A polar sea ice multi-scale prediction method and system

Abstract

This application discloses a multi-scale forecasting method and system for polar sea ice, relating to the field of climate forecasting technology. The method includes: acquiring raw sea ice concentration observation data; preprocessing the raw sea ice concentration observation data to obtain preprocessed data; extracting multi-scale spatial features from the preprocessed data to obtain feature tokens at multiple scales; aligning the feature tokens at multiple scales by feature dimensions to obtain aligned features; performing cross-scale attention fusion on the aligned features to obtain cross-scale fused features; performing intra-scale modeling optimization on the cross-scale fused features to obtain intra-scale optimized features; and performing linear mapping restoration and spatial decoding on the intra-scale optimized features to obtain sea ice concentration forecast results. The sea ice concentration forecast results include future daily, weekly, and monthly sea ice concentration forecasts. This application can solve the problems of single-scale modeling and information silos in existing technologies.

Description

Technical Field

[0001] This application relates to the field of climate forecasting technology, and in particular to a multi-scale forecasting method and system for polar sea ice. Background Technology

[0002] Arctic sea ice plays a crucial role in the global climate system, significantly impacting polar ecological stability and human activities along the coast. In recent years, artificial intelligence-based sea ice forecasting technologies have made remarkable progress, with data-driven methods demonstrating the potential to surpass traditional physical numerical models in forecast accuracy, computational efficiency, and forecast timeliness.

[0003] However, existing deep learning forecasting models still have significant technical bottlenecks: First, the lead time of most advanced forecasts is limited to the sub-seasonal scale, making it difficult to achieve continuous daily forecasts from the sub-seasonal to the seasonal scale; second, existing schemes lack the ability to collaboratively model the characteristics of sea ice changes at multiple time scales, and cannot effectively utilize the complementarity between information at different scales; third, current methods mostly adopt independent modeling at a single scale and have failed to establish a cross-scale information interaction mechanism, resulting in limited long-term forecast accuracy.

[0004] Therefore, how to solve the problems of single-scale modeling and information silos in existing technologies has become an urgent technical problem to be solved in this field. Summary of the Invention

[0005] The purpose of this application is to provide a multi-scale forecasting method and system for polar sea ice, which can solve the problems of single-scale modeling and information silos in the existing technology.

[0006] To achieve the above objectives, this application provides the following solution.

[0007] In a first aspect, this application provides a multi-scale forecasting method for polar sea ice, which includes the following steps.

[0008] Obtain raw sea ice concentration observation data.

[0009] The original sea ice concentration observation data is preprocessed to obtain preprocessed data; the preprocessed data includes: daily sea ice concentration data, weekly sea ice concentration data and monthly sea ice concentration data.

[0010] Multi-scale spatial feature extraction is performed on the preprocessed data to obtain feature tokens at multiple scales.

[0011] The feature tokens at the multiple scales are aligned in terms of feature dimensions to obtain aligned features; the aligned features include: daily scale aligned features, weekly scale aligned features and monthly scale aligned features.

[0012] The aligned features are then subjected to cross-scale attention fusion to obtain cross-scale fused features.

[0013] The cross-scale fused features are then subjected to intra-scale modeling optimization to obtain intra-scale modeling optimized features.

[0014] Linear mapping and spatial decoding are performed on the optimized features within the specified scale to obtain sea ice concentration forecast results. The sea ice concentration forecast results include: future daily-scale sea ice concentration forecasts, future weekly-scale sea ice concentration forecasts, and future monthly-scale sea ice concentration forecasts.

[0015] Optionally, multi-scale spatial feature extraction is performed on the preprocessed data to obtain feature tokens at multiple scales, specifically including the following steps.

[0016] The preprocessed data is divided into image blocks for pre-initial embedding to obtain initial embedding features.

[0017] Based on the initial embedded features, multi-scale feature extraction is performed by a spatial feature encoder to obtain the final 2D spatial features; the spatial feature encoder includes several pairs of continuous blocks.

[0018] The final 2D spatial features are linearly projected and transformed to obtain feature tokens of multiple scales.

[0019] Optionally, the calculation formula for each pair of consecutive blocks is as follows.

[0020] .

