CN109359623B - Hyperspectral image migration classification method based on depth joint distribution adaptive network - Google Patents

Hyperspectral image migration classification method based on depth joint distribution adaptive network Download PDF

Info

Publication number
CN109359623B
CN109359623B CN201811347067.7A CN201811347067A CN109359623B CN 109359623 B CN109359623 B CN 109359623B CN 201811347067 A CN201811347067 A CN 201811347067A CN 109359623 B CN109359623 B CN 109359623B
Authority
CN
China
Prior art keywords
probability distribution
domain
target domain
network
hyperspectral
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
CN201811347067.7A
Other languages
Chinese (zh)
Other versions
CN109359623A (en
Inventor
耿杰
马晓瑞
王洪玉
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.)
Dalian University of Technology
Northwestern Polytechnical University
Original Assignee
Dalian University of Technology
Northwestern Polytechnical 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 Dalian University of Technology, Northwestern Polytechnical University filed Critical Dalian University of Technology
Priority to CN201811347067.7A priority Critical patent/CN109359623B/en
Publication of CN109359623A publication Critical patent/CN109359623A/en
Application granted granted Critical
Publication of CN109359623B publication Critical patent/CN109359623B/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
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • 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)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Computer Vision & Pattern Recognition (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

A hyperspectral image migration classification method based on a depth joint distribution adaptive network comprises the following steps: inputting hyperspectral images of a source domain and a target domain, and performing characteristic normalization and dimension unification; combining the characteristics of the source domain and target domain hyperspectral images; constructing an edge probability distribution adaptation network, and performing edge probability distribution adaptation of the source domain and target domain hyperspectral images; selecting training samples of hyperspectral images of a source domain and a small number of target domains according to a one-to-many classification principle; constructing a conditional probability distribution adaptation network, and performing conditional probability distribution adaptation of the hyperspectral images of the source domain and the target domain; and performing one-to-many classification on the hyperspectral images of the target domain. The invention provides a depth-based joint distribution adaptation network, which realizes the characteristic adaptation of the high-spectrum images of the source domain and the target domain and reduces the joint probability distribution difference of the source domain and the target domain; meanwhile, the one-to-many classification model is adopted, so that the distinguishing performance between classes is improved, and the accuracy of the migration and classification of the hyperspectral images is further improved.

