Antarctic amplification effect prediction method and system based on multi-scale space-time fusion and physical constraint

Abstract

The invention relates to the technical field of earth climate systems, in particular to an Antarctic amplification effect prediction method and system based on multi-scale space-time fusion and physical constraint, and the method comprises the steps: collecting data: collecting multi-source data related to an Antarctic ice cover in real time, and obtaining dynamic information of the ice cover; the multi-source data comprises satellite remote sensing data, ground observation data and meteorological reanalysis data; data preprocessing: cleaning the collected multi-source data, and removing noise, abnormal values and missing values; through a multi-scale space-time fusion model, physical constraint deep learning and a dynamic training strategy, the method comprehensively exceeds a traditional physical model and a single detection technology in the aspects of prediction accuracy, dynamic adaptability, calculation efficiency, application range and the like; the method fills the blank of dynamic prediction and multi-source fusion of the ice cover amplification effect in the prior art, provides an efficient and reliable technical tool for coping with global climate changes, and has important scientific significance and practical value.

Description

Technical Field

[0001] The present invention relates to the technical field of the Earth's climate system, and in particular to a method and system for predicting the Antarctic amplification effect based on multi-scale spatio-temporal fusion and physical constraints. Background Art

[0002] As a core component of the Earth's climate system, the Antarctic ice sheet's melting not only directly causes sea-level rise but also triggers the "Antarctic amplification effect" through mechanisms such as ocean heat flux, circulation changes, and atmospheric feedback, significantly exacerbating global climate change. The Antarctic amplification effect refers to the high sensitivity of the Antarctic ice sheet's response to global climate change, with its melting rate and feedback effect far exceeding other regions, significantly exacerbating sea-level rise and climate system changes. Traditional prediction methods mainly rely on physical models (such as ice sheet dynamics models and climate simulations), based on image or video recognition or understanding, integrating data from various sources, obtaining data and records, and estimating the ice sheet melting rate and its contribution to sea level through numerical simulations. However, these methods have significant limitations: 1. Strong data dependence: Physical models require high-precision observational data (such as ice thickness, temperature, snowfall), while data in the Antarctic region is scarce, resulting in relatively high prediction uncertainty (MSE is often higher than 0.08); 2. High computational complexity: Involves solving complex partial differential equations, with a single simulation taking several hours and being difficult to update in real time; 3. Insufficient non-linear response: The non-linear characteristics of the ice sheet amplification effect (such as accelerated melting under extreme climates) are difficult to capture, and the prediction accuracy is limited (R² is usually lower than 0.85).

[0003] In recent years, machine learning techniques have been applied in environmental science. For example: Chinese invention patent CN108733909A, "A method for detecting the melting area of an ice sheet", uses a fully convolutional network (FCNN) to detect the melting area based on the brightness temperature data of a microwave radiometer, with an accuracy of 93%, but is limited to static distribution analysis, relies on a single data source, and cannot predict future trends or amplification effects; Chinese invention patent CN114219979A, "A short-term precipitation prediction model based on multi-scale spatio-temporal fusion", realizes short-term precipitation prediction (from minutes to hours) through U-Net and ConvLSTM. Although it involves multi-scale spatio-temporal fusion, it lacks physical constraints, and the prediction range is limited to precipitation; Chinese invention patent CN119206537A, "A physically-guided deep learning satellite remote sensing cloud detection method", combines the SAM model and radiation transfer constraints to detect cloud cover, with high static analysis accuracy (IoU is about 0.88), but has no dynamic prediction function and only serves cloud monitoring.

[0004] The existing technologies generally lack a systematic method for integrating multi-source data and achieving high-precision and dynamic prediction of the Antarctic amplification effect. Summary of the Invention

[0005] The present invention provides a method and system for predicting the Antarctic amplification effect based on multi-scale spatio-temporal fusion and physical constraints, which overcomes the deficiencies of the above-mentioned existing technologies and can effectively solve the problems that the existing technologies lack the integration of multi-source data and cannot achieve high-precision and dynamic prediction of the Antarctic amplification effect.

[0006] To solve the above problems, one of the technical solutions of the present invention is achieved through the following means: A method for predicting the Antarctic amplification effect based on multi-scale spatio-temporal fusion and physical constraints, comprising the following steps: Data collection: Real-time collect multi-source data related to the Antarctic ice sheet and obtain ice sheet dynamic information; wherein, the multi-source data includes satellite remote sensing data, ground observation data, and meteorological reanalysis data; Data preprocessing: Clean the collected multi-source data to remove noise, outliers, and missing values; extract multi-scale spatio-temporal features related to the Antarctic amplification effect according to feature engineering; construct a dynamic graph based on the ice sheet spatial grid, and the structure represents the heat transfer and melting diffusion relationship between regions; standardize the data; wherein, the multi-scale spatio-temporal features related to the Antarctic amplification effect include the ice sheet thickness change rate, ocean heat flux, and atmospheric circulation index; Model construction and training: Construct a multi-scale spatio-temporal fusion model, adopt a multi-task collaborative training method, simultaneously optimize the prediction of ice sheet melting volume, sea level rise contribution rate, and climate feedback intensity through a weighted loss function, combine a generative adversarial network to generate extreme climate scenario data to enhance the training set, and dynamically update the model parameters on distributed nodes through online federated learning; Prediction: Input the real-time collected multi-source data into the trained multi-scale spatio-temporal fusion model, and output the prediction result of the Antarctic amplification effect; wherein, the prediction result includes the ice sheet melting volume, the change in sea level height, and the quantitative value of the climate feedback effect; Result analysis and verification: Conduct time series analysis and spatial distribution analysis on the prediction result of the Antarctic amplification effect, combine sensitivity analysis and uncertainty assessment, verify the model accuracy, and generate a visualization report.

