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

Abstract

This invention relates to the field of satellite remote sensing image processing and meteorological big data analysis technology, specifically a satellite sea surface temperature (SST) reconstruction method based on a spatiotemporally constrained diffusion model. First, it acquires missing Himawari L3 satellite data and OSTIA L4 global data covering a global sea area. Then, it preprocesses the missing Himawari L3 and OSTIA L4 data to obtain several slice data points. Finally, it inputs the preprocessed slice data into a trained spatiotemporally constrained diffusion model, which reconstructs the slice data using a weight file. The reconstructed slice data undergoes texture enhancement and large-scale reconstruction optimization to obtain complete reconstructed data covering the entire sea area. This invention significantly improves the texture restoration accuracy and physical consistency of satellite SST under continuous, large-scale cloud cover.

Description

A satellite sea surface temperature reconstruction method based on a spatiotemporal constrained diffusion model Technical Field

[0001] This invention relates to the fields of satellite remote sensing image processing and meteorological big data analysis, and in particular to a satellite sea surface temperature reconstruction method based on a spatiotemporal constrained diffusion model. Background Technology

[0002] Sea surface temperature (SST) is a key physical parameter characterizing ocean thermal conditions and air-sea energy exchange, playing a crucial role in global climate change monitoring and ocean numerical forecasting. Currently, next-generation geostationary meteorological satellites, such as the Japan Meteorological Agency's Himawari, can provide high spatiotemporal resolution Level 3 skin-level SST products. Leveraging the observational advantages of their infrared sensors, this type of Level 3 data boasts extremely high observational accuracy, clearly capturing intricate dynamic textures such as mesoscale eddies and ocean fronts.

[0003] However, infrared remote sensing technology is not only limited by cloud cover but also susceptible to interference from aerosol scattering and heavy precipitation, resulting in large-area, non-random data gaps in L3 SST products during practical applications. To obtain spatiotemporally continuous, fully covered SST fields, existing operational methods typically employ L4 (Level 4) multi-source fusion analysis products (such as the OSTIA system developed by the UK Met Office). L4 data, through statistical methods such as optimal interpolation (OI), fuses data from multiple satellites and in-situ buoys, filling in the gaps below the clouds.

[0004] Although Level 4 products achieve full spatial coverage of data, the statistical interpolation algorithm they employ is essentially a smoothing estimate, tending to preserve low-frequency climate background information while severely suppressing high-frequency change signals. This results in Level 4 data being "complete" but with lower accuracy, often losing the sharp texture and instantaneous dynamic details originally present in Level 3 data. Under complex weather processes such as typhoons and frontal transitions, the smoothed fields provided by Level 4 data cannot accurately reflect the dramatic changes in the ocean surface.

[0005] To address this issue, existing deep learning reconstruction methods (such as those based on convolutional neural networks or conventional generative adversarial networks) attempt to repair missing regions, but most fail to effectively utilize the historical high-precision observations of L3-level data as prior guidance, and also lack reasonable utilization of the low-frequency constraints of the L4-level background field. When faced with large-scale missing L3-level data, existing technologies often fall into a dilemma: either the reconstruction result is too smooth (degrading to L4 level), or the generated texture is messy and does not conform to physical laws. Therefore, how to combine the high-precision texture advantage of L3-level data with the full-coverage background advantage of L4-level data to achieve high-precision reconstruction while ensuring physical realism is a key technical challenge that urgently needs to be solved. Summary of the Invention

[0006] This invention provides a satellite sea surface temperature (SST) reconstruction method based on a spatiotemporal constrained diffusion model, which can reconstruct SST data missing from satellite cloud layers while maintaining the original accuracy and texture. This solves the problem that existing technologies mainly use spatial statistical interpolation to forcibly fill in the missing data and use the weighted average of surrounding effective observation points to characterize the sea surface temperature of the missing area, resulting in the smoothing and erasure of high-frequency texture details.

[0007] This invention provides a satellite sea surface temperature reconstruction method based on a spatiotemporal constrained diffusion model, comprising the following steps:

[0008] Step 1: Acquire missing Himawari L3 satellite data and OSTIA L4 global data for global sea areas;

[0009] Step 2: Preprocess the missing Himawari L3 satellite data and OSTIA L4 global data to obtain several slice data.

[0010] Step 3: Input the preprocessed slice data into the trained spatiotemporal constrained diffusion model. The spatiotemporal constrained diffusion model reconstructs the slice data using a weight file. The reconstructed slice data is then subjected to texture enhancement and large-scale reconstruction optimization to obtain complete reconstructed data for the entire sea area.

[0011] As a preferred technical solution, the preprocessing steps for the missing Himawari L3 satellite data and OSTIA L4 global data in step two are as follows:

[0012] Step 21: Perform bicubic interpolation on the OSTIA L4 global data to magnify 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.

[0013] Step 22: For the corresponding sea area, perform a sliding window tiling of 128×128 pixels with a step size of 64 pixels to obtain several uniform tiling data, save them as NPZ files, and name them with the corresponding date.

[0014] Set the step size to 64 pixels, meaning that there is a 50% overlap between adjacent slices in both the horizontal and vertical directions.

[0015] As a preferred technical solution, the construction and training process of the spatiotemporal constraint diffusion model in step three is as follows:

[0016] Step 31, multi-source data acquisition;

[0017] Step 32: Slice the multi-source data;

[0018] Step 33: Construct and train a spatiotemporal constrained diffusion model; obtain a weight file; the weight file is used to reconstruct the missing Himawari L3 satellite data.

[0019] As a preferred technical solution, in step 31, the multi-source data includes Himawari L3 satellite data and OSTIA L4 global data; wherein, the spatial latitude and longitude resolution of Himawari L3 satellite data is 0.02°; and the spatial latitude and longitude resolution of OSTIA L4 global data is 0.05°.

[0020] As a preferred technical solution, the slice preprocessing steps in step 32 are as follows:

[0021] Step 321: Select the sea area corresponding to the multi-source data;

[0022] Select the sea area and set the image size to 1024×896 pixels.

[0023] Step 322: Perform bicubic interpolation on the OSTIA L4 global data to magnify 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.

