CN110826458A - Multispectral remote sensing image change detection method and system based on deep learning - Google Patents

Multispectral remote sensing image change detection method and system based on deep learning Download PDF

Info

Publication number
CN110826458A
CN110826458A CN201911050821.5A CN201911050821A CN110826458A CN 110826458 A CN110826458 A CN 110826458A CN 201911050821 A CN201911050821 A CN 201911050821A CN 110826458 A CN110826458 A CN 110826458A
Authority
CN
China
Prior art keywords
network
change detection
remote sensing
sample set
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.)
Withdrawn
Application number
CN201911050821.5A
Other languages
Chinese (zh)
Inventor
石爱业
高文静
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Hohai University HHU
Original Assignee
Hohai University HHU
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Hohai University HHU filed Critical Hohai University HHU
Priority to CN201911050821.5A priority Critical patent/CN110826458A/en
Publication of CN110826458A publication Critical patent/CN110826458A/en
Withdrawn legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/10—Terrestrial scenes
    • G06V20/13—Satellite images
    • 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
    • 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
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/20—Image preprocessing
    • G06V10/30—Noise filtering

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Computing Systems (AREA)
  • Multimedia (AREA)
  • Molecular Biology (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Astronomy & Astrophysics (AREA)
  • Remote Sensing (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a multispectral remote sensing image change detection method and a multispectral remote sensing image change detection system based on deep learning, which belong to the technical field of image processing, and comprise the following steps: calculating the change vector amplitude of the remote sensing image; obtaining a pseudo training sample set according to a change vector magnitude (EM) algorithm: a marked sample set (comprising a change type sample set and a non-change type sample set) and a non-marked sample set; constructing two network student networks and a teacher network, constructing a cross entropy loss function for a marked sample set, and constructing a mean square error function for a non-marked sample set; optimizing the student network by adopting a random gradient descent optimization algorithm, and updating the weight parameters of the teacher network in each training turn; and acquiring a corresponding final change detection result according to the final teacher network. In addition, a label-free sample set is added in the training of the network to participate in the training, so that the result of the change detection is more reliable and has more robustness.

Description

Multispectral remote sensing image change detection method and system based on deep learning
Technical Field
The invention belongs to the technical field of image processing, and particularly relates to a multispectral remote sensing image change detection method and system based on deep learning.
Background
The change detection of the remote sensing image is to quantitatively analyze and determine the characteristics and the process of the surface change from the remote sensing data in different periods. Various scholars put forward a plurality of effective detection algorithms from different angles and application researches, and generally speaking, according to whether a training sample is needed in the detection process, the change detection can be divided into three main categories: unsupervised change detection algorithms, semi-supervised change detection algorithms and supervised change detection algorithms. The unsupervised change detection algorithm does not need a training sample, and the modeling process does not need prior knowledge, so the algorithm is widely applied. The unsupervised change detection algorithm generally first constructs a difference image of two time phases, and then processes the difference image to obtain a change detection result. Conventional Change detection methods Change Vector Analysis (CVA), detection methods based on principal component transformation, clustering methods based on Fuzzy C-means (FCM), multivariate Change detection algorithms, Change detection algorithms based on undirected graphs (including MRF, CRF, etc.), and the like. In the unsupervised change detection modeling based on the deep learning, a difference image does not need to be constructed by a conventional change detection method, the depth information of two time phase images is directly utilized, the change information among different images can be better depicted, and the method has advantages in the field of multi-time phase change detection.
The university of Western electronic technology proposed a SAR image change detection based on unsupervised DBN neural network in the patent of its application, namely, unsupervised deep neural network-based SAR image change detection (patent application No. 201410818305.3, publication No. CN104517124A) and in the published "Change detection in synthetic aperture images based on deep neural network" (IEEE Transactions on Geoscience and removal Sensing,2016,27(1):125- & 137). The method mainly aims at the radar images with multiple time phases, the sample selection of the pseudo label is based on the FCM algorithm, and the reliability of the sample is not high.
A method for detecting Multispectral changes of a generation countermeasure network (GAN) is provided in an article A general characterization classification for Change Detection in multisection image (IEEE Journal of selected Topics in Applied Earth requirements and Remote Sensing,2019,12(1):321 addition 333). The method comprises the steps of firstly, jointly selecting a pseudo mark sample according to a CVA technology and an Otsu threshold value method, and in the training of generating the countermeasure network, obtaining three types of data: and training the network by the pseudo-labeled sample, the non-pseudo-labeled sample and the generated false data together, and finally obtaining a final change detection result according to the trained discrimination network. The disadvantage of this approach is that training of the network is prone to collapse of the pattern.
Disclosure of Invention
The invention aims to overcome the defects in the prior art and provides a multispectral remote sensing image change detection method and system based on deep learning, so that the final change detection result of a double-temporal multispectral remote sensing image is more reliable and stable.
In order to achieve the purpose, the technical scheme adopted by the invention is as follows: a multispectral remote sensing image change detection method based on deep learning comprises the following steps:
a. carrying out image registration on the remote sensing image, carrying out radiation correction by using a multivariate change detection method, and then calculating the change vector amplitude of the remote sensing image;
b. acquiring a pseudo training sample set by utilizing a maximum expectation algorithm according to the change vector amplitude;
c. two identical fully-connected deep learning networks were constructed: a teacher network and a student network;
d. training a teacher network and a student network through a pseudo training sample set;
e. and acquiring a final change detection result according to the teacher network.
The image registration comprises geometric coarse correction and geometric fine correction
The number of neurons in the first layer of the teacher network and the number of neurons in the second layer of the student network are 294, the number of neurons in the second layer of the teacher network and the number of neurons in the third layer of the student network are 64, the number of neurons in the third layer of the teacher network and the number of neurons in the fourth layer of the student network are 2, the input dimension of the teacher network and the input dimension of the student network are 294, and the output dimension of the teacher.
The teacher network and the student network take samples in a 7 x 7 square window as input, and the multispectral images of two time phases of each sample have 6 wave bands.
The teacher network and the student network both use batch normalization processing except for the output layer, and the middle layer uses the LeakyReLU (x) activation function:
LeakyReLU(x1)＝max(0,x1)+negative_slope*min(0,x1) (7)
in the formula, x1Is the output matrix of the output intermediate layer of the network and negative slope represents the angle controlling the negative slope.
The step d comprises the following steps:
da. Labeled data set of pseudo training sample set
Figure BDA0002255290400000031
And
Figure BDA0002255290400000032
inputting student network Ns and calculating loss function L of labeled data set1：
Wherein x represents the spectrum normalization characteristic value of the corresponding sample point, y is the label corresponding to the sample, and wsIs the weight of the student network, f represents the probability output of the student network Ns, lCERepresents a cross entropy loss function, which is defined as follows:
Figure BDA0002255290400000034
in the formula, k represents a tag in change detection, the value is {0, 1}, 0 represents no change, 1 represents a change, and w represents the weight of the network.
db. Unlabeled data set of pseudo training sample set
Figure BDA0002255290400000041
Input into student network Ns to obtain
Figure BDA0002255290400000042
Corresponding tag output
dc. Unlabeled data set of pseudo training sample set
Figure BDA0002255290400000044
Input to teacher network Nt to obtain
Figure BDA0002255290400000045
Corresponding output
Figure BDA0002255290400000046
g represents the probability output, w, of the teacher network NttA weight representing the teacher network Nt;
dd. Aiming at the unlabeled data set, expecting that the student network Ns is consistent with the predicted labels of the teacher network Nt, and constructing an unlabeled loss function L2：
L2＝||f(x；ws)-g(x；wt)2(10)
de. According to the total loss function L1+L2Updating the weight w of the student network Ns by adopting a back propagation algorithm and a random gradient descent algorithms。
df. Updating parameter w of teacher network Nt by using parameter of student network Nst：
Figure BDA0002255290400000047
In the formula, epc represents the number of times of training of the whole training sample, and α represents the decay rate;
dg. Repeating the steps da to df until the student network Ns converges, and obtaining the final weight w of the teacher network Nt according to the formula (11)t。
The multispectral remote sensing image change detection system based on deep learning comprises a processor and a storage device, wherein a plurality of instructions are stored in the storage device, and the storage device is used for the processor to load and execute the steps of any one of the methods.
Compared with the prior art, the invention has the following beneficial effects: according to the multispectral remote sensing image change detection method and system based on deep learning, the marked sample set and the unmarked sample set are obtained in an unsupervised mode, so that the defects of high cost and strong randomness of manual sample selection are overcome; the change detection system comprises two networks: the student network and the teacher network update parameters of the teacher network in each training step and guide learning of the student network in time, so that the final change detection result of the double-time-phase multispectral remote sensing image is more reliable and stable, and the detection precision is higher.
Drawings
Fig. 1 is a schematic flow chart illustrating an implementation process of a multispectral remote sensing image change detection method based on deep learning according to an embodiment of the present invention;
FIG. 2 is a 3 rd band schematic of a high resolution IKONOS image taken in Saudi Arabia Mina area, month 1, 2007 as used in an embodiment of the present invention;
FIG. 3 is a schematic 3 band diagram of a Mina region high resolution IKONOS image of Saudi Arabia, month 12, 2007 as used in an embodiment of the present invention;
FIG. 4 is a change detection reference (Ground Truth) image;
FIG. 5 is an EM-MRF algorithm detection result image;
FIG. 6 is a DBN algorithm detection result image;
FIG. 7 is an image of the detection results of the GAN algorithm;
fig. 8 is an image of a detection result obtained by the method of the embodiment of the present invention.
Detailed Description
A multispectral remote sensing image change detection method based on deep learning comprises the following steps: carrying out image registration on the remote sensing image, carrying out radiation correction by using a multivariate change detection method, and then calculating the change vector amplitude of the remote sensing image; obtaining a pseudo training sample set according to the magnitude of the variation vector and by utilizing a maximum expectation algorithm: a marked sample set (comprising a change type sample set and a non-change type sample set) and a non-marked sample set; two identical fully-connected deep learning networks were constructed: a teacher network and a student network; training a teacher network and a student network through a pseudo training sample set, wherein a cross entropy loss function is constructed for a marked sample set, and a mean square error function is constructed for a non-marked sample set; optimizing the student network by adopting a random gradient descent optimization algorithm, and updating the weight parameters of the teacher network in each training turn; and acquiring a corresponding final change detection result according to the final teacher network. In addition, a label-free sample set is added in the training of the network to participate in the training, so that the result of the change detection is more reliable and has more robustness.
The invention is further described below with reference to the accompanying drawings. The following examples are only for illustrating the technical solutions of the present invention more clearly, and the protection scope of the present invention is not limited thereby.
As shown in fig. 1, the implementation process of the multispectral remote sensing image change detection method based on deep learning of the present invention is schematically illustrated, and the implementation process includes the following steps:
step 1: inputting two high-resolution optical remote sensing images in the same area and different time phases and respectively recording as X1And X2。
Step 2: using remote sensing software ENVI to X1And X2Performing image registration, which comprises two steps of coarse correction and fine correction:
for geometric coarse correction, the method is realized by using related functions in ENVI4.8 software, and the specific operation steps are as follows: (1) displaying a reference image and an image to be corrected; (2) collecting ground control points GCPs; the GCPs should be uniformly distributed in the whole image, and the number of the GCPs is at least more than or equal to 9. (3) Calculating an error; (4) selecting a polynomial model; (5) resampling and outputting by adopting bilinear interpolation;
bilinear difference method, if the unknown function f is found to be (x, y) at point P, let us assume that the function f is known to be Q11＝(x1,y1),Q12＝(x1,y2),Q21＝(x2,y1),And Q22＝(x2,y2) Values of four points. If a coordinate system is chosen such that the coordinates of these four points are (0,0), (0,1), (1,0), and (1,1), respectively, then the bilinear interpolation formula can be expressed as:
f(x,y)≈f(0,0)(1-x)(1-y)+f(1,0)x(1-y)+f(0,1)(1-x)y+f(1,1)xy (1)
for geometric fine correction, performing geometric fine correction on the multispectral remote sensing image data subjected to geometric coarse correction by using an automatic matching and triangulation method;
the triangle subdivision method is that a Delaunay triangulation network is built by adopting a point-by-point insertion method, for each triangle, affine transformation model parameters in the triangle are determined by utilizing the row and column numbers of three vertexes of the triangle and the geographic coordinates of the same-name points of the corresponding reference image, and the image to be corrected is corrected to obtain the corrected remote sensing image.
And step 3: detection of X by Multivariate Change Detection (MAD) method1And X2And carrying out radiation normalization correction. The method comprises the steps of firstly finding a linear combination of brightness values of wave bands of two-stage images to obtain a difference image with enhanced change information, determining changed and unchanged areas through a threshold value, and then completing relative radiation correction through a mapping equation of two time phase pixel pairs corresponding to the unchanged areas.
And 4, step 4: for the input multi-temporal high-resolution image, the variation vector amplitude XMIs calculated as follows:
Figure BDA0002255290400000071
wherein, XM(i, j) represents the magnitude of the change vector at coordinate (i, j); b represents the wave band number of each time phase remote sensing image; (i, j) coordinates representing the remote sensing image; b represents the serial number of the band; x1b(i, j) represents a pixel value at the b-th band image (i, j) of the previous phase; x2b(i, j) represents the pixel value at the b-th band image (i, j) of the subsequent phase.
And 5: using Bayes principles and based on Expectation-Max (Expectation-Max)imzation, EM) algorithm obtains the optimal segmentation threshold T. Will | XMAnd taking the area with the value of-T | ≦ delta as a pseudo training sample set. Wherein δ is selected as XM15% of the dynamic range, the calculation process of the optimal segmentation threshold T and the construction of the pseudo training sample set are as follows:
1) suppose XMUnchanged class omega on imagenAnd variation class omegacObey the following gaussian distribution, namely:
Figure BDA0002255290400000072
wherein p (X)M(i,j)|ωl) Representing image XMThe pixel value at the upper coordinate (i, j) belongs to the class ωlConditional probability of, ωl∈{ωn,ωc}；σl，mlRespectively represent images XMUpper belongs to the class omegalVariance and mean of σl∈{σn,σc}，ml∈{mn,mcThe mean and standard deviation of the unchanged classes are mnAnd σnThe mean and standard deviation of the variation classes are mcAnd σc(ii) a Using EM algorithm, m can be estimatedn、σn、mcAnd σcThe following description will be given only by taking parameter estimation of unchanged classes as an example, and the parameter estimation of changed classes is similar.
Figure BDA0002255290400000081
Figure BDA0002255290400000082
Figure BDA0002255290400000083
Wherein, I and J respectively represent the number of rows and columns of the image, and t represents the number of iterations; p is a radical oft(XM(i, j)) represents the pixel value X at the number of iterations tM(i, j) overviewRate; p is a radical oft(ωn) Representing the prior probability of the unchanged class when the iteration number is t; p is a radical oft(XM(i,j)|ωn) Representing the pixel value X at the number of iterations tM(i, j) conditional probability of belonging to an unchanged class;
Figure BDA0002255290400000084
representing the unchanged class mean value when the iteration number is t;
Figure BDA0002255290400000085
representing the unchanged class variance when the iteration number is t + 1;
2) solving the change vector magnitude image X according to Bayes minimum error criterionMA segmentation threshold value T of (1);
3) determining a training sample set:
3.1) based on the threshold T estimated by the EM algorithm, XMThe samples greater than T are preliminarily determined as a change sample set (the change class is marked as 1), and the samples less than or equal to T are preliminarily determined as unchanged samples (the unchanged class is marked as 0), so that an initial change detection result C is obtained0。
3.2) to C0Mean filtering is performed with a window size ψ (ψ is set to an odd number greater than 1), samples at the center of the corresponding window with a filtering result of 1 are set to a changed class (denoted as 1), samples at the center of the corresponding window with a filtering result of 0 are set to a non-changed class (denoted as 0), and corresponding samples other than the above two cases are set to unmarked samples. The change class sample set at this time is recorded as
Figure BDA0002255290400000091
Set of invariant class samples as
Figure BDA0002255290400000092
Set of unlabeled samples
Figure BDA0002255290400000093
3.3) recording XMSet of coordinates greater than T is i1，XMSeat less than or equal to TThe standard set is i2. Respectively calculate XMAt i1And i2And are respectively expressed as
Figure BDA0002255290400000094
And
Figure BDA0002255290400000095
mixing XMIs greater than
Figure BDA0002255290400000096
Is marked as a change class, X is markedMIs less than
Figure BDA0002255290400000097
The sample of (1) is marked as a non-change class. The change class sample set at this time is recorded asSet of invariant class samples as
3.4) solving
Figure BDA00022552904000000910
And
Figure BDA00022552904000000911
obtaining a sample set of variation classes
Figure BDA00022552904000000912
To findAnd
Figure BDA00022552904000000914
obtaining the invariant sample set
Figure BDA00022552904000000915
Step 6: two identical fully-connected deep learning networks, namely a teacher network Nt and a student network Ns, are constructed. The number of neurons in the first layer of the network is 294 (samples in a 7 × 7 square window are used as input, and multispectral images of two time phases of each sample have 6 wave bands in total), the number of neurons in the second layer is 64, the number of neurons in the third layer is 128, and the number of neurons in the fourth layer is 2. The teacher (student) network has an input dimension of 294 and an output dimension of 2 (corresponding to both classes changed and unchanged). All layers of the teacher (student) network except the output layer use batch normalization, the middle layer uses the LeakyReLU (x) activation function:
LeakyReLU(x1)＝max(0,x1)+negative_slope*min(0,x1) (7)
in the formula, x1Is the output matrix of the output intermediate layer of the network and negative slope represents the angle controlling the negative slope.
And 7: network parameters, including network weights and biases, of the student network Ns and the teacher network Nt are randomly initialized.
And 8: starting network training, comprising the steps of:
8.1: labeled data set of pseudo training sample set
Figure BDA00022552904000000916
Andinputting student network Ns and calculating loss function L of labeled data set1：
Wherein x represents the spectrum normalization characteristic value of the corresponding sample point, y is the label corresponding to the sample, and wsIs the weight of the student network, f represents the probability output of the student network Ns, lCERepresents a cross entropy loss function, which is defined as follows:
Figure BDA0002255290400000102
in the formula, k represents a tag in change detection, the value is {0, 1}, 0 represents no change, 1 represents a change, and w represents the weight of the network.
8.2 unlabeled data set of pseudo training sample set
Figure BDA0002255290400000103
Input into student network Ns to obtainCorresponding tag output
Figure BDA0002255290400000105
f represents the probabilistic output of the student network Ns;
8.3 unlabeled dataset of pseudo-training sample setInput to teacher network Nt to obtain
Figure BDA0002255290400000107
Corresponding output
Figure BDA0002255290400000108
g represents the probability output, w, of the teacher network NttA weight representing the teacher network;
8.4, aiming at the unlabeled data set, expecting that the student network Ns is consistent with the predicted labels of the teacher network Nt, and constructing an unlabeled loss function L according to the steps 8.3 and 8.42：
L2＝||f(x；ws)-g(x；wt)||2(10)
8.5 according to the total loss function L1+L2Updating the weight w of the student network Ns by adopting a back propagation algorithm and a random gradient descent algorithms；
8.6 updating parameter w of teacher network Nt by using parameter of student network Nst：
Figure BDA0002255290400000109
Where epc indicates the number of training samples as a whole, and α indicates the decay rate, which is generally set to a value between 0.9 and 0.999.
And step 9: repeating the steps 8.1-8.6 until the student network Ns converges, and obtaining the final weight w of the teacher network Nt according to the formula (11)t。
Step 10: and normalizing the two multispectral images in different time phases, and inputting the normalized multispectral images into a trained teacher network to divide the classes of change and non-change.
The present invention is described in further detail below with reference to specific experimental data. The experimental data used in this experiment were multi-temporal IKNOS high-resolution image data in the Mina region of saudi arabia, with an image size of 700 × 950, using three bands B1, B2, and B3. The specific parameter settings in the deep learning training are as follows: (1) the number of batch samples was 128; (2) the number of iterative training times is 100; (3) when a pseudo training sample is constructed, setting the local window size psi of the sample to be 7; (4) the learning rate is set to 0.01.
To verify the effectiveness of the present invention, the change detection method of the present invention was compared to the following change detection methods:
(1) CVA-based EM-MRF method (EM-MRF) [ detection methods mentioned in the article "Automatics analysis of differential image for unsupervised change detection" (EETransmission on Geoscience and Remote Sensing,2000,38(3):1171-1182 ], by Bruzzone L. et al, Italy ].
(2) Deep belief network detection method (DBN) proposed by Gong et al [ Maoguo Gong et al, detection methods proposed in the article "Change detection in synthetic architecture radiation images based on deep neural networks. (IEEE Transactions on Geoscience and remove Sensing,2016,27(1):125-
(3) A method for detecting a challenge Network (GAN) [ Maoguo Gong et al, in the article "AGENTATIVE DISCRIMINATION CLASSIFIED NETWORK FOR CHANGE DETECTION IN MULTI-SPECTRAL IMAGERY ] (IEEE Journal of Selected Topics in Applied Earth Observation and removal Sensing,2019,12(1):321-
(4) The method of the invention.
The detection performance is measured by four indexes of error detection number FP, missing detection number FN, total error number OE and Kappa coefficient. The closer FP, FN and OE are to 0 and the closer Kappa coefficient k is to 1, indicating the better performance of the change detection method. The results are shown in Table 1.
TABLE 1 comparison of multi-temporal IKONOS image change detection results in Mina area
Method of producing a composite material FP FN OE k
EM-MRF 13373 1367 14740 0.821
DBN 1944 2214 4158 0.942
GAN 1719 3997 5716 0.918
The method of the invention 1611 2314 3925 0.943
Ideal for 0 0 0 1
As can be seen from Table 1, the Kappa coefficient k of the detection method provided by the invention is maximum and is closer to 1 than the other three detection algorithms. In addition, the total error number OE of the present invention is the smallest in the comparison algorithm, closer to 0. In conclusion, the performance of the change detection algorithm of the invention is superior to that of the other three detection methods, which shows that the change detection method provided by the invention is effective.
Fig. 2 is a time-phase multispectral IKONOS image of the Mina region, fig. 3 is a time-phase multispectral IKONOS image of the Mina region, and fig. 4 is a reference image of change detection. Fig. 5 is a change detection result of the EM-MRF algorithm, fig. 6 is a change detection result of the DBN algorithm, fig. 7 is a change detection result of the GAN algorithm, and fig. 8 is a change detection result of the method according to the embodiment of the present invention. From the comparison between the reference diagram of fig. 4 and fig. 5 to 8, the detection effect of the algorithm of the present invention is the best in visual effect.
As will be appreciated by one skilled in the art, embodiments of the present application may be provided as a method, system, or computer program product. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and the like) having computer-usable program code embodied therein.
The present application is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the application. It will be understood that each flow and/or block of the flow diagrams and/or block diagrams, and combinations of flows and/or blocks in the flow diagrams and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
The above description is only a preferred embodiment of the present invention, and it should be noted that, for those skilled in the art, several modifications and variations can be made without departing from the technical principle of the present invention, and these modifications and variations should also be regarded as the protection scope of the present invention.

Claims (7)

1. A multispectral remote sensing image change detection method based on deep learning is characterized by comprising the following steps:
a. carrying out image registration on the remote sensing image, carrying out radiation correction by using a multivariate change detection method, and then calculating the change vector amplitude of the remote sensing image;
b. acquiring a pseudo training sample set by utilizing a maximum expectation algorithm according to the change vector amplitude;
c. two identical fully-connected deep learning networks were constructed: a teacher network and a student network;
d. training a teacher network and a student network through a pseudo training sample set;
e. and acquiring a final change detection result according to the teacher network.
2. The deep learning-based multispectral remote sensing image change detection method according to claim 1, wherein the image registration comprises geometric coarse correction and geometric fine correction.
3. The deep learning-based multispectral remote sensing image change detection method as claimed in claim 1, wherein the number of neurons in a first layer of the teacher network and the student network is 294, the number of neurons in a second layer of the teacher network is 64, the number of neurons in a third layer of the teacher network is 128, the number of neurons in a fourth layer of the teacher network is 2, and the input dimension and the output dimension of the teacher network and the student network are 294 and 2, respectively.
4. The method for detecting the change of the multispectral remote sensing image based on the deep learning as claimed in claim 3, wherein the teacher network and the student network take samples in a 7 x 7 square window as input, and the multispectral image of each sample in two time phases has 6 wave bands.
5. The deep learning based multispectral remote sensing image change detection method as claimed in claim 3, wherein the teacher network and the student network use batch normalization processing except for output layer, and the middle layer uses LeakyReLU (x) activation function:
LeakyReLU(x1)＝max(0,x1)+negative_slope*min(0,x1) (7)
in the formula, x1Is the output matrix of the output intermediate layer of the network and negative slope represents the angle controlling the negative slope.
6. The method for detecting the change of the multispectral remote sensing image based on the deep learning as claimed in claim 1, wherein the step d comprises the following steps:
da. Labeled data set of pseudo training sample set
Figure FDA0002255290390000021
And
Figure FDA0002255290390000022
inputting student network Ns and calculating loss function L of labeled data set1：
Wherein x represents the spectrum normalization characteristic value of the corresponding sample point, y is the label corresponding to the sample, and wsIs the weight of the student network, f represents the probabilistic output of the student network Ns,
Figure FDA00022552903900000211
represents a cross entropy loss function, which is defined as follows:
in the formula, k represents a tag in change detection, the value is {0, 1}, 0 represents no change, 1 represents a change, and w represents the weight of the network.
db. Unlabeled data set of pseudo training sample set
Figure FDA0002255290390000025
Input into student network Ns to obtain
Figure FDA0002255290390000026
Corresponding tag output
Figure FDA0002255290390000027
dc. Unlabeled data set of pseudo training sample set
Figure FDA0002255290390000028
Input to teacher network Nt to obtain
Figure FDA0002255290390000029
Corresponding output
Figure FDA00022552903900000210
g represents the probability output, w, of the teacher network NttA weight representing the teacher network Nt;
dd. Aiming at the unlabeled data set, expecting that the student network Ns is consistent with the predicted labels of the teacher network Nt, and constructing an unlabeled loss function L2：
L2＝||f(x；ws)-g(x；wt)||2(10)
de. According to the total loss function L1+L2Updating the weight w of the student network Ns by adopting a back propagation algorithm and a random gradient descent algorithms。
df. Updating parameter w of teacher network Nt by using parameter of student network Nst：
Figure FDA0002255290390000031
In the formula, epc represents the number of times of training of the whole training sample, and α represents the decay rate;
dg. Repeating the steps da to df until the student network Ns convergesAnd obtaining the final weight w of the teacher network Nt according to the formula (11)t。
7. A multispectral remote sensing image change detection system based on deep learning is characterized by comprising a processor and a storage device, wherein a plurality of instructions are stored in the storage device, and the instructions are used for the processor to load and execute the steps of the method according to any one of claims 1 to 6.
CN201911050821.5A 2019-10-31 2019-10-31 Multispectral remote sensing image change detection method and system based on deep learning Withdrawn CN110826458A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201911050821.5A CN110826458A (en) 2019-10-31 2019-10-31 Multispectral remote sensing image change detection method and system based on deep learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201911050821.5A CN110826458A (en) 2019-10-31 2019-10-31 Multispectral remote sensing image change detection method and system based on deep learning

Publications (1)

Publication Number Publication Date
CN110826458A true CN110826458A (en) 2020-02-21

Family

ID=69551760

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201911050821.5A Withdrawn CN110826458A (en) 2019-10-31 2019-10-31 Multispectral remote sensing image change detection method and system based on deep learning

Country Status (1)

Country Link
CN (1) CN110826458A (en)

Cited By (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111797703A (en) * 2020-06-11 2020-10-20 武汉大学 Multi-source remote sensing image classification method based on robust deep semantic segmentation network
CN111959495A (en) * 2020-06-29 2020-11-20 北京百度网讯科技有限公司 Vehicle control method and device and vehicle
CN112465884A (en) * 2020-11-25 2021-03-09 西北工业大学 Multi-element remote sensing image change detection method based on generating type feature representation network
CN112926624A (en) * 2021-01-23 2021-06-08 西北工业大学 Robust multi-temporal multispectral image change detection method
CN113255451A (en) * 2021-04-25 2021-08-13 西北工业大学 Method and device for detecting change of remote sensing image, electronic equipment and storage medium
CN113326764A (en) * 2021-05-27 2021-08-31 北京百度网讯科技有限公司 Method and device for training image recognition model and image recognition
CN113705646A (en) * 2021-08-18 2021-11-26 西安交通大学 Radio frequency fine characteristic information extraction method and system based on semi-supervised meta learning
CN113869464A (en) * 2021-12-02 2021-12-31 深圳佑驾创新科技有限公司 Training method of image classification model and image classification method
CN115019180A (en) * 2022-07-28 2022-09-06 北京卫星信息工程研究所 SAR image ship target detection method, electronic device and storage medium
CN115761518A (en) * 2023-01-10 2023-03-07 云南瀚哲科技有限公司 Crop classification method based on remote sensing image data
CN115937568A (en) * 2022-09-29 2023-04-07 中国地质大学（北京） Basalt structure background classification method, system and device and storage medium
CN117830324A (en) * 2023-12-28 2024-04-05 深圳大学 3D medical image segmentation method based on multi-dimensional and global local combination
US12027162B2 (en) 2021-03-03 2024-07-02 Google Llc Noisy student teacher training for robust keyword spotting

Cited By (22)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111797703A (en) * 2020-06-11 2020-10-20 武汉大学 Multi-source remote sensing image classification method based on robust deep semantic segmentation network
CN111959495A (en) * 2020-06-29 2020-11-20 北京百度网讯科技有限公司 Vehicle control method and device and vehicle
CN111959495B (en) * 2020-06-29 2021-11-12 阿波罗智能技术(北京)有限公司 Vehicle control method and device and vehicle
CN112465884A (en) * 2020-11-25 2021-03-09 西北工业大学 Multi-element remote sensing image change detection method based on generating type feature representation network
CN112465884B (en) * 2020-11-25 2024-01-16 西北工业大学 Multi-element remote sensing image change detection method based on generated characteristic representation network
CN112926624A (en) * 2021-01-23 2021-06-08 西北工业大学 Robust multi-temporal multispectral image change detection method
CN112926624B (en) * 2021-01-23 2024-05-24 西北工业大学 Robust multi-temporal multi-spectral image change detection method
US12027162B2 (en) 2021-03-03 2024-07-02 Google Llc Noisy student teacher training for robust keyword spotting
CN113255451B (en) * 2021-04-25 2023-04-07 西北工业大学 Method and device for detecting change of remote sensing image, electronic equipment and storage medium
CN113255451A (en) * 2021-04-25 2021-08-13 西北工业大学 Method and device for detecting change of remote sensing image, electronic equipment and storage medium
JP7331171B2 (en) 2021-05-27 2023-08-22 ベイジン バイドゥ ネットコム サイエンス テクノロジー カンパニー リミテッド Methods and apparatus for training image recognition models, methods and apparatus for recognizing images, electronic devices, storage media, and computer programs
JP2022058915A (en) * 2021-05-27 2022-04-12 ベイジン バイドゥ ネットコム サイエンス テクノロジー カンパニー リミテッド Method and device for training image recognition model, method and device for recognizing image, electronic device, storage medium, and computer program
CN113326764A (en) * 2021-05-27 2021-08-31 北京百度网讯科技有限公司 Method and device for training image recognition model and image recognition
CN113705646B (en) * 2021-08-18 2024-05-10 西安交通大学 Radio frequency fine feature information extraction method and system based on semi-supervised meta learning
CN113705646A (en) * 2021-08-18 2021-11-26 西安交通大学 Radio frequency fine characteristic information extraction method and system based on semi-supervised meta learning
CN113869464A (en) * 2021-12-02 2021-12-31 深圳佑驾创新科技有限公司 Training method of image classification model and image classification method
CN115019180B (en) * 2022-07-28 2023-01-17 北京卫星信息工程研究所 SAR image ship target detection method, electronic device and storage medium
CN115019180A (en) * 2022-07-28 2022-09-06 北京卫星信息工程研究所 SAR image ship target detection method, electronic device and storage medium
CN115937568A (en) * 2022-09-29 2023-04-07 中国地质大学（北京） Basalt structure background classification method, system and device and storage medium
CN115937568B (en) * 2022-09-29 2024-05-07 中国地质大学（北京） Basalt structure background classification method, basalt structure background classification system, basalt structure background classification device and storage medium
CN115761518A (en) * 2023-01-10 2023-03-07 云南瀚哲科技有限公司 Crop classification method based on remote sensing image data
CN117830324A (en) * 2023-12-28 2024-04-05 深圳大学 3D medical image segmentation method based on multi-dimensional and global local combination

Similar Documents

Publication Publication Date Title
CN110826458A (en) Multispectral remote sensing image change detection method and system based on deep learning
Ding et al. Semi-supervised locality preserving dense graph neural network with ARMA filters and context-aware learning for hyperspectral image classification
CN109754078B (en) Method for optimizing a neural network
US10713563B2 (en) Object recognition using a convolutional neural network trained by principal component analysis and repeated spectral clustering
Ghosh et al. Integration of Gibbs Markov random field and Hopfield-type neural networks for unsupervised change detection in remotely sensed multitemporal images
CN108537102B (en) High-resolution SAR image classification method based on sparse features and conditional random field
CN110827330B (en) Time sequence integrated multispectral remote sensing image change detection method and system
Grana et al. Two lattice computing approaches for the unsupervised segmentation of hyperspectral images
CN109919241B (en) Hyperspectral unknown class target detection method based on probability model and deep learning
CN107392940B (en) SAR image change detection method based on stack semi-supervised self-adaptive denoising self-encoder
CN111986262B (en) Image area positioning method and device
CN110827332A (en) Registration method of SAR image based on convolutional neural network
CN107481235A (en) The multi-temporal remote sensing image change detecting method that a kind of mathematical morphology filter converts with reference to card side
Oliva et al. Multilevel thresholding by fuzzy type II sets using evolutionary algorithms
Carrara et al. On the robustness to adversarial examples of neural ode image classifiers
CN112634149A (en) Point cloud denoising method based on graph convolution network
CN109191503B (en) Remote sensing image change detection method and system based on conditional random field
CN109242832B (en) Multi-temporal multispectral remote sensing image change detection method and system
CN110837787B (en) Multispectral remote sensing image detection method and system for three-party generated countermeasure network
Syafeeza et al. Convolutional neural networks with fused layers applied to face recognition
CN112837320A (en) Remote sensing image semantic segmentation method based on parallel hole convolution
CN111160142B (en) Certificate bill positioning detection method based on numerical prediction regression model
CN116935122A (en) Image classification method and system based on 3D-WGMobileNet
CN116258877A (en) Land utilization scene similarity change detection method, device, medium and equipment
Hong et al. Random forest fusion classification of remote sensing PolSAR and optical image based on LASSO and IM factor

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
WW01 Invention patent application withdrawn after publication

Application publication date: 20200221