CN105069468B - Hyperspectral image classification method based on ridge ripple and depth convolutional network - Google Patents
Hyperspectral image classification method based on ridge ripple and depth convolutional network Download PDFInfo
- Publication number
- CN105069468B CN105069468B CN201510451326.0A CN201510451326A CN105069468B CN 105069468 B CN105069468 B CN 105069468B CN 201510451326 A CN201510451326 A CN 201510451326A CN 105069468 B CN105069468 B CN 105069468B
- Authority
- CN
- China
- Prior art keywords
- image
- training sample
- layer
- network
- wave filter
- 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.)
- Active
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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
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 hyperspectral image classification method based on ridge ripple and depth convolutional network, mainly solves the problems, such as that the prior art is low to classification hyperspectral imagery precision and computation complexity is high.Implementation step is：1. training sample is selected in high spectrum image；2. extract the spectral information and spatial information of training sample；3. combine spectral information and spatial information composition training sample set；4. constructing five layer depth convolutional networks, and design ridge ripple wave filter and network is initialized；5. utilize the neutral net of training sample set training construction；6. pair remaining sample is classified with trained neutral net, image classification is completed.The present invention has the advantages that nicety of grading is high and classification speed is fast, available for weather monitoring, environmental monitoring, urban planning and preventing and reducing natural disasters.
Description
Technical field
The invention belongs to technical field of image processing, more particularly to a kind of hyperspectral image classification method, available for meteorology
Monitoring, environmental monitoring, land use, urban planning and prevent and reduce natural disasters.
Background technology
High spectral resolution remote sensing refers to obtain from object interested using a lot very narrow electromagnetic wave bands related
Data.Its maximum feature is exactly while target image two-dimensional space scene information is obtained, and can also obtain high-resolution
One-dimensional its physical attribute of characterization spectral information, that is, there is the characteristic of " collection of illustrative plates ", be represent remote sensing last word new
One of type technology.The main distinction of high-spectrum remote-sensing and conventional remotely-sensed data is that high-spectrum remote-sensing is narrow-band imaging, can be with
Continuous spectral information is obtained, detects material not detectable in conventional remote sensing.Therefore, when a broadband system can only
When substantially distinguishing different material species, bloom spectrum sensor but can be the detailed identification of material and more accurately estimate the abundant of it
Degree provides potential possibility.
One main contents of hyperspectral data processing are exactly ground object target classification.Classification be a kind of description ground object target or
The analytical technology of species, its main task are to assign a category label to each pixel of data volume to produce thematic maps
A kind of process, it is that people extract one of important channel of useful information from remote sensing image.The special topic ground produced after classification
Figure can clearly reflect the spatial distribution of atural object, therefrom recognize easy to people and find its rule, make high-spectrum remote-sensing figure
As there is real use value and effectively put into practical application.
Traditional image classification method is visual interpretation, and visual technology make use of the outstanding thinking ability of the mankind to carry out qualitative comment
Spatial model in valency image.This method has some disadvantages, it needs image visual interpretation person to have abundant ground
Knowledge and visual interpretation experience, and labor intensity is big, it is necessary to devote a tremendous amount of time.In addition spectral characteristic can not all be used
What the method for visual interpretation was evaluated comprehensively.In order to improve the quality of classification and efficiency, from the seventies in last century, people start weight
Technique study depending on obtaining thematic information in remote sensing images automatically by computer.Mainly known at that time using traditional statistical model
Other method carries out remote sensing computer interpretation, and nicety of grading can not be satisfactory, right with the continuous development and change of remote sensing image
Sorting algorithm is also constantly proposing new requirement, therefore improves existing sorting algorithm, and it is always that remote sensing should to find new method
With one of hot spot in research.
The feature extraction of high spectrum image is a pith of classification hyperspectral imagery, is had to nicety of grading very big
Influence.At present, the high spectrum image feature extracting method that in the market uses mainly has the feature extracting method based on spectral information,
Feature extracting method based on spatial information, and the feature extracting method of combining space information and spectral information.
