CN110866363A - Method for inverting arctic fusion pool distribution by using artificial neural network - Google Patents

Method for inverting arctic fusion pool distribution by using artificial neural network Download PDF

Info

Publication number
CN110866363A
CN110866363A CN201911098932.3A CN201911098932A CN110866363A CN 110866363 A CN110866363 A CN 110866363A CN 201911098932 A CN201911098932 A CN 201911098932A CN 110866363 A CN110866363 A CN 110866363A
Authority
CN
China
Prior art keywords
proportion
mod09ga
neural network
pool
layer
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN201911098932.3A
Other languages
Chinese (zh)
Other versions
CN110866363B (en
Inventor
张渊智
冯佳俊
何宜军
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Nanjing University of Information Science and Technology
Original Assignee
Nanjing University of Information Science and Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Nanjing University of Information Science and Technology filed Critical Nanjing University of Information Science and Technology
Priority to CN201911098932.3A priority Critical patent/CN110866363B/en
Publication of CN110866363A publication Critical patent/CN110866363A/en
Application granted granted Critical
Publication of CN110866363B publication Critical patent/CN110866363B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • 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
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/048—Activation functions
    • 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
    • G06N3/084—Backpropagation, e.g. using gradient descent

Abstract

The invention discloses a method for inverting the distribution of a north pole fusion pool by using an artificial neural network, which comprises the steps of firstly, collecting a high-resolution fusion pool image and a contemporaneous medium-resolution imaging spectrometer MOD09GA, and carrying out re-projection to obtain sinusoidal projection; then, a grid consistent with MOD09GA is constructed; calculating the proportion of a molten pool, the proportion of sea ice and the proportion of boiled water and drained water in each grid, and then deleting pixels covered by cloud in the MOD09 GA; the acquired data and the corresponding day-to-day surface reflectivities, MOD09GA 1 to 7 band reflectivities, form a data set; and finally, constructing a melt pool model, taking the well-matched MOD09GA 1-7 wave band reflectivity as an input set, and taking the corresponding sea ice proportion, melt pool proportion and open water proportion as an output sample set to train an artificial neural network to find an optimal solution. The invention makes up the defects of field observation and navigation observation in time and space, has the characteristics of rapidness and simplicity, improves the remote sensing monitoring level of the polar region in China, and can better reflect the real molten pool condition of the arctic.

