Arctic sea surface satellite-based hourly temperature reconstruction correction method, device and medium

Abstract

This application provides a method, apparatus, and medium for hourly reconstruction and correction of Arctic sea surface temperature using satellite-based data, relating to the field of spatiotemporal reconstruction technology of satellite remote sensing data. The method includes: acquiring relevant data and performing data preprocessing to obtain an initial dataset; performing multi-scale decomposition of temperature at the annual, seasonal, monthly, weekly, and daily scales, using a multi-scale temperature background field and multi-source auxiliary environmental factors as input, and progressively filling in missing pixels using LightGBM to obtain a spatially continuous basic temperature field; constructing a residual correction module, using the basic temperature field and multi-source auxiliary environmental factors as input, outputting a correction term through residual learning and adding it to the basic field to obtain the final temperature data; and outputting a high spatiotemporal resolution near-sea surface temperature product after verification using radiosonde data. This application can effectively improve the spatiotemporal continuity and physical consistency of satellite-based data, and accurately characterize the intraday variation, intermonthly evolution, and local thermal structure characteristics of temperature in polar near-sea areas.

Description

Technical Field

[0001] This application relates to the field of satellite remote sensing parameter inversion technology, and in particular to a method, device and medium for hourly reconstruction and correction of Arctic sea surface satellite-based air temperature. Background Technology

[0002] The Arctic region is a sensitive area and amplifier of global climate change. Changes in its near-surface air temperature (NSAT) not only directly affect sea ice melting, air-sea interaction, and regional energy balance, but also exert long-range influences on mid- and low-latitude climates by affecting atmospheric circulation. Therefore, obtaining high-precision, high-temporal-resolution Arctic near-surface air temperature data is of great significance for understanding the mechanisms of polar climate change and its global impacts. As a key parameter characterizing near-surface thermal state, near-surface air temperature plays an irreplaceable role in polar energy budget, sea ice process simulation, and climate model validation.

[0003] Currently, the acquisition of Arctic near-sea surface temperature data mainly relies on three methods: ground observation, reanalysis data, and satellite remote sensing inversion. Ground observation data has high accuracy, but it is extremely sparse in the Arctic Ocean, making it difficult to meet the needs of continuous spatial analysis. Reanalysis data (such as ERA5) assimilates multi-source observation data through numerical models, providing a continuous spatiotemporal meteorological field, but it still suffers from certain model dependence and systematic errors in high-latitude regions. Satellite remote sensing, with its high spatial resolution and wide-area observation capabilities, has become an important means of acquiring regional temperature information. Polar-orbiting satellites (such as Terra and Aqua) have the advantage of high-frequency transits in polar regions due to orbital intersections; a single MODIS sensor can obtain approximately 7-8 transit observations per day in high-latitude polar regions. Its MODIS atmospheric profile products (MOD07_L2 / MYD07_L2) are L2-level strip data with a raw spatial resolution of 5 km and a temporal granularity of 5 minutes. However, due to cloud cover, orbital limitations, and the complex polar environment, there are many data gaps in the raw MODIS atmospheric profile temperature products. The average coverage of raw MODIS temperature data north of 60°N is only 7.93%. In particular, in high-latitude regions such as the central Arctic Ocean, the Greenland Sea, and the Norwegian Sea, the missing rate is close to 1, forming a significant spatial observation gap. This severely restricts its application at the regional scale, especially in polar seas, and the accuracy of subsequent climate analysis.

[0004] To date, various methods have been proposed to reconstruct satellite remote sensing temperature data products. Initially, spatial or spatiotemporal interpolation methods were widely used, such as inverse distance weighting, ordinary kriging, and time-series interpolation methods based on annual temperature cycles, utilizing the spatial proximity information of the temperature products themselves to fill gaps. However, these methods suffer from significantly reduced temporal similarity and spatial correlation of neighboring points when cloud cover is frequent or the data gap is large, leading to increased uncertainty in the reconstruction results. Another common approach is to introduce multi-source covariates through machine learning to improve reconstruction capabilities. Many scholars have used algorithms such as gradient boosting and random forests to construct temperature estimation models; however, these methods still mainly rely on empirical mapping and are sensitive to the coverage of training samples and the quality of auxiliary variables. Accuracy and generalization ability are easily affected by regional changes or insufficient samples. In recent years, the rise of deep learning methods has further promoted the development of temperature reconstruction towards spatiotemporal field reconstruction. Some scholars have used deep neural networks to construct polar temperature reconstruction datasets, which can obtain spatiotemporally continuous and highly accurate temperature products. However, most existing deep learning methods output products at monthly or daily scales, and their temporal resolution is often limited to the daily scale, making it difficult to capture the high-frequency variation characteristics of temperature.

[0005] Generally, near-shore temperatures in the Arctic are influenced by factors such as solar radiation, sea ice melting, and atmospheric circulation on short timescales, exhibiting significant diurnal variations. Therefore, developing high temporal resolution (e.g., 3-hour intervals) temperature retrieval algorithms is most suitable for practical applications. However, most current algorithms output monthly or daily-scale products, failing to fully utilize the high-frequency transit advantages of polar-orbiting satellites in polar regions. Significant shortcomings remain in achieving high-precision, 3-hour interval, and full-coverage temperature retrieval over large areas simultaneously. Summary of the Invention

[0006] This application provides a method, device, and medium for hourly reconstruction and correction of Arctic sea surface satellite-based air temperature, which can realize near-sea surface air temperature inversion by integrating multi-source physical constraints and deep spatiotemporal networks, thereby breaking the trade-off between spatiotemporal resolution and coverage integrity of existing products.