In the feature extracting method based on spectral information, each pixel in high-spectral data shows in spectral space
For a spectrum response curve.Different atural object has different wave spectrum reflectivity and absorption characteristic；Identical atural object is in difference
Wave band also there is different reflectance spectrum rates, different radiation intensity is shown as in remotely-sensed data.Therefore different atural object
Spectral profile form it is different；The spectral profile of same atural object is not flat but fluctuations yet, is usually had multiple
Peak dot and valley point.Sorting technique based on spectral information is exactly to be classified using the spectral profile of different atural objects to atural object, often
Feature extraction algorithm has sparse PCA (Principal Components Analysis), ICA (Independent
Component Analysis) and LDA (Latent Dirichlet Allocation) etc..But object spectrum response meeting
Influenced be subject to several factors, such as solar illumination, atmospheric transparency and wind speed, and it is difficult accurate survey that these factors, which are usually all,
Amount, so very big difference may be had with actual curve by actually measuring obtained object spectrum response curve.Such case
Under, spectral space describing mode cannot well adapt to the analysis of high-spectral data, so the feature extraction based on spectral information
The nicety of grading that method usually obtains is not ideal enough.
Feature extracting method based on spatial information is only to be classified using the spatial information of high spectrum image, typically
The spy of the method such as feature extracting method based on variance, the feature extracting method based on gray level co-occurrence matrixes and wavelet analysis
Levy extracting method.Such method is a kind of feature extracting method of artificial experience, it is necessary to be known a priori by the feature of image, then correspond to
Selection appropriate method, so such method needs preferable priori to can be only achieved preferable classifying quality.
For this reason, Many researchers propose the feature extracting method with reference to spatially and spectrally information, by means of high-spectrum
The spectrum and spatial information of picture improve nicety of grading, such as IFRF (Image Fusion and Recursive
Filtering), EPF (Edge-Preserving Filters) and NMFL (Nonlinear and Linear Multiple
Feature Learning) method.Although such method overcomes to a certain extent is only used only spectral information or space letter
Atural object mistake caused by breath divides problem problem, but preferable nicety of grading can be obtained by still needing more priori.
Neutral net is the method for the empty spectrum signature of a kind of effective extraction, and a kind of method of the feature learning of active,
Need not there are priori, typical neutral net such as BP neural network, wavelet neural network and ridge ripple nerve net to image
Network.But these are all the neutral nets of shallow-layer, 3 layers are all only included, in order to preferably excavate the feature of image deeper,
The model of deep neural network is suggested, and typical deep neural network has own coding depth network, limited Boltzmann machine depth
Spend network, depth convolutional network etc..Since depth convolutional network is a really two-dimentional neutral net, for the image of two dimension
For, depth convolutional network can preferably represent the feature of image.But traditional depth convolutional network wave filter is initial
Change is all use random initializtion, or Gaussian function initialization, good initialization for network performance with approach speed and have
Very big influence, and these traditional initial methods are extremely difficult to a preferable effect.
The content of the invention
It is an object of the invention to the deficiency for above-mentioned prior art, proposes that one kind is based on ridge ripple and depth convolutional network
Hyperspectral image classification method, study is difficult to solve the problems, such as the prior art in classification hyperspectral imagery to effective classification
Feature, and the problem of traditional depth convolutional network is difficult to the computation complexity for reaching smaller, carry classification hyperspectral imagery
Accuracy and speed.
To achieve the above object, it is of the invention to realize that step is as follows：
1) category of terrestrial object information in a panel height spectrum picture and the image is inputted, 10% is selected from the high spectrum image
Pixel as training sample；
2) spectral information of training sample is extracted：Along the dimension of high spectrum image spectrum, each training sample is extracted
Spectral information, composition spectral vector fj, j=1 ..., J, J are the numbers of training sample；
3) dimension-reduction treatment is carried out to high spectrum image, 4 principal components before reservation, form the image after dimensionality reduction；
4) spatial information of training sample is extracted：Centered on each training sample, per one-dimensional in the image after dimensionality reduction
On, the window of 7 × 7 sizes is chosen, obtains spatial information of the sample in the dimension
5) by the spatial information of each training sampleWith spectral vector fjForm a square training sample image
Block, and the image is normalized, the training sample square image blocks F after being normalizedj；
6) one 5 layers of depth convolutional network is constructed, and with the training sample square image blocks F after normalizingjAs
The input of the convolutional network, is trained the network, obtains trained network；
7) using the pixel of residue 90% in high spectrum image as test sample, the spectral vector of each sample is extracted
fq, and space vector ni, i=1 ..., 4, forms a square test sample image block, and the image block is returned
One changes, the test sample square image blocks F after being normalizedq, q=1 ..., Q, Q are the numbers of test sample；
8) by the test sample square image blocks F after normalizationqIt is input in the trained network of step 6), according to net
Trained parameter value in network, obtains the class scale value of the sample, completes classification.