[0007] The above data collection also includes quality verification of the multi-source data to ensure that the missing rate is less than 2%, and unify the data to daily resolution through spatio-temporal alignment to ensure the integrity, consistency, and relevance of the Antarctic amplification effect.

[0008] The feature engineering in the above data preprocessing further includes: Generate time series features based on a sliding window, extract multi-scale spatio-temporal features, and the window length is not less than 30 days; Dynamically adjust the feature weights through reinforcement learning, adaptively screen the most important feature combinations for the prediction of the amplification effect, and the weight adjustment frequency is not less than once a day.

[0009] In the above model construction and training, a multi-scale spatio-temporal fusion model is constructed, and the model includes: A multi-modal Transformer module that uses a multi-head self-attention mechanism to fuse the temporal features of multi-source data, with no less than 4 layers and no less than 128 hidden dimensions per layer; A graph neural network module that captures the spatial dependence of the ice sheet based on a dynamic graph structure, with a grid resolution of not less than 100km×100km; A physical constraint layer that optimizes the prediction results by embedding the mass conservation equation Δ𝑀=𝑀 𝑖𝑛 −𝑀 𝑜𝑢𝑡 and the heat flux balance equation 𝑄=𝜌𝑐Δ𝑇.

[0010] In the above model construction and training, multi-task collaborative training includes achieving multi-objective optimization through a weighted loss function, where the loss function is defined as Loss=𝑤1⋅MSE 融化量 +𝑤2⋅MSE 海平面 +𝑤3⋅MSE 反馈 , with weights 𝑤1, 𝑤2, 𝑤3; or / and, in the above model construction and training, online federated learning includes collaboratively updating the model parameters through at least 5 distributed nodes, using the FedAvg algorithm, fusing global real-time observation data every 24 hours, and maintaining the dynamic adaptability of the model to climate change.

[0011] The above prediction also includes outputting the confidence interval of the prediction result and predicting the long-term trend of the Antarctic ice sheet method effect through time series analysis.

[0012] The above result analysis and verification also include calculating the model performance metrics using an independent test set and comparing them with the prediction results of traditional physical models; or / and, it also includes outputting the prediction results and analysis reports in the form of charts, heatmaps, or 3D visualizations, with a heatmap resolution of not less than 100km×100km.

[0013] The second technical solution of the present invention is achieved through the following: A prediction system for the Antarctic amplification effect based on multi-scale spatio-temporal fusion and physical constraints, including, A data collection unit: Collect multi-source data related to the Antarctic ice sheet in real time and obtain the dynamic information of the ice sheet; among them, the multi-source data includes satellite remote sensing data, ground observation data, and meteorological reanalysis data; Data preprocessing unit: Clean the multi-source data collected, removing noise, outliers, and missing values; Extract multi-scale spatio-temporal features related to the Antarctic amplification effect according to feature engineering; Construct a dynamic graph based on the ice sheet spatial grid, with the structure representing the relationship of heat transfer and melting diffusion between regions; Standardize the data; Among them, the multi-scale spatio-temporal features related to the Antarctic amplification effect include the ice sheet thickness change rate, ocean heat flux, and atmospheric circulation index; Model construction and training unit: Construct a multi-scale spatio-temporal fusion model, adopt a multi-task collaborative training method, simultaneously optimize the prediction of ice sheet melting volume, sea level rise contribution rate, and climate feedback intensity through a weighted loss function, combine a generative adversarial network to generate extreme climate scenario data to enhance the training set, and dynamically update the model parameters on distributed nodes through online federated learning; Prediction unit: Input the multi-source data collected in real time into the trained multi-scale spatio-temporal fusion model, and output the prediction results of the Antarctic amplification effect; Among them, the prediction results include the ice sheet melting volume, the change in sea level height, and the quantitative values of the impact of climate feedback; Result analysis and verification unit: Conduct time series analysis and spatial distribution analysis on the prediction results of the Antarctic amplification effect, combine sensitivity analysis and uncertainty assessment, verify the model accuracy, and generate a visualization report.

[0014] The third technical solution of the present invention is implemented in the following way: A computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, it realizes the dynamic prediction method of the Antarctic amplification effect based on multi-scale spatio-temporal fusion and physical constraints.

[0015] The fourth technical solution of the present invention is implemented in the following way: An electronic device, including a processor and a memory, wherein a computer program is stored in the memory, and the computer program is loaded and executed by the processor to realize the dynamic prediction method of the Antarctic amplification effect based on multi-scale spatio-temporal fusion and physical constraints.

