CN110276365A - A kind of training method and its classification method of the convolutional neural networks for the classification of SAR image sea ice - Google Patents
A kind of training method and its classification method of the convolutional neural networks for the classification of SAR image sea ice Download PDFInfo
- Publication number
- CN110276365A CN110276365A CN201810219426.4A CN201810219426A CN110276365A CN 110276365 A CN110276365 A CN 110276365A CN 201810219426 A CN201810219426 A CN 201810219426A CN 110276365 A CN110276365 A CN 110276365A
- Authority
- CN
- China
- Prior art keywords
- sea ice
- slice
- training
- classification
- convolutional neural
- 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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/29—Graphical models, e.g. Bayesian networks
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)
Abstract
The application provides the training method and its classification method of a kind of convolutional neural networks for the classification of SAR image sea ice.The method of SAR image sea ice classification based on convolutional neural networks are as follows: obtain the slice group of pixel in SAR image data to be sorted;The convolutional neural networks for being used for the classification of SAR image sea ice that slice group input has been trained obtain the sea ice classification results of the slice group;Wherein, the structure of the convolutional neural networks for the classification of SAR image sea ice includes: at least three convolutional layer and the full articulamentum of at least two, it is pond layer and normalization layer after each convolutional layer, is dropout layers after each full articulamentum, is finally the Softmax layer for exporting result.It is big to solve artificial participation SAR image sea ice classification difficulty, low efficiency, the problem of data inaccuracy passes through the automatic classification that artificial intelligence system realizes SAR image sea ice.
Description
Technical field
This application involves remote sensing fields, and in particular to the training side of the convolutional neural networks for the classification of SAR image sea ice
Method, and the method for the classification of the SAR image sea ice based on convolutional neural networks.
Background technique
The distribution of sea ice is of great significance to the design and shipping safety of seaway, and the change in long term of sea ice distribution is anti-
Atmosphere-ice-water circle interaction and whole world change have been reflected, therefore, there is important scientific meaning to sea ice distribution monitoring and answer
With value.
Field observation is mainly passed through to the observation of sea ice all the time.But the case where due to field observation, is complicated, number
According to acquisition low efficiency.Many difficulties are brought for sea ice monitoring and research.
Summary of the invention
The application provides a kind of training method of convolutional neural networks for the classification of SAR image sea ice, and one kind is based on volume
The method of the SAR image sea ice classification of product neural network.Big to solve artificial participation sea ice classification difficulty, low efficiency, data are not
Accurate problem.
In order to solve the above-mentioned technical problem, the embodiment of the present application provides the following technical solution:
The application provides a kind of training method of convolutional neural networks for the classification of SAR image sea ice, comprising:
Sample image data is fabricated to training slice file using predefined Sea Ice Types figure layer;
Convolutional neural networks using the training slice file training for the classification of SAR image sea ice reach default essence
Degree;
Wherein, it is described for SAR image sea ice classification convolutional neural networks structure include: at least three convolutional layer and
The full articulamentum of at least two, each convolutional layer are pond layer and normalization layer later, are dropout layers after each full articulamentum, most
It is the Softmax layer for exporting result afterwards.
It is optionally, described that sample image data is fabricated to training slice file using predefined Sea Ice Types figure layer,
Include:
Each pixel in sample image data is traversed, and generates and each of matches with predefined Sea Ice Types figure layer
The slice group with Sea Ice Types information of pixel;
All slice groups with Sea Ice Types information are fabricated at least one training slice file.
Further, the pixel that the generation matches with predefined Sea Ice Types figure layer has Sea Ice Types information
Slice group, comprising:
Judging that the slice of the default specification formed centered on the pixel and predefined Sea Ice Types figure layer compare is
It is no to meet preset condition;
If it is, executing following steps:
The slice of n different size is cut out centered on the pixel;
N slice is compressed into the default specification and is fabricated to the slice group with Sea Ice Types information.
Further, the slice group with Sea Ice Types information, comprising: slice group head, n are compressed into described preset
The slice of specification;Wherein, slice group head includes Sea Ice Types information.
Optionally, the default specification, equal to the specification in the convolutional neural networks channel.
Optionally, the slice of the n different size is specifically: the default specification multiplies cutting for the specification of 2 m power
Piece, the m are 0 to n-1 continuous integral numbers, and n is greater than 1.
Optionally, training parameter of the setting for the convolutional neural networks of SAR image sea ice classification, it is described to utilize the instruction
The convolutional neural networks for practicing slice file training for the classification of SAR image sea ice reach default precision, comprising:
From the training slice file acquisition training slice;
The convolutional neural networks that training slice input is used for the classification of SAR image sea ice are obtained into output result;
Judge to export whether result reaches default precision;
If it is not, then adjusting training parameter, continues to execute above-mentioned steps;
If it is, training terminates.
It is further, described from the training slice file acquisition training slice, comprising:
Training slice is obtained at random from the training slice file.
It is optionally, described that all slice groups with Sea Ice Types information are fabricated at least one training slice file,
Include:
All slice group random alignments with Sea Ice Types information are fabricated at least one training slice file.
It optionally, further include being pre-processed to described to sample image data before the method, comprising: radiation school
Positive processing, thermal noise correction process, filtering processing.
Optionally, dropout layers of the probability value is set as 0.5-0.7.
