A method for predicting Arctic sea ice density based on deep learning

Abstract

The present invention discloses a method for predicting Arctic sea ice density based on deep learning, specifically, step 1: obtaining Arctic sea ice density data and ERA5 meteorological factor data; step 2: preprocessing the meteorological factor data; step 3: splicing the Arctic sea ice density data and meteorological factor data at the same time in the channel dimension to obtain multi-source fusion data; step 4: dividing the multi-source fusion data in step 3 into n groups according to time, and establishing a feature extraction network; the feature extraction network includes n feature extraction modules, step 5: inputting the nth group of multi-source fusion data into the nth feature extraction module of the feature extraction network, extracting the spatiotemporal features of the multi-source fusion data group, and step 6: inputting the output of the feature extraction network into the adaptive weight allocation network to obtain the sea ice density prediction value. The present invention solves the problem that the traditional time series prediction model focuses on the end, and effectively improves the accuracy of the prediction.

Description

Technical Field

[0001] The invention belongs to the technical field of smart ocean. Background Art

[0002] In recent decades, global warming has become a common topic of concern for many countries around the world and the entire scientific community. The Arctic climate system has also undergone a series of changes. According to the future predictions of the CMIP6 model, due to continued anthropogenic warming, the Arctic Ocean may be ice-free for the first time in September 2050 (sea ice area is less than 1 million square kilometers). The melting of Arctic sea ice directly affects human production and life, and also brings unprecedented opportunities for humans to develop Arctic resources. In areas with low sea ice density, the navigable time of waterways can be effectively extended, which is conducive to scientific exploration and commercial development. Therefore, the accuracy of the prediction of sea ice density in the Arctic region is crucial.

[0003] Traditional sea ice density prediction methods are mainly numerical models and statistical models. Numerical models can simulate the formation and evolution of sea ice based on the physical processes of the ocean and atmosphere. These models are based on physical equations and parameterization methods, which require high computing resources and cannot achieve lightweight predictions. Statistical methods require processing each grid one by one. In order to reduce computing costs, researchers reduced the grid resolution of sea ice density by nearly 10 times (from 25km to 225km) to make coarse-grained predictions. Therefore, in practical applications, statistical methods are difficult to support predictions for all grids in the Arctic region. Deep learning models use the powerful learning ability of neural networks to extract features from a large amount of sea ice observation data and meteorological factors, automatically capture the nonlinear relationship between them, and thus make accurate predictions. There are currently many deep learning-based methods for predicting sea ice density, but these methods still face some challenges. Current sea ice density prediction models have some limitations in capturing the spatiotemporal changes of sea ice, resulting in large errors and low accuracy. One of the main problems is that the model does not take into account meteorological factors, which play an important role in the melting and freezing of sea ice. The lack of comprehensive consideration of meteorological factors can lead to instability in forecast results. In addition, deep learning-based models usually require a lot of computing resources and long training time, which makes them demanding on equipment. This is a challenge for lightweight forecasting or for deploying models in resource-constrained environments. Summary of the invention

[0004] Purpose of the invention: In order to solve the problems existing in the above-mentioned prior art, the present invention provides a method for predicting Arctic sea ice density based on deep learning.

[0005] Technical solution: The present invention provides a method for predicting Arctic sea ice density based on deep learning, which specifically includes the following steps:

[0006] Step 1: Obtain Arctic sea ice density data and ERA5 meteorological factor data;

[0007] Step 2: Preprocess the meteorological factor data to align the meteorological factor data with the Arctic sea ice density data;

[0008] Step 3: Splice the Arctic sea ice density data and meteorological factor data at the same time in the channel dimension to obtain multi-source fusion data;

[0009] Step 4: Divide the multi-source fusion data in step 3 into n groups according to time, and establish a feature extraction network; the feature extraction network includes n feature extraction modules, one feature extraction module corresponds to a group of multi-source fusion data groups, and each group of multi-source fusion data groups is input into the corresponding feature extraction module in time sequence;

[0010] Step 5: Input the t-th group of multi-source fusion data into the t-th feature extraction module of the feature extraction network to extract the spatiotemporal features of the multi-source fusion data; t = 1, 2, ..., n;

[0011] Step 6: Input the output of the feature extraction network into the adaptive weight allocation network to obtain the predicted value of sea ice density.

[0012] Furthermore, the ERA5 meteorological factors are monthly reanalysis data, including 2m temperature, 2m dew point temperature, average sea surface temperature, and u and v components of 10m wind.

[0013] Furthermore, in step 2, the polar projection method is used to preprocess the meteorological factor data.

[0014] Furthermore, in step 4, the multi-source fusion data is grouped according to months with one year as a cycle to obtain multi-source fusion data groups for 12 consecutive months.

[0015] Furthermore, the feature extraction module includes 7 layers of feature extraction layers mainly based on ConvLSTM, and the expression is:

[0016] L1=MaxPool{ConvLSTM(x t )}

[0017] L2=MaxPool{ConvLSTM(L1)}

[0018] L3=MaxPool{ConvLSTM(L2)}

[0019] L4=ConvLSTM(L3)

[0020] L5=ConvLSTM{UpSample(L4+L3)}

[0021] L6=ConvLSTM{UpSample(L5+L2)}

[0022] L7=ConvLSTM{UpSample(L6+L1)}

[0023] Among them, L l represents the output features of the lth feature extraction layer, l = 1, 2, 3…, 7, x t It represents the input data of the t-th feature extraction module, ConvLSTM represents the convolutional long short-term memory network, MaxPool represents the maximum pooling, UpSample represents upsampling, and the ConvLSTM units in the same feature extraction layer of two adjacent feature extraction modules are connected to each other.

[0024] Furthermore, a context module is introduced into the ConvLSTM of each feature extraction layer, and the context module includes: a global pooling layer, a Softmax layer and a weighted summation layer; the structure of the ConvLSTM in the first feature extraction layer is as follows:

[0025] f t =σ(W f *[H t-1 ,L l-1 ]+b f )

[0026] i t =σ(W i *[H t-1 ,L l-1 ]+b i )

[0027] c t '=tanh(W c *[H t-1 ,L l-1 ]+b c )

[0028] o t =σ(W o *[H t-1 ,L l-1 ]+b o )

[0029] C t =f t *C t-1 +i t *c t '

[0030] H l=o t *tanh(c t ')

[0031] H t '=W x *[H t-1 ,L l-1 ]

[0032]

[0033] Among them, t represents the t-th feature extraction module, σ represents the sigmoid activation function, Tanh represents the hyperbolic tangent activation function, and f t represents the output of the forget gate, b f The bias term of the forget gate, W f represents the weight of the forget gate, L l-1 H represents the output of the l-1th feature extraction layer of the tth feature extraction module, t-1 H represents the output of ConvLSTM in the lth feature extraction layer of the t-1th feature extraction module, l represents the output of ConvLSTM in the lth feature extraction layer of the tth feature extraction module, i t Represents the output result of the input gate, W i represents the weight of the input gate, b i represents the bias term of the input gate, c t ' represents the output result of the cell state, W c represents the weight of the cell state, b c represents the cell state bias term, o t Represents the output result of the output gate, W o represents the weight of the output gate, b o represents the bias term of the output gate, C t Represents the output cell state information, C t-1 represents the cell state output by the lth feature extraction layer of the t-1th feature extraction module, Softmax is normalized, H t ' indicates a temporary variable, W x To calculate the weight of temporary variables, H t ' k Indicates H t ' is the data on the kth channel dimension, GAP represents the global average pooling, and ∑ represents the weighted sum operation.

[0034] Furthermore, the expression of the adaptive weight allocation network is as follows:

[0035]

[0036] Among them, t represents the tth feature extraction module, h t It represents the output of the t-th feature extraction module, Concat represents the concatenation operation, GAP represents the global average pooling, MLP represents the multi-layer perceptron, Softmax represents the normalization, and ∑ represents the weighted sum operation.