Compared with prior art, the present invention have the effect that：
(a) present invention uses the hyperspectral image classification method that spatial information and spectral information combine, tradition is overcome
Hyperspectral image classification method in only with spectral information, have ignored the problem of efficiently using spatial information, improve point
Class precision.
(b) present invention is under the frame of conventional depth convolutional neural networks, the initialization to convolutional layer wave filter in network
Using ridge ripple function, overcome space that conventional filter initial method is difficult to effectively and rapidly to approach high spectrum image and
The problem of spectral information, improve the speed of classification.
Brief description of the drawings
Fig. 1 is the hyperspectral image classification method FB(flow block) based on ridge ripple and depth convolutional network of the present invention；
Fig. 2 is that present invention experiment uses image and its true atural object classification chart；
Fig. 3 is the classification results comparison diagram with the prior art to Fig. 2 with the present invention；
Fig. 4 is to decline comparison diagram to the error of Fig. 2 with the present invention and existing method.
Embodiment
Referring to the drawings, technical solutions and effects of the present invention is described in further detail.
With reference to Fig. 1, step is as follows for of the invention realizing：
Step 1, input picture.
A panel height spectrum picture is inputted, as shown in the figure, wherein 2 (a) is the high spectrum image of input, Fig. 2 (b) is right for 2 (a)
The class logo image answered, the pixel of selection 10% is as training sample from 2 (a).
Step 2, the spectral information of training sample is extracted.
If the Spectral dimension of the high spectrum image inputted in step 1 is V, to each training sample, it is every to extract the sample
One-dimensional spectral value, composition spectral vector fj, j=1 ..., J, J are the numbers of training sample, spectral vector fjDimension be V.
Step 3, to high-spectrum image dimensionality reduction.
The method of dimensionality reduction is carried out to image the methods of sparse PCA, LDA, PCA, ICA, and the present invention is using PCA methods to step
The high spectrum image inputted in rapid 1 carries out dimension-reduction treatment, comprises the following steps that：
3a) obtain the covariance matrix of the high spectrum image inputted in step 1；
The component map of high spectrum image inputted in step 1 3b) is obtained according to covariance matrix；
The component map of 4 energy value maximums before 3c) finally retaining, by the image after this 4 component map composition dimensionality reductions.
Step 4, the spatial information of training sample is extracted.
In step 2 in each component map in the image after dimensionality reduction, centered on each training sample, 7 × 7 are chosen
The window of size, obtains spatial information of the sample in component map
Step 5, spatial information is combined with spectral information.
5a) by the image block of the representative image spatial information of 47 × 7 sizesIt is 14*14 to be combined into a size
Spatial information square image blocks
5b) by spectral vector fjIt is H*l to be rearranged into a size2Spectral information rectangle image block F, H is rectangular
The length of shape image block, l2For the width of rectangle image block, H*l2=V, wherein V are the dimension of spectral vector；
It is (H-14) * (H-l that a size 5c) is randomly selected in spatial information square image blocks G2) image block B,
According to image block B and spatial information square image blocks G and spectral information rectangle image block F, the instruction that size is H*H is built
Practice sample square image blocks：
5d) obtained training sample square image blocks A is normalized, obtains normalized training sample square
Image block Fj, j=1 ..., J, J are the numbers of training sample.
Step 6,5 layer depth convolutional networks are constructed and it is trained.
6a) construct 5 layer depth convolutional networks：Wherein the 1st layer is input layer, and layers 2 and 3 is convolutional layer, and the 4th layer is
Full linking layer, the 5th layer is softmax graders；1st layer of input is normalized training sample square image blocks Fj；
2nd layer includes L1A wave filter；3rd layer includes L2A wave filter；4th layer includes 100 node units；The layer 5
Export the category of training sample.