Optionally, the Sea Ice Types, comprising: put down dark sea surface, rough surface seawater, plaque-like sea ice, blocky sea ice,
Put down dark sea ice, striped sea ice.
The application provides a kind of method of SAR image sea ice classification based on convolutional neural networks, comprising:
Obtain the slice group of pixel in SAR image data to be sorted;
The convolutional neural networks for being used for the classification of SAR image sea ice that slice group input has been trained obtain the slice
The sea ice classification results of group;
Wherein, it is described for SAR image sea ice classification convolutional neural networks structure include: at least three convolutional layer and
The full articulamentum of at least two, each convolutional layer are pond layer and normalization layer later, are dropout layers after each full articulamentum, most
It is the Softmax layer for exporting result afterwards.
Optionally, the slice group for obtaining pixel in SAR image data to be sorted, comprising:
The slice of n different size is cut out centered on the pixel;
N slice is compressed into the default specification and is fabricated to slice group.
Further, the default specification, equal to the specification in the convolutional neural networks channel.
Optionally, the slice of the n different size is specifically: the default specification multiplies cutting for the specification of 2 m power
Piece, the m are 0 to n-1 continuous integral numbers, and n is greater than 1.
It optionally, further include that image to be classified data are pre-processed to described before the method, comprising: radiation
Correction process, thermal noise correction process, filtering processing.
Optionally, dropout layers of the probability value is set as 0.5-0.7.
Optionally, the Sea Ice Types, comprising: put down dark sea surface, rough surface seawater, plaque-like sea ice, blocky sea ice,
Put down dark sea ice, striped sea ice.
Disclosure based on the above embodiment can know, the embodiment of the present application have it is following the utility model has the advantages that
The application provides a kind of method of SAR image sea ice classification based on convolutional neural networks.Obtain SAR figure to be sorted
As the slice group of pixel in data;The convolutional Neural net for being used for the classification of SAR image sea ice that slice group input has been trained
Network obtains the sea ice classification results of the slice group;Wherein, the knot of the convolutional neural networks for the classification of SAR image sea ice
Structure includes: at least three convolutional layer and the full articulamentum of at least two, is pond layer and normalization layer, Mei Gequan after each convolutional layer
It is dropout layers after articulamentum, is finally the Softmax layer for exporting result.
This test is using Bering periphery as trial zone, with the SAR data of European Space Agency's Sentinel-1 satellite acquisition
Convolutional Neural metanetwork is utilized using the Tensorflow of Google company as Computational frame for research object
(Convolutional Neural Network, abbreviation CNN) technology, using multi-scale method.Solves artificial participation sea ice
Difficulty of classifying is big, low efficiency, the problem of data inaccuracy, and the automatic classification of sea ice is realized by artificial intelligence system.
Detailed description of the invention
Fig. 1 is a kind of flow chart of the training method of convolutional neural networks for the classification of SAR image sea ice;
Fig. 2 is Sea Ice Types: (a) putting down dark sea surface, (b) rough surface seawater, (c) plaque-like sea ice, (d) blocky sea
Ice (e) puts down dark sea ice, (f) striped sea ice;
Fig. 3 is the convolutional neural networks structural schematic diagram classified for SAR image sea ice;
Fig. 4 is the flow chart of a kind of method of SAR image sea ice classification based on convolutional neural networks.
Specific embodiment
In the following, being described in detail in conjunction with specific embodiment of the attached drawing to the application, but not as the restriction of the application.
It should be understood that various modifications can be made to disclosed embodiments.Therefore, description above should not regard
To limit, and only as the example of embodiment.Those skilled in the art will expect within the scope and spirit of this
Other modifications.
The attached drawing being included in the description and forms part of the description shows embodiment of the disclosure, and with it is upper
What face provided is used to explain the disclosure together to substantially description and the detailed description given below to embodiment of the disclosure
Principle.
By the description of the preferred form with reference to the accompanying drawings to the embodiment for being given as non-limiting example, the application's
These and other characteristic will become apparent.
It is also understood that although the application is described referring to some specific examples, those skilled in the art
Member realizes many other equivalents of the application in which can determine, they have feature as claimed in claim and therefore all
In the protection scope defined by whereby.
When read in conjunction with the accompanying drawings, in view of following detailed description, above and other aspect, the feature and advantage of the disclosure will become
It is more readily apparent.
The specific embodiment of the disclosure is described hereinafter with reference to attached drawing;It will be appreciated, however, that the disclosed embodiments are only
Various ways implementation can be used in the example of the disclosure.Known and/or duplicate function and structure and be not described in detail to avoid
Unnecessary or extra details makes the disclosure smudgy.Therefore, specific structural and functionality disclosed herein is thin
Section is not intended to restrictions, but as just the basis of claim and representative basis be used to instructing those skilled in the art with
Substantially any appropriate detailed construction diversely uses the disclosure.
This specification can be used phrase " in one embodiment ", " in another embodiment ", " in another embodiment
In " or " in other embodiments ", it can be referred to one or more of the identical or different embodiment according to the disclosure.
In the following, the embodiment of the present application is described in detail in conjunction with attached drawing,
The application provides a kind of training method of convolutional neural networks for the classification of SAR image sea ice;The application also mentions
For a kind of method of SAR image sea ice classification based on convolutional neural networks.It carries out one by one in the following embodiments specifically
It is bright.