[0037] Beneficial effects:

[0038] First, the present invention makes full use of the influence of meteorological factors on sea ice movement, and uses five major meteorological factors, namely, 2m temperature, 2m dew point temperature, average sea surface temperature, and u and v components of 10m wind, to assist in predicting sea ice density, thereby improving the predictability of sea ice movement;

[0039] Second, the present invention uses a feature map scaling mechanism in the feature extraction module, which greatly reduces the consumption of computing resources. Under the same computing resources, this model can use more meteorological factors to predict sea ice density;

[0040] Third, the adaptive weight allocation module proposed in the present invention utilizes the periodic laws of meteorology to solve the problem that the traditional time series prediction model focuses on the end, and effectively improves the accuracy of the prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 is a flow chart of the present invention;

[0042] Figure 2 It is the AS-LSTM network model structure diagram of the present invention;

[0043] Figure 3 It is a structural diagram of the feature extraction network of the present invention;

[0044] Figure 4 It is a structural diagram of the feature extraction module of the present invention;

[0045] Figure 5 The internal structure diagram of the ConvLSTM with context module of the present invention;

[0046] Figure 6 This is a comparison chart of the MSE experiments for the monthly prediction of sea ice density for the next month;

[0047] Figure 7 This is a comparison chart of MAE experiments for monthly prediction of sea ice density for the next month;

[0048] Figure 8 This is a comparison chart of the SSIM experiments for the monthly prediction of sea ice density for the next month;

[0049] Fig. 9 This is a comparison chart of PSNR experiments for monthly prediction of sea ice density for the next month. DETAILED DESCRIPTION

[0050] The accompanying drawings, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0051] like Figure 1 As shown, the present invention proposes a method for predicting Arctic sea ice density based on deep learning; the overall architecture of the AS-LSTM network model of the present invention is as follows Figure 2 As shown in FIG, the network includes an input module for fusing multi-source data, a feature extraction module based on ConvLSTM, an adaptive weight allocation module constructed according to meteorological periodicity, and an output module.

[0052] The method of this embodiment includes the following steps:

[0053] S1, obtain the Arctic sea ice density monthly data and ERA5 meteorological factor data set; the Arctic sea ice density monthly data is the monthly average Arctic sea ice density data, which is provided by NSIDC, and the time range of the data is from January 1979 to December 2022, with a total of 528 NC data, and the original area of ​​the data is 31.1°N to 89.84°N, 180°E to -180°W, with a spatial resolution of 25km×25km; the official sea ice density data is the result after polar projection, with a size of 304×448; the ERA5 meteorological factors are provided by ECMWF, with a time range from 1950 to the present, and monthly reanalysis data are selected, with a horizontal resolution of 0.25°×0.25°, including 2m temperature, 2m dew point temperature, sea surface average temperature, and 10m wind u component and v component.

[0054] S2, preprocess the meteorological factor data. The initial meteorological factor data size is 721×1440, and the projection method is equal longitude and latitude. In order to align the meteorological factor with the sea ice density data in spatial position, the meteorological factor needs to be projected and transformed. The present invention uses the polar projection method and cuts off the redundant data. The retained data size is 304×448, which is consistent with the sea ice density data size in S1 and corresponds one to one in geographical location.

[0055] S3, based on the preprocessing performed in step S2, the multi-source fusion data is formed by splicing the sea ice density and the meteorological factors, and the data size becomes 304×448×6. The multi-source fusion data is obtained by the following formula:

[0056] x=Concat(sic,t2m,d2m,sst,10u,10v)

[0057] Among them, sic represents sea ice concentration data, t2m represents 2-meter temperature, d2m represents 2-meter dew point temperature, sst represents the average sea surface temperature, 10u represents the u component of 10-meter wind, 10v represents the v component of 10-meter wind, Concat represents the splicing operation, and x represents the obtained multi-source fusion data.

