CN109447089A - High-resolution Sea Ice Model type-collection method based on oversubscription technology - Google Patents
High-resolution Sea Ice Model type-collection method based on oversubscription technology Download PDFInfo
- Publication number
- CN109447089A CN109447089A CN201811203230.2A CN201811203230A CN109447089A CN 109447089 A CN109447089 A CN 109447089A CN 201811203230 A CN201811203230 A CN 201811203230A CN 109447089 A CN109447089 A CN 109447089A
- Authority
- CN
- China
- Prior art keywords
- image
- oversubscription
- resolution
- model
- sea ice
- 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
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/46—Descriptors for shape, contour or point-related descriptors, e.g. scale invariant feature transform [SIFT] or bags of words [BoW]; Salient regional features
- G06V10/462—Salient features, e.g. scale invariant feature transforms [SIFT]
-
- 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
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10032—Satellite or aerial image; Remote sensing
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
Oversubscription reconstruction is carried out to arctic passive microwave remote sensing image using oversubscription technology the present invention relates to a kind of, and the polar region Sea Ice Types remote-sensing monitoring method of high-resolution Sea Ice Types extraction is carried out to it, specially a kind of high-resolution Sea Ice Model type-collection method based on oversubscription technology.It is characterized in that first carrying out oversubscription reconstruction to passive microwave image, then based on obtaining the oversubscription Extraction of Image Sea Ice Types of high-resolution.The present invention utilizes the strategy of super-resolution rebuilding, first rebuilds to obtain the image of high-resolution to the passive microwave remote sensing image oversubscription of polar region sea ice, then further realizes the high-resolution extraction of polar region Sea Ice Types on this basis.
Description
Technical field
The present invention relates to it is a kind of using oversubscription technology to arctic passive microwave remote sensing image carry out oversubscription reconstruction, and to its into
The polar region Sea Ice Types remote-sensing monitoring method that row high-resolution Sea Ice Types extract.
Background technique
The type of polar region sea ice is one of cryosphere and the key parameter of polar region environmental change, in global climate change study
In have a decisive role.Therefore, polar region Sea Ice Types are accurately extracted, is conducive to become to the variation of polar region sea ice
Gesture carries out accurate description and reasonable prediction, provides authentic data for the application such as polar region waterway effect.
Remotely-sensed data for sea ice monitoring mainly includes that optical image data, passive microwave data, SAR and satellite survey height
Data etc..Wherein, passive microwave data have many advantages, such as strong, all weather operations, wide coverage to earth's surface penetration capacity, make up
The defect that optical image (such as: MODIS) is influenced by cloud and weather provides especially in the phase at polar night of south poles for sea ice
Important data source.But the spatial resolution of passive microwave data is low (12.5km -25km), it is difficult to realize to polar region sea ice
The high-resolution of type is extracted.
Super resolution ratio reconstruction method is using low resolution image, by the frequent aliasing in elimination frequency domain or in space
Motor deterioration model is constructed in domain, obtains high resolution image, is had in remote sensing fields and is widely applied.
Summary of the invention
The purpose of the present invention is to overcome the deficiency in the prior art, provides a kind of Sea Ice Model class based on passive microwave image
Type high-resolution extracting method.The present invention utilizes the strategy of super-resolution rebuilding, first to the passive microwave remote sensing shadow of polar region sea ice
As oversubscription rebuilds to obtain the image of high-resolution, then the high-resolution of polar region Sea Ice Types is further realized on this basis and is mentioned
It takes.
To solve foregoing invention task, the technical solution adopted by the present invention are as follows:
A kind of high-resolution Sea Ice Model type-collection method based on oversubscription technology, which is characterized in that first to passive micro-
Wave image carries out oversubscription reconstruction, then based on obtaining the oversubscription Extraction of Image Sea Ice Types of high-resolution.The passive microwave shadow
As being AMSR2 image, raw video resolution ratio is 10km；It is the oversubscription reconstruction side based on confrontation net is generated that the oversubscription, which is rebuild,
The resolution ratio of raw video can be improved 4 times, obtain the image of 2.5km resolution ratio by method；The Sea Ice Types are one year ice and more
Two kinds of ice of year；The Sea Ice Types extraction is oversubscription of the deep learning method based on semantic segmentation to Sea Ice Model passive microwave
As a result Sea Ice Types extraction is carried out, the sea ice distribution of survey region one year ice and many years ice is obtained.
Specific implementation step is as follows:
Step 1: passive microwave image data prepares
Passive microwave image data used is in research and development institution, Japan Airlines (JAXA) whole world change observation mission (GCOM)
The data product of passive microwave scanning radiometer (AMSR2) on GCOM-W polar-orbiting satellite platform.
