A method and system for predicting sea ice concentration

Abstract

This invention discloses a method and system for predicting sea ice concentration. The method includes: preprocessing raw historical data to obtain input data; inputting the input data into a prediction model for prediction to obtain a sea ice concentration prediction result for a target time; inputting the input data into a trained prediction model for prediction, including: inputting the input data into an encoder for encoding to output encoded features; using 3D Hilbert scan processing and state-space modeling to model the encoded features and output spatiotemporal features, and obtaining frequency features based on the encoded features; processing the spatiotemporal features and the frequency features through pooling concatenation, channel shuffling, weighted fusion, and 3D Hilbert inverse scan to output fused depth features; inputting the fused depth features into a decoder to output the sea ice concentration prediction result for the target time. Using this invention can improve the accuracy of sea ice concentration prediction.

Description

Technical Field

[0001] This invention belongs to the field of marine environment prediction technology, specifically, it relates to a method and system for predicting sea ice concentration. Background Technology

[0002] Sea ice, a key component of the global climate system, plays a vital role in regulating regional and even global temperatures through its albedo effect. It helps maintain Earth's energy balance by reflecting incident solar radiation. However, in recent decades, sea ice has melted at a much faster rate than expected, indicating an alarming shift in global climate patterns. Therefore, accurate prediction of sea ice concentration dynamics is crucial for deepening our understanding of climate change and ensuring safe navigation at sea.

[0003] Existing sea ice concentration prediction methods can be broadly categorized into two types: physics-driven methods and data-driven methods. Physics-driven methods rely on numerical models to simulate sea ice behavior by solving coupled dynamic and thermodynamic equations. These models incorporate physical laws along with external forcing factors such as wind, ocean currents, and atmospheric conditions to predict sea ice changes. While they provide interpretable and physically consistent forecasts, their accuracy is often limited by model complexity, parameterization uncertainties, and high computational costs. In contrast, data-driven methods, particularly those based on deep learning, have made significant progress in sea ice concentration prediction due to their ability to automatically extract spatiotemporal features and nonlinear patterns from massive amounts of observational data.

[0004] Despite the aforementioned successes of deep learning technology, many mainstream architectures, especially Transformer-based models, encounter computational bottlenecks when dealing with the massive spatiotemporal data involved in climate modeling, which limits their efficiency in long sequence modeling tasks. To address this challenge, State Space Models (SSMs), particularly the Mamba model, have recently demonstrated superiority in natural language processing and computer vision due to their linear complexity in handling long sequences, and have begun to be introduced into spatiotemporal prediction tasks. For example, the Mamba-based Arctic sea ice concentration prediction method disclosed in Chinese patent application CN118298327A presents a method for predicting Arctic sea ice concentration. However, when this method directly applies the Mamba architecture, originally designed for one-dimensional sequences, to high-dimensional spatiotemporal prediction tasks such as sea ice, there are serious technical bottlenecks: (1) Inadequate temporal correlation modeling: The conventional practice of flattening the sea ice concentration image sequence frame by frame into a one-dimensional vector and then inputting it into the Mamba model will unexpectedly destroy the physical causal relationship between temporally adjacent frames in the sea ice sequence. In the dynamic changes of sea ice, the state of the next frame is usually the result of the evolution of the previous few frames. This simple flattening process interrupts this natural time flow, making it difficult for the model to accurately capture the temporal pattern of sea ice changes. (2) Loss of boundary details in edge regions: The Mamba model performs its attention calculation through a scanning mechanism, which easily erodes the inherent local spatial details in the data. For the edge regions of sea ice, the boundary details are usually very fine and scattered in the image. Existing Mamba-based methods are difficult to retain these key spatial cues, resulting in unclear prediction results in the edge regions of sea ice and reduced accuracy. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for predicting sea ice concentration, thereby improving prediction accuracy.

[0006] To achieve the above-mentioned objectives, the present invention employs the following technical solution: A method for predicting sea ice concentration includes: The raw historical data is preprocessed to obtain input data including historical time series and target time series; The input data is fed into the trained prediction model for prediction, and the sea ice concentration prediction result for the target time is obtained. The input data is fed into the trained prediction model for prediction, including: The input data is input into the encoder for encoding, and the encoded features are output. The encoded features are input into the frequency-enhanced state space module, which outputs the fused deep features. The fused deep features are input into the decoder to output the sea ice concentration prediction result for the target time. The encoded features are input into the frequency-enhanced state space module, and the fused deep features are output, including: The encoded features are processed using a 3D Hilbert scan to output one-dimensional sequence features; the one-dimensional sequence features are then input into a state-space model for modeling, and spatiotemporal features are output. The encoded features are sequentially processed by wavelet transform, depth convolution, inverse wavelet transform, and 3D Hilbert scan to output frequency features. The spatiotemporal features and frequency features are pooled and concatenated to output pooled features; the pooled features are shuffled by channel to output rearranged features; the rearranged features are grouped by convolution and deshuffled to output spatiotemporal feature attention weights and frequency feature attention weights; the spatiotemporal features and frequency features are weighted and fused according to the spatiotemporal feature attention weights and the frequency feature attention weights to output one-dimensional fused features; the one-dimensional fused features are processed by 3D Hilbert inverse scanning to obtain the fused depth features.

