A method and system for predicting arctic sea ice concentration and ice edge

Abstract

The application discloses a method and system for predicting arctic sea ice concentration and ice edge, and relates to the field of intelligent marine prediction. When the prediction time span of the sea ice concentration is less than or equal to half a year, a first sea ice concentration prediction model is used to obtain the monthly sea ice concentration of each spatial grid within the prediction time span; or a second sea ice concentration prediction model is used to obtain the monthly sea ice concentration of each spatial grid within the prediction time span; when the prediction time span of the sea ice concentration is greater than half a year, a third sea ice concentration prediction model is used to obtain the monthly sea ice concentration of each spatial grid within the prediction time span, so as to determine the monthly sea ice range and ice edge position of the Arctic within the prediction time span. The first sea ice concentration prediction model comprises an ice edge perception module, the second sea ice concentration prediction model comprises an edge perception modulation module, and the third sea ice concentration prediction model comprises an adaptive physics information gate expert hybrid mechanism. The application can improve the prediction accuracy of the sea ice concentration and the ice edge.

Description

Technical Field

[0001] This application relates to the field of intelligent marine prediction, and in particular to a method and system for predicting Arctic sea ice concentration and ice perimeter. Background Technology

[0002] Arctic sea ice plays a vital role in the global climate system, and the prediction of sea ice concentration and ice edge on monthly to seasonal scales is of practical significance for shipping, safety, resource development, and ecological protection.

[0003] Existing technologies for predicting sea ice concentration and ice edge on monthly to seasonal scales mainly include: 1) Physical numerical model: High computational cost, sensitive to initial values / boundaries / parameters, and large prediction uncertainty; 2) Statistical / machine learning methods: have lower overhead, but are difficult to express the nonlinear evolution of sea ice and its spatiotemporal dependence across scales; 3) Deep learning-based sea ice concentration (SIC) prediction has several problems: it is difficult to balance multi-scale dependence and regional heterogeneity; errors are concentrated in the periglacial region; direct fusion of multiple climate variables may mask or interfere with SIC boundary clues; and the complexity of Transformer self-attention on high-resolution grids is too high.

[0004] Therefore, there is a need for an efficient prediction method and apparatus that can model long-range dependencies, explicitly enhance periglacial information, and be robust to multi-source inputs. Summary of the Invention

[0005] The purpose of this application is to provide a method and system for predicting Arctic sea ice concentration and ice perimeter, which can improve the accuracy of sea ice concentration and ice perimeter prediction.

[0006] To achieve the above objectives, this application provides the following solution: Firstly, this application provides a method for predicting Arctic sea ice concentration and ice perimeter, including: Acquire historical monthly mean sea ice concentrations and / or historical data of climate variables at multiple time scales in the Arctic within each spatial grid; The historical monthly average sea ice concentration at multiple time scales within each spatial grid is encoded as the first feature sequence, or the historical data of historical monthly average sea ice concentration and climate variables at multiple time scales within each spatial grid are encoded as the second feature sequence. When the sea ice concentration prediction time span is less than or equal to six months, the first sea ice concentration prediction model is used to obtain the monthly sea ice concentration of each spatial grid within the prediction time span based on the first feature sequence; or, the second sea ice concentration prediction model is used to obtain the monthly sea ice concentration of each spatial grid within the prediction time span based on the second feature sequence; the first sea ice concentration prediction model includes a state space model backbone network and an ice edge sensing module connected in sequence; the second sea ice concentration prediction model includes a state space model backbone network and an edge sensing modulation module connected in sequence. When the sea ice concentration prediction time span is greater than half a year, the third sea ice concentration prediction model is used to obtain the monthly sea ice concentration of each spatial grid within the prediction time span based on the second feature sequence; the third sea ice concentration prediction model includes a state space model backbone network and an adaptive physics information gating expert hybrid mechanism connected in sequence. Based on the monthly sea ice concentration of each spatial grid within the forecast time span, a forecast field for the monthly sea ice concentration in the Arctic within the forecast time span is formed. Based on the sea ice concentration prediction field, the monthly sea ice extent and ice edge location in the Arctic during the prediction time span are determined.

[0007] Secondly, this application provides an Arctic sea ice concentration and ice edge prediction system, including: The data acquisition unit is used to acquire historical monthly average sea ice concentrations and / or historical data of climate variables at multiple time scales in the Arctic within each spatial grid. The coding unit is used to encode the historical monthly average sea ice concentration at multiple time scales within each spatial grid into a first feature sequence, or to encode the historical data of the historical monthly average sea ice concentration and climate variables at multiple time scales within each spatial grid into a second feature sequence. A short-term prediction unit is used to obtain the monthly sea ice concentration for each spatial grid within the prediction time span based on a first feature sequence and a first sea ice concentration prediction model when the sea ice concentration prediction time span is less than or equal to six months; or, based on a second feature sequence, to obtain the monthly sea ice concentration for each spatial grid within the prediction time span using a second sea ice concentration prediction model. The first sea ice concentration prediction model includes a state-space model backbone network and an ice edge sensing module connected in sequence; the second sea ice concentration prediction model includes a state-space model backbone network and an edge sensing modulation module connected in sequence. The long-term prediction unit is used to obtain the monthly sea ice concentration of each spatial grid within the prediction time span based on the second feature sequence and the third sea ice concentration prediction model when the sea ice concentration prediction time span is greater than half a year. The third sea ice concentration prediction model includes a state space model backbone network and an adaptive physics information gating expert hybrid mechanism connected in sequence. Prediction field forming units are used to form a prediction field of the Arctic sea ice concentration for each month within the prediction time span, based on the sea ice concentration for each month within the prediction time span of each spatial grid. The ice edge determination unit is used to determine the extent and location of the Arctic sea ice edge each month within the predicted time span, based on the sea ice concentration prediction field.

[0008] According to the specific embodiments provided in this application, this application has the following technical effects: This application provides a method and system for predicting Arctic sea ice concentration and ice edge. By explicitly extracting and modulating high-frequency features of the ice edge through an ice edge sensing module and an edge sensing modulation module, the prediction model is able to allocate more representational power to the edge region, reducing the overall ice edge error and improving edge consistency, resulting in more accurate ice edge / boundary predictions. The state-space model backbone network possesses long-range dependency modeling capabilities with controllable overhead, balancing global dependency and computational efficiency. An adaptive physics-information gating expert hybrid mechanism suppresses noisy edges and further mitigates performance degradation caused by misalignment of heterogeneous variable boundaries, making it more robust to multi-source inputs. In summary, this application improves the accuracy of sea ice concentration and ice edge prediction. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 A flowchart illustrating a method for predicting Arctic sea ice concentration and ice perimeter provided in this application embodiment; Figure 2 A schematic diagram illustrating the evolution of a sea ice concentration prediction model provided in an embodiment of this application; Figure 3 A schematic diagram of the functional modules of an Arctic sea ice concentration and ice edge prediction system provided in an embodiment of this application. Detailed Implementation

[0011] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0012] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0013] In one exemplary embodiment, such as Figure 1 As shown, a method for predicting Arctic sea ice concentration and ice perimeter is provided. This method is executed by computer equipment, specifically by a terminal or server alone, or by both a terminal and a server. In the embodiments of this application, it includes the following steps 101 to 106.

[0014] Step 101: Obtain historical monthly mean sea ice concentrations and / or historical data of climate variables at multiple time scales in the Arctic within each spatial grid.

[0015] Step 102: Encode the historical monthly average sea ice concentration at multiple time scales within each spatial grid as a first feature sequence, or encode the historical data of the historical monthly average sea ice concentration and climate variables at multiple time scales within each spatial grid as a second feature sequence.

[0016] Step 103: When the sea ice concentration prediction time span is less than or equal to half a year, the first sea ice concentration prediction model is used to obtain the monthly sea ice concentration of each spatial grid within the prediction time span based on the first feature sequence; or, the second sea ice concentration prediction model is used to obtain the monthly sea ice concentration of each spatial grid within the prediction time span based on the second feature sequence; the first sea ice concentration prediction model includes a state space model backbone network and an ice edge sensing module connected in sequence; the second sea ice concentration prediction model includes a state space model backbone network and an edge sensing modulation module connected in sequence.

[0017] Step 104: When the sea ice concentration prediction time span is greater than half a year, the third sea ice concentration prediction model is used to obtain the monthly sea ice concentration of each spatial grid within the prediction time span based on the second feature sequence; the third sea ice concentration prediction model includes a state space model backbone network and an adaptive physics information gating expert hybrid mechanism connected in sequence.

[0018] Step 105: Based on the monthly sea ice concentration of each spatial grid within the prediction time span, form a prediction field of monthly sea ice concentration in the Arctic within the prediction time span.

[0019] Step 106: Based on the sea ice concentration prediction field, determine the monthly sea ice extent and ice edge location in the Arctic within the prediction time span.

[0020] By implementing steps 101 to 106 above, predictions of Arctic sea ice concentration and ice edge are achieved on monthly to seasonal scales. Based on historical SIC sequences and optional multi-source climate reanalysis variables, a monthly average SIC prediction field for the next 1 to 12 months (expandable) is generated using a State Space Model (SSM) backbone network and an Ice Edge-Aware Block (IEAB) module.

[0021] In another exemplary embodiment of this application, step 101 described above may be replaced by steps 201 to 206.

[0022] Step 201: Obtain the historical monthly average sea ice concentration at multiple time scales in the Arctic within each polar stereo projection grid.

[0023] In this application, the grid where the SIC data is located is uniformly used as the target grid, that is, the 25km×25km grid under the NSIDC polar stereo projection (EPSG:3411), which covers a two-dimensional spatial grid of 448×304.

[0024] Step 202: Obtain historical data of climate variables at multiple time scales.

[0025] Climate variables include: 2m air temperature, sea surface temperature, sea level pressure, wind field, radiation, ocean heat content, and mixed layer depth.

[0026] Step 203: Calculate and standardize the anomalies of historical data of climate variables at multiple time scales on a monthly basis to obtain standardized climate data.

[0027] Step 204: Regrid the standardized climate data into each polar stereo projection grid using bilinear interpolation to obtain multi-timescale climate data within each polar stereo projection grid.

[0028] Climate variables were derived from ERA5 and ORAS5 reanalysis data, whose original spatial resolution and projection method differed from the SIC grid. To achieve spatial alignment of multiple variables, all climate variables were first re-grid from their original grid to the 25km polar stereo projection grid of the SIC through bilinear interpolation, thus ensuring that a set of climate variable values ​​corresponded to each SIC grid point.

[0029] After the above processing, SIC and each climate variable are aligned point by point on the same spatial grid (H×W) and spliced ​​in the channel dimension to form the H×W×C three-dimensional tensor of the model input.

[0030] Step 205: Mask the land and coast in the polar stereo projection grid of the Arctic to obtain the polar stereo projection grid of the ocean.

[0031] SIC data only has physical significance in the ocean region; the land region is not involved in prediction. By using a masking process, effective ocean grids can be clearly distinguished from invalid grids, preventing the model from learning incorrect patterns in meaningless regions such as land.

[0032] The unified land mask provided by NSIDCSIC was used and applied to all variables (including SIC and climate variables after regrinding). For invalid land and coastal grid points, their values ​​were uniformly set to 0, and these areas were ignored by the mask during model training and loss calculation. Errors were only calculated on valid ocean grid points.

[0033] Step 206: Based on the historical monthly average sea ice concentration and climate data at multiple time scales within each polar stereo projection grid of the ocean, interpolation is performed on the polar regions in the Arctic Ocean that lack data using time-dimension-based interpolation methods and / or spatial neighborhood-based interpolation methods to obtain historical data on historical monthly average sea ice concentration and climate variables at multiple time scales within each spatial grid of the Arctic.

[0034] In high-latitude regions (especially near the North Pole), there are sensor observation blind spots (polar holes) or occasional missing measurements. Without interpolation, the input data will have spatially discontinuous regions, which is detrimental to convolutional or patch-level modeling. Interpolation is used to fill in these missing values, ensuring that the input at each time step is a complete two-dimensional grid in spatial dimensions.

[0035] For the missing measurement areas near the North Pole caused by satellite observation limitations, the standard correction process built into the NSIDCCDR product is adopted: First, an interpolation method based on the time dimension is tried to interpolate the missing measurement values ​​in time; if the time interpolation still cannot fill the missing measurement, an interpolation method based on spatial neighborhood is further adopted to infer the missing measurement values ​​from the surrounding effective grid points.

[0036] By using the above two interpolation steps, we can ensure that the SIC field at each time step has complete spatial coverage.

[0037] After masking and interpolation are completed, all variables correspond to the same H×W mesh structure in space, and then serve as model inputs to participate in patch partitioning, embedding, and subsequent encoding and decoding processes.

[0038] In another exemplary embodiment of this application, the historical data of historical monthly average sea ice concentration and climate variables at multiple time scales within each spatial grid, which is encoded into a second feature sequence in step 102 above, can be replaced by steps 301 to 302.

[0039] Step 301: Use the historical monthly average sea ice concentration of the first preset time period and the historical data of climate variables of the second preset time period as input data; the first preset time period is longer than the second preset time period.

[0040] Constructing multi-timescale inputs: For example, using the previous 12 months of SIC as the long-term background, and using the previous few months (such as 3 months or 1 month) of climate variables as recent forcing information.

[0041] Step 302: Divide the input data into multiple overlapping patches and map them into a second feature sequence through patch embedding.

[0042] The input H×W×C 2D grid data is divided into overlapping blocks to obtain patches, which are then embedded. Each patch contains pixels / features and their channel information within a local spatial region. The vectors of multiple patches are arranged sequentially to form a feature sequence, which serves as the input to the subsequent encoder. Specifically: Before embedding: The input is a 2D grid (H×W×C), representing the spatial domain. After embedding: The output is a 1D sequence (N×D), representing the feature vector space. Here, N is the number of patches, and D is the feature dimension after embedding.

[0043] Patch Merging is used for downsampling in the encoder and Patch Expanding is used for upsampling in the decoder, forming an encoder-decoder structure and cross-layer jump connections.

[0044] In another exemplary embodiment of this application, Figure 2 This illustrates the improvement process of the three sea ice concentration prediction models used in this application. The first sea ice concentration prediction model corresponds to... Figure 2 The IEAB-V1 model, the second sea ice concentration prediction model, corresponds to... Figure 2 The IEAB-V2 model, the third sea ice concentration prediction model, corresponds to... Figure 2 The IEAB-V3 in the model. All three sea ice concentration prediction models use a state-space model backbone network. The obtained feature sequences are input into the state-space model backbone network (such as the visual state-space module VSSM / Vmamba structure) to obtain global spatiotemporal dependency modeling capabilities and output intermediate features Y.

[0045] 1) IEAB-V1: Basic edge attention.

[0046] The first sea ice concentration prediction model consists of a state-space model backbone network and an ice edge sensing module connected sequentially. The ice edge sensing module includes: a lightweight module, normalization and 1×1 convolutional learning gating (…). Figure 2The system consists of Conv 1×1+BN, a lightweight attention module, and a decoding unit. The lightweight module is used to subtract the local average of the intermediate features (AvgPool) from the intermediate features. 3×3 (Y)) to obtain the edge map; the edge map is normalized and gating is learned by 1×1 convolution and the Sigmoid function to generate the spatial attention weight map W. edg According to the spatial attention weight map W edg Intermediate features Y and feature sequences F in Through a lightweight attention module and decoding unit, based on formula F out =SimAM(Y⊙W+Y)+F in Output sea ice concentration; where F out The output is the sea ice concentration. SimAM is the SimAM module, and W is the spatial attention weight map. edg The abbreviation of .

[0047] 2) IEAB-V2: Powerful edge characteristics.

[0048] Extracting high-frequency structures at the periglacial level in IEAB and forming attention weights: a) For each channel feature Y c Calculate the edge response, for example, by using the Sobel operator to obtain the horizontal / vertical gradients and synthesizing them into the edge magnitude E'. c ;Y c It is the c-th channel feature obtained by slicing the intermediate feature map Y according to the channel dimension; b) Channel normalization of edge response E c = E' c / (Std(E' c )+ε), to obtain scale-independent marginal representations; Std() represents the calculation of variance, and ε represents the small constant to prevent division by 0; c) By learning gating through normalization and 1×1 convolution, we can obtain edge modulation weights W (e.g., tanh activation) that can be positive or negative, so as to enhance consistent edges and suppress noisy edges. d) Apply W to the intermediate features: Y' = Y + W ⊙ Y; and can be connected to a lightweight attention module (such as SimAM) to further highlight the discriminative response; e) Using residual connection output: F out = SimAM(Y') + F in .

[0049] Therefore, the edge-aware modulation module in IEAB-V2 includes: a Sobel operator, a channel normalization module, a learning gating module, a lightweight attention module, and a decoding unit; the state-space model backbone network outputs intermediate features based on the feature sequence; the Sobel operator is used to calculate the edge response for each channel feature in the intermediate features to obtain the edge amplitude; the channel normalization module is used to perform channel normalization on the edge amplitude to obtain a scale-independent edge representation; the learning gating module is used to perform instance normalization, 1×1 convolution, and normalization on the edge representation to obtain the edge modulation weights; based on the edge modulation weights and intermediate features, the sea ice concentration is output through the lightweight attention module, residual connections, and the decoding unit.

[0050] The improved design of IEAB-V2 is particularly well-suited for lead times as short as six months (1-6 months) in SIC-only input configurations, where the ice edge gradient is directly aligned with the top-gauge date. Compared to IEAB-V1, the normalized bias and signed gating in IEAB-V2 provide more selective edge-aware modulation, helping to suppress spurious edge responses when auxiliary variables are introduced. Nevertheless, IEAB-V2 still applies a shared edge enhancement rule across all channels, and this "one-size-fits-all" strategy remains a fundamental limitation of multivariate fusion. This inspired the introduction of the adaptive physics-information gated expert hybrid mechanism in IEAB-V3.

[0051] 3) IEAB-V3: Physics-Informed Mixture-of-Experts (PI-MoE).

[0052] The PI-MoE proposed in this application is a physics-driven, channel-specific gated hybrid. Its main purpose is to carefully tailor the feature representations according to the core physical meaning. A major limitation of the IEAB-V2 design lies in the fundamental statistical differences between SIC and climate variables: sharp spatial discontinuities constitute meaningful signals for SIC but represent noise in smooth atmospheric and oceanic fields. To address this issue, a hybrid expert framework explicitly guided by physical priors is introduced, enabling adaptive, per-channel routing to specialized processing units. This design allows edge enhancements to be applied only where physically meaningful, while preserving large-scale climate information elsewhere. The PI-MoE consists of two parts: a dual-expert framework and a lightweight routing network. The dual experts reflect two main physical mechanisms in the input variables: the SIC field requiring edge-sensitive processing, and the smooth atmospheric / oceanic field, where such operations can introduce spurious artifacts. The routing mechanism does not rely on abstract learned statistics but rather on the feature map Y. c Two physically interpretable per-channel statistics are computed to drive routing decisions.

[0053] The adaptive physics-gated expert hybrid mechanism includes: the Sobel operator, edge experts, global experts, a routing network, a lightweight attention module, and a decoding unit. The Sobel operator is used to calculate the edge response for each channel feature in the intermediate features to obtain the edge magnitude.

[0054] Edge experts are used to obtain edge features based on the edge magnitude. Edge experts are optimized for SIC-class channels where sharp boundaries are crucial. It directly applies this to the Sobel edge map E. c The above operation is performed, followed by depthwise convolution, and then tanh activation to improve gradient information while preserving edge polarity: F edge =tanh(Conv 3×3 (E c )).

[0055] The Global Expert is used to obtain global features based on intermediate features; specifically designed for climate-like channels, it features spatial smoothness and large-scale consistency. It processes intermediate features Y using convolutions with a large receptive field (5×5), followed by grouped normalization and sigmoid activation to generate a stable attention map emphasizing broad climate patterns. F global = σ(GroupNorm(Conv 5×5 (Y))).

[0056] The spatial smoothness of intermediate features is used as the first input data, and the edge energy of edge magnitudes is used as the second input data. Both are input into the routing network, which outputs edge weights and global weights. The first statistic of the routing network—edge energy—is defined as the spatial average of the Sobel gradient magnitudes. .

[0057] The second statistic is spatial smoothness, which measures spatial homogeneity by comparing local variability and global variability: .

[0058] in This represents the set of all valid 3×3 patches extracted from Y, with a stride of 1. `Var()` represents the variance operator. For numerical stability, these two statistics are linearly normalized to [0,1] in the current batch before being input into the router. Routing is performed at the channel level rather than the spatial level, naturally aligning with the semantic organization of the learned feature maps, where different channels encode different physical variables. Feature vectors The data is passed to a lightweight MLP router, which outputs the expert weight per channel. α = .

[0059] Based on edge weights and global weights, edge features and global features are adaptively fused, and then intermediate features are modulated to obtain modulated features. Based on the modulated features and feature sequence, sea ice concentration is output through a lightweight attention module and decoding unit. This process can be expressed by the following formula: .

[0060] Therefore, multi-expert periglacial sensing: Addressing the differences in statistical characteristics between SIC and climate variables, "edge experts" and "global experts" are established, and weights are generated via a routing network based on physically interpretable indicators such as edge energy and spatial smoothness for each channel. α The expert output is adaptively fused before Y is modulated to reduce the negative interference of multiple source variables on ice edge prediction. The best results are achieved when the prediction window reaches the annual prediction (predicting the monthly average SIC for the next 12 months in one go).

[0061] The decoder combines skip connections to restore spatial resolution step by step; finally, the features are mapped to SIC prediction values ​​through 1×1 convolution and range constraints are applied (e.g., cropped to [0,1]). The monthly average SIC prediction field for multiple future lead times is output; and the sea ice extent and ice edge location can be derived from a threshold (e.g., SIC>0.15).

[0062] The PI-MoE design aims to reduce the negative interference of various auxiliary climate variables by sending large-scale information related to the margins to expert witnesses. Therefore, the model can incorporate additional climate predictors while maintaining a consistent representation of the boundaries, which is particularly beneficial for multivariate and longer-term sea ice forecasts.

[0063] In another exemplary embodiment of this application, the first, second, and third sea ice concentration prediction models all include a single-month prediction mode and a multi-month prediction mode. The single-month prediction mode means that the model predicts the sea ice concentration for the following month only, repeating the prediction multiple times to obtain the sea ice concentration for each month within the prediction time span; in the single-month prediction mode, planned sampling and / or teacher coercion are introduced during model training. The multi-month prediction mode means that the model generates the sea ice concentration for each month within the prediction time span all at once.

[0064] In another exemplary embodiment of this application, training and long-term rolling prediction robustness are improved: The method of this application is used to enhance the prediction robustness of the model. When the model is configured to output the monthly average SIC for the next month, this method can be used to autoregressively predict the SIC for multiple months in the future during the training phase, giving it long-term predictive capabilities. Mask loss is used to calculate the error on ocean pixels (e.g., a combination of MAE and RMSE), and ice edge-related indicators (such as Integrated Ice-Edge Error, IIEE) can be added as evaluation or auxiliary constraints. Planned sampling / teacher coercion is introduced in the single-month prediction mode: During training, the model selects to use either the true SIC or the model prediction as the next input with a probability that decays with training progress (decaying linearly or inversely according to sigmoid from 1 to 0), thereby improving the stability of multi-step rolling prediction.

[0065] Compared with the prior art, this application has at least the following advantages: (1) More accurate ice edge / boundary prediction: By explicitly extracting and modulating high-frequency features of the ice edge through IEAB, the model can allocate more representational capabilities in the marginal ice zone (MIZ), which can reduce the overall ice edge error and improve edge consistency.

[0066] (2) Balancing global dependency and computational efficiency: The state-space model backbone has the ability to model long-range dependencies and has controllable overhead, making it suitable for seasonal-scale prediction of high-resolution grid data.

[0067] (3) More robust to multi-source input: channel standardization + gating suppresses noise edges; multi-expert routing further alleviates the performance degradation caused by misalignment of heterogeneous variable boundaries.

[0068] (4) Long-term rolling prediction is more stable: planned sampling / teacher forced reduction of error accumulation and edge drift in multi-step rolling.

[0069] The above advantages stem from the following key technologies: edge extraction, standardization, modulated weights that can be positive or negative and residual attention structure; multi-expert routing and fusion; and planned sampling training strategy.

[0070] Based on the same inventive concept, this application also provides an Arctic sea ice concentration and perimeter prediction system for implementing the methods described above. The solution provided by this system is similar to the implementation scheme described in the above methods; therefore, the specific limitations of one or more Arctic sea ice concentration and perimeter prediction system embodiments provided below can be found in the limitations of the above methods described above, and will not be repeated here.

[0071] In one exemplary embodiment, such as Figure 3As shown, an Arctic sea ice concentration and ice edge prediction system is provided, comprising: a data acquisition unit, an encoding unit, a short-term prediction unit, a long-term prediction unit, a prediction field formation unit, and an ice edge determination unit.

[0072] The data acquisition unit is used to acquire historical monthly mean sea ice concentrations and / or historical data of climate variables at multiple time scales within each spatial grid in the Arctic. The encoding unit is used to encode the historical monthly mean sea ice concentrations at multiple time scales within each spatial grid into a first feature sequence, or to encode the historical data of historical monthly mean sea ice concentrations and climate variables at multiple time scales within each spatial grid into a second feature sequence.

[0073] A short-term prediction unit is used to obtain the monthly sea ice concentration of each spatial grid within the prediction time span based on a first feature sequence and a first sea ice concentration prediction model when the sea ice concentration prediction time span is less than or equal to six months; or, based on a second feature sequence, to obtain the monthly sea ice concentration of each spatial grid within the prediction time span using a second sea ice concentration prediction model. The first sea ice concentration prediction model includes a state-space model backbone network and an ice edge sensing module connected in sequence; the second sea ice concentration prediction model includes a state-space model backbone network and an edge sensing modulation module connected in sequence.

[0074] The long-term prediction unit is used to obtain the monthly sea ice concentration of each spatial grid within the prediction time span based on the second feature sequence and the third sea ice concentration prediction model when the sea ice concentration prediction time span is greater than half a year. The third sea ice concentration prediction model includes a state space model backbone network and an adaptive physics information gating expert hybrid mechanism connected in sequence.

