CN108319693A - A kind of geomorphic feature clustering method based on three-dimensional Remote Sensing Database - Google Patents
A kind of geomorphic feature clustering method based on three-dimensional Remote Sensing Database Download PDFInfo
- Publication number
- CN108319693A CN108319693A CN201810103634.8A CN201810103634A CN108319693A CN 108319693 A CN108319693 A CN 108319693A CN 201810103634 A CN201810103634 A CN 201810103634A CN 108319693 A CN108319693 A CN 108319693A
- Authority
- CN
- China
- Prior art keywords
- layer
- remote sensing
- fourier
- shielding layer
- characterization
- 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.)
- Withdrawn
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
- G06F18/231—Hierarchical techniques, i.e. dividing or merging pattern sets so as to obtain a dendrogram
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biophysics (AREA)
- General Health & Medical Sciences (AREA)
- Bioinformatics & Computational Biology (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Evolutionary Biology (AREA)
- Computational Linguistics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Multimedia (AREA)
- Databases & Information Systems (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of geomorphic feature clustering method based on three-dimensional Remote Sensing Database, implementation step is：The three-dimensional remote sensing landforms of image sample to be tested, reduction, feature coarse extraction are called, training sample and test sample is chosen, solves characteristic distance, details extraction, utilizes training sample training neural network, clustering.The present invention trains continuous Fourier neural network using successively change mode, avoid the network number of plies it is more when there is the problem of diffusion, and reflection data intrinsic propesties can be extracted, data details feature is portrayed, protrude the high dimensional feature of difference between different types of ground objects.Since the present invention utilizes the deep layer high dimensional feature of continuous Fourier neural network extraction data, the problem that characteristic is less or feature learning is insufficient, unreasonable present in sorting technique is avoided, the clustering precision of UAV flight's remote sensor remote sensing landforms image is improved.
Description
Technical field
The invention belongs to remote sensing landforms identification technology fields, and in particular to a kind of landforms based on three-dimensional Remote Sensing Database are special
Levy clustering method.
Background technology
Remote sensing technology is from remote perception target reflection or electromagnetic wave, visible light, the infrared ray etc. that itself radiate, to mesh
The technology that is detected and identified of marking, broad range of data data can be obtained by having, and the speed for obtaining information is fast, the period is short and information
Measure the features such as big.According to the difference of acquisition modes, two class of optical remote sensing and microwave remote sensing can be classified as.Wherein, optical remote sensing
Corresponding equipment is simple, and acquired image space geometry high resolution, image is easy to interpret, but can only be used on daytime, by
Weather condition influences more serious；Since microwave has good penetration capacity, microwave remote sensing (SAR) can round-the-clock, round-the-clock right
Ground is imaged, but coherent imaging characteristic makes its image there is multiplying property coherent speckle noise.The two respectively has feature, is complementary to one another, in state
Increasingly important role is just being played in people's economy and national defense construction.
With the continuous development of sensor technology, space technology and computer technology, numerous machines/space remote sensing platform input
Operation, novel, high resolution sensor mounted can carry out earth surface uninterrupted, lasting observation, obtain a large amount of
Wide cut, high-definition remote sensing data.But face growing acquisition capability, therewith it is unbefitting be to remote sensing image processing with
The research and development of Interpretation Technology relatively lag behind, and cannot meet the active demand of practical application.Wherein, Multitemporal Remote Sensing Images variation detection
Refer to two width or more by being obtained to the same area, different times as one of processing and the basis of interpretation and key technology
Width remote sensing images are compared analysis, and then the change of interested atural object, scene or target is obtained according to difference between image
Change information, be limited by mankind's activity aggravation and natural calamity takes place frequently, research and development are increasingly paid attention to by people, are increasingly becoming distant
The research hotspot in sense field.
In recent years, domestic and foreign scholars propose many effective change detecting methods, and summing up to be divided into has supervision to become
Change detection and unsupervised variation detection.It is limited to have ground real change classification sample needed for supervision variation detection to be difficult to obtain, mesh
Preceding work focuses primarily upon unsupervised change detection class, is roughly divided into：(1) the multi-temporal remote sensing figure based on Cluster Distribution Divergence
As variation detection；(2) Multitemporal Remote Sensing Images based on differential image analysis change detection；And (3) are melted based on Markov
The Multitemporal Remote Sensing Images variation detection of conjunction.Especially the most universal to the research of the second class algorithm, core concept is will to change
Test problems are considered as binary classification/segmentation problem of image, and can be subdivided into clustering, intelligent optimization, Threshold segmentation, limited
Mixed model, markov random file, active profile and level set etc. are tactful.Wherein, clustering it is simple by it, effectively and by
To generally approving.
With the development of space technology and sensor technology, the acquisition modes of remotely-sensed data are more and more.Remotely-sensed data is logical
Regular data amount is huge, and information is interweaved, and how to be to make full use of these data effective and reasonable the organizing of these data
Key.
Traditional data organization is simply to be put in storage, and lookup and acquisition of information are carried out to database according to application demand,
The ability of one side search efficiency and acquisition of information is restricted, and on the other hand cannot intuitively check the information in database,
Significantly limit the utility value of data.Therefore, efficient data organization technique is the basis of remote sensing information process, is current
Need solve key technology.
