CN109858557A - A kind of new hyperspectral image data semisupervised classification algorithm - Google Patents
A kind of new hyperspectral image data semisupervised classification algorithm Download PDFInfo
- Publication number
- CN109858557A CN109858557A CN201910113156.3A CN201910113156A CN109858557A CN 109858557 A CN109858557 A CN 109858557A CN 201910113156 A CN201910113156 A CN 201910113156A CN 109858557 A CN109858557 A CN 109858557A
- Authority
- CN
- China
- Prior art keywords
- data
- label
- cnn1
- primary
- new
- 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.)
- Granted
Links
Landscapes
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of new hyperspectral image data semisupervised classification algorithms, and this method is to get all high-spectral datas first, and as primary data;Secondly using sparse subspace clustering as clustering method polymerize all primary datas, and obtains the pseudo label of all primary datas;Secondly pre-training is carried out to all primary datas and its pseudo label in CNN1;For the objective function in layer of classifying in CNN1, the new objective function based on metric learning is proposed;Last classification layer is removed using CNN1 later, along with full articulamentum constitutes CNN2;The data of tape label and their known true tag are finely adjusted CNN2 later.The present invention is the semi-supervised deep learning mechanism by carrying out classification hyperspectral imagery using pseudo label;With two-part training, training information is extended;The problem of in order to preferably grasp EO-1 hyperion composition of sample feature, new objective function is proposed in the pre-training stage.
Description
Technical field
The invention belongs to classification hyperspectral imagery fields, are related to data semisupervised classification technology, specifically a kind of new height
Spectral image data semisupervised classification algorithm.
Background technique
By the development of the second half in 20th century, great variation is had occurred in theoretical, technology and application aspect in remote sensing technology.
Wherein, high spectrum image remote sensing technology is undoubtedly highly important one aspect in remote sensing technology great change.High spectrum image
Remote sensing technology, it is same to target area with tens of to hundreds of wave bands continuously segmented using bloom spectrum sensor/imaging spectrometer
When be imaged, by image in conjunction with spectrum, while obtaining space and spectral information, obtain the high spectrum image as unit of pixel.
The image contains a large amount of spectral information, can acquire target in the ultraviolet of electromagnetic spectrum, visible light, near-infrared and middle infrared
The radiation in domain or reflective information.
High-spectrum remote sensing data contains spatial information and spectral information abundant.The most important feature of high spectrum image is just
It is the continuous spectrum and spatial surface image that can obtain atural object simultaneously, spectrum is tieed up into information and traditional space dimension image information
It is mused together.The process of classification hyperspectral imagery is exactly to be believed according to the spectral signature information and spatial character of image picture elements
Breath divides the process of classification to the pixel group of each pixel or more homogeneous.
Summary of the invention
The invention proposes a kind of hyperspectral image data semisupervised classification algorithm, which is integrated in probabilistic framework
Spectral information and spatial information;The pseudo label progress pre-training that sparse subspace clustering obtains total data is first passed through, and pre-
Training stage has added metric learning regularization term to construct new objective function;Then it is carried out with the data of tape label and its label
Fine tuning;Then it uses Markov random field as post processing of image, further utilizes spatial information.
The purpose of the present invention is to provide a kind of new hyperspectral image data semisupervised classification algorithms.
The purpose of the present invention can be achieved through the following technical solutions:
A kind of new hyperspectral image data semisupervised classification algorithm, this method include the following steps:
Step 1: all high-spectral datas are got, and as primary data;
Step 2: using sparse subspace clustering as clustering method and polymerize all primary datas, and obtains all
The pseudo label of primary data;The pseudo label of primary data is directly obtained by sparse subspace clustering;
Step 3: pre-training is carried out to all primary datas and its pseudo label in CNN1;
For the objective function in layer of classifying in CNN1, the new objective function based on metric learning is proposed;
Step 4: removing last classification layer using CNN1, along with full articulamentum constitutes CNN2;
Step 5: the data of tape label and their label are finely adjusted on CNN2;
Step 6: after layer of classifying, by Markov random field, using the spatial information of high spectrum image, into one
Step post-processes classification results, improves nicety of grading.
Further, the primary data includes label high-spectral data and unmarked high-spectral data;Mark is utilized simultaneously
Note high-spectral data and unmarked high-spectral data are trained CNN1.