[0007] In some embodiments of this application, the one-dimensional sequence features are input into the state space model for modeling, and the spatiotemporal features are output, including: The one-dimensional sequence features are preprocessed and then input into the state space model for modeling and output. The output of the state space model is then subjected to layer normalization to obtain the main path result. The one-dimensional sequence features are also modulated to output a modulated signal; The main path result is multiplied element by element with the modulation signal, and the product is used as the spatiotemporal feature and output.

[0008] In some embodiments of this application, the one-dimensional sequence features are preprocessed, including: The one-dimensional sequence features are sequentially processed by linear processing, convolution processing, and SiLU activation function.

[0009] In some embodiments of this application, the state space model employs forward scanning modeling and backward scanning modeling for the preprocessed data of the one-dimensional sequence features, respectively; the spatiotemporal features include a first spatiotemporal feature corresponding to the forward scanning modeling and a second spatiotemporal feature corresponding to the backward scanning modeling.

[0010] In some embodiments of this application, the one-dimensional fusion feature is processed using 3D Hilbert inverse scan to obtain the fused depth feature, including: The one-dimensional fused features are processed using 3D Hilbert inverse scan to output the fused 3D features; The fused 3D features are concatenated with the encoded feature residuals to output the fused initial features; The initial fused features are subjected to layer normalization and depthwise convolution processing in sequence to output the fused processed features. The initial fused features are concatenated with the residual fused features to output the fused depth features.

[0011] In some embodiments of this application, the frequency enhancement state space module includes multiple modules connected in series, and the encoded features are input into the multiple modules connected in series to output the fused depth features.

[0012] In some embodiments of this application, the original historical data is preprocessed to obtain the input data, which includes historical time series and target time series, including: The original historical data is subjected to spatiotemporal alignment and missing value repair to obtain processed historical data; A sliding window is used to extract temporally continuous data segments from the processed historical data to construct the input data, which includes historical time series and target time series.

[0013] In some embodiments of this application, when training the prediction model, a composite loss function for model training is constructed based on pixel-level loss and gradient loss as the optimization objective.

[0014] A sea ice concentration prediction system, comprising: The historical data preprocessing module is used to preprocess the raw historical data to obtain input data including historical time series and target time series; A prediction model is used to make predictions based on the input data to obtain the sea ice concentration prediction result for the target time. The prediction model includes: The encoder unit is used to encode the input data and output the encoded features; The frequency enhancement state space module unit is used to process the encoded features and output the fused deep features; The decoder unit is used to decode the fused depth features and output the sea ice concentration prediction result for the target time. The frequency enhancement state space module unit includes: The spatiotemporal feature acquisition subunit is used to process the encoded features using 3D Hilbert scanning to output one-dimensional sequence features; the one-dimensional sequence features are input into a state space model for modeling and output spatiotemporal features. The frequency feature acquisition subunit is used to sequentially perform wavelet transform processing, depth convolution processing, inverse wavelet transform processing, and 3D Hilbert scan on the encoded features to output frequency features. The feature fusion subunit is used to pool and concatenate the spatiotemporal features and the frequency features to output pooled features; to shuffle the pooled features by channels to output rearranged features; to perform grouped convolution and deshuffling on the rearranged features to output spatiotemporal feature attention weights and frequency feature attention weights; to perform weighted fusion of the spatiotemporal features and frequency features according to the spatiotemporal feature attention weights and the frequency feature attention weights to output one-dimensional fused features; and to perform 3D Hilbert inverse scanning processing on the one-dimensional fused features to obtain the fused depth features.

[0015] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the above-described sea ice concentration prediction method is implemented.

[0016] Compared with the prior art, the advantages and positive effects of the present invention are: The sea ice concentration prediction method and system provided by this invention employ 3D Hilbert scanning to process the encoded features output by the encoder, which are then input into a state-space model for modeling. The output spatiotemporal features effectively map high-dimensional spatiotemporal data into a one-dimensional sequence while preserving the local proximity of the original data in both time and space dimensions to the greatest extent possible, avoiding disruption of spatiotemporal continuity. This allows the subsequent state-space model to more effectively capture the inherent temporal patterns and dynamic processes of sea ice evolution, significantly improving the model's ability to model temporal correlations. Furthermore, wavelet processing and 3D Hilbert scanning are used on the encoded features to output frequency features, effectively separating high-frequency information representing fine structures such as sea ice edges, resulting in accurate... The method preserves detailed information about the sea ice edge; then, the spatiotemporal and frequency features are shuffled through channels to obtain attention weights for the two features, ensuring that the calculation of attention weights occurs within the corresponding channel group. This allows for more accurate capture of the interrelationships between different features. Based on the attention weights, the spatiotemporal and frequency features are adaptively fused to obtain fused depth features, fully utilizing the complementary information of the two features. Finally, the fused depth features are decoded to obtain the sea ice concentration prediction results. This effectively overcomes the limitation of a single feature in taking into account both macroscopic trends and microscopic details, significantly improving the prediction accuracy of the sea ice edge region and the ability to reconstruct boundary details, thus achieving high-precision sea ice concentration prediction.