[0016] The present invention comprehensively surpasses traditional physical models and single detection technologies in terms of prediction accuracy (MSE < 0.03, R² > 0.95), dynamic adaptability (robustness improved by 30%), computational efficiency (speeded up by 10 times), and application scope through a multi-scale spatio-temporal fusion model (Transformer + GNN), physical constraint deep learning, and dynamic training strategies (GAN + federated learning); This method fills the gaps in the dynamic prediction of the ice sheet amplification effect and multi-source fusion in the existing technology, provides an efficient and reliable technical tool for coping with global climate change, and has important scientific significance and practical value. Description of the Drawings

[0017] The following further details the specific implementation manners of the present invention in conjunction with the drawings.

[0018] Figure 1 This is the overall schematic diagram of the method in Embodiment 1 of the present invention.

[0019] Figure 2 This is the flowchart of the method in Embodiment 1 of the present invention.

[0020] Figure 3 This is the system block diagram of Embodiment 2 of the present invention. Detailed implementation manners

[0021] The present invention is not limited by the following embodiments, and the specific implementation manners can be determined according to the technical solutions of the present invention and the actual situation.

[0022] Embodiment 1: As Figure 1 , 2 shown, Embodiment 1 of the present invention discloses a prediction method for the Antarctic amplification effect based on multi-scale spatio-temporal fusion and physical constraints, including the following steps: S101, Data collection: Collect multi-source data related to the Antarctic ice sheet in real time and obtain ice sheet dynamic information; among them, the multi-source data includes satellite remote sensing data, ground observation data, and meteorological reanalysis data; S102, Data preprocessing: Clean the collected multi-source data, remove noise, outliers, and missing values; extract multi-scale spatio-temporal features related to the Antarctic amplification effect according to feature engineering; construct a dynamic graph based on the ice sheet spatial grid, and the structure represents the heat transfer and melting diffusion relationship between regions; standardize the data; among them, the multi-scale spatio-temporal features related to the Antarctic amplification effect include the ice sheet thickness change rate, ocean heat flux, and atmospheric circulation index; S103, Model construction and training: Construct a multi-scale spatio-temporal fusion model, adopt a multi-task collaborative training method, simultaneously optimize the prediction of ice sheet melting volume, sea level rise contribution rate, and climate feedback intensity through a weighted loss function, combine a generative adversarial network (GAN) to generate extreme climate scenario data to enhance the training set, and dynamically update the model parameters on distributed nodes through online federated learning; S104, Prediction: Input the multi-source data collected in real time into the trained multi-scale spatio-temporal fusion model, and output the prediction results of the Antarctic amplification effect; among them, the prediction results include the quantification values of ice sheet melting volume (unit: Gt / year, ±30 Gt), sea level height change (unit: mm / year, ±0.2 mm), and climate feedback impact (unit: W / m², ±0.1 W / m²); S105, Result analysis and verification: Conduct time series analysis and spatial distribution analysis on the prediction results of the Antarctic amplification effect, combine sensitivity analysis and uncertainty assessment, verify the model accuracy, and generate a visualization report.

[0023] In the above data collection, the multi-source data collected includes: Satellite remote sensing data: GRACE gravity data (monthly resolution, 2002 - 2023, unit: Gt) to measure the change in ice sheet mass; ICESat altimeter data (spatial resolution 100m, unit: m) to obtain the change in ice thickness; Ground observation data: Temperature (daily average, unit: °C), wind speed (hourly average, unit: m / s), and snowfall (monthly cumulative, unit: mm) recorded at Antarctic research stations (such as McMurdo Station, Amundsen - Scott Station); Meteorological reanalysis data: Atmospheric circulation index (monthly average, dimensionless), ocean heat flux (daily average, unit: W / m²), and precipitation (daily cumulative, unit: mm) provided by ERA5.

[0024] The above data collection also includes quality verification of multi-source data to ensure that the missing rate is less than 2%, and the data is unified to daily resolution through spatio-temporal alignment to ensure the integrity, consistency, and relevance of the Antarctic amplification effect. Among them, data is downloaded through NASA Earthdata and ECMWF CDS API, stored in NetCDF format, and spatio-temporally aligned to daily resolution. The data preprocessing script (Python, xarray library) unifies multi-source data to daily resolution, verifies integrity (missing rate < 2%, outliers < 1%), and ensures data integrity.

[0025] Data cleaning in the above data preprocessing includes removing noise, outliers (Z-score threshold ±3σ), and missing values (linear interpolation or MICE algorithm). Among them, the Z-score method (threshold ±3σ) is used for outlier processing to identify outliers (such as sudden changes in ICESat ice thickness > 500m), and abnormal records are deleted (about 0.5%). For missing value filling, when the missing rate < 5%, linear interpolation (numpy.interp) is used; when the missing rate > 5%, multiple imputation (MICE algorithm, Python fancyimpute library) is used to ensure the continuity of the time series.

[0026] Feature engineering in the above data preprocessing also includes: Generating time series features based on a sliding window, extracting multi-scale spatio-temporal features, with the window length not less than 30 days (step size 1 day); among them, multi-scale spatio-temporal features include the change rate of ice sheet thickness (Δℎ / Δ𝑡, unit: m / year), ocean heat flux (unit: W / m²), and atmospheric circulation index (dimensionless); Dynamically adjust the feature weights through reinforcement learning (Q - learning algorithm, learning rate 0.01, discount factor 0.9) (such as ocean heat flux 0.45 ± 0.05, ice thickness 0.35 ± 0.05), adaptively screen the most important feature combinations for the amplification effect prediction, and the weight adjustment frequency is not less than once a day.

[0027] In the above data preprocessing, constructing the dynamic graph includes constructing the physical associations of heat transfer and melting diffusion between regions based on a 100km×100km grid (about 4000 nodes, 12000 edges) (the edge weights are based on the heat transfer coefficient 𝑘 = 0.5 𝑊 / 𝑚 2 / 𝐾), and generating the graph structure (PyTorch Geometric library).

[0028] In the above data preprocessing, data standardization processing includes using Min - Max normalization to the [0, 1] interval (formula: 𝑋′=(𝑋−𝑋 𝑚𝑖𝑛 ) / (𝑋 𝑚𝑎𝑥 −𝑋 𝑚𝑖𝑛 ), where X´ is the dimensionless standardized value, X is the original observed value, X min and 𝑋 𝑚𝑎𝑥 define the data range, and X max −X min is the scaling factor; this processing process ensures the compatibility of multi - source data in this patent and lays a foundation for subsequent spatio - temporal fusion and physical constraint modeling).

[0029] In the above model construction and training, the data flow is: multi - source feature input → temporal fusion → spatial modeling → physical constraint output; in model construction and training, construct a multi - scale spatio - temporal fusion model, where the model includes: Multi - modal Transformer module, which uses the multi - head self - attention mechanism to fuse the temporal features of multi - source data, with no less than 4 layers, and the hidden dimension of each layer is no less than 128; it can fuse the temporal features of multi - source data through 4 - layer multi - head self - attention (4 heads, hidden dimension 128, Dropout rate 0.1) to capture long - term dependencies over 10 years (long - term trends from 2000 - 2023) (PyTorch nn.Transformer); Graph Neural Network (GNN) module, which captures the spatial dependencies of the ice sheet based on the dynamic graph structure, and the grid resolution is not less than 100km×100km; it can model the spatial dependencies and melting diffusion effects of the ice sheet based on the dynamic graph through 2 - layer GraphSAGE convolution (64 neurons per layer, activation function ReLU) (PyTorch Geometric GraphSAGE); Physical constraint layer, by embedding the mass conservation equation Δ𝑀 = 𝑀 𝑖𝑛−M 𝑜𝑢𝑡 (M 𝑖𝑛 / 𝑜𝑢𝑡 is the ice mass input / output, unit: Gt) and the heat flux balance equation \(Q = \rho c\Delta T\) to optimize the prediction results. Here, \(Q\) is the heat flux, representing the energy transfer rate, unit \(W / m^2\) 2 (watts per square meter, SI unit); \(\rho\) is the density, reflecting the mass distribution of the substance, unit: \(kg / m^3\) 3 (kilograms per cubic meter), \(\rho = 917 kg / m^3\) 3 ; \(c\) is the specific heat capacity, measuring the heat absorption capacity, unit: \(J / kg·K\) (joules per kilogram per kelvin), \(c = 2090 J / kg / K\); \(\Delta T\) is the temperature difference, driving the heat change, unit: \(K\) (kelvin). The prediction is optimized through a hybrid loss function (\(Loss = 0.7\cdot MSE+0.3\cdot Physics Loss\)), where \(Physics Loss\) is the sum of the squares of the physical deviations.

[0030] Output layer: fully connected layer (\(256\rightarrow3\) neurons), predicting the ice sheet melting amount (\(Gt / year\)), the contribution rate of sea level rise (\(mm / year\)), and the climate feedback intensity (\(W / m²\)).

[0031] When the above model is implemented, the model is implemented using PyTorch 2.0, with a total of about 5.2 million parameters, running on an NVIDIA A100 GPU (\(40GB\) video memory, CUDA 11.8), and the single forward inference takes about 0.5 seconds.

[0032] In the construction and training of the above model, multi-task collaborative training includes multi-objective optimization through a weighted loss function. Here, the loss function is defined as \(Loss = w_1\cdot MSE\) 融化量 \(+ w_2\cdot MSE\) 海平面 \(+ w_3\cdot MSE\) 反馈 , where \(Loss\) is the overall error; \(w_1\), \(w_2\), \(w_3\) are task weights to balance the importance; \(MSE\) 融化量 、\(MSE\) 海平面 、\(MSE\) 反馈 are the errors of each task, quantifying the prediction accuracy of the melting amount, sea level, and feedback respectively.

[0033] Dynamically adjusted according to the task importance, with initial values of 𝑤1 = 0.4, 𝑤2 = 0.3, and 𝑤3 = 0.3 respectively, and dynamically adjusted according to the validation set error every 10 rounds (±0.05); for multi-objective optimization, the optimizer is Adam (learning rate 0.001, 𝛽1 = 0.9, 𝛽2 = 0.999, halved every 50 rounds); the GAN includes a generator (at least 3 fully connected networks, 64→128→64 dimensions, activation function ReLU) and a discriminator (at least 2 convolutional networks, 64→32→1 dimension, activation function Sigmoid). The generator outputs simulated extreme climate data (such as a 5°C increase in temperature and a +50% increase in snowfall). The training set is increased from 100,000 to 150,000 (PyTorch torch.nn) to enhance the model's prediction ability under future uncertain conditions, that is, to enhance the model's robustness.

