Satellite sea temperature reconstruction method based on space-time constraint diffusion model

Abstract

The invention relates to the technical field of satellite remote sensing image processing and meteorological big data analysis, in particular to a satellite sea temperature reconstruction method based on a space-time constraint diffusion model, and the method comprises the steps: firstly obtaining Himawari L3-level satellite data and OSTIA L4-level global data with missing in a global sea area; then, the missing Himawari L3-level satellite data and the OSTIA L4-level global data are preprocessed, and a plurality of pieces of slice data are obtained; and finally, inputting the plurality of preprocessed slice data into the trained space-time constraint diffusion model, and reconstructing the plurality of slice data by the space-time constraint diffusion model through the weight file. And performing texture enhancement and large-scale reconstruction optimization on the plurality of reconstructed slice data to obtain complete reconstruction data of the whole sea area. According to the method, the texture reduction degree and the physical consistency precision of the satellite SST under continuous large-range cloud shielding are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of satellite remote sensing image processing and meteorological big data analysis, and particularly relates to a satellite sea surface temperature reconstruction method based on a space-time constraint diffusion model. BACKGROUND

[0002] Sea surface temperature (SST) is a key physical parameter representing the thermal state of the ocean and the energy exchange between the ocean and the atmosphere, and plays an important role in global climate change monitoring and ocean numerical prediction. At present, the new generation of geostationary meteorological satellites represented by the Himawari of the Japan Meteorological Agency can provide high temporal and spatial resolution L3-level (Level 3) skin sea temperature products. Such L3-level data, with the observation advantage of infrared sensors, has very high observation accuracy and can clearly capture fine dynamic texture features such as mesoscale eddies and ocean fronts.

[0003] However, infrared remote sensing technology is not only limited by cloud cover, but also easily disturbed by factors such as aerosol scattering and heavy precipitation, resulting in large-area, non-random data missing of L3-level SST products in actual application. In order to obtain a spatiotemporally continuous full-coverage SST field, the existing business means usually adopts L4-level (Level 4) multi-source fusion analysis products (such as the OSTIA system developed by the UK Meteorological Bureau). The L4-level data, through statistical methods such as optimal interpolation (OI), fuses data from multiple satellites and in-situ buoy data to fill in the missing data under the cloud.

[0004] Although the L4-level product realizes spatial full coverage of data, the statistical interpolation algorithm used by the L4-level product is essentially a smoothing estimation, which tends to retain low-frequency climate background information and severely suppresses high-frequency change signals. This results in the L4-level data being "complete" but low in accuracy, often losing the sharp texture and instantaneous dynamic details of the original L3-level data. In complex weather processes such as typhoons and frontal passages, the smoothed field provided by the L4-level data cannot truly reflect the severe changes on the ocean surface.

[0005] To solve this problem, existing deep learning reconstruction methods (such as methods based on convolutional neural networks or conventional generative adversarial networks) attempt to repair missing areas, but most of them fail to effectively utilize the historical high-precision observations of L3-level data as prior guidance and lack reasonable use of the low-frequency constraints of L4-level background fields. The existing technology often faces a dilemma when facing large-scale missing of L3-level data: either the reconstruction result is too smooth (degrading to the level of L4), or the generated texture is chaotic and does not conform to the physical law. Therefore, how to combine the high-precision texture advantage of L3-level data with the full-coverage background advantage of L4-level data, while ensuring the physical authenticity, to achieve high-precision reconstruction, is a key technical problem to be solved at present. SUMMARY

[0006] The application provides a satellite sea surface temperature reconstruction method based on a space-time constraint diffusion model, which can reconstruct SST data with missing satellite cloud layers and maintain original accuracy and texture, so as to solve the problem that the prior art mainly uses a space statistical interpolation method to forcibly fill in the missing part, uses a weighted average value of surrounding effective observation points to represent the sea surface temperature of the missing area, and causes high-frequency texture details to be smoothed and removed.

[0007] The application provides a satellite sea surface temperature reconstruction method based on a space-time constraint diffusion model, which comprises the following steps: Step 1: obtaining Himawari L3 satellite data with missing parts and OSTIA L4 global data in the global sea area; Step 2: preprocessing the missing Himawari L3 satellite data and the OSTIA L4 global data to obtain a plurality of slice data; Step 3: inputting the plurality of preprocessed slice data into a trained space-time constraint diffusion model, and reconstructing the plurality of slice data by the space-time constraint diffusion model through a weight file; and performing texture enhancement and large-scale reconstruction optimization on the plurality of reconstructed slice data to obtain complete reconstruction data in the global sea area.

[0008] As a preferred technical solution, in step 2, the preprocessing steps of the missing Himawari L3 satellite data and the OSTIA L4 global data are as follows: Step 21: performing bicubic interpolation on the OSTIA L4 global data to enlarge the same accuracy as the Himawari L3 satellite data, so that the Himawari L3 satellite data and the OSTIA L4 global data are aligned in longitude and latitude.

[0009] Step 22: performing 128x128 pixel sliding window slicing with a step length of 64 pixels for the corresponding sea area range to obtain a plurality of uniform slice data, saving the slice data as NPZ files, and naming the slice data according to the corresponding date; The step length is set to 64 pixels, that is, there is 50% area overlap between adjacent slices in the horizontal and vertical directions.

[0010] As a preferred technical solution, in step 3, the construction and training process of the space-time constraint diffusion model is as follows: Step 31: obtaining multi-source data; Step 32: slicing the multi-source data; Step 33: constructing a space-time constraint diffusion model and training the model to obtain a weight file; the weight file is used for reconstructing the missing Himawari L3 satellite data.

[0011] As a preferred technical solution, in step 31, the multi-source data includes Himawari L3 satellite data and OSTIAL4 global data; wherein the spatial latitude and longitude resolution of the Himawari L3 satellite data is 0.02°; and the spatial latitude and longitude resolution of the OSTIA L4 global data is 0.05°.