[0017] Other features and advantages of the present invention will become clearer after reading the detailed embodiments of the invention in conjunction with the accompanying drawings. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 Flowcharts of some embodiments of the sea ice concentration prediction method provided by the present invention; Figure 2 Flowcharts of some other embodiments of the sea ice concentration prediction method provided by the present invention; Figure 3 Flowcharts of some embodiments of the sea ice concentration prediction method provided by the present invention; Figure 4 Structural block diagrams of some embodiments of the sea ice concentration prediction system provided by the present invention; Figure 5 This is a visual comparison chart of the prediction results and deviations of various prediction methods, including the sea ice concentration prediction method provided in this invention, based on the first dataset. Figure 6 This is a visual comparison chart of prediction results using multiple prediction methods, including the sea ice concentration prediction method provided in this invention, based on the second dataset. Figure 7 This is a visual comparison chart of prediction results using multiple prediction methods, including the sea ice concentration prediction method provided in this invention, based on a third dataset. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0021] It should be noted that the technical solutions of the various embodiments of the present invention can be combined with each other, but only if they can be implemented by those skilled in the art. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0022] To address the technical problem of low prediction accuracy in existing methods for predicting sea ice concentration, this invention proposes a novel sea ice concentration prediction method that can effectively model spatiotemporal correlations and accurately preserve detailed information about sea ice edges, thereby improving the overall performance of sea ice concentration prediction.

[0023] Figure 1 The diagram shows flowcharts of some embodiments of the sea ice concentration prediction method provided by the present invention.

[0024] like Figure 1 As shown, this embodiment uses the following process to predict sea ice concentration.

[0025] Step S11: Preprocess the raw historical data to obtain input data including historical time series and target time series.

[0026] This embodiment uses historical sea ice concentration data as a basis and employs a prediction model to predict sea ice concentration at a target time. The original historical sea ice concentration data can be obtained from remote sensing images containing sea ice concentration data; this embodiment is not limited to this. After obtaining the original historical data, preprocessing is performed to obtain input data including historical time series and target time series. The historical time series includes several historical times corresponding to historical sea ice concentrations, and the target time series includes several future times to be predicted; the data corresponding to the target times is the sea ice concentration data to be predicted. Preprocessing of the original historical data can employ existing techniques; this embodiment is not limited to this.

[0027] In some embodiments, the raw historical data is preprocessed to obtain input data including historical time series and target time series, including: Spatiotemporal alignment and missing value repair are performed on the original historical data to obtain processed historical data. In some embodiments, the original historical data is often missing due to sensor limitations or cloud cover. A spatiotemporal inverse distance weighting algorithm combined with a Gaussian kernel can be used for interpolation to reconstruct the missing data and ensure the spatiotemporal integrity of the data.

[0028] A sliding window is used to extract temporally continuous data segments from the processed historical data to construct input data including historical time series and target time series. In some embodiments, a 14-day window length is used to take a 14-day continuous sea ice concentration data sequence from the historical data as the historical time series, and the sea ice concentration for the next 14 days is predicted using the historical 14-day sea ice concentration data.

[0029] Then, the input data is fed into the trained prediction model, and the prediction process of steps S12-S14 is performed to obtain the sea ice concentration prediction result for the target time.

[0030] Step S12: Input the input data into the encoder for encoding, and output the encoded features.

[0031] The encoder encodes the input data, which can be achieved using existing technologies, and this embodiment does not limit this.

[0032] In some embodiments, the encoder employs a structure consisting of convolutional layers, normalization layers, and activation layers to extract features and reduce the dimensionality of the input data, thereby obtaining encoded features with deep feature representation.

[0033] Step S13: Input the encoded features into the frequency enhancement state space module and output the fused deep features.

[0034] The frequency enhancement state space module includes a spatiotemporal dependency modeling branch for acquiring spatiotemporal features and a high-frequency detail enhancement branch for acquiring frequency features. The two branches are processed in parallel, and the features obtained after processing are fused to form fused deep features.