6b) 5 layer depth convolutional networks of training, its step are as follows：
The wave filter of two convolutional layers 6b1) is initialized using ridge ripple function, i.e.,：
First, discretization is carried out to the scale parameter a, displacement parameter b and directioin parameter θ of continuous ridge ripple function：If ruler
Spend the value range of parameter a for a ∈ (0,3], discretization at intervals of 1, the value range of directioin parameter θ for θ ∈ [0, π), it is discrete
Change is shown below at intervals of π/18, the value range of displacement parameter b：
Discretization is at intervals of 1；
Then, the wave filter set according to layers 2 and 3 in depth convolutional network, and the parameter of above-mentioned discretization, obtain
The wave filter group for including K wave filter to one, K are the number of the wave filter group median filter；
Finally, L is randomly choosed respectively from wave filter group1And L2A wave filter, as above-mentioned the 2nd layer of depth convolutional network
With the initial value of the 3rd layer of wave filter, wherein L1＜ K, L2＜ K；
6b2) by each normalized training sample square image blocks FjAs the input of input layer, by preceding to biography
Broadcast, obtain the output category of network；
Network 6b3) is exported into the least mean-square error of category and the true category of training sample as cost function；
Cost function 6b4) is minimized using back-propagation algorithm, obtains trained network parameter.
Step 7, the spectral information and spatial information of test sample are extracted.
7a) using the pixel of residue 90% in the high spectrum image that step 1 inputs as test sample；
7b) extract the spectral information of test sample：If the Spectral dimension of the high spectrum image inputted in step 1 is V, to every
One test sample, extracts the test sample per one-dimensional spectral value, composition spectral vector fq, q=1 ..., Q, Q are test specimens
This number, spectral vector fqDimension be V；
7c) extract the spatial information of test sample：In step 2 in each component map in the image after dimensionality reduction, with
Centered on each test sample, the window of 7 × 7 sizes is chosen, obtains spatial information n of the test sample in component mapi,i
=1 ..., 4；
7d) according to method in step 5, by the spatial information and n of each test samplei, i=1 ..., 4 spectral vector fqGroup
It is normalized, obtains normalized into a test sample square image blocks, and to the test sample square image blocks
Test sample square image blocks Fq, q=1 ..., Q, Q are the numbers of test sample.
Step 8, by normalized test sample square image blocks FqIt is input in step 6 in trained network, into
Row propagated forward.
The propagated forward process is：
First, by normalized test sample square image blocks FqWave filter with the 2nd layer carries out convolution, obtains the 2nd
The characteristic pattern of layer output；
Then, convolution is carried out with the 2nd layer of characteristic pattern exported and the 3rd layer of wave filter, obtains the feature of the 3rd layer of output
Figure；
Then, the characteristic pattern by the 3rd layer of output is input to the 4th layer, by the 4th layer of the output for being calculated the 4th layer；
Finally, the 4th layer be input in the 5th layer of softmax graders is obtained into the class scale value of test sample, it is complete
Constituent class.
The effect of the present invention can be further illustrated with following emulation experiment：
(1) simulated conditions
The hardware condition of emulation of the present invention is：Windows XP, SPI, CPU Pentium (R) 4, fundamental frequency are
2.4GHZ；Software platform is：MatlabR2012a；
The image credit that emulation is selected is the high spectrum image of Pavia University, and 9 class atural objects are shared in the image,
As shown in Fig. 2 (a), Fig. 2 (b) is the corresponding class logo images of Fig. 2 (a).
Data in table 1 are the training sample and the number of test sample to being selected in the image per one kind atural object.
Table 1
Emulation mode uses the method for the present invention and existing PCA, sparse PCA, IFRF, EPF and NMFL method respectively.
(2) emulation content and result
Emulation 1, Fig. 2 (a) is carried out classification emulation with of the invention and described existing five kinds of methods, as a result such as Fig. 3, its
In：
Fig. 3 (a) is the classification results figure with PCA methods,
Fig. 3 (b) be with the classification results figure of sparse PCA methods,
Fig. 3 (c) is the classification results figure with IFRF methods,
Fig. 3 (d) is the classification result figure with EPF,
Fig. 3 (e) is the classification results figure with NMFL methods,
Fig. 3 (f) is the classification results figure with the method for the present invention.
From the classification results figure of Fig. 3 (a) -3 (f) as it can be seen that the sorting technique precision and classifying quality of the present invention are more preferable.
Emulation 2, with the method and existing random initializtion method and Gauss of ridge ripple of the present invention initialization convolutional layer wave filter
Both initial methods of initial method carry out classification emulation to Fig. 2 (a), obtain error decline figure as shown in Figure 4.Fig. 4
Abscissa be iterations, ordinate is the output category of training sample and the least mean-square error of true category, with repeatedly
The increase of generation number, least mean-square error value are gradually reduced.