Different from common data, remotely-sensed data has the characteristics that oneself.On the one hand, the mesh paid close attention in military remote sensing application
It is relatively fixed to mark type, such as certain air base, certain type naval vessel.On the other hand, target itself also has very strong Layer semantics
Characteristic, such as certain aircraft carrier fleet are subordinate to Mr. Yu naval force.So these data are suitble to be described with multistratum classification system.
In addition, the target that majority GIS and remote sensing application software can describe all is static, and in fact, many need
The target of expression and processing is not unalterable, including very strong time and space information.With the accumulation of data volume
With the development of situation, taxonomic hierarchies may also face modification at any time.Therefore must pay close attention to can to time-space process and when null object into
The new data organization model of row description.
It can be seen that by the above pertinent literature information：The existing geomorphic feature cluster point based on three-dimensional Remote Sensing Database
Analysis method is most of to be converted into plane remote sensing landforms by three-dimensional remote sensing landforms, is then extracted again to plane remote sensing landforms image procossing
The characteristic point of remote sensing landforms.First, perspective transformations are that plane computations amount is big, and process processing time is longer；Secondly, by perspective transformations
There can be certain error for the corresponding remote sensing landforms image of plane, influence the accuracy of identification of three-dimensional remote sensing landforms.Therefore, how
The problem of characteristic point of the three-dimensional remote sensing landforms of direct extraction fast, accurately and comprehensively is a urgent need research.
Invention content
In order to solve the problems in the prior art, the present invention proposes a kind of distant it is not necessary that three-dimensional remote sensing landforms are converted to plane
Feel landforms, complicated calculation amount and perspective transformations can be avoided at error caused by plane, and can fast and accurately complete pair
A kind of geomorphic feature clustering method based on three-dimensional Remote Sensing Database of three-dimensional remote sensing topographic feature extraction.
In order to achieve the above object, the technical solution adopted in the present invention is：Include the following steps：
A kind of geomorphic feature clustering method based on three-dimensional Remote Sensing Database, it is characterised in that：Include the following steps：
1) plane characteristic for calling image sample to be tested, calls UAV flight's remote sensor remote sensing to be clustered
The cross matrix of landforms image sample, wherein cross matrix is the matrix that size is 5 × 5 × M, and N is that UAV flight's remote sensing passes
The sum of sensor remote sensing landforms image pixel, pretreatment, use window size for 9 × 9 electromagnetic interface filter to cross matrix into
Row filtering, is obtained filtered cross matrix, restores the plane remote sensing landforms of sample to be tested using the cross matrix and stood
Body remote sensing landforms carry out feature coarse extraction according to the three-dimensional remote sensing landforms restored to three-dimensional remote sensing landforms, obtain a series of vertical
The height value on body remote sensing landforms surface extracts the maximum height value on three-dimensional remote sensing landforms surface as sample to be tested solid remote sensing
The characteristic point of looks feature, using the characteristic point as the feature of UAV flight's remote sensor remote sensing landforms image, composition
The sample set of one size of N × 11 randomly selects 8% sample as UAV flight's remote sensor remote sensing from sample set
Landforms image training sample, using the sample of residue 92% as UAV flight's remote sensor remote sensing landforms image test specimens
This；
2) it by the characteristic value of UAV flight's remote sensor remote sensing landforms image training sample feature obtained above, asks
Go out the characteristic distance with contrast mould's outer contour shape feature, then by sample to be tested and contrast mould's each sample outer contour shape
After the characteristic distance of feature sorts from big to small, statistical nature distance is less than the number of predetermined threshold value, is recorded with variable x, x
Initial value be equal to 0, characteristic distance is more primary with predetermined threshold value, be less than threshold value if, k just from add 1, final variables x etc.
In 1, then illustrate that UAV flight's remote sensor remote sensing landforms image training sample obtained above can be uniquely identified, it is complete
At the extraction of three-dimensional remote sensing geomorphic feature；
3) position feature for extracting three-dimensional remote sensing landforms isobaric terminal and crosspoint, is completed thin to three-dimensional remote sensing landforms
The extraction for saving feature directly extracts three-dimensional remote sensing relief detail characteristic point to realize；Continuous Fourier's nerve net is generated at random
The initial weight and Fourier's activation primitive scaling variable and offset variable of network surface layer network and time layer network；By training sample tune
In the Fourier neural network for using surface layer, the initial weight T of surface layer network shielding layer and calling node layer is utilized1', characterize layer
With the initial weight T of masking node layer1", the scaling variable m of Fourier's activation primitive1With offset variable n1Calculate separately cover web
The characterization ψ of network shielding layer1With the characterization value h of characterization layer1；Utilize the table of training sample in absolute error formula computational chart layer network
Levy error E1；Using least square method, obtain the optimal weights of surface layer network, the optimal scaling variable of Fourier's activation primitive and
Optimum displacement variable and optimal shielding layer characterization；It regard the shielding layer characterization of surface layer Fourier neural network as sublevel Fourier
The calling of neural network, and utilize the initial weight T ' of time layer network shielding layer and characterization node layer2, characterize layer and shielding layer section
The initial weight T " of point2, Fourier's activation primitive scaling variable m2With offset variable n2Calculate time characterization ψ of layer network shielding layer2
With the characterization h of characterization layer2；The characterization error E of training sample in time layer network is calculated using absolute error formula2；Using minimum two
Multiplication obtains time optimal weights of layer network, the optimal scaling variable of Fourier's activation primitive and optimum displacement variable and most
Excellent shielding layer characterization；
4) training sample and test sample are called respectively in trained continuous Fourier neural network, is trained
Sample characteristics collection and test sample feature set call training sample feature set and test sample feature set to linSVM tools
Case obtains the final cluster result of UAV flight's remote sensor remote sensing landforms image, calculates clustering precision, and statistics is to be clustered
UAV flight's remote sensor remote sensing landforms image in pixel number identical with class label in cluster result, calculate
Class label same pixel point number accounts for the hundred of UAV flight's remote sensor remote sensing landforms image total pixel number to be clustered
Divide ratio, obtains clustering precision.