[0075] A prediction field forming unit is used to form a monthly sea ice concentration prediction field for the Arctic within the prediction time span, based on the monthly sea ice concentration of each spatial grid. An ice edge determination unit is used to determine the monthly sea ice extent and ice edge location of the Arctic within the prediction time span, based on the predicted sea ice concentration field.

[0076] As an optional implementation, the system also includes: a visualization unit; the visualization unit is used to display the predicted sea ice concentration field, sea ice extent, and ice edge location for each month in the Arctic over the predicted time span.

[0077] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0078] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for predicting Arctic sea ice concentration and ice perimeter, characterized in that, include: Acquire historical monthly mean sea ice concentrations and / or historical data of climate variables at multiple time scales in the Arctic within each spatial grid; The historical monthly average sea ice concentration at multiple time scales within each spatial grid is encoded as the first feature sequence, or the historical data of historical monthly average sea ice concentration and climate variables at multiple time scales within each spatial grid are encoded as the second feature sequence. When the sea ice concentration prediction time span is less than or equal to six months, the first sea ice concentration prediction model is used to obtain the monthly sea ice concentration of each spatial grid within the prediction time span based on the first feature sequence; or, the second sea ice concentration prediction model is used to obtain the monthly sea ice concentration of each spatial grid within the prediction time span based on the second feature sequence; the first sea ice concentration prediction model includes a state space model backbone network and an ice edge sensing module connected in sequence; the second sea ice concentration prediction model includes a state space model backbone network and an edge sensing modulation module connected in sequence. When the sea ice concentration prediction time span is greater than half a year, the third sea ice concentration prediction model is used to obtain the monthly sea ice concentration of each spatial grid within the prediction time span based on the second feature sequence; the third sea ice concentration prediction model includes a state space model backbone network and an adaptive physics information gating expert hybrid mechanism connected in sequence. Based on the monthly sea ice concentration of each spatial grid within the prediction time span, a prediction field for the monthly sea ice concentration in the Arctic within the prediction time span is formed. Based on the sea ice concentration prediction field, the monthly sea ice extent and ice edge location in the Arctic during the prediction time span are determined.

2. The method for predicting Arctic sea ice concentration and ice perimeter according to claim 1, characterized in that, Historical data on monthly mean sea ice concentrations and climate variables at multiple time scales in the Arctic were obtained within each spatial grid, specifically including: Obtain historical monthly average sea ice concentrations at multiple time scales within each polar stereo projection grid in the Arctic; Acquire historical data of climate variables across multiple time scales; Anomalies of climate variables at multiple time scales are calculated monthly and standardized to obtain standardized climate data. Standardized climate data is re-grid into each polar stereo projection grid through bilinear interpolation to obtain climate data at multiple time scales within each polar stereo projection grid. Anomalies were calculated and standardized monthly for climate data at multiple time scales within each polar stereo projection grid to obtain standardized climate data within each polar stereo projection grid. Masking is applied to the land and coast in the polar stereo projection grid of the Arctic to obtain the polar stereo projection grid of the ocean. Based on historical monthly mean sea ice concentration and climate data at multiple time scales within each polar stereo projection grid of the ocean, interpolation methods based on time dimension and / or based on spatial neighborhood are used to interpolate the polar regions in the Arctic Ocean that lack data, thereby obtaining historical data on historical monthly mean sea ice concentration and climate variables at multiple time scales within each spatial grid of the Arctic.

3. The method for predicting Arctic sea ice concentration and ice perimeter according to claim 1, characterized in that, Historical data on monthly average sea ice concentrations and climate variables at multiple time scales within each spatial grid are encoded into a second feature sequence, specifically including: The historical monthly average sea ice concentration for the first preset time period and the historical data of climate variables for the second preset time period are used together as input data; the first preset time period is longer than the second preset time period; The input data is divided into multiple overlapping patches, and mapped to a second feature sequence through patch embedding.

4. The method for predicting Arctic sea ice concentration and ice perimeter according to claim 1, characterized in that, The ice edge sensing module includes: a lightweight module, a normalization and 1×1 convolutional learning gating, a lightweight attention module, and a decoding unit; The state-space model backbone network outputs intermediate features based on the feature sequence; The lightweight module is used to subtract the local average value of the intermediate features from the intermediate features to obtain the edge map; The edge map is normalized and learned by 1×1 convolution gating and Sigmoid function to generate a spatial attention weight map; Based on the spatial attention weight map, intermediate features, and feature sequences, a lightweight attention module and a decoding unit are used, according to formula F. out =SimAM(Y⊙W+Y)+F in Output sea ice concentration; where F out For the output sea ice concentration, SimAM is the SimAM module, Y is the intermediate feature, W is the spatial attention weight map, and F is the output sea ice concentration. in It is a characteristic sequence.

5. The method for predicting Arctic sea ice concentration and ice perimeter according to claim 1, characterized in that, The edge-aware modulation module includes: a Sobel operator, a channel normalization module, a learning gating module, a lightweight attention module, and a decoding unit; The state-space model backbone network outputs intermediate features based on the feature sequence; The Sobel operator is used to calculate the edge response for each channel feature in the intermediate features to obtain the edge magnitude. The channel normalization module is used to normalize the edge amplitude to obtain scale-independent edge representations; The learning gating module is used to perform 1×1 convolution and normalization on the edge representation to obtain edge modulation weights; Based on edge modulation weights and intermediate features, the sea ice concentration is output through a lightweight attention module, residual connections, and a decoding unit.

6. The method for predicting Arctic sea ice concentration and ice perimeter according to claim 1, characterized in that, The adaptive physics-based information gating expert hybrid mechanism includes: Sobel operator, edge expert, global expert, routing network, lightweight attention module, and decoding unit; The Sobel operator is used to calculate the edge response for each channel feature in the intermediate features to obtain the edge magnitude. Edge experts are used to obtain edge features based on the edge amplitude; Global experts are used to obtain global features based on intermediate features; The spatial smoothness of intermediate features is used as the first input data, and the edge energy of edge magnitude is used as the second input data. Both are input into the routing network, and the edge weights and global weights are output. Based on edge weights and global weights, edge features and global features are adaptively fused, and then intermediate features are modulated to obtain modulated features. Based on the modulated features and feature sequences, the sea ice concentration is output through a lightweight attention module and a decoding unit.

7. The method for predicting Arctic sea ice concentration and ice perimeter according to claim 1, characterized in that, The first sea ice concentration prediction model, the second sea ice concentration prediction model and the third sea ice concentration prediction model all include a single-month prediction mode and a multi-month prediction mode; The single-month prediction mode refers to the model predicting the sea ice concentration for the next month only each time, and repeating the model prediction multiple times to obtain the sea ice concentration for each month within the prediction time span; in the single-month prediction mode, planned sampling and / or teacher coercion are introduced during model training. A multi-month forecast model refers to a model that generates the monthly sea ice concentration for a single forecast period.

8. The method for predicting Arctic sea ice concentration and ice perimeter according to claim 1, characterized in that, During training, the first sea ice concentration prediction model, the second sea ice concentration prediction model, and the third sea ice concentration prediction model use mask loss to calculate the error on ocean pixels.

9. A system for predicting Arctic sea ice concentration and ice perimeter, characterized in that, include: The data acquisition unit is used to acquire historical monthly average sea ice concentrations and / or historical data of climate variables at multiple time scales in the Arctic within each spatial grid. The coding unit is used to encode the historical monthly average sea ice concentration at multiple time scales within each spatial grid into a first feature sequence, or to encode the historical data of the historical monthly average sea ice concentration and climate variables at multiple time scales within each spatial grid into a second feature sequence. A short-term prediction unit is used to obtain the monthly sea ice concentration for each spatial grid within the prediction time span based on a first feature sequence and a first sea ice concentration prediction model when the sea ice concentration prediction time span is less than or equal to six months; or, based on a second feature sequence, to obtain the monthly sea ice concentration for each spatial grid within the prediction time span using a second sea ice concentration prediction model. The first sea ice concentration prediction model includes a state-space model backbone network and an ice edge sensing module connected in sequence; the second sea ice concentration prediction model includes a state-space model backbone network and an edge sensing modulation module connected in sequence. The long-term prediction unit is used to obtain the monthly sea ice concentration of each spatial grid within the prediction time span based on the second feature sequence and the third sea ice concentration prediction model when the sea ice concentration prediction time span is greater than half a year. The third sea ice concentration prediction model includes a state space model backbone network and an adaptive physics information gating expert hybrid mechanism connected in sequence. Prediction field forming units are used to form a prediction field of the Arctic sea ice concentration for each month within the prediction time span, based on the sea ice concentration for each month within the prediction time span of each spatial grid. The ice edge determination unit is used to determine the extent and location of the Arctic sea ice edge each month within the predicted time span, based on the sea ice concentration prediction field.

10. The Arctic sea ice concentration and perimeter prediction system according to claim 9, characterized in that, Also includes: Visual unit; The visualization unit is used to display the predicted sea ice concentration field, sea ice extent, and ice edge location for each month in the Arctic over the predicted time span.

Images