namely a 1 × H one-dimensional vector, into a two-dimensional matrix to obtain W × L two-dimensional matrices, and storing the W × L two-dimensional matrices, namely, imwrite (the imwrite is a function in matlab) into a sample set containing W × L gray-scale pictures, namely, gray-scale image data, wherein each picture in the sample set represents full-spectrum information of a certain pixel, and a certain pixel point of a single picture represents data value representation of the pixel under a specific waveband. The texture characteristics of the whole picture are very obvious and are enough to express the data information of the target pixel, and the texture characteristics also reflect the change of the information between data layers;
classifying the gray image data samples by autonomously learning the texture characteristics of the samples in the sample set through a convolution classification model;
classifying the small sample data of 50-200 gray level images by using a Convolutional Neural Network (CNN); and analyzing the classification effect, and finally finishing comparison between the spectral image classification based on the CNN and other methods.
Experimental protocol and data preprocessing results
The hyperspectral databases proposed in the experimental part are respectively: KSC, Pavia U, Indianpens.
By using the method for classifying hyperspectral data based on CNN introduced by the invention, the classification accuracy based on CNN method in small training samples is compared with KNMF and RBF SVM methods, and the performance of the CNN based method in small training samples is analyzed. And then, a 6:2:2 data sample division ratio is adopted for testing, classification capability research based on a CNN method in the aspect of hyperspectral data classification is carried out through a comparison experiment, and all generated pictures in a training set, a verification set and a test set in the experiment are randomly distributed.
The gray level image obtained through a series of preprocessing has obvious texture features and fluctuation features, and the visualization result of the gray level image is as shown in fig. 4(a) to (e) by taking the preprocessing result of the indianpins data set as an example:
analysis of small samples
To analyze the sensitivity of the convolutional neural network-based classification model to the number of training samples, the following experiment was performed. Firstly, the performance of two CNN-based methods and KNMF, RBF SVM and other methods in small training samples are compared, the experiments are carried out on the above mentioned Pavia U data set and KSC data set, and the number of the training samples of each category is selected to be 50, 100, 150 and 200 for the experiments. The results of the KSC experiment are shown in Table 1, which are plotted in a line graph as FIG. 2. The results of the Pavia U experiment are shown in table 2, which is plotted as a line graph in fig. 3.
TABLE 1 comparison of overall accuracy of CCN versus methods on KSC data
TABLE 2 comparison of overall accuracy of CNN on Pavia U data with each method
From table 1 and fig. 2, it can be seen that the overall accuracy of both methods of CNN, KNMF, RBF-SVM on KSC data set increases with the number of training samples. In the gray-scale map method of CNN, the training results of KNMF and RBF-SVM are basically the same as the training results of the RBF-SVM, and the number of each training sample is almost the same as the training result of the RBF-SVM increases along with the increase of the training samples, because although each type of training sample is seen rarely, the KSC data set itself has few effective samples, and when each type takes 200 as the training sample, the total proportion of the KSC data set is already large, so that the method is almost the same as the method of RBF-SVM.
From table 2 and fig. 3, it can be seen that the overall accuracy of the CNN, KNMF and RBF-SVM methods on the Pavia U data set increases with the number of training samples. Among these, the CNN gray graph method works best when the training samples are 50, but the training result of the RBF-SVM rises fastest with the increase of the training samples, and the overall accuracy of the RBF-SVM has leveled with the CNN gray graph when the training samples are increased to 100. When the next section of the Pavia U data set is divided in a 6:2:2 manner, the method of CNN will be better than that of RBF-SVM because the training samples are already large.
By combining the above experiments, because the CNN is based on a probability model, the advantage of the hyperspectral classification method based on the convolutional neural network can be better exerted by adopting a large sample data amount and increasing the input information amount compared with adopting small sample data.
Compared with the traditional classification method
For comparison experiments, parameters of the network are adjusted on the basis of the finally selected five-layer network, a stable CNN classification model is obtained through training, and the classification model is compared with a linear kernel SVM which is good in hyperspectral image classification performance, a recently proposed support vector machine model RBF-SVM which is based on an automatic coding machine model SAE-LR and an RBF kernel and other classification methods. The CNN classification model will classify the test data after 20000 iterative training passes.
TABLE 3 statistics of the accuracy of various classification algorithms on KSC data sets
Tables 3 and 4 use the CNN-gray scale, CNN-wave form, linear kernel support vector machine, RBF kernel support vector machine, PCA transformation support vector machine, automatic coding and logistic regression SAE-LR methods to classify the KSC and Pavia U data, respectively, and compare the classification results. The average accuracy and the total accuracy of each method classification are given in the two tables, and the classification accuracy of all the subdivided types in the two data sets is respectively given, so that the results are clearer, and the parts where the best results are obtained by the classification method provided by the invention are displayed in bold.
