CN107529647A - A kind of cloud atlas cloud amount computational methods based on the unsupervised sparse learning network of multilayer - Google Patents
A kind of cloud atlas cloud amount computational methods based on the unsupervised sparse learning network of multilayer Download PDFInfo
- Publication number
- CN107529647A CN107529647A CN201611107944.4A CN201611107944A CN107529647A CN 107529647 A CN107529647 A CN 107529647A CN 201611107944 A CN201611107944 A CN 201611107944A CN 107529647 A CN107529647 A CN 107529647A
- Authority
- CN
- China
- Prior art keywords
- cloud
- multilayer
- unsupervised
- layer
- network
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Landscapes
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The present invention relates to a kind of cloud atlas cloud amount computational methods based on the unsupervised sparse learning network of multilayer, first with forward direction, successively the sparse automatic coding machine successively feature coding unsupervised to picture progress obtains high-order semantic information, then cloud atlas is divided into spissatus, thin cloud and clear sky using high-order semantic information, finally utilizes " Spatial coherence method " to calculate the total amount of cloud in cloud atlas.The present invention is higher than the degree of accuracy that conventional satellite cloud atlas cloud amount calculates, and the time that the time of sample training and cloud amount calculate under the conditions of same hardware obtains very big reduction.
Description
Technical field
It is more particularly to a kind of to be based on the unsupervised sparse learning network of multilayer the invention belongs to cloud technical field of image processing
Cloud atlas cloud amount computational methods.
Background technology
Cloud covers more than 50% earth surface, is one of important meteorological and climatic elements.In order to obtain accurately
Cloud amount is distributed, and the detection and classification racked first is entered to satellite cloud picture, then cloud amount calculating is carried out on the basis of cloud classification.Mesh
Before, satellite cloudiness computational methods mainly have ISCCP methods, CLAVR methods, APOLLO methods etc., the vacation of ISCCP algorithms in the world
If observation radiation value comes solely from cloud and clear sky one of both, pixel observes radiation value compared with the radiation value of clear sky pair, if two
When difference between person is more than the maximum changing amplitude of clear sky radiance in itself, it is cloud to judge the pixel；CLAVR algorithms be with
2x2 matrix-blocks are as detection unit, and when four pixels all do not pass through cloud detection, trip current is cloudless, when all passing through detection
Cloud is judged to, otherwise it is assumed that being mixed type.For mixed type matrix, if cloud and clear sky, existing matrix meets other jointly
Such as ice/snow clear sky judgment condition, the matrix are judged to clear sky again；APOLLO algorithms use the form of logical AND, that is, only have
When pixel meets all threshold test conditions, it is clear sky just to think the pixel, denies thinking that the pixel is cloud.In addition also
There is the methods of MODIS, NIR/VIS.
Above-mentioned cloud amount computational methods can be divided mainly into two classes：First, by having the ratio between cloud pixel and total pixel in region
Calculate cloud amount；Second, equivalent cloud amount is calculated by the ratio between pixel amount of radiation and reflectivity.First method is easy to operate, but by
In sub-pixed mapping cloud amount can not be analyzed, cloud amount result of calculation can be caused higher；Although second method solves Asia to a certain extent
Pixel cloud amount problem, but some situations still less be applicable, such as have in region multi layer cloud or ground surface type change acutely etc..Yun Jian
Survey is the basis that cloud amount calculates, therefore the accuracy rate for wanting to improve cloud amount calculating must first obtain preferable cloud detection result.
The technology of cloud detection at present is broadly divided into two classes:Threshold method and clustering methodology.Threshold method mainly uses infrared temperature
Threshold value, visible photo threshold etc. are spent, but because satellite image is extremely complex changeable, image is examined using fixed global threshold
Survey can produce bigger error.Clustering method mainly has histogram cluster, adaptive threshold cluster, dynamic threshold cluster
Deng, but the classification of cloud generally has a lot of features, research at this stage is carried out mainly for a certain feature of cloud, it is impossible to very well
Extraction cloud atlas picture on effective information, therefore the cloud detection effect of above method is not fine.It is additionally based on machine learning
Cloud detection method of optic is also widely applied, mainly including SVMs, k nearest neighbor, Fuzzy strategy and neutral net, wherein nerve
The accuracy of detection of network is more preferable than other method effect, but it equally exists defect, is mainly shown as and does not use cloud fully
Figure feature, effective information extraction are inadequate.Deep learning development in recent years is very fast, wherein just including being developed by extreme learning machine
And the depth limit learning machine come, it all shows powerful adaptability and robustness in many application fields, and learns speed
Degree is fast.Multilayer perceptron containing more hidden layers of the depth limit learning machine as particular design, can fully extract validity feature, right
Cloud atlas is classified.But most depth convolutional neural networks of current application are due to needing error back propagation, and
Parameter amount is especially big, therefore the efficiency for learning and classifying is very low, have impact on the application on site of this method.