Description

Method for inverting arctic fusion pool distribution by using artificial neural network
Technical Field
The invention belongs to the technical field of polar region remote sensing, and relates to a method for inverting the distribution of a molten arctic pool by using an artificial neural network.
Background
The melting pool can cover 50% -60% of the range of the arctic sea ice in the arctic summer. The formation of a sea ice surface molten pool is caused by short-wave sunshine in summer and surface air temperature higher than the freezing point, so that the sea ice surface albedo is reduced, and the sea ice absorbs more extra heat. The melting pool is used as the most important parameter for the change of the albedo of the sea ice in the spring and summer of the arctic, and the melting pool with low albedo can accelerate the melting of the sea ice to cause the coming of the ice-free season of the arctic earlier. Therefore, the monitoring of arctic molten pool change has high practical significance.
Conventional monitoring methods include navigational observation, and measurement in the field. However, these methods are limited by the laboratory equipment and field observers, and the data obtained are not amenable to extensive arctic melt mapping. Until Tschudi et al proposed a spectral decomposition method to extract the boftt/cuckoo molten pool distribution. The algorithm solves 4 sets of linear equations based on four different types of reflectivity characteristics (molten pool, open water, snow and bare ice).
Figure BDA0002269242250000011
The algorithm of Tschudi is extended to the entire north pole and the four types of reflectivity features are grouped into three, but this algorithm uses the same fixed reflectivity as Tschudi.
Figure BDA0002269242250000012
Compared with the high-resolution melting image root mean square error issued by the American snow ice center (NSIDC), the method of (1) achieves more than 10.7%. Modeling with fixed or insufficient a priori conditions may result in large errors that may even affect the accuracy of the melt pool distribution as input to the model.
Under the condition, the invention provides a remote sensing detection method for arctic molten pool distribution based on an artificial neural network by using measured data as a prior condition.
Disclosure of Invention
The purpose of the invention is as follows: the method for inverting the distribution of the arctic molten pool by using the artificial neural network makes up for the defects of the remote sensing monitoring molten pool in time and space distribution and precision, and improves the application level of the polar region monitoring business in China.
The technical scheme is as follows: the invention relates to a method for inverting the distribution of a molten north pole pool by using an artificial neural network, which comprises the following steps of:
(1) collecting a high-resolution fusion pool image and a contemporaneous medium-resolution imaging spectrometer MOD09GA, wherein the high-resolution fusion pool image is projected in a mode of mercator projection, and the re-projection is sinusoidal projection, and the MOD09GA projection is sinusoidal projection;
(2) constructing a grid consistent with MOD09 GA;
(3) calculating a molten pool proportion, a sea ice proportion and a water opening and discharging proportion in each grid according to the MOD09GA grids, and then deleting pixels covered by the cloud in the MOD09 GA;
(4) counting the data obtained in the step (3) and the corresponding MOD09GA 1 to 7 wave band daily surface reflectivity to form a data set, and dividing the data set into a training set and a verification set;
(5) and (3) constructing a melting tank model, taking the well-matched MOD09GA 1-7 wave band reflectivity as an input set, and taking the corresponding sea ice proportion, melting tank proportion and open water proportion as output samples to train an artificial neural network to find an optimal solution.
Further, the projection mode and resolution of the grids are consistent with those of MOD09GA, and each grid corresponds to a daily earth surface reflectivity
Further, the ratio of the training set to the validation set in step (4) is 7: 3.
Further, the artificial neural network in the step (5) is a stacked self-coding and is composed of a plurality of self-codings; the self-coding comprises a three-layer neural network consisting of an input layer, a hidden layer and an output layer.
Further, the objective function of the molten pool model in the step (5) is as follows:
Figure BDA0002269242250000021
wherein the first term is a basic error term, the second term is a regular term, xiTo input samples, yiTo output a sample, nlIs the number of network layers, slAnd sl+1Is the number of neurons in layer l and layer l +1,
Figure BDA0002269242250000022
the weight between the ith neuron on the l th layer and the jth neuron on the l +1 th layer is obtained, lambda is a weight attenuation parameter, the learning rate of the stacked self-coding is set to be 0.15, the momentum is set to be 0.5, the linear is selected as the output function, and the sigmoid is selected as the activation function; the first self-coding layer was set with 25 neurons and trained 5000 times, the second with 15 neurons and trained 5000 times, and the back propagation layer was also trained 5000 times.
Has the advantages that: compared with the prior art, the invention has the beneficial effects that: 1. the invention provides a new remote sensing method of the arctic molten pool coverage range, which can be used for large-range drawing, by utilizing the matching of high-resolution molten pool images and MODIS surface reflectivity data and modeling through an artificial neural network; 2. the MODSI surface reflectivity product is input into the molten pool model, so that the molten pool distribution condition can be directly obtained, the method has the characteristics of quickness and simplicity, the defects of field observation and navigation observation in time and space are overcome, and the remote sensing monitoring level of the polar region in China is improved; 3. the fusion data training set of the invention is from real data, and can reflect the real fusion condition of the arctic.
Drawings
FIG. 1 is a flow chart of the present invention;
FIG. 2 is a grid map consistent with MOD09 GA;
FIG. 3 is a diagram of a self-encoder structure;
FIG. 4 is a flow chart of stacked self-encoding;
FIG. 5 is a graph of the error between the melt ratio and the real data obtained using the model.
Detailed Description
The present invention is described in further detail below with reference to the attached drawing figures.
The invention provides a new remote sensing monitoring method through the relations between the reflectivity of a medium resolution imaging spectrometer (MODIS) with the resolution of 500m and the reflectivity of a 1-7 wave band of a MODIS surface reflectivity product (MOD09GA) day by day and a molten pool, sea ice and seawater, and the method can be applied to an MODIS surface reflectivity product (MOD09A1) with the resolution of 500m for 8 days to successfully reverse the distribution characteristics of an arctic molten pool. The method makes up the defects of the remote sensing monitoring fusion pool in time and space distribution and precision, and improves the application level of the polar monitoring service in China.
As shown in fig. 1, the research area is 60 ° to 90 ° north latitude, and the specific steps are as follows:
1. high resolution melt pool images from the american snow ice center from 2000 to 2001 and contemporary medium resolution imaging spectrometers earth surface reflectance product images day by day were collected (MOD09 GA). The projection mode of the high resolution fusion image is the mercator projection (UTM), and MOD09GA uses the Sinusoidal projection (sinussoid). The high resolution melt pool image was re-projected as a sinusoidal projection consistent with MOD09GA using nearest neighbor interpolation of the ENVI software.
2. And constructing a grid consistent with the daily earth surface reflectivity product image. The tool box in the Arcgis software selects a data management tool-element class-creates an area vector fishing net, constructs grids consistent with MOD09GA, the projection mode of the grids is consistent with the daily surface reflectivity product image, the resolution of the grids is consistent with the daily surface reflectivity product image, and each grid corresponds to one daily surface reflectivity, as shown in fig. 2.
3. The tool box in the Arcgis software selects the conversion tool-out of the grid-out as a face vector. And converting the high-resolution molten pool image into a planar element. The tool box in the Arcgis software selects the analysis tool-statistics-intersection tabulation. The input space element selects the created planar fishing net, and the input type element selects the high-resolution pool-melting image converted into the planar element. The results output by Arcgis are the proportion of the melting pool, the proportion of sea ice and the proportion of boiled water and drained water in each fishing net. Table 1 shows the calculated and counted proportion of the partial melting pool, the sea ice proportion and the boiled water and drained water proportion of Arcgis.
TABLE 1 proportion of melting pool, sea ice and open water under part of MODIS grid
B1 B2 B3 B4 B5 B6 B7 Sea ice Melting tank Water boiling and discharging
78 46 300 194 74 62 7 0.070 0.098 0.83
301 217 450 387 85 8 8 0 0.034 0.95
1075 706 1253 1618 -58 -18 3022 0.10 0.10 0.78
2250 1747 2198 2713 306 195 3056 0.072 0.084 0.83
1293 945 1438 1796 30 115 3011 0 0.02 0.97
192 167 347 262 55 30 -3 0.02 0.05 0.92
363 329 453 381 96 27 18 0.02 0.04 0.93
2250 1747 2198 2713 306 195 3056 0.04 0.07 0.88
787 554 1155 1129 206 98 943 0.03 0.05 0.99
6662 4955 7881 7603 1167 94 28 0.84 0.15 0
5689 4264 6955 6637 1431 400 1110 0.78 0.20 0.01
5544 4109 6889 6525 1294 338 1119 0.75 0.20 0.04
4. And (4) counting the data obtained in the step 3 and the corresponding MOD09GA 1 to 7 wave band surface reflectivity, and dividing the data into a training set and a verification set.
Using a Matlab neural network toolbox to count the reflectivity and the melting pool proportion, the sea ice proportion and the proportion of open water below each grid, taking the reflectivity of the first 7 wave bands of MOD09GA which is matched by 70% as an input set, taking the corresponding melting pool proportion, the sea ice proportion and the open water proportion as an output set to train the neural network, and taking the rest 30% as a verification set for verifying the correctness of the network.
5. And (3) constructing a melting pool model, taking the matched MOD09GA reflectivity as an input set, and taking the corresponding sea ice proportion, melting pool proportion and open water proportion as an output sample set to train an artificial neural network to find an optimal solution.
The neural network used in the invention is a Stacked self-encoding (Stacked auto encoder), which is a deep learning network based on an artificial neural network. It consists of a plurality of self-encoders. The self-coding is a three-layer neural network composed of an input layer, a hidden layer, and an output layer, as shown in fig. 3.
The training of the self-encoder network comprises the encoding and decoding processes, the encoding process is to use a nonlinear activation function to map an input layer to a hidden layer, the decoding process is to reconstruct the data of the hidden layer, and the decoding and encoding formula is as follows:
h＝f(x)＝sf(wx+b)
Figure BDA0002269242250000051
wherein f (x) and g (x) represent the encoding function and the decoding function, respectively, sfWhich represents the function of the activation of the code,
sgrepresenting a decoding activation function. w is ax+ b and
Figure BDA0002269242250000052
the weights of the encoder and decoder, respectively. Selected by the inventionThe activation function is a Sigmoid function, and the formula is as follows:
Figure BDA0002269242250000053
wherein x is wp+b。
Given an input vector, the goal of the auto-encoder is to minimize the input x and the output y
The difference between the two, the reconstruction error, can be described by a cross entropy function, which is given by the following equation:
Figure BDA0002269242250000054
for the training set S, the average reconstruction error can be represented as:
Figure BDA0002269242250000055
after the self-encoder is encoded and decoded, the network parameters need to be trained by using a back propagation algorithm, so that the input value is reconstructed. The key to the training of the self-encoder is to obtain the weights and biases that minimize the objective function. When the N samples are trained by the self-encoder, the optimal weights and offsets can be achieved by optimizing the objective function. The calculation formula of the objective function is as follows:
Figure BDA0002269242250000056
the first term is a basic error term and the second term is a regularization term. x is the number ofiTo input samples, yiTo output a sample, nlIs the number of network layers, slAnd sl+1Is the number of neurons in layer l and layer l +1,
Figure BDA0002269242250000057
is the weight between the i-th neuron of layer l and the j-th neuron of layer l +1, and λ is a weight decay parameter.
FIG. 4 is a flow chart of stacked self-coding, which comprises inputting a surface reflectivity product, training 5000 times with the input data to train the first self-coding grid with 25 neurons to obtain weights and offsets, training 5000 times with the first self-coding output as the second self-coding input with 15 neurons, and training the weights and offsets of the second self-coder. And finally, after the initialization of the parameters in the network is finished, performing 'fine tuning' on the parameters by using a back propagation layer, and training for 5000 times. The learning rate of the stack type self-coding is set to be 0.15, the momentum is set to be 0.5, the function is output to select linear, and the function is activated to select sigmoid. The first self-coding layer is provided with 25 neurons and trained 5000 times, the second self-coding layer is provided with 15 neurons and trained 5000 times, and the back propagation layer is also trained 5000 times. Fig. 5 shows the error between the melt pool ratio and the real data obtained by applying the model, which is a total of 837 groups, it can be seen that the root mean square error is 6% and the correlation coefficient is 0.77.
It should be noted that the above examples are intended to facilitate the understanding and use of the invention by those skilled in the art, and are not intended to be limiting in any way. All equivalent changes or modifications made in accordance with the spirit of the present disclosure are intended to be covered by the scope of the present disclosure. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims (5)

