CN109993220A - Multi-source Remote Sensing Images Classification method based on two-way attention fused neural network - Google Patents

Multi-source Remote Sensing Images Classification method based on two-way attention fused neural network Download PDF

Info

Publication number
CN109993220A
CN109993220A CN201910224541.5A CN201910224541A CN109993220A CN 109993220 A CN109993220 A CN 109993220A CN 201910224541 A CN201910224541 A CN 201910224541A CN 109993220 A CN109993220 A CN 109993220A
Authority
CN
China
Prior art keywords
sample
lidar
feature
layer
hyperion
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN201910224541.5A
Other languages
Chinese (zh)
Other versions
CN109993220B (en
Inventor
张向荣
焦李成
梁婷
唐旭
李阳阳
古晶
侯彪
马文萍
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.)
Xidian University
Original Assignee
Xidian University
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 Xidian University filed Critical Xidian University
Priority to CN201910224541.5A priority Critical patent/CN109993220B/en
Publication of CN109993220A publication Critical patent/CN109993220A/en
Application granted granted Critical
Publication of CN109993220B publication Critical patent/CN109993220B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

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
    • G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
    • G06F18/24133—Distances to prototypes
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G06F18/253—Fusion techniques of extracted features

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)
  • Image Processing (AREA)

Abstract

The invention discloses a kind of multi-source Remote Sensing Images Classification method based on two-way attention fused neural network, mainly solve the problems, such as that the prior art is low to multi-source Remote Sensing Images Classification precision.Its implementation are as follows: 1) high-spectral data and laser radar data are pre-processed and divided, obtain training sample and test sample；2) attention fused layer of the design based on attention mechanism is weighted screening and fusion to spectroscopic data and laser radar data, and it constructs two-way and connects convolutional neural networks, 3) convolutional neural networks that connect are trained using multiclass cross entropy as loss function, obtain trained network model, 4) test sample is predicted using trained model, obtains final classification results.The present invention can extract the feature of multi- source Remote Sensing Data data and carry out effective integrated classification to it, improve the problem that dimension is excessively high in fusion, improve average nicety of grading, can be used for merging the remote sensing images that two width different sensors obtain.

