CN105787501B - Power transmission line corridor region automatically selects the vegetation classification method of feature - Google Patents
Power transmission line corridor region automatically selects the vegetation classification method of feature Download PDFInfo
- Publication number
- CN105787501B CN105787501B CN201510953921.4A CN201510953921A CN105787501B CN 105787501 B CN105787501 B CN 105787501B CN 201510953921 A CN201510953921 A CN 201510953921A CN 105787501 B CN105787501 B CN 105787501B
- Authority
- CN
- China
- Prior art keywords
- feature
- remote sensing
- cross validation
- sample
- sensing image
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Expired - Fee Related
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of power transmission line corridor regional vegetation classification methods, comprising: step 1, extracts the feature of training sample and cross validation sample, the various features constitutive characteristic collection of all training samples；Step 2, it is based on training sample and cross validation sample, characteristic optimization selection is carried out using cross-validation method, to obtain preferred feature；Step 3, vegetation classification is carried out to remote sensing image test data using preferred feature.The method of the present invention is in the feature selecting stage without largely being iterated to calculate, gained assemblage characteristic has robustness after optimized selection, assemblage characteristic is used for the vegetation classification in remote sensing image power transmission line corridor region, is remarkably improved algorithm computational efficiency and nicety of grading.
Description
Technical field
The invention belongs to remote sensing image intelligent analysis technical field, in particular to a kind of power transmission line corridor region is automatic
Select the vegetation classification method of feature.
Background technique
With national economy rapid growth, power grid construction is grown rapidly, and the safe and stable operation of transmission line of electricity is to ensure people
The essential condition lived.China's transmission line of electricity power transmission distance, on the way landform and complex geologic conditions wide across region.It plants
It is the main body of terrestrial ecosystems, as green space system chief component, playing in the entire ecosystem can not
The effect of substitution.For electric system medium-high voltage transmission lines road, design of the trend and distribution of tree cover to transmission line of electricity
There is significant impact with operation.For example, electric power line pole tower height, shaft tower addressing and the design of transmission line of electricity are raw by vegetation
The influence of long situation.Especially in recent years, power transmission line corridor nearby brushfire, the initiations such as burn the grass on waste land line tripping accident frequency
Hair, causes huge economic loss.In addition to human factor, fire caused by forest vegetation spontaneous combustion is the main original for causing tripping
Cause.With the continuous development of power grid construction, high pressure, UHV transmission line will cover the area of more complex environments, either
The arable land that has developed and utilized, coniferous forest, or the nature reserve area without exploitation, to the prison of vegetation type and growing state
The critical issue that transmission line construction must be faced with safe operation will all be become by surveying.China's transmission line of electricity wide coverage,
Vegetation type is various on the way, and inflammable vegetation pattern and vegetation area can effectively be judged by carrying out vegetation classification by remote sensing,
So that it is determined that brushfire Yi Faqu, to prevention brushfire, prevents line tripping, guarantee transmission line safety operation meaning weight
Greatly.
Resource three (ZY-3) satellites are first autonomous civilian high-resolution stereo mapping satellites of China, pass through solid
Observation, can survey 1: 5 ten thousand topographic maps of system, be not necessarily to field measurement, the accurate acquisition of data can be realized, while also real
The overall digital for having showed image processing and having arranged.The main task of satellite is long-term, continuous, stable, rapidly acquisition covering
The high-resolution stereopsis and multispectral image of China for survey of territorial resources and are monitored, are prevented and reduced natural disasters, agriculture, forestry, water conservancy, life
Apply offer service in the fields such as state environment, city construction planning and building, traffic, national Important Project.
From the point of view of Transmission Line Design and operating condition, suburb, outdoor transmission line of electricity pass through vegetation to some extent
Overlay area, this directly determines the design and operation of transmission line of electricity.Therefore, the vegetation in power transmission line corridor region is investigated further
Distribution and characteristic are particularly important.Classification is an important link of remote sensing image interpretation and the heat in Remote Sensing Study field
Point.Nowadays greatly improving with image resolution ratio, the ground object target details for being included is more obvious, shapes textures structure
Etc. information it is also more prominent.In face of characteristic information abundant, selection is targetedly characterized in improving a pass of classification performance
Key link.Grandson is aobvious et al.[1]Result of study show effectively extract high-resolution remote sensing image using shape and color characteristic
Middle building target.Et al.[2]By the semantic tagger of remote sensing image and LDA (Latent Dirichlet
Allocation) model[3]In conjunction with achieving certain effect.LIU C et al.[4]Result of study show several feature groups
It closes, classifying quality is better than single feature, but not better using the more classifying qualities of type of feature.For how to have selected
Targetedly feature, LIU C et al.[4]It proposes to complete the optimal of feature using augmentation LDA model (augmented LDA, aLDA)
Selection.