1. A method for inverting the distribution of a molten arctic pool by using an artificial neural network is characterized by comprising the following steps of:
(1) collecting a high-resolution fusion pool image and a contemporaneous medium-resolution imaging spectrometer MOD09GA, wherein the projection mode of the high-resolution fusion pool image is Mokat projection, and the re-projection is sinusoidal projection; the MOD09GA projection is a sinusoidal projection;
(2) constructing a grid consistent with MOD09 GA;
(3) calculating a molten pool proportion, a sea ice proportion and a water opening and discharging proportion in each grid according to the MOD09GA grids, and then deleting pixels covered by the cloud in the MOD09 GA;
(4) counting the data obtained in the step (3) and the corresponding day-by-day earth surface reflectivity, forming a data set by MOD09GA 1-7 wave band reflectivity, and dividing the data set into a training set and a verification set;
(5) and (3) constructing a melting pool model, taking the well matched MOD09GA 1 to 7 wave band reflectivity as an input set, and taking the corresponding sea ice proportion, melting pool proportion and open water proportion as an output sample set to train an artificial neural network to find an optimal solution.
2. The method of claim 1, wherein the grids are projected in a manner and at a resolution consistent with MOD09GA, and each grid corresponds to a MOD09GA reflectivity.
3. The method for inverting the distribution of arctic molten pools by using an artificial neural network as claimed in claim 1, wherein the ratio of the training set to the validation set in step (4) is 7: 3.
4. The method for inverting the arctic molten pool distribution by using the artificial neural network as claimed in claim 1, wherein the artificial neural network of step (5) is a stacked self-coding consisting of a plurality of self-codings; the self-coding comprises a three-layer neural network consisting of an input layer, a hidden layer and an output layer.
5. The method for inverting the arctic molten pool distribution by using the artificial neural network as claimed in claim 1, wherein the objective function of the molten pool model in the step (5) is:
Figure FDA0002269242240000011
wherein the first term is a basic error term, the second term is a regular term, xiTo input samples, yiTo output a sample, nlIs the number of network layers, slAnd sl+1Is the number of neurons in layer l and layer l +1,
Figure FDA0002269242240000012
is i at the l layerThe weight between a neuron and the jth neuron on the l +1 layer is obtained, lambda is a weight attenuation parameter, the learning rate of the stacked self-coding is set to be 0.15, the momentum is set to be 0.5, the output function selects linear, and the activation function selects sigmoid; the first self-coding layer was set with 25 neurons and trained 5000 times, the second with 15 neurons and trained 5000 times, and the back propagation layer was also trained 5000 times.
CN201911098932.3A 2019-11-12 2019-11-12 Method for inverting arctic fusion pool distribution by using artificial neural network Active CN110866363B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201911098932.3A CN110866363B (en) 2019-11-12 2019-11-12 Method for inverting arctic fusion pool distribution by using artificial neural network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201911098932.3A CN110866363B (en) 2019-11-12 2019-11-12 Method for inverting arctic fusion pool distribution by using artificial neural network