Figure 4, it is seen that the present invention error curve there is faster fall off rate, can with minimum calculating when
Between reach preferable nicety of grading.
Above test result indicates that：Compared with the technology of the prior art, the present invention is solving the problems, such as classification hyperspectral imagery
In adaptive learning Characteristic Problem on, there is obvious advantage, and shorten calculate the time.
Claims (3)
1. the hyperspectral image classification method based on ridge ripple and depth convolutional network, includes the following steps：
1) category of terrestrial object information in a panel height spectrum picture and the image is inputted, 10% picture is selected from the high spectrum image
Element is used as training sample；
2) spectral information of training sample is extracted：Along the dimension of high spectrum image spectrum, the light of each training sample is extracted
Spectrum information, composition spectral vector fj, j=1 ..., J, J are the numbers of training sample；
3) dimension-reduction treatment is carried out to high spectrum image, 4 principal components before reservation, form the image after dimensionality reduction；
4) spatial information of training sample is extracted：Centered on each training sample, per on one-dimensional in the image after dimensionality reduction, choosing
The window of 7 × 7 sizes is taken, obtains spatial information of the sample in the dimension
5) by the spatial information of each training sampleWith spectral vector fjForm a square training sample image block：
5a) by the image block of the representative image spatial information of 47 × 7 sizesIt is combined into the sky that a size is 14*14
Between information square image blocks
5b) by spectral vector fjIt is H*l to be rearranged into a size2Rectangle image block F, H is rectangle image block
It is long, l2For the width of rectangle image block, H*l2=V, wherein V are the dimension of spectral vector；
It is (H-14) * (H-l that a size 5c) is randomly selected in spatial information square image blocks G2) image block B, according to
Image block B and spatial information square image blocks G and spectral information rectangle image block F, the training sample that structure size is H*H
This square image blocks：
The square image blocks are normalized, the training sample square image blocks F after being normalizedj；
6) one 5 layers of depth convolutional network is constructed, and with the training sample square image blocks F after normalizingjAs the convolution
The input of network, is trained the network, obtains trained network：
The wave filter of two convolutional layers 6a) is initialized using ridge ripple function；
6b) by each training sample square image blocks F after normalizingjAs the input of input layer, by propagated forward,
Obtain the output category of network；
Network 6c) is exported into the least mean-square error of category and the true category of training sample as cost function；
Cost function 6d) is minimized using back-propagation algorithm, obtains trained network parameter；
7) using the pixel of residue 90% in high spectrum image as test sample, the spectral vector f of each sample is extractedq, and
Space vector ni, i=1 ..., 4, forms a square test sample image block, and the image block is normalized,
Test sample square image blocks F after being normalizedq, q=1 ..., Q, Q are the numbers of test sample；
8) by the test sample square image blocks F after normalizationqIt is input in the trained network of step 6), according in network
Trained parameter value, obtains the class scale value of the sample, completes classification.
2. the hyperspectral image classification method according to claim 1 based on ridge ripple and depth convolutional network, wherein described
5 layer depth convolutional networks in step 6), its 1st layer is input layer, and layers 2 and 3 is convolutional layer, and the 4th layer is full link
Layer, the 5th layer is softmax graders；
Described 1st layer of input is training sample square image blocks Fj；
Described 2nd layer includes L1A wave filter；
Described 3rd layer includes L2A wave filter；
Described 4th layer includes 100 node units；
The category of the 5th layer of output training sample.