[0021] in, For the first Intermediate features of a layer after layer normalization and attention calculation; For layer normalization; For multi-head self-attention of windows; For the first Features of the layer output; For the first Features of the layer output; It is a multilayer perceptron; For the first Intermediate features of a layer after layer normalization and attention calculation; For multi-head self-attention in shifted windows; For the first Features of the layer output.

[0022] Optionally, the expression for the linear projection transformation is as follows.

[0023] .

[0024] in, Feature tokens for multiple scales; For flattening operation; This is a linear transformation operation; The first Features of the layer output; This represents the total number of layers in the encoder.

[0025] Optionally, the expression for feature dimension alignment is as follows.

[0026] .

[0027] in, Features after alignment; This is a linear transformation operation; For feature tokens of multiple scales.

[0028] Optionally, the expression for the cross-scale attention fusion is as follows.

[0029] .

[0030] .

[0031] .

[0032] in, It is a multi-granularity feature sequence; This represents the feature representation after cross-scale fusion. This is a multi-head attention mechanism; This is a query, key, and value matrix obtained by linear transformation of granular features; The dimension of the key vector; This is a daily-scale alignment feature; This is a periodic-scale alignment feature; Alignment features at the monthly scale; It is a normalized exponential function.

[0033] Optionally, the expression for the intra-scale modeling optimization is as follows.

[0034] .

[0035] in, Features optimized for in-scale modeling; This represents the feature representation after cross-scale fusion. This is the first learnable weight in FFN; This is the second learnable weight in FFN; This is the first learnable bias in FFN; This is the second learnable bias in FFN; This is the activation function.

[0036] Optionally, the expression for the linear mapping restoration is as follows.

[0037] .

[0038] in, This is the restored original scale feature sequence; Features optimized for in-scale modeling; This is the first learnable parameter of the reduction layer; This is the second learnable parameter of the reduction layer; It is a daytime scale; It is a periodic scale; It is measured on a lunar scale.

[0039] Optionally, the expression for the spatial decoding is as follows.

[0040] .

[0041] in, For scale Sea ice concentration forecast results; This is the restored original scale feature sequence; It is a daytime scale; It is a periodic scale; Measured on a lunar scale; This is a spatial decoding function.

[0042] Secondly, this application provides a polar sea ice multi-scale forecasting system, which is used to implement the polar sea ice multi-scale forecasting method described in any one of the first aspects, and the polar sea ice multi-scale forecasting system includes the following modules.

[0043] The data acquisition module is used to acquire raw sea ice concentration observation data.

[0044] The data preprocessing module is used to preprocess the original sea ice concentration observation data to obtain preprocessed data; the preprocessed data includes: daily sea ice concentration data, weekly sea ice concentration data and monthly sea ice concentration data.

[0045] The spatial feature encoding module is used to extract multi-scale spatial features from the preprocessed data to obtain feature tokens at multiple scales.

[0046] The feature dimension alignment module is used to align the feature tokens of the multiple scales to obtain aligned features; the aligned features include: daily scale aligned features, weekly scale aligned features and monthly scale aligned features.

[0047] The cross-scale attention fusion module is used to perform cross-scale attention fusion on the aligned features to obtain cross-scale fused features.

[0048] The intra-scale feature modeling module is used to perform intra-scale modeling optimization on the cross-scale fused features to obtain intra-scale modeling optimized features.

[0049] The feature restoration and decoding module is used to perform linear mapping restoration and spatial decoding on the optimized features within the specified scale to obtain sea ice concentration forecast results. These sea ice concentration forecast results include: future daily-scale sea ice concentration forecasts, future weekly-scale sea ice concentration forecasts, and future monthly-scale sea ice concentration forecasts. Based on the specific embodiments provided in this application, the following technical effects are disclosed.

[0050] This application provides a method and system for multi-scale forecasting of polar sea ice. The method includes: acquiring raw sea ice concentration observation data; preprocessing the raw sea ice concentration observation data to obtain preprocessed data; the preprocessed data includes: daily-scale sea ice concentration data, weekly-scale sea ice concentration data, and monthly-scale sea ice concentration data; by processing the raw observation data into daily, weekly, and monthly standard-scale data, a data foundation framework for multi-scale forecasting can be constructed. Multi-scale spatial feature extraction is performed on the preprocessed data to obtain feature tokens at multiple scales; feature dimension alignment is performed on the feature tokens at multiple scales to obtain aligned features; the aligned features include: daily-scale aligned features, weekly-scale aligned features, and monthly-scale aligned features; by extracting spatial feature information from data at different scales, the raw data is transformed into feature tokens with spatial distinctiveness; and by unifying the dimensional standards of feature tokens at different scales, feature fusion barriers caused by scale differences can be eliminated. Cross-scale attention fusion is performed on the aligned features to obtain cross-scale fused features; by integrating complementary information of features at different scales through a cross-scale attention mechanism, the multi-scale correlation of features can be enhanced. The cross-scale fused features are then optimized through intra-scale modeling to obtain optimized intra-scale features. By refining the intra-scale modeling of the fused features, the forecast accuracy of single-scale features can be improved. The optimized intra-scale features are then linearly mapped and spatially decoded to obtain sea ice concentration forecast results. These forecast results include future daily, weekly, and monthly sea ice concentration forecasts. By restoring and decoding the optimized features into specific sea ice concentration forecast results, multi-scale polar sea ice forecasts at daily, weekly, and monthly scales can be achieved. This application constructs a polar sea ice concentration forecast method covering daily, weekly, and monthly scales through a complete process design of multi-scale data processing, feature extraction and fusion, and modeling optimization. This method effectively integrates spatial feature information of sea ice at different scales, eliminates scale dimensional differences, and improves the accuracy and multi-scale adaptability of polar sea ice concentration forecasts, providing comprehensive and reliable technical support for polar sea ice disaster early warning and marine resource development. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, 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.

[0052] Figure 1 This is an application environment diagram of a multi-scale forecasting method for polar sea ice in one embodiment of this application.

[0053] Figure 2 This is a flowchart illustrating a multi-scale forecasting method for polar sea ice provided in an embodiment of this application.

[0054] Figure 3 This is a schematic diagram of the functional modules of a polar sea ice multi-scale forecasting system provided in an embodiment of this application.

[0055] Figure 4 This is a schematic diagram of the functional modules of a polar sea ice multi-scale forecasting system provided in another embodiment of this application. Detailed Implementation

[0056] 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.

[0057] This application discloses a multi-scale forecasting method for polar sea ice: by using sea ice concentration observation data products, extracting data row features through a multi-scale modeling AI model, and simultaneously completing the forecasting of sea ice concentration at multiple time scales.

[0058] Specifically, this application discloses a method for predicting Arctic sea ice concentration based on multi-scale artificial intelligence, belonging to the fields of earth science and climate forecasting technology. The method includes: acquiring sea ice concentration observation data products; constructing a dataset after preprocessing data at multiple time scales; extracting sea ice spatial distribution features at each time scale (e.g., daily, weekly, and monthly averages) using a unified spatial feature encoder; aligning the encoded feature sequences of sea ice concentration distribution at multiple time scales (e.g., 7 days, 8 weeks, and 6 months) along the feature dimension to establish a unified multi-time scale latent space representation; fusing multi-scale latent space features through a cross-scale attention mechanism, with intra-scale information modeled by an artificial neural network; and finally, simultaneously outputting sea ice concentration changes at multiple time scales. Compared with existing technologies, this application can simultaneously perform cross-scale modeling of sea ice concentration at multiple time scales, improving the forecasting skill at each scale. Experiments show that performance improvements have been achieved on multiple timescale forecasting tasks using the G02202 dataset from the National Snow and Ice Data Center (NSIDC) from 2016 to 2023, enabling better forecasting of extreme events (sea ice area during the melting period).

[0059] 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.

[0060] The multi-scale forecasting method for polar sea ice provided in this application can be applied to, for example... Figure 1 In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be set up independently, integrated into server 104, or placed in the cloud or on other servers. Terminal 102 can send raw sea ice concentration observation data to server 104. After receiving the raw sea ice concentration observation data, server 104 preprocesses the raw sea ice concentration observation data to obtain preprocessed data. The preprocessed data includes: daily-scale sea ice concentration data, weekly-scale sea ice concentration data, and monthly-scale sea ice concentration data. Multi-scale spatial feature extraction is performed on the preprocessed data to obtain feature tokens at multiple scales. Feature dimension alignment is performed on the feature tokens at multiple scales to obtain aligned features. The aligned features include: daily-scale aligned features, weekly-scale aligned features, and monthly-scale aligned features; the aligned features are fused across scales to obtain cross-scale fused features; the cross-scale fused features are optimized by intra-scale modeling to obtain optimized intra-scale modeling features; the optimized intra-scale modeling features are linearly mapped and spatially decoded to obtain sea ice concentration forecast results; the sea ice concentration forecast results include: future daily-scale sea ice concentration forecasts, future weekly-scale sea ice concentration forecasts, and future monthly-scale sea ice concentration forecasts. Server 104 can feed back the obtained sea ice concentration forecast results to terminal 102. In addition, in some embodiments, the polar sea ice multi-scale forecasting method can also be implemented by server 104 or terminal 102 alone. For example, terminal 102 can directly perform polar sea ice multi-scale forecasting based on the original sea ice concentration observation data, or server 104 can obtain the original sea ice concentration observation data from the data storage system and perform polar sea ice multi-scale forecasting based on the original sea ice concentration observation data.

[0061] The terminal 102 can be, but is not limited to, various desktop computers, laptops, smartphones, and tablets. Portable wearable devices can be smartwatches, smart bracelets, head-mounted devices, etc. The server 104 can be implemented using a standalone server or a server cluster composed of multiple servers, or it can be a cloud server.

[0062] In one exemplary embodiment, such as Figure 2As shown, a multi-scale forecasting method for polar sea ice 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 this embodiment, the method is applied to... Figure 1 Taking server 104 as an example, the following steps are included.

[0063] S1: Obtain raw sea ice concentration observation data.

[0064] S2: Preprocess the original sea ice concentration observation data to obtain preprocessed data; the preprocessed data includes: daily sea ice concentration data, weekly sea ice concentration data and monthly sea ice concentration data.

[0065] S3: Perform multi-scale spatial feature extraction on the preprocessed data to obtain feature tokens at multiple scales.

[0066] S4: Align the feature tokens of the multiple scales according to their feature dimensions to obtain aligned features; the aligned features include: daily scale aligned features, weekly scale aligned features and monthly scale aligned features.

[0067] S5: Perform cross-scale attention fusion on the aligned features to obtain cross-scale fused features.

[0068] S6: Perform intra-scale modeling optimization on the cross-scale fused features to obtain intra-scale modeling optimized features.

[0069] S7: Perform linear mapping restoration and spatial decoding on the optimized features within the scale to obtain sea ice concentration forecast results; the sea ice concentration forecast results include: future daily-scale sea ice concentration forecast, future weekly-scale sea ice concentration forecast and future monthly-scale sea ice concentration forecast.

[0070] By implementing steps S1 to S7 above, and through a full-process design of multi-scale data processing, feature extraction and fusion, and modeling optimization, a polar sea ice concentration forecasting method covering daily, weekly, and monthly scales was constructed. This method effectively integrates spatial feature information of sea ice at different scales, eliminates scale dimensional differences, and improves the accuracy and multi-scale adaptability of polar sea ice concentration forecasting. It provides comprehensive and reliable technical support for polar sea ice disaster early warning and marine resource development.

[0071] As an optional implementation, in step S1, the computer system acquires raw sea ice concentration observation data products from institutions such as the U.S. National Snow and Ice Data Center.

[0072] As an optional implementation, in step S2, the acquired data is preprocessed to construct the model input. Preprocessing includes cleaning the raw data at different time scales and removing invalid values.

[0073] This step transforms the raw observation data into a normalized input suitable for model processing.

[0074] As an optional implementation, in step S3, multi-scale spatial feature extraction is performed on the preprocessed data to obtain feature tokens at multiple scales, specifically including the following steps.

[0075] S31: Perform image block division and pre-initial embedding on the preprocessed data to obtain initial embedding features.

[0076] S32: Based on the initial embedded features, multi-scale feature extraction is performed through a spatial feature encoder to obtain the final 2D spatial features; the spatial feature encoder includes several pairs of continuous blocks.

[0077] S33: Perform linear projection transformation on the final 2D spatial features to obtain feature tokens of multiple scales.

[0078] Specifically, the computer system inputs the preprocessed daily, weekly, and monthly sea ice concentration data from S2 into a shared spatial feature encoder. This encoder, based on the Swing Transformer V2 architecture, performs hierarchical spatial feature extraction.

[0079] The image block segmentation and initial embedding formulas are shown below.

[0080] .

[0081] in, These are the initial embedded features; Preprocessed input data (two-dimensional spatial data, with dimensions of [height × width × number of channels]) at the daily (d), weekly (w), and monthly (m) scales. For image patch partitioning, the two-dimensional spatial data is divided into non-overlapping image patches of a fixed size (e.g., 4×4 pixels). Then, each image patch is mapped to a feature vector of a fixed dimension through linear projection to obtain the initial embedded features. (The dimension is [number of image patches × embedding dimension]).

[0082] Hierarchical feature encoding.

[0083] pass L Multi-scale feature extraction is performed using the Swing Transformer layer blocks, and the calculation for each pair of consecutive blocks is shown below.

[0084] .

[0085] in, For the first Intermediate features of a layer after layer normalization and attention calculation; Layer normalization is a process that normalizes each dimension of the feature vector to reduce the risk of gradient vanishing and accelerate model training convergence. (Window-based Multi-Head Self-Attention) divides the feature map into multiple local windows and calculates self-attention only within each window, reducing computational complexity while capturing local spatial correlations. For the first Features of the layer output; For the first Features of the layer output; The Multi-Layer Perceptron (MLP) consists of linear transformations and activation functions. It performs nonlinear transformations on the features output by the attention mechanism to enhance feature representation capabilities. For the first Intermediate features of a layer after layer normalization and attention calculation; (Shifted Window Multi-Head Self-Attention) is a shifted window multi-head self-attention that shifts the window position in adjacent layers to achieve information interaction between different windows, making up for the locality limitation of W-MSA and capturing long-distance spatial dependencies; For the first Features of the layer output.

[0086] Spatial feature tokenization.

[0087] The final 2D spatial features are converted into 1D tokens through linear projection, as shown in the following formula.

[0088] .

[0089] in, Feature tokens for multiple scales; For the flattening operation, the 2D feature map (with dimensions of [number of image patches × feature dimension]) is converted into a 1D vector; This is a linear transformation operation that maps the flattened vector to feature tokens of a uniform dimension. The first Features of the layer output; The total number of layers in the encoder (set according to data complexity and forecasting requirements, such as 6 or 8 layers) is determined by the number of layers. The more layers there are, the higher the level of abstraction in feature extraction.

[0090] This step maps sea ice concentration data from different time scales to a unified feature space using a shared encoder, ensuring consistency in feature representation while reducing model parameters. The shifted window attention mechanism effectively captures the local spatial correlations of sea ice, providing high-quality spatial feature representations for subsequent cross-scale fusion.

[0091] As an optional implementation, in step S4, the computer system receives the feature tokens of multiple scales output in step S3. Since daily, weekly, and monthly scales encompass different time spans, their sequence dimensions exhibit mismatch issues. The computer system utilizes a linear transformation to project these feature sequences onto a unified feature dimension. The expression for this feature dimension alignment is shown below.

[0092] .

[0093] in, For aligned features, including day-scale aligned features Weekly-scale alignment features Alignment features with lunar scale ; This is a linear transformation operation; For feature tokens of multiple scales (which are the original feature tokens of scale s (day, week, month).

[0094] This step can solve the problem that features at different scales cannot be directly fused due to dimensional differences, and provides a unified feature platform for the next step of cross-scale information interaction.

[0095] As an optional implementation, in step S5, the computer system defines the aligned multi-scale features from step S4 as... The input is then fed into a cross-scale attention fusion module. This module employs an attention mechanism to automatically calculate and fuse the dependencies between features at different time scales. The expression for the cross-scale attention fusion is shown below.

[0096] .

[0097] .

[0098] in, It is a multi-granularity feature sequence; This represents the feature representation after cross-scale fusion. This is a multi-head attention mechanism; This is a query, key, and value matrix obtained by linear transformation of granular features; The dimension of the key vector; This is a daily-scale alignment feature; This is a periodic-scale alignment feature; Alignment features at the monthly scale; It is a normalized exponential function.

[0099] This step is the core of cross-scale modeling. Its role is to enable the feature information of daily, weekly and monthly scales to complement and enhance each other (for example, to use long-term trend information of monthly scale to correct short-term fluctuations of daily scale), thereby significantly improving the overall forecasting ability of the model.

[0100] As an optional implementation, in step S6, the computer system fuses the features obtained from the cross-scale fusion in step S5. The input is an artificial neural network (such as a feedforward network FFN), which performs nonlinear transformations and deep modeling on the features at each scale. The expression for the scale-based modeling optimization is shown below.

[0101] .

[0102] in, Features optimized for in-scale modeling; This represents the feature representation after cross-scale fusion. This is the first learnable weight in FFN (obtained through model training); This is the second learnable weight in FFN (obtained through model training); This is the first learnable bias in FFN (obtained through model training); This is the second learnable bias in FFN (obtained through model training); The Gaussian Error Linear Unit (Gaussian Error Linear Unit) is an activation function with smooth gradient properties, which can better capture nonlinear relationships in features and improve the model's ability to model deep features.

[0103] This step further optimizes and refines the feature representation of each time scale based on the global information fused across scales, thereby enhancing the expressive power of the model.

[0104] As an optional implementation, in step S7, the computer system processes the features after step S6. The feature sequences are restored to their original time scales through linear mapping, and a spatial decoding module is used to reconstruct the specific sea ice concentration values ​​from these feature sequences. The expression for the linear mapping restoration is shown below.

[0105] .

[0106] The expression for spatial decoding is shown below.

[0107] .

[0108] in, This is the restored original scale feature sequence; Features optimized for in-scale modeling; This is the first learnable parameter of the reduction layer; This is the second learnable parameter of the reduction layer; It is a daytime scale; It is a periodic scale; Measured on a lunar scale; For scale Sea ice concentration forecast results; This is a spatial decoding function.

[0109] The computer system outputs the final decoding result of step S7, which simultaneously generates a future daily-scale sea ice concentration forecast. Weekly sea ice concentration forecast Lunar-scale sea ice concentration forecast .

[0110] This application addresses the single-scale modeling and information silo problems in existing technologies by introducing a cross-scale attention fusion mechanism and a unified multi-scale processing framework. Specifically, traditional methods process each time scale independently, while this application establishes an explicit cross-scale information interaction channel; existing schemes lack guidance from long-term scales for short-term forecasts, while this application achieves multi-scale collaborative optimization; compared to single-output models, this application simultaneously outputs multi-scale forecast results, meeting the needs of different application scenarios.

[0111] These technological breakthroughs give this application significant advantages in practical applications such as polar environment monitoring and Arctic shipping planning, and provide reliable technical support for sea ice forecasting to move from research to operational application.

[0112] Based on the same inventive concept, this application also provides a polar sea ice multi-scale forecasting system for implementing the aforementioned polar sea ice multi-scale forecasting method. The solution provided by this system is similar to the implementation described in the above method; therefore, the specific limitations of one or more polar sea ice multi-scale forecasting system embodiments provided below can be found in the limitations of the polar sea ice multi-scale forecasting method described above, and will not be repeated here.

[0113] In one exemplary embodiment, such as Figure 3 As shown, a multi-scale forecasting system for polar sea ice is provided, which includes the following modules.

[0114] The data acquisition module is used to acquire raw sea ice concentration observation data.

[0115] The data preprocessing module is used to preprocess the original sea ice concentration observation data to obtain preprocessed data; the preprocessed data includes: daily sea ice concentration data, weekly sea ice concentration data and monthly sea ice concentration data.

[0116] The spatial feature encoding module is used to extract multi-scale spatial features from the preprocessed data to obtain feature tokens at multiple scales.

[0117] The feature dimension alignment module is used to align the feature tokens of the multiple scales to obtain aligned features; the aligned features include: daily scale aligned features, weekly scale aligned features and monthly scale aligned features.

[0118] The cross-scale attention fusion module is used to perform cross-scale attention fusion on the aligned features to obtain cross-scale fused features.

[0119] The intra-scale feature modeling module is used to perform intra-scale modeling optimization on the cross-scale fused features to obtain intra-scale modeling optimized features.

[0120] The feature restoration and decoding module is used to perform linear mapping restoration and spatial decoding on the optimized features within the scale to obtain sea ice concentration forecast results. The sea ice concentration forecast results include: future daily-scale sea ice concentration forecasts, future weekly-scale sea ice concentration forecasts, and future monthly-scale sea ice concentration forecasts.

