CN109389106B - A hyperspectral image unmixing method and system based on 3D convolutional neural network - Google Patents

A hyperspectral image unmixing method and system based on 3D convolutional neural network Download PDF

Info

Publication number
CN109389106B
CN109389106B CN201811563969.4A CN201811563969A CN109389106B CN 109389106 B CN109389106 B CN 109389106B CN 201811563969 A CN201811563969 A CN 201811563969A CN 109389106 B CN109389106 B CN 109389106B
Authority
CN
China
Prior art keywords
neural network
image
convolutional neural
training sample
unmixing
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.)
Expired - Fee Related
Application number
CN201811563969.4A
Other languages
Chinese (zh)
Other versions
CN109389106A (en
Inventor
李杏梅
王心宇
陈雪晴
刘晓杰
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.)
China University of Geosciences
Original Assignee
China University of Geosciences
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 China University of Geosciences filed Critical China University of Geosciences
Priority to CN201811563969.4A priority Critical patent/CN109389106B/en
Publication of CN109389106A publication Critical patent/CN109389106A/en
Application granted granted Critical
Publication of CN109389106B publication Critical patent/CN109389106B/en
Expired - Fee Related legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • 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
    • 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
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • 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)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Biophysics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Astronomy & Astrophysics (AREA)
  • Remote Sensing (AREA)
  • Multimedia (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Evolutionary Biology (AREA)
  • Computational Linguistics (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Image Analysis (AREA)

Abstract

本发明提供了一种基于3D卷积神经网络的高光谱图像解混方法及系统，其方法包括：首先利用真实遥感图像的光谱特征库构造训练样本集，并利用训练样本集和softmax分类器对3D卷积神经网络进行训练；然后将真实图像输入训练好的3D神经网络中进行图像解混，得到真实图像对应的丰度矩阵，即解混结果。本发明的有益效果是：本发明提出一种高光谱图像解混的新方法，利用本发明所提出的3D卷积神经网络对高光谱图像进行解混，几乎不需要调节参数，就可以获得令人满意的结果，相比于其他方法，本发明所提出的技术方案更加简单实用。

Figure 201811563969

The present invention provides a hyperspectral image unmixing method and system based on a 3D convolutional neural network. The method includes: firstly constructing a training sample set by using a spectral feature library of real remote sensing images, and using the training sample set and a softmax classifier to compare The 3D convolutional neural network is trained; then the real image is input into the trained 3D neural network for image unmixing, and the abundance matrix corresponding to the real image is obtained, that is, the unmixing result. The beneficial effects of the present invention are as follows: the present invention proposes a new method for unmixing hyperspectral images, using the 3D convolutional neural network proposed by the present invention to unmix the hyperspectral images, almost without adjusting parameters, the Satisfactory results, compared with other methods, the technical solution proposed by the present invention is simpler and more practical.