[0024] Step 323: Perform cloud removal filtering on the Himawari L3 satellite data after aligning latitude and longitude, and ensure that its effective pixels match the sea area range of the OSTIA L4 global data;

[0025] The process of cloud removal filtering for Himawari L3 satellite data after latitude and longitude alignment includes temperature scale conversion and outlier filtering.

[0026] Temperature scale conversion: The initial temperature scale is determined by detecting the average value of pixels;

[0027] Outlier filtering: Identify and remove outliers that are outside the reasonable ocean temperature range, and uniformly label them as NaN;

[0028] Step 324: Perform a 128×128 pixel sliding window slice for the corresponding sea area to obtain multiple uniform slices, save them as NPZ files, and name them with the corresponding date.

[0029] As a preferred technical solution, the spatiotemporal constraint diffusion model construction process in step 33 is as follows:

[0030] Step 331: First, perform linear normalization on the multi-source data, and then perform data augmentation;

[0031] Linear normalization: Mapping the original sea surface temperature values ​​to... Interval:

[0032]

[0033] in, : Normalized multi-source data values;

[0034] : Raw multi-source data physical values ​​(degrees Celsius);

[0035] The lower limit for multi-source data is set to 8.0.

[0036] The upper limit for multi-source data is set to 34.0.

[0037] Data augmentation: Perform random horizontal and vertical flips on the slices;

[0038] Step 332: Extract the multidimensional physical time feature vector corresponding to the date in the slice file name. ;

[0039] Step 333: Forward noise addition is performed on the Himawari L3 satellite data to obtain the Himawari L3 satellite data to be denoised;

[0040] Noise addition using a cosine noise schedule: For a given ground truth image of Himawari L3 satellite data In the Noise addition result of each step The calculation is as follows:

[0041]

[0042] in, True-value image of normalized Himawari L3 satellite data;

[0043] : Noisy latent variables generated at any given time;

[0044] : indicates a Gaussian distribution (normal distribution);

[0045] : Noise scaling factor defined by the cosine schedule;

[0046] The result of multiplying them together;

[0047] : Identity matrix, representing standard normal distribution noise;

[0048] Step 334: The Himawari L3 satellite data to be denoised is concatenated with the OSTIA L4 global data as background field constraints in the channel dimension to perform denoising and predict the noise distribution.

[0049] The Himawari L3 satellite data to be denoised is set as follows: OSTIA L4 global data setting is Predict from noise using formula (3-1) :

[0050]

[0051] in, : Probability distribution of the reverse generation process learned by the improved UNet model;

[0052] : Latent variables of the reconstructed image at the moment preceding the denoising process;

[0053] : Number of diffusion steps;

[0054] :current Latent variables of the noisy image at any given time;

[0055] : Full-coverage background field constraints provided by OSTIA data;

[0056] : A physical time feature vector containing date and year offsets;

[0057] : Indicates a Gaussian distribution;

[0058] The mean of the Gaussian distribution predicted by the improved UNet model represents the expected value of the reconstructed image.

[0059] The Gaussian variance predicted by the improved UNet model represents the uncertainty in the reconstruction process;

[0060] Step 335: By training and comparing the difference between predicted noise and real noise, the real pixels and textures in the Himawari L3 satellite data are identified, and a spatiotemporal constrained diffusion model is constructed.

[0061] The identification of real pixels and textures in Himawari L3 satellite data is achieved through the loss function formula (3-2):

[0062]

[0063] As a preferred technical solution, steps 331 to 335 are repeated to iteratively train the spatiotemporal constraint diffusion model, and the model parameters are continuously corrected using the gradient descent algorithm until the loss function reaches a certain value on the validation set. The magnitude remains stable, i.e., training stops and a well-trained spatiotemporal constraint diffusion model is obtained; the weight file is then obtained.

[0064] As a preferred technical solution, 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 temporal dual embedding layer.

[0065] The ConvNeXt V2 module employs a global response normalization (GRN) mechanism in each ResBlock:

[0066]

[0067]

[0068]

[0069] in, Input feature map, dimension 1 ;

[0070] Feature maps in spatial dimensions ( On Norm, representing channel strength;

[0071] The normalized response value reflects the competitive position between channels;

[0072] Total number of channels; : To prevent extremely small constants from being divided by zero;

[0073] Element-wise multiplication;

[0074] Learnable scaling and translation parameters, where, and Initialize them to 0 respectively, so that the module behaves as an identity mapping in the early stage of training;

[0075] The linear attention module with memory key-value pairs extracts features using formula (1), reducing the complexity from... Reduce to ;

[0076] (1)

[0077] This includes incorporating Memory Key-Value pairs to learn global physics priors:

[0078]

[0079] in, The query matrix represents the pixel features that need to be reconstructed.

[0080] The key matrix and value matrix represent the spatial context;

[0081] : Memory KV, which is the pre-set learnable static parameter in the model, used to store the prior global sea surface temperature distribution;

[0082] : Perform Softmax normalization on the matrix columns;

[0083] : Perform Softmax normalization on the matrix rows;

[0084] : Matrix transpose operator;

[0085] The time-based dual embedding layer encodes the dates in the NPZ file names using formulas (2) and (3), and uses a global MLP for feature adaptation.

[0086] (2)

[0087] (3)

[0088] in, Location encoding;

[0089] By using sine and cosine waves of different frequencies, time is mapped to a multidimensional space;

[0090] Physical time signal;

[0091] : The total dimension of the embedded vector;

[0092] Dimension index;

[0093] : The base number for controlling wavelength.

[0094] As a preferred technical solution, in step three, texture enhancement and large-scale reconstruction optimization are performed using Monte Carlo ensemble sampling and dynamic boundary fusion algorithms.

[0095] Monte Carlo ensemble sampling, utilizing Execute with different random seeds Each independent denoising sampling process yields a set of sample distributions for the reconstructed image. ;

[0096] The generated samples are then weighted and averaged.

[0097]

[0098] in, The final reconstructed sea surface temperature field;

[0099] Number of samples;

[0100] : No. Samples generated through independent denoising;

[0101] : The weighted weight of each sample.