Step 2: the oversubscription of passive microwave image is rebuild
Step 2.1, deep learning frame is built: Ubuntu16.4+NVIDIA GTX 1080Ti GPUs+Python 2.7
+CUDA8.0+Cudnn5.1+Tensorflow 1.2。
Step 2.2, training sample prepares
Specific step is as follows:
(1) triple channel RGB image is changed into the original passive microwave batch data downloaded in step 1；
(2) RGB image obtained in (1) can be overlapped the image for being cut to certain pixel size, and meets every image
Cover the requirement in sea ice region；
Step 2.3, oversubscription model training
It is carried out using confrontation network (SRGAN, a kind of depth network model, be set as model 1) is generated in passive microwave image
What oversubscription was rebuild, it is therefore an objective to cheat the arbiter for having resolution capability, which is trained to distinguish be true picture
Or super-resolution image.Based on this method, generator can learn to generate similar as a result, therefore with true picture height
It allows arbiter to be difficult to distinguish, there are optimal solution is natural image in subspace.
Specific training process is as follows:
(1) network parameter is arranged: initial weight is VGG19 network initial weight, and learning rate 0.0001, the number of iterations is
40000 times, batch processing image parameters is set as 16, and the up-sampling factor is 4；
(2) low point of image capturing: network inputs high-resolution image (being denoted as HR), i.e. step 2.2 intermediate-resolution are 10km quilt
Dynamic microwave image, training start to obtain corresponding low point of image (being denoted as LR), resolution ratio 40km for four times of its down-sampling；
(3) generator training: low point of image LR in (2) is inputted into training generator as net is generated, carries out super-resolution
Rate is rebuild, and network output is corresponding oversubscription image SR；
(4) arbiter training: the oversubscription image SR that generator in (3) is exported is as net input is differentiated, to differentiate these
Whether image is passive microwave shadow HR in (2), is exported to judge that input picture is the probability value of high score image；
(5) model alternative optimization training: the network parameter of Maker model (3) and arbiter model (4) is by alternately excellent
Change training, two kinds of models can get a promotion, until generator generation all cannot be distinguished and nature by the study of best arbiter
The image of image (high partial image), i.e. training terminate.
Step 2.4, oversubscription model measurement:
(1) low point of image oversubscription is rebuild: the quilt downloaded based on trained network model in step 2.3, input step one
Dynamic microwave image, original spatial resolution 10km obtain the oversubscription of 2.5km spatial resolution under four times of up-sampling factors
Image.
(2) oversubscription result qualitative evaluation.
(3) oversubscription result quantitative assessment:
1) oversubscription evaluation of result standard selects, including Y-PSNR (PNSR) and structural similarity (SSIM), and value is got over
It is big then indicate that oversubscription test result more tends to image truth.
2) low/high point of image is to acquisition
The passive microwave image that step 1 is obtained is inputted as high score image, by four times as low point of image input.
3) it in the trained SRGAN network model of the low point of image input step 2.5 obtained previous step, and uses
Two kinds of evaluation indexes of PNSR and SSIM carry out quantitative analysis.
Step 3: the Sea Ice Types based on oversubscription image extract
Step 3.1, Sea Ice Model type: it is related to one year ice, many years ice and without ice formation three classes.
Step 3.2, training sample prepares: the sample one of FCN is divided into two classes, and respectively oversubscription image has language with corresponding
The mark figure of adopted label.It is produced as follows:
(1) oversubscription image obtains: the passive microwave image that step 1 is downloaded, the processing method of applying step 2.2 obtain
RGB triple channel image is then based on the trained oversubscription model (model 1) of step 2.3, and output is differentiated after oversubscription is rebuild to be certain
The oversubscription image of rate；
(2) mark figure obtains: being labeled based on three types, is respectively as follows: one year ice, many years ice and without ice formation.Mark
Software is Arcgis10.2, and specific step is as follows:
1) coordinate is converted, by coordinate used in AARI sea ice distribution map origin coordinate system transform to step 1 passive microwave image
System NSIDC_Sea_Ice_Polar_Stereographic_North, is denoted as aari.shp；
2) new face element to be created, backgroud.shp is denoted as, range size is consistent with step 1 passive microwave image,
Coordinate system is also consistent；
3) vector lattice of turnstiling differentiate the obtained aari.shp and backgroud.shp rasterizing of upper step after conversion
Rate is set as 2500m as oversubscription image spatial resolution, i.e. size is 3040 × 4480 pixels, and output format is
Tiff file, respectively aari.GIF and backgroud.GIF；
4) classification of arri.GIF, with reference to its attribute list and frequency histogram, one shares six class values, corresponds to six kinds of sea ice
Type exports as arri_clasify.GIF；
5) figure layer is inlayed, and using Mosaic tool, backgroud.GIF and arri_clasify.GIF is carried out figure layer conjunction
And it exports as arri_new.GIF；
6) image reclassification shares six kinds of Sea Ice Types in arri_new.GIF, and this time Sea Ice Types extract target are as follows: nothing
Ice formation, one year ice and many years ice three classes, respective value 0,1,2 export as arri_new_reclassify.GIF；
7) mark figure export, arri_new_reclassify.GIF is exported, png format, resolution ratio and oversubscription are saved as
Image is consistent, is 2.5km.