[0007] In a first aspect, this application provides a method for hourly reconstruction and correction of Arctic sea surface satellite-based air temperature, the method comprising: MODIS atmospheric profile temperature data, ERA5 reanalysis data, and multi-source auxiliary environmental factors were acquired. The missing data were processed by linear interpolation and then spatiotemporal matching and preprocessing were performed to obtain the initial dataset. Based on the initial dataset, the temperature signal is decomposed into multiple scales to establish a multi-scale temperature background field. Using the multi-scale temperature background field and multi-source auxiliary environmental factors as input, the LightGBM model is used to progressively fill in missing pixels, and a physical constraint-driven full-coverage reconstruction module is constructed to obtain a spatially continuous basic temperature field. A deep spatiotemporal residual correction module is constructed. The basic temperature field and multi-source auxiliary environmental factors are input into a network composed of a graph convolutional network, a temporal convolutional network and a Transformer. The correction term is output through residual learning and added to the basic field to obtain the final temperature data. The accuracy of the final temperature data was verified using radiosonde data to obtain near-sea surface temperature products.

[0008] In one possible design, MODIS atmospheric profile temperature data, ERA5 reanalysis data, and multi-source auxiliary environmental factors are acquired. The missing data are then processed using linear interpolation followed by spatiotemporal matching and preprocessing to obtain an initial dataset, including: Acquire 1000 hPa temperature layer data from MOD07_L2 and MYD07_L2 products of Terra and Aqua satellites, extract effective pixels in the area north of 60°N, and merge them according to a 3-hour time window to obtain partial coverage data of MODIS near-sea surface air temperature at 3-hour intervals. The following data were obtained from the hourly reanalysis data of ERA5 as the multi-source auxiliary environmental factors: 10 m wind speed, 2 m dew point temperature, net solar radiation at the surface, 2 m air temperature, mean sea level pressure, and downward longwave radiation flux at the surface. The MODIS near-sea surface temperature partial coverage data and the multi-source auxiliary environmental factors were uniformly resampled to a spatial resolution of 5km and spatiotemporally matched to obtain the initial dataset.

[0009] In one possible design, based on the initial dataset, the temperature signal is decomposed into a multi-scale temperature background field; using the multi-scale temperature background field and multi-source auxiliary environmental factors as input, the LightGBM model is used to progressively fill in missing pixels, constructing a physically constrained, full-coverage reconstruction module to obtain a spatially continuous basic temperature field, including: A long-term background field is constructed based on annual temperature statistics, and seasonal, monthly and weekly temperature change information are added in sequence. At the same time, sea level pressure and wind field data from the multi-source auxiliary environmental factors are introduced as auxiliary constraints to obtain temperature background at different time scales. Based on the LightGBM model, using temperature background at different time scales and multi-source auxiliary environmental factors as inputs, a nonlinear mapping relationship between temperature and various influencing factors is established through a nonlinear mapping function. Missing pixels in the original data are gradually filled in layer by layer to obtain a spatially continuous basic temperature field.

[0010] In one possible design, the nonlinear mapping function is expressed as: in, , , , , These represent the statistical background of temperature on annual, seasonal, monthly, weekly, and daily scales, respectively. Temperature products obtained by interpolating irregular triangular meshes from raw MODIS hourly-scale data. For a multi-source auxiliary environmental factor set, Nonlinear mapping function constructed for LightGBM, This is the residual term.

[0011] In one possible design, a deep spatiotemporal residual correction module is constructed. The base temperature field and multi-source auxiliary environmental factors are input into a network composed of a graph convolutional network, a temporal convolutional network, and a Transformer, connected in series. Through residual learning, a correction term is output and added to the base field to obtain the final temperature data, including: The multi-source auxiliary environmental factors of each grid cell are used as nodes in the graph structure. An adjacency matrix is ​​constructed based on physical similarity. Then, a graph convolutional network is used to spatially aggregate the node features to obtain spatially enhanced features. The spatial augmentation features are fed into a temporal convolutional network, and causal convolution and dilated convolution are used to capture the dependencies across time steps to obtain temporal augmentation features. The spatial enhancement features and the temporal enhancement features are concatenated along the feature dimension and fused through a linear projection layer to obtain a fused feature sequence. The fused feature sequence is input into the Transformer encoder, and global dependencies are modeled through a multi-head self-attention mechanism to obtain global features; The global features are mapped into temperature residual correction terms using a multilayer perceptron. The temperature residual correction terms are then added to the base temperature field to obtain the final temperature data.

[0012] In one possible design, the computational process of using a graph convolutional network to spatially aggregate node features to obtain spatially enhanced features can be represented as follows: in, For spatial enhancement features, For activation function, To introduce a self-connected adjacency matrix, For the corresponding degree matrix, For learnable weights, For the input feature matrix, This is an adjacency matrix constructed based on physical similarity. It is the identity matrix; The spatial augmentation features are fed into a temporal convolutional network, and causal convolution and dilated convolution are used to capture dependencies across time steps. The computation process of obtaining the temporal augmentation features is expressed as follows: in, for t Temporal enhancement features at any given moment For convolution kernel, As the expansion factor, The kernel size is [size]. The index of the convolution kernel (from 0 to ...) ), for GCN output characteristics at time step; The calculation process of concatenating the spatial enhancement features and the temporal enhancement features along the feature dimension and fusing them through a linear projection layer to obtain the fused feature sequence is expressed as follows: in, To fuse feature sequences, Used to normalize the fused features. The weight matrix is ​​a learnable matrix. It is a learnable bias vector; The global features are mapped to a temperature residual correction term using a multilayer perceptron. The temperature residual correction term is then added to the base temperature field to obtain the final temperature data. The calculation process is as follows: in, This is the temperature residual correction term. It is a multilayer perceptron. As a global feature, Based on the basic temperature field, This is the final temperature data.

[0013] In one possible design, radiosonde data is used to verify the accuracy of the final temperature data, resulting in near-sea surface temperature products, including: Four strategies—direct validation, sample cross-validation, time cross-validation, and site cross-validation—along with monthly validation, latitude zone validation, and sensitivity testing under different data coverage, were employed to evaluate the accuracy of the final temperature data and output Arctic near-sea surface temperature products that meet the accuracy requirements.

[0014] Secondly, this application provides an hourly reconstruction and correction device for Arctic sea surface satellite-based air temperature. The device includes: The data acquisition module is configured to acquire MODIS atmospheric profile temperature data, ERA5 reanalysis data, and multi-source auxiliary environmental factors. The missing data is processed by linear interpolation and then spatiotemporal matching and preprocessing are performed to obtain the initial dataset. The full-coverage reconstruction module is configured to perform multi-scale decomposition of the temperature signal based on the initial dataset to establish a multi-scale temperature background field; using the multi-scale temperature background field and multi-source auxiliary environmental factors as input, the LightGBM model is used to progressively fill in missing pixels to construct a physically constrained full-coverage reconstruction module and obtain a spatially continuous basic temperature field. The residual correction module is configured to construct a deep spatiotemporal residual correction module. It inputs the basic temperature field and multi-source auxiliary environmental factors into a network composed of a graph convolutional network, a temporal convolutional network, and a Transformer. Through residual learning, it outputs a correction term and adds it to the basic field to obtain the final temperature data. The verification output module is configured to use radiosonde data to verify the accuracy of the final temperature data, thereby obtaining near-sea surface temperature products.

[0015] Thirdly, embodiments of this application provide an electronic device, including: at least one processor and a memory; the memory stores computer-executable instructions; the at least one processor executes the computer-executable instructions stored in the memory, causing the at least one processor to perform the Arctic sea surface satellite-based temperature hourly reconstruction and correction method as described in the first aspect and various possible designs of the first aspect.

[0016] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions. When a processor executes the computer-executable instructions, it implements the hourly reconstruction and correction method for Arctic sea surface satellite-based air temperature as described in the first aspect and various possible designs of the first aspect.

[0017] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the hourly reconstruction and correction method for Arctic sea surface satellite-based air temperature as described in the first aspect and various possible designs of the first aspect.

[0018] The method, apparatus, and medium for hourly reconstruction and correction of Arctic sea surface satellite-based air temperature provided in this application have at least the following beneficial effects: This application enables the acquisition of near-surface air temperature products with full coverage of the Arctic Ocean, 3-hour intervals, and 5km spatial resolution. Validated by independent radiosonde data, the final product R... 2 The accuracy reached 0.83, with a root mean square error of 2.34 K. After restoration, the original MODIS temperature data improved from an average coverage of 7.93% to 100% coverage, exhibiting good spatiotemporal continuity and physical consistency. Furthermore, the corrected temperature products clearly reproduce the latitudinal gradient distribution of Arctic temperatures, the diurnal variation characteristics on a 3-hour scale, and the inter-monthly evolution patterns during the warm season. They accurately capture the local thermal structures driven by ocean current transport and land-sea thermal differences. Attached Figure Description

[0019] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0020] Figure 1 A flowchart of an hourly reconstruction and correction method for Arctic sea surface satellite-based air temperature according to an embodiment of the present invention is shown; Figure 2 A spatiotemporal distribution diagram comparing the accuracy before and after using the depth spatiotemporal residual correction module according to an embodiment of this application is shown; Figure 3 A scatter plot comparing the accuracy before and after using the depth spatiotemporal residual correction module according to an embodiment of this application is shown; Figure 4 The residual distribution diagram of the physical guidance depth network according to an embodiment of this application is shown in different latitudinal regions; Figure 5 The following diagram shows the verification results of the Physically Guided Deep Network according to an embodiment of this application at different coverage levels; Figure 6 The spatial distribution map of the nearshore surface temperature product recovered by the model according to an embodiment of this application in the target area is shown. Figure 7 A spatial distribution map of the monthly average nearshore surface temperature product in the target area, reconstructed from the model according to an embodiment of this application, is presented. Figure 8 The spatial distribution pattern of nearshore surface air temperature products in the target area and the reconstruction results of typical sea areas, which are restored by the model according to the embodiments of this application, are presented. Figure 9 A structural diagram of an hourly reconstruction and correction device for Arctic sea surface satellite temperature according to an embodiment of this application is shown.

[0021] It should be noted that, Figures 2 to 8 The Arctic land-sea boundary data involved in the data comes from the Global Self-Consistent Hierarchical High-Resolution Geographic Database (GSHHG) provided by the National Center for Environmental Information (NCEI) of the National Oceanic and Atmospheric Administration (NOAA), version 2.3.7 (https: / / www.ngdc.noaa.gov / mgg / shorelines / ).

[0022] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concepts of this application to those skilled in the art through reference to specific embodiments. Detailed Implementation

[0023] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0024] The collection, storage, use, processing, transmission, provision, and disclosure of financial data or user data involved in the technical solution of this application all comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0025] It should be noted that in the embodiments of this application, certain software, components, models and other existing solutions in the industry may be mentioned. These should be regarded as exemplary and are only intended to illustrate the feasibility of implementing the technical solution of this application. However, it does not mean that the applicant has used or necessarily used the solution.

[0026] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0027] This application provides a method for hourly reconstruction and correction of Arctic sea surface satellite-based air temperature, such as... Figure 1 As shown, the method includes the following steps S10 to S40, which are described in detail below.

