Sea surface height abnormal data downscaling method based on deep learning

Abstract

The invention provides a sea surface height abnormal data downscaling method based on deep learning, and relates to the technical field of data processing, and the method specifically comprises the following steps: preprocessing data including AVISO data, SWOT data and key auxiliary variables; an enhanced super-resolution generative adversarial network ESRGAN is constructed, a high-resolution feature map is generated, and the ESRGAN comprises a generative network and a discrimination network; using a plurality of loss function combinations to optimize the performance of the generative network and the discriminant network, including adversarial loss, perception loss and pixel loss; and evaluating the pixel-level error and the spatial detail consistency output by the ESRGAN by adopting a root-mean-square error and a structural similarity index respectively. According to the technical scheme, the problems that in the prior art, a data downscaling method is limited in applicability and generalization ability, and is difficult to adapt to fusion requirements of different regions and multi-source data are solved.

Description

Technical Field

[0001] The present invention relates to the field of data processing, and in particular to a method for downscaling sea surface height anomaly data based on deep learning. Background Art

[0002] Sea surface height anomaly (SSHA) data are of great significance in marine scientific research and are widely used to reveal ocean circulation, tidal changes, small and medium-scale dynamic processes (such as eddies, fronts, etc.) and climate change-related phenomena. However, due to the limitations of observation technology, the currently commonly used SSHA data have obvious deficiencies in resolution and accuracy. AVISO satellite altimeter products are one of the most commonly used SSHA observation data in the world, with advantages such as wide coverage and complete time series, but their spatial resolution is usually tens of kilometers, which is only suitable for the study of large-scale phenomena and cannot describe the details of small and medium-scale dynamic processes. In contrast, SWOT satellites, as a new generation of ocean observation platforms, can provide high-precision SSHA data with a resolution of several hundred meters, which is particularly suitable for the detailed study of small and medium-scale dynamic processes and nearshore areas. However, due to the limited coverage of SWOT data, the long observation period (revisit rate of 21 days), and the high acquisition and processing costs, it is difficult to be widely used in global long-term observations.

[0003] Traditional data downscaling methods, such as linear interpolation and empirical orthogonal function (EOF) decomposition, can improve data resolution to a certain extent, but have obvious limitations in capturing the complex nonlinear characteristics and spatiotemporal dynamic correlations of SSHA data. These methods usually rely on assumed linear relationships and ignore the essential characteristics of nonlinearity and multi-scale coupling in ocean dynamic processes, resulting in poor results in dealing with small and medium-scale phenomena. In addition, the applicability and generalization capabilities of traditional methods are limited, making it difficult to adapt to the fusion needs of different regions and multi-source data. These problems make the existing downscaling techniques unable to meet actual needs in terms of accuracy, efficiency and applicability.

[0004] Therefore, a high-precision and efficient deep learning-based downscaling method for sea surface height anomaly data is needed. Summary of the invention

[0005] The main purpose of the present invention is to provide a sea surface height anomaly data downscaling method based on deep learning to solve the problem that the data downscaling method in the prior art has limited applicability and generalization ability and is difficult to adapt to the fusion needs of different regions and multi-source data.

[0006] To achieve the above object, the present invention provides a method for downscaling sea surface height anomaly data based on deep learning, which specifically includes the following steps: S1, preprocess the data, the data include: AVISO data, SWOT data and key auxiliary variables, the key auxiliary variables include: AVISO, SST, wind speed, air pressure, seabed topography and tidal data.

[0007] S2, builds an enhanced super-resolution generative adversarial network ESRGAN to generate high-resolution feature maps , ESRGAN includes: generation network and discrimination network.

[0008] S3 uses a combination of multiple loss functions to optimize the performance of the generation network and the discriminator network, including adversarial loss, perceptual loss, and pixel loss.

[0009] S4, the pixel-level error and spatial detail consistency of ESRGAN output are evaluated using the root mean square error MSE and the structural similarity index SSIM, respectively.

[0010] Furthermore, step S1 specifically includes the following steps: S1.1, use AVISO data as model input and SWOT data as model labels.

[0011] S1.2, time-match the data.

[0012] S1.3, spatially match and align the data.

[0013] S1.4, clean the data and process missing values.

[0014] S1.5, normalize the data.

[0015] S1.6, divide the data into training set, validation set and test set according to time.

[0016] Furthermore, step S1.2 is specifically as follows: According to the observation time of SWOT, the AVISO, SST, wind speed, air pressure, seabed topography and tide data of the corresponding time point were screened.

[0017] Furthermore, step S1.3 specifically includes the following steps: S1.3.1, cut out regular rectangular areas from the SWOT strip-shaped distribution data, and extract the AVISO data of the corresponding areas to generate paired data blocks.

[0018] S1.3.2, align the grid coordinates of the cropped regular rectangular area to ensure that AVISO and SWOT are at the same spatial resolution and coordinate system.