Publications (2)

Publication Number Publication Date
CN110866363A true CN110866363A (en) 2020-03-06
CN110866363B CN110866363B (en) 2023-04-07

Family

ID=69653756

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201911098932.3A Active CN110866363B (en) 2019-11-12 2019-11-12 Method for inverting arctic fusion pool distribution by using artificial neural network

Country Status (1)

Country Link
CN (1) CN110866363B (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113553766A (en) * 2021-07-22 2021-10-26 南京信息工程大学 Method for inverting arctic snow depth by using machine learning

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104007443A (en) * 2014-04-24 2014-08-27 南京大学 Method for extracting polar region sea ice melting tank by utilizing spectral feature
CN105116464A (en) * 2015-08-12 2015-12-02 南京大学 Polar sea ice melting pool extraction method based on neural network model
CN109447089A (en) * 2018-10-16 2019-03-08 同济大学 High-resolution Sea Ice Model type-collection method based on oversubscription technology

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104007443A (en) * 2014-04-24 2014-08-27 南京大学 Method for extracting polar region sea ice melting tank by utilizing spectral feature
CN105116464A (en) * 2015-08-12 2015-12-02 南京大学 Polar sea ice melting pool extraction method based on neural network model
CN109447089A (en) * 2018-10-16 2019-03-08 同济大学 High-resolution Sea Ice Model type-collection method based on oversubscription technology

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
MARKO MÄKYNEN等: ""On the Estimation of Melt Pond Fraction on the Arctic Sea Ice With ENVISAT WSM Images"" *
RÖSEL, A.等: ""Melt ponds on Arctic sea ice determined from MODIS satellite data using an artificial neural network"" *

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113553766A (en) * 2021-07-22 2021-10-26 南京信息工程大学 Method for inverting arctic snow depth by using machine learning
CN113553766B (en) * 2021-07-22 2023-04-28 南京信息工程大学 Method for inverting North snow depth by using machine learning

Also Published As

Publication number Publication date
CN110866363B (en) 2023-04-07

Similar Documents

Publication Publication Date Title
CN111310889B (en) Evaporation waveguide profile estimation method based on deep neural network
CN111665575B (en) Medium-and-long-term rainfall grading coupling forecasting method and system based on statistical power
CN114239422B (en) Method for improving marine chlorophyll a concentration prediction accuracy based on machine learning
CN114966692B (en) Transformer-based InSAR technology frozen soil area multivariable time sequence deformation prediction method and device
CN113255972B (en) Short-term rainfall prediction method based on Attention mechanism
CN113240170A (en) Air quality prediction method based on seasonal cyclic neural network
CN111242351A (en) Tropical cyclone track prediction method based on self-encoder and GRU neural network
CN114049545B (en) Typhoon intensity determining method, system, equipment and medium based on point cloud voxels
CN115222163A (en) Multi-factor medium-long term real-time forecasting method and system for harbor basin inlet waves and application
CN115526413A (en) Prediction method based on daily maximum temperature of fully-connected neural network
CN110866363B (en) Method for inverting arctic fusion pool distribution by using artificial neural network
CN114241331B (en) Remote sensing modeling method for ground biomass of reed in wetland by taking UAV as ground and Septinel-2 medium
CN113408644B (en) Satellite data reconstruction method and method for detecting response of upper ocean to typhoon
CN113297904B (en) Method and system for estimating alpine grassland biomass based on satellite driving model
CN104217260A (en) Combined filling system for measured wind speed loss values of multiple neighboring wind motors in wind field
CN117131991A (en) Urban rainfall prediction method and platform based on hybrid neural network
CN116467933A (en) Storm surge water increasing prediction method and system based on deep learning
Wang et al. Study on Remote Sensing of Water Depths Based on BP Artificial Neural Network.
CN104899464A (en) Sampling learning machine remote sensing quantitative inversion method applicable to noise condition
CN115271221A (en) City expansion prediction method, device, medium and equipment
CN113344290A (en) Correcting method for sub-season rainfall weather forecast based on U-Net network
CN113988153A (en) High-resolution aerosol estimation method based on condition generation countermeasure network
Tripathi et al. Southern Indian Ocean SST indices as early predictors of Indian summer monsoon
Yang et al. Simulation System of Lake Eutrophication Evolution based on RS & GIS Technology—a Case Study in Wuhan East Lake
CN112380984B (en) Remote sensing-based salt-biogas vegetation slow-flow capacity space evaluation method

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