The distribution of sea ice is of great significance to the design and shipping safety of seaway, and the change in long term of sea ice distribution is anti-
Atmosphere-ice-water circle interaction and whole world change have been reflected, therefore, there is important scientific meaning to sea ice distribution monitoring and answer
With value.Due to the difficulty of field observation, satellite remote sensing is the important means of sea ice monitoring, especially synthetic aperture radar
(Synthetic Aperture Radar, abbreviation SAR), do not influenced by sunshine and cloud block, high resolution, be polar region sea
The important means of ice Remote Sensing Study.
This test is using Bering periphery as trial zone, with the SAR data of European Space Agency's Sentinel-1 satellite acquisition
Convolutional neural networks are utilized using the artificial intelligence learning system Tensorflow of Google as Computational frame for research object
(Convolutional Neural Network, abbreviation CNN) technology realizes the automatic classification of sea ice using multi-scale method.
CNN has the neuron composition of the weight that can learn and biasing constant.Each neuron receives some inputs, and
It does some dot products to calculate, output is the score of each classification.It is a kind of feedforward neural network, its artificial neuron can respond
Surrounding cells in a part of coverage area have outstanding performance for large-scale image procossing.The characteristics of CNN is to tie up body with multidimensional
Long-pending neuron.The characteristics of using input being picture, neuron is designed to multiple dimensions: width, height, depth.
The depth of depth not instead of neural network is used to describe neuron.For example, the slice group size of input is 32 × 32 × 3
(rgb), then input neuron just also has 32 × 32 × 3 dimension.
Convolutional neural networks are made of the full articulamentum on one or more convolutional layers and top, while also including associated weights
With pond layer etc..This structure enables convolutional Neural network to utilize the two-dimensional structure of input data.
Convolutional layer carries out feature extraction with it.
Full articulamentum plays the role of classifier in entire convolutional neural networks.
Classifier is the general designation for the method classified in data mining to sample.
Pond layer is a form of down-sampled.Using non-linear pond function, the image of input is divided into several
Rectangular area exports maximum value to each subregion.It is steadily decreasing the space size of data, therefore the quantity and calculating of parameter
Amount can also decline, and also control over-fitting to a certain extent.
The over-fitting refers in order to obtain unanimously hypothesis and makes to assume to become over stringent.Generally use increase data
The method of amount and test sample collection evaluates classifier performance.
Layer is normalized, what which was substantially carried out is that local do subtracts and do except normalization, it can be forced in characteristic pattern
Adjacent feature carries out local competition, forces and is at war in the feature of the same space position of different characteristic figure.It is given at one
Position carry out subtraction normalization operation, the value of the actually position subtracts the value after the weighting of each pixel of neighborhood, weight
Be it is different from the positional distance Different Effects in order to distinguish, weight can be determined by a Gauss weighting windows.Division normalization
Each characteristic pattern is actually first calculated in the value of the weighted sum of the neighborhood of the same spatial position, then takes all characteristic patterns
The mean value of this value, then the value of the position of each characteristic pattern is recalculated as being the value of the point divided by maximum value, the maximum
Value refers to this in the value of the weighted sum of the neighborhood of the figure.What denominator indicated is the same spatial neighborhood in all characteristic patterns
Weighting standard is poor.It is exactly mean value and normalized square mean, that is, feature normalization for an image.
Dropout layers, in order to prevent CNN over-fitting.In one specific network of training, when the number of iterations increases
When more, in fact it could happen that network is fitted fine, but very poor to the fitting degree of verifying collection situation to training set.Institute
To introduce dropout layers, allow and fall every time for random update network parameter, increase the general ability of network.
Softmax layers, belong to multi classifier, inputs as sample characteristics, export the probability for belonging to each classification for sample.
Classification belonging to maximum probability value is classification results.
First embodiment, i.e., the training of a kind of convolutional neural networks for the classification of SAR image sea ice are provided to the application
The embodiment of method.
The present embodiment is described in detail below with reference to Fig. 1, wherein Fig. 1 is a kind of for the classification of SAR image sea ice
The flow chart of the training method of convolutional neural networks.
Sample image data is pre-processed, comprising: radiant correction processing, thermal noise correction process, filtering processing.
Pretreated purpose is to improve the quality of sample image data, keeps consistency and comparability between data.
Radiant correction processing, refer to it is to the system for causing data acquisition and Transmission system to generate due to extraneous factor, with
The process for causing image distortion because of radiation error is eliminated or is corrected in the correction that the radiation distortion of machine or distortion carry out.
Thermal noise correction process refers to the unnecessary or extra interference information progress being present in image data
Processing, can be divided into the noise due to caused by periodic excursion from noise Producing reason and be made an uproar due to caused by electromagnetic interference
Sound.The algorithm of removal noise can be divided into filter in spatial domain method and frequency filtering method.
Filtering processing inhibits the noise of target image under conditions of retaining image minutia as far as possible.
For example, making radiant correction processing to Sentinel-1IW GRD first class product using European Space Agency's SNAP software, by picture
First DN value is converted to backscattering coefficient σ, then makees thermal noise correction process, then does RefinedLee filtering processing, reduces spot
Spot noise.
Sample image data is fabricated to training slice file using predefined Sea Ice Types figure layer by step S101.