Further, CNN1 is to refer to initial deep layer convolutional neural networks in the step 3.
Further, pre-training refers to whole primary data and its pseudo label in the step 3, is put into nerve net
It is trained in network CNN1;Pre-training can by a large amount of primary datas come it is preparatory training network CNN1 and obtain one group it is appropriate
Parameter may learn feature;Because the data marked in primary data are few, used here unlabelled high-spectral data and
The high-spectral data of label.
Further, CNN2 is to refer to new deep layer convolutional neural networks in the step 4;It is eliminated by CNN1
Last classification layer, then plus full articulamentum constitute;It is exactly softmax cross entropy for the objective function for layer of classifying in CNN2
Loss function.
Further, in the step 5 fine tuning refer to data with true tag and its label as primary data,
It is put into CNN2 and is trained.
Further, its spatial information, specific manifestation are further utilized by Markov random field in the step 6
For in image classification task, Markov random field encourages adjacent pixel class label having the same;This can further be mentioned
The classification accuracy of high classification hyperspectral imagery task;The effect of Markov random field assumes that adjacent pixel is more likely to belong to
Same class encourages neighborhood pixels to belong to similar label.
Beneficial effects of the present invention:
The present invention is by utilizing the pseudo- semi-supervised deep learning mechanism for marking progress classification hyperspectral imagery;Use two-part
Training, carrys out spread training information.Simultaneously the invention proposes by optimizing new target loss function, it joined and portray EO-1 hyperion
The metric learning regular conditions item of data characteristics;Classification results are post-processed finally by Markov random field, energy
Enough accomplish the spatial information for further utilizing high spectrum image, to realize spatial information and the spectrum letter to high-spectral data
The combined extracting of breath, i.e. " unification of sky spectrum ", therefore effectively improve final classifying quality.
Detailed description of the invention
In order to facilitate the understanding of those skilled in the art, the present invention will be further described below with reference to the drawings.
Fig. 1 is hyperspectral image data semisupervised classification algorithm flow chart of the present invention;
Fig. 2 is the structure chart of the semi-supervised deep learning frame of the present invention;
Fig. 3 is pre-training of the present invention and the frame diagram for finely tuning the stage.
Specific embodiment
As shown in Figure 1-3, a kind of new hyperspectral image data semisupervised classification algorithm, this method include the following steps:
Step 1: all high-spectral datas are got, and as primary data;The primary data includes label
High-spectral data and unmarked high-spectral data;
Step 2: using sparse subspace clustering as clustering method and polymerize all primary datas, and obtains all
The pseudo label of primary data;
Wherein, the pseudo label of primary data is directly obtained by sparse subspace clustering;Sparse subspace clustering algorithm
Main thought is that any one sample can be realized infinite number of possible expression by other sample points, then from same
The sample of subspace just can be carried out mutual expression, just there is expression coefficient, it is meant that most of to be in this infinite number of expression
Number is 0, is unrelated sample, just produces rarefaction representation;
Step 3: pre-training is carried out to all primary datas and its pseudo label in CNN1;Pre-training refers to all
Primary data and its pseudo label, be put into neural network CNN1 and be trained.
Pre-training come preparatory training network CNN1 and can obtain one group of parameter appropriate by mass data, can learn
To feature;Because the data marked in EO-1 hyperion are few, the mass data that pre-training uses here has also used unlabelled data;
For the objective function in layer of classifying in CNN1, the new objective function based on metric learning is proposed;It is wherein new
Objective function refer to the J in Fig. 3;CNN1 is to refer to initial deep layer convolutional neural networks;
Step 4: removing last classification layer using CNN1, along with full articulamentum constitutes CNN2;CNN2 is to refer to newly
Deep layer convolutional neural networks;It is last classification layer is eliminated by CNN1, then plus full articulamentum constitute.For in CNN2
The objective function of classification layer, is softmax cross entropy loss function.
Step 5: the data of tape label and their label are finely adjusted in CNN2;Fine tuning refers to true mark
The data and its label of label are put into CNN2 and are trained as primary data.
Step 6: after high spectrum image carries out classifying step, its sky is further utilized by Markov random field
Between information.