[0035] Specifically, the encoded features are processed using 3D Hilbert scanning to output one-dimensional sequence features. These one-dimensional sequence features are then input into a state-space model for modeling, outputting spatiotemporal features. The 3D Hilbert scanning process effectively preserves the continuity of the spatiotemporal neighborhood within the encoded features, thereby maximizing the preservation of the local proximity of sea ice concentration data in the original three-dimensional spatiotemporal context. This allows the subsequent state-space model to more effectively capture the inherent temporal patterns and dynamic processes of sea ice evolution, significantly improving the model's ability to model temporal correlations. The state-space modeling process can utilize existing technologies, and this embodiment is not limited to this. In some other embodiments, the state-space model employs the Mamba model. By combining 3D Hilbert scanning with a state-space model to obtain spatiotemporal features, the linear computational complexity advantage of the state-space model in processing long-sequence data is inherited. While achieving long-range spatiotemporal dependency modeling, the computational load and memory consumption are significantly reduced, resulting in higher computational efficiency and better scalability when processing large-scale, high-resolution marine remote sensing data, making it more suitable for practical operational forecasting scenarios.

[0036] Simultaneously, the encoded features are sequentially processed using wavelet transform, deep convolution, inverse wavelet transform, and 3D Hilbert scan to output frequency features. By performing wavelet transform on the encoded features, the features can be decomposed into different frequencies to achieve multi-scale decoupling, thereby effectively separating high-frequency information representing fine structures such as sea ice edges. In some embodiments, wavelet transform is used to distribute the encoded features into a low-frequency component representing the main structure and three high-frequency components capturing horizontal, vertical, and diagonal details. Subsequently, deep convolution is used to simultaneously enhance the local features of the decomposed components, accurately extracting spatial patterns from high-frequency details while reducing computational overhead. Then, inverse wavelet transform is used to remap the processed features back to the original spatial dimension, completing the effective reconstruction of frequency information and transforming them into features with the original resolution. This mechanism combining time-frequency analysis and convolution processing enables the model to focus on learning the fine features of sea ice edges, significantly improving its ability to represent and reconstruct complex spatial structures. Finally, a 3D Hilbert scan is used to convert the features to one dimension, achieving semantic alignment with the spatiotemporal features in the spatial context for fusion.

[0037] The model fuses spatiotemporal and frequency features: Spatiotemporal and frequency features are pooled and concatenated to output pooled features. Then, the pooled features are shuffled by channel, arranging features from the same channel from different branches adjacently, outputting rearranged features. These rearranged features are then subjected to grouped convolution and deshuffling, with channel attention weights calculated on a group basis, outputting spatiotemporal and frequency feature attention weights. This design ensures that attention weight calculation occurs within the corresponding channel group, more accurately capturing the relationships between different features. Next, the spatiotemporal and frequency features are weighted and fused based on their respective attention weights, enabling the model to intelligently and adaptively identify and fuse complementary information from different branches, outputting a one-dimensional fused feature. Finally, the one-dimensional fused feature is processed using 3D Hilbert inverse scanning, restoring the sequence to a three-dimensional spatial feature map based on the mapping relationship of the scanning path, ultimately obtaining the fused depth feature.

[0038] Step S14: Input the fused depth features into the decoder and output the sea ice concentration prediction result for the target time.

[0039] The decoder adopts a structure symmetrical to the encoder, and gradually upsamples the fused depth features output by the frequency enhancement state space module to restore the original spatial resolution, and finally generates the sea ice concentration prediction result for the target time.

[0040] The method described in the above embodiments for predicting sea ice concentration uses a 3D Hilbert scan to process the encoded features output by the encoder, which are then used as input to a state-space model for modeling. The output spatiotemporal features effectively map high-dimensional spatiotemporal data into a one-dimensional sequence while preserving the local proximity of the original data in both time and space, significantly improving the model's ability to model temporal correlations. By applying wavelet processing and 3D Hilbert scanning to the encoded features, frequency features are output, effectively separating high-frequency information representing fine structures such as sea ice edges and accurately preserving detailed information about the sea ice edges. The spatiotemporal and frequency features are then shuffled through channels to obtain attention weights for the two features. Based on these attention weights, the two features are fused to form a fused deep feature, ensuring that the calculation of attention weights occurs within the corresponding channel group, enabling more accurate capture of the relationships between different features. Therefore, the method described in the above embodiments effectively overcomes the limitation of a single feature in simultaneously capturing macroscopic trends and microscopic details, significantly improving the prediction accuracy of sea ice edge regions and the ability to reconstruct boundary details, achieving high-precision sea ice concentration prediction.

[0041] In other embodiments, the frequency enhancement state space module comprises multiple modules connected in series, forming a multi-level feature extraction and fusion structure. Encoded features are sequentially processed by these multiple connected frequency enhancement state space modules, outputting the final fused deep features. Using a multi-level connected frequency enhancement state space module structure to construct the prediction model can continuously reduce prediction errors through deep feature mining, significantly enhancing the model's ability to model complex spatiotemporal dependencies, thereby achieving higher-fidelity sea ice dynamic simulation.

[0042] Figure 2 The diagram shows flowcharts of some other embodiments of the sea ice concentration prediction method provided by the present invention. In this embodiment, with... Figure 1 Similarly, in this embodiment, the encoded features output by the encoder are input to the frequency enhancement state space module. This module includes a spatiotemporal dependency modeling branch for acquiring spatiotemporal features and a high-frequency detail enhancement branch for acquiring frequency features. These two branches process in parallel, and the resulting features are fused to form fused depth features. Specifically, the spatiotemporal dependency modeling branch in the frequency enhancement state space module performs 3D Hilbert scanning on the encoded features and outputs one-dimensional sequence features. Then, based on these one-dimensional sequence features, spatiotemporal features are acquired using the following process.

[0043] Step S21: Preprocess the one-dimensional sequence features.

[0044] Specifically, linear processing, convolution processing, and SiLU activation function activation are performed sequentially on the one-dimensional sequence features.

[0045] Step S22: Model using a state-space model and output the results.

[0046] The preprocessed data from step S21 is input into a state-space model to efficiently capture long-range dependencies in the sequence with linear complexity, and the results are output.

[0047] Step S23: Layer normalization processing to obtain the main path result.

[0048] The output results after modeling the state-space model are subjected to layer normalization to output the main path results.

[0049] Step S24: Modulate the one-dimensional sequence features and output the modulated signal.

[0050] Specifically, linear processing and SiLU activation function activation are performed sequentially on the one-dimensional sequence features to output a modulated signal.

[0051] Step S25: Multiply the main path result element by element with the modulated signal, and output the product as a spatiotemporal feature.

[0052] In other embodiments, the state-space model employs both forward and backward scan modeling for the data after preprocessing the one-dimensional sequence features. Correspondingly, the spatiotemporal features include a first spatiotemporal feature corresponding to the forward scan modeling and a second spatiotemporal feature corresponding to the backward scan modeling. Utilizing the state-space model for bidirectional scanning enables the model to comprehensively perceive spatiotemporal contextual relationships, significantly enhancing the completeness and robustness of sea ice dynamic feature extraction.

[0053] Figure 3 The diagram shows flowcharts of some embodiments of the sea ice concentration prediction method provided by the present invention. Figure 1 Similarly, in the implementation, the encoded features output by the encoder are input to the frequency enhancement state space module. The frequency enhancement state space module obtains spatiotemporal features and frequency features through two branch distributions, and then performs weighted fusion of the two features using attention weights obtained through channel shuffling, group convolution, and deshuffling to obtain a one-dimensional fused feature. Then, based on the one-dimensional fused feature, the following process is used to obtain the fused deep feature.

[0054] Step S31: Apply 3D Hilbert inverse scan processing to the one-dimensional fused features and output the fused 3D features.

[0055] The specific implementation of this process can employ existing technologies, and this embodiment does not limit it in this regard.

[0056] Step S32: Connect the fused 3D features with the encoded feature residuals to output the fused initial features.

[0057] Step S33: Perform layer normalization and depthwise convolution on the initial fused features in sequence, and output the fused processed features.

[0058] Step S34: Connect the initial features after fusion with the residual features after fusion processing to output the fused depth features.

[0059] This embodiment obtains the final fused depth features through two-level residual connections, which can avoid the problems of gradient vanishing and degradation, and improve the prediction speed and accuracy of the prediction model.

[0060] The prediction model used for sea ice concentration prediction requires training with fumigation samples. During model training, the AdamW optimizer is used to update the model parameters. In some embodiments, to effectively guide the updating of model parameters, a composite loss function is constructed based on pixel-level loss and gradient loss as the optimization objective. The loss value guides the backpropagation process during training to update all learnable parameters of the model. Using this composite loss function ensures both the accuracy of the prediction results at the pixel level and the fidelity in spatial structural details. The pixel-level loss can be calculated using L1 loss, which measures the absolute difference between the predicted and actual results at each pixel to ensure the overall accuracy of the predicted values. The gradient loss penalizes inconsistencies in edge and texture details by calculating the L1 difference in spatial gradient between the predicted and actual results, thereby strengthening the model's ability to preserve spatial structural details such as sea ice fronts and edges. The specific calculation formula can employ existing techniques.

[0061] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0062] Figure 4 The diagram shows a structural block diagram of some embodiments of the sea ice concentration prediction system provided by the present invention.

[0063] like Figure 4 As shown, the prediction system of this embodiment includes structural units, the functions of the structural units, and the relationships between them, as detailed below: Prediction systems include: The historical data preprocessing module 41 is used to preprocess the raw historical data to obtain input data including historical time series and target time series.

[0064] Prediction model 42 is used to make predictions based on the input data output by the historical data preprocessing module 41, and to obtain the sea ice concentration prediction results for the target time.

[0065] Prediction model 42 includes: The encoder unit 421 is used to encode the input data output by the historical data preprocessing module 41 and output the encoded features.

[0066] The frequency enhancement state space module unit 422 is used to process the encoded features output by the encoder unit 421 and output the fused depth features.

[0067] Decoder unit 423 is used to decode the fused depth features output by frequency enhancement state space module unit 422 and output the sea ice concentration prediction result for the target time.

[0068] The frequency enhancement state space module unit 422 further includes: The spatiotemporal feature acquisition subunit 4221 is used to process the encoded features output by the encoder unit 421 using 3D Hilbert scanning to output one-dimensional sequence features; and inputs the one-dimensional sequence features into the state space model to model and output spatiotemporal features.

[0069] The frequency feature acquisition subunit 4222 is used to sequentially perform wavelet transform processing, depth convolution processing, inverse wavelet transform processing, and 3D Hilbert scan on the encoded features output by the encoder unit 421 to output frequency features.

[0070] The feature fusion subunit 4223 is used to pool and concatenate the spatiotemporal features output by the spatiotemporal feature acquisition subunit 4221 and the frequency features output by the frequency feature acquisition subunit 4222, and output the pooled features; shuffle the pooled features by channels, and output the rearranged features; perform group convolution and deshuffling on the rearranged features, and output the spatiotemporal feature attention weights and the frequency feature attention weights; perform weighted fusion of the spatiotemporal features and the frequency features according to the spatiotemporal feature attention weights and the frequency feature attention weights, and output the one-dimensional fused features; and perform 3D Hilbert inverse scanning processing on the one-dimensional fused features to obtain the fused depth features.

[0071] The prediction system described above runs the corresponding software program, performs the corresponding functions, and follows the... Figure 1 , Figure 2 , Figure 3 The methods described in the embodiments and other embodiments predict sea ice concentration and achieve the corresponding technical effects of the method embodiments.

[0072] To intuitively present the technical effects of the sea ice concentration prediction method proposed in this invention and to demonstrate its performance in spatial detail, the prediction results can be visualized and illustrated from multiple angles.

[0073] Figure 5The image shows a visual comparison of the prediction results and biases of various prediction methods, including the sea ice concentration prediction method provided in this invention, based on the first dataset.

[0074] The first dataset is a historical dataset obtained from a known database. Based on the first dataset, the prediction method provided by this invention is applied. Figure 5 The prediction results and biases of sea ice concentration on September 1, 2020, September 7, 2020, and September 14, 2020, using the existing VMRNN and FCNet methods (represented as FH-Mamba) and other methods.

[0075] Figure 5 The method for generating visualizations of prediction results and biases is briefly described below: The trained prediction model is used to predict the sea ice concentration on the first input dataset, generating sea ice concentration prediction maps for the target time. These prediction maps are two-dimensional fields of continuous values, where each pixel value represents the percentage of sea ice concentration at that location. These numerical fields are rendered into pseudo-color images using appropriate color gradations (e.g., from dark blue for open water to bright white for completely frozen water) to visually represent the predicted spatial distribution of sea ice. Simultaneously, to more finely evaluate the prediction accuracy, the model's predictions are compared with real observation data from the same period (…). Figure 5 The bias map is generated by subtracting pixels from the ground truth (represented in the image). The bias maps corresponding to the three prediction methods are represented as Difference-FH-Mamba, Difference-VMRNN, and Difference-FCNet, respectively. Figure 5 As shown, the bias plot uses a diverging color map for coloring. Red indicates predicted values ​​higher than actual values ​​(positive bias, Over Predicted), blue indicates predicted values ​​lower than actual values ​​(negative bias, Under Predicted), areas close to white indicate minimal prediction error (Sea Ice), yellow represents land, and dark blue represents water. By analyzing the bias plot, it is possible to clearly identify which specific regions (such as the sea ice edge) exhibit systematic bias in the model.

[0076] Depend on Figure 5As can be seen, in the visualization results of the sea ice concentration prediction in September 2020, the prediction map generated by the FH-Mamba method of this invention is in high agreement with the actual ground truth. Compared with the obvious prediction bias (red indicates overestimation, blue indicates underestimation) that appears in the sea ice edge region (shown in the enlarged area in the figure) of the VMRNN method and FCNet method, the method of this invention demonstrates excellent accuracy, and can accurately predict these complex and dynamically changing regions. It is particularly outstanding in preserving the fine structure and detailed features of the sea ice edge, thus making the prediction results more accurate and reliable.

[0077] Figure 6 The image shows a visual comparison of prediction results based on the second dataset, using various prediction methods, including the sea ice concentration prediction method provided in this invention.