[0012] As a preferred technical solution, in step 32, the operation steps of the slice preprocessing are as follows: Step 321, select the sea area range corresponding to the multi-source data; Select the sea area range, set the image size to 1024x896 pixels, Step 322, perform bicubic interpolation on the OSTIA L4 global data to enlarge it to the same precision as the Himawari L3 satellite data, so that the latitude and longitude of the Himawari L3 satellite data and the OSTIA L4 global data are aligned; Step 323, perform cloud removal filtering on the Himawari L3 satellite data after aligning the latitude and longitude, and ensure that the valid pixels thereof match the sea area range of the OSTIA L4 global data; The process of performing cloud removal filtering on the Himawari L3 satellite data after aligning the latitude and longitude includes temperature scale conversion and outlier filtering; Temperature scale conversion: determine the initial temperature scale by detecting the pixel mean value; Outlier filtering: identify and remove outliers outside the reasonable sea temperature range, and mark them as NaN; Step 324, slice the corresponding sea area range with a sliding window of 128x128 pixels to obtain multiple uniform slices, save them as NPZ files, and name them according to the corresponding date.

[0013] As a preferred technical solution, in step 33, the construction process of the spatio-temporal constraint diffusion model is as follows: Step 331, first perform linear normalization on the multi-source data, and then perform data augmentation; Linear normalization: map the original sea temperature value to the interval [-1, 1]:

[0014] wherein, : the normalized multi-source data value; : the original multi-source data physical value (in Celsius); : the lower limit value of the multi-source data, which is 8.0; : the upper limit value of the multi-source data, which is 34.0; Data augmentation: random horizontal and vertical flipping is performed on the slices Step 332, extract the multi-dimensional physical time feature vector corresponding to the date in the slice file name ;

[0015] Step 333, forward noise is added to the Himawari L3 satellite data to obtain the Himawari L3 satellite data to be denoised;

[0016] Cosine noise schedule is used for noise addition: for a given true value image of Himawari L3 satellite data , the noise addition result at the th step is , which is calculated as:

[0017]

[0018] Wherein, : the normalized true value image of the Himawari L3 satellite data; : : the noisy latent variable generated at the moment; : represents a Gaussian distribution (normal distribution); : noise scaling factor defined by the cosine schedule; is the cumulative result of its multiplication; : unit matrix, representing standard normal distribution noise; Step 334, perform concatenation operation in channel dimension between the Himawari L3 satellite data to be denoised and the OSTIA L4 global data as background field constraint condition, denoising and predicting noise distribution at the same time;

[0019] The Himawari L3 satellite data to be denoised is set as , the OSTIA L4 global data is set as , and the :

[0020]

[0021] Wherein, : the probability distribution of the improved UNet model learning reverse generation process; : reconstructed image latent variable at the previous moment of the denoising process; : diffusion step number; : current : noisy image latent variable at the current time instant; : full coverage background field constraint provided by OSTIA data; : physical time feature vector containing date and year offset; : represents a Gaussian distribution; : Gaussian distribution mean predicted by the improved UNet model, representing the expected value of the reconstructed image; : Gaussian distribution variance predicted by the improved UNet model, representing the uncertainty of the reconstruction process; Step 335, by training the difference between the predicted noise and the real noise, the real pixels and textures in the Himawari L3 satellite data are identified, and the spatio-temporal constraint diffusion model is constructed; The identification of real pixels and textures in Himawari L3 satellite data is realized through loss function formula (3-2):

[0022] As a preferred technical solution, steps 331 to 335 are repeated to iteratively train the spatio-temporal constraint diffusion model, and the model parameters are continuously corrected using the gradient descent algorithm until the value of the loss function on the validation set reaches order of magnitude and remains stable, that is, the training is stopped and the trained spatio-temporal constraint diffusion model is obtained; the weight file is obtained.

[0023] As a preferred technical solution, in step 334, the improved UNet architecture is used for denoising, and the improved UNet architecture includes a ConvNeXt V2 module, a linear attention module with Memory KV, and a time double embedding layer. The ConvNeXt V2 module uses a global response normalization GRN mechanism in each ResBlock:

[0024]

[0025]

[0026]

[0027] wherein, : input feature map, dimension ; : feature map in spatial dimension ) on the norm, representing the channel strength; : normalized response value, reflecting the competition status between channels; : total number of channels; : a small constant to prevent division by zero; : element-wise multiplication; : learnable scaling and translation parameters, where, and are initialized to 0, so that the module behaves as an identity mapping at the beginning of training;

[0028] Linear attention module with Memory KV, which extracts features through formula (1), reduces the complexity from to ;

[0029] (1)

[0030] where Memory KV is added to learn the global physical prior:

[0031] where, : Query, representing the pixel features that need to be reconstructed; : Key and Value, representing the spatial context; : Memory KV, i.e., the pre-set learnable static parameters in the model, used to store the global sea temperature distribution prior; : Softmax normalization on the matrix row; : Softmax normalization on the matrix row; : matrix transpose operator; Temporal double embedding layer, which encodes the date in the NPZ file name through formula (2) and (3), and uses global MLP for feature adaptation;

[0032] (2)

[0033] (3)

[0034] where, : position encoding; : mapping time to multi-dimensional space by using sine and cosine waves of different frequencies; : current diffusion step or physical time signal (e.g. seasonal information); : total dimension of embedding vectors; : dimension index; : cardinality of control wavelengths.

[0035] As a preferred technical solution, in step three, texture enhancement and large-scale reconstruction optimization are performed by Monte Carlo integrated sampling and dynamic boundary fusion algorithm; Monte Carlo integrated sampling uses different random seeds to perform independent denoising sampling processes, obtaining a set of sample distributions about the reconstructed image ; Weighted mean processing is performed on the generated multiple samples:

[0036]

[0037] wherein, : final reconstructed sea surface temperature field; : number of samplings; : the independent denoising generated sample; : the weighted weight of each sample.

[0038] As a preferred technical solution, the dynamic boundary fusion algorithm determines the spatial topological position of the reconstruction block in real time through the adaptive boundary anchoring algorithm, performs forced correction on the weight operator that fits the global physical boundary, and sets its weight value to 1.0. According to the intersection gradient of the reconstruction region and the known observation region, the fusion width is gradually shrunk in the denoising iteration process; Fusion formula:

[0039]

[0040] wherein, : fused pixel value; : spatial weight mask (Mask), the weight approaches 0 at the edge and approaches 1 at the center; : reconstruction result of the current cut block; : the overlapping area result of adjacent cut blocks.

[0041] The application provides a satellite sea surface temperature (SST) missing data reconstruction method based on a space-time constraint diffusion model, which generates and reconstructs missing data from the mechanism of SST dynamic residual distribution and historical texture priori, and significantly improves the texture restoration degree and physical consistency accuracy of satellite SST under continuous large-scale cloud shielding compared with prior art. BRIEF DESCRIPTION OF DRAWINGS

[0042] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced as follows. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0043] Figure 1 The flowchart of the satellite sea surface temperature reconstruction method based on the space-time constraint diffusion model of the present application; Figure 2 The improved Unet architecture diagram; Figure 3 The training loss convergence analysis diagram; Figure 4 The average performance scene diagrams under different shielding scenes; Figure 5 The actual existing broken cloud shielding diagram in the South China Sea satellite observation; Figure 6 The average power spectral density (PSD) curve diagram; Figure 7 The South China Sea region Himawari true value diagram; Figure 8 The reconstructed SST diagram in the South China Sea region; Figure 9 The Indian Ocean sea area Himawari true value diagram; Figure 10 The reconstructed SST diagram in the Indian Ocean sea area. DETAILED DESCRIPTION

[0044] In order to make the purpose, technical solutions and advantages of the embodiments of the present application more clear, the technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0045] In addition, the term "and / or" in this document is only a description of the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B can represent the existence of A alone, the existence of A and B at the same time, and the existence of B alone. In addition, the character " / " in this document generally represents an "or" relationship between the associated objects unless otherwise specified.

[0046] The embodiments of the present application will be further described in detail below with reference to the accompanying drawings.

[0047] Referring to Figure 1 The present application provides a satellite sea surface temperature reconstruction method based on a space-time constraint diffusion model, comprising the following steps: Step 1, obtain Himawari L3 satellite data with missing data and OSTIA L4 global data in the global sea area.

[0048] The missing Himawari L3 satellite data is satellite data obscured by clouds, i.e. real satellite data.

[0049] The OSTIA L4 global data is complete and does not have missing data.

[0050] Step 2, pre-process the missing Himawari L3 satellite data and the OSTIA L4 global data to obtain a plurality of slice data.

[0051] Step 21, perform bicubic interpolation on the OSTIA L4 global data to enlarge it to the same precision as the Himawari L3 satellite data, so that the Himawari L3 satellite data and the OSTIA L4 global data are aligned in longitude and latitude.

[0052] Step 22, slice the corresponding sea area range with a sliding window of 128x128 pixels and a step size of 64 pixels to obtain a plurality of uniform slice data, save it as an NPZ file, and name it with the corresponding date.

[0053] The step size is set to 64 pixels, i.e. there is 50% area overlap in the horizontal and vertical directions between adjacent slices.

[0054] Specifically, the processed paired slices (Himawari L3 satellite data and OSTIA L4 global data) are packaged into independent NPZ files for efficient loading by PyTorch DataLoader.

[0055] Step three, input the pre-processed slice data into the trained spatiotemporal constraint diffusion model, and the spatiotemporal constraint diffusion model reconstructs the slice data through the weight file; the reconstructed slice data is subjected to texture enhancement and large-scale reconstruction optimization to obtain complete reconstruction data of the entire sea area. Further, the Himawari L3 satellite data is reconstructed.

[0056] In step three, the construction and training process of the spatiotemporal constraint diffusion model is as follows: Step 31, multi-source data acquisition; Step 32: slice processing of multi-source data; Step 33: construction and training of the spatiotemporal constraint diffusion model, and obtaining of the weight file; the weight file is used for reconstructing the missing Himawari L3 satellite data.

[0057] In step 31, the multi-source data includes Himawari L3 satellite data (Target Data) and OSTIA L4 global data (Background Data).

[0058] Among them, the Himawari L3 satellite data has a spatial resolution of up to 0.02° (about 2 km), which can sensitively capture the mesoscale features and small temperature gradients of the ocean surface layer. It is obtained from the Himawari third-generation geostationary meteorological satellite released by the Japan Meteorological Agency (JMA).

[0059] OSTIA L4 global data integrates multiple observation sources such as infrared, microwave, ships, and buoys through the optimal interpolation (OI) technique, providing daily global seamless sea surface temperature fields with a spatial resolution of 0.05°. It is obtained from the OSTIA (Operational Sea Surface Temperature and Sea Ice Analysis) L4 multi-source fusion product released by the UK Meteorological Office (MetOffice).

[0060] Both the Himawari L3 satellite data and the OSTIA L4 global data use 85% of the total data obtained as the training set of the spatiotemporal constraint diffusion model to participate in the training. The OSTIA L4 global data serves as the "spatial constraint background" in the reconstruction process, and through the dual-channel (Himawari L3 satellite data and OSTIA L4 global data) input method, it guides the spatiotemporal constraint diffusion model to restore the basic field that conforms to the large-scale physical law in the missing area.

[0061] The South China Sea and its surrounding sea areas are selected as the core test field for algorithm verification. The time span is from January 1, 2020 to December 31, 2022, and the specific geographical range covers latitude 4.0°N~24.48°N, longitude 106.0°E~123.92°E. The sea area has complex terrain and is distributed with rich mesoscale eddies, Kuroshio branches and their edge front activities. The complex texture generated by this highly dynamic dynamic environment is the best scene for testing whether the diffusion model can restore high-frequency details while maintaining original accuracy.

[0062] In step 32, the obtained multi-source data is finely sliced and preprocessed to adapt to the calculation requirements of the diffusion model for high-dimensional tensor input. The operation steps of the slicing preprocessing are as follows:

[0063] In step 321, the sea area range corresponding to the multi-source data is selected.

[0064] The South China Sea and its surrounding sea areas are selected, and the image size is set to 1024x896 pixels to match the spatial resolution of 1km.