Figure 201811563969

Description

Hyperspectral image unmixing method and system based on 3D convolutional neural network
Technical Field
The invention relates to the field of images, in particular to a hyperspectral image unmixing method and a hyperspectral image unmixing system based on a 3D convolutional neural network.
Background
The research on the decomposition of the mixed pixels of the hyperspectral images starts in the middle of the 90 s of the 20 th century, is limited to the hardware level of computers at that time, and researchers in the period only put forward a mixed pixel decomposition algorithm with a short process and simple calculation, but the thought for researching the decomposition problem of the mixed pixels and the proposed linear spectrum mixed model are still the theoretical basis of the field until 20 years later.
The mixed model of the spectrum is classified into a linear model and a nonlinear model.
The linear spectrum hybrid model assumes that incident solar radiation only interacts with one type of surface of an object, there is no interaction between the objects, and each photon can only superpose a signal of one type of substance on the image element spectrum. Assuming that when the reflected light of three objects in the same scene is simultaneously incident into one sensor, the measured spectrum is a weighted average of the reflected spectra of the three objects, and the corresponding weight represents the proportion of each feature.
The nonlinear spectrum hybrid model considers the interaction of reflected light among various ground objects in the same scene. The Hapke mixed spectrum model is provided for the surface of a planet, and is difficult to be applied to the ground surface with a plant cover, and data collection is difficult; the K-M theory is mostly applied to spectral unmixing in the mining field, but only converts the reflectivity to a proportional amount of the material absorption coefficient in a limited range; the bilinear fitting model (bilinear fitting model) considers the multiple interactions of photons with the ground object, and adds an interaction part in the linear model to explain the interaction between the multiple photons. Although many scholars began to expand research more into the nonlinear domain, nonlinear hybrid models were still far from perfect compared to linear hybrid models.
Disclosure of Invention
In order to solve the problems, the invention provides a hyperspectral image unmixing method and a hyperspectral image unmixing system based on a 3D convolutional neural network, and the hyperspectral image unmixing method based on the 3D convolutional neural network mainly comprises the following steps:
s101: constructing n training samples with the size of S multiplied by L, setting the size of abundance samples of end members in the training samples to be S multiplied by M, taking each pixel as a center of each training sample, extracting spatial characteristics of S1 multiplied by S1, and adding abundance values of the center pixels of each training sample to synthesize a training sample set; wherein S is the space size of the training sample, and L is the number of wave bands of the training sample; m is the ground object type of the training sample; s1 is the window size; the number of the wave bands of the training sample and the real image is equal;
s102: training the 3D convolutional neural network by using the training sample set obtained in the step S101 to obtain a trained 3D convolutional neural network;
s103: unmixing the real image to be unmixed by using the trained 3D convolutional neural network to obtain an abundance matrix of the real image; the abundance matrix is the unmixing result.
Further, in step S102, the 3D convolutional neural network has two 3D convolutional layers C1 and C2, and one fully connected layer F1; there are 2 convolution kernels in C1 and 4 convolution kernels in C2.
Further, in step S102, the 3D convolutional neural network is trained by using a softmax classifier, the loss of the 3D convolutional neural network is minimized by propagating the error back by using a gradient descent method, and the convolution kernel of the 3D convolutional neural network is updated by using an additional momentum method, as shown in formula (1) and formula (2):
Figure BDA0001914042980000021
wi+1＝wi-εmi+1 (2)
in the above formula, i is greater than 0 and is the iteration number, m is the momentum, epsilon represents the learning rate, w represents the weight, and mc is the momentum factor.
Further, in step S103, the specific step of unmixing the real image by using the trained 3D convolutional neural network includes:
s201: inputting the real image as input data into the first 3D convolutional layer to obtain 2 three-dimensional data C1 out;
s202: processing the three-dimensional data C1out by using a ReLU function to obtain 2 processed three-dimensional data C1REout, and sending the 2 processed three-dimensional data C1REout serving as input to a second convolutional layer to obtain 4 three-dimensional data C2 out;
s203: processing the three-dimensional data C2out by using a ReLU function to obtain 4 processed three-dimensional data C2REout, and arranging the four three-dimensional data C2REout in a row in sequence to obtain 1 eigenvector;
s204: and feeding the feature vectors into a full-link layer F1 as input to obtain a corresponding abundance matrix.
Further, in step S103, reconstructing the image according to the obtained unmixing result to obtain a reconstructed image; then comparing the reconstructed image with a real image, and calculating a quantitative error value RMSE of the reconstructed image and the real image; judging the image unmixing effect of the current 3D convolutional neural network according to the quantitative error value RMSE; the smaller the RMSE is, the better the image unmixing effect is; the formula for RMSE is shown in equation (3):
Figure BDA0001914042980000031
in the above formula, yiFor the ith picture element in the real image,
Figure BDA0001914042980000032
the image element is the ith image element of the reconstructed image, wherein i is 1,2, …, and n is the total number of the image elements of the reconstructed image; the total number of the image elements of the reconstructed image and the real image is equal.
Further, a hyperspectral image unmixing system based on 3D convolutional neural network, its characterized in that: the system comprises the following modules:
the training sample construction module is used for constructing n training samples with the size of S multiplied by L, setting the size of each end member abundance sample in the training samples to be S multiplied by M, taking each pixel of each training sample as a center, extracting the spatial characteristics of S1 multiplied by S1, and combining the abundance values of the central pixels of each training sample to synthesize a training sample set; wherein S is the space size of the training sample, and L is the number of the training sample wave bands; m is the ground object type of the training sample; s1 is the window size; the number of the wave bands of the training sample and the real image is equal;
a network training module, configured to train the 3D convolutional neural network by using the training sample set obtained in step S101, to obtain a trained 3D convolutional neural network;
the unmixing module is used for unmixing the real image to be unmixed by utilizing the trained 3D convolutional neural network to obtain an abundance matrix of the real image; the abundance matrix is the unmixing result.
The technical scheme provided by the invention has the beneficial effects that: the invention provides a novel hyperspectral image unmixing method, the 3D convolutional neural network provided by the invention is used for unmixing the hyperspectral image, the satisfactory result can be obtained almost without adjusting parameters, and compared with other methods, the technical scheme provided by the invention is simpler and more practical.
Drawings
The invention will be further described with reference to the accompanying drawings and examples, in which:
FIG. 1 is a flowchart of a hyperspectral image unmixing method based on a 3D convolutional neural network in an embodiment of the present invention;
FIG. 2 is a schematic diagram of a module composition of a hyperspectral image unmixing system based on a 3D convolutional neural network in an embodiment of the present invention;
FIG. 3 is a schematic diagram of an experimental flow chart in an embodiment of the present invention;
FIG. 4 is a schematic diagram of the corresponding abundances of the infinite end-members of the FCLS algorithm in the embodiment of the present invention;
FIG. 5 is a schematic representation of the corresponding abundance of the FCLS algorithm chalcedony end-member in the embodiment of the present invention;
FIG. 6 is a schematic representation of the corresponding abundances of the sparse decomposition of the alenite end-members of L1 in an embodiment of the present invention;
FIG. 7 is a schematic representation of L1 sparsely decomposed chalcedony end-member correspondence abundance in an embodiment of the invention;
FIG. 8 is a schematic diagram of the corresponding abundance of the infinite end-members of the 3D-CNN algorithm in the embodiment of the present invention;
FIG. 9 is a schematic diagram of the corresponding abundance of the 3D-CNN algorithm chalcedony end-member in the example of the present invention.
Detailed Description
For a more clear understanding of the technical features, objects and effects of the present invention, embodiments of the present invention will now be described in detail with reference to the accompanying drawings.
The embodiment of the invention provides a hyperspectral image unmixing method and a hyperspectral image unmixing system based on a 3D convolutional neural network.
Referring to fig. 1, fig. 1 is a flowchart of a hyperspectral image unmixing method based on a 3D convolutional neural network in an embodiment of the present invention, which specifically includes the following steps:
s101: constructing n training samples with the size of S multiplied by L, setting the size of abundance samples of end members in the training samples to be S multiplied by M, taking each pixel as a center of each training sample, extracting spatial characteristics of S1 multiplied by S1, and adding abundance values of the center pixels of each training sample to synthesize a training sample set; wherein S is the space size of the training sample, and L is the number of wave bands of the training sample; m is the ground object type of the training sample; s1 is the window size; the number of the wave bands of the training sample and the real image is equal;
s102: training the 3D convolutional neural network by using the training sample set obtained in the step S101 to obtain a trained 3D convolutional neural network;
s103: unmixing the real image to be unmixed by using the trained 3D convolutional neural network to obtain an abundance matrix of the real image; the abundance matrix is the unmixing result.
In step S102, the 3D convolutional neural network has two 3D convolutional layers C1 and C2, and a fully connected layer F1; there are 2 convolution kernels in C1 and 4 convolution kernels in C2.
In step S102, training the 3D convolutional neural network by using a softmax classifier, performing back propagation on the error by using a gradient descent method to minimize the loss of the 3D convolutional neural network, and updating the convolution kernel of the 3D convolutional neural network by using an additional momentum method, as shown in formula (1) and formula (2):
Figure BDA0001914042980000051
wi+1＝wi-εmi+1 (2)
in the above formula, i is greater than 0 and is the iteration number, m is the momentum, epsilon represents the learning rate, w represents the weight, and mc is the momentum factor.
In step S103, the specific step of unmixing the real image by using the trained 3D convolutional neural network includes:
s201: inputting the real image as input data into the first 3D convolutional layer to obtain 2 three-dimensional data C1 out; the method comprises the following specific steps:
the C1 comprises 2 3D convolution kernels with the size of H1 xW 1 xD 1, and after each sample is respectively subjected to 3D convolution with the 2 convolution kernels, 2 pieces of three-dimensional data C1out with the size of (S2-H1+1) × (S2-W1+1) × (L-D1+1) are obtained; s2 is the real image space size; l is the number of spectral bands of the real image;
s202: processing the three-dimensional data C1out by using a ReLU function to obtain 2 processed three-dimensional data C1REout, and sending the 2 processed three-dimensional data C1REout serving as input to a second convolutional layer to obtain 4 three-dimensional data C2 out; the method comprises the following specific steps:
the second 3D convolutional layer C2 involved 4 3D convolutional kernels of size H2 × W2 × D2. After each input three-dimensional data is respectively subjected to 3D convolution with 4 convolution kernels, the results of the 4 three-dimensional data obtained by the same convolution kernel are accumulated and summed to obtain 4 three-dimensional data C2out with the size of (S2-H1-H2+2) × (S2-W1-W2+2) × (L-D1-D2+ 2);
s203: processing the three-dimensional data C2out by using a ReLU function to obtain 4 processed three-dimensional data C2REout, and arranging the four three-dimensional data C2REout into a row according to the sequence (the sequence is set according to the requirement), so as to obtain 1 (S2-H1-H2+2) × (S2-W1-W2+2) × (L-D1-D2+2) dimensional feature vector;
s204: and inputting the feature vectors into a full-link layer F1 as input to obtain a corresponding abundance matrix.
In step S103, reconstructing an image corresponding to the abundance matrix according to the obtained unmixing result to obtain a reconstructed image; then comparing the reconstructed image with a real image, and calculating a quantitative error value RMSE of the reconstructed image and the real image; judging the image unmixing effect of the current 3D convolutional neural network according to the quantitative error value RMSE; the smaller the RMSE is, the better the image unmixing effect is; the formula for RMSE is shown in equation (3):
Figure BDA0001914042980000061
in the above formula, yiFor the ith picture element in the real image,
Figure BDA0001914042980000062
the image element is the ith image element of the reconstructed image, wherein i is 1,2, …, and n is the total number of the image elements of the reconstructed image; the total number of the image elements of the reconstructed image and the real image is equal.