[0078] The second dataset is also a historical dataset obtained from a known database. Using the prediction method FH-Mamba provided in this invention, as well as the existing VMRNN method and FCNet, sea ice concentration predictions for multiple dates in September of different years are made, and the sea ice extent is calculated based on the prediction results. The sea ice extent is defined as the total area of ​​the region with a sea ice concentration of not less than 15%. In addition, the sea ice extent is also calculated based on actual observation data from the same period. Then, the curves showing the sea ice extent predicted by different methods and the sea ice extent from actual observation data changing over time are plotted on the same graph for comparison, forming... Figure 6 The time-series graph shown clearly reveals the accuracy of the prediction model in forecasting the overall melting and growth trends of sea ice.

[0079] from Figure 6 As can be seen, when assessing the sea ice extent in September of different years, the sea ice extent curve (red line) predicted by the FH-Mamba method of this invention matches the actual sea ice extent curve (black line) significantly better than the other two methods. This indicates that the prediction method of this invention effectively reduces the prediction error and significantly improves the accuracy of the overall results.

[0080] Figure 7 The image shows a visual comparison of prediction results using multiple prediction methods, including the sea ice concentration prediction method provided in this invention, based on a third dataset.

[0081] The third dataset is also a historical dataset obtained from a known database. Using the prediction method FH-Mamba provided in this invention, as well as the existing VMRNN method and FCNet, sea ice concentration predictions were made for multiple dates in September for 2007, 2012, and 2020. Sea ice extent was calculated based on the prediction results, and the curves showing the sea ice extent predicted by different methods and the actual sea ice extent over time were plotted on the same graph for comparison, forming... Figure 7 The timing curves shown are as follows.

[0082] Figure 7 The three years shown are the years of the historical minimum sea ice extent, from Figure 7 It can be seen that even in years with abnormally low sea ice extent, the prediction boundary of the method of this invention is still closest to the actual observation value, demonstrating superior robustness.

[0083] In addition to visualizing the results, the technical effectiveness of the sea ice concentration prediction method proposed in this invention can also be demonstrated by comparing multiple performance indicators.

[0084] Table 1 below shows a performance comparison of sea ice concentration prediction results obtained using the proposed FH-Mamba prediction method and existing methods such as ConvLSTM, PredRNN, PredRNNv2, IceNet, SimVP, SwinLSTM, VMRNN, and FCNet on the same dataset. The performance metrics in the table include Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Nash Efficiency Coefficient (NSE). RMSE is the square root of the average of the sum of squares of the differences between predicted and true values. It assigns higher weight to larger errors and is therefore very sensitive to extreme biases in the model's predictions; a smaller RMSE value indicates higher prediction accuracy. MAE is the average of the absolute differences between predicted and true values. It intuitively represents the average magnitude of the prediction error and, compared to RMSE, is less sensitive to outliers; a smaller MAE value indicates a smaller average deviation between the model's predictions and the actual situation. The Nash efficiency coefficient (NSE) compares the model's predictions with the mean of the observed data; the closer the NSE value is to 1, the better the model's prediction performance.

[0085]

[0086] As can be seen from Table 1, the method proposed in this invention achieves the lowest RMSE, MAE and the highest NSE in all three performance indicators, demonstrating its superior stability and accuracy compared to other methods. It can more accurately capture the dynamic changes of sea ice, thereby significantly improving the reliability of prediction.

[0087] Table 2 below shows a performance comparison of the prediction results obtained by using the proposed prediction method FH-Mamba and existing methods IceNet, SimVP, SwinLSTM, VMRNN, and FCNet to perform sea ice concentration prediction for another dataset with higher resolution.

[0088]

[0089] As can be seen from Table 2, when predicting higher resolution datasets, the method proposed in this invention also outperforms other methods in three performance metrics, effectively demonstrating the powerful generalization ability and robust performance under different data sources of this invention, further highlighting the advantages and value of this invention in complex marine spatiotemporal data prediction tasks.

[0090] Other embodiments of the present invention also provide a computer storage medium on which a computer program is stored, and when the computer program is executed by a processor, it implements... Figure 1 , Figure 2 , Figure 3 The embodiments and other embodiments provide sea ice concentration prediction methods and achieve the technical effects of the corresponding embodiments.

[0091] The aforementioned computer storage media can be implemented using any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The computer storage media can be any available storage medium accessible to general-purpose or special-purpose computers.

[0092] In some embodiments, a computer storage medium is coupled to a processor, enabling the processor to read information from and write information to the storage medium. Alternatively, the storage medium can be an integral part of the processor. Both the processor and the storage medium can reside in application-specific integrated circuits (ASICs). Of course, the processor and storage medium can also exist as discrete components in the device.

[0093] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions claimed by the present invention.