[0102] As a preferred technical solution, the dynamic boundary fusion algorithm determines the spatial topology position of the reconstructed blocks in real time through an adaptive boundary anchoring algorithm, performs forced correction on the weight operators that conform to the global physical boundary, and keeps their weight values ​​constant at 1.0; and gradually shrinks the fusion width during the denoising iteration process based on the boundary gradient between the reconstructed region and the known observation region.

[0103] Fusion formula:

[0104]

[0105] in, : The merged pixel values;

[0106] Spatial weight mask: the weights at the edges approach 0 and at the center approach 1.

[0107] : The reconstruction result of the current segment;

[0108] : The result of the overlapping area of ​​adjacent slices.

[0109] This invention provides a satellite sea surface temperature (SST) missing data reconstruction method based on a spatiotemporal constrained diffusion model. It performs generative reconstruction of missing data from the mechanism of SST dynamic residual distribution and historical texture prior. Compared with the prior art, it significantly improves the texture restoration accuracy and physical consistency accuracy of satellite SST under continuous large-scale cloud cover. Attached Figure Description

[0110] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0111] Figure 1 is a flowchart of the satellite sea surface temperature reconstruction method based on the spatiotemporal constrained diffusion model of the present invention;

[0112] Figure 2 is a schematic diagram of the improved Unet architecture;

[0113] Figure 3 shows the convergence analysis of training loss.

[0114] Figure 4 shows the average performance of scenarios one, two, and three under different occlusion conditions;

[0115] Figure 5 is a schematic diagram of the actual cloud obstruction observed in satellite observations of the South China Sea.

[0116] Figure 6 shows the average power spectral density (PSD) curve;

[0117] Figure 7 is a schematic diagram of the Himawari true values ​​in the South China Sea region;

[0118] Figure 8 is a schematic diagram of the reconstructed SST in the South China Sea region;

[0119] Figure 9 is a schematic diagram of the true values ​​of Himawari in the Indian Ocean.

[0120] Figure 10 is a schematic diagram of the reconstructed SST in the Indian Ocean. Detailed Implementation

[0121] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0122] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article, unless otherwise specified, generally indicates that the preceding and following related objects have an "or" relationship.

[0123] The embodiments of the present invention will now be described in further detail with reference to the accompanying drawings.

[0124] Referring to Figure 1, this invention provides a satellite sea surface temperature reconstruction method based on a spatiotemporal constrained diffusion model, comprising the following steps:

[0125] Step 1: Acquire missing Himawari L3 satellite data and OSTIA L4 global data for global sea areas.

[0126] The missing Himawari L3 satellite data is satellite data obscured by cloud cover, i.e., actual satellite data.

[0127] OSTIA Level 4 global data is complete, with no missing data.

[0128] Step two involves preprocessing the missing Himawari L3 satellite data and OSTIA L4 global data to obtain several slice data.

[0129] Step 21: Perform bicubic interpolation on the OSTIA L4 global data to magnify 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.

[0130] Step 22: For the corresponding sea area, perform a sliding window tiling of 128×128 pixels with a step size of 64 pixels to obtain several uniform tiling data, save them as NPZ files, and name them with the corresponding date.

[0131] Set the step size to 64 pixels, meaning that there is a 50% overlap between adjacent slices in both the horizontal and vertical directions.

[0132] 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.

[0133] Step three involves inputting the preprocessed slice data into a trained spatiotemporal constrained diffusion model. The model then reconstructs the slice data using a weight file. The reconstructed slice data is then subjected to texture enhancement and large-scale reconstruction optimization to obtain complete reconstructed data covering the entire sea area. This ultimately achieves the reconstruction of Himawari L3 satellite data.

[0134] In step three, the construction and training process of the spatiotemporal constraint diffusion model is as follows:

[0135] Step 31, multi-source data acquisition;

[0136] Step 32: Slice the multi-source data;

[0137] Step 33: Construct and train a spatiotemporal constrained diffusion model; obtain a weight file; the weight file is used to reconstruct the missing Himawari L3 satellite data.

[0138] In step 31, the multi-source data includes Himawari L3 satellite data (Target Data) and OSTIA L4 global data (Background Data).

[0139] The Himawari L3 satellite data boasts a spatial resolution of up to 0.02° (approximately 2 kilometers), enabling it to keenly capture sub-mesoscale features and minute temperature gradients in the ocean surface. It was acquired using the Himawari third-generation geostationary weather satellite, released by the Japan Meteorological Agency (JMA).

[0140] OSTIA Level 4 global data integrates multiple observation sources, including infrared, microwave, ship, and buoy data, through optimal interpolation (OI) technology, providing a seamless global sea surface temperature field with a daily spatial resolution of 0.05°. It is acquired using the OSTIA (Operational Sea Surface Temperature and Sea Ice Analysis) Level 4 multi-source fusion product published by the UK Met Office.

[0141] Both Himawari L3 satellite data and OSTIA L4 global data comprised 85% of the total acquired data and were used as the training set for the spatiotemporal constrained diffusion model. During the reconstruction process, OSTIA L4 global data served as the "spatial constrained background," guiding the spatiotemporal constrained diffusion model to recover the fundamental field conforming to large-scale physical laws in the missing regions through a dual-channel input method (Himawari L3 satellite data and OSTIA L4 global data).

[0142] This invention selects the South China Sea and its surrounding waters as the core test field for algorithm verification. The time span is from January 1, 2020 to December 31, 2022, specifically covering the geographical area from latitude 4.0°N to 24.48°N and longitude 106.0°E to 123.92°E. This sea area has complex topography and abundant mesoscale eddies, branches of the Kuroshio Current, and their marginal frontal activities. The complex textures generated by this highly dynamic dynamic environment are the best scenario to test whether the diffusion model can recover high-frequency details while maintaining its original accuracy.

[0143] In step 32, the acquired multi-source data undergoes refined slice preprocessing to adapt to the computational requirements of the diffusion model for high-dimensional tensor inputs. The slice preprocessing steps are as follows:

[0144] Step 321: Select the sea area corresponding to the multi-source data.