It is the basis of the present embodiment by sea ice classification.The present embodiment is analysis to SAR image as starting point, by point
It analyses image different manifestations and obtains analysis result.Therefore, classification is also to pass through the comparison with actual conditions and combination from image
The purpose of research obtains the classification results of sea ice.Sea is divided into seawater and sea ice, since sea ice be unable to do without seawater, in many cases
Seawater and sea ice are accompanied, and in order to achieve the purpose that distinguish sea ice, the present embodiment is added seawater as non-sea ice class
In Sea Ice Types.The Sea Ice Types, comprising: put down dark sea surface, rough surface seawater, plaque-like sea ice, is put down secretly at blocky sea ice
Sea ice, striped sea ice.Referring to figure 2..
Put down dark sea surface, smooth, based on mirror-reflection.
Rough surface seawater, stormy waves region, seawater surface roughness increase.
Plaque-like sea ice, internal modification is obvious, there is pressure ridge or crack between block, and back scattering is uneven;
Blocky sea ice, inside is uniform, does not deform, and back scattering is uniform;
Dark sea ice is put down, is weak in the young ice or thin ice of ice crack gap or fringe of land formation or polynia, after image scattering;
Striped sea ice, then be the mixture of ice and water at large stretch of sea ice edge and open seawater interaction, scattered in seawater
Sea ice is in striated due to drift motion.
It is described that sample image data is fabricated to training slice file using predefined Sea Ice Types figure layer, comprising:
Step S101-1 traverses each pixel in sample image data, generates and predefined Sea Ice Types figure layer phase
The slice group with Sea Ice Types information of matched pixel.
Pixel, also known as pixel or pixel point.That is image unit.It is the minimum unit for forming digitized image.
In remote sensing data acquiring, when such as scanning imagery, it is that the minimum that sensor is scanned sampling to ground scenery is single
Member;In Digital Image Processing, it is sampled point when being scanned digitlization to analog image.It is to constitute remote sensing digital image
Basic unit, be the sampled point during remotely sensed image.
Each pixel in the traversal sample image data, refers to the arrangement regulation according to pixel in sample image data
All pixels being successively read in sample data.
The predefined Sea Ice Types figure layer, refers to the Sea Ice Types in Manual definition's image, generates different sea ice
Type figure layer.
The slice group with Sea Ice Types information for the pixel that the generation matches with predefined Sea Ice Types figure layer,
Include:
Step S101-1-1, judge the default specification formed centered on the pixel slice and predefined sea ice class
Whether the comparison of type figure layer meets preset condition.
The slice refers to a piece of figure layer intercepted from sample image data.
Any specification can be set into the default specification, but for convenience of calculation, by default specification according to convolution mind
Specification is preset described in featured configuration through network.For example, the default specification, equal to the rule in the convolutional neural networks channel
Lattice.The specification in the convolutional neural networks channel is 32 × 32.For convenience of calculation, default specification is also equal to 32 × 32.The picture
The slice of the default specification formed centered on member, 32 × 32 slice formed centered on the exactly described pixel.
It is predefined to refer to whether the slice of the default specification formed centered on the pixel belongs to for the preset condition
Sea Ice Types.
Step S101-1-2, if it is, executing following steps:
Step S101-1-2-1 cuts out the slice of n different size centered on the pixel.
The slice of the n different size centered on same pixel, characteristic dimension included in each slice are different.?
More accurately output result can be obtained when training.The specification of each slice can be any specification, but for convenience of calculation, institute
The slice for stating n different size is specifically: the default specification multiplies the slice of the specification of 2 m power, and the m is 0 to n-1
Continuous integral number, n are greater than 1.For example, default specification is equal to the input of the Computational frame Tensorflow of 32 × 32, Google company
Image channel number only provides 1,3 both of which, therefore chooses n and be equal to 3, then the slice of 3 different sizes be 32 × 32,64 ×
64,128 × 128.And other Computational frame n can be the value more than or equal to 1.
N slice is compressed into the default specification and is fabricated to Sea Ice Types information by step S101-2-2-2
Slice group.
The slice group with Sea Ice Types information, comprising: slice group head, n are compressed into cutting for the default specification
Piece;Wherein, slice group head includes Sea Ice Types information.
For example, default specification, which is equal to 32 × 32, n, is equal to 3, then the slice of 3 different sizes is 32 × 32,64 × 64,128
×128.64 × 64 will be then sliced and be sliced 128 × 128 and be compressed into 32 × 32.In the present embodiment, with Sea Ice Types information
Slice group forms 1 group 3 layers 32 × 32 of slice.Wherein the 1st layer is original 32 × 32 data, and the 2nd layer is the 32 of 64 × 64 contractions
× 32 data, the 3rd layer is 128 × 128 32 × 32 data shunk.Each slice group totally 12292 byte, wherein 1-4 word
Section is floating type Sea Ice Types code, and 4096 × 3 bytes are 3 layer 32 × 32 of real-coded GA respectively later.
All slice groups with Sea Ice Types information are fabricated at least one training slice file by step S101-2.
In the present embodiment, it is preferred that all slice group random alignments with Sea Ice Types information are fabricated at least one
A training slice file.It is used for training data to avoid similar categorical data concentrations, so that training result be caused to lose
Very.