[0065] In step 322, the OSTIA L4 global data is bi-cubic interpolated to be enlarged to the same accuracy as the Himawari L3 satellite data, so that the longitude and latitude of the Himawari L3 satellite data and the OSTIA L4 global data are aligned.

[0066] Because the resolution of the OSTIA L4 global data ( ) and the Himawari L3 satellite data ( ) is inconsistent, the bi-cubic interpolation technology is adopted to upsample the background source slice to the same dimension as the target source, i.e. 128x128, to ensure the spatial correlation of the dual-channel input.

[0067] The latitude and longitude axes are rearranged (latitude in descending order, longitude in ascending order) to ensure that the physical space direction of all slices is strictly consistent.

[0068] In order to let the model learn to use the OSTIA L4 global data as a constraint to reconstruct the Himawari texture, pixel-level alignment of the two must be achieved.

[0069] In step 323, the Himawari L3 satellite data after longitude and latitude alignment is cloud filtered, and the effective pixels are ensured to match the sea area range (no land) of the OSTIA L4 global data.

[0070] ​The process of cloud filtering for the Himawari L3 satellite data after aligning the latitude and longitude includes temperature scale conversion and outlier filtering.

[0071] Temperature scale conversion: Determine the initial temperature scale by detecting the average pixel value. Determine the initial temperature scale according to the Himawari L3 satellite data and OSTIA L4 global data, and if it is in Kelvin, convert it to Celsius. For example, if the average temperature is detected to be greater than , it is determined to be in Kelvin and automatically subtracts , and is converted to Celsius (°C).

[0072] Outlier filtering: Identify and remove outliers outside the reasonable sea temperature range (−2.0°C to 45.0°C), and mark them as NaN.

[0073] Specifically, to ensure the convergence of model training, land mask removal and 100% cloud-free constraint are used for filtering.

[0074] Land mask removal: Use the completeness of OSTIA L4 global data. If there are any NaN values (representing land or no data area) in the slice, discard the sample directly.

[0075] 100% cloud-free constraint (for Target Data): As the "ground truth" (GT) of the spatiotemporal constraint diffusion model, the Himawari L3 satellite data training slice must be completely cloud-free. By checking whether there are missing values in the slice, high-quality observation samples are selected to ensure that the spatiotemporal constraint diffusion model learns the true sea texture rather than cloud artifacts.

[0076] Step 324: Slice the 128x128 pixel sliding window for the corresponding sea area range to obtain multiple uniform slices, save them as NPZ files, and name them according to the corresponding date.

[0077] Specifically, the processed paired slices (Himawari L3 satellite data and OSTIA L4 global data (background data)) are packaged into independent NPZ files to facilitate efficient loading by PyTorch DataLoader.

[0078] Use 128x128 pixel sliding window slicing to ensure that each small block maintains local texture correlation while reducing memory pressure.

[0079] Total number of samples: A total of 22093 effective samples without land interference were screened out.

[0080] Dataset matching: The data from 2020 to 2022 is divided into 85% training set and 15% validation set for model learning.

[0081] Cross-year testing: Himawari L3 satellite data and corresponding OSTIA L4 global data for the whole year of 2023 are selected independently as the test set to evaluate the long-term robustness of the model under real climate variability.

[0082] In the slice preprocessing stage, the all_reduce operator is used to synchronize the statistical information across processes. The SST temperature range of the full sample is calculated to provide accurate parameter support for the normalization of the input layer of the subsequent spatiotemporal constraint diffusion model.

[0083] In this step 32, given the large amount of satellite data (covering multi-year daily data), a parallel preprocessing scheme based on PyTorchDistributed is used to significantly improve computational efficiency. The core of multi-source data preprocessing is to ensure that each slice learned by the model has high scientific value through physical cleaning, spatial alignment, and strict filtering.

[0084] Step 33: Build and train the spatiotemporal constraint diffusion model.

[0085] The spatiotemporal constraint diffusion model construction process is as follows: Step 331, first linearly normalize the multi-source data, then perform data augmentation. Ensure the stability of the diffusion process in mathematics.

[0086] Linear normalization: map the original sea surface temperature value (8.0°C ~ 34.0°C) to the interval.

[0087]

[0088] where, : normalized multi-source data value; : original multi-source data physical value (in Celsius); : lower limit value of multi-source data, take 8.0; : upper limit value of multi-source data, take 34.0.

[0089] Data augmentation: perform random horizontal and vertical flipping on the slice to improve the generalization ability of the spatiotemporal constraint diffusion model under different ocean current directions.

[0090] Step 332, extract the multi-dimensional physical time feature vector corresponding to the date in the slice file name , so that the spatiotemporal constraint diffusion model has the ability to process the seasonal fluctuations and interannual evolution of sea surface temperature (SST); wherein the day of year (doy) within a year is mapped to a two-dimensional unit circle coordinate space using sine and cosine transformation, which effectively solves the contradiction in traditional linear time representation that the numerical distance between the end of December (365 days) and the beginning of January (1 day) is far away, but in physical season it is adjacent, ensuring the smooth continuity of the seasonal cycle characteristics of SST in the feature space; at the same time, the year offset relative to the 2020 base year is introduced as the third dimension feature, so that the model has the ability to capture the climate warming trend or abnormal evolution of a specific year.

[0091] Step 333, forward noise is added to the Himawari L3 satellite data to obtain the Himawari L3 satellite data to be denoised.

[0092] Forward noise addition: cosine noise schedule (Cosine Noise Schedule) is used for noise addition, which can more evenly distribute the decay process of image information compared with linear schedule.

[0093] For a given true value image of Himawari L3 satellite data , the noise addition result at the th step is directly calculated as:

[0094] , wherein : the true value image of the normalized Himawari L3 satellite data; : : the noisy latent variable generated at the moment; : represents a Gaussian distribution (normal distribution); : noise scaling factor defined by cosine schedule; is the cumulative result thereof.

[0095] : identity matrix, representing standard normal distribution noise.

[0096] Step 334, concatenate the Himawari L3 satellite data to be denoised and the OSTIA L4 global data as the background field constraint condition in the channel dimension to perform denoising and predict noise distribution.