The content of the invention
The purpose of the present invention is to overcome the shortcomings of above-mentioned background technology, there is provided one kind is based on the unsupervised sparse study net of multilayer
The cloud atlas cloud amount computational methods of network, overcome the defects of traditional neural network is inadequate to Cloud-Picture Characteristics utilization rate.On realizing
Technical purpose is stated, the technical scheme is that：
Based on the cloud atlas cloud amount computational methods of the unsupervised sparse learning network of multilayer, comprise the following steps：
(1) training of the unsupervised sparse learning network model structure of multilayer：Neutral net is set as one containing m hidden layer
Network, utilize the sample (X markedi,Yi) image block, the neutral net is learnt, operating limit learning machine model enters
The unsupervised feature learning of row multilayer, obtains optimal network parameter, wherein, XiFor n × n image block, YiRepresent XiIt is corresponding
Cloud classification, n meets that 10≤n≤50, i represent i-th of sample, i=1,2,3 ..., p, p be total sample number；The network
The output layer of model realizes that cloud classification learns by the way of supervised learning；
(2) satellite photo cloud classification：Satellite photo is divided into the fritter that each pixel size is n × n, as neutral net
Input data, and obtain the output of whole neutral net, final classification carried out by output valve, sentenced according to the maximum of output
Determine the species of cloud；
(3) satellite photo amount calculates：According to the species of step (2) medium cloud, cloud atlas is calculated using space correlation method
On total amount of cloud.
The further design of the cloud atlas cloud amount computational methods based on the unsupervised sparse learning network of multilayer is, described
The unsupervised feature extraction network model of multilayer includes：
The unsupervised characteristic extraction part of multilayer, including by the limit learn based on automatic sparse coding realize the nothing of feature
Supervision extraction, the feature of input data is extracted by partially connected；
Supervised learning part based on extreme learning machine, using the characteristic vector of last layer of extraction of hidden layer as last
One layer of input, based on the theoretical progress tagsort study of extreme learning machine, the network parameter of last layer is obtained,
The further design of the cloud atlas cloud amount computational methods based on the unsupervised sparse learning network of multilayer is, step
(1) the unsupervised feature extraction network model of multilayer contains m hidden layer, wherein i-th of hidden layer, which uses, has viIndividual implicit node
Extreme learning machine carries out automatic sparse coding study, and the hidden layer of the extreme learning machine is exported as the unsupervised feature of multilayer
Extract network i+1 layer input, using the connection weight between the hidden layer and output layer of extreme learning machine as multilayer without prison
Superintend and direct i-th layer of connection weight between i+1 of feature extraction network.
The further design of the cloud atlas cloud amount computational methods based on the unsupervised sparse learning network of multilayer is, every layer
Network carries out feature extraction and specifically comprised the following steps：
A) the satellite photo sample X of input passes through H1(X)=g (w1·X+b1)·β1After obtain random character mapping, wherein
w1It is that unsupervised multilayer feature extracts the weights between first autocoding limit study input layer and hidden layer, b in network1
It is biasing, g is excitation function, β1It is the weight between hidden layer and output layer；
B) output of the hidden layer of extreme learning machine can be inputted by the connection between hidden layer and output layer
Reconstruct, that is, cause H1(X)=X；
C) β is obtained by way of limit study and sparse coding1Estimate β1, unsupervised multilayer feature extraction network
The feature output K of first layer1=g (w1·X+b1) the just input as the unsupervised multilayer feature extraction network second layer, β1With regard to making
The connection weight between network first tier and the second layer is extracted for unsupervised multilayer feature；
D) the output K of unsupervised multilayer feature extraction network jth layerjCan be used as unsupervised multilayer feature extraction network jth+
1 layer of input, βjAs unsupervised multilayer feature extract+1 layer of network jth layer and unsupervised multilayer feature extraction network jth it
Between connection weight, repeat step d) networks and carry out feature extraction to finish.