[0058] S4, according to the periodic law of meteorology, this embodiment selects multi-source fusion data of 12 consecutive months as input data with a year as a cycle, and inputs them into the feature extraction network in chronological order; the feature extraction network of the present invention

[0059] like Figure 3 As shown in the figure, it includes 12 feature extraction modules. The main part of the network is a 7-layer 12-step ConvLSTM. Corresponding upsampling or downsampling operations are performed between each feature extraction layer, and the data size of each layer is scaled accordingly. The above 12 multi-source fusion data are input into the corresponding feature extraction modules in chronological order, and finally output the spatiotemporal feature information of the 12 feature extraction modules after passing through the feature extraction network. The operation formula of feature extraction is:

[0060] h1,...,h t ...,h 12 =Extract(x1,...,x t ...,x 12 )

[0061] Extract represents the feature extraction network, x t represents the multi-source fusion data at time t, that is, the t-th multi-source fusion data, h t Represents x t The spatiotemporal feature information obtained after the feature extraction module; the input of the feature extraction network is 12 multi-source fusion data, and the output is 12 spatiotemporal feature information.

[0062] The main body of the feature extraction module is a ConvLSTM multi-layer structure. The multi-source fusion data input to the feature extraction module undergoes 7 layers of feature extraction layers based on the ConvLSTM model. The feature map scaling mechanism exists in the feature extraction layer. The structure diagram of the feature extraction module is shown in the figure. Figure 4 As shown, the overall operation of the feature extraction module is:

[0063] L1=MaxPool{ConvLSTM(x t )}

[0064] L2=MaxPool{ConvLSTM(L1)}

[0065] L3=MaxPool{ConvLSTM(L2)}

[0066] L4=ConvLSTM(L3)

[0067] L5=ConvLSTM{UpSample(L4+L3)}

[0068] L6=ConvLSTM{UpSample(L5+L2)}

[0069] L7=ConvLSTM{UpSample(L6+L1)}

[0070] Among them, L l represents the output features of the lth feature extraction layer, l = 1, 2, 3…, 7, x t Represents the input data of the t-th feature extraction module, ConvLSTM represents a convolutional long short-term memory network, its convolution kernel size is 3×3, MaxPool is maximum pooling, the filter size is 2×2, UpSample is upsampling, the scaling factor is 2, and the ConvLSTM units in the same feature extraction layer of two adjacent feature extraction modules are connected to each other.

[0071] The first layer obtains data of size 152×224 and sends it to the seventh layer through a skip connection; the second layer obtains data of size 76×112 and sends it to the sixth layer through a skip connection; the third layer obtains data of size 38×56 and sends it to the fifth layer through a skip connection.

[0072] There is a context module in each ConvLSTM to solve the information loss problem caused by long sequences; it consists of a global pooling layer, a Softmax layer, and a weighted sum layer. The structure of a single ConvLSTM is as follows: Figure 5 As shown, the formula is as follows:

[0073] f t =σ(W f *[H t-1 ,L l-1 ]+b f )

[0074] i t =σ(W i *[H t-1 ,L l-1 ]+b i )

[0075] c t '=tanh(W c *[H t-1 ,L l-1 ]+b c )

[0076] ot =σ(W o *[H t-1 ,L l-1 ]+b o )

[0077] C t =f t *C t-1 +i t *c t '

[0078] H l =o t *tanh(c t ')

[0079] H t '=W x *[H t-1 ,L l-1 ]

[0080]

[0081] Among them, t represents the t-th feature extraction module, σ represents the sigmoid activation function, Tanh represents the hyperbolic tangent activation function, and f t represents the output of the forget gate, b f The bias term of the forget gate, W f represents the weight of the forget gate, L l-1 H represents the output of the l-1th feature extraction layer of the tth feature extraction module, t-1 H represents the output of ConvLSTM in the lth feature extraction layer of the t-1th feature extraction module, l represents the output of ConvLSTM in the lth feature extraction layer of the tth feature extraction module, i t Represents the output result of the input gate, W i represents the weight of the input gate, b i represents the bias term of the input gate, c t ' represents the output result of the cell state, W c represents the weight of the cell state, b c represents the cell state bias term, o t Represents the output result of the output gate, W o represents the weight of the output gate, b o represents the bias term of the output gate, C t Represents the output cell state information, C t-1 represents the cell state output by the lth feature extraction layer of the t-1th feature extraction module, Softmax is normalized, H t ' indicates a temporary variable, W xTo calculate the weight of temporary variables, H t ' k Indicates H t ' is the data on the kth channel dimension, GAP represents the global average pooling, and ∑ represents the weighted sum operation.