Step S102, the convolutional neural networks using the training slice file training for the classification of SAR image sea ice reach
To default precision;
Wherein, it is described for SAR image sea ice classification convolutional neural networks structure include: at least three convolutional layer and
The full articulamentum of at least two, each convolutional layer are pond layer and normalization layer later, are dropout layers after each full articulamentum, most
It is the Softmax layer for exporting result afterwards.Dropout layers of the probability value is set as 0.5-0.7.
Since the depth of convolutional neural networks is deeper, brought calculation amount is bigger, and the time of consuming is more, rationally controls
The depth of convolutional neural networks is also the option that must be taken into consideration.For example, referring to figure 3., it is preferred that described for SAR image sea
The structure of the convolutional neural networks of ice classification includes: 3 convolutional layers and 2 full articulamentums, is pond layer after each convolutional layer
With normalization layer, it is each dropout layers after full articulamentum, is finally the Softmax layer for exporting result.It is dropout layers described
Probability value be preferably 0.6.It does not only reach the purpose of sea ice classification using the structure, and greatly reduces calculation amount, save
Analysis time.
Before training, training parameter of the setting for the convolutional neural networks of SAR image sea ice classification, comprising: sliding is flat
Decay, number, learning rate decaying, initial learning rate, every wheel number of training, every wheel verifying are taken turns needed for learning rate decaying
Sample number, training batch.
The convolutional neural networks using the training slice file training for the classification of SAR image sea ice reach default
Precision, comprising:
Step S102-1, from the training slice file acquisition training slice.
It is described to be sliced from the training slice file acquisition training, comprising: to obtain instruction at random from the training slice file
Practice slice.It can be distorted in this way to avoid training result.
The convolutional neural networks that training slice input is used for the classification of SAR image sea ice are obtained output knot by step S102-2
Fruit.
Step S102-3 judges to export whether result reaches default precision.
In the present embodiment, default precision reaches 0.945-0.922.
Step S102-4, if it is not, then adjusting training parameter, continues to execute above-mentioned steps.
Step S102-5, if it is, training terminates.
The present embodiment reaches the training parameter obtained when default precision are as follows: sliding average decaying=0.9999, learning rate
Wheel number=100 needed for decaying, learning rate decaying=0.6, initial learning rate=0.05, every wheel number of training=8192,
Every wheel verifies sample number=128, training batch=60000.
The application also provides second embodiment, i.e., the method for a kind of SAR image sea ice classification based on convolutional neural networks.
Since this method embodiment is substantially similar to first method embodiment, so describing fairly simple, relevant part is referred to
The corresponding explanation of first method embodiment.Embodiment of the method described below is only schematical.
Fig. 4 shows a kind of reality of the method for SAR image sea ice classification based on convolutional neural networks provided by the present application
Apply example.Fig. 4 is the flow chart of a kind of method of SAR image sea ice classification based on convolutional neural networks.
Image to be classified data are pre-processed, comprising: radiant correction processing, thermal noise correction process, filtering processing.
Referring to FIG. 4, the application provides a kind of method of SAR image sea ice classification based on convolutional neural networks, comprising:
Step S201 obtains the slice group of pixel in SAR image data to be sorted.
The slice group for obtaining pixel in SAR image data to be sorted, comprising:
The slice of n different size is cut out centered on the pixel;
N slice is compressed into the default specification and is fabricated to slice group.
Any specification can be set into the default specification, when inputted to convolutional neural networks be sliced when, need to will be described pre-
If specification inputs convolutional neural networks as parameter.The default specification is usually arranged to 8 × 8, if the default specification
It is arranged too small, then it is very few to be sliced contained feature, influences classification results.For convenience of calculation, by default specification according to convolutional Neural
Specification is preset described in the featured configuration of network.For example, the default specification, equal to the specification in the convolutional neural networks channel.
The specification in the convolutional neural networks channel is 32 × 32.For convenience of calculation, default specification is also equal to 32 × 32.The pixel is
The slice for the default specification being centrally formed, 32 × 32 slice formed centered on the exactly described pixel.
Optionally, the slice of the n different size centered on same pixel, characteristic dimension included in each slice
It is different.More accurately output result can be obtained in training.The specification of each slice can be any specification, but in order to calculate
Convenient, the slice of the n different size is specifically: the default specification multiplies the slice of the specification of 2 m power, and the m is 0
To the continuous integral number of n-1, n is greater than 1.
For example, default specification is equal to the input picture channel of the Computational frame Tensorflow of 32 × 32, Google company
Several 1,3 both of which of offer, therefore choose n and be equal to 3, then the slice of 3 different sizes be 32 × 32,64 × 64,128 ×
128.And other Computational frame n can be the value more than or equal to 1.In the present embodiment, the slice group forms 1 group 3 layers 32
× 32 slice.Wherein the 1st layer is original 32 × 32 data, and the 2nd layer is 64 × 64 32 × 32 data shunk, and the 3rd layer is 128
× 128 32 × 32 data shunk.3 layer 32 × 32 of real-coded GA.
Step S202, the convolutional neural networks for being used for the classification of SAR image sea ice that slice group input has been trained obtain
Take the sea ice classification results of the slice group;
Wherein, it is described for SAR image sea ice classification convolutional neural networks structure include: at least three convolutional layer and
The full articulamentum of at least two, each convolutional layer are pond layer and normalization layer later, are dropout layers after each full articulamentum, most
It is the Softmax layer for exporting result afterwards.Dropout layers of the probability value is set as 0.5-0.7.