The further design of the cloud atlas cloud amount computational methods based on the unsupervised sparse learning network of multilayer is, described
The specific algorithm of sparse coding is as follows in step c)：
First, coding input is set as R, there is L hidden node, connection weights of the W between input layer and hidden layer, β
Connection weight between hidden layer and output layer, biIt is the biasing (b=(b of i-th of implicit node1,b2,...,bL)) and it is full
Sufficient g (WR+b) β=X；
Then, according toUsing l1Norm constraint β so that β is sparse, wherein W and b
It is the value between any given 0 to 1；
Finally, estimate β is obtained by rapid desufflation threshold value iterative method.
The further design of the cloud atlas cloud amount computational methods based on the unsupervised sparse learning network of multilayer is, described
Cloud classification learns, and concretely comprises the following steps：The feature output of unsupervised m-th of the hidden layer of feature extraction network of multilayer is learned as supervision
The input of part is practised, is exported by the use of the target of mark Y as the supervised learning of sample, and supervised by extreme learning machine method
Learn the network parameter of part.
The further design of the cloud atlas cloud amount computational methods based on the unsupervised sparse learning network of multilayer is, described
The cloud classification result of neutral net output includes three kinds：Clear sky, Bao Yun and spissatus.
The further design of the cloud atlas cloud amount computational methods based on the unsupervised sparse learning network of multilayer is, described
The result of the cloud classification of neutral net output has three kinds, and setting output sample is set as corresponding three-dimensional vector.
The further design of the cloud atlas cloud amount computational methods based on the unsupervised sparse learning network of multilayer is, step
(3) specifically include：The cloud classification result exported by neutral net obtains the lower bound I of spissatus pixel intensitycldWith clear sky pixel
The upper bound I of brightnessclr, IcldIt is the minimum value of all spissatus brightness, IclrIt is the maximum of all clear sky brightness, it is every on cloud atlas
The cloud amount of individual pixel is tried to achieve according to following formula：
Ac=(I-Iclr)/(Icld-Iclr)
Wherein I is the brightness of single pixel receiver, finally gives the total amount of cloud of all pixels on cloud atlas.
Beneficial effects of the present invention are：
The present invention ensure that network has well general by the unsupervised feature extraction of multilayer of the sparse study of successively forward direction
Change performance, the cloud feature being fully extracted in cloud atlas, cloud amount is calculated using improved space correlation method, result of calculation is more smart
Really.The thought of limit of utilization study of the present invention realizes the autocoding of feature, and utilizes l1Norm constraint realizes sparse coding, this
Invention does not need reverse error to learn, and is a kind of unsupervised forward coding mode, and changing conventional depth study needs error
The mode of backpropagation, and sparse coding can be realized.Which is on the premise of the degree of accuracy improves, in equal hardware condition
The training speed of lower sample improves more than 100 times than traditional depth convolutional neural networks, and the classification speed of sample improves 10 times
More than.
Brief description of the drawings
Fig. 1 is the basic flow sheet of the present invention.
Fig. 2 is the unsupervised feature extraction network structure of multilayer for the sparse study of successively forward direction used in the present invention.
Embodiment
The application is described in further detail with reference to the accompanying drawings and detailed description.
Below with reference to accompanying drawing, technical scheme is described in detail.
As shown in figure 1, the Cloud of the unsupervised feature extraction network of multilayer of the sparse study of successively forward direction of the present embodiment
Figure cloud amount computational methods, comprise the following steps：
Step 1：The training of the unsupervised sparse learning network model structure of multilayer：It is hidden containing m as one to set neutral net
The network of layer, utilizes the sample (X markedi,Yi), XiFor the image block of a n × n (10≤n≤50, n are positive integer),
Show XiCorresponding cloud classification (is divided into spissatus, thin cloud and clear sky) in the present invention, and i represents i-th of sample, i=1,2,3 ...,
P, p are total sample number) network is learnt, operating limit learning machine and sparse representation model carry out the unsupervised feature of multilayer
Study, obtains optimal network parameter.Output layer realizes that cloud classification learns by the way of supervised learning.The unsupervised spy of multilayer
Sign extraction network model is mainly made up of the unsupervised characteristic extraction part of multilayer and the supervised learning part based on extreme learning machine.