In recent years, it expresses and learns for that can obtain facilitating the multi-level features that progress image is carried out by bottom to high level
It practises, deep learning feature has received widespread attention in machine learning.Hinton study group proposes deepness belief network (deep
belief network)[5].From structure, deepness belief network and the difference of traditional multi-layer perception (MLP) less, and with have
Supervised learning algorithm is the same.Unlike unique, deepness belief network needs first to carry out unsupervised learning before doing supervised learning,
Then it is trained using weight obtained by unsupervised learning as the initial value of supervised learning.Deepness belief network is in comparison disagreement algorithm
CD-n algorithm is proposed on the basis of (contrastive divergence, CD), need to only sample n times may be updated a weight.
Weight is just fixed after having learnt a limited Boltzmann machine model；It is superimposed one layer of new Hidden unit again, makes to be limited glass originally
The hidden layer of the graceful machine model of Wurz becomes input layer, thus constructs a new limited Boltzmann machine model.It uses again later
The weight of the new limited Boltzmann machine model of same method study.The rest may be inferred, stackable multiple limited Boltzmanns out
Machine model, to constitute deepness belief network.The weight that limited Boltzmann machine model learning is arrived is as deepness belief network
Initial weight, then algorithm is learnt with backpropagation (Back Propagation), to form deepness belief network
Learning method.
Following bibliography involved in text:
[1] grandson is aobvious, and the object-based Boosting method of Wang Hongqi, Zhang Zheng automatically extracts in high-resolution remote sensing image
Building target [J] electronics and information journal, 2009,31 (1): 177-181.
[2]DATCU M.Semantic annotation of satellite
Images using latent dirichlet allocation [J] .IEEE, Geoscience and Remote
Sensing Letters, 2009,7 (1): 28-32.
[3] BLEI.D.M, NGA.Y, JORDAN.M.I.Latent dirichlet allocation [J] .Journal
Of Machine Learning Research, 2003,3:993-1022.
[4] LIU C, SHARANL, ADELSONEH, et al.Exploring features in a Bayesian
framework for material recognition[C].IEEE Conference on Computer Vision and
Pattern Recognition (CVPR) .Washington, DC:IEEE Computer Society, 2010:239-246.
[5]G.E.Hinton,S.Osindero,Y.-W.Teh,A fast learning algorithm for deep
Belief nets, Neural Computation, 2006, vol.18, pp.1527~1554
[6] FAN Rong-En, CHANG Kai-Wei, HSIEH Cho-Jui, et al.LIBLINEAR:A library
For large linear classification [J] .The Journal of Machine Learning Research,
2008,9:1871-1874.
Summary of the invention
The object of the present invention is to provide a kind of power transmission line corridor regions that sorting algorithm efficiency and nicety of grading can be improved
Vegetation classification method.
Present invention introduces a variety of characteristics of remote sensing image including deep learning feature, and optimize choosing to these features
It selects, is used to carry out vegetation classification after the feature after optimum choice is combined.
In order to achieve the above objectives, the present invention adopts the following technical scheme:
A kind of power transmission line corridor regional vegetation classification method, comprising:
Step 1, the feature of training sample and cross validation sample is extracted, this step further comprises:
1.1 extract scene unit from remote sensing image training sample, and scene unit is divided into training sample and cross validation sample
This；
1.2 are defined the scene type of each scene unit and are marked category label using artificial visual mode, are denoted as definition class
Other label, scene type include vegetation class and non-vegetation class；
1.3 extract the various features of each scene unit；
Step 2, characteristic optimization selection is carried out based on cross-validation method, this step further comprises:
Features various in feature set are quantified as visual vocabulary by 2.1 respectively, form the corresponding visual vocabulary table of various features
Di, by each visual vocabulary table DiLDA model is inputted, the corresponding latent semantic distribution probability vector θ of various features is obtainedi；The spy
Collection is the set that the various features of cross validation sample are constituted；
2.2 by the corresponding θ of the various features of cross validation sampleiAnd define category label input regularization logistic regression respectively
Classifier carries out cross validation, obtains the classification accuracy of various features, and the highest feature of classification accuracy is denoted as optimal characteristics,
R is initialized using the classification accuracy of optimal characteristics；
A kind of other features and optimal characteristics are carried out tandem compound by 2.3, assemblage characteristic are denoted as, using a kind of other features
Visual vocabulary table D corresponding to optimal characteristics expands, and obtains new vision vocabulary D', and a kind of described other features refer to spy
Any feature in collection in addition to optimal characteristics；
New vision vocabulary D' is inputted LDA model by 2.4, and must dive semantic probability of happening vector θ '；
The definition category label of θ ' and cross validation sample input regularization logistic regression classifier is carried out intersection by 2.5 to be tested
Card obtains the classification accuracy r of present combination feature according to the cross validation sample predictions category label of outputnew；
If 2.6 rnew> r, by rnewIt is assigned to r, present combination feature is denoted as optimal characteristics, continues step 2.3；Otherwise,
Present combination feature, that is, preferred feature；
Step 3, the classification of remote sensing image test data, this step further comprise:
3.1 use the corresponding latent semantic distribution probability vector of the optimization feature of training sample and cross validation sample and determine
Adopted category label training classifier；
3.2 extract scene unit from remote sensing image test data, extract the optimization feature of each scene unit；
3.3 based on the corresponding latent semantic distribution probability of each scene unit optimization feature, and the classifier that use has been trained is to distant
Sense image test data are classified, wherein the scene unit for being divided into vegetation class constitutes vegetation area；
Above-mentioned latent semantic distribution probability vector inputs the acquisition of LDA model by that will optimize the corresponding visual vocabulary table of feature.
Above-mentioned various features include SIFT feature, DAISY feature, LBP feature, BRIEF feature and CNN feature.
Use k averaging method or sparse coding method by characteristic quantification for visual vocabulary in sub-step 2.1, i.e., to various features point
It is not clustered, the cluster centre of various features i.e. its corresponding visual vocabulary.
Scene unit is extracted in sub-step 1.1 and sub-step 3.1, specifically:
Remote sensing image is divided using uniform grid, a grid represents a scene unit, nothing between adjacent scene unit
Overlapping, the remote sensing image are remote sensing image training sample or remote sensing image test data.
The classifier used in step 3 is regularization logistic regression classifier or SVM classifier.
Compared to the prior art, the invention has the advantages that and the utility model has the advantages that
The method of the present invention is not necessarily to largely be iterated to calculate in the feature selecting stage, gained assemblage characteristic after optimized selection
With robustness, assemblage characteristic is used for the vegetation classification in remote sensing image power transmission line corridor region, is remarkably improved algorithm meter
Calculate efficiency and nicety of grading.
Detailed description of the invention
Fig. 1 is the specific flow diagram of the method for the present invention.
Specific embodiment
Below in conjunction with attached drawing, detailed description of the preferred embodiments.Technical solution of the present invention can be used
Computer software automatic running.
Step 1, the pretreatment of remote sensing image.
The pretreatment of remote sensing image includes the pretreatment of remote sensing image training sample and remote sensing image test data.
The pretreatment of remote sensing image training sample includes that extraction scene unit and artificial visual mode define scene type.It is first
First, substantially the identical image subblock of several sizes, i.e. scene unit will be divided by remote sensing image training sample；Using artificial visual
Mode defines the scene type of each scene unit and marks category label, and the category label of the label is denoted as definition category label.
The scene unit that remote sensing image training sample divides is divided into training sample and cross validation sample.Scene type is only in the present invention
Including vegetation class and non-vegetation class, forest and meadow belong to vegetation, and arable land and bare area belong to non-vegetation, remote sensing image test data
Pretreatment extract its scene unit.
When it is implemented, remote sensing image is " resource three " panchromatic image, remote sensing image, grid are divided using uniform grid
Just represent a scene unit, it is non-overlapping between adjacent scene unit.Final goal of the present invention is to give remote sensing image test data
All scene units assign a scene type label, and are distinguished using different colours.In the present embodiment, scene unit size is
100 pixel *, 100 pixel.
Step 2, the various features of scene unit are extracted.
Scene classification is carried out to remote sensing image, first have to extract scenery element images Expressive Features.Iamge description feature
It is various, it respectively gives priority in expression, has his own strong points in classification application.Due to power transmission line corridor area in remote sensing image
Vegetation density is not substantially uniformity in domain, and edge shape multiplicity has stronger texture, color characteristics.In view of remote sensing image spy
Levy it is varied, in order to illustrate the validity of this method, for the special characteristic set with 5 kinds of representative feature constructions,
It is not limited to this when using the present invention.
(1) SIFT (Scale Invariant Feature Transform, Scale invariant features transform) feature extraction.
SIFT feature is locality characteristic, all has invariance for image translation, scaling, rotation, and for affine
Transformation, visual angle change, illumination variation and noise etc. have very strong matching robustness, can be by finding in graphical rule space
To Local Extremum position, direction and dimensional information be indicated and obtain.
The present invention extracts the SIFT feature of scene unit using existing SIFT feature kit, and specific embodiment is such as
Under:
It detects scale spatial extrema point and obtains the accurate positioning of extreme point and the direction distribution of key point, generate characteristic point and retouch
State operator；When generating feature point description operator, the pixel region that size is 16 × 16 is chosen centered on key point, by the pixel
Region division is the sub-block that size is 4 × 4, calculates the histogram of gradients in 8 directions, i.e. SIFT feature in each sub-block respectively.
Extracted SIFT feature vector has 16 × 8=128 dimension.The crucial point image is detected under different scale space to be had
The Local Extremum of directional information.
(2) DAISY (daisy formula) feature extraction.
DAISY be towards dense characteristic extract can quickly calculate local image characteristics description son, essential idea and
SIFT is similar: the gradient orientation histogram of each sub-block of block statistics.Unlike, DAISY is changed on partition strategy
Into, using Gaussian convolution carry out gradient orientation histogram piecemeal converge, using Gaussian convolution can quickly it is computational can be quick
Densely carry out the extraction of DAISY feature.
DAISY feature can be used existing DAISY feature tools packet and extract.
(3) LBP (Local Binary Pattern, Local Binary Pattern) feature extraction.
LBP feature is a kind of operator for describing image local textural characteristics, has rotational invariance and gray scale constant
The remarkable advantages such as shape.LBP feature can be carried out using existing LBP feature tools packet, and specific embodiment is as follows:
(a) it will test the zonule that window is divided into 16 × 16, be denoted as cell.
(b) this step carries out pixel each in cell one by one: by the gray value of 8 adjacent pixels of current pixel point
It is compared with current pixel point, if neighbor pixel gray value is greater than the gray value of current pixel point, current pixel point quilt
It is otherwise 0 labeled as 1.In this way, adjacent 8 pixels in 3 × 3 neighborhood of current pixel point, which are compared, can produce 8 binary systems
Number, i.e. the LBP value of current pixel point.
(c) it calculates each pixel in cell and is labeled the frequency that numerical value (0 and 1) occurs, obtain all pixels point LBP in cell
The histogram of all cell is normalized in the histogram of value.
(d) feature vector that the histogram after all cell normalization connects into, i.e. LBP feature vector.
(4) BRIEF (Binary Robust Independent Element Feature, two-value robust independent element)
Feature extraction.
BRIEF feature describes operator using the feature that binary coding method extracts characteristic point peripheral region, and BRIEF feature is retouched
It states operator to calculate simply, memory space describes operator again smaller than SIFT feature.It is used since BRIEF feature describes operator
Hamming distance is from being compared, and matching speed is faster.
BRIEF feature can be extracted using existing BRIEF feature tools packet, and specific embodiment is as follows:
(a) region of square is selected centered on characteristic point, the characteristic point is the tool directional information detected
Local Extremum.
(b) Gauss nuclear convolution is carried out to the region, eliminates partial noise.
(c) the random point that generates is to < x, y > in the area, if the pixel value of x is greater than y, return value 1；Otherwise it returns
0。
(d) step (c) n times are repeated, N is that empirical value describes to get the binary-coded feature to one 256 dimensions, i.e.,
BRIEF feature.
(5) CNN feature (convolutional neural networks feature) is extracted.
CNN feature includes convolutional layer (convolution layer), pond layer (pooling layer) and full articulamentum
(fully connected layer).Complete depth convolutional neural networks are by multiple convolution, pond and full articulamentum series connection group
It closes and constitutes.In convolutional layer, input picture or input feature vector figure and several filter groups (also referred to as convolution kernel) carry out convolutional filtering
Obtain characteristic pattern；Then, to pond layer is sent into after gained characteristic pattern nonlinear processing, above-mentioned nonlinear processing is using non-linear
Function carries out, and common nonlinear function has Sigmoid function, Tanh function, ReLU (Rectified Linear Unit) letter
Number etc..
Pondization operation is carried out to the characteristic pattern of input in the layer of pond.It is identical that characteristic pattern is divided into size by pondization operation
Square region, to each region Counting statistics amount, such as the average value of fixed size window all pixels response in characteristic pattern or most
Big value.By pond layer, it is equivalent to and down-sampling is carried out to characteristic pattern, obtain the lesser characteristic pattern of size.Pond layer
Output can reconnect a convolutional layer, and the output of the convolutional layer reconnects another pond layer, in reasonable network number of plies premise
Under, according to convolutional layer-pond layer sequential iteration.The statistic of the last layer pond layer output is sent into full articulamentum.
Full articulamentum is made of several full connection hidden layers and Softmax Regression decision-making level.Convolutional neural networks
Training utilize back-propagation algorithm complete.Training is completed after obtaining each layer of network parameter, to any one width input picture
Can be calculated by feedforward (Feed-forward) mode the input picture feature, i.e. CNN feature.CNN feature has good
Translation and scale invariance.
Step 3, the characteristic optimization selection based on semantic model
For training sample and cross validation sample, it is based on semantic model, remote sensing shadow is directed to from the extracted feature of step 2
As the optimization feature of test data, will be used to classify after the combination of selected optimization feature later.
Semantic model comes from the data processing model in natural language processing, for expressing labyrinth and abundant language
Justice.Whether the image scene classification work based on semantic model is generally completed by including latent semanteme in analysis image.This reality
Example is applied using LDA (Latent Dirichlet Allocation, potential Di Li Cray distribution) model, the model is by latent semanteme
Hybrid weight is considered as the potential stochastic variable of k dimension parameter.