[0034] In the model construction and training, online federated learning includes at least 5 distributed nodes (5 Antarctic research stations (such as McMurdo, equipped with 8GB GPUs)) and 3 cloud servers (AWS EC2 g5.xlarge)) collaborating to update the model parameters daily (communication bandwidth ≥ 10Mbps). The FedAvg algorithm is used to fuse global real-time observation data (about 500MB / day) every 24 hours to maintain the model's dynamic adaptability to climate change.

[0035] In the specific implementation of the above model construction and training, training data (from 2000 to 2020, 80% training set), validation set (from 2021 to 2022, 20%), batch size 64, trained for 100 rounds, and the validation MSE < 0.03.

[0036] The above prediction also includes outputting the confidence interval of the prediction result (confidence level not less than 95%) and predicting the long-term trend of the Antarctic ice sheet method effect through time series analysis (such as changes in the next 1 - 10 years). For example: input real-time data in 2023 (updated daily, about 1GB), and predict the results from 2024 to 2033: ice sheet melting volume: 500 ± 30 Gt / year; sea level rise: 1.5 ± 0.2 mm / year; climate feedback: 0.8 ± 0.1 W / m² (95% confidence interval). The prediction results are stored in the cloud (AWS S3) in a CSV file (fields: timestamp, predicted value, lower limit of confidence interval, upper limit).

[0037] In the above result analysis and verification, the analysis and evaluation include sensitivity analysis (identifying key factors such as ocean heat flux) and uncertainty analysis (confidence interval).

[0038] The above result analysis and verification also include calculating the model performance metrics using an independent test set (data from 2021 - 2022), such as the mean squared error (MSE) not being higher than 0.03 and the coefficient of determination (R²) not being lower than 0.95, and comparing with the prediction results of traditional physical models (MSE≈0.08) to verify the superiority of this method.

[0039] The above also includes outputting the prediction results and analysis reports in the form of charts, heatmaps, or 3D visualizations. The resolution of the heatmap is not lower than 100km×100km for reference in climate research, policy - making, and for the public. Among them, the charts include time - series charts (10 - year Matplotlib line charts, X - axis: year, Y - axis: melt volume / sea level / feedback), spatial heatmaps (Seaborn library, 100km resolution, color scale: blue - low to red - high), and 3D change models. Verify the above results: MSE = 0.028, R² = 0.96, and compare with traditional models (MSE = 0.08, R² = 0.85).

[0040] When performing the above visualizations, for the time - series chart: 10 - year trend, with confidence intervals marked; for the heatmap: Antarctic map, grid coverage > 95%; for the error chart: predicted values vs. observed values at Antarctic scientific research stations (line comparison, error ±5%).

[0041] The effects of the above embodiments are as follows: Taking the Antarctic Peninsula as an example, inputting data from 2000 - 2020 (about 50GB), predicting the trend from 2021 - 2030: the annual melt volume increases by 5% (about 25 Gt / year), the sea - level contribution rate is 1.8±0.2 mm / year, and the climate feedback intensity is 0.9±0.1 W / m². The verification accuracy R² = 0.96, which is better than traditional physical models (R² = 0.85), static classification of the prior art (accuracy 93%), short - term prediction of the prior art (MSE = 0.05), and cloud detection of the prior art (IoU = 0.88).

[0042] Based on the above, integrating the prediction results and analysis to generate a report, outlining trends (such as a 5% annual increase in melt volume), impacts (such as a 1.8 - mm / year sea - level rise), and emission reduction suggestions for reference in scientific research, decision - making, and for the public.

[0043] In summary, the embodiments of the present invention have the following effects: 1. Significantly improve the prediction accuracy Capturing complex spatio-temporal dynamic relationships: The present invention constructs a multi-scale spatio-temporal fusion model through a multi-modal Transformer and a graph neural network (GNN), which efficiently integrates multi-source heterogeneous data (such as GRACE gravity data, ICESat altimeter data, ERA5 reanalysis data), and accurately captures the non-linear interactions and dynamic evolution laws among the ice sheet thickness change rate, ocean heat flux, and atmospheric circulation index. Compared with the simplified assumptions and single data source methods of traditional physical models, as well as static cloud detection based on satellite remote sensing, the prediction accuracy of the present invention is significantly improved, the mean square error (MSE) is reduced to below 0.03, the coefficient of determination (R²) exceeds 0.95, and it is better than the existing cloud detection IoU (Intersection over Union) index.

[0044] Physical constraint optimization for prediction: By embedding the mass conservation equation (Δ𝑀 = 𝑀 𝑖𝑛 −𝑀 𝑜𝑢𝑡 ), and the heat flux balance equation (𝑄 = 𝜌𝑐Δ𝑇), and combining a hybrid loss function (Loss = 0.7⋅MSE + 0.3⋅Physics Loss) to optimize the prediction results, ensuring that the output conforms to the physical laws of the ice sheet. Different from the radiation transfer physical constraint (Bayesian probability calculation of cloud probability) which only serves static detection, the present invention applies physical constraints to multi-objective dynamic prediction (melting volume, sea level, climate feedback), further improving the accuracy and scientific reliability.