The unsupervised characteristic extraction part of multilayer, including by the limit learn based on automatic sparse coding realize that the unsupervised of feature carries
Take, the feature of input data is extracted by partially connected.Supervised learning part, by hidden layer last layer extraction feature to
The input as last layer is measured, based on the theoretical progress tagsort study of extreme learning machine, obtains the network ginseng of last layer
Number, as Fig. 2 gives the structure of the unsupervised feature extraction network of multilayer of the sparse study of successively forward direction.
Pass through H in the satellite photo sample X of the unsupervised characteristic extraction part input of multilayer1(X)=g (w1·X+b1)·β1
After obtain random character mapping, wherein w1It is first autocoding limit study input in the unsupervised feature extraction network of multilayer
Weights between layer and hidden layer, b1It is biasing, g is that (conventional neutral net excitation letter can be selected in the present invention to excitation function
Number), β1It is the weight between hidden layer and output layer；Extreme learning machine autocoder uses own coding mould in the present invention
Formula, that is to say, that what the output of the hidden layer of extreme learning machine can be inputted by the connection between hidden layer and output layer
Reconstruct, that is, cause H1(X)=X.The w in this learning process1And b1It is the value at random between given 0 to 1.In order that
Network it is more efficient, arithmetic speed faster, the present invention use l1Norm carrys out constraint beta1So that β1To be sparse.By quickly receiving
Contracting threshold value iterative method1Seek β1Estimate β1So thatIt is minimum.So multilayer is unsupervised
The feature output K of feature extraction network first tier1=g (w1·X+b1) just it is used as the unsupervised feature extraction network second layer of multilayer
Input, β1With regard to as the connection weight between the unsupervised feature extraction network first tier of multilayer and the second layer.
The output K of the unsupervised feature extraction network jth layer of multilayerjThe unsupervised feature extraction network jth+1 of multilayer can be used as
The input of layer, βj(estimate obtained by the automatic sparse coding of extreme learning machine) can be used as the unsupervised feature extraction net of multilayer
Connection weight between network jth layer and unsupervised+1 layer of the feature extraction network jth of multilayer.So unsupervised feature extraction net of multilayer
The output of+1 layer of network jth is Kj+1=g (wj+1·Kj+bj+1).Still X is reconstructed and causes g (wj+1·X+bj+1)·βj+1
=X, β is asked by rapid desufflation threshold value iterative methodj+1Estimate βj+1So thatIt is minimum.So Kj+1=g (wj+1·Kj+bj+1) multilayer can be used as without prison
Superintend and direct the input of+2 layers of feature extraction network jth, βj+1Can be used as+1 layer of the unsupervised feature extraction network jth of multilayer and multilayer without
The connection weight supervised between+2 layers of feature extraction network jth.
In supervised learning part, the feature output of unsupervised m-th of the hidden layer of feature extraction network of multilayer is learned as supervision
The input of habit, exported by the use of the target of mark Y as the supervised learning of sample.Limit of utilization learning machine method obtains supervised learning portion
The network parameter divided.
Step 2：Satellite photo cloud classification, satellite photo is divided into each pixel size, and for n × n, (10≤n≤50, n are
Positive integer) fritter, as the input data of neutral net, and obtain the output of whole network.Carried out by output valve final
Classification, the present invention are classified using softmax methods, i.e., the species of cloud is judged according to the maximum of output.In the present embodiment
The cloud classification result of neutral net output includes three kinds：Clear sky, Bao Yun and spissatus.And setting output sample is set as corresponding three
Dimensional vector.
Step 3：Satellite photo amount calculates, according to the classification results of step (2) medium cloud, using space correlation method meter
Calculate the total amount of cloud on cloud atlas.Specific method is as follows：The cloud classification result exported by neutral net
Obtain the lower bound I of spissatus pixel intensitycldWith the upper bound I of clear sky pixel intensityclr, IcldIt is all spissatus brightness
Minimum value, IclrIt is the maximum of all clear sky brightness, herein mainly makes a distinction thin cloud.Own so on satellite image
Pixel can be tried to achieve by following formula：
Ac=(I-Iclr)/(Icld-Iclr)
I is the brightness of single pixel receiver in above formula.1 is considered if result of calculation is more than 1, if result of calculation
It is considered 0 less than 0.In this case, spissatus cloud amount is defaulted as 1, and bright day cloud amount is arranged to zero.