Referring to fig. 2, fig. 2 is a schematic diagram illustrating a module composition of a hyperspectral image unmixing system based on a 3D convolutional neural network in an embodiment of the present invention, including a training sample construction module 11, a network training module 12, and an unmixing module 13, which are connected in sequence;
a training sample construction module 11, configured to construct n training samples with a size of sxsxsx L, set the size of each end member abundance sample in the training samples to sx M, extract spatial features of S1 × S1 for each training sample with each pixel as a center, and combine the abundance values of the center pixels of each training sample to form a training sample set; wherein S is the space size of the training sample, and L is the number of the training sample wave bands; m is the ground object type of the training sample; s1 is the window size; the number of the wave bands of the training sample and the real image is equal;
a network training module 12, configured to train the 3D convolutional neural network by using the training sample set obtained in step S101, so as to obtain a trained 3D convolutional neural network;
the unmixing module 13 is configured to unmix the real image to be unmixed by using the trained 3D convolutional neural network to obtain an abundance matrix of the real image; the abundance matrix is the unmixing result.
In the network training module 12, the 3D convolutional neural network has two 3D convolutional layers C1 and C2 and a fully connected layer F1; there are 2 convolution kernels in C1 and 4 convolution kernels in C2.
In this example, the actual data used was AVIRIS (aircraft visual not obtained imaging spectrometer) data imaged in 1995 in Cuprite, Nevada, USA, and experiments were carried out to demonstrate the effectiveness and creativity of the proposed technical method. In the area, minerals such as Alunite (Alunite), Kaolinite (Kaolinite), Chalcedony (Chalcedony), Muscovite (Muscovite), Montmorillonite (Montmorillonite), Jarosite (Jarosite), Calcite (Calcite) and the like are mainly included, and typical pure pixels exist; and small amount of mineral substances such as Buddingtonite, Nontronite, Halloysite, Dickite, etc. The group of data is widely applied to the research of the hyperspectral image end member extraction method and has good representativeness.
The spectral library used in the embodiment of the present invention is usgs (united States Geological survey) mineral spectral library M, which has a size of 224 × 498, where 224 is the number of wavelength bands of the spectrum, and 498 is the number of ground feature types, i.e. the number of end members.
The USGS spectral library and the Cuprice area hyperspectral image are removed from low signal-to-noise ratio and water vapor absorption wave bands, and 196 wave bands are left. For convenience of calculation, the image used in this experiment is a part of the entire image, and the size thereof is 143 × 123.
In the experiment, 11 end members contained in the end member set of the image are firstly determined, and corresponding 11 spectrums are extracted from the mineral spectrum library M of the USGS to form the experiment end member set. Thus, the end-member sample set size is 196 × 11. In practical situations, the mineral mixture generally does not include all end members, so when the hyperspectral image sample is synthesized, the mixture situation that the number of the end members is not more than 5 is mainly considered. Therefore, the sample randomly selects a number p as the number of end members, where p is a positive integer between 1 and 5. Abundance samples generated from p random end-members of the 11 end-members obey dirichlet distribution and meet the "sum is one" and "non-negative" constraints. In generating the training samples, spatially adjacent pixels are set to the same or similar abundance. And finally, synthesizing the hyperspectral image according to the linear model, and enabling the signal-to-noise ratio of the synthesized hyperspectral image to be 60 dB.
In the experiment, data of known ground object types need to be processed, so that the estimation of the number of end members by an algorithm is not needed. Because a general hyperspectral mixed image lacks the abundance of reference, when a network is trained, an abundance map sample of each end member needs to be constructed by self, and then a hyperspectral image sample is constructed by a known spectrum library.
The size of the hyperspectral image sample is set to 100 × 100 × 196, and the size of the abundance sample of each end member is 100 × 100 × 11. And (3) taking each pixel as a center to extract 5 multiplied by 196 characteristics of the hyperspectral image sample, and combining the unmixing results of the center pixels to form a training sample set. After the 3D convolutional neural network parameters are debugged by the samples, the samples are used for de-mixing real data, and an experimental flow chart is shown in FIG. 3.
The experimental results are shown in table 1, the algorithm for experimental comparison has FCLS and L1 norm sparse decomposition, and table 1 shows the reconstruction errors of the three algorithms on the hyperspectral data in the Cuprite region. Fig. 4 to 9 show the abundance results of typical figure Alunite (Alunite) and Chalcedony (Chalcedony) in 11 figures solved by 3 algorithms.
When the solution is solved by using the fully constrained least square method, the iteration is carried out for 2000 times.
When a sparse mixed pixel decomposition model based on L1 norm is used for solving, lambda is set to be 10-4, and the iteration number is 2000.
When the 3D convolutional neural network is used for training data, mini-batch is adopted, the batch number is 20, iteration is carried out for 2000 times, and the learning rate is 0.01.
Table 1 reconstruction errors of different algorithms on high spectral data of Cuprice area
FCLS L1 sparse decomposition 3D-CNN
RMSE 0.0453 0.0427 0.0394
According to the comparison of the spectrum reconstruction errors RMSE of the three algorithms in the table 1, the unmixing result based on the 3D convolutional neural network is better than the other two algorithms. Since there is no standard abundance map to compare, it can only be roughly estimated visually and numerically. But overall the results of the 3D convolutional neural network are roughly satisfactory. The most important advantage is that few adjustments of the parameters are required to obtain good results.
The invention has the beneficial effects that: the invention provides a novel hyperspectral image unmixing method, the 3D convolutional neural network provided by the invention is used for unmixing the hyperspectral image, the satisfactory result can be obtained almost without adjusting parameters, and compared with other methods, the technical scheme provided by the invention is simpler and more practical.
The above description is only for the purpose of illustrating the preferred embodiments of the present invention and is not to be construed as limiting the invention, and any modifications, equivalents, improvements and the like that fall within the spirit and principle of the present invention are intended to be included therein.