Since the depth of convolutional neural networks is deeper, brought calculation amount is bigger, and the time of consuming is more, rationally controls
The depth of convolutional neural networks is also the option that must be taken into consideration.Referring to figure 3., it is preferred that described to classify for SAR image sea ice
The structures of convolutional neural networks include: 3 convolutional layers and 2 full articulamentums, be pond layer and normalizing after each convolutional layer
Change layer, is dropout layers after each full articulamentum, is finally the Softmax layer for exporting result.Dropout layers of the probability
Value preferably 0.6.It does not only reach the purpose of sea ice classification using the structure, and greatly reduces calculation amount, saved analysis
Time.The Sea Ice Types, comprising: dark sea surface, rough surface seawater are put down, plaque-like sea ice, puts down dark sea ice at blocky sea ice,
Striped sea ice.
For example, using 4 scape Sentinel-1 data for sea ice classification drawing, wherein 3 scapes are to have neither part nor lot in trained data,
1 scape is the data for participating in microsection manufacture and convolutional neural networks training.Image, the data of each sliding window are traversed by sliding window
Input as convolutional neural networks is classified, and finally realizes the classification of whole scape image.Following table is to sample to examine to 4 scape classification charts
The sea ice classification results looked into:
As can be seen from the above table, the data for taking part in convolutional neural networks training on April 21st, 2017, put down dark seawater and item
Line sea ice preferably identified, respectively reaches 99.99% and 94.25%, and blocky sea ice recognition accuracy rate is up to 81.42%, portion
Divide and coarse seawater (14.97%) is divided by mistake.For having neither part nor lot in trained data analysis, plaque-like sea ice and coarse seawater classification knot
Fruit is preferable, can achieve 83.37% and 97.31% respectively.But put down dark sea ice and be easy to be divided by mistake to put down dark seawater, blocky sea
Ice is easy to be divided into coarse seawater by mistake, and in the data statistics region on April 9th, 2017, blocky sea ice majority is divided into slightly by mistake
Rough seawater.
Although the application is disclosed as above with preferred embodiment, it is not for limiting the application, any this field skill
Art personnel are not departing from spirit and scope, can make possible variation and modification, therefore the guarantor of the application
Shield range should be subject to the range that the claim of this application defined.
Claims (19)
1. a kind of training method of the convolutional neural networks for the classification of SAR image sea ice characterized by comprising
Sample image data is fabricated to training slice file using predefined Sea Ice Types figure layer;
Convolutional neural networks using the training slice file training for the classification of SAR image sea ice reach default precision;
Wherein, the structure of the convolutional neural networks for the classification of SAR image sea ice includes: at least three convolutional layer and at least 2
A full articulamentum, each convolutional layer are pond layer and normalization layer later, are dropout layers after each full articulamentum, are finally
Export the Softmax layer of result.
2. the method according to claim 1, wherein described utilize predefined Sea Ice Types figure layer by sample graph
As data creating is sliced file at training, comprising:
Each pixel in sample image data is traversed, and generates each pixel to match with predefined Sea Ice Types figure layer
The slice group with Sea Ice Types information;
All slice groups with Sea Ice Types information are fabricated at least one training slice file.
3. according to the method described in claim 2, it is characterized in that, the generation matches with predefined Sea Ice Types figure layer
Pixel the slice group with Sea Ice Types information, comprising:
Whether the slice and the comparison of predefined Sea Ice Types figure layer for judging the default specification formed centered on the pixel accord with
Close preset condition;
If it is, executing following steps:
The slice of n different size is cut out centered on the pixel;
N slice is compressed into the default specification and is fabricated to the slice group with Sea Ice Types information.
4. according to the method described in claim 3, it is characterized in that, the slice group with Sea Ice Types information, comprising: cut
Piece group head, n are compressed into the slice of the default specification;Wherein, slice group head includes Sea Ice Types information.
5. according to the method described in claim 3, it is characterized in that, the default specification, it is logical to be equal to the convolutional neural networks
The specification in road.
6. according to the method described in claim 3, it is characterized in that, the slice of the n different size is specifically: described default
Specification multiplies the slice of the specification of 2 m power, and the m is 0 to n-1 continuous integral number, and n is greater than 1.
7. the method according to claim 1, wherein convolutional Neural net of the setting for the classification of SAR image sea ice
The training parameter of network, the convolutional neural networks using the training slice file training for the classification of SAR image sea ice reach
To default precision, comprising:
From the training slice file acquisition training slice;
The convolutional neural networks that training slice input is used for the classification of SAR image sea ice are obtained into output result;
Judge to export whether result reaches default precision;
If it is not, then adjusting training parameter, continues to execute above-mentioned steps;
If it is, training terminates.
8. the method according to the description of claim 7 is characterized in that it is described from it is described training slice file acquisition training slice,
Include:
Training slice is obtained at random from the training slice file.
9. according to the method described in claim 2, it is characterized in that, described by all slice group systems with Sea Ice Types information
It is made at least one training slice file, comprising:
All slice group random alignments with Sea Ice Types information are fabricated at least one training slice file.
10. -9 described in any item methods according to claim 1, which is characterized in that before the method further include to described
Sample image data is pre-processed, comprising: radiant correction processing, thermal noise correction process, filtering processing.