The thought of the present embodiment limit of utilization study realizes the autocoding of feature, and realizes connection weight by sparse constraint
The rarefaction of value.The present invention does not need reverse error to learn, and is a kind of unsupervised forward coding mode, changes conventional depth
Study needs the mode of error back propagation, and can realize sparse coding.Which the degree of accuracy improve on the premise of,
The training speed of sample improves more than 100 times than traditional convolution deep learning under equal hardware condition, and the classification speed of sample carries
It is high more than 10 times.On the basis of cloud detection result, total amount of cloud computational methods of the present invention use " Spatial coherence method ".Experimental result
Show, the result that the present invention obtains is preferable, is adapted to the cloud amount of satellite cloud picture to calculate research.
The technological thought of above example only to illustrate the invention, it is impossible to protection scope of the present invention is limited with this, it is every
According to technological thought proposed by the present invention, any change done on the basis of technical scheme, the scope of the present invention is each fallen within
Within.
Claims (9)
1. a kind of cloud atlas cloud amount computational methods based on the unsupervised sparse learning network of multilayer, it is characterised in that including following step
Suddenly：
(1) training of the unsupervised sparse learning network model structure of multilayer：Neutral net is set as a net containing m hidden layer
Network, utilize the sample (X markedi,Yi), the neutral net is learnt, operating limit learning machine model carries out multilayer without prison
Feature learning is superintended and directed, obtains optimal network parameter, wherein, XiFor n × n image block, YiRepresent XiPoint of corresponding cloud
Class, n meet that 10≤n≤50, i represent i-th of sample, i=1,2,3 ..., p, p be total sample number；The network model it is defeated
Go out layer and realize that cloud classification learns by the way of supervised learning；
(2) satellite photo cloud classification：Satellite photo is divided into the fritter that each pixel size is n × n, as the defeated of neutral net
Enter data, and obtain the output of whole neutral net, final classification is carried out by output valve, cloud is judged according to the maximum of output
Species；
(3) satellite photo amount calculates：According to the species of step (2) medium cloud, calculated using space correlation method on cloud atlas
Total amount of cloud.
2. the cloud atlas cloud amount computational methods according to claim 1 based on the unsupervised sparse learning network of multilayer, its feature
It is that the unsupervised feature extraction network model of the multilayer includes：
The unsupervised characteristic extraction part of multilayer, including by the limit learn based on automatic sparse coding realize the unsupervised of feature
Extraction, the feature of input data is extracted by partially connected；
Supervised learning part based on extreme learning machine, using the characteristic vector of last layer of extraction of hidden layer as last layer
Input, carry out tagsort study based on extreme learning machine is theoretical, obtain the network parameter of last layer.
3. the cloud atlas cloud amount computational methods according to claim 2 based on the unsupervised sparse learning network of multilayer, its feature
It is：The unsupervised feature extraction network model of multilayer of step (1) contains m hidden layer, wherein i-th of hidden layer, which uses, has viIt is individual hidden
Extreme learning machine containing node carries out automatic sparse coding study, and using the hidden layer output of the extreme learning machine as multilayer without
Supervise feature extraction network i+1 layer input, using the connection weight between the hidden layer and output layer of extreme learning machine as
The unsupervised i-th layer of connection weight between i+1 of feature extraction network of multilayer.
4. the cloud atlas cloud amount computational methods according to claim 3 based on the unsupervised sparse learning network of multilayer, its feature
It is：Feature extraction is carried out per layer network to specifically comprise the following steps：
A) the satellite photo sample X of input passes through H1(X)=g (w1·X+b1)·β1After obtain random character mapping, wherein w1It is
Weights in unsupervised multilayer feature extraction network between first autocoding limit study input layer and hidden layer, b1It is inclined
Put, g is excitation function, β1It is the weight between hidden layer and output layer；
B) reconstruct that the output of the hidden layer of extreme learning machine can be inputted by the connection between hidden layer and output layer,
I.e. so that H1(X)=X；
C) β is obtained by way of limit study and sparse coding1Estimate β1, unsupervised multilayer feature extraction network first
The feature output K of layer1=g (w1·X+b1) the just input as the unsupervised multilayer feature extraction network second layer, β1With regard to as nothing
The connection weight supervised between multilayer feature extraction network first tier and the second layer；
D) the output K of unsupervised multilayer feature extraction network jth layerj+ 1 layer of network jth can be extracted as unsupervised multilayer feature
Input, βjThe company between+1 layer of network jth layer and unsupervised multilayer feature extraction network jth is extracted as unsupervised multilayer feature
Weights are connect, the progress feature extraction of step d) networks is repeated and finishes.