3. the hyperspectral image classification method according to claim 1 based on ridge ripple and depth convolutional network, wherein described
Step 6a) the middle wave filter using ridge ripple function two convolutional layers of initialization, carry out in accordance with the following steps：
Discretization 6a1) is carried out to the parameter of continuous ridge ripple function：
The continuous ridge ripple functional expression is：Wherein (x1,x2) be wave filter coordinate
Value, a is scale parameter, and b is displacement parameter, and θ is directioin parameter, and ψ (x) is wavelet function,
If the value range of scale parameter a for a ∈ (0,3], discretization at intervals of 1, the value range of directioin parameter θ for θ ∈ [0,
π), discretization is shown below at intervals of π/18, the value range of displacement parameter b：
Discretization is at intervals of 1；
6a2) according in depth convolutional network layers 2 and 3 set wave filter, and step 6a1) in set parameter, obtain
The wave filter group for including K wave filter to one, K are the number of the wave filter group median filter；
6a3) L is randomly choosed respectively from wave filter group1And L2A wave filter, as the 2nd layer of above-mentioned depth convolutional network and the 3rd
The initial value of layer wave filter, wherein L1＜ K, L2＜ K.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510451326.0A CN105069468B (en) | 2015-07-28 | 2015-07-28 | Hyperspectral image classification method based on ridge ripple and depth convolutional network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510451326.0A CN105069468B (en) | 2015-07-28 | 2015-07-28 | Hyperspectral image classification method based on ridge ripple and depth convolutional network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN105069468A CN105069468A (en) | 2015-11-18 |
| CN105069468B true CN105069468B (en) | 2018-04-17 |
Family
ID=54498829
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201510451326.0A Active CN105069468B (en) | 2015-07-28 | 2015-07-28 | Hyperspectral image classification method based on ridge ripple and depth convolutional network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN105069468B (en) |
Families Citing this family (20)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105631480B (en) * | 2015-12-30 | 2018-10-26 | 哈尔滨工业大学 | The Hyperspectral data classification method folded based on multilayer convolutional network and data recombination |
| CN105718957A (en) * | 2016-01-26 | 2016-06-29 | 西安电子科技大学 | Polarized SAR image classification method based on nonsubsampled contourlet convolutional neural network |
| CN106023065B (en) * | 2016-05-13 | 2019-02-19 | 中国矿业大学 | A kind of tensor type high spectrum image spectral-spatial dimension reduction method based on depth convolutional neural networks |
| CN106203482A (en) * | 2016-06-30 | 2016-12-07 | 东南大学 | Characteristics of The Remote Sensing Images dimension reduction method based on mRMR and KPCA |
| CN106326899A (en) * | 2016-08-18 | 2017-01-11 | 郑州大学 | Tobacco leaf grading method based on hyperspectral image and deep learning algorithm |
| CN106529570B (en) * | 2016-10-14 | 2019-06-18 | 西安电子科技大学 | Image classification method based on depth ridge ripple neural network |
| CN108229515A (en) * | 2016-12-29 | 2018-06-29 | 北京市商汤科技开发有限公司 | Object classification method and device, the electronic equipment of high spectrum image |
| CN107122708A (en) * | 2017-03-17 | 2017-09-01 | 广东工业大学 | Classification hyperspectral imagery algorithm based on convolutional neural networks and the learning machine that transfinites |
| CN106997380B (en) * | 2017-03-21 | 2019-07-12 | 北京工业大学 | Imaging spectrum safe retrieving method based on DCGAN depth network |
| CN107194373A (en) * | 2017-06-16 | 2017-09-22 | 河海大学 | A kind of target in hyperspectral remotely sensed image feature extraction and classifying method |
| CN107292343B (en) * | 2017-06-23 | 2020-05-22 | 中南大学 | Hyperspectral remote sensing image classification method based on six-layer convolutional neural network and spectrum-space information combination |
| CN107862324B (en) * | 2017-10-19 | 2021-11-02 | 北京化工大学 | MWSPCA-based CBR prediction model intelligent early warning method |
| CN108171130A (en) * | 2017-12-15 | 2018-06-15 | 安徽四创电子股份有限公司 | A kind of EO-1 hyperion terrain classification recognition methods |
| CN108460342B (en) * | 2018-02-05 | 2021-01-01 | 西安电子科技大学 | Hyperspectral image classification method based on convolutional neural network and cyclic neural network |
| CN108428021B (en) * | 2018-05-21 | 2021-10-12 | 国网山东省电力公司青岛供电公司 | Micro-grid short-term load prediction model based on HSA-RRNN |
| CN108830243A (en) * | 2018-06-22 | 2018-11-16 | 西安电子科技大学 | Hyperspectral image classification method based on capsule network |
| CN109389080B (en) * | 2018-09-30 | 2022-04-19 | 西安电子科技大学 | Hyperspectral image classification method based on semi-supervised WGAN-GP |