[0082] S5, the adaptive weight allocation module receives 12 spatiotemporal feature information from the feature extraction network, concatenates them and sends them to the global pooling layer to generate 12 weights, and then enters the multi-layer perceptron layer and the Softmax layer to obtain 12 normalized weight values; each weight value is weighted and summed with the feature information at the corresponding time to obtain the sea ice density prediction data. The specific structure is as follows: Figure 2 As shown, the formula of the adaptive weight allocation module is:

[0083]

[0084] Among them, Concat represents data concatenation, GAP represents global average pooling, Softmax is the normalization layer, MLP is a multi-layer perceptron, which is used to adaptively learn the 12 input weights, and ∑ represents the weighted summation of each weight value and its corresponding feature information at the corresponding moment.

[0085] Experimental plan and results

[0086] The latitude of the research area selected by the present invention is 31.1°N to 89.84°N, and the longitude is 180°E to 180°W. It covers the entire Arctic Ocean, and the data size is 304×448. The present invention uses the data from January 1979 to December 2014 as the training set, and the data from January 2015 to December 2022 as the validation set to predict the monthly average data of sea ice density in the next month.

[0087] In order to verify the impact of each module in the present invention on the experiment, this experiment added the feature map scaling mechanism and context module (S-LSTM), adaptive weight allocation module (AS-LSTM), ERA5 meteorological factor data (AS-LSTM with meteorological factor) to the traditional ConvLSTM network, and conducted an ablation experiment. The selected evaluation indicators are MSE, MAE, PSNR, and SSIM. At the same time, the video memory size required for the experiment is also compared. The experimental results are shown in Table 1:

[0088] Table 1

[0089]

[0090]

[0091] The experimental results show that by adding modules or meteorological factors, the model effect is getting better and better. Compared with the original ConvLSTM, the AS-LSTM model proposed in this invention has improved MSE and MAE by 17.4% and 9.1% respectively. SSIM has increased from 0.951 to 0.960, PSNR has also increased from 26.60 to 27.32, and the video memory usage has been reduced by 24.0%.

[0092] AS-LSTM was compared with ConvLSTM, PredRNN, and CNN to predict the sea ice density data for the next month. The experimental results are shown in Table 2:

[0093] Table 2

[0094] Model MSE(%) MAE(%) SSIM PSNR Video Memory(Mb) CNN 0.28 1.17 0.946 25.66 Null ConvLSTM 0.23 0.99 0.951 26.60 8432 PredRNN 0.21 0.97 0.954 27.17 10647 AS-LSTM 0.19 0.90 0.960 27.32 6410

[0095] The results show that in the single-step prediction experiment, the model of the present invention has achieved a certain degree of leadership in the four evaluation indicators of MSE, MAE, SSIM, and PSNR, and the video memory occupancy of the present invention is also the lowest.

[0096] The results of the experiment are plotted as a line graph. Figure 6 ,7,8,9, it can be clearly seen that other models are prone to significant errors in the early stages of ice formation. The method of the present invention adopts an innovative method of feature map scaling mechanism, which significantly enhances the ability to obtain multi-scale spatiotemporal information. By combining meteorological physical elements, AS-LSTM can better capture the complex changes of sea ice in the early stages of ice formation and make more accurate predictions. Therefore, the model of the present invention did not show a significant increase in errors in the early stages of ice formation. This result further verifies the stability and reliability of AS-LSTM in short-term sea ice forecasting (the next month).

[0097] It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. To avoid unnecessary repetition, the present invention will not further describe various possible combinations.