TABLE 4 statistics of the accuracy of various classification algorithms on the Pavia U data set
The accuracy of the spectral information gray level image classification method based on the convolutional neural network on the Pavia U data set exceeds that of a classification method based on a support vector machine; however, the performance on KSC datasets is less than optimal because the number of network layers and parameters chosen are considered based on the most stable balance, rather than the number of network layers and parameters that achieve the best results under each dataset. In addition, the convolutional neural network is a probabilistic model, and the classification accuracy generally decreases as the class of the classification increases.
Advantages of the design
(1) Converting one-dimensional data to two-dimensional data can accommodate larger data volumes and spectral dimensions.
(2) By the method, the classification accuracy of the hyperspectral images is improved.
(3) The time used for the classification process is shortened by using the GPU.
(4) And the abstract characteristics of the data are independently learned by utilizing the deep convolutional network, so that a data model with deficient logic is avoided.
The present invention is capable of other embodiments and its several details are capable of modifications in various obvious respects, all without departing from the spirit and scope of the present invention.
Claims (5)
1. The hyperspectral data classification method for converting hyperspectral data into a gray scale map based on the convolutional neural network is characterized by comprising the following steps of:
firstly, preprocessing hyperspectral original data to obtain hyperspectral spectral vectors, and converting the hyperspectral spectral vector data into grayscale image data;
step one, performing layer-by-layer normalization on hyperspectral original data to obtain normalized hyperspectral data;
step two, extracting the normalized highlightForming a full-wave-band value of each pixel in the spectrum data into one-dimensional vector data
Step three, repeating the step two and the step two to traverse in other pixels in the normalized hyperspectral data, and extracting W × L1 × H one-dimensional pixel spectrum vectors to obtain hyperspectral spectrum vectors;
step four, transposing the spectrum vector of each pixel in the hyperspectral spectrum vector, namely a 1 XH one-dimensional vector into a two-dimensional matrix to obtain WxL two-dimensional matrices, and storing the WxL two-dimensional matrices into a sample set containing WxL gray level pictures, namely gray level image data;
and step two, autonomously learning the texture characteristics of the samples in the sample set through a convolution classification model, and classifying the gray level image data samples.
2. The hyperspectral data classification method by hyperspectral data to gray-scale map based on the convolutional neural network as claimed in claim 1, wherein: the step one, the preprocessing of the hyperspectral raw data specifically comprises the following steps:
performing layer-by-layer normalization on hyperspectral original data:
in the formula,the normalized hyperspectral data is obtained;hyperspectral raw data at the k layer (i, j) position; w is the width of hyperspectral original data; l is the length of the hyperspectral original data; h is the depth of hyperspectral original data; w, L, H is a positive integer.
3. The hyperspectral data classification method for the hyperspectral data to gray scale map based on the convolutional neural network as claimed in claim 2, wherein: one-dimensional vector data is formed in the first and second stepsThe method specifically comprises the following steps:
wherein,the hyperspectral raw data of the depth H at the position (i, j) are obtained;and (4) obtaining a spectrum vector of a pixel at the (i, j) position in the normalized hyperspectral data.
4. The hyperspectral data classification method by hyperspectral data to gray-scale map based on the convolutional neural network as claimed in claim 3, wherein: in the second step, the classification of the gray level image data samples specifically comprises the following steps:
classifying sample data of the gray level image by using a convolutional neural network; the sample data includes small sample data and large sample data.
5. The hyperspectral data classification method for the hyperspectral data to gray scale map based on the convolutional neural network as claimed in claim 1 or 2 is characterized in that: in the second step, the classification of the gray level image data samples specifically comprises the following steps:
the sample data of the grayscale image is classified using a convolutional neural network.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610805454.5A CN106650765B (en) | 2016-09-06 | 2016-09-06 | High-spectral data based on convolutional neural networks turns the Hyperspectral data classification method of grayscale image |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610805454.5A CN106650765B (en) | 2016-09-06 | 2016-09-06 | High-spectral data based on convolutional neural networks turns the Hyperspectral data classification method of grayscale image |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN106650765A CN106650765A (en) | 2017-05-10 |
| CN106650765B true CN106650765B (en) | 2019-06-11 |
Family
ID=58851895
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201610805454.5A Active CN106650765B (en) | 2016-09-06 | 2016-09-06 | High-spectral data based on convolutional neural networks turns the Hyperspectral data classification method of grayscale image |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106650765B (en) |
