Multi-source observation guided cascading diffusion downscaling method for marine element field
By using a cascaded diffusion downscaling method guided by multi-source observations, the residuals are decomposed into multi-scale residuals, and coarse- and fine-scale branch models are established. This solves the problem of insufficient recovery of the fine-scale structure of the ocean element field and enables the generation of high-resolution results and uncertainty analysis.
Patent Information
- Application Number
- CN202611532060.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-30
- Publication Date
- 2026-11-03
AI Technical Summary
Existing technologies are unable to effectively recover the fine-scale structure of ocean element fields. The generated results rely excessively on low-resolution backgrounds, and the mixed effects of multi-source observation scales lead to fragmented spatial morphology and weak details.
A cascaded diffusion downscaling method guided by multi-source observations is adopted. By decomposing the residuals into multi-scale residuals, a cascaded physical space diffusion model with coarse-scale and fine-scale branches is established. Prior noise estimation is constructed using multi-source observation data and static data, a conditional propagation path is established, and reverse denoising training is used to generate ocean structures.
It restores continuous frontal, vortex, and vortex filament structures, coordinates anomalies at different scales, improves the accuracy of structure localization and local variations, generates multiple physically plausible high-resolution results, and provides uncertainty distribution.
Smart Images
Figure CN123030965A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of marine environmental information processing, specifically to a multi-source observation-guided cascade diffusion downscaling method for marine element fields. Background Technology
[0002] Oceanographic data fields are crucial foundational data for understanding ocean dynamic processes, forecasting marine environments, and ensuring the safety of offshore production activities. Elements such as sea surface temperature, sea surface height, sea surface salinity, and ocean currents exhibit significant multi-scale characteristics, encompassing both large-scale, gently varying backgrounds and fine-scale structures such as ocean fronts, eddy edges, filamentary eddies, and local anomalies. Due to limitations in computational resources, numerical discretization methods, and observational coverage, the spatial resolution of ocean numerical models, reanalysis products, and some remote sensing products is often insufficient to fully describe these fine-scale structures. Therefore, spatial downscaling is necessary to obtain higher-resolution oceanographic data fields.
[0003] Traditional ocean downscaling methods mainly include statistical regression, spatial interpolation, and deterministic deep learning. Statistical regression and spatial interpolation typically estimate high-resolution results based on the correlation between neighboring grids, making it difficult to represent the nonlinear spatial variations generated by complex ocean dynamic processes. Deterministic deep learning methods can establish a mapping relationship between low-resolution and high-resolution fields, but their training objectives are usually based on mean squared error or mean absolute error. When the same low-resolution background may correspond to multiple fine-scale states, the model tends to output the average result of multiple possible states, manifesting as widened fronts, weakened gradients, vanishing vortices, and insufficient detail energy. Diffusion generation methods learn the probability distribution of high-resolution ocean element fields through progressive noise addition and reverse denoising, providing a new technical approach for recovering uncertain fine-scale structures.
[0004] However, existing methods typically use a low-resolution background as a strong constraint to directly generate a complete high-resolution field. Under this condition, the model tends to rely on the smooth structure in the background field, resulting in weak amplitudes of the generated anomalous components. Increasing the detail energy by increasing the high-frequency loss may lead to problems such as scattered spatial distribution, lack of directional continuity, and inaccurate phase positions. Therefore, simply recovering the spectral energy cannot guarantee that the generated details have a reasonable marine spatial morphology.
[0005] Actual ocean observations exhibit significant heterogeneity. Satellite remote sensing covers a wide area, but its spatial resolution, missing data range, and observation errors vary depending on the sensor and environmental conditions. Moorings and buoys provide continuous and highly accurate point-to-point observations, but their spatial coverage is limited. Underwater gliders can acquire high-precision observations along their movement trajectories, but their position and sampling range change over time. Different observations represent different spatial and temporal scales. If they are uniformly interpolated into a continuous grid, sparse observations may be artificially diffused to areas lacking observational data; if observations are directly written into the generated results, local abrupt changes may occur near the observation location; if used only as ordinary input variables, they may be masked by background field characteristics due to their scarcity. Furthermore, existing multi-source observation fusion methods mainly focus on the numerical error between observed and predicted values, with less emphasis on distinguishing the constraints of observations on structures at different spatial scales. Low-resolution backgrounds are more suitable for determining large-scale states, continuous satellite observations are more suitable for correcting mesoscale structural locations, and in-situ observations can provide information on local variations. If these scale effects are not distinguished, information from different sources may compete with each other, causing the model to compromise the surrounding structure while satisfying local observations, or to maintain overall smoothness without effectively utilizing high-precision field observations.
[0006] Therefore, there is a need for a multi-source observation-guided method for cascading diffusion downscaling of ocean element fields, which first learns the fine-scale structure distribution of the ocean and then implements cascading guidance based on the observation source, error, time, and representative scale. Summary of the Invention
[0007] The main objective of this invention is to provide a multi-source observation-guided cascade diffusion downscaling method for ocean element fields, in order to solve the problems of weak fine-scale structure, fragmented spatial morphology, excessive reliance on low-resolution background in the generated results, and mixed effects of multi-source observation scales in the existing technology.
[0008] To achieve the above objectives, this invention provides a multi-source observation-guided cascade diffusion downscaling method for ocean element fields, specifically including the following steps: S1: Acquire marine element fields, multi-source observation data, and static data of the target sea area, perform preprocessing, and obtain standardized data.
[0009] S2, construct background components based on standardized data, calculate residuals, and decompose residuals to obtain multi-scale residuals.
[0010] S3. Establish corresponding diffusion generation branches for residuals at each scale, construct a cascaded physical space diffusion model including coarse-scale branches and fine-scale branches, apply random noise to residuals at each scale to obtain noisy residual states, calculate prior noise estimates, establish conditional transmission paths for transmitting structural information from coarse-scale branches to fine-scale branches, and obtain ocean structure generation priors through reverse denoising training.
[0011] S4 obtains multi-time background conditions based on the actual background field and calculates satellite observation conditions using multi-source satellite remote sensing data.
[0012] S5 constructs dynamic sparse observation features based on field observation data and calculates the scale effect coefficients of each field observation on different scale diffusion branches.
[0013] S6 inputs multi-time background conditions, satellite observation conditions, and weighted dynamic sparse observation features into the corresponding diffusion branch, which, together with the ocean structure generation prior, constrains the generation of residuals at each scale.