5. the cloud atlas cloud amount computational methods according to claim 4 based on the unsupervised sparse learning network of multilayer, its feature
It is：The specific algorithm of sparse coding is as follows in the step c)：
First, coding input is set as R, there is L hidden node, connection weights of the W between input layer and hidden layer, and β is hidden
Containing the connection weight between layer and output layer, biIt is the biasing (b=(b of i-th of implicit node1,b2,...,bL)) and meet g
(WR+b) β=X；
Then, according toUsing l1Norm constraint β so that β is sparse, and wherein W and b are to appoint
Meaning it is given 0 to 1 between value；
Finally, estimate β is obtained by rapid desufflation threshold value iterative method.
6. the cloud atlas cloud amount computational methods according to claim 1 based on the unsupervised sparse learning network of multilayer, its feature
It is：The cloud classification study, is concretely comprised the following steps：The feature of unsupervised m-th of the hidden layer of feature extraction network of multilayer is exported
As the input of supervised learning part, exported by the use of the target of mark Y as the supervised learning of sample, and pass through extreme learning machine side
Method obtains the network parameter of supervised learning part.
7. the cloud atlas cloud amount computational methods according to claim 1 based on the unsupervised sparse learning network of multilayer, its feature
It is：The cloud classification result of the neutral net output includes three kinds：Clear sky, Bao Yun and spissatus.
8. the cloud atlas cloud amount computational methods according to claim 1 based on the unsupervised sparse learning network of multilayer, its feature
It is：The result of the cloud classification of the neutral net output has three kinds, and setting output sample is set as corresponding three-dimensional vector.
9. the cloud atlas cloud amount computational methods according to claim 1 based on the unsupervised sparse learning network of multilayer, its feature
It is, step (3) specifically include：The cloud classification result exported by neutral net obtains the lower bound I of spissatus pixel intensitycld
With the upper bound I of clear sky pixel intensityclr, IcldIt is the minimum value of all spissatus brightness, IclrIt is the maximum of all clear sky brightness
It is worth, the cloud amount of each pixel is tried to achieve according to following formula on cloud atlas：
Ac=(I-Iclr)/(Icld-Iclr)
Wherein I is the brightness of single pixel receiver, finally gives the total amount of cloud of all pixels on cloud atlas.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201611107944.4A CN107529647B (en) | 2016-12-02 | 2016-12-02 | Cloud picture cloud amount calculation method based on multilayer unsupervised sparse learning network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201611107944.4A CN107529647B (en) | 2016-12-02 | 2016-12-02 | Cloud picture cloud amount calculation method based on multilayer unsupervised sparse learning network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107529647A true CN107529647A (en) | 2018-01-02 |
| CN107529647B CN107529647B (en) | 2021-02-09 |
Family
ID=60765443
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201611107944.4A Active CN107529647B (en) | 2016-12-02 | 2016-12-02 | Cloud picture cloud amount calculation method based on multilayer unsupervised sparse learning network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107529647B (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108108720A (en) * | 2018-01-08 | 2018-06-01 | 天津师范大学 | A kind of ground cloud image classification method based on depth multi-modal fusion |
| CN108846474A (en) * | 2018-05-18 | 2018-11-20 | 南京信息工程大学 | The satellite cloud picture cloud amount calculation method of convolutional neural networks is intensively connected based on multidimensional |
| CN110533063A (en) * | 2019-07-17 | 2019-12-03 | 赛德雷特(珠海)航天科技有限公司 | A kind of cloud amount calculation method and device based on satellite image and GMDH neural network |
| CN116861217A (en) * | 2023-07-18 | 2023-10-10 | 菏泽学院 | Identity recognition method and system for mobile terminal |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20030002731A1 (en) * | 2001-05-28 | 2003-01-02 | Heiko Wersing | Pattern recognition with hierarchical networks |
| CN103699902A (en) * | 2013-12-24 | 2014-04-02 | 南京信息工程大学 | Sorting method of ground-based visible light cloud picture |
| CN104504389A (en) * | 2014-12-18 | 2015-04-08 | 南京信息工程大学 | Satellite cloud amount computing method based on convolution neural network |