Families Citing this family (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP3662271B1 (en) * | 2017-07-31 | 2023-10-11 | Smiths Detection Inc. | System for determining the presence of a substance of interest in a sample |
| CN109447184A (en) * | 2018-11-28 | 2019-03-08 | 南京理工大学 | Android application network behavior classification method and system based on deep learning |
| CN109961073A (en) * | 2019-03-29 | 2019-07-02 | 云南电网有限责任公司电力科学研究院 | The acquisition methods and device of a kind of transmission line of electricity and shaft tower information |
| CN111178425B (en) * | 2019-12-26 | 2022-11-29 | 哈尔滨工业大学 | Surface type classification method for large-scale high-speed rotation equipment based on convolutional neural network |
| CN114076637B (en) * | 2020-08-12 | 2023-06-09 | 舜宇光学（浙江）研究院有限公司 | Hyperspectral acquisition method and system, electronic equipment and coded broad spectrum imaging device |
| CN118093794B (en) * | 2024-04-28 | 2024-07-23 | 青岛理工大学 | Novel graphic data conversion and visualization method and system |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104102929A (en) * | 2014-07-25 | 2014-10-15 | 哈尔滨工业大学 | Hyperspectral remote sensing data classification method based on deep learning |
| WO2015017796A3 (en) * | 2013-08-02 | 2015-04-02 | Digimarc Corporation | Learning systems and methods |
| WO2015077493A1 (en) * | 2013-11-20 | 2015-05-28 | Digimarc Corporation | Sensor-synchronized spectrally-structured-light imaging |
| 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 |
| CN105865624A (en) * | 2016-03-22 | 2016-08-17 | 清华大学 | Spectrum extraction method and device for hyperspectral collection system |
-
2016
- 2016-09-06 CN CN201610805454.5A patent/CN106650765B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2015017796A3 (en) * | 2013-08-02 | 2015-04-02 | Digimarc Corporation | Learning systems and methods |
| WO2015077493A1 (en) * | 2013-11-20 | 2015-05-28 | Digimarc Corporation | Sensor-synchronized spectrally-structured-light imaging |
| CN104102929A (en) * | 2014-07-25 | 2014-10-15 | 哈尔滨工业大学 | Hyperspectral remote sensing data classification method based on deep learning |
| 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 |
| CN105865624A (en) * | 2016-03-22 | 2016-08-17 | 清华大学 | Spectrum extraction method and device for hyperspectral collection system |
Non-Patent Citations (1)
| Title |
|---|
| "基于高光谱成像技术的茄子叶片色差值检测和早疫病识别方法研究";谢传奇;《中国博士学位论文全文数据库 信息科技辑》;20160515;正文第5章 |
Also Published As
| Publication number | Publication date |
|---|---|
| CN106650765A (en) | 2017-05-10 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106650765B (en) | High-spectral data based on convolutional neural networks turns the Hyperspectral data classification method of grayscale image | |
| CN107316013B (en) | Hyperspectral image classification method based on NSCT (non-subsampled Contourlet transform) and DCNN (data-to-neural network) | |
| CN109978041B (en) | Hyperspectral image classification method based on alternative updating convolutional neural network | |
| Wang et al. | A pixel-based color image segmentation using support vector machine and fuzzy C-means | |
| CN109684922B (en) | Multi-model finished dish identification method based on convolutional neural network | |
| CN109766858A (en) | Three-dimensional convolution neural network hyperspectral image classification method combined with bilateral filtering | |
| CN101059425A (en) | Method and device for identifying different variety green tea based on multiple spectrum image texture analysis | |
| CN112836671B (en) | Data dimension reduction method based on maximized ratio and linear discriminant analysis | |
| CN111738114B (en) | Vehicle target detection method based on anchor-free accurate sampling remote sensing image | |
| CN107274387A (en) | The end member extraction method of target in hyperspectral remotely sensed image based on Evolutionary multiobjective optimization | |
| CN109190511A (en) | Hyperspectral classification method based on part Yu structural constraint low-rank representation | |
| CN112733736A (en) | Class imbalance hyperspectral image classification method based on enhanced oversampling | |
| Li et al. | An improved lightweight network architecture for identifying tobacco leaf maturity based on Deep learning | |
| dos Santos et al. | Less is more: Accelerating faster neural networks straight from jpeg | |
| CN110390341B (en) | Convolutional neural network-based image recognition method for radioactive source of green traffic vehicle carried goods | |
| Luan et al. | Sunflower seed sorting based on convolutional neural network | |
| Li et al. | Evaluation the performance of fully convolutional networks for building extraction compared with shallow models | |
| An et al. | Tensor based low rank representation of hyperspectral images for wheat seeds varieties identification | |
| Song et al. | Using dual-channel CNN to classify hyperspectral image based on spatial-spectral information | |
| Mavaddati | Rice classification and quality detection based on sparse coding technique | |
| CN116310870A (en) | Object detection classification method, system and storage medium based on hyperspectral image | |
| CN116580252A (en) | Hyperspectral image classification method based on multiscale dense connection and feature aggregation network | |
| Enawugaw et al. | Machine Learning Based Soil-Type Classification | |
| CN113469084B (en) | Hyperspectral image classification method based on contrast generation countermeasure network | |
| CN115115878A (en) | Hyperspectral image classification method and device combining random shielding and BYOL structure |
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 |