[0014] S7. Based on the loss function, the trainable parameters of the conditional propagation path and diffusion generation branch are updated to obtain the trained multi-source observation-guided cascade diffusion model.
[0015] S8, guided by multi-source observations, the cascaded diffusion model performs reverse diffusion from coarse to fine scales, synthesizing the residuals generated at each scale with the actual background field to obtain a single downscaling member, and repeating sampling to determine the final downscaling result and uncertainty.
[0016] Furthermore, step S1 specifically includes the following steps: S1.1 The data acquired for the target sea area includes: multi-source remote sensing data, on-site observation data, time attribute conditions, static auxiliary conditions, historical first grid precision ocean element field and corresponding historical second grid precision background field, with the first grid precision being higher than the second grid precision.
[0017] S1.2, Preprocessing includes: time matching, spatial alignment, outlier handling, land-sea masking, unit unification and normalization of data from different sources; for data with known observation errors, the error magnitude is determined in the original physical units, and the observations and error information are converted to the data range used by the cascaded physical space diffusion model according to the same rules.
[0018] Furthermore, step S2 specifically includes the following steps: S2.1, based on standardized data, let the first... The historical first grid-precision ocean element field at that moment was Background component for: ; Among them, subscript Represents the background, and the background degradation operator is: The background resampling operator is .
[0019] S2.2, the first The residuals of the first-time historical grid-accuracy ocean feature field relative to the background components: ; in, For the first The residuals of the first-precision ocean feature field at a given moment relative to the background components.
[0020] S2.3, decompose the residuals and set... Individual scale decomposition operator , The number of scale layers is not less than 2. For the first The scale decomposition operator; the first Individual scale residuals for: ; The residuals at each scale together constitute the multi-scale residuals. : ; in, This indicates that approximate errors within a preset range are allowed during the scaling and reconstruction process.
[0021] S2.4, divides multi-scale residuals into coarse-scale residuals. and fine-scale residuals : ; Among them, superscript Indicates a coarse scale, superscript Indicates a finer scale.
[0022] Furthermore, step S3 specifically includes the following steps: S3.1, Construct a cascaded physical space diffusion model including coarse-scale and fine-scale branches; assuming For the first The true residuals at each scale without noise. For the first The first scale The residual states at each diffusion time step are given, and random noise is applied to the residuals at each scale. The forward diffusion process is as follows: ; in, For the first The retention factor of the original residual for each diffusion time step. For the first The noise addition factor for each diffusion time step, This represents random noise.
[0023] S3.2, p. Prior noise estimation at each scale for: ; Among them, subscript Indicates prior knowledge; For the first A cascaded physical space diffusion model at multiple scales; These are trainable parameters; These are static auxiliary conditions or time-related attribute conditions; Provides structural information for other scales.
[0024] S3.3, In the prior training phase of marine structures, establish a conditional transfer path for transmitting structural information from coarse-scale branches to fine-scale branches: ; in, For the first Each scale residual For the first Multi-scale structure extraction operators at various scales. To be passed to the first Structural information at each scale.
[0025] Furthermore, step S4 specifically includes the following steps: S4.1, Obtain the actual background field at the target time and the time adjacent to the target time, and calculate the background conditions at multiple times: ; in, For the first Background conditions at any given moment; superscript Indicates the actual numerical pattern, reanalysis data product, or business context; offset relative to the target time The actual background scene; A set of time offsets; Extract operators for background conditions.
[0026] S4.2, let the first... Multi-source satellite remote sensing data at a given time are The satellite observation conditions are: ; Among them, superscript Indicates satellite remote sensing, For the first Satellite observation conditions at that moment, Extracting operators for satellite observation conditions; To effectively cover the mask, This is information about observation errors.
[0027] Furthermore, step S5 specifically includes the following steps: S5.1, let the first... The field observation value is The relationship between this and the historical first-grid precision ocean feature field is as follows: ; in, For the first The observation operator corresponding to each observation, For the first Measurement error of the observation; For the first The first-ever grid-precision ocean element field in history at this moment.
[0028] S5.2, each field observation is converted into dynamic sparse observation features: ; in, For the first Dynamic sparse observation characteristics formed by a number of observations; Operator for extracting dynamic sparse observation features; For the observation location; The time interval between the observation time and the target time; The standard deviation of the observation error; Category of observation source; This is the representative scale for observation.
[0029] S5.3, for observations at fixed locations such as moorings and buoys, the observation sequence is divided into slowly varying components and short-term varying components: ; Among them, superscript Indicates slowly varying components, superscript It represents the short-term variation component.
[0030] The slowly varying component is: ;
[0031] The short-time variation components are: ; in, For the first A time-scale separation operator for fixed observations.
[0032] S5.4, Calculate the first... The observation on the first The effect coefficient of each scale: ; in, For the first The observation on the first The scaling factor of each branch, For the first The function for determining the observation action coefficients at each scale; For the first The background conditions at each moment; For the first Satellite observation conditions at a given moment.
[0033] The action coefficient satisfies: .
[0034] Furthermore, step S6 specifically includes the following steps: S6.1, let the first... The current spatial characteristics at each scale are The correction amount generated by the multi-time background conditions is: ; The correction amount caused by satellite observation conditions is: ; The correction amount generated by the on-site observation is: ; in, , and These are multi-temporal background conditions, satellite observation conditions, and on-site observation conditions for the first time. The correction amount generated by each scale branch; , and These are the operators that apply to the corresponding conditions; For the first The background conditions at each moment; For the first Satellite observation conditions at a given moment; For the first The dynamic sparse observation characteristics formed by the observations.
[0035] S6.2, the conditional noise estimation form is: ; in, For the first The final noise estimate after incorporating multi-source conditions at each scale; , and These are the conditional factors of background, satellite remote sensing, and field observation in the diffusion time step, respectively. For the first Prior noise estimation at each scale.
[0036] S6.3 generates prior prediction output based on the current diffusion state and ocean structure, recovering the first... Noise-free residuals at each scale : ; in, In order to be with the first Each diffusion time step and the corresponding recovery operator for the diffusion form; For the first The first scale The residual state of each diffusion time step.
[0037] Furthermore, step S7 specifically includes the following steps: S7.1 The loss function includes: diffusion base loss and auxiliary loss. The auxiliary loss includes: residual reconstruction loss, low-pass consistency loss, frequency band structure loss, frequency band energy loss, spatial gradient loss, power spectrum loss, background consistency loss and observation consistency loss.