-
2016
- 2016-12-02 CN CN201611107944.4A patent/CN107529647B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20030002731A1 (en) * | 2001-05-28 | 2003-01-02 | Heiko Wersing | Pattern recognition with hierarchical networks |
| CN103699902A (en) * | 2013-12-24 | 2014-04-02 | 南京信息工程大学 | Sorting method of ground-based visible light cloud picture |
| CN104504389A (en) * | 2014-12-18 | 2015-04-08 | 南京信息工程大学 | Satellite cloud amount computing method based on convolution neural network |
Non-Patent Citations (3)
| Title |
|---|
| 刘念等: "基于核函数极限学习机的分布式光伏短期功率预测", 《农业工程学报》 * |
| 李林等: "基于极限学习机的地基可见光云图的云分类", 《大气科学学报》 * |
| 魏杰: "深度极限学习机的研究与应用", 《中国优秀硕士学位论文全文数据库信息科技辑》 * |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108108720A (en) * | 2018-01-08 | 2018-06-01 | 天津师范大学 | A kind of ground cloud image classification method based on depth multi-modal fusion |
| CN108846474A (en) * | 2018-05-18 | 2018-11-20 | 南京信息工程大学 | The satellite cloud picture cloud amount calculation method of convolutional neural networks is intensively connected based on multidimensional |
| CN110533063A (en) * | 2019-07-17 | 2019-12-03 | 赛德雷特(珠海)航天科技有限公司 | A kind of cloud amount calculation method and device based on satellite image and GMDH neural network |
| CN116861217A (en) * | 2023-07-18 | 2023-10-10 | 菏泽学院 | Identity recognition method and system for mobile terminal |
Also Published As
| Publication number | Publication date |
|---|---|
| CN107529647B (en) | 2021-02-09 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Tong et al. | Channel-attention-based DenseNet network for remote sensing image scene classification | |
| CN113011499B (en) | Hyperspectral remote sensing image classification method based on double-attention machine system | |
| Rao et al. | Siamese transformer network for hyperspectral image target detection | |
| Bai et al. | Edge-guided recurrent convolutional neural network for multitemporal remote sensing image building change detection | |
| Xu et al. | High-resolution remote sensing image change detection combined with pixel-level and object-level | |
| CN109508655A (en) | The SAR target identification method of incomplete training set based on twin network | |
| CN110119728A (en) | Remote sensing images cloud detection method of optic based on Multiscale Fusion semantic segmentation network | |
| CN109086700A (en) | Radar range profile's target identification method based on depth convolutional neural networks | |
| CN110866287B (en) | Point attack method for generating countercheck sample based on weight spectrum | |
| CN104182985B (en) | Remote sensing image change detection method | |
| CN108520202A (en) | Confrontation robustness image characteristic extracting method based on variation spherical projection | |
| CN108846474A (en) | The satellite cloud picture cloud amount calculation method of convolutional neural networks is intensively connected based on multidimensional | |
| CN107529647A (en) | A kind of cloud atlas cloud amount computational methods based on the unsupervised sparse learning network of multilayer | |
| CN103440505A (en) | Spatial neighborhood information weighted hyper-spectral remote sensing image classification method | |
| Chawan et al. | Automatic detection of flood using remote sensing images | |
| CN104732244A (en) | Wavelet transform, multi-strategy PSO (particle swarm optimization) and SVM (support vector machine) integrated based remote sensing image classification method | |
| CN106127198A (en) | A kind of image character recognition method based on Multi-classifers integrated | |
| CN105913081A (en) | Improved PCAnet-based SAR image classification method | |
| CN110084812A (en) | A kind of terahertz image defect inspection method, device, system and storage medium | |
| CN113408340B (en) | Dual-polarization SAR small ship detection method based on enhanced feature pyramid | |
| Chen et al. | Automatic variogram parameter extraction for textural classification of the panchromatic IKONOS imagery | |
| CN116168295B (en) | Lithology remote sensing intelligent interpretation model establishment method and interpretation method | |
| CN107563430A (en) | A kind of convolutional neural networks algorithm optimization method based on sparse autocoder and gray scale correlation fractal dimension | |
| CN110414513A (en) | Vision significance detection method based on semantically enhancement convolutional neural networks | |
| CN114612937A (en) | Single-mode enhancement-based infrared and visible light fusion pedestrian detection method |
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 |