[0045] 2. Enhancing dynamic adaptability and generalization ability Handling non-linearity and extreme scenarios: The Antarctic amplification effect involves highly non-linear processes such as glacier flow, meltwater runoff, and climate feedback. The present invention utilizes the long-term dependence modeling of Transformer (time span ≥ 10 years) and the spatial correlation analysis of GNN (grid resolution 100km × 100km), combined with extreme climate scenario data generated by a generative adversarial network (GAN) (such as a 5℃ increase in temperature and a 50% increase in snowfall), significantly enhancing the model's generalization ability for complex processes and future uncertainty conditions, with the robustness improved by more than 30% compared to traditional methods. Compared with the existing static feature extraction (SAM model) only for cloud detection and single weather scenario adaptation, the present invention has stronger dynamics and multi-scenario adaptability.

[0046] Real-time dynamic update: Through online federated learning, the model fuses global real-time observation data every 24 hours on at least 5 distributed nodes (such as Antarctic research stations, cloud servers), dynamically adjusts parameters, and adapts to climate mutation scenarios. Although existing technologies combine physical prior knowledge to improve cloud detection accuracy, they rely on pre-trained SAM models and have no dynamic update mechanism. The present invention has a significant advantage in real-time performance.

[0047] 3. Greatly improving prediction efficiency Efficient Processing of Multi-source Big Data: With the progress of remote sensing technology, the amount of data related to the Antarctic ice sheet has increased exponentially. Through the efficient parallel computing of Transformer (4-head self-attention, hidden dimension 128) and cloud computing deployment, the present invention can quickly process large-scale multi-source data sets, shortening the prediction response time to the second level, which is more than 10 times faster than traditional numerical simulations. Compared with the existing billion-parameter calculations and integer linear programming optimizations (with relatively high computational complexity) relying on the SAM model, the present invention is superior in terms of multi-source fusion and real-time prediction efficiency.

[0048] Support for Real-time Early Warning: For events such as ice sheet disintegration caused by extreme weather, the present invention can achieve a minute-level prediction response (such as outputting results within 1 - 5 minutes), providing timely support for emergency decision-making. Although the existing cloud detection is accurate, it is limited to static analysis and has no real-time early warning ability, so the present invention has stronger timeliness.

[0049] 4. Breaking through the Limitations of Traditional Physical Models and Single Detection Methods Reducing Dependence on Assumptions: Traditional physical models need to rely on a large number of simplified assumptions (such as uniform melting rate), while the present invention reduces the assumption conditions through the fusion of multi-source data-driven and physical constraints, making the prediction results closer to the actual dynamic changes. Although the existing technology introduces radiative transfer constraints, it only optimizes the cloud detection probability and does not involve dynamic prediction, so the multi-objective dynamic optimization of the present invention is more comprehensive.

[0050] Surpassing Single Detection Technologies: The existing technology only classifies the melting area using brightness temperature data and only detects cloud cover using satellite data, with limited information dimensions. The present invention integrates multi-source data (satellite + ground + reanalysis) and simultaneously predicts the melting volume (Gt / year), sea level rise (mm / year), and climate feedback (W / m²) through multi-task collaborative training. The comprehensiveness of information and the depth of prediction are significantly superior to single detection methods.

[0051] 5. Promoting Cross-disciplinary Technology Integration and Wide Application Technical Innovation Integration: The present invention deeply combines computer science (Transformer, GNN, GAN), geophysics (physical constraints), and distributed computing (federated learning) to form a unique cross-disciplinary technology system. Compared with the single application of only combining SAM and radiative transfer in the existing technology, the present invention has a higher degree of technology integration, promoting the integrated development of climate science and artificial intelligence.

[0052] Expanded application scenarios: This method is not only applicable to the prediction of the Antarctic amplification effect, but can also be extended to the dynamic analysis of the Arctic ice sheet, the assessment of global sea-level rise, and the early warning of extreme climate events. For example, it can predict the inundation risk of coastal cities (accurate to ±0.2 mm / year) or optimize international emission reduction policies (feedback intensity accurate to ±0.1 W / m²). The existing technologies are only limited to cloud detection for weather monitoring, while the application scope of the present invention is wider, with significant socio-economic value and global influence.

[0053] In summary, through the multi-scale spatio-temporal fusion model (Transformer + GNN), physics-constrained deep learning, and dynamic training strategies (GAN + federated learning), the present invention comprehensively surpasses traditional physical models and single detection technologies in terms of prediction accuracy (MSE < 0.03, R² > 0.95), dynamic adaptability (robustness improved by 30%), computational efficiency (speeded up by 10 times), and application scope. This method fills the gaps in the dynamic prediction of the ice sheet amplification effect and multi-source fusion in the existing technologies, providing an efficient and reliable technical tool for coping with global climate change, and having important scientific significance and practical value.