Claims

1. A method for predicting sea ice concentration, characterized in that, The method includes: The raw historical data is preprocessed to obtain input data including historical time series and target time series; The input data is fed into the trained prediction model for prediction, and the sea ice concentration prediction result for the target time is obtained. The input data is fed into the trained prediction model for prediction, including: The input data is input into the encoder for encoding, and the encoded features are output. The encoded features are input into the frequency-enhanced state space module, which outputs the fused deep features. The fused deep features are input into the decoder to output the sea ice concentration prediction result for the target time. The encoded features are input into the frequency-enhanced state space module, and the fused deep features are output, including: The encoded features are processed using a 3D Hilbert scan to output one-dimensional sequence features; the one-dimensional sequence features are then input into a state-space model for modeling, and spatiotemporal features are output. The encoded features are sequentially processed by wavelet transform, depth convolution, inverse wavelet transform, and 3D Hilbert scan to output frequency features. The spatiotemporal features and frequency features are pooled and concatenated to output pooled features; the pooled features are shuffled by channel to output rearranged features; the rearranged features are grouped by convolution and deshuffled to output spatiotemporal feature attention weights and frequency feature attention weights; the spatiotemporal features and frequency features are weighted and fused according to the spatiotemporal feature attention weights and the frequency feature attention weights to output one-dimensional fused features; the one-dimensional fused features are processed by 3D Hilbert inverse scanning to obtain the fused depth features.

2. The sea ice concentration prediction method according to claim 1, characterized in that, The one-dimensional sequence features are input into the state-space model for modeling, and the spatiotemporal features are output, including: The one-dimensional sequence features are preprocessed and then input into the state space model for modeling and output. The output of the state space model is then subjected to layer normalization to obtain the main path result. The one-dimensional sequence features are also modulated to output a modulated signal; The main path result is multiplied element by element with the modulation signal, and the product is used as the spatiotemporal feature and output.

3. The sea ice concentration prediction method according to claim 2, characterized in that, The one-dimensional sequence features are preprocessed, including: The one-dimensional sequence features are sequentially processed by linear processing, convolution processing, and SiLU activation function.

4. The sea ice concentration prediction method according to claim 2, characterized in that, The state-space model employs forward scanning modeling and backward scanning modeling for the preprocessed data of the one-dimensional sequence features; the spatiotemporal features include a first spatiotemporal feature corresponding to the forward scanning modeling and a second spatiotemporal feature corresponding to the backward scanning modeling.

5. The sea ice concentration prediction method according to claim 1, characterized in that, The one-dimensional fused features are processed using 3D Hilbert inverse scan to obtain the fused depth features, including: The one-dimensional fused features are processed using 3D Hilbert inverse scanning to output the fused 3D features; The fused 3D features are concatenated with the encoded feature residuals to output the fused initial features; The initial fused features are subjected to layer normalization and depthwise convolution processing in sequence to output the fused processed features. The initial fused features are concatenated with the residual fused features to output the fused depth features.

6. The sea ice concentration prediction method according to any one of claims 1 to 5, characterized in that, The frequency enhancement state space module comprises multiple modules connected in series. The encoded features are input into the multiple modules connected in series, and the fused depth features are output.

7. The sea ice concentration prediction method according to any one of claims 1 to 5, characterized in that, The original historical data is preprocessed to obtain the input data, which includes historical time series and target time series, including: The original historical data is subjected to spatiotemporal alignment and missing value repair to obtain processed historical data; A sliding window is used to extract temporally continuous data segments from the processed historical data to construct the input data, which includes historical time series and target time series.

8. The sea ice concentration prediction method according to any one of claims 1 to 5, characterized in that, When training the prediction model, a composite loss function for model training is constructed based on pixel-level loss and gradient loss as the optimization objective.

9. A sea ice concentration prediction system, characterized in that, The system includes: The historical data preprocessing module is used to preprocess the raw historical data to obtain input data including historical time series and target time series; A prediction model is used to make predictions based on the input data to obtain the sea ice concentration prediction result for the target time. The prediction model includes: The encoder unit is used to encode the input data and output the encoded features; The frequency enhancement state space module unit is used to process the encoded features and output the fused deep features; The decoder unit is used to decode the fused depth features and output the sea ice concentration prediction result for the target time. The frequency enhancement state space module unit includes: The spatiotemporal feature acquisition subunit is used to process the encoded features using 3D Hilbert scanning to output one-dimensional sequence features; the one-dimensional sequence features are input into a state space model for modeling and output spatiotemporal features. The frequency feature acquisition subunit is used to sequentially perform wavelet transform processing, depth convolution processing, inverse wavelet transform processing, and 3D Hilbert scan on the encoded features to output frequency features. The feature fusion subunit is used to pool and concatenate the spatiotemporal features and the frequency features to output pooled features; to shuffle the pooled features by channels to output rearranged features; to perform grouped convolution and deshuffling on the rearranged features to output spatiotemporal feature attention weights and frequency feature attention weights; to perform weighted fusion of the spatiotemporal features and frequency features according to the spatiotemporal feature attention weights and the frequency feature attention weights to output one-dimensional fused features; and to perform 3D Hilbert inverse scanning processing on the one-dimensional fused features to obtain the fused depth features.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the sea ice concentration prediction method according to any one of claims 1 to 8.

Images