[0038] S7.2, the total loss is : ; in, For diffusion-based losses; For the first Class structure auxiliary loss; This is the weight of the auxiliary loss; The loss adjustment function is determined based on the diffusion time step or signal-to-noise ratio; Loss of observation consistency; Weights for observation consistency loss; Background consistency loss; The weights for background consistency loss.
[0039] S7.3, the diffusion fundamental loss is: ; in, for The error measurement function used for each scale; Indicates random noise; For the first The final noise estimate after adding multi-source conditions at each scale.
[0040] Based on the noise-free residuals at each scale recovered in step S6.3, the noise-free residuals at each scale are synthesized with the background components to obtain the ocean element field estimate for the training phase: ; in, Estimate the ocean element field recovered during the training phase; For background components; For the recovery of the first Noise-free residuals at each scale.
[0041] The observation consistency loss is: ; in, For the first The weight of each observation in the loss; In order to be with the first An error measurement function adapted to the characteristics of the observed error; For the first The observation operator corresponding to each observation; For the first Item of on-site observation value.
[0042] S7.4 After updating the trainable parameters of the conditional propagation path and diffusion generation branch based on the loss function, the trained multi-source observation-guided cascade diffusion model is obtained.
[0043] Furthermore, step S8 specifically includes the following steps: S8.1 is based on a multi-source observation-guided cascade diffusion model, which performs reverse diffusion from coarse scale to fine scale. The coarse and fine scales are obtained by dividing them according to the spatial frequency characteristics of the residuals. The spatial frequency of the coarse scale residuals is lower than that of the fine scale residuals.
[0044] No. The scale in the th... The update process for each time step is as follows: ; in, The updated diffusion state; For the first The back-diffusion update operator at each time step, For the first Optional random terms used at each time step; For the first The first scale The residual state at each diffusion time step; For the first The final noise estimate after adding multi-source conditions at each scale.
[0045] S8.2, a single downscaling member is: ; in, For the first Descaled members at any given moment; This represents the actual background field on the target mesh. For the first Noise-free residuals generated at each scale; For the first Amplitude coefficients at each scale.
[0046] S8.3, under the same background and observation conditions, using different random initial states to perform... Second sampling, obtained One downscaling member: ; in, The total number of members in the set. For the set member index, For the first One downscaling member.
[0047] S8.4 uses the ensemble mean to represent the final downscaling result. The formula for calculating the ensemble mean is: ; in, For the first The average value of all members of the set at any given time.
[0048] S8.5 uses set dispersion to represent uncertainty. The formula for calculating set dispersion is: ; in, For the first The set discreteness at each time point.
[0049] The present invention has the following beneficial effects: This invention separates prior learning of ocean structures from background and observational guidance. Fine-scale structures are no longer entirely dominated by low-resolution background; the background and observations are primarily used to correct position, phase, and amplitude, thereby better recovering continuous frontal, vortex, and filamentary structures. Cascaded diffusion can coordinate anomalies at different scales, reducing details, structural fragmentation, and spectral imbalance. Differentiated guidance based on the characteristics of multi-source observations from satellites, moorings, buoys, and underwater gliders can also improve the accuracy of structure localization and local changes, avoiding abrupt changes caused by sparse observations. Under the same observational conditions, multiple physically plausible high-resolution results can be generated, and further, ensemble mean and uncertainty distribution can be obtained, providing support for ocean environmental analysis, refined forecasting, and subsequent data assimilation. Attached Figure Description
[0050] Figure 1 This is a flowchart of the multi-source observation-guided cascade diffusion downscaling method for ocean element fields according to the present invention.
[0051] Figure 2 This is the sea surface temperature field before downscaling.
[0052] Figure 3 This invention utilizes the multi-source observation-guided cascade diffusion downscaling method of the present invention to downscale the sea surface temperature field.
[0053] Figure 4 This is a high-resolution true value sea surface temperature field.
[0054] Figure 5 For the target coarse-scale residual plot.
[0055] Figure 6 This is a coarse-scale residual map generated using the multi-source observation-guided cascade diffusion downscaling method for ocean element fields according to the present invention.
[0056] Figure 7 For the target fine-scale residual plot.
[0057] Figure 8 This is a fine-scale residual map generated using the multi-source observation-guided cascade diffusion downscaling method for ocean element fields according to the present invention. Detailed Implementation
[0058] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0059] like Figure 1 The multi-source observation-guided cascade diffusion downscaling method for ocean element fields, as shown, specifically includes the following steps: S1: Acquire marine element fields, multi-source observation data, and static data of the target sea area, perform preprocessing, and obtain standardized data.
[0060] S2, construct background components based on standardized data, calculate residuals, and decompose residuals to obtain multi-scale residuals.
[0061] S3. Establish corresponding diffusion generation branches for residuals at each scale, construct a cascaded physical space diffusion model including coarse-scale branches and fine-scale branches, apply random noise to residuals at each scale to obtain noisy residual states, calculate prior noise estimates, establish conditional transmission paths for transmitting structural information from coarse-scale branches to fine-scale branches, and obtain ocean structure generation priors through reverse denoising training.
[0062] S4 obtains multi-time background conditions based on the actual background field and calculates satellite observation conditions using multi-source satellite remote sensing data.
[0063] S5 constructs dynamic sparse observation features based on field observation data and calculates the scale effect coefficients of each field observation on different scale diffusion branches.
[0064] S6 uses the scale action coefficient to weight the dynamic sparse observation features to obtain the weighted dynamic sparse observation features. The multi-time background conditions, satellite observation conditions and weighted dynamic sparse observation features are input into the corresponding diffusion branches, which together with the ocean structure generation prior constrain the generation of residuals at each scale.
[0065] S7. Based on the loss function, the trainable parameters of the conditional propagation path and diffusion generation branch are updated to obtain the trained multi-source observation-guided cascade diffusion model.
[0066] S8, guided by multi-source observations, the cascaded diffusion model performs reverse diffusion from coarse to fine scales, synthesizing the residuals generated at each scale with the actual background field to obtain a single downscaling member, and repeating sampling to determine the final downscaling result and uncertainty.
[0067] Specifically, step S1 includes the following steps: S1.1, the data acquired for the target sea area includes: multi-source remote sensing data, on-site observation data, time attribute conditions, static auxiliary conditions, historical first-grid precision ocean element field and corresponding historical second-grid precision background field, with the first grid precision being higher than the second grid precision. The first grid precision is high resolution, and the second grid precision is low resolution.