It can be obtained by LDA model:
Marginal probability p (D | α, β) is calculated according to formula (1):
In formula (1)~(2): D indicates corpus；D indicates document serial number；M indicates total number of documents in corpus D；N is indicated
Document Length；NdIndicate the length of d-th of document；W is vocabulary distribution statistics, wnFor n-th of vocabulary, wdnFor in d-th of document
N-th of vocabulary；θ is latent semantic distribution probability, θdIndicate semantic distribution of diving in d-th of document；Z indicates latent semanteme, znIt is n-th
A latent semanteme, zdnFor n-th in d-th of document latent semanteme；α, β are respectively hyper parameter；P (θ, z, w | α, β) it indicates in α, β condition
The joint probability of lower θ, z, w；P (θ | α) indicates the distribution of α condition dive semanteme；p(θd| α) it indicates under the conditions of α in d-th of document
Latent semantic distribution；p(zn| θ) indicate latent semanteme znProbability of happening；p(zdn|θd) it is based on semantic distribution of diving in d-th of document
Latent semanteme znProbability of happening；p(wn|zn, β) and indicate zn, vocabulary w under the conditions of βnThe probability of generation；p(wdn|zdn, β) and it indicates d-th
Document is in zn, vocabulary w under the conditions of βnThe probability of generation.
When estimating hyper parameter α, β, variation reasoning (variational inference) can be used or Markov Chain covers spy
The methods of Caro sampling method (Markov Chain Monte Carlo, MCMC).
Augmentation LDA model (aLDA) proposed in document [4] completes the optimum choice to feature using greedy algorithm.
Its core concept are as follows: in the cross validation stage, select from feature set makes a kind of maximum feature of classification accuracy rate every time, i.e., optimal
The optimal characteristics and other features are combined by feature, the classification accuracy rate of feature after combination are calculated, until classification accuracy rate
Until no longer rising.The method of the present invention then completes the optimum choice of feature using LDA model.
In sorting phase, document [4] obtains category label by maximum a posteriori principle using the parameter of aLDA model itself,
That is:
In formula (3): λc=log πc, πcIndicate parameter π corresponding to the C classification；C is true category label, is obeyed with π
For the multinomial distribution of parameter, π is the parameter of multinomial distribution here；C*For the scene type label of estimation；L(αc, η) and it is that model is joined
The maximization lower bound of variation reasoning, α in number estimationcIt is function lower bound for the corresponding hyper parameter α of the C classification in LDA model, η
Parameter.
When feature combines, if there is m kind feature available, it is corresponding with m visual vocabulary table, vocabulary table size is denoted as |
Di|=Vi., vocabulary table size, that is, Document Length, ViIndicate the quantity that visual vocabulary is corresponded to after i-th kind of characteristic quantification, DiIt is i-th kind
The visual vocabulary table length of visual vocabulary composition is corresponded to after characteristic quantification.Visual vocabulary corresponding to i-th kind of feature is represented byNiFor the corresponding visual vocabulary number of i-th kind of feature.Due to various features it is independent into
Row cluster operation, by the visual vocabulary of different characteristic Combination,
It indicates are as follows:
Wherein N1、N2、…NmValue be manually set.
It in document [4], needs based on formula (3), using LDA model inherent parameters, is constantly iterated calculating until classification
Accuracy rate λcIt is no longer changed, it is longer to calculate the time.Simultaneously as directly completing classification using model inherent parameters, handing over
Fork verifying and last test phase must individually be trained for every a kind of sample, to obtain corresponding parameter value of all categories.
It will so make the calculating time in the stage linearly increasing with the increase of sample class number.
In view of the above problems, the present invention is in the cross validation stage by LDA model and regularization logistic regression classifier
(Logistic Regression Classifier) is combined, and the cross validation stage only need to be based on all categories sample to LDA mould
Type carries out the primary optimization and automatic selection trained and can be completed to feature.Semantic vector of diving in LDA model is inputted into regularization
Logistic regression classifier is to replace formula (3), to directly acquire category label corresponding to cross validation sample.Such one
Come, just no longer needs to carry out a large amount of interative computations in the cross validation stage.Meanwhile after introducing regularization logistic regression classifier,
All samples once train to semantic model, do not need to be iterated computation model one by one further according to specimen types
Parameter, so as to effectively improve computational efficiency.
The characteristic optimization choice phase carries out following steps based on scene cell data in cross validation sample:
(1) features various in feature set are carried out respectively: is visual vocabulary by characteristic quantification, visual vocabulary is connected into this kind
The visual vocabulary table D of featurei, DiIndicate the visual vocabulary table of i-th kind of feature；By each visual vocabulary table DiLDA model is inputted, is obtained
To the corresponding latent semantic distribution probability vector θ of various featuresi.Here feature set refers to that the various features of cross validation sample are constituted
Set.
Characteristic quantification mode has diversity, such as k averaging method or sparse coding method can be used.This example utilizes k mean value
(k-means) features various in feature set are clustered respectively, corresponding cluster centre this kind of spy in series of various features
Levy corresponding visual vocabulary table Di。
(2) according to the corresponding θ of various features of cross validation sampleiAnd category label is defined, it is returned using regularization logic
Return classifier to carry out cross validation respectively, it is accurate that the classification classified using various features is calculated according to cross validation results
The highest feature of classification accuracy is denoted as optimal characteristics by rate, initializes r using the classification accuracy of optimal characteristics.
In this sub-step, cross validation carries out respectively for various features: using current kind of feature pair of cross validation sample
The θ answerediRegularization logistic regression classifier is trained with category label is defined, the regularization logistic regression classification that use has been trained
Device predicts that the category label of cross validation sample, the category label predicted are denoted as prediction category label, compare cross validation sample
This definition category label and prediction category label, obtains the classification accuracy of current kind of feature.
(3) by a kind of other features and optimal characteristics tandem compound, it is denoted as assemblage characteristic, it is corresponding using a kind of other features
Visual vocabulary table expand optimal characteristics visual vocabulary table, obtain new vision vocabulary D'.A kind of described other features refer to spy
Any feature in collection in addition to optimal characteristics；
(4) new vision vocabulary D' is inputted to the LDA model trained, obtains the corresponding latent semantic distribution of current assemblage characteristic
Probability vector θ '.
(5) the definition category label of θ ' and cross validation sample input regularization logistic regression classifier intersect and be tested
Card obtains the classification accuracy r of present combination feature according to the cross validation sample predictions category label of outputnew；
(6) judge rnewWith the size of r, if rnew> r, by rnewValue is assigned to r, and present combination feature is denoted as optimal characteristics,
It repeats step (3)；Conversely, stopping characteristic optimization selection, present combination feature optimizes feature.
Step 4, the classification of remote sensing image test data is carried out based on optimization feature.
Various features in the optimization feature of training sample and cross validation sample are sequentially connected end to end, as training sample
Vocabulary inputs LDA model, obtains latent semantic ProbabilityDistribution Vector θcom；By latent semantic distribution probability vector θcomInput classification
Device, meanwhile, the definition category label of training sample and cross validation sample is inputted, also for being trained to classifier.Using
The classifier trained classifies to the scene unit of remote sensing image test data, wherein being divided into the scene unit of vegetation class
Constitute vegetation area.
It is pointed out in document [6] in Finite Samples and higher characteristic dimension using regularization logistic regression classifier, classification
Effect outline is better than Linear SVM classifier, and has a clear superiority in speed.Therefore, the present invention selects Liblinear work
Existing regularization logistic regression classifier in tool packet.
Logistic regression is study f:X → Y equation or the method for P (Y | X), and Y is discrete value, X=< X here1,
X2,...,Xn> it is that wherein each variable is discrete or continuous value for any one vector.Logistic regression classifier through overfitting,
It is one group of weight w0,w1,...,wm.When test sample concentrate test data come then, this group of weight according to test data
The mode of linear weighted function finds out a z value, and the process for solving z is to be learnt corresponding linear out by way of machine learning
Curved surface is fitted training data sample:
Z=w0+w1·X1+w2·X2+...+wm·Xm (5)
Wherein, X1,X2,...,XmIt is each feature of certain sample data, dimension m, according to the form of sigmoid function
It finds out:
It can be seen that the value range of σ (z) (a possibility that representing certain classification) between [0,1], input is entire reality
Number field, curve can continuously be led.When solving two classification problems, a threshold value can be set and distinguished classification.
Specific embodiment described herein is only an example for the spirit of the invention.The neck of technology belonging to the present invention
The technical staff in domain can do the similar mode of various modify or supplement or adopts to described specific embodiment and substitute,
However, it does not deviate from the spirit of the invention or beyond the scope of the appended claims.