Concrete principle of the invention may be interpreted as, and the process of classification hyperspectral imagery is exactly the Spectral Properties according to image picture elements
Reference breath and spatial character information, the process of classification is divided to the pixel group of each pixel or more homogeneous;
Firstly, because the sample of high-spectral data label is few, the data only marked on a small quantity are for training, so we make
Sparse subspace clustering is used as clustering method to polymerize all primary datas and obtain the pseudo label of all data, all is first
Beginning data include the high-spectral data not marked and the high-spectral data of tape label, because sparse subspace clustering is a kind of excellent
The elegant clustering method based on sparse expression has good application effect in many fields;
Next, all primary datas and its pseudo label are used for pre-training deep layer convolutional neural networks;With only make cross entropy
Traditional CNN model of minimization of loss is different, it is proposed that model be to be trained by optimizing new identification objective function
's;For this purpose, we are also added into a metric learning regularization term other than minimizing softmax cross entropy loss function;
Metric learning regularization term refers to the comparison loss in utilization measure study, it is combined with cross entropy then, is constructed
New objective function in pre-training stage CNN1;The object of metric learning, study is usually the distance of sampling feature vectors, measurement
The distance between the destination of study is by training and to learn, reduce or limit similar sample, while increasing different classes of sample
The distance between.Because there are different objects in high spectrum image with same spectra and same object has showing for different spectrum
As.Similitude is still two big challenges between diversity and class in class, these challenges can generally make the performance of classification hyperspectral imagery
It degenerates.It can make to gather in class by being added after metric learning regularization term, disperse between class.
Then, we delete the last layer from network trained in advance, and add the layer connected entirely and softmax classification
Layer constructs new depth network C NN2;It is new to finely tune this to only use the label data with true class label for we at this time
Depth network C NN2;
Finally, we are further believed using space after classification hyperspectral imagery step using Markov random field
Breath;In image classification task, Markov random field encourages adjacent pixel class label having the same;This has been demonstrated can be with
Greatly improve the classification accuracy of classification hyperspectral imagery task;The effect of Markov random field assumes that adjacent pixel more
Same class may be belonged to, neighborhood pixels is encouraged to belong to similar label.
The present invention, which passes through, utilizes the pseudo- semi-supervised deep learning for marking progress classification hyperspectral imagery;It is trained with two-part,
Carry out spread training information.Simultaneously the invention proposes by optimizing new target loss function, it is explicitly emphasized that the degree of CNN feature
Amount study regularization term, and by combining Markov random field and CNN, it can be realized so more preferable that utilize space
Information realizes the empty spectrum information combined extracting in hyperspectral classification.
Above content is only to structure of the invention example and explanation, affiliated those skilled in the art couple
Described specific embodiment does various modifications or additions or is substituted in a similar manner, without departing from invention
Structure or beyond the scope defined by this claim, is within the scope of protection of the invention.
Claims (7)
1. a kind of new hyperspectral image data semisupervised classification algorithm, which is characterized in that this method includes the following steps:
Step 1: all high-spectral datas are got, and as primary data;
Step 2: using sparse subspace clustering as clustering method and polymerize all primary datas, and obtains all initial
The pseudo label of data;The pseudo label of primary data is directly obtained by sparse subspace clustering;
Step 3: pre-training is carried out to all primary datas and its pseudo label in CNN1;
For the objective function in layer of classifying in CNN1, the new objective function based on metric learning is proposed;
Step 4: removing last classification layer using CNN1, along with full articulamentum constitutes CNN2;
Step 5: the data of tape label and their label are finely adjusted on CNN2;
Step 6: further right using the spatial information of high spectrum image by Markov random field after layer of classifying
Classification results are post-processed, and nicety of grading is improved.
2. the new hyperspectral image data semisupervised classification algorithm of one kind according to claim 1, which is characterized in that described
Primary data includes label high-spectral data and unmarked high-spectral data;Label high-spectral data and unmarked height are utilized simultaneously
Spectroscopic data is trained CNN1.
3. the new hyperspectral image data semisupervised classification algorithm of one kind according to claim 1, which is characterized in that described
CNN1 is to refer to initial deep layer convolutional neural networks in step 3.