[0028] S10: Acquire MODIS atmospheric profile temperature data, ERA5 reanalysis data, and multi-source auxiliary environmental factors. After processing the missing data using linear interpolation, perform spatiotemporal matching and preprocessing to obtain the initial dataset.

[0029] In this embodiment, atmospheric profile products (MOD07_L2 and MYD07_L2) provided by the MODIS sensors on the Terra and Aqua satellites are used as the primary satellite data source. These products are L2-level strip data with a raw spatial resolution of 5 km and a temporal granularity of 5 minutes. Temperature data from the 1000 hPa pressure layer are extracted, and high-confidence effective pixels are selected using the product's built-in quality assurance (QA) information. Due to the orbital convergence advantage of polar-orbiting satellites in polar regions, a single MODIS sensor can obtain approximately 7-8 overpass observations per day in areas north of 60°N. The observation results from Terra and Aqua are merged into a 3-hour time window to construct a 3-hour interval gridded data.

[0030] We obtained hourly reanalysis data from ERA5 provided by the European Centre for Medium-Range Weather Forecasts (ECMWF), selecting the following variables: 10 m wind speed, 2 m dew point temperature, net solar radiation at the surface, 2 m air temperature, mean sea level pressure, and downward longwave radiation flux at the surface. The original spatial resolution of this dataset is 0.25° × 0.25°, and the temporal resolution is 1 hour.

[0031] The site data used was the IGRA (Integrated Global Radiosonde Archive) radiosonde dataset published by the National Center for Environmental Information (NCEI) as an independent validation data source. This dataset directly acquires vertical atmospheric profiles via radiosonde balloons, possessing extremely high vertical resolution and scientific authority. For the Arctic Ocean study area, 30 radiosonde stations north of 60°N, spatially evenly distributed, and with complete observation sequences were selected, and their 1000 hPa standard isobaric surface temperature observations were extracted. These stations will be used for independent evaluation in subsequent accuracy validation and will not participate in the model training process.

[0032] For missing values ​​in the MODIS and ERA5 data, linear interpolation was first used for preliminary filling. Then, all data (including the location information of radiosonde stations) were uniformly resampled to a spatial resolution of 5 km and spatiotemporal matching was performed to obtain the initial dataset. In terms of the spatiotemporal matching strategy, the time dimension was based on the satellite observation time, and a ±1 hour time window was set to average the temperature observations of each station. In the spatial dimension, a 10 km neighborhood was constructed centered on the latitude and longitude of the station, and the corresponding temperature product pixels were spatially averaged to establish matching relationships.

[0033] S20: Based on the initial dataset, a physical constraint-driven full-coverage reconstruction module is constructed by adopting a multi-scale temperature signal decomposition and hierarchical gradient boosting strategy. LightGBM is used to progressively fill in missing pixels to obtain a spatially continuous basic temperature field.

[0034] In this embodiment, considering that the changes in near-shore temperature in the Arctic are influenced by long-term climate trends, seasonal sea ice melting, and short-term weather processes, exhibiting a typical multi-scale nested structure, a reconstruction strategy of multi-scale progressive decomposition and fusion was designed.

[0035] An irregular triangular mesh is used to spatially aggregate the original sparse observations, extracting annual, seasonal, monthly, and weekly temperature statistical backgrounds as stable physical reference surfaces. Then, a four-layer LightGBM cascade model is constructed: The first layer uses annual and seasonal temperature background as the main explanatory variables to calculate the monthly average temperature value: The second layer uses annual, seasonal, and monthly temperature products to fill in the weekly temperature information: The third layer uses annual, seasonal, monthly, and weekly temperature products to fill in the daily temperature information: The fourth layer uses annual, seasonal, monthly, weekly, and daily temperature backgrounds, as well as spatial auxiliary variables (including 10-meter wind speed, mean sea level pressure, net surface solar radiation, surface downdraft longwave radiation flux, and 2-meter dew point temperature) as inputs to perform nonlinear approximation on missing pixels at the hourly scale: in, , , , , These represent the statistical background of temperature on annual, seasonal, monthly, weekly, and daily scales, respectively. Temperature products obtained by interpolating irregular triangular meshes from raw MODIS hourly-scale data. For a multi-source auxiliary environmental factor set, Nonlinear mapping function constructed for LightGBM, This is the residual term.

[0036] Model training relies solely on high-confidence original satellite observation pixels. During the product generation stage, a lossless mosaicking strategy is employed to retain the original observation pixels while filling in missing pixels, ultimately yielding a spatially continuous basic temperature field.

[0037] S30: Construct a deep spatiotemporal residual correction module, which connects the base temperature field and the auxiliary factor input graph through convolution, temporal convolution, and Transformer in a concatenated network. The correction term is output through residual learning and added to the base field to obtain the final temperature data.

[0038] In this embodiment, a deep spatiotemporal residual correction module was designed to correct the system residuals output by the full-coverage reconstruction module. This module adopts a residual learning strategy, and the overall process is divided into two stages: multi-dimensional spatiotemporal feature extraction and global feature correction learning.

[0039] In the multidimensional spatiotemporal feature extraction stage, the multi-source auxiliary environmental factors of each grid cell are treated as nodes in a graph structure, and an adjacency matrix A is constructed based on the physical similarity between variables (such as K-nearest neighbor relationships). Spatial feature aggregation is performed through a graph convolutional network (GCN), and its single-layer propagation rule is as follows: in, For spatial enhancement features, For activation function, To introduce a self-connected adjacency matrix, For the corresponding degree matrix, For learnable weights, For the input feature matrix, This is an adjacency matrix constructed based on physical similarity. This is the identity matrix. Through this process, GCN can learn the spatial collaborative weights of various meteorological factors in the unique geographical environment of the polar regions, thereby transforming isolated grid data into a physical feature stream with spatial topological attributes. .