[0068] The data used in this invention includes at least a historical high-resolution ocean feature field and a corresponding low-resolution background field. Depending on the specific application conditions, it may also include satellite remote sensing observations, fixed-site observations, mobile platform observations, temporal attributes, and static auxiliary conditions. The low-resolution background field can originate from ocean numerical models, reanalysis products, operational forecast products, or be obtained from high-resolution historical fields after degradation processing. Satellite remote sensing observations can be continuous coverage data, partially missing data, or data with different spatial resolutions. Field observations can originate from moorings, buoys, underwater gliders, survey vessels, shore-based stations, and other ocean observation platforms.
[0069] S1.2, Preprocessing includes: time matching, spatial alignment, outlier handling, land-sea masking, unit unification, and normalization for data from different sources; for data with known observation errors, the error magnitude is determined in the original physical units, and the observations and error information are converted to the data range used by the cascaded physical spatial diffusion model according to the same rules. Static auxiliary conditions include information such as water depth, land-sea distribution, distance from the shore, seabed slope, and geographical location. The types and number of static conditions are selected according to the target sea area and target elements; it is not required to use all conditions simultaneously.
[0070] Specifically, step S2 includes the following steps: S2.1, based on standardized data, let the first... The historical first grid-precision ocean element field at that moment was Background component for: ; Among them, subscript Represents the background, and the background degradation operator is: The background resampling operator is . Background components that are consistent with the target field mesh. Background degradation operator. This may include low-pass, downsampling, spatial averaging, sensor response simulation, or combinations thereof; background resampling operators. Used to convert degraded low-resolution fields to the target mesh.
[0071] S2.2, the first The residuals of the first-time historical grid-accuracy ocean feature field relative to the background components: ; in, For the first The residuals of the historical first-grid-precision ocean feature field relative to the background components at a given time point cannot be resolved by the historical second-grid-precision background field. This invention preferably uses diffusion modeling on these residuals, rather than directly generating a complete high-resolution field. The background components are responsible for providing the large-scale ocean state, while the detailed residuals are used to describe fronts, vortex edges, vortex filaments, and local anomalies. In other embodiments, anomaly fields after trend removal, climatological removal, or local mean removal can also be modeled. The formation of the background components is not limited to the aforementioned degradation processes.
[0072] S2.3, to describe ocean anomalies at different spatial scales, decomposes the residuals and sets... Individual scale decomposition operator , The number of scale layers is not less than 2. For the first The scale decomposition operator; the first Individual scale residuals for: ; The residuals at each scale together constitute the multi-scale residuals. : ; in, This indicates that approximate errors within a preset range are allowed during the scaling and reconstruction process.
[0073] S2.4, divides multi-scale residuals into coarse-scale residuals. and fine-scale residuals : ; Among them, superscript Indicates a coarse scale, superscript This represents a fine-scale representation. Coarse-scale residuals primarily describe wider fronts, vortex edges, and regional anomaly zones, while fine-scale residuals mainly describe slender vortex filaments, frontal sharpening, and localized small-scale variations. Scale decomposition can employ spatial filtering, frequency decomposition, wavelet transform, multi-resolution transform, learnable decomposition, or combinations thereof. Different scales can be completely complementary or retain appropriate overlap. The number of scale layers and the cutoff scale are determined based on background resolution, target resolution, and the dynamic characteristics of the target sea area.
[0074] Specifically, step S3 includes the following steps: S3.1, Construct a cascaded physical space diffusion model including coarse-scale and fine-scale branches; assuming For the first The true residuals at each scale without noise. For the first The first scale The residual states at each diffusion time step are given, and random noise is applied to the residuals at each scale. The forward diffusion process is as follows: ; in, For the first The retention factor of the original residual for each diffusion time step. For the first The noise addition factor for each diffusion time step, Represents random noise. Coefficients and It can be determined by discrete noise sequences, continuous noise functions, or other diffusion patterns. Random noise can be Gaussian noise or other noise forms suitable for the statistical characteristics of the target marine elements.
[0075] S3.2, p. Prior noise estimation at each scale for: ; Among them, subscript Indicates prior knowledge; For the first A cascaded physical space diffusion model at multiple scales; These are trainable parameters; These are static auxiliary conditions or time-related attribute conditions; Provides structural information for other scales. Diffusion denoising models can employ convolutional structures, attention structures, transformer structures, or combinations thereof. Diffusion parameters can be in the form of noise prediction, original residual prediction, velocity prediction, or other equivalent forms.
[0076] In section S3.3, during the prior training phase of ocean structure analysis, the actual low-resolution background and real-time observations at the target time are not input, or these conditions are masked according to preset probabilities. The model can retain auxiliary conditions that do not directly determine the phase of the target structure, such as land-sea masking, month, season, geographical location, and static topography. By restricting the use of real-time conditions, the diffusion model learns the morphological distribution of structures such as fronts, vortices, and vortex filaments from historical high-resolution ocean fields, rather than downscaling by replicating the low-resolution background. A conditional transfer path is established to transmit structural information from the coarse-scale branch to the fine-scale branch: ; in, For the first Each scale residual For the first Multi-scale structure extraction operators at various scales. To be passed to the first Structural information at multiple scales. Cross-scale structure extraction can employ feature extraction, conditional modulation, attention computation, or other methods capable of conveying spatial structure, maintaining connections between fine-scale structures and larger-scale fronts and vortices. Fine-scale branches can also be trained independently first, and then jointly trained to establish connections with coarser-scale structures.
[0077] Specifically, step S4 includes the following steps: S4.1, Obtain the actual background field at the target time and the time adjacent to the target time. The actual background field is the actual low-resolution background field. Calculate the background conditions at multiple time intervals: ; in, For the first Background conditions at any given moment; superscript Indicates the actual numerical pattern, reanalysis data product, or business context; offset relative to the target time The actual background scene; A set of time offsets; Background condition extraction operators are used. Multi-temporal backgrounds are used to describe the large-scale ocean state, the location of major structures, and their temporal variation trends. Background conditions preferentially apply to coarser-scale diffusion branches. For finer-scale branches, the intensity of background conditions can be reduced, or only scale-processed background features can be passed. The prior parameters of ocean structures can remain entirely unchanged during background-guided training, or only some parameters can be adjusted. Alternatively, an independent background correction component can be established to reduce the impact of background guidance on the original structure generation capability.