4. the new hyperspectral image data semisupervised classification algorithm of one kind according to claim 1, which is characterized in that described
Pre-training refers to whole primary data and its pseudo label in step 3, is put into neural network CNN1 and is trained;Pre- instruction
White silk come preparatory training network CNN1 and can obtain one group of parameter appropriate by a large amount of primary datas, may learn feature;
Because the data marked in primary data are few, the high-spectral data of unlabelled high-spectral data and label has been used here.
5. the new hyperspectral image data semisupervised classification algorithm of one kind according to claim 1, which is characterized in that described
CNN2 is to refer to new deep layer convolutional neural networks in step 4;It is that last classification layer, then plus Quan Lian are eliminated by CNN1
Connect layer composition;It is exactly softmax cross entropy loss function for the objective function for layer of classifying in CNN2.
6. the new hyperspectral image data semisupervised classification algorithm of one kind according to claim 1, which is characterized in that described
Fine tuning refers to that data with true tag and its label as primary data, are put into CNN2 and are trained in step 5.
7. the new hyperspectral image data semisupervised classification algorithm of one kind according to claim 1, which is characterized in that described
Its spatial information is further utilized by Markov random field in step 6, is embodied in image classification task, horse
Er Kefu random field encourages adjacent pixel class label having the same;This can be further improved classification hyperspectral imagery task
Classification accuracy;The effect of Markov random field assumes that adjacent pixel is more likely to belong to same class, encourages neighborhood pixels category
In similar label.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910113156.3A CN109858557B (en) | 2019-02-13 | 2019-02-13 | Novel semi-supervised classification method for hyperspectral image data |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910113156.3A CN109858557B (en) | 2019-02-13 | 2019-02-13 | Novel semi-supervised classification method for hyperspectral image data |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109858557A true CN109858557A (en) | 2019-06-07 |
CN109858557B CN109858557B (en) | 2022-12-27 |
Family
ID=66897911
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910113156.3A Active CN109858557B (en) | 2019-02-13 | 2019-02-13 | Novel semi-supervised classification method for hyperspectral image data |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109858557B (en) |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110309868A (en) * | 2019-06-24 | 2019-10-08 | 西北工业大学 | In conjunction with the hyperspectral image classification method of unsupervised learning |
CN111401426A (en) * | 2020-03-11 | 2020-07-10 | 西北工业大学 | Small sample hyperspectral image classification method based on pseudo label learning |
CN112541904A (en) * | 2020-12-16 | 2021-03-23 | 西安电子科技大学 | Unsupervised remote sensing image change detection method, storage medium and computing device |
CN112784611A (en) * | 2021-01-21 | 2021-05-11 | 阿里巴巴集团控股有限公司 | Data processing method, device and computer storage medium |
CN113361628A (en) * | 2021-06-24 | 2021-09-07 | 海南电网有限责任公司电力科学研究院 | CNN insulator aging spectrum classification method under multi-task learning |
WO2022199636A1 (en) * | 2021-03-24 | 2022-09-29 | Ping An Technology (Shenzhen) Co., Ltd. | Method, device, and storage medium for semi-supervised learning for bone mineral density estimation in hip x-ray images |
CN115546567A (en) * | 2022-12-01 | 2022-12-30 | 成都考拉悠然科技有限公司 | Unsupervised field adaptive classification method, system, equipment and storage medium |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2017158058A1 (en) * | 2016-03-15 | 2017-09-21 | Imra Europe Sas | Method for classification of unique/rare cases by reinforcement learning in neural networks |
CN107958216A (en) * | 2017-11-27 | 2018-04-24 | 沈阳航空航天大学 | Based on semi-supervised multi-modal deep learning sorting technique |
-
2019
- 2019-02-13 CN CN201910113156.3A patent/CN109858557B/en active Active
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2017158058A1 (en) * | 2016-03-15 | 2017-09-21 | Imra Europe Sas | Method for classification of unique/rare cases by reinforcement learning in neural networks |
CN107958216A (en) * | 2017-11-27 | 2018-04-24 | 沈阳航空航天大学 | Based on semi-supervised multi-modal deep learning sorting technique |
Non-Patent Citations (1)
Title |
---|
李绣心等: "基于卷积神经网络的半监督高光谱图像分类", 《电子测量与仪器学报》 * |