(3) oversubscription image and corresponding mark figure image cropping: can be overlapped cutting respectively.
Step 3.3, full convolutional neural networks (FCN, a kind of depth network model, this hair semantic segmentation model training: are applied
It is bright to be set as model 2) semantic segmentation method carry out Sea Ice Types extractions.FCN introduces full convolutional network end to end, i.e., will
Three full articulamentums in traditional convolutional neural networks (CNN) are completely converted into convolutional layer, make to calculate highly efficient.
Step 3.4, semantic segmentation model measurement: based on trained network model in step 3.3, in input step 3.2
Oversubscription image, obtain Sea Ice Model Map of Distributions of Types.
(1) classification results qualitative evaluation.
(2) classification results quantitative analysis.
Detailed description of the invention
In order to more clearly explain the embodiment of the invention or the technical proposal in the existing technology, to embodiment or will show below
There is attached drawing needed in technical description to be briefly described.It should be evident that the accompanying drawings in the following description is only this
Some embodiments of invention for those of ordinary skill in the art without creative efforts, can be with
It obtains other drawings based on these drawings.
Fig. 1 overview flow chart.
Fig. 2 step 2 flow chart.
Fig. 3 step 3 flow chart.
Fig. 4 step 3.2 flow chart.
Fig. 5 oversubscription reconstructed results.It (a) is the passive microwave image that resolution ratio is 10km to input low partial image.(b) and
It (d) is respectively green frame and red frame the corresponding region enlarged drawing in oversubscription reconstructed results in input picture.(c) and (e) is corresponding super
Image after point, resolution ratio is 2.5km.
Fig. 6 AARI sea ice distribution map.
Fig. 7 Sea Ice Types extract result: (a) be passive microwave oversubscription image (b) be classify true value (c) be classification results
It (d) is error in classification figure.
Specific embodiment
The preferred embodiment of the present invention is described in detail with reference to the accompanying drawing, so that advantages and features of the invention energy
It is easier to be readily appreciated by one skilled in the art, so as to make a clearer definition of the protection scope of the present invention.
As shown in Figure 1, the overall procedure of the embodiment of the present invention includes two parts: the oversubscription of passive microwave image is rebuild and base
It is extracted in the Sea Ice Types of oversubscription image, specific implementation step is as follows:
Step 1: passive microwave image data prepares
Passive microwave image data used in this embodiment is research and development institution, Japan Airlines (JAXA) whole world change observation mission
(GCOM) data product of the passive microwave scanning radiometer (AMSR2) on GCOM-W polar-orbiting satellite platform, can freely download.
The bright temperature product for 3 grades of horizontal polarization 36.5GHZ that product data are AMSR2 is downloaded, image size is 760 × 1120 pixels, empty
Between be covered as the Northern Hemisphere, spatial resolution 10km.
Step 2: the oversubscription of passive microwave image is rebuild
Step 2.1, deep learning frame is built: Ubuntu16.4+NVIDIA GTX 1080Ti GPUs+Python2.7+
CUDA8.0+Cudnn5.1+Tensorflow 1.2。
Step 2.2, training sample prepares: deep learning needs a large amount of external data collection to carry out training pattern, and to guarantee to instruct
Practice efficiency, the specific steps are as follows:
(1) the original passive microwave data downloaded in step 1 are 16bit single channel gray level image, having a size of 760 ×
The image of 1120 pixels changes into 760 × 1120 × 3 triple channel RGB image in batches；
(2) RGB image obtained in (1) can be overlapped the image for being cut to 400*400 pixel size, and meets every shadow
Requirement as covering sea ice region；
Final training set is 912, and test set is 72.