[0078] S4.2, let the first... Multi-source satellite remote sensing data at a given time are The satellite observation conditions are: ; Among them, superscript Indicates satellite remote sensing, For the first Satellite observation conditions at that moment, Extracting operators for satellite observation conditions; To effectively cover the mask, This operator takes into account observation error information, including satellite spatial resolution, sensor response, effective coverage, missing locations, and observation errors. Satellite remote sensing observations are primarily used to correct for the locations of fronts, vortices, and anomalies that are consistent with their resolution capabilities. For smaller-scale structures that satellites cannot effectively resolve, the generated results are not required to conform to satellite observations pixel-by-pixel; instead, they are generated by a fine-scale diffusion prior under larger-scale constraints. Satellite observation errors are preferably determined according to physical units before normalization. When constructing pseudo-satellite observations, resolution degradation, spatial missing locations, coverage variations, and observation errors can be simulated simultaneously.
[0079] Specifically, step S5 includes the following steps: S5.1, let the first... The field observation value is The relationship between this and the historical first-grid precision ocean feature field is as follows: ; in, For the first The observation operator corresponding to each observation, For the first Measurement error of the observation; For the first The first-ever grid-precision oceanographic feature field at any given moment. For fixed stations, the observation operator... It can represent the spatial mapping from the target grid to a fixed location; for underwater gliders, the observation operator can represent the spatiotemporal mapping from the target grid to the movement trajectory; for observations with a certain spatial coverage, the observation operator can also include area averaging, spatial response, or other device observation processes.
[0080] S5.2, each field observation is converted into dynamic sparse observation features: ; in, For the first Dynamic sparse observation characteristics formed by a number of observations; Operator for extracting dynamic sparse observation features; For the observation location; The time interval between the observation time and the target time; The standard deviation of the observation error; Category of observation source; This serves as the representative scale for observations. Comprehensive observation characteristics may also include platform status, direction of movement, observation depth, quality indicators, and other attributes that describe the reliability and representativeness of the observations.
[0081] S5.3, for observations at fixed locations such as moorings and buoys, the observation sequence is divided into slowly varying components and short-term varying components: ; Among them, superscript Indicates slowly varying components, superscript It represents the short-term variation component.
[0082] The slowly varying component is: ; The short-time variation components are: ; in, For the first A time-scale separation operator is used for fixed observations. Moving average, low-pass filtering, trend estimation, or other time decomposition methods can be employed. Slow-varying components are primarily used to correct local background and coarse-scale residuals, while short-time varying components are mainly used to constrain fine-scale residuals. The time window and attenuation scale are determined based on the observation frequency, target elements, and the dynamic time scale of the target sea area. For mobile platforms such as underwater gliders, the change in observation position over time is preserved, allowing the movement trajectory to constrain the position, orientation, and trend of local structures.
[0083] S5.4, Calculate the first [item] based on the observation source, spatial resolution, error magnitude, and representative scale. The observation on the first The effect coefficient of each scale: ; in, For the first The observation on the first The effect coefficient of each scale For the first The function for determining the observation action coefficients at each scale; For the first The background conditions at each moment; For the first Satellite observation conditions at a given moment.
[0084] The action coefficient satisfies: .
[0085] In other implementations, normalization may not be performed; only a reasonable range for the scale action coefficients may be set. The scale action coefficients can be determined by preset rules, learned by the model, or jointly determined by preset rules and trainable parameters. Low-resolution background typically acts primarily on the coarse-scale branch; satellite observations act on scales consistent with their resolution; slowly varying components from fixed observations primarily correct the coarse-scale state; short-term varying components and motion trajectory observations can have a strong effect on the fine-scale branch.
[0086] Specifically, step S6 includes the following steps: S6.1, let the first... The current spatial characteristics at each scale are The correction amount generated by the multi-time background conditions is: ; The correction amount caused by satellite observation conditions is: ; The correction amount generated by the on-site observation is: ; in, , and These are multi-temporal background conditions, satellite observation conditions, and on-site observation conditions for the first time. The correction amount generated by each scale branch; , and These are the operators that apply to the corresponding conditions; For the first The background conditions at each moment; For the first Satellite observation conditions at a given moment; For the first The dynamic sparse observation features formed by the observations. The action operator can be feature concatenation, conditional modulation, attention computation, neighborhood weighting, kernel function propagation, or other methods that can establish a relationship between conditional information and spatial features.
[0087] S6.2, an optional form of conditional noise estimation is as follows: ; in, For the first The final noise estimate after incorporating multi-source conditions at each scale; , and These are the conditional factors of background, satellite remote sensing, and field observation in the diffusion time step, respectively. For the first Prior noise estimation at various scales. The conditional action coefficient can vary with scale level, diffusion time step, observation error, and observation timeliness. To prevent excessive correction from a single observation, the magnitude of the action coefficient or correction amount can be limited. The above formula represents one possible implementation method. Multi-source conditions can also be applied to intermediate features of the diffusion model, noise-free residual estimates, or reverse diffusion states. In-situ observations do not directly overwrite the corresponding grid values in the generated field, but rather implement flexible corrections to the diffusion process through observational features to maintain continuity between the observation location and the surrounding ocean structure.
[0088] S6.3 generates prior prediction output based on the current diffusion state and ocean structure, recovering the first... Noise-free residuals at each scale : ; in, In order to be with the first Each diffusion time step and the corresponding recovery operator for the diffusion form; For the first The first scale The residual state at each diffusion time step. This recovery operator can calculate the original residual based on noise estimation, or it can directly output the residual estimate from the denoising model. The coarse-scale residual estimate is used to constrain the subsequent fine-scale generation, while the fine-scale residual estimate supplements smaller-scale changes while maintaining the continuity of the coarse-scale structure.
[0089] Specifically, step S7 includes the following steps: S7.1 The loss function includes: diffusion base loss and auxiliary loss. The auxiliary loss includes: residual reconstruction loss, low-pass consistency loss, frequency band structure loss, frequency band energy loss, spatial gradient loss, power spectrum loss, background consistency loss and observation consistency loss.
[0090] S7.2, the total loss is : ; in, For diffusion-based losses; For the first Class structure auxiliary loss; This is the weight of the auxiliary loss; The loss adjustment function is determined based on the diffusion time step or signal-to-noise ratio; Loss of observation consistency; Weights for observation consistency loss; Background consistency loss; The weights for background consistency loss.
[0091] S7.3, the diffusion fundamental loss is: ; in, for The error measurement function used for each scale; Indicates random noise; For the first The final noise estimate after incorporating multi-source conditions at each scale can be calculated using mean squared error, absolute error, robust error, or other metrics suitable for diffusion training.