| CN110458208A (en) * | 2019-07-24 | 2019-11-15 | 哈尔滨工业大学 | Hyperspectral image classification method based on information measure |
| CN111222576B (en) * | 2020-01-08 | 2023-03-24 | 西安理工大学 | High-resolution remote sensing image classification method |
| CN112183669B (en) * | 2020-11-04 | 2024-02-13 | 航天科工(北京)空间信息应用股份有限公司 | Image classification method, device, equipment and storage medium |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104050507A (en) * | 2014-06-30 | 2014-09-17 | 南京理工大学 | Hyper spectrum image classification method based on multilayer neural network |
| CN104102929A (en) * | 2014-07-25 | 2014-10-15 | 哈尔滨工业大学 | Hyperspectral remote sensing data classification method based on deep learning |
| CN104700116A (en) * | 2015-03-13 | 2015-06-10 | 西安电子科技大学 | Polarized SAR (synthetic aperture radar) image object classifying method based on multi-quantum ridgelet representation |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8711028B2 (en) * | 2010-08-26 | 2014-04-29 | Lawrence Livermore National Security, Llc | Buried object detection in GPR images |
-
2015
- 2015-07-28 CN CN201510451326.0A patent/CN105069468B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104050507A (en) * | 2014-06-30 | 2014-09-17 | 南京理工大学 | Hyper spectrum image classification method based on multilayer neural network |
| CN104102929A (en) * | 2014-07-25 | 2014-10-15 | 哈尔滨工业大学 | Hyperspectral remote sensing data classification method based on deep learning |
| CN104700116A (en) * | 2015-03-13 | 2015-06-10 | 西安电子科技大学 | Polarized SAR (synthetic aperture radar) image object classifying method based on multi-quantum ridgelet representation |
Also Published As
| Publication number | Publication date |
|---|---|
| CN105069468A (en) | 2015-11-18 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN105069468B (en) | Hyperspectral image classification method based on ridge ripple and depth convolutional network | |
| CN110929607B (en) | Remote sensing identification method and system for urban building construction progress | |
| Pu et al. | Object-based urban detailed land cover classification with high spatial resolution IKONOS imagery | |
| Wharton | A Spectral-Knowledge-Based Approach for Urban Land-Cover Discrmination | |
| CN103456018B (en) | Remote sensing image change detection method based on fusion and PCA kernel fuzzy clustering | |
| CN103413151B (en) | Hyperspectral image classification method based on figure canonical low-rank representation Dimensionality Reduction | |
| CN106529508A (en) | Local and non-local multi-feature semantics-based hyperspectral image classification method | |
| CN110728197B (en) | Single-tree-level tree species identification method based on deep learning | |
| CN107358203B (en) | A kind of High Resolution SAR image classification method based on depth convolution ladder network | |
| CN108460391A (en) | Based on the unsupervised feature extracting method of high spectrum image for generating confrontation network | |
| Peeters et al. | Automated recognition of urban objects for morphological urban analysis | |
| Ahmad et al. | Hybrid dense network with attention mechanism for hyperspectral image classification | |
| Asming et al. | Processing and classification of landsat and sentinel images for oil palm plantation detection | |
| Tao et al. | Deep Learning for Hyperspectral Image Analysis and Classification | |
| CN110163294A (en) | Remote Sensing Imagery Change method for detecting area based on dimensionality reduction operation and convolutional network | |
| CN110516727B (en) | Hyperspectral image classification method based on FPGA (field programmable Gate array) depth edge filter | |
| CN116664954A (en) | Hyperspectral ground object classification method based on graph convolution and convolution fusion | |
| Xie et al. | Improvement and application of UNet network for avoiding the effect of urban dense high-rise buildings and other feature shadows on water body extraction | |
| CN117115669B (en) | Object-level ground object sample self-adaptive generation method and system with double-condition quality constraint | |
| Wang et al. | Histogram matching-based semantic segmentation model for crop classification with Sentinel-2 satellite imagery | |
| CN115019174B (en) | Up-sampling remote sensing image target recognition method based on pixel recombination and attention | |
| Chen et al. | Mapping urban form and land use with deep learning techniques: a case study of Dongguan City, China | |
| Li et al. | Identifying building rooftops in hyperspectral imagery using CNN with pure pixel index | |
| Liang et al. | Local Spectral Similarity-Guided Sparse Unmixing of Hyperspectral Images With Spatial Graph Regularization | |
| Srivastava et al. | Correlation Between Built-Up and Land Surface Temperature Using Sentinel-2 and Landsat-8 Images through Semi-Supervised Deep Learning Model for Efficient Land Surface Monitoring |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |