CN105550709A - Remote sensing image power transmission line corridor forest region extraction method - Google Patents

Abstract

The present invention discloses a remote sensing image power transmission line corridor forest region extraction method. The method comprises the steps of (1) training an SVM classifier: extracting a scene unit training sample from a remote sensing training sample and using an artificial visual way to define the scene category of each scene unit training sample and carrying marking by using a mark number; extracting the global characteristic of each scene unit training sample; and combining the scene category mark number of the scene unit training sample and the global characteristic to train the SVM classifier, (2) using the trained SVM classifier to classify the remote sensing image test data: extracting a scene unit from the remote sensing image test data; extracting the global characteristic of each scene unit; and using the trained SVM classifier to extract a forest region based on the global characteristic combination of the scene units. According to the method, the completeness of extracted object description can be enhanced, and thus the accuracy of the forest area extraction is improved.

Description

A kind of remote sensing image power transmission line corridor wood land extracting method

Technical field

The invention belongs to remote sensing image intelligent analysis technical field, particularly a kind of remote sensing image power transmission line corridor wood land extracting method.

Background technology

Transmission line of electricity is the lifeblood of operation of power networks, is " lifeline " that involve the interests of the state and the people.Along with the fast development of Chinese national economy, the building cause of China's electrical network has stepped into the brand-new epoch.Extra high voltage network has become the main grid structure of China's electrical network, and UHV transmission line also in succession builds up and puts into effect.It should be noted that when transmission line of electricity is much more along the line through wood land, because in power transmission line corridor, forest weeds are numerous, suffering from a severe drought, plough in the spring when to open up wasteland etc. and very easily cause mountain fire, cause transmission line of electricity to trip.Power transmission line corridor region, Hubei has the features such as weather complexity, landform is changeable, vegetation is luxuriant, and especially in recent years, the power failure of the tripping operation on a large scale fault of stop that transmission line of electricity causes because of brushfire, burn the grass on waste land etc. inside the province gets more and more.From 2005 so far, Hubei Province's internal cause mountain fire and the power transmission and transformation line trip accident that causes has more than 10 times.Particularly winter dryness in 2013, Hubei within the border only near 500kV circuit every day all occur more than 30 and play mountain fire, make electrical network each O&M department drop into a large amount of manpower and materials and take precautions against.Therefore, carry out effective identification and extraction to the forest of the multiple sensitizing range in power transmission line corridor region, particularly mountain fire, to instructing, power transmission line corridor vegetation management meaning is important.

No. three, resource is the civilian stereo mapping satellite of China's first high precision, the useful load such as the three line scanner mapping camera that it is equipped with and multispectral camera can adopt three line scanner imaging mode, and by facing, forward sight, backsight and multispectral camera obtain the high-resolution remote sensing image of institute observation area.In image except the spectral information containing common remote sensing image, also comprise the shape and structure of a large amount of earth's surfaces object, texture information, contextual information etc.Relative low resolution multispectral image, high-resolution remote sensing image has abundant spatial information and atural object geometry, and texture information is also more obvious, is more convenient to the attributive character of cognitive atural object.Therefore more concerns are also received to the fast understanding of " No. three, resource " satellite image scene and automatic marking problem.

Along with the development of power grid construction, high pressure, UHV transmission line will cover more multi-environment complex region, especially be the emphasis of showing great attention to extensive forest region.Understand forest-covered area and growth characteristics, reliable basis can be provided for the design of power transmission line corridor and external insulation coordinate; Meanwhile, to choose reasonable transmission line of electricity power transmission path trend and vegetation management, there is important references to be worth.Because remote sensing image is subject to such environmental effects in imaging process, same classification scene is made to present larger difference on illumination, direction, yardstick.On the other hand, different classes of scene may comprise same target classification, just because the difference of target classification on locus and density degree result in the difference of scene type ^[1].Therefore, due to complicacy and the polytrope of remote sensing scene, make therefrom to be extracted in order to a challenging and far reaching job the forest atural object intellectuality in power transmission line corridor region.

Along with the universal of computer technology and development, intelligent, mechanization has become the focus received much concern.Human brain working method is copied in machine learning, allows computer learning be similar to the working method of brain.The various simple motions completed in mankind's daily life, if allow computing machine complete, then need the neural network utilizing high complexity ^[2].The people such as Hinton ^[3]the degree of depth study proposed, has started a tide of neural network research.Degree of depth study is the multi-level expression and extraction study carried out to obtain contributing to understanding data institute's elocutionary meanings such as picture, sound, text ^[4].And use simple machine learning, such as, by the neural network containing 2 or 3 hidden layers, the differentiation decision-making similar with the mankind cannot be reached.This just needs the learner of multilayer, and Level by level learning also the Knowledge delivery learning to arrive to lower one deck, so that lower floor can obtain the expression form of higher level, is expected to obtain the conclusion similar with the mankind ^[5].

Following list of references is related in literary composition:

[1] Xu Kan, Yang Wen, Chen Lijun etc. utilize the remote sensing images scene classification [J] of topic model. Wuhan University Journal: information science version, 2011, (5): 540-543.

[2]HAYKINS.NeuralNetworks:Acomprehensivefoundation[M].2 ^nded.NewYork:Prentice-Hall,1999.

[3]HINTONGE,MCCLELLANDJL,RUMELHARTDE.DistributedRepresentations[M].Cambridge:MITPress,1986.

[4]BENGIOY.LearningdeeparchitecturesforAI[J].FoundationsandTrendsinMachineLearning,2009,2(1):1-127.

[5] Li Haifeng, Li Chunguo. degree of depth study structure and method comparison analysis [J]. University Of Hebei's journal (natural science edition), 2012,05:538-544.

[6]A.Krizhevsky,I.Sutskever,andG.E.Hinton.Imagenetclassificationwithdeepconvolutionalneuralnetworks[C].InNIPS,2012.

[7]T.-H.Chan*,K.Jia,S.Gao,J.Lu,Z.Zeng,andY.Ma,"PCANet:Asimpledeeplearningbaselineforimageclassification？"[C],CoRR,2014.

Summary of the invention

The object of this invention is to provide a kind of remote sensing image power transmission line corridor wood land extracting method based on the combination of degree of depth learning characteristic.

The present invention is based on " No. three, resource " high-resolution remote sensing image and carry out the extraction of power transmission line corridor wood land.First, three kinds of degree of depth learning characteristics are carried out respectively to remote sensing image and extracts, be i.e. CNN feature, PCANet characteristic sum RandomNet feature; Then carried out tandem compound, obtained the combination of degree of depth learning characteristic; Finally according to the Feature Combination of this form, support vector machine (SVM) sorter is utilized to complete the classification of the forest to remote sensing image power transmission line corridor region.

For achieving the above object, the present invention adopts following technical scheme:

A kind of remote sensing image power transmission line corridor wood land extracting method, comprising:

Step 1, the training of SVM classifier, is specially:

1.1 extract scene unit training sample from remote sensing image training sample, and adopt artificial visual mode define the scene type of each scene unit training sample and adopt labelled notation, described scene type comprises forest class and non-forest class;

CNN feature, the PCANet characteristic sum RandomNet feature of each scene unit training sample of 1.2 extraction, then the three kinds of features extracted are carried out the global characteristics that tandem compound forms scene unit training sample;

1.3 adopt the scene type label of scene unit training sample and global characteristics to train SVM classifier;

Step 2, adopts the SVM classifier of having trained to classify to remote sensing image test data, is specially:

2.1 extract scene unit from remote sensing image test data;

CNN feature, the PCANet characteristic sum RandomNet feature of each scene unit of 2.2 extraction, then the three kinds of features extracted are carried out the global characteristics that tandem compound forms scene unit;

2.3 based on the global characteristics of scene unit, adopts the SVM classifier of having trained to classify to remote sensing image test data Scene unit, and the scene unit being wherein divided into forest class forms wood land.

Extract scene unit in sub-step 1.1 and sub-step 2.1, be specially:

Adopt uniform grid to divide remote sensing image, namely a grid represents a scene unit, zero lap between adjacent scene unit, and described remote sensing image is remote sensing image training sample or remote sensing image test data.

Compared to the prior art, tool of the present invention has the following advantages and beneficial effect:

With tradition based on hand-designed (hand-crafted) feature sorting technique compared with, the present invention utilizes the unsupervised learning feature of degree of depth learning characteristic, carries out auto adapted filtering process; , multiple degree of depth learning characteristic is combined meanwhile, merge the advantage of each degree of depth learning characteristic.The global characteristics combined by multiple degree of depth learning characteristic being introduced the forest classified in remote sensing image power transmission line corridor region, the completeness to extracting goal description can being strengthened, thus improve forest classified accuracy rate.

Accompanying drawing explanation

Fig. 1 is the concrete flow diagram of the inventive method.

Embodiment

Below in conjunction with accompanying drawing, the specific embodiment of the present invention is described in detail.Technical solution of the present invention can adopt computer software automatically to run.

Step 1, remote sensing image pre-service.

This step comprises the pre-service of remote sensing image training sample and remote sensing image test data.

The pre-service of remote sensing image training sample is comprised and extracts scene unit and define scene type in artificial visual mode.First, significantly remote sensing image training sample is divided into the identical image subblock of some sizes, obtains scene unit training sample; Then adopt artificial visual mode to define the scene type of each scene unit training sample, and adopt label to mark, Scene classification of the present invention only comprises forest class and non-forest class.The pre-service of remote sensing image test data is for extracting its scene unit.

During concrete enforcement, remote sensing image is " No. three, resource " panchromatic image, and adopt uniform grid to divide remote sensing image, namely grid just represents a scene unit, zero lap between adjacent scene unit.Final goal of the present invention gives a scene type label to all scene unit of remote sensing image test data, and adopt different colours to distinguish.In the present embodiment, scene unit size is 100 pixel * 100 pixels.

Step 2, the degree of depth learning characteristic of scene unit extracts.

Feature representation based on degree of depth study can complete automatically by low-level feature to the progressively transmission of high-level characteristic, and for manual feature, degree of depth learning characteristic more meets human visual system.The present invention is based on scene unit, CNN, PCANet and RandomNet tri-kinds of degree of depth learning characteristics extracting scene unit respectively describe scene unit, and are used for classification by after three kinds of degree of depth learning characteristic combinations of extracting.

The present invention need to extract scene unit training sample and the degree of depth learning characteristic of remote sensing image test data scene unit, will be described in detail to degree of depth learning characteristic leaching process below.

(1) CNN feature (convolutional neural networks feature) is extracted.

Convolutional neural networks comprises convolutional layer, pond layer (poolinglayer) and full articulamentum.Complete degree of depth convolutional neural networks is made up of multiple convolution, pond and full articulamentum tandem compound.In convolutional layer, input picture or input feature vector figure and some bank of filters (also referred to as convolution kernel) are carried out convolutional filtering and are obtained characteristic pattern; Then send into pond layer to after its non-linearization process, above-mentioned process calculates based on nonlinear function, and conventional nonlinear function has Sigmoid function, Tanh function, ReLU (RectifiedLinearUnit) function etc.

In the layer of pond, pondization operation is carried out to the characteristic pattern of input.Pondization operation is divided into the identical square region of size by characteristic pattern, to each region compute statistics, as mean value or the maximal value of pixel response values all in fixed size window in characteristic pattern.By pond layer, be equivalent to carry out down-sampling to characteristic pattern, obtain the characteristic pattern that a size is less.The output of pond layer can connect a convolutional layer again, and the output of this convolutional layer connects another pond layer again, under rational network number of plies prerequisite, according to the sequential iteration of convolutional layer-pond layer.The statistic that last one deck pond layer exports is sent in full articulamentum.

Full articulamentum is made up of some full connection hidden layers and SoftmaxRegression decision-making level.The training of convolutional neural networks utilizes back-propagation algorithm (BackPropagation) to complete.After having trained the network parameter obtaining every one deck, any width input picture is calculated to the feature of this input picture by feedforward (Feed-forward) mode, i.e. CNN feature.CNN feature has good translation and scale invariance.

(2) PCANet feature (PCA network characterization) is extracted.

PCANet feature and the maximum difference of CNN feature are, the convolution kernel of PCANet feature is obtained by unsupervised PCA training, but not adopt back-propagation algorithm.In the present embodiment, adopt existing PCANet feature tools bag to extract the PCANet feature of scene unit, key step is as follows:

A () be the little image block of stochastic sampling in scene module training sample, carry out PCA conversion to the little image block of sampling, and before extracting, L1 principal component vector is as convolution kernel, and convolution kernel and training image do common L1 the characteristic pattern that convolution obtains ground floor.

(b) characteristic pattern is sampled and utilize PCA to train L2 convolution kernel, with after L1 characteristic pattern convolution of ground floor the common L1 stack features figure of the second layer, often group comprises L2 characteristic pattern.

C () each stack features figure to the second layer carries out Hash coding respectively, code length is 2 ^l2, obtain L1 code pattern altogether.

D () carries out piecemeal process to each code pattern, obtain the statistic histogram of each sub-block frequency of occurrences, is connected by the histogram of L1 all sub-block of code pattern, i.e. the feature of input picture.

(3) RandomNet feature extraction.

The difference of RandomNet feature and PCANet feature is, the convolution kernel of its ground floor and the second layer is not obtained by non-supervisory PCA training, but the random random number producing one group of Gaussian distributed.In the present embodiment, existing RandomNet feature tools bag is adopted to extract the RandomNet feature of scene unit.

Step 3, degree of depth learning characteristic combines.

Three feature vectors of scene unit step 2 extracted connect successively with cascade, obtain the degree of depth learning characteristic combination of scene unit, i.e. the global characteristics of scene unit.Adopt the global characteristics of scene unit training sample to train SVM classifier, adopt the SVM classifier of having trained to classify to remote sensing image test data scene unit.

Step 4, utilizes SVM classifier to classify to remote sensing image test data.

By above step, the global characteristics of all scene unit training samples and remote sensing image test data scene unit can be obtained.Based on this, the present invention adopts SVM classifier to complete the forest classified in power transmission line corridor region, and this assorting process mainly comprises training SVM classifier and the classification of remote sensing image test data scene unit.

(1) adopt scene type label and the global characteristics training SVM classifier of scene unit training sample, obtain the model parameter of SVM classifier.

The SVM classifier used for criterion, carries out spatial division by finding optimal separating hyper plane to training sample with maximum class interval.The training process of SVM classifier finally changes into and solves following optimization problem:

$min\frac{1}{2}\left|\right|w|{|}^2+C\underset{i=1}{\overset{n}{\Σ}}{\mathrm{\ξ}}_i$

s.t,y _i(w ^Tφ(x _i)+b)≥1-ξ _i,i＝1,2,...n(1)

ξ _i≥0,i＝1,2,...n

In formula (1), w and b is defining classification lineoid w ^tφ (x _i) parameter vector of+b=0; C is used to the constant of two weights in Controlling object function; ξ _irepresent slack variable; φ (x _i) represent training sample x _inonlinear Mapping; y _irepresent training sample x _icategory label; N is number of training.

Adopt method of Lagrange multipliers, the optimization problem of belt restraining formula (1) Suo Shi is changed into unconstrained optimization problem, and its cost function L is:

$L=\frac{1}{2}\left|\right|w|{|}^2+C\underset{i=1}{\overset{n}{\Σ}}{\mathrm{\ξ}}_i-\underset{i=1}{\overset{n}{\Σ}}{\mathrm{\α}}_i\left(y_i\right(w^T\mathrm{\φ}\left(x_i\right)+b)-1+{\mathrm{\ξ}}_i)-\underset{i=1}{\overset{n}{\Σ}}{r}_i{\mathrm{\ξ}}_i---\left(2\right)$

In formula (2), α _i, r _ibe training sample x _ivariable to be optimized.

On the basis meeting Karush-Kuhn-Tucker (KKT) condition, the unconstrained problem shown in formula (2) can be changed into following optimization problem:

$max\underset{i=1}{\overset{n}{\Σ}}{\mathrm{\α}}_i-\frac{1}{2}\underset{i,j=1}{\overset{n}{\Σ}}\left({\mathrm{\α}}_i{\mathrm{\α}}_j{y}_i{y}_{j}K\right(x_i,x_j\left)\right)$

s.t.,0≤α _i≤C,i＝1,...,n(3)

$\underset{i=1}{\overset{n}{\Σ}}\left({\mathrm{\α}}_i{y}_i\right)=0$

In formula (3), meet between w and α

K (x, x _i) be defined kernel function in advance, be used for the inner product operation of training sample in feature space being changed into the kernel function of training sample in luv space and map.The present embodiment adopts radial basis function as follows as kernel function:

$K(x,x_i)=\exp \{-\frac{|x-x_i{|}^2}{256{\mathrm{\σ}}^2}\}---\left(4\right)$

Wherein, x is support vector, x _ifor training sample, σ is the variance of training sample.

Variable { α is obtained by separating optimization problem (3) _i} _{i=1 ..., n}optimal value, complete the training process of SVM classifier model parameter.

(2) SVM classifier of having trained is utilized to classify to scene unit all in remote sensing image test data, namely the prediction of scene type label is carried out to each scene unit, the classification of each scene unit can be obtained by the scene type label of prediction, adopt different colours to distinguish different scene type.

When utilizing the SVM classifier of having trained to carry out scene classification, only need, by the classification function curved surface of the proper vector of remote sensing image test data scene unit input SVM classifier, corresponding scene type label can be obtained, classification function f (H ^(t)) as follows:

$f\left(H^{\left(t\right)}\right)=\underset{i=1}{\overset{n}{\Σ}}\left({\mathrm{\α}}_i{y}_{i}K\right(H_i,H^{\left(t\right)}\left)\right)+b---\left(5\right)$

In formula (5), H _ifor training sample x _iglobal characteristics, H ^(t)represent the global characteristics of scene unit to be sorted.

Specific embodiment described herein is only to the explanation for example of the present invention's spirit.Those skilled in the art can make various amendment or supplement or adopt similar mode to substitute to described specific embodiment, but can't depart from spirit of the present invention or surmount the scope that appended claims defines.

Claims (2)

1. a remote sensing image power transmission line corridor wood land extracting method, is characterized in that, comprising:

Step 1, the training of SVM classifier, is specially:

1.1 extract scene unit training sample from remote sensing image training sample, and adopt artificial visual mode define the scene type of each scene unit training sample and adopt labelled notation, described scene type comprises forest class and non-forest class;

CNN feature, the PCANet characteristic sum RandomNet feature of each scene unit training sample of 1.2 extraction, then the three kinds of features extracted are carried out the global characteristics that tandem compound forms scene unit training sample;

1.3 adopt the scene type label of scene unit training sample and global characteristics to train SVM classifier;

Step 2, adopts the SVM classifier of having trained to classify to remote sensing image test data, is specially:

2.1 extract scene unit from remote sensing image test data;

CNN feature, the PCANet characteristic sum RandomNet feature of each scene unit of 2.2 extraction, then the three kinds of features extracted are carried out the global characteristics that tandem compound forms scene unit;

2.3 based on the global characteristics of scene unit, adopts the SVM classifier of having trained to classify to remote sensing image test data Scene unit, and the scene unit being wherein divided into forest class forms wood land.

2. remote sensing image power transmission line corridor wood land as claimed in claim 1 extracting method, is characterized in that:

Extract scene unit in sub-step 1.1 and sub-step 2.1, be specially:

Adopt uniform grid to divide remote sensing image, namely a grid represents a scene unit, zero lap between adjacent scene unit, and described remote sensing image is remote sensing image training sample or remote sensing image test data.