[0097] The Himawari L3 satellite data to be denoised is set as , and the OSTIA L4 global data is set as . The noise is predicted from the formula (3-1) :

[0098]

[0099] wherein, : the probability distribution of the improved UNet model learning the reverse generation process (denoising process); : the reconstructed image latent variable at the previous time of the denoising process; : the diffusion step number (noisy moment). It represents a specific stage of injecting noise into the original image

[0100] : the noisy image latent variable at the current moment; : the full coverage background field constraint provided by the OSTIA data; : the physical time feature vector containing the date (doy) and year offset; : representing a Gaussian distribution (normal distribution); : the Gaussian distribution mean value predicted by the improved UNet model, representing the expected value of the reconstructed image; : the Gaussian distribution variance predicted by the improved UNet model, representing the uncertainty of the reconstruction process.

[0101] In step 334, denoising uses an improved UNet architecture, as shown in Figure 2 . The improved UNet architecture predicts the noise distribution.

[0102] The improved UNet architecture includes ConvNeXt V2 modules, linear attention modules with Memory KV, and time dual embedding layers.

[0103] ​In the noise prediction distribution stage of the diffusion model, the improved UNet architecture uses multi-dimensional time encoding (TE) as the core modulator to coordinate the ConvNeXt V2 module and the linear attention module with Memory KV to achieve a "space-time-object" trinity collaborative processing. Specifically, the seasonal and interannual features extracted by TE are injected into the ConvNeXt V2 module through a parameter modulation layer, guiding it to dynamically adjust the response weight of local sub-mesoscale details according to the current physical background (such as the dynamic characteristics of a specific season); at the same time, TE provides a physical index for the linear attention module with Memory KV, enabling it to accurately retrieve the global spatial constraints that conform to the current evolution trend from the long-term sea temperature distribution rules stored in the linear attention module with Memory KV. The deep coupling of local convolution detail capture, global attention association modeling, and macroscopic time prior in the improved UNet architecture ensures that the spatiotemporal constraint diffusion model can predict physically accurate and spatially coherent noise residuals at each step t of the reverse denoising process , thereby achieving high-fidelity SST field reconstruction under extreme observation loss.

[0104] The ConvNeXt V2 module uses large kernel depth convolution to replace the standard convolution layer residual module in the prior art, thereby significantly expanding the receptive field and better capturing mesoscale ocean phenomena.

[0105] To enhance feature competition between channels and improve the texture sharpness of sea surface temperature edges, the Global Response Norm (GRN) mechanism is used in each ResBlock:

[0106]

[0107]

[0108]

[0109] wherein, : input feature map, dimension ; : the norm of the feature map in the spatial dimension , representing the channel intensity; : normalized response value, reflecting the competitive position between channels; : total number of channels; : a small constant to prevent division by zero; : element-wise multiplication; : learnable scaling and translation parameters, where, and are initialized to 0, so that the module behaves as an identity mapping at the beginning of training; Linear attention module with Memory KV, which reduces the computational complexity from to .

[0110] (1)

[0111] where Memory KV is added to learn the global physical prior:

[0112]

[0113] Letter meaning: : Query, representing the pixel feature that needs to be reconstructed; : Key and Value, representing the spatial context; : Memory KV, i.e., the pre-set learnable static parameters in the model, used to store the global sea temperature distribution prior; : Softmax normalization on matrix rows; : Softmax normalization on matrix rows; : Matrix transpose operator.

[0114] Linear attention module with Memory KV, which significantly reduces the computational complexity of training 128x128 resolution sea temperature slices, making the inference speed increase by about 50%. Shortens the training time and improves the training speed.

[0115] Temporal double embedding layer, which encodes the date in the NPZ file name obtained in step 324 through formulas (2) and (3), so that the spatiotemporal constraint diffusion model learns seasonal changes rather than fixed time, and uses global MLP for feature adaptation.

[0116] (2)

[0117] (3)

[0118] wherein, : position encoding; : mapping time into a multi-dimensional space by using sine and cosine waves of different frequencies; : current diffusion step or physical time signal (e.g. seasonal information); : total dimension of embedding vectors; : dimension index; : cardinality of control wavelength.

[0119] Step 335, by training the difference between the predicted noise and the real noise, the real pixels and textures in the Himawari L3 satellite data are identified, and the spatio-temporal constraint diffusion model is constructed.

[0120] The identification of the real pixels and textures in the Himawari L3 satellite data is realized by the loss function formula (3-2):

[0121]

[0122] wherein, : training loss value; : represents the joint distribution of the training data , the random noise sampled from the standard normal distribution , and the randomly sampled diffusion step expectation value of the joint distribution; : real noise sampled from the standard normal distribution; : noise residual predicted by the improved UNet; : noisy image at time t; : OSTIA background condition guidance; time: embedding vector integrating diffusion step and seasonal information.

[0123] Step 336, repeat steps 331 to 335, iteratively train the spatio-temporal constraint diffusion model, and constantly correct the model parameters by using the gradient descent algorithm until the loss function on the validation set reaches order of magnitude and remains stable, as shown in Figure 3 , that is, stop training and obtain the trained spatio-temporal constraint diffusion model; get the weight file.

[0124] The hyperparameter settings for training the spatiotemporal constraint diffusion model: the AdamW optimizer is used, and the initial learning rate is set to , and the ReduceLROnPlateau scheduler is used, and when the validation set Loss does not decrease for 15 epochs, the learning rate is halved.

[0125] Through this training process, the model establishes a nonlinear mapping relationship for pixel completion and texture restoration of Himawari L3 satellite data using OSTIA L4 global data and multi-dimensional time features, thereby realizing spatiotemporal continuous high-precision sea surface temperature reconstruction.

[0126] In step three, texture enhancement and large-scale reconstruction optimization are performed through Monte Carlo ensemble sampling and dynamic boundary fusion algorithm. The complete reconstruction data of the entire sea area is more consistent with the real texture, and the complete reconstruction of the South China Sea can be realized.