[0145] Specifically, the South China Sea and its surrounding waters were selected, and the image size was set to 1024×896 pixels to match... Spatial resolution.

[0146] Step 322: Perform bicubic interpolation on the OSTIA L4 global data to magnify 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.

[0147] Due to OSTIA L4 level global data ( ) and Himawari L3 satellite data ( Since the resolutions of the two sources are inconsistent, a bi-cubic interpolation technique is used to upsample the background source slice to a 128×128 dimension consistent with the target source, ensuring the spatial correlation of the dual-channel input.

[0148] The latitude and longitude axes are rearranged (latitude in descending order, longitude in ascending order) to ensure that the physical spatial orientation of all slices is strictly consistent.

[0149] In order for the model to learn to use OSTIA L4 global data as constraints to reconstruct the Himawari texture, pixel-level alignment between the two must be achieved.

[0150] Step 323: Perform cloud removal filtering on the Himawari L3 satellite data after aligning latitude and longitude, and ensure that its effective pixels match the sea area range (no land) of the OSTIA L4 global data.

[0151] The process of cloud removal filtering for Himawari L3 satellite data after latitude and longitude alignment includes temperature scale conversion and outlier filtering.

[0152] Temperature scale conversion: The initial temperature scale is determined by detecting the pixel average. The initial temperature scale is determined based on Himawari L3 satellite data and OSTIA L4 global data. If it is a Kelvin scale, it is uniformly converted to Celsius. For example, if the detected average temperature is greater than... If it is, then it is determined to be a Kelvin unit and automatically subtracted. All values ​​are converted to Celsius (°C).

[0153] Outlier filtering: Identify and remove outliers that are outside the reasonable ocean temperature range (−2.0°C to 45.0°C) and label them as NaN.

[0154] Specifically, to ensure the convergence of model training, land mask culling and 100% cloudless constraints are used for filtering.

[0155] Land masking removal: Taking advantage of the integrity of OSTIA L4 level global data, if a slice contains any NaN value (representing land or no data area), the sample is discarded directly.

[0156] 100% Cloud-Free Constraint (for Target Data): As the "ground truth" (GT) for the spatiotemporal constrained diffusion model, the training slices of Himawari L3 satellite data must be completely cloud-free. High-quality observation samples are selected by checking for missing values ​​within the slices, ensuring that the spatiotemporal constrained diffusion model learns the true ocean texture rather than cloud artifacts.

[0157] Step 324: Perform a 128×128 pixel sliding window slice for the corresponding sea area to obtain multiple uniform slices, save them as NPZ files, and name them with the corresponding date.

[0158] Specifically, the processed paired slices (Himawari L3 satellite data and OSTIA L4 global data (background data)) are packaged into independent NPZ files for efficient loading by PyTorch DataLoader.

[0159] A sliding window slice of 128×128 pixels is used to ensure that each small block maintains local texture relevance while reducing memory pressure.

[0160] Total number of samples: A total of 22,093 valid samples without land interference were screened.

[0161] Data set allocation: 85% of the data from 2020 to 2022 was used as the training set and 15% as the validation set for model learning.

[0162] Cross-year testing: Himawari L3 satellite data and corresponding OSTIA L4 global data for the entire year of 2023 were independently selected as test sets to evaluate the long-term robustness of the model under real climate variability.

[0163] At the end of the slice preprocessing stage, statistical information is synchronized across processes using the all_reduce operator. 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 constrained diffusion model.

[0164] In step 32, given the massive amount of satellite data (covering multi-year diurnal scale data), a parallel preprocessing scheme based on PyTorchDistributed is adopted to significantly improve computational efficiency. The core of multi-source data preprocessing lies in ensuring that every slice learned by the model possesses extremely high scientific value through physical cleaning, spatial alignment, and rigorous filtering.

[0165] Step 33: Construct and train a spatiotemporal constraint diffusion model.

[0166] The process of constructing the spatiotemporal constraint diffusion model is as follows:

[0167] Step 331: First, linear normalization is performed on the multi-source data, followed by data augmentation. This ensures the mathematical stability of the diffusion process.

[0168] Linear normalization: Mapping the original sea surface temperature values ​​(8.0°C ~ 34.0°C) to... Interval.

[0169]

[0170] in, : Normalized multi-source data values;

[0171] : Raw multi-source data physical values ​​(degrees Celsius);

[0172] The lower limit for multi-source data is set to 8.0.

[0173] : Upper limit of multi-source data, set to 34.0.

[0174] Data augmentation: Random horizontal and vertical flipping is performed on the slices to improve the generalization ability of the spatiotemporal constrained diffusion model under different ocean current directions.

[0175] Step 332: Extract the multidimensional physical time feature vector corresponding to the date in the slice file name. This approach enables the spatiotemporally constrained diffusion model to handle seasonal fluctuations and interannual evolution of sea surface temperature (SST). Specifically, sine and cosine transforms are used to map the day of the year (doy) to a two-dimensional unit circle coordinate space. This periodic encoding effectively resolves the contradiction in traditional linear time representation where the end of December (365 days) and the beginning of January (1 day) are numerically far apart but physically adjacent, ensuring the smooth continuity of SST seasonal cyclical features in the feature space. Simultaneously, a year offset relative to the 2020 baseline year is introduced as a third feature, enabling the model to capture global warming trends or anomalous evolution in specific years.

[0176] Step 333: Forward noise is added to the Himawari L3 satellite data to obtain the Himawari L3 satellite data to be denoised.

[0177] Forward noise addition: A cosine noise schedule is used for noise addition. Compared with a linear schedule, it can distribute the attenuation process of image information more evenly.

[0178] For a given Himawari L3 satellite data, the true image In the Noise addition result of each step The direct calculation is as follows:

[0179]

[0180] in, True-value image of normalized Himawari L3 satellite data;

[0181] : Noisy latent variables generated at any given time;

[0182] : indicates a Gaussian distribution (normal distribution);

[0183] : Noise scaling factor defined by the cosine schedule;

[0184] The result of multiplying them together.

[0185] : Identity matrix, representing standard normal distribution noise.

[0186] Step 334: The Himawari L3 satellite data to be denoised is concatenated with the OSTIA L4 global data as background field constraints in the channel dimension to perform denoising, while simultaneously predicting the noise distribution.

[0187] The Himawari L3 satellite data to be denoised is set as follows: OSTIA L4 global data setting is Predict from noise using formula (3-1) :

[0188]

[0189] in, The probability distribution of the reverse generation process (denoising process) learned by the improved UNet model;

[0190] : Latent variables of the reconstructed image at the moment preceding the denoising process;

[0191] : Diffusion steps (noise addition time). It represents the number of diffusion steps in the original image. Specific stages of noise injection.

[0192] :current Latent variables of the noisy image at any given time;

[0193] : Full-coverage background field constraints provided by OSTIA data;

[0194] : A physical time feature vector containing date (doy) and year offset;

[0195] : indicates a Gaussian distribution (normal distribution);

[0196] The mean of the Gaussian distribution predicted by the improved UNet model represents the expected value of the reconstructed image.

[0197] The Gaussian variance predicted by the improved UNet model represents the uncertainty in the reconstruction process.

[0198] In step 334, denoising employs an improved UNet architecture, as shown in Figure 2. The improved UNet architecture is used to predict the noise distribution.

[0199] The improved UNet architecture includes the ConvNeXt V2 module, a linear attention module with Memory KV, and a Time Double Embedding Layer (TE).

[0200] In the noise prediction distribution stage of the diffusion model, the improved UNet architecture uses a temporal dual embedding layer (TE) as the core modulator to coordinate the "space-time-object" three-in-one collaborative processing between the ConvNeXt V2 module and the Memory KV linear attention module. Specifically, the seasonal and interannual features extracted by TE are injected into the ConvNeXt V2 module through a parametric modulation layer, guiding it to dynamically adjust the response weights to local sub-mesoscale details based on the current physical background (such as the dynamic characteristics of a specific season). At the same time, TE provides a physical index for the Memory KV linear attention module, enabling it to accurately retrieve global spatial constraints that conform to the current evolutionary trend from the long-term sea surface temperature distribution patterns stored in the Memory KV linear attention module. This deep coupling of local convolutional detail capture, global attention correlation modeling, and macroscopic temporal prior in the improved UNet architecture ensures that the spatiotemporal constraint diffusion model can predict noise residuals with accurate physical background and coherent spatial structure at each step t of the reverse denoising process. This enables high-fidelity SST field reconstruction even under extreme observational limitations.

[0201] ConvNeXt V2 module, using Large kernel depthwise convolution, replacing the standard in existing technologies The convolutional layer residual module significantly expands the receptive field, enabling better capture of mesoscale ocean phenomena.

[0202] To enhance feature competition between channels and improve texture sharpness at sea surface temperature edges, a Global Response Normalization (GRN) mechanism is employed in each ResBlock:

[0203]

[0204]

[0205]

[0206] in, Input feature map, dimension 1 ;

[0207] Feature maps in spatial dimensions ( On Norm, representing channel strength;

[0208] The normalized response value reflects the competitive position between channels;

[0209] Total number of channels; : To prevent extremely small constants from being divided by zero;

[0210] Element-wise multiplication;

[0211] Learnable scaling and translation parameters, where, and Initialize them to 0 respectively, so that the module behaves as an identity mapping in the early stage of training;

[0212] The linear attention module with memory key-value pairs extracts features using formula (1), reducing the complexity from... Reduce to .

[0213] (1)

[0214] This includes incorporating Memory Key-Value pairs to learn global physics priors:

[0215]

[0216] Letter meanings:

[0217] The query matrix represents the pixel features that need to be reconstructed.

[0218] The key matrix and value matrix represent the spatial context;

[0219] : Memory KV, which is the pre-set learnable static parameter in the model, used to store the prior global sea surface temperature distribution;

[0220] : Perform Softmax normalization on the matrix columns;

[0221] : Perform Softmax normalization on the matrix rows;

[0222] : Matrix transpose operator.

[0223] The linear attention module with Memory KV significantly reduces the computational complexity of training on 128×128 resolution sea surface temperature slices, improving inference speed by approximately 50%. This shortens training time and increases training speed.

[0224] The time-dual embedding layer encodes the dates in the NPZ file names obtained in step 324 using formulas (2) and (3), allowing the spatiotemporal constraint diffusion model to learn seasonal changes rather than fixed times, and uses a global MLP for feature adaptation.

[0225] (2)

[0226] (3)

[0227] in, Location encoding;

[0228] By using sine and cosine waves of different frequencies, time is mapped to a multidimensional space;

[0229] Physical time signal;

[0230] : The total dimension of the embedded vector;

[0231] Dimension index;

[0232] : The base number for controlling wavelength.

[0233] Step 335: By training and comparing the difference between predicted noise and real noise, the real pixels and textures in the Himawari L3 satellite data are identified, and a spatiotemporal constrained diffusion model is constructed.

[0234] The identification of real pixels and textures in Himawari L3 satellite data is achieved through the loss function formula (3-2):

[0235]

[0236] in, Training loss value;

[0237] : indicates training data Random noise sampled from a standard normal distribution and the number of diffusion steps for random sampling Find the expected value of the joint distribution;

[0238] True noise sampled from a standard normal distribution;

[0239] Noise residuals predicted by the improved UNet;

[0240] : Noisy images at any given time;

[0241] OSTIA background conditions guidance;

[0242] time: An embedding vector that combines diffusion steps and seasonal information.

[0243] Step 336: Repeat steps 331 to 335 to iteratively train the spatiotemporal constraint diffusion model, continuously refining the model parameters using the gradient descent algorithm until the loss function reaches a certain value on the validation set. The magnitude remains stable, as shown in Figure 3, i.e., training stops and a well-trained spatiotemporal constraint diffusion model is obtained; the weight file is then obtained.

[0244] Hyperparameter settings for training the spatiotemporal constrained diffusion model: The AdamW optimizer was used, with an initial learning rate set to... In conjunction with the ReduceLROnPlateau scheduler, the learning rate is halved when the validation set loss no longer decreases within 15 epochs.

[0245] 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 multidimensional temporal features, thereby achieving high-precision spatiotemporal continuous sea surface temperature reconstruction.

[0246] In step three, texture enhancement and large-scale reconstruction optimization are performed using Monte Carlo ensemble sampling and dynamic boundary fusion algorithms. This makes the reconstructed data of the entire sea area more closely resemble the real texture and enables complete reconstruction of the entire South China Sea region.

[0247] Monte Carlo ensemble sampling, at the numerical level, suppresses high-frequency random noise through multi-sample averaging, significantly reducing the root mean square error (RMSE) between the reconstructed field and Himawari L3 satellite data. At the physical level, it preserves the robust dynamic characteristics of cross-path coexistence, resulting in a smoother spatial gradient in the final SST field that conforms to the actual ocean thermodynamic evolution. Monte Carlo ensemble sampling achieves a leap from "probabilistic generation" to "steady-state reconstruction," laying a statistical foundation for the high-precision engineering application of sea surface temperature fields across the entire ocean.

[0248] The operation steps are as follows:

[0249] In the inference phase, using Execute with different random seeds A separate denoising sampling process.

[0250] Obtain a set of sample distributions for the reconstructed image. .

[0251] The generated samples are then weighted and averaged.

[0252]

[0253] in, The final reconstructed sea surface temperature field;

[0254] Number of samples;

[0255] : No. Samples generated through independent denoising;

[0256] : The weighted weight of each sample.

[0257] Monte Carlo ensemble sampling addresses the issue of random artifacts or anomalous spots that may be generated in a single denoising process. It can enhance stable ocean feature textures (such as constant vortex edges) while canceling out uncertain spots generated in a single sampling, significantly reducing reconstruction errors.

[0258] The dynamic boundary fusion algorithm addresses the potential risk of weight collapse at the geometric boundaries of the entire field. It uses an adaptive boundary anchoring algorithm to determine the spatial topological position of the reconstructed blocks in real time and forcibly corrects the weight operators that conform to the global physical boundary, keeping their weight values ​​constant at 1.0. This ensures that even in scenarios where adjacent overlapping blocks are missing compensation, the edge regions maintain the original physical numerical magnitude of the reconstructed field, thus mathematically eliminating non-physical edge numerical decay and guaranteeing the physical consistency of the overall field reconstruction results.

[0259] The dynamic boundary fusion algorithm gradually shrinks the fusion width during the denoising iteration process based on the boundary gradient between the reconstructed region and the known observation region.

[0260] Fusion formula:

[0261]

[0262] in, : The merged pixel values;

[0263] Spatial weight mask: the weights at the edges approach 0 and at the center approach 1.

[0264] : The reconstruction result of the current segment;

[0265] : Results of overlapping areas of adjacent slices;

[0266] The dynamic boundary fusion algorithm can effectively smooth the sampling variance between different slices, making the temperature gradient transition at the splicing point natural and eliminating the "mesh effect" produced by the traditional block method.

[0267] For the reconstruction of the entire South China Sea, the sea area needs to be divided into 128×128 pixel blocks, reconstructed separately, and then stitched together. To eliminate the "seams" or "discontinuities" at the stitching points, this embodiment abandons "hard shearing" and instead adopts the aforementioned dynamic boundary fusion algorithm. Even when generating ultra-large images spanning thousands of kilometers, the edges still maintain smooth texture transitions.

[0268] This invention uses the South China Sea as an example for experiments.

[0269] To comprehensively evaluate the robustness of the model, this invention conducts stress tests using four typical occlusion scenarios:

[0270] Scenario 1: Full Image Random Reconstruction: Simulates the situation where the entire image is completely obscured by clouds, and reconstructs it using only the OSTIA background field and time encoder.

[0271] 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.

[0272] 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.

[0273] 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:

[0274] RMSE (Root Mean Square Error): Measures the absolute temperature deviation between the reconstructed value and the true value.

[0275] SSIM (Structural Similarity): Evaluates the degree of similarity between a reconstructed image and the ground truth in terms of brightness, contrast, and structure.

[0276] 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.

[0277] Himawari Truth Value: This refers to the raw data from the Himawari L3 satellite as the comparison data.

[0278] The average performance under different occlusion scenarios is shown in Figure 4 for scenarios 1, 2, and 3; and in Figure 5 for scenario 4.

[0279] Figure 4 shows the reconstruction performance under three typical idealized masking conditions, intuitively reflecting 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 the temporal dual embedding layer (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.

[0280] Figure 5 illustrates the actual cloud obstruction observed in satellite observations of the South China Sea, highlighting the model's ability to integrate and preserve the fidelity of fragmented observation signals. Even with extremely sparse and irregularly fragmented satellite observations, this invention demonstrates strong robustness, accurately extracting thermodynamic features from scattered points and using high-quality filling based on global spatial constraints provided by OSTIA L4 level global data. The reconstructed results not only maintain consistency with the true values ​​in macroscopic distribution but also successfully recreate the temperature step characteristics at the edge of mesoscale eddies.

[0281] To verify whether the reconstructed field retains the original high-frequency texture details of Himawari, this invention uses power spectral density (PSD) analysis.

[0282] The PSD curves show the energy distribution at different spatial scales (wavelengths), as shown in Figure 6.

[0283] The black line (Himawari truth value) represents a high-precision true structure.

[0284] The blue line (OSTIA) shows a significant energy deficiency at small scales (high frequency region), and the curve is smooth.

[0285] The red line (reconstructed SST) closely matches the black line across the entire wavelength range, especially in the small-scale region. The high-frequency details lost by OSTIA were successfully recovered (km).

[0286] 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 real physical field.

[0287] Full reconstruction display of the South China Sea region: The Himawari ground truth value of the South China Sea region is shown in Figure 7, and the SST after reconstruction of the South China Sea region is shown in Figure 8.

[0288] Land Boundary Filling: For the complex coastline and island / reef areas of the South China Sea, this invention employs a land filling algorithm based on physical distance. Since neural networks cannot directly handle NaN (NaN) missing values ​​on land, this algorithm utilizes Euclidean distance transformation to replace invalid pixels in the slice affected by land with valid observations from the nearest geographically located sea area. This operation ensures the numerical stability of the convolution operator at the land-sea boundary, avoiding edge collapse.

[0289] Visualization: The reconstructed SST field not only perfectly fills the gap left by the cloud (White Cloud Area), but also eliminates the seams of the block splicing through the dynamic boundary fusion algorithm.

[0290] Indian Ocean Migration Experiment Demonstration: To further verify the spatial robustness and migration capability of this invention under different geodynamic environments, this study directly applies the model weights trained in the South China Sea to the Indian Ocean for generalization experiments. The core objective of this step is to explore whether this invention has truly learned the general physical characteristics of sea surface temperature reconstruction, and its reliability in performing "zero-sample" migration tasks in unseen geographical areas.

[0291] The true value of Himawari in the Indian Ocean is shown in Figure 9, and the reconstructed SST in the Indian Ocean is shown in Figure 10.

[0292] Leveraging the high-frequency texture generation capabilities learned in the South China Sea, this invention successfully extracted natural sea surface temperature fluctuations within these "blind zones." Although the model had not studied samples from the Indian Ocean, the reconstruction results accurately reproduced the vortex topology within the Bay of Bengal and the frontal features along the Arabian Sea coast. This demonstrates that the spatiotemporally constrained diffusion model captures the universal physical mapping logic of "recovering high-frequency dynamic details from background field residuals," rather than merely memorizing geographical coordinates of the South China Sea, exhibiting strong spatial robustness.

[0293] This invention, through stress testing on three scenarios—full-map randomization, half-masking, and local missing data—achieved high-fidelity reconstruction of sub-mesoscale features in the South China Sea region. In the fully random reconstruction scenario, the average RMSE was 0.3358°C, SSIM was 0.9718, and PSNR was 38.3 dB. Power spectral density (PSD) analysis showed that the reconstruction curve (red) maintained the highest consistency with the Himawari ground truth (black) across the entire wavelength range, particularly in the high-frequency band of 10–30 km, effectively correcting the energy collapse phenomenon caused by excessive smoothing in the OSTIA L4 level global data (blue). This indicates that the invention not only approximates the ground truth at the pixel level but also reconstructs the energy cascade characteristics of ocean turbulence at the physical energy spectrum level, proving that its reconstructed field possesses a realistic ocean dynamic mechanism. Through extensive comparative experiments and field data verification, the spatiotemporal constrained diffusion model of this invention has been demonstrated to have outstanding performance in solving the problem of "high-precision data missing". Its reconstruction product has high physical reliability and dynamic fidelity, and has significant scientific value and engineering application prospects.

[0294] This invention provides a satellite sea surface temperature (SST) missing data reconstruction method based on a spatiotemporal constrained diffusion model. It performs generative reconstruction of missing data from the mechanism of SST dynamic residual distribution and historical texture prior. Although this method is limited by the validity of previous day historical data in extreme cases, it significantly improves the texture restoration accuracy and physical consistency accuracy of satellite SST under continuous large-scale cloud cover compared with the prior art.

[0295] Of course, the present invention may have other various embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art can make various corresponding changes and modifications according to the present invention, but these corresponding changes and modifications should all fall within the protection scope of the appended claims.

Claims

1. A satellite sea surface temperature reconstruction method based on a spatiotemporally constrained diffusion model, characterized in that, Includes the following steps: Step 1: Acquire missing Himawari L3 satellite data and OSTIA L4 global data for the entire global ocean area. Step 2: Preprocess the missing Himawari L3 satellite data and OSTIA L4 global data to obtain several data slices. Step 3: Input the preprocessed data slices into a trained spatiotemporal constrained diffusion model. The spatiotemporal constrained diffusion model reconstructs the data slices using a weight file. The reconstructed data slices are then subjected to texture enhancement and large-scale reconstruction optimization to obtain complete reconstructed data for the entire ocean area. In Step 3, the construction and training process of the spatiotemporal constrained diffusion model is as follows: Step 31: Acquire multi-source data; Step 32: Process the multi-source data into slices; Step 33: Construct and train the spatiotemporal constrained diffusion model; obtain the weight file. The weight file is used to reconstruct the missing Himawari L3 satellite data; in step 33, the spatiotemporal constrained diffusion model construction process is as follows: step 331, firstly, the multi-source data is linearly normalized, and then data augmentation is performed; Linear normalization: Mapping the original sea surface temperature values ​​to... Interval: in, : Normalized multi-source data values; : Original multi-source data physical values; The lower limit for multi-source data is set to 8.

0. : Upper limit of multi-source data, set to 34.0; Data augmentation: Perform random horizontal and vertical flipping on the slices; Step 332, extract the multidimensional physical time feature vector corresponding to the date in the slice file name. Step 333: Forward noise addition is performed on the Himawari L3 satellite data to obtain the Himawari L3 satellite data to be denoised; noise addition is performed using a cosine noise schedule: for a given ground truth image of Himawari L3 satellite data In the Noise addition result of each step The calculation is as follows: in, True-value image of normalized Himawari L3 satellite data; ： Noisy latent variables generated at any given time; : Indicates a Gaussian distribution; : Noise scaling factor defined by the cosine schedule; The result of multiplying them together; The identity matrix represents the standard normally distributed noise; step 334 involves concatenating the Himawari L3 satellite data to be denoised with the OSTIA L4 global data (used as background field constraints) along the channel dimension to perform denoising and simultaneously predict the noise distribution; the Himawari L3 satellite data to be denoised is set as... OSTIA L4 global data setting is Predict from noise using formula (3-1) ： in, : Probability distribution of the reverse generation process learned by the improved UNet model; : Latent variables of the reconstructed image at the moment preceding the denoising process; : Number of diffusion steps; :current Latent variables of the noisy image at any given time; : Full-coverage background field constraints provided by OSTIA data; : A physical time feature vector containing date and year offsets; : Indicates a Gaussian distribution; The mean of the Gaussian distribution predicted by the improved UNet model represents the expected value of the reconstructed image. The Gaussian variance predicted by the improved UNet model represents the uncertainty in the reconstruction process; in step 335, by comparing the difference between predicted noise and real noise during training, the real pixels and textures in the Himawari L3 satellite data are identified, and a spatiotemporal constrained diffusion model is constructed; the identification of real pixels and textures in the Himawari L3 satellite data is achieved through the loss function formula (3-2): in, Training loss value; : indicates training data Random noise sampled from a standard normal distribution and the number of diffusion steps for random sampling Find the expected value of the joint distribution; True noise sampled from a standard normal distribution; Noise residuals predicted by the improved UNet; ： Noisy images at any given time; OSTIA background conditions guide; time: embedding vector integrating diffusion steps and seasonal information; in step 334, denoising adopts an improved UNet architecture, which includes a ConvNeXt V2 module, a linear attention module with memory KV, and a temporal dual embedding layer; the ConvNeXt V2 module uses a global response normalization GRN mechanism in each ResBlock: in, Input feature map, dimension 1 ； Feature maps in spatial dimensions Norm, representing channel strength; The normalized response value reflects the competitive position between channels; Total number of channels; : To prevent extremely small constants from being divided by zero; Element-wise multiplication; Learnable scaling and translation parameters, where, and Initialize them to 0 respectively, so that the module behaves as an identity mapping in the early stage of training; the linear attention module with memory KV extracts features through formula (1), reducing the complexity from Reduce to ； (1) Among them, Memory KV is added to learn global physical priors: in, : Query matrix, representing the pixel features that need to be reconstructed; The key matrix and value matrix represent the spatial context; : Memory KV, which is the pre-set learnable static parameter in the model, used to store the prior global sea surface temperature distribution; : Perform Softmax normalization on the matrix columns; : Perform Softmax normalization on the matrix rows; : Matrix transpose operator; Temporal dual embedding layer, which encodes the dates in the NPZ file names using formulas (2) and (3), and uses global MLP for feature adaptation; （2） (3) Among them, Location encoding; By using sine and cosine waves of different frequencies, time is mapped to a multidimensional space; Physical time signal; : The total dimension of the embedded vector; Dimension index; : The base number for controlling wavelength.

2. The satellite sea surface temperature reconstruction method based on a spatiotemporal constrained diffusion model according to claim 1, characterized in that, In step two, the preprocessing steps for the missing Himawari L3 satellite data and OSTIA L4 global data are as follows: 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 latitude and longitude of the Himawari L3 satellite data and OSTIA L4 global data are aligned. Step 22: For the corresponding sea area, perform a sliding window tiling of 128×128 pixels with a step size of 64 pixels to obtain several uniform tiling data, save them as NPZ files and name them with the corresponding date; set the step size to 64 pixels, that is, there is 50% overlap between adjacent tilings in both the horizontal and vertical directions.

3. The satellite sea surface temperature reconstruction method based on a spatiotemporal constrained diffusion model according to claim 2, characterized in that, In step 31, the multi-source data includes Himawari L3 satellite data and OSTIA L4 global data; the spatial latitude and longitude resolution of the Himawari L3 satellite data is 0.02°; the spatial latitude and longitude resolution of the OSTIA L4 global data is 0.05°.

4. The satellite sea surface temperature reconstruction method based on a spatiotemporal constrained diffusion model according to claim 2, characterized in that, In step 32, the slice preprocessing steps are as follows: Step 321, select the sea area corresponding to the multi-source data; select the sea area and set the image size to 1024×896 pixels; Step 322, perform bicubic interpolation on the OSTIA L4 global data and enlarge it to the same precision as the Himawari L3 satellite data to align the latitude and longitude of the Himawari L3 satellite data and the OSTIA L4 global data. Step 323: Perform cloud removal filtering on the aligned latitude and longitude Himawari L3 satellite data and ensure that its effective pixels match the sea area range of the OSTIA L4 global data; the process of cloud removal filtering on the aligned latitude and longitude Himawari L3 satellite data includes temperature scale conversion and outlier filtering; temperature scale conversion: determine the initial temperature scale by detecting the pixel mean; Outlier filtering: Identify and remove outliers that are outside the reasonable ocean temperature range and mark them as NaN; Step 324: Perform a 128×128 pixel sliding window slice for the corresponding sea area to obtain multiple uniform slices, save them as NPZ files and name them with the corresponding date.

5. The satellite sea surface temperature reconstruction method based on a spatiotemporally constrained diffusion model according to claim 1, characterized in that, Repeat steps 331 to 335 to iteratively train the spatiotemporal constraint diffusion model, continuously refining the model parameters using the gradient descent algorithm until the loss function reaches a certain value on the validation set. The magnitude remains stable, i.e., training stops and a well-trained spatiotemporal constraint diffusion model is obtained; the weight file is then obtained.

6. The satellite sea surface temperature reconstruction method based on a spatiotemporal constrained diffusion model according to claim 1, characterized in that, In step three, texture enhancement and large-scale reconstruction optimization are performed using Monte Carlo ensemble sampling and dynamic boundary fusion algorithms; Monte Carlo ensemble sampling utilizes... Execute different random seeds Each independent denoising sampling process yields a set of sample distributions for the reconstructed image. The generated samples are then weighted and averaged. in, The final reconstructed sea surface temperature field; Number of samples; : No. Samples generated through independent denoising; : The weighted weight of each sample.

7. The satellite sea surface temperature reconstruction method based on a spatiotemporal constrained diffusion model according to claim 6, characterized in that, The dynamic boundary fusion algorithm determines the spatial topological position of the reconstructed blocks in real time through an adaptive boundary anchoring algorithm, and performs forced correction on the weight operators that conform to the global physical boundary, keeping their weight values ​​constant at 1.

0. Based on the boundary gradient between the reconstructed region and the known observation region, the fusion width is gradually reduced during the denoising iteration process. The fusion formula is: in, : The merged pixel values; Spatial weight mask, with weights approaching 0 at the edges and approaching 1 at the center; : The reconstruction result of the current segment; : The result of the overlapping area of ​​adjacent slices.

Images