Step 2.3, oversubscription model training: this embodiment is using generation confrontation network (SRGAN, a kind of depth network mould
Type is set as model 1) oversubscription reconstruction is carried out in passive microwave image, it is therefore an objective to the arbiter for having resolution capability is cheated,
The arbiter is trained to distinguish true picture or super-resolution image.Based on this method, generator can learn to give birth to
At with true picture height it is similar as a result, arbiter is therefore allowed to be difficult to distinguish, there are optimal solutions in subspace is
Natural image.Specific training process is as follows:
(1) network parameter is arranged: initial weight is VGG19 network initial weight, and learning rate 0.0001, the number of iterations is
40000 times, batch processing image parameters is set as 16, and the up-sampling factor is 4；
(2) low point of image capturing: network inputs high-resolution image (being denoted as HR), i.e. step 2.2 intermediate-resolution are 10km quilt
Dynamic microwave image, training start to obtain corresponding low point of image (being denoted as LR), resolution ratio 40km for four times of its down-sampling；
(3) generator training: low point of image LR in (2) is inputted into training generator as net is generated, carries out super-resolution
Rate is rebuild, and network output is corresponding oversubscription image SR；
(4) arbiter training: the oversubscription image SR that generator in (3) is exported is as net input is differentiated, to differentiate these
Whether image is passive microwave shadow HR in (2), is exported to judge that input picture is the probability value of high score image；
(5) model alternative optimization training: the network parameter of Maker model (3) and arbiter model (4) is by alternately excellent
Change training, two kinds of models can get a promotion, until generator generation all cannot be distinguished and nature by the study of best arbiter
The image of image (high partial image), i.e. training terminate.
Step 2.4, oversubscription model measurement:
(1) low point of image oversubscription is rebuild: the quilt downloaded based on trained network model in step 2.3, input step one
Dynamic microwave image, original spatial resolution 10km obtain the oversubscription of 2.5km spatial resolution under four times of up-sampling factors
Image.
(2) oversubscription result qualitative evaluation: Fig. 5 is that oversubscription is rebuild as a result, after the original low point of image of comparison and oversubscription reconstruction
Obtained high-resolution image, the oversubscription image that SRGAN is obtained is finer, and reconstruction understands.
(3) oversubscription result quantitative assessment:
1) oversubscription evaluation of result standard selects, including Y-PSNR (PNSR) and structural similarity (SSIM), and value is got over
It is big then indicate that oversubscription test result more tends to image truth.
2) low/high point of image is to acquisition, and experimental data AMSR2 is top-quality in passive microwave data at present, temporarily
High-resolution passive microwave data can not be found and carry out the oversubscription image that quantitative analysis resolution ratio rises to 2.5km.Therefore, it will walk
Rapid one resolution ratio obtained is that 10km passive microwave image is inputted as high score image, by its bi-cubic interpolation method down-sampling four
Again to get the image for arriving 40km, inputted as low point of image.
3) in the trained SRGAN network model of low point of image (40km) input step 2.5 obtained previous step, and
Quantitative analysis is carried out using two kinds of evaluation indexes of PNSR and SSIM, as shown in 1 result of table, shows that image is with higher after rebuilding
Signal-to-noise ratio and structural similarity.
Step 3: the Sea Ice Types based on oversubscription image extract
Step 3.1, Sea Ice Model type: in polar region region, one year ice and many years ice both Sea Ice Types relatively have generation
Table and be major influence factors, and under large scale and remote sensing monitoring means in, be main Sea Ice Types, so this time
Sea Ice Types extraction is mainly concerned with one year ice, many years ice and without ice formation three classes.
Step 3.2, training sample prepares: the sample one of FCN is divided into two classes, and respectively oversubscription image has language with corresponding
The mark figure of adopted label, as shown in Figure 4.It is produced as follows:
(1) oversubscription image obtains: the passive microwave image (resolution ratio 10km) that step 1 is downloaded, applying step 2.2
Processing method, obtain RGB triple channel image, be then based on the trained oversubscription model (model 1) of step 2.3, oversubscription is rebuild
Output is the oversubscription image of 2.5km resolution ratio afterwards；
(2) mark figure obtains: carrying out classification mark pixel-by-pixel for the oversubscription image exported in (1), the present embodiment is main
It is labeled based on three types, is respectively as follows: one year ice, many years ice and without ice formation.Wherein mark figure is distributed by AARI sea ice
Figure is the sea ice product of State Scientific Centre north and south poles research institute, Russian Federation publication, and temporal resolution is 7 days, base
This covering Sea Ice Model region, data format Shapefile, as shown in Figure 6.Marking software is Arcgis10.2, specific to walk
It is rapid as follows:
1) coordinate is converted, by coordinate used in AARI sea ice distribution map origin coordinate system transform to step 1 passive microwave image
System NSIDC_Sea_Ice_Polar_Stereographic_North, is denoted as aari.shp；
2) new face element to be created, backgroud.shp is denoted as, range size is consistent with step 1 passive microwave image,
Coordinate system is also consistent；
3) vector lattice of turnstiling differentiate the obtained aari.shp and backgroud.shp rasterizing of upper step after conversion
Rate is set as 2500m as oversubscription image spatial resolution, i.e. size is 3040 × 4480 pixels, and output format is
Tiff file, respectively aari.GIF and backgroud.GIF；
4) classification of arri.GIF, with reference to its attribute list and frequency histogram, one shares six class values, corresponds to six kinds of sea ice
Type exports as arri_clasify.GIF；
5) figure layer is inlayed, and using Mosaic tool, backgroud.GIF and arri_clasify.GIF is carried out figure layer conjunction
And it exports as arri_new.GIF；
6) image reclassification shares six kinds of Sea Ice Types in arri_new.GIF, and this time Sea Ice Types extract target are as follows: nothing
Ice formation, one year ice and many years ice three classes, respective value 0,1,2 export as arri_new_reclassify.GIF；
7) mark figure export, arri_new_reclassify.GIF is exported, png format, resolution ratio and oversubscription are saved as
Image is consistent, is 2.5km.