Description

Hyperspectral image migration classification method based on depth joint distribution adaptive network
Technical Field
The invention belongs to the technical field of remote sensing image processing, and particularly relates to hyperspectral image migration classification based on a depth joint distribution adaptive network.
Background
The hyperspectral image has high spectral resolution and wide wave band coverage range, and the obtained feature target spectrum detail information is rich, thereby being beneficial to fine feature analysis. The hyperspectral image classification is an important content for the hyperspectral image interpretation and is widely applied to the fields of mineral exploration, vegetation investigation, agricultural monitoring and the like. Due to the fact that the hyperspectral image is large in data quantity and redundancy exists, the same target has spectrum difference on different data, and the classification effect is affected.
The hyperspectral image classification is a process of classifying the characteristics of the acquired target spectral signals and the like by analysis, and a supervised classification method is mainly adopted. The supervised classification models all need a large number of training samples to obtain excellent classification results. Usually, training samples are obtained by manual labeling, which requires a large labor cost. In practical application, a new marked sample of remote sensing data is not easy to obtain. In order to realize automatic classification, the new remote sensing data should be classified by using a classifier trained by other data. However, in the hyperspectral image classification process, the data migration capability of the classifier is affected due to different factors such as sensors and waveband coverage. A classifier trained on one data set has a higher classification accuracy on that data, but hardly achieves the same effect on other data.
Aiming at the problem of limited data migration capability in hyperspectral image migration classification, a domain adaptation model based on migration learning is proposed to reduce the influence of different sensor data differences. For example, Kernel-based domain-invariant feature selection in hyperspectral images for transfer learning, published by IEEE Transactions on Geoscience and Remote Sensing, volume 54, 5, of 2016 C.et al, proposes a method for domain-invariant feature selection based on nuclear learning, in which the stability of data migration is measured by calculating the distance of conditional probability distributions of source and target domains in the regenerated nuclear Hilbert space, and experiments prove the effectiveness of feature selection. In 2018, published in volume 56 and phase 10 of IEEE Transactions on science and motion Sensing, Deep feature alignment neural networks for domain adaptation of hyperspectral data, a Deep convolution cyclic neural network was proposed to perform feature learning and domain adaptation of source domain and target domain hyperspectral images, and experiments prove that the method can improve the precision of target domain hyperspectral image classification by using information of the source domain. The method does not analyze the joint probability distribution of the hyperspectral images of the source domain and the target domain, and does not effectively utilize the joint probability distribution of data to further improve the migration classification effect.
Disclosure of Invention
The invention aims to provide a model capable of obtaining a better migration classification effect, and provides a hyperspectral image migration classification method based on a depth joint distribution adaptive network.
The technical scheme of the invention is as follows:
the hyperspectral image migration classification method based on the depth joint distribution adaptive network comprises the following steps:
(1) inputting hyperspectral images of a source domain and a target domain, and performing characteristic normalization and dimension unification:
(1a) respectively normalizing the spectral characteristics of the two hyperspectral images of the source domain and the target domain to enable the spectral characteristics to be distributed between 0 and 1;
(1b) if the spectral dimensions of the two hyperspectral images are different, zero padding is carried out on the hyperspectral images with low dimensions, so that the hyperspectral images with low dimensions and the hyperspectral images with high dimensions are unified in dimension;
(2) combining the characteristics of the source domain and target domain hyperspectral images:
(2a) respectively vectorizing the spectral characteristics of the source domain hyperspectral image and the target domain hyperspectral image;
(2b) combining the spectral characteristics of the source domain and the target domain into a vector set;
(3) constructing a three-layer marginal probability distribution adaptation network, and carrying out marginal probability distribution adaptation of the hyperspectral images of the source domain and the target domain:
(3a) constructing a first layer of edge probability distribution adaptation network by utilizing a linear denoising encoder and a nonlinear encoder, solving the weight of the linear denoising encoder, and adapting the spectral characteristics of a source domain and a target domain;
(3b) constructing a second layer edge probability distribution adaptation network by utilizing a linear denoising encoder and a nonlinear encoder, solving the weight of the linear denoising encoder, and adapting the spectral characteristics of a source domain and a target domain;
(3c) constructing a third layer edge probability distribution adaptation network by utilizing a linear denoising encoder and a nonlinear encoder, solving the weight of the linear denoising encoder, and adapting the spectral characteristics of a source domain and a target domain;
(4) selecting training samples of hyperspectral images of a source domain and a target domain according to a one-to-many classification principle:
classifying the class C, selecting a source domain hyperspectral image sample and p% of a target domain hyperspectral image sample, forming C training sample sets according to a one-to-many classification principle, wherein each training set sample label belongs to the class or does not belong to the class, and the sample characteristic is a characteristic output by an edge probability distribution adaptation network;
(5) c three layers of conditional probability distribution adaptation networks are constructed, and the conditional probability distribution adaptation of the hyperspectral images of the source domain and the target domain is carried out:
(5a) respectively initializing the weight and the offset parameter of the C three-layer conditional probability distribution adaptation networks, and respectively pre-training the C networks layer by utilizing each training sample set;
(5b) after pre-training, the implicit output of the third layer conditional probability distribution adaptation network is used as the optimized sample characteristic;
(5c) in each network, a softmax classifier is connected after the third layer of conditional probability distribution adaptation network, characteristics and class labels after the training samples are optimized are input into the classifier, and the weight and the bias parameters of the softmax classifier are optimized;
(5d) fine-tuning the C conditional probability distribution adaptation networks from the top layer to the bottom layer, and further optimizing parameters of each network;
(6) performing one-to-many classification on the hyperspectral image of the target domain:
(6a) respectively inputting a test sample set of a target domain hyperspectral image into C conditional probability distribution adaptation networks, wherein the optimized characteristics of the networks are implicit output of a third layer of conditional probability distribution adaptation networks;
(6b) classifying the test samples by using the trained C softmax classifiers to obtain prediction probabilities of the test samples belonging to and not belonging to each class;
(6c) comparing the prediction probabilities belonging to each category, and taking the category corresponding to the maximum probability as the prediction label of each sample;
(6d) and outputting a classification result graph of the hyperspectral image of the target domain according to the prediction label vector and the spatial position of the test sample.
Compared with the prior art, the invention mainly has the following advantages:
firstly, the invention provides a deep joint distribution adaptation network, which respectively utilizes an edge probability distribution adaptation network and a conditional probability distribution adaptation network to realize the characteristic adaptation of the hyperspectral images of a source domain and a target domain and reduce the joint probability distribution difference of the source domain and the target domain.
Secondly, the method utilizes the source domain samples and a small amount of target domain samples to train and learn the classifier, thereby reducing the requirements of the target domain training samples; meanwhile, a one-to-many classification model is adopted, and the distinguishing performance between classes in the classes is improved and the migration classification precision is improved by learning a plurality of classifiers for two-classification.
Drawings
FIG. 1 is a flow chart of an implementation of hyperspectral image migration classification;
FIG. 2 is source domain Pavia University hyperspectral data;
FIG. 3 is target domain Pavia Center hyperspectral data;
FIG. 4 is a graph of classification results without migration using a one-to-many softmax classifier;
FIG. 5 is a graph of the classification results using the method of the present invention.
Detailed Description
The present invention will now be described in detail with reference to the following examples and accompanying drawings.
According to fig. 1, the hyperspectral image migration classification method based on the depth joint distribution adaptive network comprises the following steps:
(1) reading hyperspectral images of a source domain and a target domain, and performing characteristic normalization and dimension unification:
(1a) respectively carrying out linear normalization on the spectral characteristics of the hyperspectral images of the source domain and the target domain, so that the spectral characteristics are distributed between 0 and 1;
(1b) if the spectral dimensions of the hyperspectral images of the source domain and the target domain are different, zero padding is carried out on the hyperspectral images with low dimensions, so that the hyperspectral images with low dimensions are unified with the hyperspectral images with high dimensions;
(2) combining the characteristics of the source domain and target domain hyperspectral images:
(2a) quantizing the spectral feature vector of the source domain hyperspectral image into XSQuantizing the spectral feature vector of the hyperspectral image of the target domain into XT；
(2b) Combining spectral features of a source domain and a target domain into a vector set X ═ XS XT]；
(3) Constructing three layers of conditional probability distribution adaptation networks, and carrying out edge probability distribution adaptation of the hyperspectral images of the source domain and the target domain:
(3a) constructing a first layer edge probability distribution adaptation network by utilizing a linear denoising coder and a nonlinear coder:
for sample xiObtaining M disturbed versions by randomly setting zero to each dimension characteristic, wherein the mth disturbed version is x'i,m(ii) a Pair x 'by a Linear De-noising encoder'i,mThe linear implicit rays are to be restored to original,
Figure BDA0001864015160000051
the objective equation for optimizing the weight W is as follows:
Figure BDA0001864015160000052
wherein the first term is an average reconstruction error, the second term is an edge penalty term, lambda represents a balance factor, NSAnd NTRespectively representing the sample numbers of the source domain hyperspectral image and the target domain hyperspectral image, wherein N is the total sample number, namely N is equal to NS+NT(ii) a Let the perturbed sample matrix X'm＝[x′1,m,x′2,m,…,x′N,m]Original copy matrix of M times
Figure BDA0001864015160000053
M perturbed sample matrix combinations X '═ X'1,X′2,…,X′M]Thus, the above formula objective equation is converted into
Figure BDA0001864015160000061
Wherein D represents a disparity index matrix defined as
Figure BDA0001864015160000062
Thus, the linear de-noising encoder weight W has the following closed solution:
Figure BDA0001864015160000063
then, a nonlinear encoder is adopted to carry out nonlinear implicit reflection on the characteristics, and the formula is as follows
Figure BDA0001864015160000064
Here, ,
Figure BDA0001864015160000065
representing an output of the first tier edge probability distribution adaptation network;
(3b) according to the step (3a), a linear de-noising encoder and a non-linear encoder are utilized to construct a second layer edge probability distribution adaptation network, the weight of the linear de-noising encoder is solved, the spectral characteristics of a source domain and a target domain are adapted, and the output of a second layer network is obtained
Figure BDA0001864015160000066
(3c) According to the step (3a), a third layer edge probability distribution adaptation network is constructed by utilizing a linear denoising encoder and a nonlinear encoder, the weight of the linear denoising encoder is solved, the spectral characteristics of a source domain and a target domain are adapted, and the output of a third layer network is obtained
Figure BDA0001864015160000067
(4) Selecting training samples of hyperspectral images of a source domain and a target domain according to a one-to-many classification principle:
classifying 9 classes, selecting 500 samples of each class of the source domain hyperspectral image and 1% of samples of each class of the target domain hyperspectral image, forming 9 training sample sets according to a one-to-many classification principle, wherein the labels of the samples in each training sample set belong to the class (represented as 1) and do not belong to the class (represented as 0), and the sample characteristics are output of a third-layer network
Figure BDA0001864015160000068
(5) Constructing 9 three layers of conditional probability distribution adaptation networks, and performing conditional probability distribution adaptation of the hyperspectral images of the source domain and the target domain:
(5a) initializing the weight and the bias parameters of the three layers of conditional probability distribution adaptive networks, and pre-training each layer of network layer by utilizing a c training sample set; the encoding process and the decoding process of the k-th layer conditional probability distribution adaptation network are as follows:
Figure BDA0001864015160000071
Figure BDA0001864015160000072
wherein,
Figure BDA0001864015160000073
and
Figure BDA0001864015160000074
representing the input, implicit representation, and output, respectively, of the k-th layer network, f (-) and g (-) being nonlinear functions of encoding and decoding, respectively,
Figure BDA0001864015160000075
and
Figure BDA0001864015160000076
are respectively the right of the encoding processThe weight and the offset are added, and the weight and the offset are added,
Figure BDA0001864015160000077
and
Figure BDA0001864015160000078
weight and bias for the decoding process, respectively, representing S (source domain) or T (target domain);
the target equation for the pre-training weights and bias parameters is:
Figure BDA0001864015160000079
wherein the first term represents the average reconstruction error on the training samples, the second term is a weight penalty term to prevent the weight from being too large,
Figure BDA00018640151600000710
Ntrainrepresenting the number of training samples, and lambda' representing a weight penalty factor; solving the above formula by using a back propagation algorithm;
(5b) implicit output of third-layer conditional probability distribution adaptation network after pre-training
Figure BDA00018640151600000711
As optimized sample features;
(5c) connecting a softmax classifier after the third layer of conditional probability distribution adaptation network, inputting the optimized features and class labels of the training samples into the classifier, and optimizing the weight and bias parameters of the softmax classifier;
(5d) fine-tuning the conditional probability distribution adaptive network from the top layer to the bottom layer, and further optimizing parameters of each network; the target equation for reverse fine tuning the entire network weight and bias parameters is:
Figure BDA00018640151600000712
wherein the first term is the total reconstruction error of the coding layer and the second term is a conditionA penalty term for reducing the difference of the conditional probability distribution of the source domain and the target domain samples;
Figure BDA00018640151600000713
and
Figure BDA00018640151600000714
respectively counting samples of a c-th source domain hyperspectral region and a target domain hyperspectral region; similarly, solving the above formula by using a back propagation algorithm to obtain parameters of each layer of the trimmed conditional probability distribution adaptive network and the softmax classifier;
(5e) according to the steps (5a) - (5d), training by utilizing 9 training sample sets respectively to obtain 9 three-layer conditional probability distribution adaptation networks, and realizing the conditional probability distribution adaptation of the hyperspectral images of the source domain and the target domain;
(6) performing one-to-many classification on the hyperspectral image of the target domain:
(6a) inputting a test sample set of a hyperspectral image of a target domain into a c conditional probability distribution adaptation network, wherein the characteristic after network optimization is the implicit output of a third layer of conditional probability distribution adaptation network
Figure BDA0001864015160000081
(6b) Classifying the test samples by using a trained softmax classifier to obtain the prediction probability belonging to the class c:
Figure BDA0001864015160000082
wherein,
Figure BDA0001864015160000083
and
Figure BDA0001864015160000084
partial weights and offsets corresponding to the softmax classifier, respectively;
(6c) according to the steps (6a) - (6b), respectively inputting the test sample set of the target domain hyperspectral image into 9 conditional probability distribution adaptation networks and softmax classifiers to obtain the probability of the sample belonging to each category;
(6d) comparing the prediction probabilities belonging to each category, and taking the category corresponding to the maximum probability as the prediction label of each sample, as follows
Figure BDA0001864015160000085
(6e) And outputting a classification result graph of the hyperspectral image of the target domain according to the prediction label vector and the spatial position of the test sample.
The technical effects of the invention are explained by simulation experiments as follows:
1. simulation conditions and content
The experimental data of the invention are Pavia University hyperspectral data and Pavia Center hyperspectral data, which are both obtained by a ROSIS sensor. As shown in fig. 2, the size of the source domain Pavia University hyperspectral data is 610 × 340 pixels, 103 bands, fig. 2(a) is an image synthesized by 57 th, 34 th and 3 rd bands, and fig. 2(b) is a corresponding real terrain labeling map, which has 9 types including asphalt, grass, gravel, trees, metal plates, soil, asphalt, bricks and shadows. As shown in fig. 3, the target domain Pavia Center hyperspectral data has a size of 1096 × 715 pixels and 102 bands, fig. 3(a) is an image synthesized by 57 th, 34 th and 3 rd bands, and fig. 3(b) is a corresponding real ground object labeled map, and has 9 types including water, tree, grass, brick, land, asphalt, tile and shadow. Fig. 4 is a result chart of classification of a Pavia Center hyperspectral image without migration by using a one-to-many softmax classifier, fig. 5 is a result chart of migration classification of a Pavia Center hyperspectral image by using the method of the invention, and table one shows the correspondence between training samples of a source domain and a target domain and the classification precision comparison thereof. In simulation experiments, the method and the comparison method are both realized in Matlab R2017a by programming.
2. Analysis of simulation results
Table-classification accuracy comparison
Figure BDA0001864015160000091
Figure BDA0001864015160000101
As can be seen from the table I, the deep joint distribution adaptive network of the invention is adopted to carry out migration classification to classification without migration, so that higher classification precision is obtained, and the effectiveness of the method of the invention in migration classification is proved. Comparing fig. 4 and fig. 5, it can be known that the classification result graph of the method of the present invention has fewer misclassification points and is closer to the real ground object label graph. In a word, the method can effectively improve the migration and classification effect of the hyperspectral images.

Claims (1)

1. A hyperspectral image migration classification method based on a depth joint distribution adaptive network is characterized by comprising the following steps:
(1) reading hyperspectral images of a source domain and a target domain, and performing characteristic normalization and dimension unification:
(1a) respectively carrying out linear normalization on the spectral characteristics of the hyperspectral images of the source domain and the target domain, so that the spectral characteristics are distributed between 0 and 1;
(1b) if the spectral dimensions of the hyperspectral images of the source domain and the target domain are different, zero padding is carried out on the hyperspectral images with low dimensions, so that the hyperspectral images with low dimensions are unified with the hyperspectral images with high dimensions;
(2) combining the characteristics of the source domain and target domain hyperspectral images:
(2a) quantizing the spectral feature vector of the source domain hyperspectral image into XSQuantizing the spectral feature vector of the hyperspectral image of the target domain into XT；
(2b) Combining spectral features of a source domain and a target domain into a vector set X ═ XS XT]；
(3) Constructing three layers of conditional probability distribution adaptation networks, and carrying out edge probability distribution adaptation of the hyperspectral images of the source domain and the target domain:
(3a) constructing a first layer edge probability distribution adaptation network by utilizing a linear denoising coder and a nonlinear coder:
for sample xiObtaining M disturbed versions by randomly setting zero to each dimension characteristic, wherein the mth disturbed version is x'i,m(ii) a Pair x 'by a Linear De-noising encoder'i,mThe linear implicit rays are to be restored to original,
Figure FDA0002965080290000011
the objective equation for optimizing the weight W is as follows:
Figure FDA0002965080290000012
wherein the first term is an average reconstruction error, the second term is an edge penalty term, lambda represents a balance factor, NSAnd NTRespectively representing the sample numbers of the source domain hyperspectral image and the target domain hyperspectral image, wherein N is the total sample number, namely N is equal to NS+NT(ii) a Let the perturbed sample matrix X'm＝[x′1,m,x′2,m,…,x′N,m]Original copy matrix of M times
Figure FDA0002965080290000021
M perturbed sample matrix combinations X '═ X'1,X′2,…,X′M]Thus, the above formula objective equation is converted into
Figure FDA0002965080290000022
Wherein D represents a disparity index matrix defined as
Figure FDA0002965080290000023
Thus, the linear de-noising encoder weight W has the following closed solution:
Figure FDA0002965080290000024
then, a nonlinear encoder is adopted to carry out nonlinear implicit reflection on the characteristics, and the formula is as follows
Figure FDA0002965080290000025
Here, ,
Figure FDA0002965080290000026
representing an output of the first tier edge probability distribution adaptation network;
(3b) according to the step (3a), a linear de-noising encoder and a non-linear encoder are utilized to construct a second layer edge probability distribution adaptation network, the weight of the linear de-noising encoder is solved, the spectral characteristics of a source domain and a target domain are adapted, and the output of a second layer network is obtained
Figure FDA0002965080290000027
(3c) According to the step (3a), a third layer edge probability distribution adaptation network is constructed by utilizing a linear denoising encoder and a nonlinear encoder, the weight of the linear denoising encoder is solved, the spectral characteristics of a source domain and a target domain are adapted, and the output of a third layer network is obtained
Figure FDA0002965080290000028
(4) Selecting training samples of hyperspectral images of a source domain and a target domain according to a one-to-many classification principle:
classifying 9 classes, selecting 500 samples of each class of the source domain hyperspectral image and 1% of samples of each class of the target domain hyperspectral image, forming 9 training sample sets according to a one-to-many classification principle, wherein the labels of the samples in each training sample set belong to the class to which the sample belongs and do not belong, the labels of the samples in each training sample set belong to the class to which the sample belongs are expressed as 1, and the labels of the samples in each training sample set do not belong to the class to which theThe class of the sample belongs to is represented as 0, and the sample characteristics are output of a third-layer network
Figure FDA0002965080290000029
(5) Constructing 9 three layers of conditional probability distribution adaptation networks, and performing conditional probability distribution adaptation of the hyperspectral images of the source domain and the target domain:
(5a) initializing the weight and the bias parameters of the three layers of conditional probability distribution adaptive networks, and pre-training each layer of network layer by utilizing a c training sample set; the encoding process and the decoding process of the k-th layer conditional probability distribution adaptation network are as follows:
Figure FDA0002965080290000031
Figure FDA0002965080290000032
wherein,
Figure FDA0002965080290000033
and
Figure FDA0002965080290000034
representing the input, implicit representation and output of the k-th network, respectively, f (-) and g (-) being nonlinear functions of encoding and decoding, respectively, W1 kAnd
Figure FDA0002965080290000035
respectively the weight and the bias of the encoding process,
Figure FDA0002965080290000036
and
Figure FDA0002965080290000037
weight and bias of decoding process, respectively, representing S(source domain) or T (target domain);
the target equation for the pre-training weights and bias parameters is:
Figure FDA0002965080290000038
wherein the first term represents the average reconstruction error on the training samples, the second term is a weight penalty term to prevent the weight from being too large,
Figure FDA0002965080290000039
Ntrainrepresenting the number of training samples, and lambda' representing a weight penalty factor; solving the above formula by using a back propagation algorithm;
(5b) implicit output of third-layer conditional probability distribution adaptation network after pre-training
Figure FDA00029650802900000310
As optimized sample features;
(5c) connecting a softmax classifier after the third layer of conditional probability distribution adaptation network, inputting the optimized features and class labels of the training samples into the classifier, and optimizing the weight and bias parameters of the softmax classifier;
(5d) fine-tuning the conditional probability distribution adaptive network from the top layer to the bottom layer, and further optimizing parameters of each network; the target equation for reverse fine tuning the entire network weight and bias parameters is:
Figure FDA00029650802900000311
the first term is the total reconstruction error of the coding layer, and the second term is a conditional penalty term and is used for reducing the difference of the conditional probability distribution of the source domain samples and the target domain samples;
Figure FDA0002965080290000041
and
Figure FDA0002965080290000042
respectively counting samples of a c-th source domain hyperspectral region and a target domain hyperspectral region; similarly, solving the above formula by using a back propagation algorithm to obtain parameters of each layer of the trimmed conditional probability distribution adaptive network and the softmax classifier;
(5e) according to the steps (5a) - (5d), training by utilizing 9 training sample sets respectively to obtain 9 three-layer conditional probability distribution adaptation networks, and realizing the conditional probability distribution adaptation of the hyperspectral images of the source domain and the target domain;
(6) performing one-to-many classification on the hyperspectral image of the target domain:
(6a) inputting a test sample set of a hyperspectral image of a target domain into a c conditional probability distribution adaptation network, wherein the characteristic after network optimization is the implicit output of a third layer of conditional probability distribution adaptation network
Figure FDA0002965080290000043
(6b) Classifying the test samples by using a trained softmax classifier to obtain the prediction probability belonging to the class c:
Figure FDA0002965080290000044
wherein,
Figure FDA0002965080290000045
and
Figure FDA0002965080290000046
partial weights and offsets corresponding to the softmax classifier, respectively;
(6c) according to the steps (6a) - (6b), respectively inputting the test sample set of the target domain hyperspectral image into 9 conditional probability distribution adaptation networks and softmax classifiers to obtain the probability of the sample belonging to each category;
(6d) comparing the prediction probabilities belonging to each category, and taking the category corresponding to the maximum probability as the prediction label of each sample, as follows
Figure FDA0002965080290000047
(6e) And outputting a classification result graph of the hyperspectral image of the target domain according to the prediction label vector and the spatial position of the test sample.
CN201811347067.7A 2018-11-13 2018-11-13 Hyperspectral image migration classification method based on depth joint distribution adaptive network Expired - Fee Related CN109359623B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201811347067.7A CN109359623B (en) 2018-11-13 2018-11-13 Hyperspectral image migration classification method based on depth joint distribution adaptive network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201811347067.7A CN109359623B (en) 2018-11-13 2018-11-13 Hyperspectral image migration classification method based on depth joint distribution adaptive network

Publications (2)

Publication Number Publication Date
CN109359623A CN109359623A (en) 2019-02-19
CN109359623B true CN109359623B (en) 2021-05-11

Family

ID=65344913

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201811347067.7A Expired - Fee Related CN109359623B (en) 2018-11-13 2018-11-13 Hyperspectral image migration classification method based on depth joint distribution adaptive network

Country Status (1)

Country Link
CN (1) CN109359623B (en)

Families Citing this family (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109961096B (en) * 2019-03-19 2021-01-05 大连理工大学 Multimode hyperspectral image migration classification method
CN110033045A (en) * 2019-04-17 2019-07-19 内蒙古工业大学 A kind of method and apparatus of trained identification image atomization
CN110321941A (en) * 2019-06-24 2019-10-11 西北工业大学 The Compression of hyperspectral images and classification method of identifiable feature learning
CN110598636B (en) * 2019-09-09 2023-01-17 哈尔滨工业大学 Ship target identification method based on feature migration
CN111898635A (en) * 2020-06-24 2020-11-06 华为技术有限公司 Neural network training method, data acquisition method and device
CN112016392B (en) * 2020-07-17 2024-05-28 浙江理工大学 Small sample detection method for soybean insect pest degree based on hyperspectral image
CN112331313B (en) * 2020-11-25 2022-07-01 电子科技大学 Automatic grading method for sugar net image lesions based on label coding
CN113030197B (en) * 2021-03-26 2022-11-04 哈尔滨工业大学 Gas sensor drift compensation method
CN113505856B (en) * 2021-08-05 2024-04-09 大连海事大学 Non-supervision self-adaptive classification method for hyperspectral images
CN113947126A (en) * 2021-09-07 2022-01-18 广东工业大学 Ceramic tile color classification method and device based on transfer learning
CN113947725B (en) * 2021-10-26 2024-06-14 中国矿业大学 Hyperspectral image classification method based on convolution width migration network
CN114494762B (en) * 2021-12-10 2024-09-24 南通大学 Hyperspectral image classification method based on depth migration network
CN115410088B (en) * 2022-10-10 2023-10-31 中国矿业大学 Hyperspectral image field self-adaption method based on virtual classifier

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2011026167A1 (en) * 2009-09-03 2011-03-10 National Ict Australia Limited Illumination spectrum recovery
CN106530246A (en) * 2016-10-28 2017-03-22 大连理工大学 Image dehazing method and system based on dark channel and non-local prior
CN107239759A (en) * 2017-05-27 2017-10-10 中国科学院遥感与数字地球研究所 A kind of Hi-spatial resolution remote sensing image transfer learning method based on depth characteristic
JP2018081404A (en) * 2016-11-15 2018-05-24 パナソニック インテレクチュアル プロパティ コーポレーション オブ アメリカＰａｎａｓｏｎｉｃ Ｉｎｔｅｌｌｅｃｔｕａｌ Ｐｒｏｐｅｒｔｙ Ｃｏｒｐｏｒａｔｉｏｎ ｏｆ Ａｍｅｒｉｃａ Discrimination method, discrimination device, discriminator generation method and discriminator generation device
CN108171232A (en) * 2017-11-15 2018-06-15 中山大学 The sorting technique of bacillary and viral children Streptococcus based on deep learning algorithm
CN108280396A (en) * 2017-12-25 2018-07-13 西安电子科技大学 Hyperspectral image classification method based on depth multiple features active migration network

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2011026167A1 (en) * 2009-09-03 2011-03-10 National Ict Australia Limited Illumination spectrum recovery
CN106530246A (en) * 2016-10-28 2017-03-22 大连理工大学 Image dehazing method and system based on dark channel and non-local prior
JP2018081404A (en) * 2016-11-15 2018-05-24 パナソニック インテレクチュアル プロパティ コーポレーション オブ アメリカＰａｎａｓｏｎｉｃ Ｉｎｔｅｌｌｅｃｔｕａｌ Ｐｒｏｐｅｒｔｙ Ｃｏｒｐｏｒａｔｉｏｎ ｏｆ Ａｍｅｒｉｃａ Discrimination method, discrimination device, discriminator generation method and discriminator generation device
CN107239759A (en) * 2017-05-27 2017-10-10 中国科学院遥感与数字地球研究所 A kind of Hi-spatial resolution remote sensing image transfer learning method based on depth characteristic
CN108171232A (en) * 2017-11-15 2018-06-15 中山大学 The sorting technique of bacillary and viral children Streptococcus based on deep learning algorithm
CN108280396A (en) * 2017-12-25 2018-07-13 西安电子科技大学 Hyperspectral image classification method based on depth multiple features active migration network

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
Reduced Encoding Diffusion Spectrum Imaging Implemented With a Bi-Gaussian Model;Chun-Hung Yeh;《IEEE Transactions on Medical Imaging 》;20080722;第1415 - 1424页 *
基于分类器集成的高光谱遥感图像分类方法;樊利恒;《光学学报》;20140930;第99-109页 *

Also Published As

Publication number Publication date
CN109359623A (en) 2019-02-19

Similar Documents

Publication Publication Date Title
CN109359623B (en) Hyperspectral image migration classification method based on depth joint distribution adaptive network
CN110321963B (en) Hyperspectral image classification method based on fusion of multi-scale and multi-dimensional space spectrum features
CN111461258B (en) Remote sensing image scene classification method of coupling convolution neural network and graph convolution network
Lek et al. Artificial neuronal networks: application to ecology and evolution
CN113095409B (en) Hyperspectral image classification method based on attention mechanism and weight sharing
CN107145836B (en) Hyperspectral image classification method based on stacked boundary identification self-encoder
CN112949416B (en) Supervised hyperspectral multiscale graph volume integral classification method
CN111639587B (en) Hyperspectral image classification method based on multi-scale spectrum space convolution neural network
Zhao et al. Toward a better understanding of coastal salt marsh mapping: A case from China using dual-temporal images
CN107590515A (en) The hyperspectral image classification method of self-encoding encoder based on entropy rate super-pixel segmentation
CN113255972B (en) Short-term rainfall prediction method based on Attention mechanism
CN109145832A (en) Polarimetric SAR image semisupervised classification method based on DSFNN Yu non local decision
Dalal et al. Optimized deep learning with learning without forgetting (LwF) for weather classification for sustainable transportation and traffic safety
Novelli et al. Improving the ANN classification accuracy of landsat data through spectral indices and linear transformations (PCA and TCT) aimed at LU/LC monitoring of a river basin
CN109671019A (en) A kind of remote sensing image sub-pixed mapping drafting method based on multi-objective optimization algorithm and sparse expression
CN108460326A (en) A kind of high spectrum image semisupervised classification method based on sparse expression figure
CN112836736B (en) Hyperspectral image semi-supervised classification method based on depth self-encoder composition
CN117611908A (en) Accurate classification of crops of sky ground integration large scene hyperspectral remote sensing image
CN116994071A (en) Multispectral laser radar point cloud classification method based on self-adaptive spectrum residual error
Niang et al. Digital Mapping of Soil Drainage Classes Using Multitemporal RADARSAT‐1 and ASTER Images and Soil Survey Data
CN117132884A (en) Crop remote sensing intelligent extraction method based on land parcel scale
CN115965867A (en) Remote sensing image earth surface coverage classification method based on pseudo label and category dictionary learning
CN110533118A (en) Remote sensing images sparse expression classification method based on Multiple Kernel Learning
Li et al. Change detection in synthetic aperture radar images based on log-mean operator and stacked auto-encoder
Walter et al. Mapping of Phragmites in estuarine wetlands using high-resolution aerial imagery

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

Granted publication date: 20210511

Termination date: 20211113