[0054] Example 2: As Figure 3 shown, the embodiment of the present invention discloses a prediction system for the Antarctic amplification effect based on multi-scale spatio-temporal fusion and physical constraints, including Receipt collection unit: Real-time collect multi-source data related to the Antarctic ice sheet and obtain ice sheet dynamic information. Among them, the multi-source data includes satellite remote sensing data, ground observation data, and meteorological reanalysis data. Data preprocessing unit: Clean the collected multi-source data to remove noise, outliers, and missing values; perform multi-scale spatio-temporal feature extraction related to the Antarctic amplification effect according to feature engineering; construct a dynamic graph based on the ice sheet spatial grid, and the structure represents the heat transfer and melting diffusion relationship between regions; standardize the data. Among them, the multi-scale spatio-temporal features related to the Antarctic amplification effect include the ice sheet thickness change rate, ocean heat flux, and atmospheric circulation index. Model construction and training unit: Construct a multi-scale spatio-temporal fusion model, adopt a multi-task collaborative training method, simultaneously optimize the prediction of ice sheet melting volume, sea-level rise contribution rate, and climate feedback intensity through a weighted loss function, combine the generative adversarial network (GAN) to generate extreme climate scenario data to enhance the training set, and dynamically update the model parameters on distributed nodes through online federated learning. Prediction unit: Input the real-time collected multi-source data into the trained multi-scale spatio-temporal fusion model, and output the prediction results of the Antarctic amplification effect. Among them, the prediction results include the quantification values of ice sheet melting volume (unit: Gt / year), sea-level height change (unit: mm / year), and climate feedback impact (unit: W / m²). Result analysis and verification unit: Perform time series analysis and spatial distribution analysis on the prediction results of the Antarctic amplification effect. Combine sensitivity analysis and uncertainty assessment to verify the model accuracy and generate a visualization report.

[0055] Example 3: The embodiment of the present invention discloses a computer-readable storage medium, on which a computer program is stored. When the program is executed by a processor, it implements the dynamic prediction method of the Antarctic amplification effect based on multi-scale spatio-temporal fusion and physical constraints.

[0056] Example 4: The embodiment of the present invention discloses an electronic device, including a processor and a memory. A computer program is stored in the memory, and the computer program is loaded and executed by the processor to implement the dynamic prediction method of the Antarctic amplification effect based on multi-scale spatio-temporal fusion and physical constraints.

[0057] The above-mentioned electronic device further includes a transmission device and an input / output device, wherein both the transmission device and the input / output device are connected to the processor.

[0058] The above-mentioned processor may be a central processing unit CPU, a general-purpose processor, a digital signal processor DSP, an ASIC, an FPGA or other programmable logic devices, transistor logic devices, hardware components or any combination thereof. It can implement or execute various exemplary logical blocks, modules and circuits described in combination with the disclosure content of the present application. It can also be a combination that realizes computing functions, such as a combination including one or more microprocessors, a combination of DSP and microprocessors, etc. The memory may include, but is not limited to: various media that can store computer programs such as USB flash drives, read-only memories, mobile hard disks, magnetic disks or optical discs.

[0059] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memories, CD-ROMs, optical memories, etc.) containing computer-usable program codes. The solutions in the embodiments of the present application can be implemented in various computer languages, for example, object-oriented programming languages such as Java and interpreted scripting languages such as JavaScript.

[0060] This application is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, as well as the combination of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to the processors of general-purpose computers, special-purpose computers, embedded processors, or other programmable data processing devices to generate a machine, such that the instructions executed by the processors of the computer or other programmable data processing devices generate means for implementing the functions specified in one or more of the flows Figure 1 one or more of the flows and / or blocks Figure 1 or means for implementing the functions specified in one or more of the blocks.

[0061] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to operate in a specific manner, such that the instructions stored in the computer-readable memory generate a manufactured article including instruction means for implementing the functions specified in one or more of the flows Figure 1 one or more of the flows and / or blocks Figure 1 or means for implementing the functions specified in one or more of the blocks.

[0062] Embodiments of the present application also provide a computer program product. The computer program product includes a non-transitory computer-readable storage medium storing a computer program. The computer program is operable to cause a computer to execute some or all of the steps of any of the methods described in the above method embodiments. The computer program product can be a software installation package, and the computer includes an electronic device.

Claims

1. A method for predicting the Antarctic amplification effect based on multi-scale spatiotemporal fusion and physical constraints, characterized in that: The following steps are involved: Data collection: Real-time collection of multi-source data related to the Antarctic ice sheet and acquisition of ice sheet dynamic information; multi-source data includes satellite remote sensing data, ground observation data, and meteorological reanalysis data; Data preprocessing: Clean the collected multi-source data to remove noise, outliers and missing values; extract multi-scale spatiotemporal features related to the Antarctic amplification effect based on feature engineering; construct a dynamic graph based on the ice sheet spatial grid, and the structure represents the relationship between heat transfer and melting diffusion between regions; standardize the data; among them, the multi-scale spatiotemporal features related to the Antarctic amplification effect include the ice sheet thickness change rate, ocean heat flux and atmospheric circulation index; Model construction and training: Construct a multi-scale spatiotemporal fusion model, adopt a multi-task collaborative training method, and use a weighted loss function to simultaneously optimize the prediction of ice sheet melting, sea level rise contribution rate, and climate feedback intensity. Combined with a generative adversarial network, generate an extreme climate scenario data enhancement training set, and dynamically update model parameters on distributed nodes through online federated learning; Prediction: Input the multi-source data collected in real time into the trained multi-scale spatiotemporal fusion model to output the prediction results of the Antarctic amplification effect; the prediction results include the quantified values ​​of ice sheet melting, sea level change and climate feedback impact; Result analysis and verification: Time series analysis and spatial distribution analysis are performed on the predicted results of the Antarctic amplification effect, combined with sensitivity analysis and uncertainty assessment, to verify the model accuracy and generate a visual report.