Beneficial effects of the present invention are：The deep layer high dimensional feature that data are extracted using continuous Fourier neural network, is avoided
The problem that characteristic is less or feature learning is insufficient, unreasonable present in sorting technique, improves UAV flight
The clustering precision of remote sensor remote sensing landforms image.
Description of the drawings
In order to more clearly explain the embodiment of the invention or the technical proposal in the existing technology, to embodiment or will show below
There is attached drawing needed in technology description to be briefly described, it should be apparent that, the accompanying drawings in the following description is only this
Some embodiments of invention without having to pay creative labor, may be used also for those of ordinary skill in the art
With obtain other attached drawings according to these attached drawings.
Fig. 1 is the flow chart of the present invention.
Specific implementation mode
With reference to embodiment, the present invention will be further described.
The technical solution adopted in the present invention is：Include the following steps：
1) plane characteristic for calling image sample to be tested, calls UAV flight's remote sensor remote sensing to be clustered
The cross matrix of landforms image sample, wherein cross matrix is the matrix that size is 5 × 5 × M, and N is that UAV flight's remote sensing passes
The sum of sensor remote sensing landforms image pixel, pretreatment, use window size for 9 × 9 electromagnetic interface filter to cross matrix into
Row filtering, is obtained filtered cross matrix, restores the plane remote sensing landforms of sample to be tested using the cross matrix and stood
Body remote sensing landforms carry out feature coarse extraction according to the three-dimensional remote sensing landforms restored to three-dimensional remote sensing landforms, obtain a series of vertical
The height value on body remote sensing landforms surface extracts the maximum height value on three-dimensional remote sensing landforms surface as sample to be tested solid remote sensing
The characteristic point of looks feature, using the characteristic point as the feature of UAV flight's remote sensor remote sensing landforms image, composition
The sample set of one size of N × 11 randomly selects 8% sample as UAV flight's remote sensor remote sensing from sample set
Landforms image training sample, using the sample of residue 92% as UAV flight's remote sensor remote sensing landforms image test specimens
This；
2) it by the characteristic value of UAV flight's remote sensor remote sensing landforms image training sample feature obtained above, asks
Go out the characteristic distance with contrast mould's outer contour shape feature, then by sample to be tested and contrast mould's each sample outer contour shape
After the characteristic distance of feature sorts from big to small, statistical nature distance is less than the number of predetermined threshold value, is recorded with variable x, x
Initial value be equal to 0, characteristic distance is more primary with predetermined threshold value, be less than threshold value if, k just from add 1, final variables x etc.
In 1, then illustrate that UAV flight's remote sensor remote sensing landforms image training sample obtained above can be uniquely identified, it is complete
At the extraction of three-dimensional remote sensing geomorphic feature；
3) position feature for extracting three-dimensional remote sensing landforms isobaric terminal and crosspoint, is completed thin to three-dimensional remote sensing landforms
The extraction for saving feature directly extracts three-dimensional remote sensing relief detail characteristic point to realize；Continuous Fourier's nerve net is generated at random
The initial weight and Fourier's activation primitive scaling variable and offset variable of network surface layer network and time layer network；By training sample tune
In the Fourier neural network for using surface layer, the initial weight T of surface layer network shielding layer and calling node layer is utilized1', characterize layer
With the initial weight T of masking node layer1", the scaling variable m of Fourier's activation primitive1With offset variable n1Calculate separately cover web
The characterization ψ of network shielding layer1With the characterization value h of characterization layer1；Utilize the table of training sample in absolute error formula computational chart layer network
Levy error E1；Using least square method, obtain the optimal weights of surface layer network, the optimal scaling variable of Fourier's activation primitive and
Optimum displacement variable and optimal shielding layer characterization；It regard the shielding layer characterization of surface layer Fourier neural network as sublevel Fourier
The calling of neural network, and utilize the initial weight T ' of time layer network shielding layer and characterization node layer2, characterize layer and shielding layer section
The initial weight T " of point2, Fourier's activation primitive scaling variable m2With offset variable n2Calculate time characterization ψ of layer network shielding layer2
With the characterization h of characterization layer2；The characterization error E of training sample in time layer network is calculated using absolute error formula2；Using minimum two
Multiplication obtains time optimal weights of layer network, the optimal scaling variable of Fourier's activation primitive and optimum displacement variable and most
Excellent shielding layer characterization；
4) training sample and test sample are called respectively in trained continuous Fourier neural network, is trained
Sample characteristics collection and test sample feature set call training sample feature set and test sample feature set to linSVM tools
Case obtains the final cluster result of UAV flight's remote sensor remote sensing landforms image, calculates clustering precision, and statistics is to be clustered
UAV flight's remote sensor remote sensing landforms image in pixel number identical with class label in cluster result, calculate
Class label same pixel point number accounts for the hundred of UAV flight's remote sensor remote sensing landforms image total pixel number to be clustered
Divide ratio, obtains clustering precision.