11. -9 described in any item methods according to claim 1, which is characterized in that dropout layers of the probability value is set as
0.5-0.7。
12. -9 described in any item methods according to claim 1, which is characterized in that the Sea Ice Types, comprising: put down dark surface
Seawater, rough surface seawater, plaque-like sea ice, put down dark sea ice, striped sea ice at blocky sea ice.
13. a kind of method of the SAR image sea ice classification based on convolutional neural networks characterized by comprising
Obtain the slice group of pixel in SAR image data to be sorted;
The convolutional neural networks for being used for the classification of SAR image sea ice that slice group input has been trained obtain the slice group
Sea ice classification results;
Wherein, the structure of the convolutional neural networks for the classification of SAR image sea ice includes: at least three convolutional layer and at least 2
A full articulamentum, each convolutional layer are pond layer and normalization layer later, are dropout layers after each full articulamentum, are finally
Export the Softmax layer of result.
14. according to the method for claim 13, which is characterized in that described to obtain pixel in SAR image data to be sorted
Slice group, comprising:
The slice of n different size is cut out centered on the pixel;
N slice is compressed into the default specification and is fabricated to slice group.
15. according to the method for claim 14, which is characterized in that the default specification is equal to the convolutional neural networks
The specification in channel.
16. according to the method for claim 14, which is characterized in that the slice of the n different size is specifically: described pre-
If specification multiplies the slice of the specification of 2 m power, the m is 0 to n-1 continuous integral number, and n is greater than 1.
17. the described in any item methods of 3-16 according to claim 1, which is characterized in that before the method further include to institute
It states and image to be classified data is pre-processed, comprising: radiant correction processing, thermal noise correction process, filtering processing.
18. the described in any item methods of 3-16 according to claim 1, which is characterized in that dropout layers of the probability value is set as
0.5-0.7。
19. the described in any item methods of 3-16 according to claim 1, which is characterized in that the Sea Ice Types, comprising: put down dark table
Face seawater, rough surface seawater, plaque-like sea ice, put down dark sea ice, striped sea ice at blocky sea ice.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810219426.4A CN110276365B (en) | 2018-03-16 | 2018-03-16 | Training method of convolutional neural network for SAR image sea ice classification and classification method thereof |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810219426.4A CN110276365B (en) | 2018-03-16 | 2018-03-16 | Training method of convolutional neural network for SAR image sea ice classification and classification method thereof |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110276365A true CN110276365A (en) | 2019-09-24 |
CN110276365B CN110276365B (en) | 2021-12-28 |
Family
ID=67957847
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810219426.4A Active CN110276365B (en) | 2018-03-16 | 2018-03-16 | Training method of convolutional neural network for SAR image sea ice classification and classification method thereof |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110276365B (en) |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110996055A (en) * | 2019-12-02 | 2020-04-10 | 国网辽宁省电力有限公司检修分公司 | Self-cleaning device and method for lens of image monitoring system of power transmission line |
CN111652289A (en) * | 2020-05-15 | 2020-09-11 | 中国科学院空天信息创新研究院 | Sea ice and seawater segmentation method for synthetic aperture radar image |
CN112818851A (en) * | 2020-09-02 | 2021-05-18 | 河南工业大学 | Method for detecting icebound lake based on FY-3MWRI data |
CN112906645A (en) * | 2021-03-15 | 2021-06-04 | 山东科技大学 | Sea ice target extraction method with SAR data and multispectral data fused |
CN113376327A (en) * | 2021-07-08 | 2021-09-10 | 海南绿能环境工程有限公司 | Environmental monitoring information management method and system based on big data |
Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2005291782A (en) * | 2004-03-31 | 2005-10-20 | National Institute Of Information & Communication Technology | Ice thickness estimation method by sar |
CN101894275A (en) * | 2010-06-29 | 2010-11-24 | 武汉大学 | Weakly supervised method for classifying SAR images |
US20120328161A1 (en) * | 2011-06-22 | 2012-12-27 | Palenychka Roman | Method and multi-scale attention system for spatiotemporal change determination and object detection |
CN104392463A (en) * | 2014-12-16 | 2015-03-04 | 西安电子科技大学 | Image salient region detection method based on joint sparse multi-scale fusion |
CN105512661A (en) * | 2015-11-25 | 2016-04-20 | 中国人民解放军信息工程大学 | Multi-mode-characteristic-fusion-based remote-sensing image classification method |
US20160196672A1 (en) * | 2015-01-05 | 2016-07-07 | Superfish Ltd. | Graph image representation from convolutional neural networks |
CN107180225A (en) * | 2017-04-19 | 2017-09-19 | 华南理工大学 | A kind of recognition methods for cartoon figure's facial expression |
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 |
-
2018
- 2018-03-16 CN CN201810219426.4A patent/CN110276365B/en active Active
Patent Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2005291782A (en) * | 2004-03-31 | 2005-10-20 | National Institute Of Information & Communication Technology | Ice thickness estimation method by sar |
CN101894275A (en) * | 2010-06-29 | 2010-11-24 | 武汉大学 | Weakly supervised method for classifying SAR images |