Claims (3)

1. A hyperspectral image unmixing method based on a 3D convolutional neural network is characterized by comprising the following steps: the method comprises the following steps:
s101: constructing n training samples with the size of S multiplied by L, setting the size of abundance samples of end members in the training samples to be S multiplied by M, taking each pixel as a center of each training sample, extracting spatial characteristics of S1 multiplied by S1, and adding abundance values of the center pixels of each training sample to synthesize a training sample set; wherein S is the space size of the training sample, and L is the number of wave bands of the training sample; m is the ground object type of the training sample; s1 is the window size; the number of the wave bands of the training sample and the real image is equal;
s102: training the 3D convolutional neural network by using the training sample set obtained in the step S101 to obtain a trained 3D convolutional neural network;
s103: unmixing the real image to be unmixed by using the trained 3D convolutional neural network to obtain an abundance matrix of the real image; the abundance matrix is the unmixing result;
in step S102, the 3D convolutional neural network has two 3D convolutional layers C1 and C2, and a fully connected layer F1; there are 2 convolution kernels in C1 and 4 convolution kernels in C2;
in step S102, training the 3D convolutional neural network by using a softmax classifier, performing back propagation on the error by using a gradient descent method to minimize the loss of the 3D convolutional neural network, and updating the convolution kernel of the 3D convolutional neural network by using an additional momentum method, as shown in formula (1) and formula (2):
Figure FDA0003024851810000011
wi+1＝wi-εmi+1 (2)
in the above formula, i is greater than 0 and is iteration times, m is momentum, epsilon represents learning rate, w represents weight, and mc is a momentum factor;
in step S103, the specific step of unmixing the real image by using the trained 3D convolutional neural network includes:
s201: inputting the real image as input data into the first 3D convolutional layer to obtain 2 three-dimensional data C1 out;
s202: processing the three-dimensional data C1out by using a ReLU function to obtain 2 processed three-dimensional data C1REout, and sending the 2 processed three-dimensional data C1REout serving as input to a second convolutional layer to obtain 4 three-dimensional data C2 out;
s203: processing the three-dimensional data C2out by utilizing a ReLU function to obtain 4 processed three-dimensional data C2REout, and arranging the four three-dimensional data C2REout in a row in sequence to obtain 1 eigenvector;
s204: and feeding the feature vectors into a full-link layer F1 as input to obtain a corresponding abundance matrix.
2. The hyperspectral image unmixing method based on the 3D convolutional neural network as claimed in claim 1, wherein: in step S103, reconstructing the image according to the obtained unmixing result to obtain a reconstructed image; then comparing the reconstructed image with a real image, and calculating a quantitative error value RMSE of the reconstructed image and the real image; judging the image unmixing effect of the current 3D convolutional neural network according to the quantitative error value RMSE; the smaller the RMSE is, the better the image unmixing effect is; the formula for RMSE is shown in equation (3):
Figure FDA0003024851810000021
in the above formula, yiFor the ith picture element in the real image,
Figure FDA0003024851810000022
the image element is the ith image element of the reconstructed image, wherein i is 1,2, …, and n is the total number of the image elements of the reconstructed image; the total number of the image elements of the reconstructed image and the real image is equal.
3. A hyperspectral image unmixing system based on a 3D convolutional neural network is used for realizing the hyperspectral image unmixing method based on the 3D convolutional neural network according to any one of claims 1 to 2, and is characterized in that: the system comprises the following modules:
the training sample construction module is used for constructing n training samples with the size of S multiplied by L, setting the size of each end member abundance sample in the training samples to be S multiplied by M, taking each pixel of each training sample as a center, extracting the spatial characteristics of S1 multiplied by S1, and combining the abundance values of the central pixels of each training sample to synthesize a training sample set; wherein S is the space size of the training sample, and L is the number of the training sample wave bands; m is the ground object type of the training sample; s1 is the window size; the number of the wave bands of the training sample and the real image is equal;
a network training module, configured to train the 3D convolutional neural network by using the training sample set obtained in step S101, to obtain a trained 3D convolutional neural network;
the unmixing module is used for unmixing the real image to be unmixed by utilizing the trained 3D convolutional neural network to obtain an abundance matrix of the real image; the abundance matrix is the unmixing result.
CN201811563969.4A 2018-12-20 2018-12-20 A hyperspectral image unmixing method and system based on 3D convolutional neural network Expired - Fee Related CN109389106B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201811563969.4A CN109389106B (en) 2018-12-20 2018-12-20 A hyperspectral image unmixing method and system based on 3D convolutional neural network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201811563969.4A CN109389106B (en) 2018-12-20 2018-12-20 A hyperspectral image unmixing method and system based on 3D convolutional neural network