In the step 1), continuous phase solution bag method is according to formulaIt is stood
The elevation information h (x, y), wherein l on body remote sensing landforms surface0It is distance of the image center on unmanned plane to ground surface, d is phase
Machine is to the distance of projecting apparatus, f0It is the wavelength of projection grating on ground surface.
The three-dimensional remote sensing geomorphic feature is one group of vector V that a length is T*S:
Calculating shielding layer characterization formula described in the step 3) is as follows：
Wherein,The characterization for indicating masking node layer j is a total expression of shielding layer node characterization, table herein
Layer network shielding layer node table takes over ψ for use1It indicates, secondary layer network shielding layer node table takes over ψ for use2It indicates, m is to call number of nodes, T 'jk
It indicates masking node layer j and calls the weight between node k, T ' is one of shielding layer and calling layer node weights total herein
It indicates, surface layer network shielding layer and calling layer node weights T1' indicate, secondary layer network shielding layer and calling layer node weights are used
T′2It indicates, xkIndicate the calling of calling node k, njIndicate the offset variable of Fourier's activation primitive of masking node layer j, herein
N is a total expression of Fourier's offset variable, Fourier's offset variable n of surface layer network1It indicates, Fu of secondary layer network
In leaf offset variable n2It indicates, mjIndicate the scaling variable of Fourier's activation primitive of masking node layer j, m is Fourier herein
One total expression of activation primitive scaling variable, Fourier's activation primitive scaling variable m of surface layer network1It indicates, sublevel
Fourier's activation primitive scaling variable m of network2It indicates；
It is as follows that computational representation layer characterizes formula：
Wherein, h (i) indicates the characterization of characterization node i, and h is a total expression for characterizing layer characterization, surface layer network herein
Characterization layer characterization h1It indicates, the characterization layer characterization h of secondary layer network2It indicates, p is shielding layer number of nodes, T "ijIndicate characterization
Weight between node i and masking node layer j, T " is a total table for characterizing node layer and shielding layer node weights herein
Show, the characterization node layer and shielding layer node weights T of surface layer network1It " indicates, the characterization node layer and shielding layer of secondary layer network
Node weights T "2It indicates, ψ (j) indicates the characterization of masking node layer j, and ψ is a total table of shielding layer node characterization herein
Show, surface layer network shielding layer node table takes over ψ for use1It indicates, secondary layer network shielding layer node table takes over ψ for use2It indicates.To sample to be tested
The pretreatment of plane remote sensing landforms image includes segmentation, filtering, noise reduction and the refinement of plane remote sensing landforms image.
Least square method described in the step 3 is as follows：
The first step calculates the weight between the shielding layer of continuous Fourier neural network and characterization layer according to following formula：
Wherein, T "t+1Indicate that the weight between shielding layer and characterization layer when the t+1 times recurrence, t indicate continuous Fourier god
Recursive number, T are trained through network weightt" indicate that the weight between shielding layer and characterization layer, μ indicate masking when the t times recurrence
The learning rate of weight between layer and characterization layer, usual value range are 0<μ<1,Indicate sample when the t times recurrence
Absolute error partial derivative of weight between shielding layer and characterization layer operates, and α is information coefficient, and usual value range is 0.7<α<
1, Δ TtIt " indicates shielding layer when the t times recurrence and characterizes the weight school deviator between layer.
Second step calculates the weight between the calling layer and shielding layer of continuous Fourier neural network according to following formula：
Wherein, T 't+1Indicate that the weight between calling layer and shielding layer when the t+1 times recurrence, t indicate continuous Fourier god
Recursive number, T are trained through network weightt' indicate that the weight between calling layer and shielding layer when the t times recurrence, μ indicate to call
The learning rate of weight between layer and shielding layer, usual value range are 0<μ<1,Indicate the exhausted of sample when the t times recurrence
Partial derivative operation to error weight between calling layer and shielding layer, α is information coefficient, and usual value range is 0.7<α<1,
ΔTtThe weight school deviator between calling layer and shielding layer when the t times recurrence of ' expression.
Third walks, and according to following formula, calculates the scaling variable of continuous Fourier neural network shielding layer Fourier's activation primitive：
Wherein, mt+1Indicate that the scaling variable of shielding layer Fourier activation primitive when the t+1 times recurrence, t indicate in continuous Fu
The recursive number of leaf neural network weight training, mtIndicate the scaling variable of shielding layer Fourier activation primitive when the t times recurrence,
μ indicates that the learning rate of shielding layer Fourier's activation primitive scaling variable, usual value range are 0<μ<1,It indicates the t times
The absolute error of sample operates the partial derivative of shielding layer Fourier's activation primitive scaling variable when recurrence, and α is information coefficient, leads to
Normal value range is 0.7<α<1, Δ mtIndicate the school deviator of shielding layer Fourier activation primitive scaling variable when the t times recurrence.