[0092] Based on the noise-free residuals at each scale recovered in step S6.3, the noise-free residuals at each scale are synthesized with the background components to obtain the ocean element field estimate for the training phase: ; in, Estimate the ocean element field recovered during the training phase; For background components; For the recovery of the first Noise-free residuals at each scale.
[0093] The observation consistency loss is: ; in, For the first The weight of an observation in the loss is determined by the observation error, its source, and its timeliness. In order to be with the first An error measurement function adapted to the characteristics of the observed error; For the first The observation operator corresponding to each observation; For the first Item of field observation. Background scale consistency loss. This is used to compare the difference between the generated field after background observation or scale transformation and the input background; its specific form is not limited. For example, spatial gradient loss can be expressed as: ; in, For spatial gradient operators, and These are the gradient magnitudes of the generating field and the target field, respectively. Let be the gradient error metric function. The spectral consistency loss can be expressed as: ; in, Indicates the first Statistical measures for a frequency band or wavenumber range This indicates the error metric function used in this frequency band. The structure-assisted loss is preferably adjusted based on the diffusion time step or signal-to-noise ratio, enhancing its effect when spatial structure information is relatively reliable and reducing its effect under high noise conditions. Different loss weights can be used for different scale branches, allowing the coarse-scale branch to prioritize preserving the structure position, while the fine-scale branch prioritizes preserving the gradient, spectrum, and local anomaly energy.
[0094] S7.4 After updating the trainable parameters of the conditional propagation path and diffusion generation branch based on the loss function, the trained multi-source observation-guided cascade diffusion model is obtained.
[0095] Specifically, step S8 includes the following steps: S8.1, based on a multi-source observation-guided cascade diffusion model, performs back diffusion from coarse to fine scale and vice versa. The coarse and fine scales are obtained by dividing the residuals according to their spatial frequency characteristics. Specifically, the two-dimensional power spectrum or radial average power spectrum of the historical residual field can be calculated first. One or more scale boundaries are determined based on the spectral energy distribution, spectral slope change, effective resolution of the background field, and spatial resolution of the target field. Then, low-pass filtering, band-pass filtering, wavelet decomposition, or other scale decomposition operators are used to divide the residuals into coarse-scale residuals with lower spatial frequencies and fine-scale residuals with higher spatial frequencies.
[0096] No. The scale in the th... The update process for each time step is as follows: ; in, The updated diffusion state; For the first The back-diffusion update operator at each time step, For the first Optional random terms used at each time step; For the first The first scale The residual state at each diffusion time step; For the first The final noise estimate after incorporating multi-source conditions at each scale. The sampling process is preferably performed in order from coarse to fine scales. After generating the coarse-scale residuals, they are used as the structural conditions for the fine-scale branches. For more than three scale levels, generation can be performed layer by layer from largest to smallest scale, or an alternating correction or partially parallel approach can be used.
[0097] S8.2, a single downscaling member is: ; in, For the first Descaled members at any given moment; This represents the actual background field on the target mesh. For the first Noise-free residuals generated at each scale; For the first The scale amplitude coefficients can be fixed at one or determined using independent validation data. Once determined, the parameters remain unchanged during testing and practical applications. In implementations without setting independent amplitude coefficients, the physical amplitude of the residuals at each scale can also be learned directly from the diffusion model.
[0098] S8.3, under the same background and observation conditions, using different random initial states to perform... Second sampling, obtained High-resolution downscaling member: ; in, The total number of members in the set. For the set member index, For the first One downscaling member.
[0099] S8.4 uses the ensemble mean to represent the final downscaling result, i.e., the final oceanographic field; the formula for calculating the ensemble mean is: ; in, For the first The average value of all members of the set at any given time.
[0100] S8.5 uses set dispersion to represent uncertainty. The formula for calculating set dispersion is: ; in, For the first The ensemble dispersion at each time step represents the possible range of variation for different fine-scale structures. The ensemble mean provides a comprehensive downscaling result, while ensemble dispersion, quantile range, or other ensemble statistics characterize spatial uncertainty. Acronymic differences are typically small in well-observed regions, while ensemble differences in sparsely observed regions can reflect the possible range of various reasonable fine-scale structures.
[0101] To illustrate the implementation effect of the present invention, this embodiment uses sea surface temperature as the target ocean element and utilizes low-resolution sea surface temperature background field as well as satellite, mooring and underwater glider observations for cascade diffusion downscaling. Figure 2 , Figure 3 and Figure 4 The typical downscaling results from September 7, 2023 are shown. Figure 2 The sea surface temperature field before downscaling; Figure 3 The sea surface temperature field after downscaling using the multi-source observation-guided cascade diffusion downscaling method of the present invention; Figure 4 This is a high-resolution true-value sea surface temperature field. For example... Figure 2 The low-resolution background field shown primarily reflects a smooth, large-scale temperature distribution. After cascaded diffusion downscaling guided by multi-source observations, as... Figure 3 The generated result shown restores the relationship with... Figure 4 The target fields shown have basically the same frontal position, shape, and local temperature changes.
[0102] like Figure 5 , Figure 6 , Figure 7 and Figure 8 The results of the cascade diffusion generation of sea surface temperature residuals are presented, namely the target components and generated components of sea surface temperature residuals at different spatial scales on September 7, 2023. Figure 5 For the target coarse-scale residual plot; Figure 6 This is a coarse-scale residual map generated using the multi-source observation-guided cascade diffusion downscaling method for ocean element fields according to the present invention; Figure 7 For the target fine-scale residual plot; Figure 8 This is a fine-scale residual map generated using the multi-source observation-guided cascade diffusion downscaling method for ocean element fields according to the present invention.
[0103] Coarse-scale residuals reflect wider fronts and regional anomalies, while fine-scale residuals reflect local gradients and smaller-scale variations. For example... Figure 6 and Figure 8 The results generated by the multi-source observation-guided cascade diffusion downscaling method for ocean element fields shown in this invention are respectively related to the following in terms of main structural location, extension direction, and spatial morphology: Figure 5 and Figure 7The corresponding targets shown are basically consistent, indicating that cascade diffusion can learn ocean anomalies at different scales and maintain the coordination between coarse and fine scale structures.