Publications (2)

Publication Number Publication Date
CN109389106A CN109389106A (en) 2019-02-26
CN109389106B true CN109389106B (en) 2021-06-08

Family

ID=65430614

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201811563969.4A Expired - Fee Related CN109389106B (en) 2018-12-20 2018-12-20 A hyperspectral image unmixing method and system based on 3D convolutional neural network

Country Status (1)

Country Link
CN (1) CN109389106B (en)

Families Citing this family (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109975250B (en) * 2019-04-24 2020-03-24 中国科学院遥感与数字地球研究所 Leaf area index inversion method and device
CN110378294B (en) * 2019-07-22 2022-08-19 大连海事大学 Hyperspectral target detection method and system based on local energy constraint and feature fusion
CN110852369B (en) * 2019-11-06 2022-09-27 西北工业大学 A Hyperspectral Image Classification Method Using Joint 3D/2D Convolutional Networks and Adaptive Spectral Unmixing
CN111144214B (en) * 2019-11-27 2021-06-15 中国石油大学(华东) Hyperspectral image unmixing method based on multi-layer stacked autoencoder
CN111260576B (en) * 2020-01-14 2022-07-05 哈尔滨工业大学 Hyperspectral unmixing algorithm based on de-noising three-dimensional convolution self-coding network
CN111581879A (en) * 2020-04-30 2020-08-25 北京航空航天大学 A method and system for determining the nonlinear abundance of mixed pixels in a space artificial target