4th step calculates the offset variable of continuous Fourier neural network shielding layer Fourier's activation primitive according to following formula：
Wherein, nt+1Indicate that the offset variable of shielding layer Fourier activation primitive when the t+1 times recurrence, t indicate in continuous Fu
The recursive number of leaf neural network weight training, ntIndicate the offset variable of shielding layer Fourier activation primitive when the t times recurrence,
μ indicates that the learning rate of shielding layer Fourier's activation primitive offset variable, usual value range are 0<μ<1,It indicates the t times
The absolute error of sample operates the partial derivative of shielding layer Fourier's activation primitive offset variable when recurrence, and α is information coefficient, leads to
Normal value range is 0.7<α<1, Δ ntIndicate the school deviator of shielding layer Fourier activation primitive offset variable when the t times recurrence.
5th step judges whether to reach maximum recurrence number, if it is not, returning to the first step, if so, stopping recurrence, obtains net
The optimal weights of network, the optimal scaling variable and optimum displacement variable of Fourier's activation primitive and optimal shielding layer characterization.
The present invention finds out the maximum height value on remote sensing landforms surface by the elevation information of three-dimensional remote sensing landforms；With a system
The outer contour shape for the three-dimensional remote sensing landforms of equivalent cross section reflection that the three-dimensional remote sensing landforms of row parallel plane cutting obtain；According to change
Amount x decides whether carefully to extract the three-dimensional further feature of remote sensing landforms；To three-dimensional remote sensing relief detail feature, i.e. isobar
Terminal and crosspoint, according to the plane remote sensing landforms of same person and each isobar terminal of three-dimensional remote sensing landforms and crosspoint
Transverse and longitudinal coordinate value is equal, first extracts plane remote sensing landforms isobar terminal and intersection feature, then correspond to three-dimensional remote sensing landforms
In, to complete the extraction of three-dimensional remote sensing relief detail characteristic point.The present invention directly carries out feature extraction to three-dimensional remote sensing landforms,
Avoid conventional method by three-dimensional remote sensing landforms be converted to plane remote sensing landforms extract again it is computationally intensive caused by feature and by
The shortcomings of error of perspective transformations plane；In addition the present invention determines the base in three-dimensional remote sensing geomorphic feature coarse extraction according to variable x
On plinth, if needs further carefully extract three-dimensional remote sensing geomorphic feature, to the three-dimensional remote sensing of more comprehensive and accurate extraction
The feature of looks.
In brief, the polarization remote sensing landforms image clustering method of the invention based on continuous Fourier neural network, it is main
Solve the problems, such as in the prior art due to characteristic is less or feature extraction is unreasonable and caused by clustering precision decline.In fact
Now step is：Call image；Pretreatment；Choose sample；Continuous Fourier neural network is trained using training sample；Extraction is special
Sign；Cluster；Calculate clustering precision.The present invention trains continuous Fourier neural network using successively change mode, using this network
Model and training method, can be very good to avoid the problem that occurring when the network number of plies is more diffusion, and can be from original
UAV flight's remote sensor remote sensing landforms image extracting data goes out to reflect data intrinsic propesties, depicts data details spy
Sign protrudes poor another characteristic between different types of ground objects.Since the present invention extracts data using continuous Fourier neural network
Deep layer high dimensional feature, successfully avoids that characteristic present in existing clustering technique is less or feature learning is insufficient, does not conform to
The problem of reason, improves the clustering precision of UAV flight's remote sensor remote sensing landforms image.