[0040] Spatial enhancement features of GCN output The data is fed into a Temporal Convolutional Network (TCN). TCN employs causal convolution to ensure temporal causal consistency and uses dilated convolution to exponentially expand the receptive field, thereby effectively capturing dependencies over long time spans. in, for t Temporal enhancement features at any given moment For convolution kernel, As the expansion factor, The kernel size is [size]. The index of the convolution kernel (from 0 to ...) ), for The output characteristics of the GCN at any given time. Meanwhile, the serial design from GCN to TCN ensures that the input of TCN is a semantic feature enhanced with spatial relation, thereby enabling it to more sensitively identify complex time-varying residual patterns caused by polar air-sea coupling.

[0041] In the global feature correction learning stage, the spatial static background (spatial enhancement features) is... ) and time dynamic increment (temporal enhancement feature) The features are concatenated along the feature dimension and fused into a unified feature space through a linear projection layer. in, To fuse feature sequences, Used to normalize the fused features. The weight matrix is ​​a learnable matrix. This is a learnable bias vector.

[0042] Fusion feature sequence It is fed into the Transformer encoder. The encoder consists of two sub-layers: a multi-head self-attention network (MHSA) and a feedforward network (FFN). Each sub-layer uses residual connections and layer normalization (LayerNorm).

[0043] First Through three learnable linear transformation matrices Mapped to query matrices respectively Key matrix Sum matrix : The scaling dot product attention is calculated as follows: To capture information from different feature subspaces, a multi-head attention mechanism is employed. Segmented along the feature dimension Size (dimensions of each head) After calculating the attention separately, they are concatenated and then subjected to a linear transformation. Fusion: Subsequently, the first residual connection and layer normalization are applied to obtain intermediate features. : Will The data is fed into a feedforward network, and residual connections and layer normalization are performed again to obtain global features. : Finally, a three-layer perceptron (MLP) is used to process global features. Mapped to temperature residual correction term To obtain the final corrected temperature: in, This is the temperature residual correction term. It is a multilayer perceptron. As a global feature, Based on the basic temperature field, This is the final temperature data.

[0044] S40: Utilizes radiosonde data to verify the accuracy of the final temperature data, and outputs high-precision, full-coverage Arctic near-sea surface temperature products with 3-hour intervals.

[0045] In this embodiment, the IGRA radiosonde dataset published by the National Center for Environmental Information (NCEI) was used as independent validation data. Thirty radiosonde stations north of 60°N, with uniform spatial distribution and complete observation sequences, were selected, and their 1000 hPa standard isobaric surface temperature observations were extracted. In the temporal dimension, hourly temperature observations for each station were averaged within a ±1-hour time window, using satellite observation time as the baseline. In the spatial dimension, a 10 km neighborhood was constructed centered on the station's latitude and longitude, and spatial averaging was performed on the corresponding temperature product pixels to establish a matching relationship.

[0046] For example, this embodiment evaluates the performance of the proposed model in several ways. First, the full-coverage Arctic nearshore surface air temperature product generated by the model was validated against independent radiosonde data. The accuracy of the original MODIS temperature data was compared with 1000 hPa temperature observations from independent radiosonde stations (IGRA). The accuracy improvement effect before and after using the depth spatiotemporal residual correction module was visually demonstrated through spatiotemporal distribution maps and scatter plots. Subsequently, the accuracy improvement effect was further demonstrated using data including the coefficient of determination (R²). 2 The accuracy of the proposed method is evaluated using three statistical metrics, including R0, root mean square error (RMSE), and mean absolute error (MAE). 2 The RMSE is used to measure the goodness of fit between the reconstructed temperature and the observed values, while the MAE is used to assess the magnitude of the average error. The MAE reflects the actual magnitude of the error without being overly influenced by outliers. The formulas for calculating each indicator are as follows: in, Indicates the first Reconstructed or corrected sea surface temperature values ​​for each sample. This represents the observation value at the corresponding station. and These are their average values, and N is the sample size.

[0047] from Figure 2 It can be seen that the accuracy of the product before the residual correction module exhibits significant latitudinal heterogeneity, showing larger errors in the low-latitude regions of the southern part of the study area. 2 While the overall accuracy was low, RMSE and MAE increased significantly. After residual correction, the accuracy indicators of all stations in the region achieved an overall leap. Although the corrected products still retained the distribution trend of slightly lower accuracy in low-latitude regions compared to high-latitude regions, the residual correction module effectively reduced the accuracy differences between different latitudinal zones. Based on the spatiotemporal distribution of station errors, this embodiment employs four strategies—direct verification, sample cross-validation, time cross-validation, and station cross-validation—to assess the accuracy of the final Arctic near-sea surface temperature product. Figure 3 It can be seen that significant accuracy fluctuations and systematic errors exist at different times before residual correction. Overall verification R 2 The R value is only 0.54, and the R value for each time period is... 2 The values ​​are distributed between 0.48 and 0.59, while the RMSE generally exceeds 4K, with this uncertainty being particularly pronounced around 00:00 and 06:00. After residual correction, the product's accuracy at all times achieved a significant improvement. Overall validation R 2 The RSI reached 0.83, and the RMSE dropped to 3.15K. Analysis of the results across different time periods reveals that the product's performance across all four time periods remained highly consistent, with its time-specific RSI... 2 They have remained stable at around 0.78-0.90, and the RMSE has remained at a low level of around 2.50K.