(3) oversubscription image and corresponding mark figure that size is 3040 × 4480 pixels image cropping: can be overlapped sanction respectively
It is cut to 2048*2048 pixel size, final sample number is that training set 420 is opened, and test set 42 is opened.
Step 3.3, semantic segmentation model training: this embodiment applies full convolutional neural networks (FCN, a kind of depth network
Model, the present invention are set as model 2) semantic segmentation method carry out Sea Ice Types extractions.FCN introduces full convolution end to end
Three full articulamentums in traditional convolutional neural networks (CNN) are completely converted into convolutional layer by network, make to calculate more high
Effect.Specific step is as follows:
(1) network parameter is arranged, and initial weight is VGG19 network initial weight, and learning rate 0.0001, the number of iterations is
100000 times, batch processing image parameters is set as 2.(network parameter finally determining in case study on implementation)
(2) data input, and the oversubscription image handled well in step 3.2 and corresponding mark figure are inputted network
(3) model training carries out FCN network model and trains, and a time general one day.
Step 3.4, semantic segmentation model measurement: based on trained network model in step 3.3, in input step 3.2
Oversubscription image, Sea Ice Model Map of Distributions of Types is obtained, as shown in Fig. 7-(c).
(1) classification results qualitative evaluation: manually mark schemes (Fig. 7-(b)) and testing classification result figure in contrast and experiment
(7- (c)) and Fig. 7-(b) and 7- (c) make the difference resulting error in classification figure (Fig. 7-(d)), can accurately be obtained based on FCN
The distribution situation of a wide range of one year ice and many years ice.
(2) classification results quantitative analysis: as shown in table 2, experimental result overall precision 93% or so, Kappa coefficient reaches
0.86, illustrate that classification results on the whole and true value have high consistency, further illustrates and realize high-resolution using oversubscription technology
The scheme of rate sea ice monitoring is feasible.
1 oversubscription reconstructed image accuracy assessment result of table
2 confusion matrix of table and segmentation precision
The foregoing is merely the preferred embodiment of the present invention, protection scope of the present invention is not limited in above-mentioned embodiment party
Formula, all technical solutions for belonging to the principle of the invention all belong to the scope of protection of the present invention.For those skilled in the art and
Speech, several improvements and modifications carried out without departing from the principles of the present invention, these improvements and modifications also should be regarded as this
The protection scope of invention.
Claims (5)
1. a kind of high-resolution Sea Ice Model type-collection method based on oversubscription technology, which is characterized in that first to passive microwave
Image carries out oversubscription reconstruction, then based on obtaining the oversubscription Extraction of Image Sea Ice Types of high-resolution.
2. the high-resolution Sea Ice Model type-collection method based on oversubscription technology as described in claim 1, which is characterized in that
Specific implementation step is as follows:
Step 1: passive microwave image data prepares
Step 2: the oversubscription of passive microwave image is rebuild
Step 2.1, deep learning frame is built
Step 2.2, training sample prepares
Specific step is as follows:
(1) triple channel RGB image is changed into the original passive microwave batch data downloaded in step 1；
(2) RGB image obtained in (1) can be overlapped the image for being cut to certain pixel size, and meets every image and covers
The requirement in lid sea ice region；
Step 2.3, oversubscription model training
Oversubscription is carried out in passive microwave image using confrontation network (SRGAN, a kind of depth network model, be set as model 1) is generated
Rebuild, it is therefore an objective to cheat the arbiter for having resolution capability, the arbiter be trained to distinguish true picture or
Super-resolution image；Based on this method, generator can learn to generate highly similar to be sentenced with true picture as a result, therefore allowing
Other device is difficult to distinguish, and there are optimal solution is natural image in subspace；
Step 2.4, oversubscription model measurement:
(1) low point of image oversubscription is rebuild: based on trained network model in step 2.3, input step one is downloaded passive micro-
Wave image, original spatial resolution 10km obtain the oversubscription image of 2.5km spatial resolution under four times of up-sampling factors.