Description

Multi-source Remote Sensing Images Classification method based on two-way attention fused neural network
Technical field
The invention belongs to technical field of image processing, in particular to a kind of Classifying Method in Remote Sensing Image can be used for two width The remote sensing images that different sensors obtain carry out integrated classification.
Background technique
In recent years, remote sensor quantity increases very rapid, and people can obtain the multi-source data of Same Scene simultaneously Collection, this different information for integrate different sensors capture.Such as multispectral image MSI or high spectrum image HSI is usually made of multiple spectrum channels of Same Scene, is included detailed spectrum and spatial information, is provided and accurately distinguish sense The ability of interesting material.On the other hand, laser radar LiDAR data can elevation to scene and object height information carry out table Show, helps to distinguish the object for being made of that but height is different analog material.Therefore, the integrated of both sensor informations can mention For detailed spectral-spatial-elevation information, to further improve classification performance.
The classification of remote sensing images is usually complicated and challenging, in order to make full use of high spectrum image and LiDAR Data, current integrated classification method probably have following several.A kind of method is using laser radar data as a channel heap It is laminated in high spectrum image, and classifies to the image of heap poststack, this method, can due to the identification feature of laser radar Classification performance is improved, however, the connection of a simple primary characteristic may be not enough to separate classification interested.Second method It is that traditional feature extraction is carried out to high spectrum image and LiDAR image respectively first, such as morphological feature, wavelet character, line Feature etc. is managed, support vector machines, feature of the traditional classifiers such as random forest RF to the two kinds of images extracted are then used Classify.
Limited capacity of the traditional classification method of both the above in feature learning and classification, firstly, traditional characteristic may It will lead to the underutilization to spatial information, in addition, the feature quantity that different sensors are extracted is more, although can be more detailed The information of two kinds of images is characterized, but will cause serious dimension disaster simultaneously.In recent years, the method quilt of deep learning Feature is extracted in proposition in a hierarchical fashion, provides direction for the multi-source Remote Sensing Image Fusion based on depth characteristic.Convolutional Neural Network C NN simulates " local field of view " concept in human visual system, and full connection is converted to local connection, is connected using part Processing space dependence is connect, the number of parameters for needing training is significantly reduced, calculating cost is reduced, in addition, convolutional Neural Network has the ability for learning layering and autonomous learning abundant, can extract according to different data sources is adaptive Suitable feature, therefore be adapted to carry out the integrated classification of high spectrum image and LiDAR image.
The current integrated classification method based on convolutional neural networks, usually step is: firstly, respectively to EO-1 hyperion number Depth characteristic extraction is carried out according to convolutional neural networks similar with two structures of LiDAR data building, then mentions two networks The feature of the two kinds of images taken carries out dimension variation, and cascade becomes one group of feature, and full articulamentum is finally recycled to carry out image Classification.Such methods can utilize the advantages of convolutional neural networks, effectively special to high spectrum image and LiDAR image zooming-out Sign, and these features are applied to classification problem, still, this model is when carrying out integrated classification, however it remains centainly not Foot place: high spectrum image network and LiDAR data network first is separated from each other, and cannot carry out the circulation of characteristic information well And fusion；Secondly, the feature for two kinds of images that convolutional neural networks extract is only input after simple feature cascade Classify to full articulamentum, there is no the screening and fusion that carry out feature, this will lead to feature dimension is excessively high and feature The generation of redundancy phenomena.
Summary of the invention
It is a kind of based on two-way attention fusion nerve it is an object of the invention in view of the above shortcomings of the prior art, propose The multi-source Remote Sensing Image Fusion classification method of network is reinforced circulation between data, is mentioned to reduce the degree of redundancy of characteristic information The nicety of grading of high multi-source Remote Sensing Images.
Technical thought of the invention is: the interactive connection convolutional neural networks by establishing a two-way, respectively to bloom Modal data and LiDAR data carry out feature extraction；By designing a fused layer based on attention mechanism, to by convolution mind The feature of high-spectral data and LiDAR data after network carries out study screening and fusion, and obtained fusion feature is connected It is connected to full articulamentum and carries out classification output, implementation includes the following:
(1) input the panel height spectrum picture and a width laser radar LiDAR image that have been registrated, building training sample set and Test sample collection:
(1a) inputs high-spectrum remote sensing, has in the image comprising M and marks pixel and N number of unmarked pixel, Each pixel of image is a sample, and M marked sample constitutes EO-1 hyperion training sample set H, N number of unmarked sample Constitute EO-1 hyperion test sample collection He；
(1b) inputs laser radar LiDAR image, wherein has and marks the number of pixel for M, and coordinate and high-spectrum There is the coordinate of label pixel identical as in, the number of unmarked pixel is N, and unmarked picture in coordinate and high spectrum image The coordinate of vegetarian refreshments is identical, each pixel of image is a sample, and M marked sample constitutes LiDAR training sample set R, N number of unmarked sample constitute LiDAR test sample collection Re.
(2) be based on attention mechanism method, design attention Fusion Features layer, for EO-1 hyperion sample characteristics and LiDAR sample characteristics carry out the weighting study screening and fusion of selectivity；
(3) building two-way connects convolutional neural networks:
(3a) uses dual input method, and two of the convolutional neural networks inputs that connect of design two-way be respectively size are W The EO-1 hyperion sample and size of × W × D is the LiDAR sample of W × W；
(3b) separately designs branch's convolutional neural networks that two-way is directed to EO-1 hyperion sample set and LiDAR sample set, passes through Interconnection layers in two-way branch neural network carry out the circulation fusion of feature, wherein EO-1 hyperion branching networks are rolled up by three Lamination and three pond layer compositions, LiDAR branching networks are by three convolutional layers, three pond layers and two interconnection layers groups At；
(3c) is sequentially connected an attention fused layer after each branch's convolutional neural networks, two full articulamentums and One output layer completes two-way and connects the buildings of convolutional neural networks, and the connect output of convolutional neural networks of the two-way is The generic of each sample；
(4) adaptive moments estimation optimization algorithm is used, the true of multi-class cross entropy loss function and training sample is passed through Classification p (H, R), the two-way designed in training step (3) connect convolutional neural networks:
EO-1 hyperion training sample set H and LiDAR training sample set R is separately input to two-way and connected convolutional Neural by (4a) In network, after two-way branching networks, EO-1 hyperion sample characteristics X is obtained1With LiDAR sample characteristics X2, by both features Attention fused layer is input to after progress dimension transformation to be screened and merged, and fusion feature X is obtained,
(4b) sequentially inputs fusion feature X to full articulamentum and output layer, obtains the affiliated pre- of each training sample Survey classification；
(4c) repeats (4a)-(4b) step, until the two-way convolutional neural networks that connect reach convergence, obtains comprising training Good two-way connects convolutional neural networks model；
(5) classification prediction is carried out to test sample:
EO-1 hyperion test sample collection He and LiDAR test sample collection Re is input to trained two-way convolutional neural networks In model, the generic of each test sample is obtained.
Compared with prior art, the present invention having the advantage that
First, the present invention is devised when carrying out high-spectral data and LiDAR data feature extraction containing interactive connection The feature of high-spectral data is circulated in the feature of LiDAR data by two-way branch convolutional neural networks, this makes LiDAR special The not simple elevation information comprising scene in sign, information is more abundant, improves two kinds of remotely-sensed data phases in existing method Mutually the problem of separation, so that the integrated classification result of multi-source Remote Sensing Images is more accurate.
Second, the present invention to EO-1 hyperion branch neural network network and LiDAR branch neural network extraction feature into When row fusion, the Fusion Features layer based on attention mechanism is devised, is carried out using feature of the attention method to two kinds of images The weight of weighting screening and fusion, fused layer is obtained by interactive connection convolutional neural networks by training, is effectively schemed to two kinds The feature of picture is screened, and unessential feature is filtered out, and is assigned important feature biggish weight proportion, is reduced feature Redundancy and dimension, improve remote sensing image fusion classification precision.
Detailed description of the invention
Fig. 1 is realization general flow chart of the invention；
Fig. 2 is remote sensing scene image and true tag used in the present invention emulates；
Fig. 3 is with the simulation result diagram of the invention classified to image in Fig. 2 and result enlarged drawing；
Fig. 4 is to the simulation result diagram and result enlarged drawing classified with the prior art to image in Fig. 2.
Specific embodiment
Specific example and effect of the invention are described in further detail below in conjunction with attached drawing.
Referring to Fig.1, steps are as follows for the realization of this example:
Step 1, the panel height spectrum picture and a width laser radar LiDAR image being registrated are inputted, training sample is constructed Collection and test sample collection.
High-spectrum remote sensing 1a) is inputted, has in the image comprising M and marks pixel and N number of unmarked pixel, figure Each pixel of picture is a sample, and M marked sample constitutes EO-1 hyperion training sample set H, N number of unmarked sample structure At EO-1 hyperion test sample collection He, in this example, marked sample number M is 66485, and unmarked sample number N is 598360；
1b) input laser radar LiDAR image, wherein have and mark the number of pixel for M, and coordinate and high-spectrum There is the coordinate of label pixel identical as in, the number of unmarked pixel is N, and unmarked picture in coordinate and high spectrum image The coordinate of vegetarian refreshments is identical, each pixel of image is a sample, and M marked sample constitutes LiDAR training sample set R, N number of unmarked sample constitute LiDAR test sample collection Re, and in this example, marked sample number M is 66485, unmarked Number of samples N is 598360；
Step 2, building two-way connects convolutional neural networks.
2a) use dual input method, design two-way connect two of convolutional neural networks inputs be respectively size be W × The EO-1 hyperion sample and size of W × D is the LiDAR sample of W × W, and in this example, two-way connects the two of convolutional neural networks Kind input sample size is respectively 41 × 41 × 3 and 41 × 41；
2b) separately design branch's convolutional neural networks that two-way is directed to EO-1 hyperion sample set and LiDAR sample set:
2b1) design is directed to branch's convolutional neural networks of EO-1 hyperion sample set, totally six layers of the network, is followed successively by the first volume The-the first pond of lamination the-the second pond of the-the second convolutional layer of layer layer-third convolutional layer-third pond layer, wherein each convolutional layer is equal Feature extraction is carried out to EO-1 hyperion sample, after each pond layer carries out dimension transformation to the feature that a upper convolutional layer extracts, Obtain bloom spectrum signature S；
2b2) design is directed to branch's convolutional neural networks of LiDAR sample set, totally eight layers of the network, is followed successively by the 1st convolution - the 1 the-the 3 pond of pond layer the-the 1 interconnection layers the-the 2 convolutional layer the-the 2 pond layer the-the 3 convolutional layer of the-the 2 interconnection layers of layer Change layer, wherein each convolutional layer carries out feature extraction to LiDAR sample, and each pond layer extracts a upper convolutional layer After LiDAR feature carries out dimension transformation, obtain LiDAR feature L, interconnection layers by bloom spectrum signature S and LiDAR feature L into Row connection, obtains LiDAR connection features F；
The circulation fusion of feature is carried out by the interconnection layers in above-mentioned two-way branch neural network.
It 2c) is based on attention mechanism method, designs attention Fusion Features layer:
The feature and LiDAR training sample for 2c1) obtaining EO-1 hyperion training sample by branching networks are by branching networks Obtained feature is attached, and expression formula is as follows:
xi=concat (x1i,x2i)
Wherein, i=1,2..., M, M are training sample sum, x1iIndicate that i-th of EO-1 hyperion training sample is rolled up by branch The feature obtained after product network, x2iIndicate the feature that i-th of LiDAR training sample obtains after branch's convolutional network, xiTable Show the feature vector after i-th of EO-1 hyperion training sample feature and i-th of LiDAR training sample feature are attached；
2c2) initialize a weight matrix and a bias vector, by weight matrix with connect after feature vector xiPhase Multiply, then tangent function calculating is added and carried out with bias vector, obtains the tangent vector e by tangent function operationi:
ei=tanh (xiW+b),
Wherein, w and b respectively indicates the weight matrix and bias vector of attention fused layer；
2c3) to tangent vector eiExponential function calculating is normalized, obtains normalized vector ai:
ai=softmax (ei)
Wherein, softmax indicates normalization exponential function；
2c4) by the feature vector x by connectioniWith normalized vector aiTransposition be multiplied, obtain the fusion of i-th of sample Feature fi=ai Txi, wherein T indicates transposition；
2c1-2c4 2c5) is respectively repeated steps to M training sample, obtains the fusion feature of all samples, is constituted complete Attention fused layer；
It 2d) is sequentially connected two full articulamentums and an output layer after attention fused layer, completes two-way and connect volume Product neural network building, the two-way connect convolutional neural networks output be each sample generic, in this example In, classification sum belonging to sample is 15.
Step 3, using adaptive moments estimation optimization algorithm, pass through the true of multi-class cross entropy loss function and training sample Real classification p (H, R), training two-way connect convolutional neural networks.
EO-1 hyperion training sample set H and LiDAR training sample set R 3a) is separately input to two-way to connect convolutional Neural In network, after two-way branching networks, EO-1 hyperion sample characteristics X is obtained1With LiDAR sample characteristics X2, by both features Attention fused layer is input to after progress dimension transformation to be screened and merged, and fusion feature X is obtained；
3b) fusion feature X is sequentially input to full articulamentum and output layer, obtains the prediction classification q of each training sample (H, R), using multi-class cross entropy loss function Y (p, q)=- ∑ p (H, R) log (q (H, R)) to the true class of training sample Other p (H, R) and prediction classification q (H, R) are calculated, and are obtained prediction the distance between classification and true classification Y (p, q), are passed through The distance carries out backpropagation to the convolutional neural networks that connect；
3c) repeat step 3a)-step 3b), until the two-way convolutional neural networks that connect reach convergence, trained Two-way connect convolutional neural networks model, in this example, step 3a)-step 3b) is repeated 200 times.
Step 4, classification prediction is carried out to test sample.
EO-1 hyperion test sample collection and LiDAR test sample collection are input to trained two-way convolutional neural networks model In, obtain the generic of each test sample.
Below in conjunction with emulation experiment, technical effect of the invention is described further:
1. simulated conditions
In this emulation experiment, used data are the high-spectrum remote-sensing number of University of Houston campus and its surrounding cities According to and LiDAR data, as shown in Figure 2, wherein Fig. 2 (a) is the remote sensing images that use of emulation, and Fig. 2 (b) is the remote sensing images True classification, the size of data are 349 × 1905, and high-spectral data includes 144 spectrum segments, shares 15 kinds of different ground species Not；
Emulation platform are as follows: dominant frequency is the memory of Intel Core the i5-6500 CPU, 8GB of 3.20GHz, windows10 Operating system, Keras and TensorFlow deep learning platform, Python.
2. emulation content and interpretation of result
Emulation 1, emulates remote sensing images using the present invention, as a result as shown in figure 3, wherein Fig. 3 (a) is with the present invention To the simulation result diagram of remote sensing images, Fig. 3 (b)-Fig. 3 (d)) it is amplification detail view to Fig. 3 (a)；
Emulation 2, emulates remote sensing images using the prior art, as a result as shown in figure 4, wherein Fig. 4 (a) is with existing For technology to the simulation result diagram of remote sensing images, Fig. 4 (b)-Fig. 4 (d) is the amplification detail view of Fig. 4 (a)；
From the comparison of Fig. 2 (b) and Fig. 3 (a) as can be seen that simulation result diagram of the invention has been reached substantially with true tag Unanimously.
It is more accurate in terms of details of classifying to can be seen that the present invention from the comparison of Fig. 3 and Fig. 4, and the prior art exists The sample of mistake point, the present invention are promoted by 92% to 98.5% compared to existing method, the average nicety of grading of simulation result；
Above-mentioned simulation result shows the present invention using the two-way convolutional neural networks of interactive connection and based on attention mechanism Fusion Features strategy, the information of high spectrum image and LiDAR image is circulated, study is carried out to its feature and screening and is melted It closes, eventually for remote sensing terrain classification, integrated classification result is more accurate complete, and average nicety of grading is higher.

Claims (4)

1. a kind of multi-source Remote Sensing Images Classification method based on two-way attention fused neural network, includes the following:
(1) the panel height spectrum picture and a width laser radar LiDAR image being registrated are inputted, training sample set and test are constructed Sample set:
(1a) inputs high-spectrum remote sensing, has in the image comprising M and marks pixel and N number of unmarked pixel, image Each pixel be a sample, M marked sample constitutes EO-1 hyperion training sample set H, N number of unmarked sample composition EO-1 hyperion test sample collection He；
(1b) inputs laser radar LiDAR image, wherein the number for having label pixel is M, and in coordinate and high spectrum image There is the coordinate of label pixel identical, the number of unmarked pixel is N, and unmarked pixel in coordinate and high spectrum image Coordinate it is identical, each pixel of image is a sample, M marked sample composition LiDAR training sample set R, N number of Unmarked sample constitutes LiDAR test sample collection Re.
(2) it is based on attention mechanism method, designs attention Fusion Features layer, for EO-1 hyperion sample characteristics and LiDAR sample Eigen carries out the weighting study screening and fusion of selectivity；
(3) building two-way connects convolutional neural networks:
(3a) uses dual input method, design two-way connect two of convolutional neural networks inputs be respectively EO-1 hyperion sample and LiDAR sample；
(3b) separately designs branch's convolutional neural networks that two-way is directed to EO-1 hyperion sample set and LiDAR sample set, passes through two-way Interconnection layers in branch's neural network carry out the circulation fusion of feature, wherein EO-1 hyperion branching networks are by three convolutional layers It is formed with three pond layers, LiDAR branching networks are made of three convolutional layers, three pond layers and two interconnection layers；
(3c) is sequentially connected an attention fused layer after each branch's convolutional neural networks, two full articulamentums and one Output layer completes two-way and connects the buildings of convolutional neural networks, and the connect output of convolutional neural networks of the two-way is each The generic of sample；
(4) use adaptive moments estimation optimization algorithm, by the true p of multi-class cross entropy loss function and training sample (H, R), the two-way designed in training step (3) connects convolutional neural networks:
EO-1 hyperion training sample set H and LiDAR training sample set R is separately input to two-way and connected convolutional neural networks by (4a) In, after two-way branching networks, obtain EO-1 hyperion sample characteristics X1With LiDAR sample characteristics X2, both features are carried out Attention fused layer is input to after dimension transformation to be screened and merged, and fusion feature X is obtained,
(4b) sequentially inputs fusion feature X to full articulamentum and output layer, obtains the affiliated prediction class of each training sample Not；
(4c) repeats (4a)-(4b) step, until the two-way convolutional neural networks that connect reach convergence, obtains including trained Two-way connects convolutional neural networks model；
(5) classification prediction is carried out to test sample:
EO-1 hyperion test sample collection He and LiDAR test sample collection Re is input to trained two-way convolutional neural networks model In, obtain the generic of each test sample.
2. design attention is special the method according to claim 1, wherein being based on attention mechanism method in (2) Fused layer is levied, is accomplished by
The feature and laser radar LiDAR training sample that (2a) obtains EO-1 hyperion training sample by branching networks are by branch The feature that network obtains is attached, and expression formula is as follows:
xi=concat (x1i,x2i)
Wherein, x1iIndicate the feature that i-th of EO-1 hyperion training sample obtains after branch's convolutional network, x2iIt indicates i-th The feature that LiDAR training sample obtains after branch's convolutional network, xiIndicate i-th of EO-1 hyperion training sample feature and i-th A LiDAR training sample feature be attached after feature vector, i=1,2..., M, M be training sample sum；
(2b) initializes a weight matrix and a bias vector, and weight matrix is multiplied with the feature vector after connecting, then Tangent function calculating is added and carried out with bias vector, obtains the tangent vector e by tangent function operationi:
ei=tanh (xiW+b),
Wherein, w and b respectively indicates the weight matrix and bias vector of attention fused layer；
The tangent vector e that (2c) obtains (2b)iExponential function calculating is normalized, obtains normalized vector ai:
ai=softmax (ei)
Wherein, softmax indicates normalization exponential function；
(2d) will pass through the feature x of connection in (2a)iWith normalized vector a obtained in (2c)iTransposition be multiplied, obtain i-th The fusion feature f of samplei=ai Txi, wherein T indicates transposition；
(2e) repeats (2a)-(2d) step to M training sample respectively, obtains the fusion feature of all samples, constitutes complete Attention fused layer.
3. method according to claim 1, which is characterized in that in step (3b) design two-way be directed to respectively EO-1 hyperion sample and Branch's convolutional neural networks of LiDAR sample, are accomplished by
(3b1) design is directed to branch's convolutional neural networks of EO-1 hyperion sample set, totally six layers of the network, is followed successively by the first convolution - the first the-the second pond of the-the second convolutional layer of pond layer layer-third convolutional layer-third pond layer of layer, wherein each convolutional layer is right EO-1 hyperion sample carries out feature extraction, after each pond layer carries out dimension transformation to the feature that its upper convolutional layer extracts, Obtain bloom spectrum signature S；
(3b2) design is directed to branch's convolutional neural networks of LiDAR sample set, totally eight layers of the network, is followed successively by the 1st convolutional layer- 1st the-the 3 pond of pond layer the-the 1 interconnection layers the-the 2 convolutional layer the-the 2 pond layer the-the 3 convolutional layer of the-the 2 interconnection layers Layer, wherein each convolutional layer carries out feature extraction to LiDAR sample, and each pond layer extracts its upper convolutional layer After LiDAR feature carries out dimension transformation, obtain LiDAR feature L, interconnection layers by bloom spectrum signature S and LiDAR feature L into Row connection, obtains LiDAR connection features F.
4. method according to claim 1, which is characterized in that the multi-class cross entropy loss function in step (4), expression Formula is as follows:
Y (p, q)=- ∑ p (H, R) log (q (H, R)),
Wherein, H and R respectively indicates EO-1 hyperion training sample and LiDAR training sample, and p (H, R) indicates the true class of training sample Not, q (H, R) indicates the prediction classification of training sample, and Y (p, q) is indicated between the true classification of training sample and prediction classification Distance.
CN201910224541.5A 2019-03-23 2019-03-23 Multi-source remote sensing image classification method based on double-path attention fusion neural network Active CN109993220B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910224541.5A CN109993220B (en) 2019-03-23 2019-03-23 Multi-source remote sensing image classification method based on double-path attention fusion neural network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910224541.5A CN109993220B (en) 2019-03-23 2019-03-23 Multi-source remote sensing image classification method based on double-path attention fusion neural network

Publications (2)

Publication Number Publication Date
CN109993220A true CN109993220A (en) 2019-07-09
CN109993220B CN109993220B (en) 2022-12-06

Family

ID=67130950

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910224541.5A Active CN109993220B (en) 2019-03-23 2019-03-23 Multi-source remote sensing image classification method based on double-path attention fusion neural network

Country Status (1)

Country Link
CN (1) CN109993220B (en)

Cited By (31)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110503191A (en) * 2019-08-27 2019-11-26 四川博文讯通科技有限公司 A kind of multilayer neural network model towards video analysis
CN110516596A (en) * 2019-08-27 2019-11-29 西安电子科技大学 Empty spectrum attention hyperspectral image classification method based on Octave convolution
CN110728224A (en) * 2019-10-08 2020-01-24 西安电子科技大学 Remote sensing image classification method based on attention mechanism depth Contourlet network
CN110866494A (en) * 2019-11-14 2020-03-06 三亚中科遥感研究所 Optical remote sensing image-based town group extraction method and system
CN111291622A (en) * 2020-01-16 2020-06-16 武汉汉达瑞科技有限公司 Method and device for detecting building change in remote sensing image
CN111563520A (en) * 2020-01-16 2020-08-21 北京航空航天大学 Hyperspectral image classification method based on space-spectrum combined attention mechanism
CN111666313A (en) * 2020-05-25 2020-09-15 中科星图股份有限公司 Correlation construction and multi-user data matching method based on multi-source heterogeneous remote sensing data
CN111667489A (en) * 2020-04-30 2020-09-15 华东师范大学 Cancer hyperspectral image segmentation method and system based on double-branch attention deep learning
CN111738052A (en) * 2020-04-13 2020-10-02 上海海洋大学 Multi-feature fusion hyperspectral remote sensing ground object classification method based on deep learning
CN111898662A (en) * 2020-07-20 2020-11-06 北京理工大学 Coastal wetland deep learning classification method, device, equipment and storage medium
CN112052758A (en) * 2020-08-25 2020-12-08 西安电子科技大学 Hyperspectral image classification method based on attention mechanism and recurrent neural network
CN112116563A (en) * 2020-08-28 2020-12-22 南京理工大学 Hyperspectral image target detection method and system based on spectral dimension and space cooperation neighborhood attention
CN112232343A (en) * 2020-09-03 2021-01-15 国家粮食和物资储备局科学研究院 Neural network and method for recognizing grain mildewed grains
CN112560733A (en) * 2020-12-23 2021-03-26 上海交通大学 Multitasking system and method for two-stage remote sensing image
CN112560967A (en) * 2020-12-18 2021-03-26 西安电子科技大学 Multi-source remote sensing image classification method, storage medium and computing device
CN112861978A (en) * 2021-02-20 2021-05-28 齐齐哈尔大学 Multi-branch feature fusion remote sensing scene image classification method based on attention mechanism
CN112949699A (en) * 2021-02-09 2021-06-11 国家卫星气象中心(国家空间天气监测预警中心) Remote sensing image classification model establishing and verifying method, system and electronic equipment
CN113077051A (en) * 2021-04-14 2021-07-06 广东博智林机器人有限公司 Network model training method and device, text classification model and network model
CN113095409A (en) * 2021-04-13 2021-07-09 西安电子科技大学 Hyperspectral image classification method based on attention mechanism and weight sharing
CN113222044A (en) * 2021-05-25 2021-08-06 合肥工业大学 Cervical fluid-based cell classification method based on ternary attention and scale correlation fusion
CN113283529A (en) * 2021-06-08 2021-08-20 南通大学 Neural network construction method for multi-modal image visibility detection
CN113435253A (en) * 2021-05-31 2021-09-24 西安电子科技大学 Multi-source image combined urban area ground surface coverage classification method
CN113627590A (en) * 2021-07-29 2021-11-09 中汽创智科技有限公司 Attention module and attention mechanism of convolutional neural network and convolutional neural network
CN113705718A (en) * 2021-09-06 2021-11-26 齐齐哈尔大学 Remote sensing scene image classification method based on multi-level feature dense fusion
CN113780147A (en) * 2021-09-06 2021-12-10 西安电子科技大学 Lightweight hyperspectral ground object classification method and system with dynamic fusion convolution network
CN114119444A (en) * 2021-11-29 2022-03-01 武汉大学 Multi-source remote sensing image fusion method based on deep neural network
CN114399684A (en) * 2022-03-24 2022-04-26 中国科学院西安光学精密机械研究所 Hyperspectral image openness classification method based on double loss functions
CN114972885A (en) * 2022-06-17 2022-08-30 西安电子科技大学 Multi-modal remote sensing image classification method based on model compression
CN115331110A (en) * 2022-08-26 2022-11-11 苏州大学 Fusion classification method and device for remote sensing hyperspectral image and laser radar image
CN115546569A (en) * 2022-12-05 2022-12-30 鹏城实验室 Attention mechanism-based data classification optimization method and related equipment
CN116091875A (en) * 2023-04-11 2023-05-09 合肥的卢深视科技有限公司 Model training method, living body detection method, electronic device, and storage medium

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20160307073A1 (en) * 2015-04-20 2016-10-20 Los Alamos National Security, Llc Change detection and change monitoring of natural and man-made features in multispectral and hyperspectral satellite imagery
US20170076438A1 (en) * 2015-08-31 2017-03-16 Cape Analytics, Inc. Systems and methods for analyzing remote sensing imagery
CN106845418A (en) * 2017-01-24 2017-06-13 北京航空航天大学 A kind of hyperspectral image classification method based on deep learning
CN107292343A (en) * 2017-06-23 2017-10-24 中南大学 A kind of Classification of hyperspectral remote sensing image method based on six layers of convolutional neural networks and spectral space information consolidation
CN108460342A (en) * 2018-02-05 2018-08-28 西安电子科技大学 Hyperspectral image classification method based on convolution net and Recognition with Recurrent Neural Network
CN109376804A (en) * 2018-12-19 2019-02-22 中国地质大学（武汉） Based on attention mechanism and convolutional neural networks Classification of hyperspectral remote sensing image method

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20160307073A1 (en) * 2015-04-20 2016-10-20 Los Alamos National Security, Llc Change detection and change monitoring of natural and man-made features in multispectral and hyperspectral satellite imagery
US20170076438A1 (en) * 2015-08-31 2017-03-16 Cape Analytics, Inc. Systems and methods for analyzing remote sensing imagery
CN106845418A (en) * 2017-01-24 2017-06-13 北京航空航天大学 A kind of hyperspectral image classification method based on deep learning
CN107292343A (en) * 2017-06-23 2017-10-24 中南大学 A kind of Classification of hyperspectral remote sensing image method based on six layers of convolutional neural networks and spectral space information consolidation
CN108460342A (en) * 2018-02-05 2018-08-28 西安电子科技大学 Hyperspectral image classification method based on convolution net and Recognition with Recurrent Neural Network
CN109376804A (en) * 2018-12-19 2019-02-22 中国地质大学（武汉） Based on attention mechanism and convolutional neural networks Classification of hyperspectral remote sensing image method

Cited By (54)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110516596A (en) * 2019-08-27 2019-11-29 西安电子科技大学 Empty spectrum attention hyperspectral image classification method based on Octave convolution
CN110503191A (en) * 2019-08-27 2019-11-26 四川博文讯通科技有限公司 A kind of multilayer neural network model towards video analysis
CN110516596B (en) * 2019-08-27 2023-03-24 西安电子科技大学 Octave convolution-based spatial spectrum attention hyperspectral image classification method
CN110728224B (en) * 2019-10-08 2022-03-11 西安电子科技大学 Remote sensing image classification method based on attention mechanism depth Contourlet network
CN110728224A (en) * 2019-10-08 2020-01-24 西安电子科技大学 Remote sensing image classification method based on attention mechanism depth Contourlet network
CN110866494A (en) * 2019-11-14 2020-03-06 三亚中科遥感研究所 Optical remote sensing image-based town group extraction method and system
CN110866494B (en) * 2019-11-14 2022-09-06 三亚中科遥感研究所 Urban group extraction method and system based on optical remote sensing image
CN111563520A (en) * 2020-01-16 2020-08-21 北京航空航天大学 Hyperspectral image classification method based on space-spectrum combined attention mechanism
CN111563520B (en) * 2020-01-16 2023-01-13 北京航空航天大学 Hyperspectral image classification method based on space-spectrum combined attention mechanism
CN111291622A (en) * 2020-01-16 2020-06-16 武汉汉达瑞科技有限公司 Method and device for detecting building change in remote sensing image
CN111291622B (en) * 2020-01-16 2023-07-11 武汉汉达瑞科技有限公司 Method and device for detecting building change in remote sensing image
CN111738052A (en) * 2020-04-13 2020-10-02 上海海洋大学 Multi-feature fusion hyperspectral remote sensing ground object classification method based on deep learning
CN111738052B (en) * 2020-04-13 2024-03-12 上海海洋大学 Multi-feature fusion hyperspectral remote sensing ground object classification method based on deep learning
CN111667489A (en) * 2020-04-30 2020-09-15 华东师范大学 Cancer hyperspectral image segmentation method and system based on double-branch attention deep learning
CN111667489B (en) * 2020-04-30 2022-04-05 华东师范大学 Cancer hyperspectral image segmentation method and system based on double-branch attention deep learning
CN111666313A (en) * 2020-05-25 2020-09-15 中科星图股份有限公司 Correlation construction and multi-user data matching method based on multi-source heterogeneous remote sensing data
CN111666313B (en) * 2020-05-25 2023-02-07 中科星图股份有限公司 Correlation construction and multi-user data matching method based on multi-source heterogeneous remote sensing data
CN111898662B (en) * 2020-07-20 2023-01-06 北京理工大学 Coastal wetland deep learning classification method, device, equipment and storage medium
CN111898662A (en) * 2020-07-20 2020-11-06 北京理工大学 Coastal wetland deep learning classification method, device, equipment and storage medium
CN112052758B (en) * 2020-08-25 2023-05-23 西安电子科技大学 Hyperspectral image classification method based on attention mechanism and cyclic neural network
CN112052758A (en) * 2020-08-25 2020-12-08 西安电子科技大学 Hyperspectral image classification method based on attention mechanism and recurrent neural network
CN112116563A (en) * 2020-08-28 2020-12-22 南京理工大学 Hyperspectral image target detection method and system based on spectral dimension and space cooperation neighborhood attention
CN112232343B (en) * 2020-09-03 2023-11-21 国家粮食和物资储备局科学研究院 Grain mildew grain identification neural network and identification method
CN112232343A (en) * 2020-09-03 2021-01-15 国家粮食和物资储备局科学研究院 Neural network and method for recognizing grain mildewed grains
CN112560967A (en) * 2020-12-18 2021-03-26 西安电子科技大学 Multi-source remote sensing image classification method, storage medium and computing device
CN112560967B (en) * 2020-12-18 2023-09-15 西安电子科技大学 Multi-source remote sensing image classification method, storage medium and computing device
CN112560733A (en) * 2020-12-23 2021-03-26 上海交通大学 Multitasking system and method for two-stage remote sensing image
CN112949699B (en) * 2021-02-09 2024-05-14 国家卫星气象中心(国家空间天气监测预警中心) Remote sensing image classification model building and verifying method and system and electronic equipment
CN112949699A (en) * 2021-02-09 2021-06-11 国家卫星气象中心(国家空间天气监测预警中心) Remote sensing image classification model establishing and verifying method, system and electronic equipment
CN112861978A (en) * 2021-02-20 2021-05-28 齐齐哈尔大学 Multi-branch feature fusion remote sensing scene image classification method based on attention mechanism
CN113095409A (en) * 2021-04-13 2021-07-09 西安电子科技大学 Hyperspectral image classification method based on attention mechanism and weight sharing
CN113077051B (en) * 2021-04-14 2024-01-26 南京零洞数智科技有限公司 Network model training method and device, text classification model and network model
CN113077051A (en) * 2021-04-14 2021-07-06 广东博智林机器人有限公司 Network model training method and device, text classification model and network model
CN113222044A (en) * 2021-05-25 2021-08-06 合肥工业大学 Cervical fluid-based cell classification method based on ternary attention and scale correlation fusion
CN113222044B (en) * 2021-05-25 2022-03-08 合肥工业大学 Cervical fluid-based cell classification method based on ternary attention and scale correlation fusion
CN113435253B (en) * 2021-05-31 2022-12-02 西安电子科技大学 Multi-source image combined urban area ground surface coverage classification method
CN113435253A (en) * 2021-05-31 2021-09-24 西安电子科技大学 Multi-source image combined urban area ground surface coverage classification method
CN113283529A (en) * 2021-06-08 2021-08-20 南通大学 Neural network construction method for multi-modal image visibility detection
CN113283529B (en) * 2021-06-08 2022-09-06 南通大学 Neural network construction method for multi-modal image visibility detection
CN113627590A (en) * 2021-07-29 2021-11-09 中汽创智科技有限公司 Attention module and attention mechanism of convolutional neural network and convolutional neural network
CN113705718A (en) * 2021-09-06 2021-11-26 齐齐哈尔大学 Remote sensing scene image classification method based on multi-level feature dense fusion
CN113705718B (en) * 2021-09-06 2024-04-02 齐齐哈尔大学 Remote sensing scene image classification method based on multi-level feature dense fusion
CN113780147A (en) * 2021-09-06 2021-12-10 西安电子科技大学 Lightweight hyperspectral ground object classification method and system with dynamic fusion convolution network
CN114119444A (en) * 2021-11-29 2022-03-01 武汉大学 Multi-source remote sensing image fusion method based on deep neural network
CN114119444B (en) * 2021-11-29 2024-04-16 武汉大学 Multi-source remote sensing image fusion method based on deep neural network
CN114399684B (en) * 2022-03-24 2022-08-05 中国科学院西安光学精密机械研究所 Hyperspectral image openness classification method based on double loss functions
CN114399684A (en) * 2022-03-24 2022-04-26 中国科学院西安光学精密机械研究所 Hyperspectral image openness classification method based on double loss functions
CN114972885A (en) * 2022-06-17 2022-08-30 西安电子科技大学 Multi-modal remote sensing image classification method based on model compression
CN114972885B (en) * 2022-06-17 2024-06-07 西安电子科技大学 Multi-mode remote sensing image classification method based on model compression
WO2024040828A1 (en) * 2022-08-26 2024-02-29 苏州大学 Method and device for fusion and classification of remote sensing hyperspectral image and laser radar image
CN115331110A (en) * 2022-08-26 2022-11-11 苏州大学 Fusion classification method and device for remote sensing hyperspectral image and laser radar image
CN115546569A (en) * 2022-12-05 2022-12-30 鹏城实验室 Attention mechanism-based data classification optimization method and related equipment
CN116091875B (en) * 2023-04-11 2023-08-29 合肥的卢深视科技有限公司 Model training method, living body detection method, electronic device, and storage medium
CN116091875A (en) * 2023-04-11 2023-05-09 合肥的卢深视科技有限公司 Model training method, living body detection method, electronic device, and storage medium

Also Published As

Publication number Publication date
CN109993220B (en) 2022-12-06

Similar Documents

Publication Publication Date Title
CN109993220A (en) Multi-source Remote Sensing Images Classification method based on two-way attention fused neural network
CN107220657B (en) A kind of method of high-resolution remote sensing image scene classification towards small data set
Zhu et al. A spectral-spatial-dependent global learning framework for insufficient and imbalanced hyperspectral image classification
CN110321963B (en) Hyperspectral image classification method based on fusion of multi-scale and multi-dimensional space spectrum features
Roy et al. Spectral–spatial morphological attention transformer for hyperspectral image classification
He et al. Multi-scale 3D deep convolutional neural network for hyperspectral image classification
CN113449680B (en) Knowledge distillation-based multimode small target detection method
CN107145830B (en) Hyperspectral image classification method based on spatial information enhancing and deepness belief network
CN109359559B (en) Pedestrian re-identification method based on dynamic shielding sample
CN110728192B (en) High-resolution remote sensing image classification method based on novel characteristic pyramid depth network
CN108549893A (en) A kind of end-to-end recognition methods of the scene text of arbitrary shape
CN107145836B (en) Hyperspectral image classification method based on stacked boundary identification self-encoder
CN110084159A (en) Hyperspectral image classification method based on the multistage empty spectrum information CNN of joint
CN103440505B (en) The Classification of hyperspectral remote sensing image method of space neighborhood information weighting
CN108399362A (en) A kind of rapid pedestrian detection method and device
CN106845418A (en) A kind of hyperspectral image classification method based on deep learning
CN107392130A (en) Classification of Multispectral Images method based on threshold adaptive and convolutional neural networks
CN108090447A (en) Hyperspectral image classification method and device under double branch's deep structures
CN108846426A (en) Polarization SAR classification method based on the twin network of the two-way LSTM of depth
CN111353531B (en) Hyperspectral image classification method based on singular value decomposition and spatial spectral domain attention mechanism
CN107025440A (en) A kind of remote sensing images method for extracting roads based on new convolutional neural networks
CN108734719A (en) Background automatic division method before a kind of lepidopterous insects image based on full convolutional neural networks
CN109344698A (en) EO-1 hyperion band selection method based on separable convolution sum hard threshold function
CN112347888A (en) Remote sensing image scene classification method based on bidirectional feature iterative fusion
CN107239759A (en) A kind of Hi-spatial resolution remote sensing image transfer learning method based on depth characteristic

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant