North pole airline environment risk spatio-temporal prediction method and system based on deep learning

Abstract

The invention relates to the technical field of north region ship navigation environment risk prediction application, and provides a north region ship navigation environment risk space-time prediction method and system based on deep learning, and the method comprises the steps: obtaining time sequence sea ice concentration data and sea ice thickness data in a recent T-day region, and processing the data into a T * X * W form; generating navigation risk index result distribution data based on the sea ice concentration data and the sea ice thickness data; performing parameter iteration on the initial space-time prediction model to obtain a space-time prediction model; inputting the navigation risk index result distribution data into the space-time prediction model, and outputting a grid map for risk prediction; according to the method, the deep neural network of the ViTs-CNN-CNN architecture is provided and is applied to the space-time prediction task, the calculation complexity of the Swin Transformer is remarkably reduced through a hierarchical structure and a local window self-attention mechanism, the efficiency is improved, and an excellent effect is achieved.

Description

Technical Field

[0001] The present invention belongs to the field of application technology of environmental risk prediction for ship navigation in the Arctic region, and proposes a spatiotemporal prediction method and system for environmental risks of Arctic routes based on deep learning. Background Art

[0002] In recent years, the geographical environment of the Pan-Arctic Circle has undergone profound changes, presenting new opportunities and challenges for humanity, a process that will continue for the foreseeable future. The window for access to Arctic shipping routes may be significantly extended, particularly the Northeast Passage connecting East Asia and Europe, which could significantly reduce my country's reliance on key international shipping routes such as the Strait of Malacca and the Suez Canal. While the navigability of Arctic shipping routes has significantly improved in recent years, safety remains a paramount concern, and effective prediction of risk distribution is a crucial means of ensuring this.

[0003] Currently, there are two approaches to risk prediction: (1) predicting risk factors based on historical risk data on sea ice density and sea ice thickness, and then calculating risk based on the results of each factor. This approach is highly interpretable. (2) first calculating a historical risk dataset based on historical risk data on sea ice density and sea ice thickness, and then making point-to-point predictions based on the risk dataset. While the first approach is highly interpretable and easy to imagine, the second approach is more effective based on the overall prediction results.

[0004] Despite the significant progress in deep learning-based sea ice concentration prediction, research on the spatiotemporal prediction of risk distribution remains lacking. Existing deep learning-based spatiotemporal prediction technologies generally employ one or more of three architectures: convolutional neural networks (CNNs), recurrent neural networks (RNNs), and attention mechanisms. These can be broadly categorized into four types based on the network type and structure used. The first is the RNN-RNN-RNN architecture, which offers excellent flexibility and accuracy and is often used as a benchmark for spatiotemporal prediction tasks. This architecture typically consists of a stack of recurrent neural networks (RNNs), with key models including PredRNN and MIM-LSTM. The second is the CNN-RNN-CNN architecture, which leverages RNNs to capture temporal features and CNNs to capture spatial features, combining the advantages of both recurrent and convolutional neural networks. Representative models of this architecture include ConvLSTM, VRNN, and PhyDNet. The third type is CNN-ViTs-CNN. ViTs refers to various attention mechanisms evolved from the Vision Transformer, including TimesFormer, ViViT, and MViT. The fourth type is the CNN-CNN-CNN structure, such as PredCNN and DPG. Although traditional CNN and RNN structures can process spatiotemporal data, they are computationally expensive, especially for long time series or large-scale datasets. Summary of the Invention

[0005] To solve the above problems, the present invention provides a neural network model with a ViTs-CNN-CNN framework to achieve direct prediction of Arctic navigation risks. The method includes:

[0006] Using satellite remote sensing to obtain time series sea ice concentration data and sea ice thickness data over a recent T-day period within a region, and processing the data into a T*X*W format, where T is the number of days and X and W are the sizes of the data grid; the processed time series sea ice concentration data and sea ice thickness data have a spatiotemporal relationship;

[0007] Determine the sea ice type based on sea ice density data and sea ice thickness data, calculate the risk index result according to the sea ice type and the corresponding risk index value, and generate the distribution data of the navigation risk index result based on the regional grid;

[0008] Inputting the navigation risk index result distribution data into a spatiotemporal prediction model and outputting a grid map of the predicted navigation risk index;

[0009] The spatiotemporal prediction model includes an encoder, a residual link module and a decoder;

[0010] The encoder includes a 3D window attention mechanism and a 3D sliding window attention mechanism, which is used to first capture the spatiotemporal meteorological condition characteristics of the local window through the 3D window attention mechanism, and then fuse the spatiotemporal meteorological condition characteristics of the local window through the 3D sliding window attention mechanism to capture the global risk characteristics of the route;

[0011] The residual link module is used to transfer the spatiotemporal meteorological condition features extracted from the encoder to the decoder;

[0012] The decoder is used to convert the input feature map into a risk prediction grid map.

[0013] Inputting a sample set of a historical sea ice density dataset into an initial spatiotemporal prediction model to obtain a risk prediction grid map predicted by the sample set; the historical sea ice density dataset includes a sample set and a label set corresponding to the sample set;

[0014] Based on the difference between the risk prediction grid map predicted by the sample set and the label set image corresponding to the sample set, the parameters of the initial spatiotemporal prediction model are iterated to obtain a spatiotemporal prediction model.

[0015] The encoder comprises:

[0016] A patch partition module is used to divide the navigation risk index results into blocks and to divide the original data with spatial and temporal characteristics into local window areas;

[0017] A linear embedding module is used to embed the data block from a high-dimensional representation to a low-dimensional representation;

[0018] The 3D-swim transformer block module is used to perform attention operations in both spatial and temporal dimensions to capture the spatial distribution patterns and temporal variation trends of sea ice data;

[0019] The path merging module is used to aggregate and condense high-level information in the image, integrating local spatiotemporal meteorological condition features into global risk characteristics of the route.

[0020] The residual link module consists of at least two residual blocks, which directly add the input to the output through a skip connection, and is used to alleviate the gradient vanishing and gradient exploding problems in deep networks.

[0021] The decoder comprises at least two upsampling modules;

[0022] The upsampling module includes:

[0023] The Pixelshuffle module is used to rearrange a low-resolution feature map into multiple channel information into spatial dimensions to improve the output resolution;

[0024] CNN block module, used to extract feature information;

[0025] The prediction layer is used to convert the high-resolution feature map into a raster map for risk prediction.

[0026] The decoder also includes a residual connection structure, which is arranged between the prediction layer and the upsampling module, so that the spatiotemporal prediction model learns the residual between the input and output, thereby improving the stability of the model.

[0027] The step of calculating the risk index result based on the sea ice concentration data and the sea ice thickness data according to the ice type and the corresponding risk index value, and generating the navigation risk index result distribution data specifically includes:

[0028] Spatially interpolate sea ice concentration data to align them with sea ice thickness data;

[0029] According to the corresponding relationship between ice type definition and risk index value, the sea ice thickness data is mapped to the risk index value;

[0030] Based on the sea ice density data, the risk index results are calculated according to the ice type and the corresponding risk index value, and the navigation risk index result distribution data is generated.

[0031] In a second aspect, the present invention further provides a deep learning-based spatiotemporal prediction system for environmental risks in Arctic shipping routes, comprising:

[0032] The data acquisition and processing module is used to record the sea ice density element values ​​of each coordinate grid in the daily data recording area, collect the recent T-day time series Arctic sea ice density data, and process it into the format of T*X*W, where T is the number of days, H and W are the size of the data grid, obtain the daily spatial distribution data of the navigation risk ROI and process it into a standardized format;

[0033] A model building module for implementing the VCCNet spatiotemporal prediction network model based on decomposition and sparse spatiotemporal attention mechanism based on the Pytorch deep learning framework;

[0034] The model training module is used to train the model based on the historical daily sea ice concentration data. According to the observation data and prediction requirements, the parameters of the VCCNet model are adjusted. The parameters include the number of prediction days and the spatial resolution of the sea ice concentration prediction results.

[0035] The navigation risk ROI spatiotemporal prediction module is used to predict the spatiotemporal distribution of future navigation risk ROI. The daily navigation risk ROI data of the most recent day A obtained in real time is input as parameters into the trained spatiotemporal prediction model to achieve the spatiotemporal prediction of the navigation risk ROI of the next day B.

[0036] The present invention has the following beneficial effects:

[0037] First, this invention has achieved for the first time the end-to-end direct prediction of navigation risk distribution based on deep learning technology. It obtains spatiotemporal predictions based on point-to-point predictions of risk data sets, rather than predicting risk factors separately based on sea ice density and sea ice thickness and then calculating risks. This can provide information support for future decision-making of merchant ships sailing in the Arctic region and can be extended to other types of merchant ships.

[0038] Second, the Video Swin Transformer based on the classic Transformer framework was used as the encoder, and a deep neural network with the ViTs-CNN-CNN architecture was proposed and applied to spatiotemporal prediction tasks. The Swin Transformer significantly reduced computational complexity and improved efficiency through its hierarchical structure and local window self-attention mechanism, achieving excellent results.

[0039] Third, we designed a hierarchical nested residual structure. The first layer consists of skip connections spanning the entire model, adding the original data to the last CNN module. The second layer consists of nested residual modules. This design ensures stable training. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 This is a flow chart of a method for spatiotemporal prediction of environmental risks of Arctic shipping routes in one embodiment of the present invention;

[0041] Figure 2 This is a diagram showing the architecture of a spatiotemporal prediction model for environmental risks on Arctic shipping routes in one embodiment of the present invention;

[0042] Figure 3 This is a flow chart of the basic modules in the spatiotemporal prediction model for environmental risks of Arctic routes in one embodiment of the present invention;

[0043] Figure 4 This is a diagram showing the structure of an encoder in a spatiotemporal prediction model for environmental risks on Arctic routes according to an embodiment of the present invention;

[0044] Figure 5 This is a structural diagram of the nested residual module in the spatiotemporal prediction model for environmental risks of Arctic routes in one embodiment of the present invention;

[0045] Figure 6 This is a diagram showing the structure of a decoder in the spatiotemporal prediction model for environmental risks on Arctic routes in one embodiment of the present invention;

[0046] Figure 7 This is an example diagram of the prediction effect after using the spatiotemporal prediction method for environmental risks of Arctic routes in one embodiment of the present invention. DETAILED DESCRIPTION

[0047] The best mode for carrying out the present invention will be described below by way of examples. It should be understood that the specific embodiments herein are intended to explain the present invention in detail and should not be construed as limiting the present invention. It should be noted that various changes and modifications may be made while adhering to the principles and core scope of the present invention, and these changes should all be considered to fall within the scope of protection of the present invention. The specific implementation steps of the present invention will be described in detail with reference to the accompanying drawings.

[0048] The present invention provides a spatiotemporal prediction model for a spatiotemporal prediction method. In a specific embodiment of the present invention, the method is implemented based on the spatiotemporal prediction model, wherein the encoder of the spatiotemporal prediction model is as follows: Figure 4 As shown, including:

[0049] The patch partition module is used to partition the processed three-dimensional time series sea ice density data into blocks and to divide the original data into local window regions;

[0050] A linear embedding module is used to embed the data block from a high-dimensional representation to a low-dimensional representation;

[0051] The 3D-swim transformer block module is used to perform attention operations in both spatial and temporal dimensions, thereby capturing complex spatiotemporal dependencies in three-dimensional space;

[0052] The path merging module is used to aggregate and condense high-level information in images.

[0053] The Swin Transformer block is a feature extraction and downsampling module. Figure 3 As shown in (a) and (b), it consists of two consecutive transformers. The process of one feature extraction and downsampling cycle is as follows:

[0054] x′ l =3DWMSA(LN(x l-1 ))+x l-1

[0055] x l =FFN(LN(x′ l ))+x′ l

[0056] x′ l+1 =3DSWMSA(LN(xl ))+x l

[0057] x l+1 =FFN(LN(x′ l+1 ))+x′ l+1

[0058] x l+1 =FFN(LN(x′ l+1 ))+x′ l+1

[0059] x=Patch(x l+1 )

[0060] where x l-1 Represents the input data, x′ l represents the data processed by 3D window attention, x l Represents the data processed by the entire window attention module, x′ l+1 represents the data processed by the sliding 3D window attention, x l+1 represents the data processed by the entire sliding window attention module; x represents the downsampled data; LN represents layer normalization, FFN represents feedforward neural network, 3DWMSA and 3DSWMSA represent 3D window attention and 3D sliding window attention, respectively.

[0061] Input data x l-1 After normalization and 3D window attention processing, features are learned in a single window, the attention weight of a single window feature is calculated, and the residual link method is used to obtain x′ l , and then through normalization and MLP, combined with the residual link to get x l , x l After normalization and the sliding window attention module, it slides between windows to cover all features of the entire input, and then obtains x′ through the residual link l+1 , and finally after normalization and MLP modules, combined with residual links to output the final feature map x l+1 .

[0062] The attention can be expressed as:

[0063]

[0064] Among them, Q, K, V are matrices that need to be learned, o is a nonlinear activation function, and B is a bias matrix. is the adjustment factor.

[0065] The nested residual module is as follows Figure 5As shown in the figure, it consists of several plug-and-play residual blocks, which add the input directly to the output through skip connections, alleviating the gradient disappearance and gradient explosion problems in deep networks.

[0066] In a specific embodiment, the nested residual module refers to performing a residual connection operation on three basic residual blocks, combining the input data with the data processed by the three residual blocks, and forming a nested residual connection with the basic residual block.

[0067] Among them, a basic residual block is expressed as follows:

[0068] x i =GN(Cov(σ(GN(Cov(x i-1 )))))+x i-1

[0069] where x i-1 is the data of the input residual block, x i The output data of the residual block, Cov is a convolutional layer with a filter size of 3*3, which is used to extract only the matrix features after the attention calculation, GN is a group normalization layer, and o is a nonlinear activation function. The basic residual block is designed with residual connections.

[0070] like Figure 3 As shown in (c), the data x is input to the residual module. i-1 First, convolution is performed to extract features, and then the mean and variance of each group are calculated through GroupNorm, and normalization is performed. After the leakyrelu activation function, negative value information can be retained to avoid the problem of neuron death. Then, convolution and group normalization are combined with residual links to combine the processed data with the input data x i-1 Add them together to get x i .

[0071] The decoder module is as follows Figure 6 As shown, it is an upsampling module with PixelShuffle as the core, including:

[0072] The Pixelshuffle module is used to rearrange a low-resolution feature map into multiple channel information into spatial dimensions to improve the output resolution;

[0073] CNN block module, used to extract feature information;

[0074] The prediction layer is used to convert the high-resolution feature map into a raster map for risk prediction.

[0075] Among them, the formula of the upsampling module is expressed as follows:

[0076] x z=σ(GN(Cov(Pixel(x z-1 ))))

[0077] where x z-1 is the data input to the upsampling module, x z It is the data of the output upsampling module, Pixel represents the upsampling layer, Cov is the convolution layer, GN is the group normalization layer, and o is the nonlinear activation function.

[0078] like Figure 3 As shown in (d), the input data x z-1 First, the pixel is upsampled by the pixel shuffle operation, and then it is sequentially subjected to convolution, group normalization, and leaky relu activation function operations.

[0079] The present invention provides a deep learning-based method for predicting Arctic ship navigation risks. First, the Polar Operational Restrictions Assessment Risk Index System (POLARIS) is used to calculate the risk index result (RIO). Then, a neural network is trained to predict the RIO for the next 8 days. The method specifically includes the following steps:

[0080] Step 1: Obtain time series sea ice concentration data and sea ice thickness data in the region over the recent T days, and process them into the form of T*X*W, where T is the number of days, and H and W are the sizes of the data grid; the processed time series sea ice concentration data has a temporal and spatial relationship;

[0081] In specific embodiments, the data may be obtained through satellite remote sensing or weather and ocean observation stations.

[0082] Step 2: Based on the sea ice concentration data and sea ice thickness data, calculate the risk index results according to the ice type and the corresponding risk index value, and generate the navigation risk index result distribution data;

[0083] In a specific embodiment, step 2 includes the following steps:

[0084] S201: spatially interpolate the sea ice density data to align them with the sea ice thickness data;

[0085] In a specific embodiment, the sea ice concentration data was preprocessed to an initial spatial resolution of 25 km x 25 km. Spatial interpolation was then performed to increase the spatial resolution to 12.5 km x 12.5 km, and the data was spatially aligned with the sea ice thickness data. Based on the characteristics of the data, the nearest neighbor interpolation method was used.

[0086] S202: Mapping sea ice thickness data to risk index values ​​according to the correspondence between ice type definitions and risk index values;

[0087] The sea ice thickness data was pre-processed and mapped to risk index values ​​based on the corresponding relationship between ice type definitions and risk index values. The ice type definitions and risk index values ​​are based on the China Classification Society's Polar Ship Guide.

[0088] S203: Based on the sea ice density data, the risk index result is calculated according to the ice type and the corresponding risk index value, and navigation risk index result distribution data is generated.

[0089] In a specific embodiment, based on the higher-resolution sea ice density data obtained in S201, a risk index result (RIO) is calculated according to the ice type and the corresponding risk index value, and the navigation risk index result distribution data is generated by fusion. The risk index result calculation formula is as follows:

[0090] RIO=(C1×RIV1)+(C2×RIV2)+…+(C n ×RIV n )

[0091] Among them, C1, C2, ..., C n is the density of various ice types in the ice domain; RIV1, RIV2, ..., RIV n is the risk index value of each ice type, and RIO is the sum of the risk index value of each ice type multiplied by its corresponding density.

[0092] Step 3: Inputting a sample set of the historical sea ice density dataset into the initial spatiotemporal prediction model to obtain a risk prediction grid map predicted by the sample set; the historical sea ice density dataset includes a sample set and a label set corresponding to the sample set;

[0093] Among them, the historical sea ice density is provided by the National Snow and Ice Center (NSIDC) of the United States, and the historical sea ice thickness data is provided by the Sea Ice Remote Sensing Center of the University of Bremen.

[0094] In a specific embodiment, a daily dataset of Arctic sea ice concentration and thickness was obtained. The data range is from May 1, 2010, to December 31, 2023, totaling 9,964 data points. The sea ice concentration data was provided by the National Sea Ice Data Center (NSIDC) in TIFF format, with a spatial resolution of 25 km x 25 km and a grid size of 304 x 448 pixels. The sea ice thickness data was provided by the Center for Remote Sensing of Sea Ice at the University of Bremen in NC format, with a spatial resolution of 12.5 km x 12.5 km and a grid size of 608 x 896 pixels after conversion to TIFF. Both datasets cover the area north of 30.98°N.

[0095] In a specific embodiment, the historical sea ice density and historical sea ice thickness data are processed according to the method of step 2 above to generate a daily dataset of Arctic navigation risk index results, and the dataset is made into training samples, with 16 days as a sample, the first 8 days as input values, the last 8 days as label values, and the step size of adjacent samples is 1.

[0096] Spatiotemporal prediction of route environmental risks requires consideration of both temporal data (the temporal dimension) and spatial data (the spatial dimension). Traditional convolutional neural networks (CNNs) typically struggle to directly handle spatiotemporal dependencies. However, the Transformer, particularly the SwinTransformer, effectively captures global dependencies in spatiotemporal information through its self-attention mechanism.

[0097] In a specific embodiment, the initial spatiotemporal prediction model includes an encoder, a nested residual module and a decoder.

[0098] Specifically, the daily dataset of Arctic Navigation Risk Index results is fed into the Patch Partitioning module, which partitions the data into local window regions in both spatial and temporal dimensions. Each data block is then subjected to dimensionality reduction in the Linear Embedding module to obtain a low-dimensional representation. These low-dimensional blocks are then fed into the 3D-Swim Transformer Block module, which computes three-dimensional self-attention within each window to capture local spatiotemporal dependencies (such as the temporal diffusion of sea ice within a region) and extract spatiotemporal features. After passing through multiple such 3D-Swim Transformer Blocks, the data enters the Path Merging module, which, through merging operations, gradually integrates local features into global features.

[0099] The feature map output by the encoder passes through the nested residual module, and the input features are directly added to the output features of the nested residual module, alleviating the gradient vanishing and gradient exploding problems in deep networks.

[0100] The output data of the nested residual module is input into the decoder and first reconstructed by the Pixelshuffle module. The low-resolution feature map is rearranged to recombine multiple channel information into spatial dimensions to improve the output resolution. The feature information is then extracted by the CNN block module. After repeated processing, it enters the prediction layer to convert the high-resolution feature map into a raster map for risk prediction.

[0101] Step 4: Based on the difference between the risk prediction grid map predicted by the sample set and the label set image corresponding to the sample set, iterate the parameters of the initial spatiotemporal prediction model to obtain a spatiotemporal prediction model;

[0102] Step 5: Inputting the navigation risk index result distribution data into the spatiotemporal prediction model and outputting a risk prediction grid map;

[0103] In a specific embodiment, the navigation risk index result distribution data sequentially enters the encoder, the nested residual module, and the decoder;

[0104] In the encoder, the navigation risk index result distribution data first passes through the patch partition module, which divides the three-dimensional data into data blocks and converts the original image into more structured local areas. This can reduce the computational complexity of processing and provide more representative local information for the subsequent attention mechanism.

[0105] Then, through the linear embedding module, the data block is embedded from high dimension to low dimension through linear transformation, and then through the 3D-swim transformer block module to learn features. Then, through the path merging module, the high-level information in the feature image is aggregated and concentrated, making deep feature extraction more effective. This step is repeated to complete feature extraction.

[0106] Because the Swin Transformer introduces a local window self-attention mechanism, which focuses only on pixels in a local area of ​​the image at a time, it can reduce computational complexity while maintaining high efficiency. When processing time series data, the Swin Transformer first processes local spatial information within each layer and then fuses global information across layers (through cross-window self-attention), thus simultaneously considering both local and global information. In route environment forecasting, local windows can help capture factors such as local weather conditions and route obstructions, while cross-layer global information fusion can capture widespread changes in weather systems or global risks of routes.

[0107] The feature maps processed by the encoder are passed to the decoder via nested residual modules, which effectively mitigate the vanishing gradient problem during deep network training. This is particularly important for training flight path environmental risk prediction models, as these models typically involve complex spatiotemporal data with high dimensionality. Without residual connections, training can be difficult and prone to instability.

[0108] Finally, the input feature map is upsampled through multiple upsampling modules in the decoder. Before the data features enter the last CNN module, a global residual connection structure is designed to further ensure the stability and generalization ability of training.

[0109] In a specific embodiment, in order to comprehensively evaluate the prediction performance of the VCCNet of the present invention, the following indicators are used: root mean square error (RMSE), mean absolute error (MAE), coefficient of determination (R 2 ), the formula is as follows:

[0110]

[0111] Where N is the total number of pixels in a row or column of the input image, X ij is the value of the pixel at the i-th row and j-th column of the non-land grid of the real sea ice density image, To predict the pixel values ​​corresponding to the non-land grids in the image, is the average value of the non-land grid in the real image.

[0112] RMSE, MAPE, and MAE are common indicators for error evaluation, which can directly reflect the relative or absolute error between the predicted results and the observed results.

[0113] For MAPE, if X ij is zero, replace the denominator with 0.1.

[0114] Coefficient of determination R 2 Measuring from the perspective of correlation can more clearly reflect the similarities and differences in evolutionary trends.

[0115] The data selected for this example covers the area above 30.98°N and longitudes from 180°E to 180°W. It covers the entire Arctic Ocean and has a data shape of 608*896. Data from May 1, 2010, to December 31, 2020, is used as the training set, data from 2021 as the validation set, and data from 2022 as the test set.

[0116] This example was run on an NVIDIA 4090 RTX GPU with 24GB of memory and an Intel Xeon Platinum 8352V 2.10GHz twelve-core processor. The training batch size was set to 3, and the initial learning rate was 0.001. The learning rate was dynamically adjusted to accelerate convergence using the cosine learning rate variable.

[0117] The evaluation indicators selected are RMSE, MAE, PSNR, and R2. The experimental results are shown in the following table:

[0118] index RMSE MAE R2 PSNR 2022 2.452 1.303 0.997 43.640

[0119] The experimental results show that the RMSE and MAE indicators of the model of the present invention are small, indicating that the prediction error of the present invention is small, the R2 indicator is close to 1, indicating that the input data is close to the fitting regression line and the degree of fit is good, and the PSNR indicator exceeds 40, indicating that the feature map obtained by the model used in the present invention has high accuracy and good reconstruction quality. At the same time, thanks to the sliding window mechanism, the training time required for the model is also shorter.

[0120] like Figure 7 As shown in FIG, there are example diagrams of prediction effects in one embodiment of the present invention, wherein the first row is the real space-time diagram, and the second row is the predicted space-time diagram corresponding to the first row. It can be seen that the prediction accuracy is very high.

[0121] Furthermore, the present invention also provides a deep learning-based spatiotemporal prediction system for environmental risks in Arctic shipping routes, specifically comprising:

[0122] The data acquisition and processing module is used to obtain the daily spatial distribution data of navigation risk ROI and process it into a standardized format. The data must have temporal and spatial characteristics. The daily data records the sea ice density element value of each coordinate grid in the area. The Arctic sea ice density data of the recent T days are collected in a time series and processed into the format of T*X*W, where T is the number of days and H and W are the sizes of the data grid.

[0123] The model building module uses the Pytorch deep learning framework to implement the VCCNet spatiotemporal prediction network model based on decomposition ideas and sparse spatiotemporal attention mechanism;

[0124] The model training module combines the historical daily data of sea ice density to train the model. According to the observation data and prediction requirements, the parameters of the VCCNet model are adjusted. The number of prediction days and the spatial resolution of the sea ice density prediction results can be adjusted.

[0125] The navigation risk ROI spatiotemporal prediction module combines recent observation data and the trained model to predict the future navigation risk ROI spatiotemporal distribution. The daily navigation risk ROI data of the most recent day A obtained in real time is input as parameters into the trained VCCNet model to achieve the spatiotemporal prediction of the navigation risk ROI for the next day B.

[0126] The waterway risk prediction system based on artificial intelligence spatiotemporal prediction of this embodiment corresponds to the waterway risk prediction method based on artificial intelligence spatiotemporal prediction of the first embodiment mentioned above, so it will not be repeated here.

[0127] The flow charts and block diagrams in the accompanying drawings show the possible architecture, functions and operations of the systems, methods and computer program products according to multiple embodiments of the present invention. In this regard, each box in the flow chart or block diagram can represent a part of a module, program segment or instruction, and the part of the module, program segment or instruction includes one or more executable instructions for realizing the logical function of the specification. In some alternative implementations, the functions marked in the box can also occur in a sequence different from that marked in the accompanying drawings. It should also be noted that each box in the block diagram and / or flow chart, and the combination of the boxes in the block diagram and / or flow chart can be realized by a special hardware-based system that performs the function or action of the specification, or can be realized by a combination of special hardware and computer instructions. It is well known to those skilled in the art that it is equivalent to realize by hardware, realize by software and realize by the combination of software and hardware.

[0128] The preferred embodiments of the present invention disclosed above are only used to help illustrate the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to the specific implementation methods described. Obviously, many modifications and changes can be made based on the contents of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can better understand and utilize the present invention. The present invention is only limited by the claims and their full scope and equivalents. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A spatiotemporal prediction method for environmental risks in Arctic shipping routes based on deep learning, characterized by: include: Using satellite remote sensing to obtain time series sea ice concentration data and sea ice thickness data over a recent T-day period within a region, and processing the data into a T*X*W format, where T is the number of days and X and W are the sizes of the data grid; the processed time series sea ice concentration data and sea ice thickness data have a spatiotemporal relationship; Determine the sea ice type based on sea ice density data and sea ice thickness data, calculate the risk index result according to the sea ice type and the corresponding risk index value, and generate the distribution data of the navigation risk index result based on the regional grid; Inputting the navigation risk index result distribution data into a spatiotemporal prediction model and outputting a grid map of the predicted navigation risk index; The spatiotemporal prediction model includes an encoder, a residual link module and a decoder; The encoder includes a 3D window attention mechanism and a 3D sliding window attention mechanism, which is used to first capture the spatiotemporal meteorological condition characteristics of the local window through the 3D window attention mechanism, and then fuse the spatiotemporal meteorological condition characteristics of the local window through the 3D sliding window attention mechanism to capture the global risk characteristics of the route; The residual link module is used to transfer the spatiotemporal meteorological condition features extracted from the encoder to the decoder; The decoder is used to convert the input feature map into a risk prediction grid map.

2. The method according to claim 1, characterized in that Also includes: Inputting a sample set of the historical sea ice density data set into an initial spatiotemporal prediction model to obtain a risk prediction grid map predicted by the sample set; The historical sea ice density dataset includes a sample set and a label set corresponding to the sample set; Based on the difference between the risk prediction grid map predicted by the sample set and the label set image corresponding to the sample set, the parameters of the initial spatiotemporal prediction model are iterated to obtain a spatiotemporal prediction model.

3. The method according to claim 1, characterized in that The encoder comprises: A patch partition module is used to divide the navigation risk index results into blocks and to divide the original data with spatial and temporal characteristics into local window areas; A linear embedding module is used to embed the data block from a high-dimensional representation to a low-dimensional representation; The 3D-swim transformer block module is used to perform attention operations in both spatial and temporal dimensions to capture the spatial distribution patterns and temporal variation trends of sea ice data; The path merging module is used to aggregate and condense high-level information in the image, integrating local spatiotemporal meteorological condition features into global risk characteristics of the route.

4. The method according to claim 1, wherein The residual link module consists of at least two residual blocks, which directly add the input to the output through a skip connection, and is used to alleviate the gradient vanishing and gradient exploding problems in deep networks.

5. The method according to claim 1, wherein The decoder comprises at least two upsampling modules; The upsampling module includes: The Pixelshuffle module is used to rearrange a low-resolution feature map into multiple channel information into spatial dimensions to improve the output resolution; CNN block module, used to extract feature information; The prediction layer is used to convert the high-resolution feature map into a raster map for risk prediction.

6. The method according to claim 5, characterized in that The decoder also includes a residual connection structure, which is arranged between the prediction layer and the upsampling module, so that the spatiotemporal prediction model learns the residual between the input and output, thereby improving the stability of the model.

7. The method according to claim 1, characterized in that The step of calculating the risk index result based on the sea ice concentration data and the sea ice thickness data according to the ice type and the corresponding risk index value, and generating the navigation risk index result distribution data specifically includes: Spatially interpolate sea ice concentration data to align them with sea ice thickness data; According to the corresponding relationship between ice type definition and risk index value, the sea ice thickness data is mapped to the risk index value; Based on the sea ice density data, the risk index results are calculated according to the ice type and the corresponding risk index value, and the navigation risk index result distribution data is generated.

8. A deep learning-based spatiotemporal prediction system for environmental risks in Arctic shipping routes, characterized by: For implementing any method described in claims 1-7, comprising: The data acquisition and processing module is used to record the sea ice density element values ​​of each coordinate grid in the daily data recording area, collect the recent T-day time series Arctic sea ice density data, and process it into the format of T*X*W, where T is the number of days, H and W are the size of the data grid, obtain the daily spatial distribution data of the navigation risk ROI and process it into a standardized format; A model building module for implementing the VCCNet spatiotemporal prediction network model based on decomposition ideas and sparse spatiotemporal attention mechanism based on the Pytorch deep learning framework; The model training module is used to train the model based on the historical daily sea ice concentration data. According to the observation data and prediction requirements, the parameters of the VCCNet model are adjusted. The parameters include the number of prediction days and the spatial resolution of the sea ice concentration prediction results. The navigation risk ROI spatiotemporal prediction module is used to predict the spatiotemporal distribution of future navigation risk ROI. The daily navigation risk ROI data of the most recent day A obtained in real time is input as parameters into the trained spatiotemporal prediction model to achieve the spatiotemporal prediction of the navigation risk ROI of the next day B.

Images