[0104] This invention provides a multi-source observation-guided cascaded diffusion downscaling method for ocean element fields, used to convert low-resolution ocean background fields into ocean element fields with higher spatial resolution. This method is applicable to sea surface temperature, sea surface height, sea surface salinity, ocean current fields, and other ocean elements with multi-scale spatial variation characteristics. This invention aims to address the problems of weak fine-scale structure, fragmented spatial morphology, excessive reliance on low-resolution background, and mixed effects of multi-source observation scales in existing ocean element field downscaling methods. It enables diffusion models to learn structural distributions with ocean spatial continuity even without real-time background and observational constraints. During the downscaling process, it uses the low-resolution background to determine large-scale states, uses satellite observations to correct the position of continuous structures, and uses multi-source field observations such as moorings, buoys, and underwater gliders to constrain local variations. Ultimately, it obtains a high-resolution ocean element field with reasonable numerical values, consistent spectrum, continuous morphology, and the ability to express uncertainties.
[0105] This invention directly generates ocean anomalies that the background field could not resolve within a grid space with clear physical meaning, without relying on latent space compression of the complete ocean field. Its basic idea is to separate the learning of ocean structure distribution from the constraints of real-time background and observation. The model first uses historical high-resolution ocean feature fields to learn spatially continuous ocean structure generation priors, and then uses low-resolution background, satellite remote sensing, and in-situ observations to correct the position, orientation, phase, and amplitude of the generated structures.
[0106] This invention proposes a multi-source observation-guided cascaded diffusion downscaling method for ocean feature fields, directly generating and modeling ocean residuals that cannot be resolved by the low-resolution background field in the physical grid space. This method first utilizes historical high-resolution ocean feature fields to learn ocean structure generation priors independent of real-time background and observations. Then, it divides the detailed residuals into coarse-scale and fine-scale components, recovering wider fronts, vortex edges, slender vortex filaments, and local anomalies through a coarse-to-fine diffusion process. This alleviates the limitations of low-resolution background on detail generation and maintains spatial coordination between structures at different scales. Based on this, differentiated guidance is implemented according to the coverage, error, timeliness, and representative scale of satellite remote sensing, fixed-site, and mobile platform observations. This ensures that the background and observations are primarily used to correct the position, phase, and amplitude of the generated structures, rather than directly replacing local grid values, helping to reduce spatial abrupt changes caused by sparse observations. Leveraging the multiple sampling capability of the diffusion model, multiple physically reasonable high-resolution members can be generated under the same background and observation conditions, further providing ensemble mean and spatial uncertainty, offering a new technical approach for refined analysis of ocean features, forecast correction, and subsequent data assimilation.
[0107] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.
Claims
1. A multi-source observation-guided cascade diffusion downscaling method for ocean element fields, characterized in that, Specifically, the steps include the following: S1: Acquire marine element fields, multi-source observation data, and static data of the target sea area, perform preprocessing, and obtain standardized data; S2, construct background components based on standardized data, calculate residuals, and decompose residuals to obtain multi-scale residuals; S3. Establish corresponding diffusion generation branches for residuals at each scale, construct a cascaded physical space diffusion model including coarse-scale branches and fine-scale branches, apply random noise to residuals at each scale to obtain noisy residual states, calculate prior noise estimates, establish a conditional transmission path for transmitting structural information from coarse-scale branches to fine-scale branches, and obtain ocean structure generation priors through reverse denoising training. S4, based on the actual background field, obtains background conditions at multiple times, and uses multi-source satellite remote sensing data to calculate satellite observation conditions; S5, construct dynamic sparse observation features based on field observation data, and calculate the scale effect coefficients of each field observation on different scale diffusion branches; S6 inputs the multi-time background conditions, satellite observation conditions and weighted dynamic sparse observation features into the corresponding diffusion branch, which together with the ocean structure generation prior constrains the generation of residuals at each scale; S7. Based on the loss function, update the trainable parameters of the conditional propagation path and the diffusion generation branch to obtain the trained multi-source observation-guided cascade diffusion model. S8, guided by multi-source observations, the cascaded diffusion model performs reverse diffusion from coarse to fine scales, synthesizing the residuals generated at each scale with the actual background field to obtain a single downscaling member, and repeating sampling to determine the final downscaling result and uncertainty.
2. The multi-source observation-guided cascade diffusion downscaling method for ocean element fields according to claim 1, characterized in that, Step S1 specifically includes the following steps: S1.1 The data acquired for the target sea area includes: multi-source remote sensing data, on-site observation data, time attribute conditions, static auxiliary conditions, historical first grid precision ocean element field and corresponding historical second grid precision background field, with the first grid precision being higher than the second grid precision; S1.2, Preprocessing includes: time matching, spatial alignment, outlier handling, land-sea masking, unit unification and normalization of data from different sources; for data with known observation errors, the error magnitude is determined in the original physical units, and the observations and error information are converted to the data range used by the cascaded physical space diffusion model according to the same rules.
3. The multi-source observation-guided cascade diffusion downscaling method for ocean element fields according to claim 1, characterized in that, Step S2 specifically includes the following steps: S2.1, based on standardized data, let the first... The historical first grid-precision ocean element field at that moment was Background component for: ； Among them, subscript Represents the background, and the background degradation operator is: The background resampling operator is ; S2.2, the first The residuals of the first-time historical grid-accuracy ocean feature field relative to the background components: ； in, For the first The residuals of the first-precision ocean feature field at a given moment relative to the background components; S2.3, decompose the residuals and set... Individual scale decomposition operator , The number of scale layers is not less than 2. For the first The scale decomposition operator; the first Individual scale residuals for: ； The residuals at each scale together constitute the multi-scale residuals. : ； in, This indicates that approximation errors within a preset range are allowed during the scaling and reconstruction process; S2.4, divides multi-scale residuals into coarse-scale residuals. and fine-scale residuals : ； Among them, superscript Indicates a coarse scale, superscript Indicates a finer scale.
4. The multi-source observation-guided cascade diffusion downscaling method for ocean element fields according to claim 1, characterized in that, Step S3 specifically includes the following steps: S3.1, Construct a cascaded physical space diffusion model including coarse-scale and fine-scale branches; assuming For the first The true residuals at each scale without noise. For the first The first scale The residual states at each diffusion time step are given, and random noise is applied to the residuals at each scale. The forward diffusion process is as follows: ； in, For the first The retention factor of the original residual for each diffusion time step. For the first The noise addition factor for each diffusion time step, Indicates random noise; S3.2, p. Prior noise estimation at each scale for: ； Among them, subscript Indicates prior knowledge; For the first A cascaded physical space diffusion model at multiple scales; These are trainable parameters; These are static auxiliary conditions or time-related attribute conditions; Structural information provided for other scales; S3.3, In the prior training phase of marine structures, establish a conditional transfer path for transmitting structural information from coarse-scale branches to fine-scale branches: ； in, For the first Each scale residual For the first Multi-scale structure extraction operators at various scales. To be passed to the first Structural information at each scale.