Claims

1. A method for predicting Arctic sea ice density based on deep learning, characterized in that: The specific steps include: Step 1: Obtain Arctic sea ice density data and ERA5 meteorological factor data; Step 2: Preprocess the meteorological factor data to align the meteorological factor data with the Arctic sea ice density data; Step 3: Combine the Arctic sea ice density data and meteorological factor data at the same time in the channel dimension to obtain multi-source fusion data; Step 4: Divide the multi-source fusion data in step 3 into n groups according to time and establish a feature extraction network; The feature extraction network includes n feature extraction modules, one feature extraction module corresponds to a group of multi-source fusion data sets, and each group of multi-source fusion data sets is input into the corresponding feature extraction module in chronological order; Step 5: Input the t-th group of multi-source fusion data into the t-th feature extraction module of the feature extraction network to extract the spatiotemporal features of the multi-source fusion data; t = 1, 2, ..., n; Step 6: Input the output of the feature extraction network into the adaptive weight allocation network to obtain the predicted value of sea ice density; The feature extraction module includes 7 layers of feature extraction layers mainly based on ConvLSTM. The expression of the feature extraction module is: L1=MaxPool{ConvLSTM(x t )} L2=MaxPool{ConvLSTM(L1)} L3=MaxPool{ConvLSTM(L2)} L4=ConvLSTM(L3) L5=ConvLSTM{UpSample(L4+L3)} L6=ConvLSTM{UpSample(L5+L2)} L7=ConvLSTM{UpSample(L6+L1)} Among them, L l represents the output features of the lth feature extraction layer, l = 1, 2, 3…, 7, x t represents the input data of the t-th feature extraction module, ConvLSTM represents the convolutional long short-term memory network, MaxPool represents the maximum pooling, UpSample represents upsampling, and the ConvLSTM units in the same feature extraction layer of two adjacent feature extraction modules are connected to each other; A context module is introduced into the ConvLSTM of each feature extraction layer, and the context module includes: a global pooling layer, a Softmax layer, and a weighted summation layer; the structure of the ConvLSTM in the first feature extraction layer is as follows: f t =σ(W f *[H t-1 ,L l-1 ]+b f ) I t =σ(W i *[H t-1 ,L l-1 ]+b i ) c t '=tanh(W c *[H t-1 ,L l-1 ]+b c ) the t =σ(W o *[H t-1 ,L l-1 ]+b o ) C t =f t *C t-1 +i t *c t ' H l =o t *tanh(c t ') H t '=W x *[H t-1 ,L l-1 ] Among them, t represents the t-th feature extraction module, σ represents the sigmoid activation function, Tanh represents the hyperbolic tangent activation function, and f t represents the output of the forget gate, b f The bias term of the forget gate, W f represents the weight of the forget gate, L l-1 H represents the output of the l-1th feature extraction layer of the tth feature extraction module, t-1 H represents the output of ConvLSTM in the lth feature extraction layer of the t-1th feature extraction module, l represents the output of ConvLSTM in the lth feature extraction layer of the tth feature extraction module, i t Represents the output result of the input gate, W i represents the weight of the input gate, b i represents the bias term of the input gate, c t ' represents the output result of the cell state, W c represents the weight of the cell state, b c represents the cell state bias term, o t Represents the output result of the output gate, W o represents the weight of the output gate, b o represents the bias term of the output gate, C t Represents the output cell state information, C t-1 represents the cell state output by the lth feature extraction layer of the t-1th feature extraction module, Softmax is normalized, H t ' indicates a temporary variable, W x To calculate the weight of temporary variables, H t ' k Indicates H t ' is the data on the kth channel dimension, GAP represents the global average pooling, and ∑ represents the weighted sum operation.

2. The method for predicting Arctic sea ice density based on deep learning according to claim 1, characterized in that: The ERA5 meteorological factors are monthly reanalysis data, including 2m temperature, 2m dew point temperature, average sea surface temperature, and u and v components of 10m wind.

3. The method for predicting Arctic sea ice density based on deep learning according to claim 1, characterized in that: In step 2, the polar projection method is used to preprocess the meteorological factor data.

4. The method for predicting Arctic sea ice density based on deep learning according to claim 1, characterized in that: In step 4, the multi-source fusion data is grouped according to months with one year as the cycle to obtain multi-source fusion data groups of 12 consecutive months.

5. The method for predicting Arctic sea ice density based on deep learning according to claim 1, characterized in that: The expression of the adaptive weight allocation network is as follows: Among them, t represents the tth feature extraction module, h t It represents the output of the t-th feature extraction module, Concat represents the concatenation operation, GAP represents the global average pooling, MLP represents the multi-layer perceptron, Softmax represents the normalization, and ∑ represents the weighted sum operation.

Images