Claims (5)
1. a kind of power transmission line corridor regional vegetation classification method, characterized in that include:
Step 1, the feature of training sample and cross validation sample is extracted, this step further comprises:
1.1 extract scene unit from remote sensing image training sample, and scene unit is divided into training sample and cross validation sample；
1.2 are defined the scene type of each scene unit and are marked category label using artificial visual mode, are denoted as and are defined classification mark
Number, scene type includes vegetation class and non-vegetation class；
1.3 extract the various features of each scene unit；
Step 2, characteristic optimization selection is carried out based on cross-validation method, this step further comprises:
Features various in feature set are quantified as visual vocabulary by 2.1 respectively, form the corresponding visual vocabulary table D of various featuresi, will
Each visual vocabulary table DiLDA model is inputted, the corresponding latent semantic distribution probability vector θ of various features is obtainedi；The feature set
That is the set of the various features composition of cross validation sample；
2.2 by the corresponding θ of the various features of cross validation sampleiAnd define category label input regularization logistic regression classifier into
Row cross validation obtains the classification accuracy of various features, and the highest feature of classification accuracy is denoted as optimal characteristics, and use is optimal
The classification accuracy of feature initializes r；
A kind of other features and optimal characteristics are carried out tandem compound by 2.3, are denoted as assemblage characteristic, using a kind of other features to most
The corresponding visual vocabulary table D of excellent feature is expanded, and obtains new vision vocabulary D', and a kind of described other features refer to feature set
In any feature in addition to optimal characteristics；
New vision vocabulary D' is inputted LDA model by 2.4, and must dive semantic distribution probability vector θ '；
The definition category label of θ ' and cross validation sample input regularization logistic regression classifier is carried out cross validation by 2.5,
The classification accuracy r of present combination feature is obtained according to the cross validation sample predictions category label of outputnew；
If 2.6 rnew> r, by rnewIt is assigned to r, present combination feature is denoted as optimal characteristics, continues step 2.3；Otherwise, currently
Assemblage characteristic, that is, preferred feature；
Step 3, the classification of remote sensing image test data, this step further comprise:
3.1 are sequentially connected end to end various features in the optimization feature of training sample and cross validation sample, as training sample
Vocabulary inputs LDA model, obtains latent semantic ProbabilityDistribution Vector θcom；By latent semantic distribution probability vector θcomInput classification
Device, meanwhile, the definition category label of training sample and cross validation sample is inputted, also for being trained to classifier；
3.2 extract scene unit from remote sensing image test data, extract the optimization feature of each scene unit；
3.3 based on the corresponding latent semantic distribution probability of each scene unit optimization feature, and the classifier that use has been trained is to remote sensing shadow
As test data is classified, wherein the scene unit for being divided into vegetation class constitutes vegetation area；
Above-mentioned latent semantic distribution probability vector inputs the acquisition of LDA model by that will optimize the corresponding visual vocabulary table of feature.
2. power transmission line corridor regional vegetation classification method as described in claim 1, it is characterized in that:
The various features include SIFT feature, DAISY feature, LBP feature, BRIEF feature and CNN feature.
3. power transmission line corridor regional vegetation classification method as described in claim 1, it is characterized in that:
Use k averaging method or sparse coding method by characteristic quantification for visual vocabulary in sub-step 2.1, i.e., to various features respectively into
Row cluster, the cluster centre of various features i.e. its corresponding visual vocabulary.
4. power transmission line corridor regional vegetation classification method as described in claim 1, it is characterized in that:
Scene unit is extracted in sub-step 1.1 and sub-step 3.2, specifically:
Remote sensing image is divided using uniform grid, a grid is to represent a scene unit, non-overlapping between adjacent scene unit；
In sub-step 1.1, remote sensing image is remote sensing image training sample；
In sub-step 3.2, remote sensing image is remote sensing image test data.
5. power transmission line corridor regional vegetation classification method as described in claim 1, it is characterized in that:
The classifier used in step 3 is regularization logistic regression classifier or SVM classifier.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510953921.4A CN105787501B (en) | 2015-12-17 | 2015-12-17 | Power transmission line corridor region automatically selects the vegetation classification method of feature |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510953921.4A CN105787501B (en) | 2015-12-17 | 2015-12-17 | Power transmission line corridor region automatically selects the vegetation classification method of feature |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN105787501A CN105787501A (en) | 2016-07-20 |
| CN105787501B true CN105787501B (en) | 2019-03-19 |
Family
ID=56390200
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201510953921.4A Expired - Fee Related CN105787501B (en) | 2015-12-17 | 2015-12-17 | Power transmission line corridor region automatically selects the vegetation classification method of feature |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN105787501B (en) |
Families Citing this family (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106326932A (en) * | 2016-08-25 | 2017-01-11 | 北京每刻风物科技有限公司 | Power line inspection image automatic identification method based on neural network and power line inspection image automatic identification device thereof |
| CN106548192B (en) * | 2016-09-23 | 2019-08-09 | 北京市商汤科技开发有限公司 | Image processing method, device and electronic equipment neural network based |
| CN107273807A (en) * | 2017-05-19 | 2017-10-20 | 河海大学 | A kind of Remote Image Classification |
| US10628253B2 (en) * | 2017-05-24 | 2020-04-21 | Tata Consultancy Services Limited | Systems and methods for cognitive control of data acquisition for efficient fault diagnosis |
| CN107909105A (en) * | 2017-11-13 | 2018-04-13 | 上海交通大学 | A kind of Market Site Selection method and system |
| CN108898177A (en) * | 2018-06-26 | 2018-11-27 | 广东电网有限责任公司 | A kind of power transmission line corridor earth surface classification method, system, medium and equipment |
| CN111028255B (en) * | 2018-10-10 | 2023-07-21 | 千寻位置网络有限公司 | Farmland area pre-screening method and device based on priori information and deep learning |
| CN109800790B (en) * | 2018-12-24 | 2022-08-19 | 厦门大学 | Feature selection method for high-dimensional data |
| CN110008811A (en) * | 2019-01-21 | 2019-07-12 | 北京工业职业技术学院 | Face identification system and method |
| CN110059749B (en) * | 2019-04-19 | 2020-05-19 | 成都四方伟业软件股份有限公司 | Method and device for screening important features and electronic equipment |
| CN110135627B (en) * | 2019-04-23 | 2021-07-16 | 中国农业大学 | Water resource optimization method and device |
| CN110929618B (en) * | 2019-11-15 | 2023-06-20 | 国网江西省电力有限公司电力科学研究院 | Potential safety hazard detection and assessment method for power distribution network crossing type building |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103413142A (en) * | 2013-07-22 | 2013-11-27 | 中国科学院遥感与数字地球研究所 | Remote sensing image land utilization scene classification method based on two-dimension wavelet decomposition and visual sense bag-of-word model |
| CN103810500A (en) * | 2014-02-25 | 2014-05-21 | 北京工业大学 | Place image recognition method based on supervised learning probability topic model |
| CN103984963A (en) * | 2014-05-30 | 2014-08-13 | 中国科学院遥感与数字地球研究所 | Method for classifying high-resolution remote sensing image scenes |
| CN104376326A (en) * | 2014-11-02 | 2015-02-25 | 吉林大学 | Feature extraction method for image scene recognition |
-
2015
- 2015-12-17 CN CN201510953921.4A patent/CN105787501B/en not_active Expired - Fee Related
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103413142A (en) * | 2013-07-22 | 2013-11-27 | 中国科学院遥感与数字地球研究所 | Remote sensing image land utilization scene classification method based on two-dimension wavelet decomposition and visual sense bag-of-word model |
| CN103810500A (en) * | 2014-02-25 | 2014-05-21 | 北京工业大学 | Place image recognition method based on supervised learning probability topic model |
| CN103984963A (en) * | 2014-05-30 | 2014-08-13 | 中国科学院遥感与数字地球研究所 | Method for classifying high-resolution remote sensing image scenes |
| CN104376326A (en) * | 2014-11-02 | 2015-02-25 | 吉林大学 | Feature extraction method for image scene recognition |
Non-Patent Citations (2)
| Title |
|---|
| Exploring features in a Bayesian framework for material recognition;Ce Liu等;《2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition》;20100805;第239-246页 |
| Semantic Annotation of Satellite Images Using Latent Dirichlet Allocation;Marie Lienou等;《IEEE Geoscience and Remote Sensing Letters 》;20100131;第7卷(第1期);第28 - 32页 |
Also Published As
| Publication number | Publication date |
|---|---|
| CN105787501A (en) | 2016-07-20 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN105787501B (en) | Power transmission line corridor region automatically selects the vegetation classification method of feature | |
| Khelifi et al. | Deep learning for change detection in remote sensing images: Comprehensive review and meta-analysis | |
| CN110689086B (en) | Semi-supervised high-resolution remote sensing image scene classification method based on generating countermeasure network | |
| CN107092870B (en) | A kind of high resolution image Semantic features extraction method | |
| CN108805200B (en) | Optical remote sensing scene classification method and device based on depth twin residual error network | |
| CN108021947B (en) | A kind of layering extreme learning machine target identification method of view-based access control model | |
| CN105550709B (en) | A kind of remote sensing image power transmission line corridor wood land extracting method | |
| CN112507793A (en) | Ultra-short-term photovoltaic power prediction method | |
| Deepan et al. | Object classification of remote sensing image using deep convolutional neural network | |
| Xu et al. | Multi-structure joint decision-making approach for land use classification of high-resolution remote sensing images based on CNNs | |
| Peng et al. | Battlefield image situational awareness application based on deep learning | |
| Bai et al. | Geographic mapping with unsupervised multi-modal representation learning from VHR images and POIs | |
| CN111611960B (en) | Large-area ground surface coverage classification method based on multilayer perceptive neural network | |
| Gao et al. | Road extraction using a dual attention dilated-linknet based on satellite images and floating vehicle trajectory data | |
| Tripathi et al. | Real time object detection using CNN | |
| Du et al. | Training SegNet for cropland classification of high resolution remote sensing images | |
| Pratiwi et al. | Early detection of deforestation through satellite land geospatial images based on CNN architecture | |
| Liu et al. | Contour-oriented cropland extraction from high resolution remote sensing imagery using richer convolution features network | |
| CN118135209A (en) | Weak supervision semantic segmentation method based on shape block semantic association degree | |
| Chen et al. | Class-aware domain adaptation for coastal land cover mapping using optical remote sensing imagery | |
| Guo et al. | Building-Road Collaborative Extraction From Remote Sensing Images via Cross-Task and Cross-Scale Interaction | |
| Yu et al. | Coupling dual graph convolution network and residual network for local climate zone mapping | |
| Yang et al. | ICSFF: Information Constraint on Self-Supervised Feature Fusion for Few-Shot Remote Sensing Image Classification | |
| Lu et al. | Monocular semantic occupancy grid mapping with convolutional variational auto-encoders | |
| Jing et al. | Time series land cover classification based on semi-supervised convolutional long short-term memory neural networks |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20190319 Termination date: 20211217 |
|
| CF01 | Termination of patent right due to non-payment of annual fee |