Claims (5)
1. a kind of geomorphic feature clustering method based on three-dimensional Remote Sensing Database, it is characterised in that：Include the following steps：
1) plane characteristic for calling image sample to be tested, calls UAV flight's remote sensor remote sensing landforms to be clustered
The cross matrix of image sample, wherein cross matrix is the matrix that size is 5 × 5 × M, and N is UAV flight's remote sensor
The sum of remote sensing landforms image pixel, pretreatment, uses window size to be filtered to cross matrix for 9 × 9 electromagnetic interface filter
Wave obtains filtered cross matrix, by the plane remote sensing landforms of sample to be tested restores to obtain using the cross matrix three-dimensional distant
Feel landforms, according to the three-dimensional remote sensing landforms restored, feature coarse extraction is carried out to three-dimensional remote sensing landforms, it is distant to obtain a series of solids
Feel the height value on landforms surface, the maximum height value for extracting three-dimensional remote sensing landforms surface is special as sample to be tested solid remote sensing landforms
The characteristic point of sign forms a N using the characteristic point as the feature of UAV flight's remote sensor remote sensing landforms image
The sample set of × 11 sizes randomly selects 8% sample as UAV flight's remote sensor remote sensing landforms from sample set
Image training sample, using the sample of residue 92% as UAV flight's remote sensor remote sensing landforms image test sample；
2) by the characteristic value of UAV flight's remote sensor remote sensing landforms image training sample feature obtained above, find out with
The characteristic distance of contrast mould's outer contour shape feature, then by sample to be tested and contrast mould's each sample outer contour shape feature
Characteristic distance sort from big to small after, statistical nature distance be less than predetermined threshold value number, recorded with variable x, x just
Initial value is equal to 0, and characteristic distance is more primary with predetermined threshold value, and if being less than threshold value, for k just from adding 1, final variables x is equal to 1,
Then illustrate that UAV flight's remote sensor remote sensing landforms image training sample obtained above can be uniquely identified, completes vertical
The extraction of body remote sensing geomorphic feature；
3) position feature for extracting three-dimensional remote sensing landforms isobaric terminal and crosspoint is completed to three-dimensional remote sensing relief detail spy
The extraction of sign directly extracts three-dimensional remote sensing relief detail characteristic point to realize；Continuous Fourier neural network table is generated at random
The initial weight and Fourier's activation primitive scaling variable and offset variable of layer network and time layer network；Training sample calling is arrived
In the Fourier neural network on surface layer, using the initial weight T1 ' of surface layer network shielding layer and calling node layer, layer and screening are characterized
It covers the initial weight T1 " of node layer, the scaling variable m1 of Fourier's activation primitive and offset variable n1 and calculates separately surface layer network
The characterization ψ 1 of the shielding layer and characterization value h1 for characterizing layer；Utilize the characterization of training sample in absolute error formula computational chart layer network
Error E 1；Using least square method, the optimal weights of surface layer network, the optimal scaling variable of Fourier's activation primitive and most are obtained
Excellent offset variable and optimal shielding layer characterization；By the shielding layer characterization of surface layer Fourier neural network as sublevel Fourier god
Calling through network, and utilize the initial weight T ' 2 of time layer network shielding layer and characterization node layer, characterization layer and masking node layer
Initial weight T " 2, Fourier's activation primitive scaling variable m2 and offset variable n2 calculate time characterization ψ 2 of layer network shielding layer
With the characterization h2 of characterization layer；The characterization error E 2 of training sample in time layer network is calculated using absolute error formula；Using minimum
Square law, obtain time optimal weights of layer network, the optimal scaling variable of Fourier's activation primitive and optimum displacement variable and
Optimal shielding layer characterization；
4) training sample and test sample are called respectively in trained continuous Fourier neural network, obtains training sample
Training sample feature set and test sample feature set are called to the tool boxes linSVM, are obtained by feature set and test sample feature set
To the final cluster result of UAV flight's remote sensor remote sensing landforms image, clustering precision is calculated, nothing to be clustered is counted
Pixel number identical with class label in cluster result in man-machine carrying remote sensor remote sensing landforms image calculates classification
Label same pixel point number accounts for the percentage of UAV flight's remote sensor remote sensing landforms image total pixel number to be clustered,
Obtain clustering precision.
2. a kind of geomorphic feature clustering method based on three-dimensional Remote Sensing Database according to claim 1, feature
It is：In the step 1), continuous phase solution bag method is according to formulaObtain solid
The elevation information h (x, y) on remote sensing landforms surface, wherein l0 are distance of the image center on unmanned plane to ground surface, and d is camera
To the distance of projecting apparatus, f0 is the wavelength of projection grating on ground surface.
3. solid remote sensing topographic feature extraction method according to claim 1, it is characterised in that：The three-dimensional remote sensing
Looks are characterized in that a length is one group of vector V of T*S:
4. a kind of geomorphic feature clustering method based on three-dimensional Remote Sensing Database according to claim 1, feature
It is：Calculating shielding layer characterization formula described in the step 3) is as follows：
Wherein,The characterization for indicating masking node layer j is a total expression of shielding layer node characterization, surface layer network herein
Shielding layer node table is taken over ψ 1 for use and is indicated, secondary layer network shielding layer node table requisition ψ 2 indicates that m is to call number of nodes, and T ' jk are indicated
It covers node layer j and calls the weight between node k, T ' is shielding layer and a total table of calling layer node weights herein
Show, surface layer network shielding layer and calling layer node weights are indicated with T1 ', secondary layer network shielding layer and calling layer node weights T '
2 indicate, xk indicates that the calling of node k, nj is called to indicate the offset variable of Fourier's activation primitive of masking node layer j, herein n
It is a total expression of Fourier's offset variable, Fourier's offset variable of surface layer network is indicated with n1, Fu of secondary layer network
In leaf offset variable indicate that mj indicates the scaling variable of Fourier's activation primitive of masking node layer j, and m is in Fu herein with n2
Fourier's activation primitive scaling variable of one total expression of leaf activation primitive scaling variable, surface layer network is indicated with m1, secondary
Fourier's activation primitive scaling variable of layer network is indicated with m2；
It is as follows that computational representation layer characterizes formula：
Wherein, h (i) indicates the characterization of characterization node i, and h is a total expression for characterizing layer characterization, the table of surface layer network herein
Sign layer characterization indicates that the characterization layer characterization of secondary layer network is indicated with h2 with h1, and p is that shielding layer number of nodes, T " ij indicate characterization section
Weight between point i and masking node layer j, T " is a total expression for characterizing node layer and shielding layer node weights herein,
The characterization node layer of surface layer network and the T1 " expressions of shielding layer node weights, the characterization node layer and shielding layer section of secondary layer network
Point weight T " 2 indicates that ψ (j) indicates the characterization of masking node layer j, and ψ is a total expression of shielding layer node characterization herein,
Surface layer network shielding layer node table requisition ψ 1 indicates that secondary layer network shielding layer node table requisition ψ 2 is indicated.Sample to be tested is put down
The pretreatment of face remote sensing landforms image includes cutting, filtering, noise reduction and the refinement of plane remote sensing landforms image.