5. The multi-source observation-guided cascade diffusion downscaling method for ocean element fields according to claim 1, characterized in that, Step S4 specifically includes the following steps: S4.1, Obtain the actual background field at the target time and the time adjacent to the target time, and calculate the background conditions at multiple times: ； in, For the first Background conditions at any given moment; superscript Indicates the actual numerical pattern, reanalysis data product, or business context; offset relative to the target time The actual background scene; A set of time offsets; Extract operators for background conditions; S4.2, let the first... Multi-source satellite remote sensing data at a given time are The satellite observation conditions are: ； Among them, superscript Indicates satellite remote sensing, For the first Satellite observation conditions at that moment, Extracting operators for satellite observation conditions; To effectively cover the mask, This is information about observation errors.
6. The multi-source observation-guided cascade diffusion downscaling method for ocean element fields according to claim 5, characterized in that, Step S5 specifically includes the following steps: S5.1, let the first... The field observation value is The relationship between this and the historical first-grid precision ocean feature field is as follows: ； in, For the first The observation operator corresponding to each observation, For the first Measurement error of the observation; For the first The first-ever grid-precision ocean element field at any given moment in history; S5.2, each field observation is converted into dynamic sparse observation features: ； in, For the first Dynamic sparse observation characteristics formed by a number of observations; Operator for extracting dynamic sparse observation features; For the observation location; The time interval between the observation time and the target time; The standard deviation of the observation error; Category of observation source; For observational representative scale; S5.3, for observations at fixed locations such as moorings and buoys, the observation sequence is divided into slowly varying components and short-term varying components: ； Among them, superscript Indicates slowly varying components, superscript Indicates short-time variation components; The slowly varying component is: ； The short-time variation components are: ； in, For the first A time-scale separation operator for fixed observations; S5.4, Calculate the first... The observation on the first The effect coefficient of each scale: ； in, For the first The observation on the first Each scale action coefficient For the first The function for determining the observation action coefficients at each scale; For the first The background conditions at each moment; For the first Satellite observation conditions at a given moment; The action coefficient satisfies: 。 7. The multi-source observation-guided cascade diffusion downscaling method for ocean element fields according to claim 1, characterized in that, Step S6 specifically includes the following steps: S6.1, let the first... The current spatial characteristics at each scale are The correction amount generated by the multi-time background conditions is: ； The correction amount caused by satellite observation conditions is: ； The correction amount generated by the on-site observation is: ； in, , and These are multi-temporal background conditions, satellite observation conditions, and on-site observation conditions for the first time. The correction amount generated by each scale branch; , and These are the operators that apply to the corresponding conditions; For the first The background conditions at each moment; For the first Satellite observation conditions at a given moment; For the first Dynamic sparse observation characteristics formed by a number of observations; S6.2, the conditional noise estimation form is: ； in, For the first The final noise estimate after incorporating multi-source conditions at each scale; , and These are the conditional factors of background, satellite remote sensing, and field observation in the diffusion time step, respectively. For the first Prior noise estimation at various scales; S6.3 generates prior prediction output based on the current diffusion state and ocean structure, recovering the first... Noise-free residuals at each scale : ； in, In order to be with the first Each diffusion time step and the recovery operator corresponding to the diffusion form; For the first The first scale The residual state of each diffusion time step.
8. The multi-source observation-guided cascade diffusion downscaling method for ocean element fields according to claim 1, characterized in that, Step S7 specifically includes the following steps: S7.1, the loss function includes: diffusion base loss and auxiliary loss. The auxiliary loss includes: residual reconstruction loss, low-pass consistency loss, frequency band structure loss, frequency band energy loss, spatial gradient loss, power spectrum loss, background consistency loss and observation consistency loss. S7.2, the total loss is : ； in, For diffusion-based losses; For the first Class structure auxiliary loss; This is the weight of the auxiliary loss; The loss adjustment function is determined based on the diffusion time step or signal-to-noise ratio; Loss of observation consistency; Weights for observation consistency loss; Background consistency loss; The weights for background consistency loss; S7.3, the diffusion fundamental loss is: ； in, for The error measurement function used for each scale; Indicates random noise; For the first The final noise estimate after incorporating multi-source conditions at each scale; Based on the noise-free residuals at each scale recovered in step S6.3, the noise-free residuals at each scale are synthesized with the background components to obtain the ocean element field estimate for the training phase: ; in, Estimate the ocean element field recovered during the training phase; For background components; For the recovery of the first Noise-free residuals at each scale; The observation consistency loss is: ； in, For the first The weight of each observation in the loss; In order to be with the first An error measurement function adapted to the characteristics of the observed error; For the first The observation operator corresponding to each observation; For the first Item field observation values; S7.4 After updating the trainable parameters of the conditional propagation path and diffusion generation branch based on the loss function, the trained multi-source observation-guided cascade diffusion model is obtained.
9. The multi-source observation-guided cascade diffusion downscaling method for ocean element fields according to claim 1, characterized in that, Step S8 specifically includes the following steps: S8.1, based on a multi-source observation-guided cascade diffusion model, performs reverse diffusion from coarse scale to fine scale. The coarse and fine scales are obtained by dividing them according to the spatial frequency characteristics of the residuals. The spatial frequency of the coarse scale residuals is lower than that of the fine scale residuals. No. The scale in the th... The update process for each time step is as follows: ； in, The updated diffusion state; For the first The back-diffusion update operator at each time step, For the first Optional random terms used at each time step; For the first The first scale The residual state at each diffusion time step; For the first The final noise estimate after incorporating multi-source conditions at each scale; S8.2, a single downscaling member is: ； in, For the first Descaled members at any given moment; This represents the actual background field on the target mesh. For the first Noise-free residuals generated at each scale; For the first Amplitude coefficients at each scale; S8.3, under the same background and observation conditions, using different random initial states to perform... Second sampling, obtained One downscaling member: ； in, The total number of members in the set. For the set member index, For the first One downscaled member; S8.4 uses the ensemble mean to represent the final downscaling result. The formula for calculating the ensemble mean is: ； in, For the first The average value of all members of the set at any given time; S8.5 uses set dispersion to represent uncertainty. The formula for calculating set dispersion is: ； in, For the first The set discreteness at each time point.