(2) oversubscription result qualitative evaluation；
(3) oversubscription result quantitative assessment:
1) oversubscription evaluation of result standard selects, including Y-PSNR (PNSR) and structural similarity (SSIM), value more it is big then
Indicate that oversubscription test result more tends to image truth；
2) low/high point of image is to acquisition
The passive microwave image that step 1 is obtained is inputted as high score image, by four times as low point of image input；
3) in the trained SRGAN network model of low point of image input step 2.5 for obtaining previous step, and using PNSR and
Two kinds of evaluation indexes of SSIM carry out quantitative analysis；
Step 3: the Sea Ice Types based on oversubscription image extract
Step 3.1, Sea Ice Model type: it is related to one year ice, many years ice and without ice formation three classes；
Step 3.2, training sample prepares: the sample one of FCN is divided into two classes, and respectively oversubscription image is with corresponding with semantic mark
The mark figure of label；It is produced as follows:
(1) oversubscription image obtains: the passive microwave image that step 1 is downloaded, the processing method of applying step 2.2 obtain RGB
Triple channel image is then based on the trained oversubscription model (model 1) of step 2.3, and output is certain resolution after oversubscription is rebuild
Oversubscription image；
(2) mark figure obtains: being labeled based on three types, is respectively as follows: one year ice, many years ice and without ice formation.
(3) oversubscription image and corresponding mark figure image cropping: can be overlapped cutting respectively；
Step 3.3, semantic segmentation model training: (FCN, a kind of depth network model, the present invention are set the full convolutional neural networks of application
For model 2) semantic segmentation method carry out Sea Ice Types extractions；FCN introduces full convolutional network end to end, i.e., will be traditional
Three full articulamentums in convolutional neural networks (CNN) are completely converted into convolutional layer, make to calculate highly efficient；
Step 3.4, semantic segmentation model measurement: super in input step 3.2 based on trained network model in step 3.3
Divide image, obtains Sea Ice Model Map of Distributions of Types；
(1) classification results qualitative evaluation；
(2) classification results quantitative analysis.
3. the high-resolution Sea Ice Model type-collection method based on oversubscription technology as described in claim 1, which is characterized in that
It is described Step 1: passive microwave image data prepare: passive microwave image data used be research and development institution, Japan Airlines (JAXA)
The data of passive microwave scanning radiometer (AMSR2) in whole world change observation mission (GCOM) on GCOM-W polar-orbiting satellite platform
Product.
4. the high-resolution Sea Ice Model type-collection method based on oversubscription technology as described in claim 1, which is characterized in that
The specific training process of the step 2.3, oversubscription model training is as follows:
(1) network parameter is arranged: initial weight is VGG19 network initial weight, and learning rate 0.0001, the number of iterations is
40000 times, batch processing image parameters is set as 16, and the up-sampling factor is 4；
(2) low point of image capturing: network inputs high-resolution image (being denoted as HR), i.e. step 2.2 intermediate-resolution are that 10km is passively micro-
Wave image, training start to obtain corresponding low point of image (being denoted as LR), resolution ratio 40km for four times of its down-sampling；
(3) generator training: low point of image LR in (2) is inputted into training generator as net is generated, carries out Super-resolution reconstruction
It builds, network output is corresponding oversubscription image SR；
(4) arbiter training: the oversubscription image SR that generator in (3) is exported is as net input is differentiated, to differentiate these images
Whether it is passive microwave shadow HR in (2), exports to judge that input picture is the probability value of high score image；
(5) model alternative optimization training: the network parameter of Maker model (3) and arbiter model (4) is instructed by alternative optimization
Practice, two kinds of models can get a promotion, and learn all indistinguishable and natural image by best arbiter until generator is generated
The image of (high partial image), i.e. training terminate.