5. a kind of geomorphic feature clustering method based on three-dimensional Remote Sensing Database according to claim 1, feature
It is：Least square method described in the step 3 is as follows：
The first step calculates the weight between the shielding layer of continuous Fourier neural network and characterization layer according to following formula：
Wherein, T " t+1 indicate that the weight between shielding layer and characterization layer, t indicate continuous Fourier's nerve net when the t+1 times recurrence
The recursive number of network weight training, Tt " indicate the t time recurrence when shielding layer and characterization layer between weight, μ indicate shielding layer with
The learning rate of weight between characterization layer, usual value range are 0<μ<1,Indicate the absolute mistake of sample when the t times recurrence
Difference partial derivative of weight between shielding layer and characterization layer operates, and α is information coefficient, and usual value range is 0.7<α<1, Δ
The weight school deviator between shielding layer and characterization layer when Tt " the t times recurrence of expression；
Second step calculates the weight between the calling layer and shielding layer of continuous Fourier neural network according to following formula：
Wherein, T ' t+1 indicate that the weight between calling layer and shielding layer when the t+1 times recurrence, t indicate continuous Fourier's nerve net
The recursive number of network weight training, weight when Tt ' the t times recurrence of expression between calling layer and shielding layer, μ indicate calling layer and
The learning rate of weight between shielding layer, usual value range are 0<μ<1,Indicate the absolute mistake of sample when the t times recurrence
The partial derivative of difference weight between calling layer and shielding layer operates, and α is information coefficient, and usual value range is 0.7<α<1, Δ
The weight school deviator between calling layer and shielding layer when Tt ' the t times recurrence of expression；
Third walks, and according to following formula, calculates the scaling variable of continuous Fourier neural network shielding layer Fourier's activation primitive：
Wherein, mt+1 indicates that the scaling variable of shielding layer Fourier's activation primitive when the t+1 times recurrence, t indicate continuous Fourier
The recursive number of neural network weight training, mt indicate the scaling variable of shielding layer Fourier's activation primitive, μ when the t times recurrence
Indicate that the learning rate of shielding layer Fourier's activation primitive scaling variable, usual value range are 0<μ<1,It indicates the t times
The absolute error of sample operates the partial derivative of shielding layer Fourier's activation primitive scaling variable when recurrence, and α is information coefficient, leads to
Normal value range is 0.7<α<1, Δ mt indicate the school deviator of shielding layer Fourier's activation primitive scaling variable when the t times recurrence；
4th step calculates the offset variable of continuous Fourier neural network shielding layer Fourier's activation primitive according to following formula：
Wherein, nt+1 indicates that the offset variable of shielding layer Fourier's activation primitive when the t+1 times recurrence, t indicate continuous Fourier
The recursive number of neural network weight training, nt indicate the offset variable of shielding layer Fourier's activation primitive, μ when the t times recurrence
Indicate that the learning rate of shielding layer Fourier's activation primitive offset variable, usual value range are 0<μ<1,It indicates the t times
The absolute error of sample operates the partial derivative of shielding layer Fourier's activation primitive offset variable when recurrence, and α is information coefficient, leads to
Normal value range is 0.7<α<1, Δ nt indicate the school deviator of shielding layer Fourier's activation primitive offset variable when the t times recurrence；
5th step judges whether to reach maximum recurrence number, if it is not, returning to the first step, if so, stopping recurrence, obtains network
Optimal weights, the optimal scaling variable and optimum displacement variable of Fourier's activation primitive and optimal shielding layer characterization.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810103634.8A CN108319693A (en) | 2018-02-01 | 2018-02-01 | A kind of geomorphic feature clustering method based on three-dimensional Remote Sensing Database |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810103634.8A CN108319693A (en) | 2018-02-01 | 2018-02-01 | A kind of geomorphic feature clustering method based on three-dimensional Remote Sensing Database |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108319693A true CN108319693A (en) | 2018-07-24 |
Family
ID=62890525
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810103634.8A Withdrawn CN108319693A (en) | 2018-02-01 | 2018-02-01 | A kind of geomorphic feature clustering method based on three-dimensional Remote Sensing Database |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108319693A (en) |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109770900A (en) * | 2019-01-08 | 2019-05-21 | 中国科学院自动化研究所 | Brain-computer interface based on convolutional neural networks instructs delivery method, system, device |
| CN110222697A (en) * | 2019-04-26 | 2019-09-10 | 南京航空航天大学 | A kind of planetary surface landforms actively perceive method based on intensified learning |
| CN110322326A (en) * | 2019-07-09 | 2019-10-11 | 程新宇 | A kind of economic big data sharing method of geography based on ArgGis |
| CN109253708B (en) * | 2018-09-29 | 2020-09-11 | 南京理工大学 | Stripe projection time phase unwrapping method based on deep learning |
| CN112733888A (en) * | 2020-12-25 | 2021-04-30 | 河南大学 | Deep space exploration image morphology analysis method based on multilayer MRF clustering |
| CN114626458A (en) * | 2022-03-15 | 2022-06-14 | 中科三清科技有限公司 | High-voltage rear part identification method and device, storage medium and terminal |
| CN115937007A (en) * | 2022-03-04 | 2023-04-07 | 中科三清科技有限公司 | Wind shear identification method and device, electronic equipment and medium |
| CN116030037A (en) * | 2023-02-23 | 2023-04-28 | 西南石油大学 | Ultrasonic image simulation method based on self-supervision learning |
-
2018
- 2018-02-01 CN CN201810103634.8A patent/CN108319693A/en not_active Withdrawn
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109253708B (en) * | 2018-09-29 | 2020-09-11 | 南京理工大学 | Stripe projection time phase unwrapping method based on deep learning |
| US11906286B2 (en) | 2018-09-29 | 2024-02-20 | Nanjing University Of Science And Technology | Deep learning-based temporal phase unwrapping method for fringe projection profilometry |
| CN109770900A (en) * | 2019-01-08 | 2019-05-21 | 中国科学院自动化研究所 | Brain-computer interface based on convolutional neural networks instructs delivery method, system, device |
| CN110222697A (en) * | 2019-04-26 | 2019-09-10 | 南京航空航天大学 | A kind of planetary surface landforms actively perceive method based on intensified learning |
| CN110322326A (en) * | 2019-07-09 | 2019-10-11 | 程新宇 | A kind of economic big data sharing method of geography based on ArgGis |
| CN112733888A (en) * | 2020-12-25 | 2021-04-30 | 河南大学 | Deep space exploration image morphology analysis method based on multilayer MRF clustering |
| CN115937007A (en) * | 2022-03-04 | 2023-04-07 | 中科三清科技有限公司 | Wind shear identification method and device, electronic equipment and medium |
| CN114626458A (en) * | 2022-03-15 | 2022-06-14 | 中科三清科技有限公司 | High-voltage rear part identification method and device, storage medium and terminal |
| CN116030037A (en) * | 2023-02-23 | 2023-04-28 | 西南石油大学 | Ultrasonic image simulation method based on self-supervision learning |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108319693A (en) | A kind of geomorphic feature clustering method based on three-dimensional Remote Sensing Database | |
| CN110188705B (en) | Remote traffic sign detection and identification method suitable for vehicle-mounted system | |
| CN110929607B (en) | Remote sensing identification method and system for urban building construction progress | |
| Yin et al. | Hot region selection based on selective search and modified fuzzy C-means in remote sensing images | |
| CN103413151B (en) | Hyperspectral image classification method based on figure canonical low-rank representation Dimensionality Reduction | |
| CN103971364B (en) | Remote sensing image variation detecting method on basis of weighted Gabor wavelet characteristics and two-stage clusters | |
| CN112101278A (en) | Hotel point cloud classification method based on k nearest neighbor feature extraction and deep learning | |
| CN107016357A (en) | A kind of video pedestrian detection method based on time-domain convolutional neural networks | |
| CN109948593A (en) | Based on the MCNN people counting method for combining global density feature | |
| CN103606154B (en) | Based on JSEG and the multiple dimensioned offshore spilled oil SAR image segmentation method of spectral clustering | |
| CN101877007A (en) | Remote sensing image retrieval method with integration of spatial direction relation semanteme | |
| CN104077605A (en) | Pedestrian search and recognition method based on color topological structure | |
| CN102819747B (en) | Method for automatically classifying forestry service images | |
| CN104715251B (en) | A kind of well-marked target detection method based on histogram linear fit | |
| CN105224937A (en) | Based on the semantic color pedestrian of the fine granularity heavily recognition methods of human part position constraint | |
| CN106156758B (en) | A kind of tidal saltmarsh method in SAR seashore image | |
| CN106295498A (en) | Remote sensing image target area detection apparatus and method | |
| CN109344812A (en) | A kind of improved single photon point cloud data denoising method based on cluster | |
| CN112101189B (en) | SAR image target detection method and test platform based on attention mechanism | |
| CN105069459B (en) | One kind is directed to High Resolution SAR Images type of ground objects extracting method | |
| CN105405138A (en) | Water surface target tracking method based on saliency detection | |
| CN108805057A (en) | A kind of SAR image oil depot area detection method based on joint significance analysis | |
| CN112668461A (en) | Intelligent supervision system with wild animal identification function | |
| Sutha | Object based classification of high resolution remote sensing image using HRSVM-CNN classifier | |
| Al-Ghrairi et al. | Classification of satellite images based on color features using remote sensing |
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 | ||
| WW01 | Invention patent application withdrawn after publication |
Application publication date: 20180724 |
|
| WW01 | Invention patent application withdrawn after publication |