US20120328161A1 (en) * | 2011-06-22 | 2012-12-27 | Palenychka Roman | Method and multi-scale attention system for spatiotemporal change determination and object detection |
CN104392463A (en) * | 2014-12-16 | 2015-03-04 | 西安电子科技大学 | Image salient region detection method based on joint sparse multi-scale fusion |
US20160196672A1 (en) * | 2015-01-05 | 2016-07-07 | Superfish Ltd. | Graph image representation from convolutional neural networks |
CN105512661A (en) * | 2015-11-25 | 2016-04-20 | 中国人民解放军信息工程大学 | Multi-mode-characteristic-fusion-based remote-sensing image classification method |
CN107180225A (en) * | 2017-04-19 | 2017-09-19 | 华南理工大学 | A kind of recognition methods for cartoon figure's facial expression |
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 |
Non-Patent Citations (4)
Title |
---|
AOBO ZHAI等: "SAR classification multiple scale training sample", 《REMOTE SENSING》 * |
张红 等: "面向对象的高分辨率SAR图像处理及应用", 《中国图象图形学报》 * |
涂松: "高分辨率SAR图像目标快速提取算法研究", 《中国博士学位论文全文数据库 信息科技辑》 * |
黄昕 等: "基于多尺度特征融合和支持向量机的高分辨率遥感影像分类", 《遥感学报》 * |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110996055A (en) * | 2019-12-02 | 2020-04-10 | 国网辽宁省电力有限公司检修分公司 | Self-cleaning device and method for lens of image monitoring system of power transmission line |
CN111652289A (en) * | 2020-05-15 | 2020-09-11 | 中国科学院空天信息创新研究院 | Sea ice and seawater segmentation method for synthetic aperture radar image |
CN111652289B (en) * | 2020-05-15 | 2023-01-13 | 中国科学院空天信息创新研究院 | Sea ice seawater segmentation method for synthetic aperture radar image |
CN112818851A (en) * | 2020-09-02 | 2021-05-18 | 河南工业大学 | Method for detecting icebound lake based on FY-3MWRI data |
CN112906645A (en) * | 2021-03-15 | 2021-06-04 | 山东科技大学 | Sea ice target extraction method with SAR data and multispectral data fused |
CN113376327A (en) * | 2021-07-08 | 2021-09-10 | 海南绿能环境工程有限公司 | Environmental monitoring information management method and system based on big data |
CN113376327B (en) * | 2021-07-08 | 2023-01-17 | 海南海笙信息科技有限公司 | Environmental monitoring information management method and system based on big data |
Also Published As
Publication number | Publication date |
---|---|
CN110276365B (en) | 2021-12-28 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110276365A (en) | A kind of training method and its classification method of the convolutional neural networks for the classification of SAR image sea ice | |
CN109063594A (en) | Remote sensing images fast target detection method based on YOLOv2 | |
Al Bashish et al. | A framework for detection and classification of plant leaf and stem diseases | |
CN103218831B (en) | A kind of video frequency motion target classifying identification method based on profile constraint | |
CN107016677A (en) | A kind of cloud atlas dividing method based on FCN and CNN | |
CN113392775B (en) | Sugarcane seedling automatic identification and counting method based on deep neural network | |
CN108734171A (en) | A kind of SAR remote sensing image ocean floating raft recognition methods of depth collaboration sparse coding network | |
CN104217196B (en) | A kind of remote sensing image circle oil tank automatic testing method | |
CN103942557B (en) | A kind of underground coal mine image pre-processing method | |
CN110020635A (en) | Growing area crops sophisticated category method and system based on unmanned plane image and satellite image | |
CN106971156A (en) | Rare earth mining area remote sensing information extraction method based on object-oriented classification | |
CN107247927B (en) | Method and system for extracting coastline information of remote sensing image based on tassel cap transformation | |
CN107103306A (en) | Winter wheat powdery mildew remote-sensing monitoring method based on wavelet analysis and SVMs | |
CN111598045A (en) | Remote sensing farmland change detection method based on object spectrum and mixed spectrum | |
CN110991257B (en) | Polarized SAR oil spill detection method based on feature fusion and SVM | |
CN111563430A (en) | Public welfare forest land image intelligent identification method and system based on convolutional nerves | |
CN112861810B (en) | Artificial forest planting time automatic detection method based on time sequence remote sensing observation data | |
CN111091079A (en) | TLS-based method for measuring dominant single plant structural parameters of vegetation in alpine and fragile regions | |
CN115965812B (en) | Evaluation method for classification of unmanned aerial vehicle images on wetland vegetation species and land features | |
CN112907520A (en) | Single tree crown detection method based on end-to-end deep learning method | |
CN102013095A (en) | Distribution transform-based multi-sensor image fusion method | |
CN103246894A (en) | Ground nephogram identifying method solving problem of insensitiveness in illumination | |
CN114842337A (en) | Fruit picking point identification method based on deep learning and multidimensional information fusion clustering | |
CN101533475A (en) | Method for extracting feature of shape-adaptive neighborhood based remote sensing image | |
CN117409339A (en) | Unmanned aerial vehicle crop state visual identification method for air-ground coordination |
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 | ||
TA01 | Transfer of patent application right |
Effective date of registration: 20211201 Address after: 100190 No. 19 West North Fourth Ring Road, Haidian District, Beijing Applicant after: Research Institute of aerospace information innovation, Chinese Academy of Sciences Address before: No. 9 Dengzhuang South Road, Haidian District, Beijing 100094 Applicant before: Institute of Remote Sensing and Digital Earth, Chinese Academy of Sciences |
|
TA01 | Transfer of patent application right | ||
GR01 | Patent grant | ||
GR01 | Patent grant |