[0048] This embodiment further evaluates the spatial robustness of the method across different latitude regions. The study area is divided into region 1 (60°N-67.5°N), region 2 (67.5°N-75°N), and region 3 (75°N-90°N). From Figure 4It is evident that in Region 1, influenced by complex air-sea boundary layer processes, the average deviation before residual correction reached -6.16K. As latitude shifts towards the North Pole, the systematic error improves compared to Region 1. After residual correction, the accuracy of all regions achieved a qualitative improvement. In Region 1, where correction was most challenging, the DST-RC model improved the average deviation from -6.16K to -0.38K. The average deviations after correction in Regions 2 and 3 also reached the same order of magnitude. Simultaneously, this embodiment conducted coverage sensitivity testing. Based on the distribution characteristics of the original effective observation data, the validation samples were divided into three scenarios: low coverage (<6.25%), medium coverage (6.25%~8.93%), and high coverage (≥8.93%). Figure 5 The model exhibits strong anti-interference capabilities and adaptability to extreme scenarios with insufficient data. Its performance is most outstanding when the original observations are relatively abundant. 2 The RSI reached 0.83, with an RMSE of only 3.29K. Even as coverage decreased, the method's performance remained highly stable. 2 The accuracy reached 0.80, with only minor fluctuations. Furthermore, even in low-coverage scenarios, the final product maintained reliable mapping accuracy (R²). 2 Maintain at 0.83, RMSE controlled at 2.92K.

[0049] To eliminate the interference of random weather systems and more intuitively reflect the model's reconstruction effect on the 3-hour characteristics of Arctic temperatures, Figure 6 This presentation displays the spatial distribution of average temperature across eight observation times (3-hour steps) during the example period (May-August 2024). The average distribution sequence from 00:00 to 21:00 clearly reproduces the subtle fluctuations in Arctic ocean temperature caused by the cycle of solar radiation angle. Temperatures peak around 12:00 and 15:00, and are relatively low around 00:00 and 03:00. Overall, the transitions between time points are smooth, demonstrating strong physical consistency of the model over time. Simultaneously, the time-scale mean values ​​of radiosonde data stations are overlaid, with the richest station data at 00:00 and 12:00, and the displayed values ​​are highly consistent with the map color. Although radiosonde data is sparse at other times, the existing values ​​from individual stations can still be accurately embedded within the numerical range of the reconstructed base map.

[0050] Figure 7Further evaluation of the method's ability to characterize features on a monthly scale during the warm season revealed that, from a temporal evolution perspective, May, as the beginning of the warm season, saw the Arctic region still largely controlled by cold air masses. The calibration product displayed a deep blue low-temperature zone, highly consistent with the low-value dots observed at the stations, accurately reproducing the cold environment below 270K in the core sea areas above 80°N. As solar radiation intensified, the Arctic region entered a period of accelerated melting (June-July). Although cloud formation during this period resulted in significant gaps in original observations, the product still clearly delineated temperature gradients. The high-temperature bands extending towards the poles in the Barents Sea and Chukchi Sea were accurately validated by radiosonde data along the route, reflecting the model's extremely high fidelity in responding to the thermal response caused by sea ice melting. In August, temperatures across the Arctic reached their peak, and the original central low-temperature zone significantly contracted. Even in the extremely complex thermal background of the marginal sea areas, the station data and the product background color remained highly consistent.

[0051] Figure 8The spatial distribution pattern of temperature across the entire Arctic ocean region was obtained. Overall, the study area exhibits a significant and smooth zonal gradient, with temperatures monotonically decreasing with increasing latitude. Spatially, high-temperature areas are mainly concentrated in marginal seas and landward openings near 60°N, while high-latitude seas above 80°N are controlled by perennial sea ice and the polar high pressure, maintaining relatively low temperatures of around 270K. Four typical sea areas—Barents Sea, Kara Sea, Chukchi Sea, and Beaufort Sea—were selected for comparative analysis. The Barents Sea, as a significant area of ​​Atlantic warm water input, shows markedly high temperatures in the figure. The product clearly demonstrates the path of warm air transported northeastward by the North Atlantic Current (AW), with temperatures generally exceeding 276K. This smooth and continuous temperature gradient illustration method effectively identifies ocean current-driven thermal characteristics. Located near the Siberian coast, the Kara Sea is influenced by both land-based thermal spillover and marginal sea circulation. The product shows a temperature pattern of higher temperatures in the southeast and lower temperatures in the northwest, accurately depicting the temperature gradient smoothly transitioning from the continental margin to the deep sea. In the Chukchi Sea region, warm air masses diffuse northward along the northern coast of Alaska, effectively reflecting the model's strong ability to capture local temperature characteristics influenced by the Pacific inflow and landward opening. Finally, in the Beaufort Sea region, the temperature distribution is relatively uniform and low, exhibiting a cool color tone and demonstrating high spatial consistency. Analysis of these regions shows that the method not only maintains high numerical synchronization with the average observations for each time period but also spatially fully recreates the dynamic expansion and contraction of temperature over time. This precise capture of sub-diurnal scale thermal characteristics provides a valuable high-frequency, full-coverage dataset for studying polar boundary layer processes and air-sea heat flux exchange at fine timescales.

[0052] This application also provides an embodiment of an hourly reconstruction and correction device for Arctic sea surface satellite-based air temperature, such as... Figure 9 As shown, the Arctic Ocean surface satellite-based hourly temperature reconstruction and correction device includes: The data acquisition module 901 is configured to acquire MODIS atmospheric profile temperature data, ERA5 reanalysis data and multi-source auxiliary environmental factors, and to perform spatiotemporal matching and preprocessing on the missing data using linear interpolation to obtain the initial dataset. The full-coverage reconstruction module 902 is configured to perform multi-scale decomposition on the temperature signal based on the initial dataset to establish a multi-scale temperature background field; using the multi-scale temperature background field and multi-source auxiliary environmental factors as input, the LightGBM model is used to progressively fill in missing pixels to construct a physical constraint-driven full-coverage reconstruction module and obtain a spatially continuous basic temperature field. The residual correction module 903 is configured to construct a deep spatiotemporal residual correction module, which inputs the basic temperature field and multi-source auxiliary environmental factors into a network composed of a graph convolutional network, a temporal convolutional network and a Transformer, and outputs a correction term through residual learning and adds it to the basic field to obtain the final temperature data. The verification output module 904 is configured to use radiosonde data to verify the accuracy of the final temperature data and obtain near-sea surface temperature products.

[0053] It should be noted that the modules described in the embodiments of this application can be implemented in software or hardware, and the described modules can also be located in the processor. Furthermore, the names of these modules do not necessarily constitute a limitation on the module itself.

[0054] The Arctic sea surface satellite-based temperature hourly reconstruction and correction device mentioned in this application embodiment belongs to the same technical concept as the previously described method, and its technical effect is basically the same, so it will not be repeated here.

[0055] This application provides an electronic device. The electronic device may include a processor and a memory, wherein the processor and the memory can communicate; exemplarily, the processor and the memory communicate via a communication bus.

[0056] The processor executes computer execution instructions stored in memory, causing the processor to perform the scheme in the above embodiments. The processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0057] The communication bus can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The system bus can be divided into address bus, data bus, control bus, etc. Transceivers are used to enable communication between database access devices and other computers (e.g., clients, read-write libraries, and read-only libraries). Memory may include random access memory (RAM) and may also include non-volatile memory.

[0058] The electronic device provided in this application embodiment can be the terminal device described in the above embodiments.

[0059] This application also provides a computer-readable storage medium storing computer instructions. When the computer instructions are executed on a computer, the computer performs the technical solution of the hourly reconstruction and correction method for Arctic sea surface satellite temperature described in the above embodiments.

[0060] This application also provides a computer program product, which includes a computer program stored in a computer-readable storage medium. At least one processor can read the computer program from the computer-readable storage medium. When the at least one processor executes the computer program, it can implement the technical solution of the hourly reconstruction and correction method for Arctic sea surface satellite temperature in the above embodiments.

[0061] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or modules, and may be electrical, mechanical, or other forms.

[0062] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to implement the solution of this embodiment according to actual needs.

[0063] Furthermore, the functional modules in the various embodiments of this application can be integrated into one processing unit, or each module can exist physically separately, or two or more modules can be integrated into one unit. The unit composed of the above modules can be implemented in hardware or in the form of hardware plus software functional units.

[0064] The integrated modules described above, implemented as software functional modules, can be stored in a computer-readable storage medium. These software functional modules, stored in a storage medium, include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute some steps of the methods of the various embodiments of this application.

[0065] It should be understood that the aforementioned processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. A general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly manifested as execution by a hardware processor, or execution by a combination of hardware and software modules within the processor.

[0066] The memory may include high-speed RAM, and may also include non-volatile storage (NVM), such as at least one disk storage device, and may also be a USB flash drive, external hard drive, read-only memory, disk or optical disc, etc.

[0067] Buses can be Industry Standard Architecture (ISA) buses, Peripheral Component Interconnect (PCI) buses, or Extended Industry Standard Architecture (EISA) buses, etc. Buses can be categorized into address buses, data buses, control buses, etc.

[0068] The aforementioned storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The storage medium can be any available medium that can be accessed by a general-purpose or special-purpose computer.

[0069] An exemplary storage medium is coupled to a processor, enabling the processor to read information from and write information to the storage medium. Alternatively, the storage medium can be an integral part of the processor. The processor and storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and storage medium can exist as discrete components in an electronic control unit or main control device.

[0070] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0071] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A method for hourly reconstruction and correction of Arctic sea surface satellite-based air temperature, characterized in that, The method includes: MODIS atmospheric profile temperature data, ERA5 reanalysis data, and multi-source auxiliary environmental factors were acquired. The missing data were processed by linear interpolation and then spatiotemporal matching and preprocessing were performed to obtain the initial dataset. Based on the initial dataset, the temperature signal is decomposed into multiple scales to establish a multi-scale temperature background field. Using the multi-scale temperature background field and multi-source auxiliary environmental factors as input, the LightGBM model is used to progressively fill in missing pixels, and a physical constraint-driven full-coverage reconstruction module is constructed to obtain a spatially continuous basic temperature field. A deep spatiotemporal residual correction module is constructed. The basic temperature field and multi-source auxiliary environmental factors are input into a network composed of a graph convolutional network, a temporal convolutional network and a Transformer. The correction term is output through residual learning and added to the basic field to obtain the final temperature data. The accuracy of the final temperature data was verified using radiosonde data to obtain near-sea surface temperature products.

2. The method according to claim 1, characterized in that, MODIS atmospheric profile temperature data, ERA5 reanalysis data, and multi-source auxiliary environmental factors were acquired. Missing data were processed using linear interpolation followed by spatiotemporal matching and preprocessing to obtain the initial dataset, which includes: Acquire 1000 hPa temperature layer data from MOD07_L2 and MYD07_L2 products of Terra and Aqua satellites, extract effective pixels in the area north of 60°N, and merge them according to a 3-hour time window to obtain partial coverage data of MODIS near-sea surface air temperature at 3-hour intervals. The following data were obtained from the hourly reanalysis data of ERA5 as the multi-source auxiliary environmental factors: 10 m wind speed, 2 m dew point temperature, net solar radiation at the surface, 2 m air temperature, mean sea level pressure, and downward longwave radiation flux at the surface. The MODIS near-sea surface temperature partial coverage data and the multi-source auxiliary environmental factors were uniformly resampled to a spatial resolution of 5km and spatiotemporally matched to obtain the initial dataset.

3. The method according to claim 1, characterized in that, Based on the initial dataset, the temperature signal is decomposed into a multi-scale temperature background field. Using the multi-scale temperature background field and multi-source auxiliary environmental factors as input, the LightGBM model is used to progressively fill in missing pixels, constructing a physically constrained, full-coverage reconstruction module to obtain a spatially continuous basic temperature field, including: A long-term background field is constructed based on annual temperature statistics, and seasonal, monthly and weekly temperature change information are added in sequence. At the same time, sea level pressure and wind field data from the multi-source auxiliary environmental factors are introduced as auxiliary constraints to obtain temperature background at different time scales. Based on the LightGBM model, using temperature background at different time scales and multi-source auxiliary environmental factors as inputs, a nonlinear mapping relationship between temperature and various influencing factors is established through a nonlinear mapping function. Missing pixels in the original data are gradually filled in layer by layer to obtain a spatially continuous basic temperature field.

4. The method according to claim 3, characterized in that, The nonlinear mapping function is expressed as: in, , , , , These represent the statistical background of temperature on annual, seasonal, monthly, weekly, and daily scales, respectively. Temperature products obtained by interpolating irregular triangular meshes from raw MODIS hourly-scale data. For a multi-source auxiliary environmental factor set, Nonlinear mapping function constructed for LightGBM, This is the residual term.

5. The method according to claim 1, characterized in that, A deep spatiotemporal residual correction module is constructed. The base temperature field and multi-source auxiliary environmental factors are input into a network composed of a graph convolutional network, a temporal convolutional network, and a Transformer, connected in series. Through residual learning, a correction term is output and added to the base field to obtain the final temperature data, including: The multi-source auxiliary environmental factors of each grid cell are used as nodes in the graph structure. An adjacency matrix is ​​constructed based on physical similarity. Then, a graph convolutional network is used to spatially aggregate the node features to obtain spatially enhanced features. The spatial augmentation features are fed into a temporal convolutional network, and causal convolution and dilated convolution are used to capture the dependencies across time steps to obtain temporal augmentation features. The spatial enhancement features and the temporal enhancement features are concatenated along the feature dimension and fused through a linear projection layer to obtain a fused feature sequence. The fused feature sequence is input into the Transformer encoder, and global dependencies are modeled through a multi-head self-attention mechanism to obtain global features; The global features are mapped into temperature residual correction terms using a multilayer perceptron. The temperature residual correction terms are then added to the base temperature field to obtain the final temperature data.

6. The method according to claim 5, characterized in that, The computational process of using graph convolutional networks to spatially aggregate node features to obtain spatially enhanced features is represented as follows: in, For spatial enhancement features, For activation function, To introduce a self-connected adjacency matrix, For the corresponding degree matrix, For learnable weights, For the input feature matrix, This is an adjacency matrix constructed based on physical similarity. It is the identity matrix; The spatial augmentation features are fed into a temporal convolutional network, and causal convolution and dilated convolution are used to capture dependencies across time steps. The computation process of obtaining the temporal augmentation features is expressed as follows: in, for t Temporal enhancement features at any given moment For convolution kernel, As the expansion factor, The kernel size is [size]. The index of the convolution kernel (from 0 to ...) ), for GCN output characteristics at time step; The calculation process of concatenating the spatial enhancement features and the temporal enhancement features along the feature dimension and fusing them through a linear projection layer to obtain the fused feature sequence is expressed as follows: in, To fuse feature sequences, Used to normalize the fused features. The weight matrix is ​​a learnable matrix. It is a learnable bias vector; The global features are mapped to a temperature residual correction term using a multilayer perceptron. The temperature residual correction term is then added to the base temperature field to obtain the final temperature data. The calculation process is as follows: in, This is the temperature residual correction term. It is a multilayer perceptron. As a global feature, Based on the basic temperature field, This is the final temperature data.

7. The method according to claim 1, characterized in that, The accuracy of the final temperature data was verified using radiosonde data to obtain near-sea surface temperature products, including: Four strategies—direct validation, sample cross-validation, time cross-validation, and site cross-validation—along with monthly validation, latitude zone validation, and sensitivity testing under different data coverage, were employed to evaluate the accuracy of the final temperature data and output Arctic near-sea surface temperature products that meet the accuracy requirements.

8. A satellite-based hourly temperature reconstruction and correction device for the Arctic sea surface, characterized in that, The device includes: The data acquisition module is configured to acquire MODIS atmospheric profile temperature data, ERA5 reanalysis data, and multi-source auxiliary environmental factors. The missing data is processed by linear interpolation and then spatiotemporal matching and preprocessing are performed to obtain the initial dataset. The full-coverage reconstruction module is configured to perform multi-scale decomposition of the temperature signal based on the initial dataset to establish a multi-scale temperature background field; using the multi-scale temperature background field and multi-source auxiliary environmental factors as input, the LightGBM model is used to progressively fill in missing pixels to construct a physically constrained full-coverage reconstruction module and obtain a spatially continuous basic temperature field. The residual correction module is configured to construct a deep spatiotemporal residual correction module. It inputs the basic temperature field and multi-source auxiliary environmental factors into a network composed of a graph convolutional network, a temporal convolutional network, and a Transformer. Through residual learning, it outputs a correction term and adds it to the basic field to obtain the final temperature data. The verification output module is configured to use radiosonde data to verify the accuracy of the final temperature data, thereby obtaining near-sea surface temperature products.

9. An electronic device, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes the computer execution instructions stored in the memory to implement the hourly reconstruction and correction method for Arctic sea surface satellite temperature as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the hourly reconstruction and correction method for Arctic sea surface satellite-based air temperature as described in any one of claims 1-7.

Images