[0127] Monte Carlo ensemble sampling, at the numerical level, suppresses high-frequency random noise through multiple sample means, significantly reducing the root mean square error (RMSE) between the reconstructed field and the Himawari L3 satellite data; at the physical level, it retains the robust dynamic characteristics of cross-path coexistence, making the final generated SST field more spatially smooth and consistent with the real ocean thermodynamic evolution law. Through Monte Carlo ensemble sampling, the transition from "probability generation" to "steady-state reconstruction" is realized, laying a statistical foundation for high-precision engineering application of sea surface temperature field in the entire sea area.

[0128] The operation steps are as follows:

[0129] In the inference phase, use different random seeds to perform independent denoising sampling processes.

[0130] Obtain a set of sample distributions about the reconstructed image .

[0131] Perform weighted mean processing on the generated multiple samples:

[0132]

[0133] , where : the final reconstructed sea surface temperature field; : the number of samples; : the th independently generated sample; : the weighted weight of each sample.

[0134] Monte Carlo integration sampling solves the problem of random artifacts or abnormal spots that may be generated by single denoising, can enhance stable ocean feature texture (such as constant vortex edge), and offset the uncertainty spots generated in single sampling, thereby significantly reducing the reconstruction error.

[0135] The dynamic boundary fusion algorithm can ensure that the edge region can still maintain the original physical value order of magnitude of the reconstructed field under the condition of missing adjacent overlapping compensation, so as to eliminate the non-physical edge value decay phenomenon at the mathematical level, and ensure the physical consistency of the reconstructed results of the whole field.

[0136] The dynamic boundary fusion algorithm gradually shrinks the fusion width in the denoising iteration process according to the gradient at the junction of the reconstructed region and the known observation region.

[0137] The fusion formula is:

[0138]

[0139] Wherein, : the pixel value after fusion; : spatial weight mask (Mask), the weight tends to 0 at the edge, and tends to 1 at the center; : the reconstruction result of the current cut block; : the overlapping region result of the adjacent cut block; The dynamic boundary fusion algorithm can effectively smooth the sampling variance between different slices, so that the temperature gradient at the splicing place is naturally transitioned, and the 'grid effect' generated by the traditional block method is eliminated.

[0140] For the reconstruction of the whole region of the South China Sea, the sea area needs to be cut into small blocks of 128*128 pixels for reconstruction and splicing. In order to eliminate the'seam' or 'fault' phenomenon at the splicing place, the dynamic boundary fusion algorithm is adopted in the embodiment instead of 'hard cutting'. Even if a super large image spanning thousands of kilometers is generated, the edge part can still maintain a smooth texture transition.

[0141] The present application takes the South China Sea as an example for experiment.

[0142] In order to comprehensively evaluate the robustness of the model, the present application carries out stress test through four typical shielding scenes: Scene one: whole image random reconstruction: simulate the condition that the whole image is completely shielded by cloud layer, and only rely on OSTIA background field and time encoder for reconstruction.

[0143] Scenario 2: Half-mask reconstruction: Simulate large-area continuous cloud cover (lower half mask), the model needs to refer to the effective observation information of the upper half and the OSTIA conditions for joint inference.

[0144] Scenario 3: Reconstruction of a missing central quarter: Simulate localized cloud occlusion (64x64 pixel mask) to verify the model's ability to recover fine local features.

[0145] Scenario 4: A scenario where Himawari L3 satellite data is fragmented and lacks true data, focusing on evaluating the model's generalization performance in real-world operational scenarios. This embodiment uses the following core metrics to analyze the deviation between the reconstructed results and the true values:

[0146] RMSE (Root Mean Square Error): Measures the absolute temperature deviation between the reconstructed value and the true value.

[0147] SSIM (Structural Similarity): Evaluates the degree of similarity between a reconstructed image and the ground truth in terms of brightness, contrast, and structure.

[0148] PSNR (Peak Signal-to-Noise Ratio): The signal-to-noise ratio level that reflects the quality of image reconstruction; the higher the value, the less distortion.

[0149] Himawari Truth Value: This refers to the raw data from the Himawari L3 satellite used as comparison data.

[0150] Average performance under different occlusion scenarios: Scenario 1, 2, and 3, such as... Figure 4 Scene Four Figure 5 As shown.

[0151] Figure 4 The reconstruction performance under three typical idealized masking conditions intuitively reflects the model's deep understanding of the spatial logic of the sea surface temperature field. In the central 64×64 region mask, the reconstruction result successfully captured and restored the sea surface temperature front structure spanning the central sea area, with clear edge features and a temperature gradient distribution highly aligned with the ground truth. Compared to the highly smooth OSTIA L4 global data, which loses sub-mesoscale dynamic features, this invention can effectively supplement the lost high-frequency fluctuation information using a diffusion-generated reverse evolution mechanism. In the experiment with the lower 50% mask, the model exhibited excellent spatial continuity, producing no visual breaks at the y=64 pixel dividing boundary. In the most challenging 100% full-field missing scenario, guided by the periodic sine and cosine priors provided by multidimensional time encoding (TE), the model generated a sea surface temperature field structure with rich texture features on the blurred OSTIA L4 global data. The generated vortex morphology and cold tongue distribution maintained significant consistency with the Himawari L3 satellite data in spatial topology.

[0152] Figure 5 The actual existence of broken cloud blocking in satellite observations of the South China Sea is demonstrated, with a focus on evaluating the model's ability to integrate and preserve the fragmented observation signals. When satellite observations are extremely sparse and exhibit irregular fragmentation, the invention exhibits strong robustness, enabling accurate extraction of thermodynamic features in scattered points and high-quality filling combined with global spatial constraints provided by OSTIA L4 global data. The reconstruction results not only maintain consistency with the true values in macroscopic distribution, but also successfully restore the temperature step features of mesoscale vortex edges.

[0153] To verify whether the reconstructed field retains the original high-frequency texture details of Himawari, the invention uses power spectral density (PSD) for analysis.

[0154] The PSD curve shows the distribution of energy at different spatial scales (wavelengths) as shown in Figure 6 The black line (Himawari true value) represents the high-precision true structure.

[0155] The blue line (OSTIA) significantly lacks energy at small scales (high-frequency region), and the curve is smooth.