2. The method for predicting the Antarctic amplification effect based on multi-scale spatiotemporal fusion and physical constraints according to claim 1 is characterized in that: The data collection also includes quality verification of multi-source data to ensure that the missing rate is less than 2%, and unifying the data to a daily resolution through spatiotemporal alignment to ensure the integrity, consistency and relevance of the Antarctic amplification effect.

3. The method for predicting the Antarctic amplification effect based on multi-scale spatiotemporal fusion and physical constraints according to claim 1 is characterized in that: The feature engineering in the data preprocessing also includes: Generate time series features based on sliding windows and extract multi-scale spatiotemporal features. The window length is no less than 30 days. Dynamically adjust feature weights through reinforcement learning, adaptively screen the most important feature combinations for predicting the amplification effect, and adjust weights no less frequently than once a day.

4. The method for predicting the Antarctic amplification effect based on multi-scale spatiotemporal fusion and physical constraints according to claim 1 is characterized in that: In the model construction and training, a multi-scale spatiotemporal fusion model is constructed, wherein the model includes: Multimodal Transformer module, which uses a multi-head self-attention mechanism to fuse the temporal features of multi-source data, with at least 4 layers and at least 128 hidden dimensions per layer; A graph neural network module, which captures the spatial dependency of ice sheets based on a dynamic graph structure, with a grid resolution of no less than 100km×100km; The physical constraint layer is constructed by embedding the mass conservation equation Δ𝑀=𝑀 𝑖𝑛 −𝑀 𝑜𝑢𝑡 And the heat flux balance equation 𝑄=𝜌𝑐Δ𝑇 optimizes the prediction results.

5. The method for predicting the Antarctic amplification effect based on multi-scale spatiotemporal fusion and physical constraints according to claim 1 is characterized in that: In the model construction and training, multi-task collaborative training includes achieving multi-objective optimization through a weighted loss function, where the loss function is defined as Loss=𝑤1⋅MSE 融化量 +𝑤2⋅MSE 海平面 +𝑤3⋅MSE 反馈 , weights 𝑤1, 𝑤2, 𝑤3; or / and, in the model construction and training, online federated learning includes, collaboratively updating model parameters through at least 5 distributed nodes, using the FedAvg algorithm, fusing global real-time observation data every 24 hours, and maintaining the dynamic adaptability of the model to climate change.

6. The method for predicting the Antarctic amplification effect based on multi-scale spatiotemporal fusion and physical constraints according to claim 1 is characterized in that: The prediction also includes outputting the confidence interval of the prediction results and predicting the long-term trend of the Antarctic ice sheet effect through time series analysis.

7. The method for predicting the Antarctic amplification effect based on multi-scale spatiotemporal fusion and physical constraints according to claim 1 is characterized in that: The result analysis and verification also includes calculating the model performance indicators using an independent test set and comparing them with the prediction results of the traditional physical model; or / and, also includes outputting the prediction results and analysis reports in the form of charts, heat maps or three-dimensional visualizations, with the resolution of the heat map not less than 100km×100km.

8. A prediction system for the Antarctic amplification effect based on multi-scale spatiotemporal fusion and physical constraints, characterized by: include, Data collection unit: collects multi-source data related to the Antarctic ice sheet in real time and obtains ice sheet dynamic information; the multi-source data includes satellite remote sensing data, ground observation data and meteorological reanalysis data; Data preprocessing unit: Clean the collected multi-source data to remove noise, outliers and missing values; extract multi-scale spatiotemporal features related to the Antarctic amplification effect based on feature engineering; construct a dynamic graph based on the ice sheet spatial grid, and the structure represents the relationship between heat transfer and melting diffusion between regions; standardize the data; among them, the multi-scale spatiotemporal features related to the Antarctic amplification effect include the ice sheet thickness change rate, ocean heat flux and atmospheric circulation index; Model construction and training unit: Construct a multi-scale spatiotemporal fusion model, adopt a multi-task collaborative training method, optimize the prediction of ice sheet melting, sea level rise contribution rate and climate feedback intensity through weighted loss function, combine generative adversarial network to generate extreme climate scenario data enhancement training set, and dynamically update model parameters on distributed nodes through online federated learning; Prediction unit: Input the multi-source data collected in real time into the trained multi-scale spatiotemporal fusion model and output the prediction results of the Antarctic amplification effect; the prediction results include the quantified values ​​of ice sheet melting, sea level change and climate feedback impact; Result Analysis and Verification Unit: Conduct time series analysis and spatial distribution analysis on the predicted results of the Antarctic amplification effect, combine sensitivity analysis and uncertainty assessment, verify the model accuracy and generate a visual report.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method for dynamic prediction of the Antarctic amplification effect based on multi-scale space-time fusion and physical constraints as described in any one of claims 1 to 7 is implemented.

10. An electronic device, characterized in that: It comprises a processor and a memory, wherein a computer program is stored in the memory, and the computer program is loaded and executed by the processor to implement the dynamic prediction method of the Antarctic amplification effect based on multi-scale space-time fusion and physical constraints as described in any one of claims 1 to 7.

Images