[0121] The polar sea ice multi-scale forecasting system provided in this application has the following data processing flow and core module structure as follows: Figure 4 As shown, this system achieves end-to-end automated processing from raw data input to multi-scale forecast results output through a series of interconnected dedicated modules. Each module performs its specific function and works collaboratively, as described in the following process.

[0122] First, the data input interface and the data preprocessing module constitute the data preparation layer, which is responsible for receiving raw observation data and completing cleaning, normalization and multi-scale calculations, providing high-quality daily, weekly and monthly scale standardized data for subsequent modules, and temporarily storing them in the corresponding data cache.

[0123] Subsequently, the spatial feature encoding module extracts key spatial distribution features from data at various scales; the feature dimension alignment module projects these features to a unified dimension; the cross-scale attention fusion module is responsible for mining and integrating deep dependencies between features; and the intra-scale feature modeling module further optimizes the feature representation at each scale.

[0124] Finally, the result generation layer, consisting of the feature restoration and decoding module and the forecast output interface, is responsible for restoring the fused and optimized high-dimensional features into specific sea ice concentration values ​​and simultaneously outputting the final forecast results for the daily, weekly, and monthly time scales.

[0125] In summary, this application has the following advantages over the prior art.

[0126] (1) The accuracy of multi-scale forecasts has been significantly improved.

[0127] Advantages: The accuracy of sea ice concentration forecasts at daily, weekly, and monthly scales is improved simultaneously.

[0128] Technology sources: The cross-scale attention fusion module enables deep interaction of features at different time scales; the feature dimension alignment module ensures effective fusion of multi-scale information; and the unified spatial feature encoder maintains the consistency of feature extraction.

[0129] Cause: By explicitly modeling the characteristics of sea ice changes on daily, weekly, and monthly scales and establishing cross-scale dependencies, the long-term trend information is fully utilized to correct short-term fluctuations, significantly improving the forecast accuracy at all scales.

[0130] (2) The forecast lead time is extended to the seasonal scale.

[0131] Advantages: Enables continuous 180-day daily forecasts from the sub-seasonal to the seasonal scale.

[0132] Technology sources: The multi-scale spatial feature coding module effectively extracts long-term variation patterns; the intra-scale feature modeling module enhances the expressive power of features at each scale; and the feature restoration and decoding module ensures the stability of long-term forecasts.

[0133] Cause: By combining deep latent spatial representation with temporal features, the error accumulation problem in long-term forecasts of existing models is overcome.

[0134] (3) Optimize computational efficiency.

[0135] Advantages: Reduces computational resource requirements while maintaining accuracy. Technology sources: A unified spatial feature encoder avoids redundant calculations; feature dimension alignment reduces the complexity of subsequent processing; An end-to-end joint training architecture improves model convergence speed.

[0136] Cause: Shared encoders and unified processing flow eliminate the overhead of parallel computing of multiple models and optimize the utilization of computing resources.

[0137] 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.

[0138] 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 multi-scale forecasting method for polar sea ice, characterized in that, The multi-scale forecasting method for polar sea ice includes: Obtain raw sea ice concentration observation data; The original sea ice concentration observation data is preprocessed to obtain preprocessed data; the preprocessed data includes: daily sea ice concentration data, weekly sea ice concentration data and monthly sea ice concentration data; Multi-scale spatial feature extraction is performed on the preprocessed data to obtain feature tokens at multiple scales; The feature tokens at the multiple scales are aligned in terms of feature dimensions to obtain aligned features; the aligned features include: daily scale aligned features, weekly scale aligned features, and monthly scale aligned features. The aligned features are then subjected to cross-scale attention fusion to obtain cross-scale fused features. The cross-scale fused features are then subjected to intra-scale modeling optimization to obtain intra-scale modeling optimized features. Linear mapping and spatial decoding are performed on the optimized features within the specified scale to obtain sea ice concentration forecast results. The sea ice concentration forecast results include: future daily-scale sea ice concentration forecasts, future weekly-scale sea ice concentration forecasts, and future monthly-scale sea ice concentration forecasts.

2. The multi-scale forecasting method for polar sea ice according to claim 1, characterized in that, Multi-scale spatial feature extraction is performed on the preprocessed data to obtain feature tokens at multiple scales, specifically including: The preprocessed data is divided into image blocks for pre-initial embedding to obtain initial embedding features; Based on the initial embedded features, multi-scale feature extraction is performed through a spatial feature encoder to obtain the final 2D spatial features; the spatial feature encoder includes several pairs of continuous blocks; The final 2D spatial features are linearly projected and transformed to obtain feature tokens of multiple scales.

3. The multi-scale forecasting method for polar sea ice according to claim 2, characterized in that, The formula for calculating each pair of consecutive blocks is: ； in, For the first Intermediate features of a layer after layer normalization and attention calculation; For layer normalization; For multi-head self-attention of windows; For the first Features of the layer output; For the first Features of the layer output; It is a multilayer perceptron; For the first Intermediate features of a layer after layer normalization and attention calculation; For multi-head self-attention in shifted windows; For the first Features of the layer output.

4. The multi-scale forecasting method for polar sea ice according to claim 2, characterized in that, The expression for the linear projection transformation is: ； in, Feature tokens for multiple scales; For flattening operation; This is a linear transformation operation; The first Features of the layer output; This represents the total number of layers in the encoder.

5. The multi-scale forecasting method for polar sea ice according to claim 1, characterized in that, The expression for feature dimension alignment is: ； in, Features after alignment; This is a linear transformation operation; For feature tokens of multiple scales.

6. The multi-scale forecasting method for polar sea ice according to claim 1, characterized in that, The expression for the cross-scale attention fusion is: ； ； ； in, It is a multi-granularity feature sequence; This represents the feature representation after cross-scale fusion. This is a multi-head attention mechanism; This is a query, key, and value matrix obtained by linear transformation of granular features; The dimension of the key vector; This is a daily-scale alignment feature; This is a periodic-scale alignment feature; Alignment features at the monthly scale; It is a normalized exponential function.

7. The multi-scale forecasting method for polar sea ice according to claim 1, characterized in that, The expression for the intra-scale modeling optimization is: ； in, Features optimized for in-scale modeling; This represents the feature representation after cross-scale fusion. This is the first learnable weight in FFN; This is the second learnable weight in FFN; This is the first learnable bias in FFN; This is the second learnable bias in FFN; This is the activation function.

8. The multi-scale forecasting method for polar sea ice according to claim 1, characterized in that, The expression for the linear mapping reduction is: ； in, This is the restored original scale feature sequence; Features optimized for in-scale modeling; This is the first learnable parameter of the reduction layer; This is the second learnable parameter of the reduction layer; It is a daytime scale; It is a periodic scale; It is measured on a lunar scale.

9. The multi-scale forecasting method for polar sea ice according to claim 1, characterized in that, The expression for spatial decoding is: ； in, For scale Sea ice concentration forecast results; This is the restored original scale feature sequence; It is a daytime scale; It is a periodic scale; Measured on a lunar scale; This is a spatial decoding function.

10. A multi-scale forecasting system for polar sea ice, characterized in that, The polar sea ice multi-scale forecasting system is used to implement the polar sea ice multi-scale forecasting method according to any one of claims 1-9, and the polar sea ice multi-scale forecasting system includes: The data acquisition module is used to acquire raw sea ice concentration observation data; The data preprocessing module is used to preprocess the original sea ice concentration observation data to obtain preprocessed data; the preprocessed data includes: daily sea ice concentration data, weekly sea ice concentration data and monthly sea ice concentration data; The spatial feature encoding module is used to extract multi-scale spatial features from the preprocessed data to obtain feature tokens at multiple scales. The feature dimension alignment module is used to align the feature tokens of the multiple scales to obtain aligned features; the aligned features include: daily scale aligned features, weekly scale aligned features and monthly scale aligned features; The cross-scale attention fusion module is used to perform cross-scale attention fusion on the aligned features to obtain cross-scale fused features; The intra-scale feature modeling module is used to perform intra-scale modeling optimization on the cross-scale fused features to obtain intra-scale modeling optimized features. The feature restoration and decoding module is used to perform linear mapping restoration and spatial decoding on the optimized features within the scale to obtain sea ice concentration forecast results. The sea ice concentration forecast results include: future daily-scale sea ice concentration forecasts, future weekly-scale sea ice concentration forecasts, and future monthly-scale sea ice concentration forecasts.

Images