[0156] The red line (reconstructed SST) closely matches the black line across the entire wavelength range, especially in the small scale region (high-frequency region) of km), successfully recovering the high-frequency details lost by OSTIA.

[0157] In summary, the closer the red line is to the black line, the more the sub-mesoscale features of the reconstructed image conform to the true physical field.

[0158] Full reconstruction of the South China Sea region: The Himawari true value in the South China Sea region is shown in Figure 7 , and the reconstructed SST in the South China Sea region is shown in Figure 8 .

[0159] Land boundary filling: For the complex coastline and island regions of the South China Sea, the invention uses a land filling algorithm based on physical distance. Since the neural network cannot directly process NaN missing values on land, this algorithm uses Euclidean distance transformation to replace the invalid pixels affected by land in the slice with the nearest valid observation values in the geographical location. This operation ensures the numerical stability of the convolution operator at the land-sea boundary, avoiding edge collapse phenomena.

[0160] Visualization effect: The reconstructed SST field not only perfectly fills the gaps left by clouds (White Cloud Area), but also eliminates the seams of block stitching through a dynamic boundary fusion algorithm.

[0161] ​Indian Ocean migration experiment demonstration: In order to further verify the spatial robustness and migration ability of the present application in different geographical dynamic environments, the model weight based on the South China Sea sea area training is directly applied to the Indian Ocean sea area to carry out generalization experiment. The core purpose of this step is to explore whether the present application has really learned the general physical characteristics of sea surface temperature reconstruction, and its reliability when performing "zero sample" migration task in unvisited geographical area.

[0162] Indian Ocean sea area Himawari true value as Figure 9 shown, the reconstructed SST in the Indian Ocean sea area is as Figure 10 shown.

[0163] With the high-frequency texture generation ability learned in the South China Sea, the present application successfully analyzes the natural sea temperature fluctuations in these "blind areas". Although the model has not learned the Indian Ocean samples, the reconstruction result accurately restores the vortex topology inside the Bay of Bengal and the frontal feature along the coast of the Arabian Sea. It proves that the time and space constraint diffusion model based on the present application captures the general physical mapping logic of "restoring high-frequency dynamic details from background field residuals", rather than just memorizing the South China Sea geographical coordinates, which embodies strong spatial robustness.

[0164] The present application tests the full map randomly, half mask and local missing three scenes, and the present application realizes high-fidelity restoration of sub-mesoscale features in the South China Sea area. In the completely random reconstruction scene, the average RMSE is 0.3358°C, the SSIM is 0.9718, and the PSNR is 38.3dB. Power spectral density (PSD) analysis shows that the reconstructed curve (red) is most consistent with the Himawari true value (black) in the full wavelength range, especially in the high-frequency band of 10−−30 km, effectively repairing the energy collapse phenomenon of the OSTIA L4 global data (blue) due to excessive smoothing. This shows that the present application not only approaches the true value in the pixel level, but also restores the energy cascade characteristics of ocean turbulence in the physical energy spectrum level, proving that the reconstructed field has real ocean dynamics mechanism. Through a large number of comparative experiments and field data verification, it is proved that the time and space constraint diffusion model based on the present application has excellent performance in solving the problem of "high-precision data missing", and the reconstruction product has high physical reliability and dynamic fidelity, which has significant scientific value and engineering application prospect.

[0165] The present application provides a satellite sea surface temperature missing reconstruction method based on a time and space constraint diffusion model, which generates missing data by reconstructing from the mechanism of SST dynamic residual distribution and historical texture prior. Although this method is limited by the effectiveness of the historical data in the previous day in extreme cases, compared with the prior art, it significantly improves the texture restoration degree and physical consistency accuracy of satellite SST under continuous large-scale cloud cover.

[0166] Of course, the present application can have other various embodiments, and it will be understood that various omissions and substitutions of equivalents are contemplated as circumstances can suggest or render expedient, but such omissions and substitutions are to be considered within the scope of the application.

Claims

1. A method for reconstructing sea surface temperature based on a spatiotemporal constrained diffusion model, characterized in that, The method comprises the following steps: Step one, obtaining Himawari L3 satellite data with missing data and OSTIA L4 global data in the global sea area; Step two, preprocessing the missing Himawari L3 satellite data and the OSTIA L4 global data to obtain a plurality of slice data; Step three, inputting the plurality of preprocessed slice data into the trained spatio-temporal constraint diffusion model, and reconstructing the plurality of slice data through the weight file; and performing texture enhancement and large-scale reconstruction optimization on the reconstructed plurality of slice data to obtain complete reconstruction data of the global sea area.

2. The method of claim 1, wherein the method comprises: In step two, the preprocessing steps of the missing Himawari L3 satellite data and the OSTIA L4 global data are as follows: Step 21, performing bicubic interpolation on the OSTIA L4 global data, and enlarging to the same precision as the Himawari L3 satellite data to align the longitude and latitude of the Himawari L3 satellite data and the OSTIA L4 global data; Step 22, performing 128x128 pixel, step 64 pixel sliding window slicing on the corresponding sea area range to obtain a plurality of uniform slice data, saving as NPZ files, and naming according to the corresponding date; The step size is set to 64 pixels, that is, there is 50% area overlap between adjacent slices in the horizontal and vertical directions.

3. The method of claim 1, wherein the method further comprises: In step three, the construction and training process of the spatio-temporal constraint diffusion model is as follows: Step 31, obtaining multi-source data; Step 32: slicing the multi-source data; Step 33: constructing the spatio-temporal constraint diffusion model and training; Obtaining a weight file; The weight file is used for reconstructing the missing Himawari L3 satellite data.

4. The method of claim 3, wherein the method further comprises: In step 31, the multi-source data includes Himawari L3 satellite data and OSTIA L4 global data; wherein the spatial longitude and latitude resolution of the Himawari L3 satellite data is 0.02°; and the spatial longitude and latitude resolution of the OSTIA L4 global data is 0.05°.

5. The method of claim 3, wherein the method further comprises: In step 32, the operation steps of slicing preprocessing are as follows: Step 321, selecting the sea area range corresponding to the multi-source data; Selecting the sea area range, and setting the image size to 1024x896 pixels; Step 322, performing bicubic interpolation on the OSTIA L4 global data, and enlarging to the same precision as the Himawari L3 satellite data to align the longitude and latitude of the Himawari L3 satellite data and the OSTIA L4 global data; Step 323, performing cloud filtering on the Himawari L3 satellite data after aligning the longitude and latitude, and ensuring that the effective pixels thereof match the sea area range of the OSTIA L4 global data; The process of performing cloud filtering on the Himawari L3 satellite data after aligning the longitude and latitude includes temperature scale conversion and abnormal value filtering; Temperature scale conversion: determining the initial temperature scale by detecting the pixel mean value; Abnormal value filtering: identifying and removing abnormal values outside the reasonable sea temperature range, and uniformly marking them as NaN; Step 324, a sliding window slicing of 128x128 pixels is performed for the corresponding sea area range, obtaining a plurality of uniform slices, saved as NPZ files and named with the corresponding date.

6. The method of claim 3, wherein the method further comprises: In step 33, the spatio-temporal constraint diffusion model construction process is: In step 331, first linearly normalize the multi-source data, and then perform data enhancement; Linear normalization: The raw sea surface temperature values are mapped to Interval: wherein, : normalized multi-source data value; : raw multi-source data physical value; : lower limit of multi-source data, 8.0; : upper limit of multi-source data, take 34.0; Data enhancement: perform random horizontal and vertical flipping on the slices; Step 332, extract the multi-dimensional physical time feature vector corresponding to the date in the slice file name ; Step 333, forward noise is added to the Himawari L3 satellite data to obtain the Himawari L3 satellite data to be denoised; Noise addition using a cosine noise schedule: for a given ground truth image of Himawari L3 satellite data , at the th step of noise addition , computed as: wherein, : True value image of normalized Himawari L3 satellite data; : noisy latent variables generated at the time instant : denotes a Gaussian distribution; : noise scaling factor defined by the cosine progress table; the multiplication result thereof; : identity matrix, representing standard normal distributed noise; Step 334, perform concatenation operation in channel dimension between the Himawari L3 satellite data to be denoised and the OSTIA L4 global data as the background field constraint condition, denoising and predicting noise distribution at the same time; Himawari L3 level satellite data to be denoised is set as , OSTIA L4 level global data is set as , and is predicted from the noise by equation (3-1) : wherein, : a probability distribution of a reverse generation process of the improved UNet model learning; : reconstructed image latent variable at a previous time instant of the denoising process; : number of diffusion steps; : current noisy image latent variable at the current time instant; : full coverage background field constraints provided by the OSTIA data; : physical time feature vector comprising date and year offset; : represents a Gaussian distribution; : the mean of the Gaussian distribution predicted by the improved UNet model, representing the expected value of the reconstructed image; : Gaussian distribution variance predicted by the improved UNet model, representing the uncertainty of the reconstruction process; Step 335, by training, the difference between the predicted noise and the real noise is identified to identify the real pixels and textures in the Himawari L3 satellite data, and the spatio-temporal constraint diffusion model is constructed; The identification of real pixels and textures in the Himawari L3 satellite data is realized by loss function formula (3-2): 。 7. The method of claim 6, wherein the method further comprises: Steps 331 to 335 are repeated to iteratively train the spatiotemporal constraint diffusion model, and the model parameters are constantly corrected by using the gradient descent algorithm until the value of the loss function on the validation set reaches an order of magnitude and remains stable, i.e. the training is stopped and the trained spatiotemporal constraint diffusion model is obtained; a weight file is obtained.

8. The method of claim 6, wherein the method further comprises: In step 334, the denoising adopts an improved UNet architecture, which includes a ConvNeXt V2 module, a linear attention module with Memory KV, and a time double embedding layer. The ConvNeXt V2 module adopts a global response normalization GRN mechanism in each ResBlock: wherein, : input feature map, dimension ; : feature map's spatial dimension norm, representing channel intensity; : normalized response value, reflecting the competitive position between channels; : total number of channels; : small constant to prevent division by zero; : element-wise multiplication; : learnable scaling and translation parameters, where, and are initialized to 0 so that the module behaves as an identity mapping at the beginning of training; Linear attention module with Memory KV, feature extraction is performed by formula (1), the complexity is reduced from to ; （1） Wherein, the Memory KV is added to learn the global physical prior: wherein, : query matrix, representing the pixel feature currently requiring reconstruction; : key matrix and value matrix, representing spatial context; : Memory KV, i.e. the preset learnable static parameters in the model, is used to store the global sea temperature distribution prior; : Softmax normalization on the matrix columns; Softmax normalization on matrix rows; : matrix transpose operator; The time double embedding layer encodes the date in the NPZ file name through formulas (2) and (3), and adapts the features using a global MLP; （2） （3） wherein : position encoding; : mapping time to a multi-dimensional space by using sine and cosine waves of different frequencies; : current diffusion step or physical time signal; : total dimension of embedding vectors; : dimension index; : base of the control wavelength.

9. The method of claim 1, wherein the method further comprises: In step three, texture enhancement and large-scale reconstruction optimization are performed through Monte Carlo ensemble sampling and dynamic boundary fusion algorithm; Monte Carlo integration sampling, using a different random seed to perform an independent denoising sampling process, obtaining a set of samples distributed about the reconstructed image ; The generated multiple samples are processed by weighted mean: wherein, : final reconstructed sea surface temperature field; : number of samples; : first sample generated by the first independent denoising : Weighting weight of each sample.

10. The method of claim 9, wherein the method further comprises: The dynamic boundary fusion algorithm uses an adaptive boundary anchoring algorithm to determine the spatial topological position of the reconstruction block in real time, and performs forced correction on the weight operator that fits the global physical boundary, setting its weight value to 1.

0. According to the intersection gradient of the reconstruction area and the known observation area, the fusion width is gradually shrunk during the denoising iteration process. Fusion formula: wherein : fused pixel value; : spatial weight mask, weights approach 0 at edges, 1 at center; : reconstruction result of current tile; : Overlap area result of adjacent tiles.

Images