CN-121981922-A - Sea temperature completion method and system based on space-time priori and flow matching generation model

CN121981922ACN 121981922 ACN121981922 ACN 121981922ACN-121981922-A

Abstract

The invention belongs to the technical field of sea temperature completion, in particular to a sea temperature completion method and a sea temperature completion system based on a space-time priori and flow matching generation model, wherein the method comprises the following steps of inputting sea surface continuous four-day sea surface temperature SST data Zhou Junzhi SST data And mask code The method comprises the steps of performing rough completion of space-time residual learning, introducing a light weight 3D-CNN space-time residual learning mechanism to obtain a rough complete graph, explicitly modeling a space-time evolution rule of sea temperature data, introducing a condition flow matching generation strategy based on optimal transmission in complete graph restoration generated by condition flow matching, performing refined reconstruction on the rough complete graph to obtain a complete sea surface temperature image, replacing a curved and random Markov chain sampling path in a traditional denoising diffusion probability model DDPM, solving the problem of low reasoning efficiency caused by relying on hundreds of steps of random denoising, and realizing high-efficiency reasoning under the condition of fewer steps.

Inventors

WANG YANSHAN
NIE JIE
CHEN RUIZI
Guo Chenyou
Zuo Zijie
ZHENG NAN
ZHAO YUTING
Song derui

Assignees

中国海洋大学

Dates

Publication Date: 20260505
Application Date: 20260409

Claims (7)

1. The sea Wen Buquan method for generating the model based on space-time priori and stream matching is characterized by comprising the following steps: s1, data input: Inputting SST data of sea surface for several continuous days Zhou Junzhi SST data And mask code Wherein the SST data Including historical timing SST data And a broken image on the same day ; S2, rough complement based on space-time residual error learning: The SST data With Zhou Junzhi SST data after broadcast Dynamic residual decoupling is carried out to obtain a dynamic abnormal sequence The dynamic abnormal sequence Splicing the mask with the expanded mask along the channel, and sending the spliced mask into a 3D convolutional encoder to obtain a preliminary complement image through space-time residual error reconstruction Then the mask is used to make up the preliminary full image Replacement of pixels of non-missing regions with a current day broken sea surface temperature image Generates a rough complement diagram ; S3, forward training of full graph restoration is generated by matching the conditional flows: the rough complement diagram And the Zhou Junzhi SST data As a conditional input, a linear interpolation path from pure noise to real image is constructed, from which a target velocity field is calculated And at the target speed field For supervision, multi-condition guided speed field prediction by training UNet network, output predicted speed field ; S4, reverse reasoning of the full graph restoration is generated by matching the conditional flows: solving the ordinary differential equation ODE by using the trained UNet network, and completing the graph roughly Carrying out refined reconstruction to obtain a complete sea surface temperature image 。
2. The method of claim 1, wherein the SST data With Zhou Junzhi SST data after broadcast Dynamic residual decoupling is carried out to obtain a dynamic abnormal sequence The dynamic abnormal sequence Splicing the mask with the expanded mask along the channel, and sending the spliced mask into a 3D convolutional encoder, wherein the method specifically comprises the following steps of: s2.1 SST data is processed And broadcast along the time dimension Subtracting from element to obtain dynamic abnormal sequence reflecting abnormal time sequence change ; S2.2 dynamic anomaly sequence Increasing channel dimension while masking Copying along a time dimension, the dynamic anomaly sequence Splicing with the mask in the channel dimension to form an input tensor of the 3D convolutional encoder, wherein the mask For the image of the current day breakage A binary mask is defined for identifying the missing and non-missing regions of pixel values.
3. The method of claim 2, wherein the 3D convolutional encoder obtains the preliminary complement image by spatio-temporal residual reconstruction Then the mask is used to make up the preliminary full image Replacement of pixels of non-missing regions with a current day broken sea surface temperature image Generates a rough complement diagram Comprising: S2.3, the 3D convolution encoder gradually expands the channel number and outputs a characteristic diagram Expressed as: ; Wherein, the In the form of a 3D convolution layer, For dynamic anomaly sequences Sum mask Splicing in the channel dimension to form an input tensor of the 3D convolutional encoder; s2.4 the feature map Carrying out average pooling in the time dimension, compressing time sequence information and reserving a space structure to obtain a collapsed space characteristic diagram ; S2.5 spatial feature map after collapse Input 2D convolution decoder for predicting and inputting current date broken image Temperature anomaly residual map with consistent image space size ; S2.6 utilizing the cycle average sea temperature data Reconstructing to obtain a preliminary complement image The calculation formula is as follows: ; S2.7, using the mask to make up the preliminary full image Replacement of pixels of non-missing regions with a current day broken sea surface temperature image Generates a rough complement diagram Expressed as: ; Wherein, the Is the inverse of the mask, is used to mark the missing region, Representing element-wise multiplication.
4. The method of claim 1, wherein a linear interpolation path from pure noise to real image is constructed, from which path a target velocity field is calculated And at the target speed field For supervision, multi-condition guided speed field prediction by UNet network, output speed field Comprising the following steps: S3.1 sampling noise from a Standard normal distribution and imaging with a real day SST image Defining the current state of flow matching using linear interpolation paths as a benchmark Expressed as: ; Wherein, the For time steps evenly sampled between [0,1], when The state is pure noise when The time-state is a real image and, For gaussian noise sampled from a standard normal distribution, Representing scalar multiplication; s3.2, calculating a target speed field according to the path The calculation formula is as follows: ; s3.3 at the target speed field To supervise the current state Rough complement map Mean sea temperature data Connecting in the dimension of the characteristic channel, and training through a UNet network to output a predicted speed field in combination with the time step t The calculation formula is as follows: ; Wherein, the A parameterized UNet network is represented, Representing all the learnable weight parameters in the network, Representing the join operation in the feature channel dimension.
5. The method of claim 4, wherein solving the ordinary differential equation ODE versus coarse complement map using a trained UNet network Carrying out refined reconstruction to obtain a complete sea surface temperature image Comprising: By using preset redrawing intensity super-parameters Step the starting time Initial image state initialization at time as rough complement map And in time intervals of The ODE track of the internal edge ordinary differential equation is circulated and updated iteratively, and in each iteration, the UNet network predicts the speed field And update the state until Ending the moment to obtain a final reconstructed complete sea surface temperature image 。
6. The method of claim 5, wherein the intensity super-parameters are drawn by a predetermined redraw Step the starting time Initial image state initialization at time as rough complement map And in time intervals of The ODE track of the internal edge ordinary differential equation is circulated and updated iteratively, and in each iteration, the UNet network predicts the speed field And update the state until Ending the moment to obtain a final reconstructed complete sea surface temperature image Comprising: s4.1, utilizing preset redrawing intensity super-parameters Step the starting time Initial image state initialization at time as rough complement map The calculation formula is as follows: ; Wherein, the Representing the starting time step of a conditional flow matching model in reverse reasoning Initial image state tensor at the time; S4.2 during the time interval The ODE track of the inner edge ordinary differential equation is circularly and iteratively updated, and in each iteration, the ODE track is based on an output speed field Using a set time step For the current state And updating the time step, and calculating the time step of the next moment And state The update formula is expressed as: ; ; Wherein, the The current time step is indicated and the current time step, For the current time step The corresponding state of the device is that, Indicating the time step at the next moment after the update, The state corresponding to the time step at the next moment; S4.3, calculating the time step of the next moment And state Respectively updating to the current time step and the current state of the next iteration, and then repeatedly executing the predicted speed field step S3.3 and the state updating operation step S4.2 until the condition is met Ending the cycle, and outputting the final reconstructed SST image after the cycle is ended 。
7. The sea Wen Buquan system based on the space-time priori and stream matching generation model is characterized by comprising a rough complement module for space-time residual error learning and a full graph restoration module for conditional stream matching generation; the rough complement module of the space-time residual error learning is used for continuously carrying out SST data on the sea surface for a plurality of days Zhou Junzhi SST data And mask code As input, where SST data Including historical timing SST data And a broken image on the same day ; The SST data With Zhou Junzhi SST data after broadcast Dynamic residual decoupling is carried out to obtain a dynamic abnormal sequence The dynamic abnormal sequence Splicing the mask with the expanded mask along the channel, and sending the spliced mask into a 3D convolutional encoder to obtain a preliminary complement image through space-time residual error reconstruction Then the mask is used to make up the preliminary full image Replacement of pixels of non-missing regions with a current day broken sea surface temperature image Generates a rough complement diagram ; The full-graph restoration module generated by the conditional flow matching calculates a target speed field by constructing a linear interpolation path from pure noise to a real image And at the target speed field For supervision, multi-condition guided speed field prediction by training UNet network, output predicted speed field ; Finally, solving the ordinary differential equation by using the trained UNet network For rough complement diagram Carrying out refined reconstruction to obtain a complete sea surface temperature image 。

Description

Sea temperature completion method and system based on space-time priori and flow matching generation model Technical Field The invention belongs to the technical field of sea temperature complementation, and particularly relates to a sea temperature complementation method and system based on a space-time priori and flow matching generation model. Background The sea surface temperature SST is a key physical quantity in marine science, climate prediction and marine environment monitoring, and a space-time continuous and high-precision data field has important significance for understanding marine power process, evaluating climate change and developing marine disaster early warning. However, due to the influence of factors such as coverage of an observation platform, weather conditions, sensor faults and the like, the SST data obtained by satellite and field observation have the common problems of space deficiency, time sequence discontinuity and the like, and the subsequent analysis and application are severely restricted. Therefore, developing an efficient and reliable SST data complement technology, reconstructing a space-time continuous sea temperature field, has become a core requirement in the field of sea data processing. Traditional data interpolation and assimilation means are often limited to the utilization of single-moment observation data, and have the defect of insufficient information utilization, and the method based on the deep neural network can be used for carrying out deep modeling on large-scale historical ocean data through an algorithm, so that the time-space evolution characteristics of a sea surface temperature field are effectively captured, and the reconstruction precision of missing data is remarkably improved. At present, the front edge method in the sea surface temperature finishing field based on deep learning adopts an image finishing mechanism of a diffusion model based on multi-scale physical constraint, mainly takes Zhou Junzhi as global physical constraint and average value deviation as local physical constraint, effectively captures long-term distribution characteristics and short-term details of data, ensures that finishing results accord with sea physical rules, such as space continuity of a temperature field, and secondly, endows stable training process and strong generating capacity based on a denoising diffusion probability model framework, and compared with other generating models such as a countermeasure network, the denoising diffusion probability model can better process complex mode and large-area missing data, so that higher-quality reconstruction is realized under a high Yun Zhedang rate scene, finally, the multi-scale fusion module extracts multi-scale characteristics, effectively eliminates overlapped noise through a gating fusion mechanism, and remarkably improves space consistency and detail performance of a finishing image. However, the above method still has defects, and the following will be discussed one by one: First, the time sequence evolution rule of sea temperature data is ignored, so that the inference of the missing region lacks time sequence logic support, and the completion precision is insufficient. The existing multi-scale physical constraint method essentially aims at solving the problem of static image restoration at a single moment by carrying out dimension reduction treatment on space-time completion tasks, and the model only relies on current day data and a static week average background field to infer, so that historical time sequence information before the missing moment is completely ignored. Because of lack of explicit modeling on sea temperature flowing and diffusing trend along with time, continuous change in the sea dynamics process is difficult to understand by the model, especially when complex sea conditions with obvious dynamic characteristics such as mesoscale vortex movement, sea frontal surface evolution and the like are processed, the phenomenon of texture fracture, position deviation or physical logic incoherence often occurs in the reconstruction result, and the accuracy of sea surface temperature data missing region completion is limited. Second, the generation mechanism based on the traditional denoising diffusion probability model DDPM has the problems of random reasoning paths and low generation efficiency. The core generation module of the model relies on the Markov chain denoising process of DDPM. This random differential equation based sampling approach, its inverse denoising path is highly random and curved in probability space. In order to recover high-quality texture details from pure Gaussian noise, the model usually needs hundreds or even thousands of time steps of iterative computation, so that the reasoning speed is extremely low, the computing resource consumption is huge, and the application potential of the model in large-scale sea temperature data processing tasks is limited. Disclosure of Invention In o