5. the high-resolution Sea Ice Model type-collection method based on oversubscription technology as described in claim 1, which is characterized in that
The mark figure obtains: use marking software for Arcgis10.2, specific step is as follows:
1) coordinate is converted, by coordinate system used in AARI sea ice distribution map origin coordinate system transform to step 1 passive microwave image
NSIDC_Sea_Ice_Polar_Stereographic_North is denoted as aari.shp；
2) new face element is created, is denoted as backgroud.shp, range size is consistent with step 1 passive microwave image, coordinate
System is also consistent；
3) vector is turnstiled lattice, by the obtained aari.shp and backgroud.shp rasterizing of upper step, after conversion resolution ratio and
Oversubscription image spatial resolution is the same, is set as 2500m, i.e. size is 3040 × 4480 pixels, output format TIFF
File, respectively aari.GIF and backgroud.GIF；
4) classification of arri.GIF, with reference to its attribute list and frequency histogram, one shares six class values, corresponds to six kinds of sea ice classes
Type exports as arri_clasify.GIF；
5) figure layer is inlayed, and using Mosaic tool, backgroud.GIF is carried out figure layer with arri_clasify.GIF and is merged,
Output is arri_new.GIF；
6) image reclassification shares six kinds of Sea Ice Types in arri_new.GIF, and this time Sea Ice Types extract target are as follows: without ice
Area, one year ice and many years ice three classes, respective value 0,1,2 export as arri_new_reclassify.GIF；
7) mark figure export, arri_new_reclassify.GIF is exported, png format, resolution ratio and oversubscription image are saved as
It unanimously, is 2.5km.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811203230.2A CN109447089B (en) | 2018-10-16 | 2018-10-16 | High-resolution arctic sea ice type extraction method based on super-resolution technology |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811203230.2A CN109447089B (en) | 2018-10-16 | 2018-10-16 | High-resolution arctic sea ice type extraction method based on super-resolution technology |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109447089A true CN109447089A (en) | 2019-03-08 |
| CN109447089B CN109447089B (en) | 2021-09-14 |
Family
ID=65545585
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811203230.2A Active CN109447089B (en) | 2018-10-16 | 2018-10-16 | High-resolution arctic sea ice type extraction method based on super-resolution technology |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109447089B (en) |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110321939A (en) * | 2019-06-21 | 2019-10-11 | 中国电子科技集团公司第二十八研究所 | A kind of image classification method based on super-resolution rebuilding |
| CN110866363A (en) * | 2019-11-12 | 2020-03-06 | 南京信息工程大学 | Method for inverting arctic fusion pool distribution by using artificial neural network |
| CN111784584A (en) * | 2020-07-14 | 2020-10-16 | 云南电网有限责任公司昆明供电局 | Insulator remote sensing image super-resolution method based on deep learning |
| CN112700489A (en) * | 2020-12-30 | 2021-04-23 | 武汉大学 | Ship-based video image sea ice thickness measurement method and system based on deep learning |
| CN112732960A (en) * | 2020-12-30 | 2021-04-30 | 北京理工大学 | Image classification method based on online federal learning |
| CN112907720A (en) * | 2021-02-08 | 2021-06-04 | 中国海洋大学 | Sea ice data visualization method and device for realistic rendering |
| CN112966656A (en) * | 2021-03-29 | 2021-06-15 | 国家卫星海洋应用中心 | Data processing method and device |
| CN112966580A (en) * | 2021-02-25 | 2021-06-15 | 山东科技大学 | Remote sensing image green tide information extraction method based on deep learning and super-resolution |
| CN116994025A (en) * | 2023-05-30 | 2023-11-03 | 广东省国土资源测绘院 | Method and device for improving farmland recognition precision by utilizing image super-resolution technology |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2003048803A1 (en) * | 2001-11-07 | 2003-06-12 | Communications Research Laboratory,Independent Administrative Institution | Method of observing sea ice |
| US20140341423A1 (en) * | 2013-03-13 | 2014-11-20 | Conocophillips Company | Method for tracking and forecasting marine ice bodies |
| CN107516317A (en) * | 2017-08-18 | 2017-12-26 | 上海海洋大学 | A kind of SAR image sea ice sorting techniques based on depth convolutional neural networks |
| CN107679476A (en) * | 2017-09-26 | 2018-02-09 | 南京大学 | A kind of Sea Ice Types Classification in Remote Sensing Image method |
| US20180144478A1 (en) * | 2016-06-29 | 2018-05-24 | Shenzhen University | Method and system for sea background modeling and suppression on high-resolution remote sensing sea images |
-
2018
- 2018-10-16 CN CN201811203230.2A patent/CN109447089B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2003048803A1 (en) * | 2001-11-07 | 2003-06-12 | Communications Research Laboratory,Independent Administrative Institution | Method of observing sea ice |
| US20140341423A1 (en) * | 2013-03-13 | 2014-11-20 | Conocophillips Company | Method for tracking and forecasting marine ice bodies |
| US20180144478A1 (en) * | 2016-06-29 | 2018-05-24 | Shenzhen University | Method and system for sea background modeling and suppression on high-resolution remote sensing sea images |
| CN107516317A (en) * | 2017-08-18 | 2017-12-26 | 上海海洋大学 | A kind of SAR image sea ice sorting techniques based on depth convolutional neural networks |
| CN107679476A (en) * | 2017-09-26 | 2018-02-09 | 南京大学 | A kind of Sea Ice Types Classification in Remote Sensing Image method |
Non-Patent Citations (3)
| Title |
|---|
| ZISIS I. PETROU 等: "Towards breaking the spatial resolution barriers: An optical flow and super-resolution approach for sea ice motion estimation", 《ISPRS JOURNAL OF PHOTOGRAMMETRY AND REMOTE SENSING》 * |
| 石立坚 等: "水平集方法在卫星遥感图像海冰边缘检测中的应用", 《遥感信息》 * |
| 顾玲嘉 等: "被动微波遥感数据超分辨率增强与混合像元分解研究综述", 《遥感技术与应用》 * |
Cited By (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110321939A (en) * | 2019-06-21 | 2019-10-11 | 中国电子科技集团公司第二十八研究所 | A kind of image classification method based on super-resolution rebuilding |
| CN110866363A (en) * | 2019-11-12 | 2020-03-06 | 南京信息工程大学 | Method for inverting arctic fusion pool distribution by using artificial neural network |
| CN110866363B (en) * | 2019-11-12 | 2023-04-07 | 南京信息工程大学 | Method for inverting arctic fusion pool distribution by using artificial neural network |
| CN111784584A (en) * | 2020-07-14 | 2020-10-16 | 云南电网有限责任公司昆明供电局 | Insulator remote sensing image super-resolution method based on deep learning |
| CN111784584B (en) * | 2020-07-14 | 2023-08-01 | 云南电网有限责任公司昆明供电局 | Insulator remote sensing image super-resolution method based on deep learning |
| CN112700489A (en) * | 2020-12-30 | 2021-04-23 | 武汉大学 | Ship-based video image sea ice thickness measurement method and system based on deep learning |
| CN112732960A (en) * | 2020-12-30 | 2021-04-30 | 北京理工大学 | Image classification method based on online federal learning |
| CN112732960B (en) * | 2020-12-30 | 2022-09-27 | 北京理工大学 | Image classification method based on online federal learning |
| CN112907720B (en) * | 2021-02-08 | 2022-08-05 | 中国海洋大学 | Sea ice data visualization method and device for realistic rendering |
| CN112907720A (en) * | 2021-02-08 | 2021-06-04 | 中国海洋大学 | Sea ice data visualization method and device for realistic rendering |
| CN112966580B (en) * | 2021-02-25 | 2022-07-12 | 山东科技大学 | Remote sensing image green tide information extraction method based on deep learning and super-resolution |
| CN112966580A (en) * | 2021-02-25 | 2021-06-15 | 山东科技大学 | Remote sensing image green tide information extraction method based on deep learning and super-resolution |
| CN112966656A (en) * | 2021-03-29 | 2021-06-15 | 国家卫星海洋应用中心 | Data processing method and device |
| CN116994025A (en) * | 2023-05-30 | 2023-11-03 | 广东省国土资源测绘院 | Method and device for improving farmland recognition precision by utilizing image super-resolution technology |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109447089B (en) | 2021-09-14 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109447089A (en) | High-resolution Sea Ice Model type-collection method based on oversubscription technology | |
| Wang et al. | Sea ice concentration estimation during melt from dual-pol SAR scenes using deep convolutional neural networks: A case study | |
| CN109919875B (en) | High-time-frequency remote sensing image feature-assisted residential area extraction and classification method | |
| CN110070518B (en) | Hyperspectral image super-resolution mapping method based on dual-path support | |
| CN107705309A (en) | Forest parameter evaluation method in laser point cloud | |
| CN103679675B (en) | Remote sensing image fusion method oriented to water quality quantitative remote sensing application | |
| CN111639587B (en) | Hyperspectral image classification method based on multi-scale spectrum space convolution neural network | |
| CN111738144A (en) | Surface water product generation method and system based on Google Earth Engine cloud platform | |
| CN106709517A (en) | Mangrove recognition method and system | |
| CN110532953B (en) | SAR image glacier identification method based on texture feature assistance | |
| CN111008664B (en) | Hyperspectral sea ice detection method based on space-spectrum combined characteristics | |
| CN107966210B (en) | Hyperspectral image-based thermal infrared fusion reconstruction method | |
| CN107688776B (en) | Urban water body extraction method | |
| CN113610070A (en) | Landslide disaster identification method based on multi-source data fusion | |
| CN109308451A (en) | A kind of high score data information extraction system and method | |
| CN115546656A (en) | Remote sensing image breeding area extraction method based on deep learning | |
| Li et al. | Spatial-temporal super-resolution land cover mapping with a local spatial-temporal dependence model | |
| CN113837123A (en) | Mid-resolution remote sensing image offshore culture area extraction method based on spectral-spatial information combination | |
| Lu et al. | Multi-spectral remote sensing image enhancement method based on PCA and IHS transformations | |
| Tejeswari et al. | Building footprint extraction from space-borne imagery using deep neural networks | |
| CN112529828A (en) | Reference data non-sensitive remote sensing image space-time fusion model construction method | |
| CN112365508A (en) | SAR remote sensing image water area segmentation method based on visual attention and residual error network | |
| CN110363101A (en) | A kind of flowers recognition methods based on CNN Fusion Features frame | |
| CN116385894A (en) | Coastline identification method, device and equipment based on remote sensing image | |
| CN115796045A (en) | Glacier recognition model modeling method based on deep learning network and semi-supervised learning |
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 |