Cited By (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110309868A (en) * | 2019-06-24 | 2019-10-08 | 西北工业大学 | In conjunction with the hyperspectral image classification method of unsupervised learning |
CN111401426A (en) * | 2020-03-11 | 2020-07-10 | 西北工业大学 | Small sample hyperspectral image classification method based on pseudo label learning |
CN111401426B (en) * | 2020-03-11 | 2022-04-08 | 西北工业大学 | Small sample hyperspectral image classification method based on pseudo label learning |
CN112541904A (en) * | 2020-12-16 | 2021-03-23 | 西安电子科技大学 | Unsupervised remote sensing image change detection method, storage medium and computing device |
CN112541904B (en) * | 2020-12-16 | 2023-03-24 | 西安电子科技大学 | Unsupervised remote sensing image change detection method, storage medium and computing device |
CN112784611A (en) * | 2021-01-21 | 2021-05-11 | 阿里巴巴集团控股有限公司 | Data processing method, device and computer storage medium |
WO2022199636A1 (en) * | 2021-03-24 | 2022-09-29 | Ping An Technology (Shenzhen) Co., Ltd. | Method, device, and storage medium for semi-supervised learning for bone mineral density estimation in hip x-ray images |
CN113361628A (en) * | 2021-06-24 | 2021-09-07 | 海南电网有限责任公司电力科学研究院 | CNN insulator aging spectrum classification method under multi-task learning |
CN115546567A (en) * | 2022-12-01 | 2022-12-30 | 成都考拉悠然科技有限公司 | Unsupervised field adaptive classification method, system, equipment and storage medium |
Also Published As
Publication number | Publication date |
---|---|
CN109858557B (en) | 2022-12-27 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109858557A (en) | A kind of new hyperspectral image data semisupervised classification algorithm | |
CN106203523B (en) | The hyperspectral image classification method of the semi-supervised algorithm fusion of decision tree is promoted based on gradient | |
CN106529508B (en) | Based on local and non local multiple features semanteme hyperspectral image classification method | |
CN108460391B (en) | Hyperspectral image unsupervised feature extraction method based on generation countermeasure network | |
CN110084159A (en) | Hyperspectral image classification method based on the multistage empty spectrum information CNN of joint | |
CN103886342B (en) | Hyperspectral image classification method based on spectrums and neighbourhood information dictionary learning | |
CN106503739A (en) | The target in hyperspectral remotely sensed image svm classifier method and system of combined spectral and textural characteristics | |
CN106845418A (en) | A kind of hyperspectral image classification method based on deep learning | |
CN103208011B (en) | Based on average drifting and the hyperspectral image space-spectral domain classification method organizing sparse coding | |
CN107657271B (en) | Hyperspectral image classification method based on long memory network in short-term | |
CN105261000A (en) | Hyperspectral image fusion method based on end member extraction and spectrum unmixing | |
CN110533077A (en) | Form adaptive convolution deep neural network method for classification hyperspectral imagery | |
CN105184302B (en) | A kind of high optical spectrum image end member extraction method | |
CN111222545B (en) | Image classification method based on linear programming incremental learning | |
CN113920376B (en) | Wheat seed variety identification method based on lightweight convolutional neural network | |
Dong et al. | Joint contextual representation model-informed interpretable network with dictionary aligning for hyperspectral and LiDAR classification | |
CN109002771A (en) | A kind of Classifying Method in Remote Sensing Image based on recurrent neural network | |
Li et al. | Airplane detection using convolutional neural networks in a coarse-to-fine manner | |
Nogueira et al. | Spatio-temporal vegetation pixel classification by using convolutional networks | |
CN108932474A (en) | A kind of remote sensing image cloud based on full convolutional neural networks compound characteristics sentences method | |
CN108460326A (en) | A kind of high spectrum image semisupervised classification method based on sparse expression figure | |
CN107169407A (en) | Hyperspectral image classification method based on joint bilateral filtering and extreme learning machine | |
CN104616283B (en) | A kind of pest detection method based on low-rank representation algorithm | |
Yadav et al. | Analysis and classification of hardwood species based on Coiflet DWT feature extraction and WEKA workbench | |
CN112257531A (en) | Forest land change remote sensing monitoring method based on diversity characteristic combination |
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 |