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101697008A (en) * 2009-10-20 2010-04-21 北京航空航天大学 Hyperspectral unmixing method for estimating regularized parameter automatically
CN105320959A (en) * 2015-09-30 2016-02-10 西安电子科技大学 End member learning based hyperspectral image sparse unmixing method
CN107680081A (en) * 2017-09-08 2018-02-09 西安电子科技大学 High spectrum image solution mixing method based on convolutional neural networks

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7505608B2 (en) * 2005-04-15 2009-03-17 The Boeing Company Methods and apparatus for adaptive foreground background analysis

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101697008A (en) * 2009-10-20 2010-04-21 北京航空航天大学 Hyperspectral unmixing method for estimating regularized parameter automatically
CN105320959A (en) * 2015-09-30 2016-02-10 西安电子科技大学 End member learning based hyperspectral image sparse unmixing method
CN107680081A (en) * 2017-09-08 2018-02-09 西安电子科技大学 High spectrum image solution mixing method based on convolutional neural networks

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
Combining Unmixing and Deep Feature Learning for Hyperspectral Image Classification;Fahim Irfan Alam et al.;《2017 International Conference on Digital Image Computing:Techniques and Appliations》;20171221;全文 *

Also Published As

Publication number Publication date
CN109389106A (en) 2019-02-26

Similar Documents

Publication Publication Date Title
CN109389106B (en) A hyperspectral image unmixing method and system based on 3D convolutional neural network
Xu et al. Nonlocal patch tensor sparse representation for hyperspectral image super-resolution
Lin et al. Hyperspectral image denoising via matrix factorization and deep prior regularization
CN109241843B (en) Space-spectrum combined multi-constraint optimization non-negative matrix unmixing method
CN112699838B (en) Hyperspectral mixed pixel nonlinear blind decomposition method based on spectral diagnosis characteristic weighting
CN109035154A (en) A kind of hyperspectral image nonlinear solution mixing method considering spectrum variability
CN111008975A (en) A mixed pixel unmixing method and system for a linear model of a space artificial target
CN114529830B (en) Remote sensing image spatiotemporal fusion method based on hybrid convolutional network
CN113837314A (en) A Hyperspectral Image Classification Method Based on Hybrid Convolutional Neural Networks
Wang et al. A self-supervised deep denoiser for hyperspectral and multispectral image fusion
Minh et al. Deep Recurrent Neural Networks for mapping winter vegetation quality coverage via multi-temporal SAR Sentinel-1
CN116403046A (en) Hyperspectral image classification device and method
Wang et al. Row-sparsity spectral unmixing via total variation
CN108765313A (en) High spectrum image denoising method based on low-rank representation in class
CN113094645B (en) Hyperspectral data unmixing method based on morphological component constraint optimization
CN114863173B (en) Self-mutual-attention hyperspectral image classification method for land resource audit
CN103034988A (en) Space-time quantitative remote sensing fusion method of arbitrary number of sensors
CN116051896A (en) Hyperspectral image classification method of lightweight mixed tensor neural network
Han et al. Hypothetical visible bands of advanced meteorological imager onboard the geostationary Korea multi-purpose satellite-2A using data-to-data translation
CN115272144A (en) Spatio-temporal spectrum fusion method for hyperspectral image and multispectral image
CN111062888B (en) A hyperspectral image denoising method based on multi-objective low-rank sparse and total spatial variation
Zhang et al. Three-Dimension Spatial-Spectral Attention Transformer for Hyperspectral Image Denoising
Wang et al. Hyperspectral unmixing via plug-and-play priors
Wu et al. Hyperspectral superresolution reconstruction via decomposition of low-rank and sparse tensor
Jia et al. Deep spectral unmixing framework via 3D denoising convolutional autoencoder

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
CF01 Termination of patent right due to non-payment of annual fee

Granted publication date: 20210608

Termination date: 20211220

CF01 Termination of patent right due to non-payment of annual fee