[0019] S1.3.3, interpolate all key auxiliary data to the spatial resolution of a regular rectangular area to maintain spatial consistency; intercept the seabed topography within the regular rectangular area and downsample it to the strip resolution using the resampling method.

[0020] S1.3.4, interpolate strips that cross boundaries.

[0021] Furthermore, the construction of the generation network in step S2 includes the following steps: S2.1, AVISO and key auxiliary variables are extracted through convolutional networks respectively, the extracted features are spliced ​​in the channel dimension, the spliced ​​features are used as input features, and the channel attention mechanism is introduced. The mathematical expression is: ; in, is the weighted feature, that is, the feature representation adjusted by the attention mechanism; is the input feature, and is the attention weight matrix, is the Sigmoid activation function, is the activation function.

[0022] S2.2, the input data, i.e., the weighted features, first undergo an initial convolution operation, followed by multiple layers of RRDB to extract deep multi-scale features. Then, the multi-scale features are gradually enlarged and mapped to the high-resolution space through the upsampling module, and finally the target high-resolution image is generated. , the mathematical expression is:

[0023] in represents the initial convolution operation, express The stacking operation of the layer RRDB, Represents an upsampling module.

[0024] Furthermore, the RRDB in step S2.2 includes multiple interconnected dense blocks, each of which includes multiple interconnected convolutional layers. The output of each dense block is the input of the subsequent dense block. The output features of each dense block are channel compressed by convolution and then superimposed with the input features through residual connection, thereby forming a local residual structure: ; ; in, For each dense block The output of the convolutional layer; is the convolution and activation function; is the combined feature of the internal dense blocks in RRDB, for AVISO and key auxiliary variables, is the residual scaling factor, is the output of the residual dense block RRDB.

[0025] Furthermore, in step S2.2, the low-resolution features extracted by RRDB are converted into Mapping to target high-resolution space: ; in, is the convolution output feature map, and are the convolution weights and biases, respectively. Low-resolution features.

[0026] Subsequently, a pixel rearrangement operation is used to convert the channel dimension of the feature map into a spatial dimension to obtain a high-resolution feature map: ; in, They are the output feature maps In the horizontal and vertical spatial position index, is the channel index, To round down, is a modulo operation, that is, calculating the sub-pixel position corresponding to the current pixel, is the upsampling ratio, is a high-resolution feature map.

[0027] Furthermore, the discriminant network extracts multi-scale features through multiple convolutional layers and outputs probabilities through fully connected layers. , indicating whether the input data is real SWOT data or generated data from the generation network.

[0028] Furthermore, the adversarial loss in step S3 is: ; in, To combat losses; for AVISO and key auxiliary variables; For a true SWOT; and Respectively express and expectations; represents the logarithmic function; is the discriminant network, that is, the probability of whether the output data is true.

[0029] The expression of perceptual loss is: ; in, For perceived loss; is the feature map of a layer in the VGG network; It is expectation; Represents the square of the Euclidean distance.

[0030] The expression of pixel loss is: ; in, is the pixel loss.

[0031] Total loss The expression is: ; in, , and is the weight parameter.

[0032] Furthermore, in step S4 is the average squared difference between the predicted value and the true value, calculated as: ; in, is the root mean square error to be obtained; For the The true value of samples; For the The predicted value of a sample.

[0033] Structural Similarity Index The calculation formula is: ; in, and are the reference image and the image to be compared respectively; and Respectively and The mean of and They are and The variance of for and The covariance of and is a constant.

[0034] The present invention has the following beneficial effects: This paper uses deep learning methods to combine the advantages of AVISO and SWOT data, and realizes the organic fusion of multi-source observation data through downscaling, providing a high-resolution data set with higher consistency and wider coverage for ocean dynamics research. Its specific effects and advantages are as follows:

[0035] (1) Improved spatial resolution: By learning the complex mapping relationship between AVISO data and SWOT data through a deep learning model, a high-precision SSHA field close to the SWOT resolution level can be generated. The spatial resolution of the generated data is increased from tens of kilometers to hundreds of meters, significantly enhancing the ability to describe small and medium-scale ocean phenomena.

[0036] (2) Improved temporal continuity: By utilizing the global coverage and temporal continuity of AVISO data and generating high-resolution data fields through a downscaling model, we not only fill the SWOT observation blind spots in space, but also improve the continuity of data in time, solving the problem that a single data source cannot achieve long-term continuous observation.

[0037] (3) High-precision nonlinear feature extraction: The deep learning model can capture the complex nonlinear relationship and multi-scale spatiotemporal dynamic characteristics between AVISO and SWOT data. Compared with traditional downscaling methods, it can effectively reduce calculation errors, and the generated data has more physical meaning in terms of spatial details and overall trends.

[0038] (4) Cost and efficiency advantages: Maximize the use of existing observation data, avoid additional hardware or data acquisition costs, and reduce dependence on high-cost, high-resolution SWOT data. High-resolution SSHA data fields can be generated only through existing AVISO data. At the same time, the deep learning model can quickly generate high-resolution data after training, without relying on complex manual processing procedures, and its efficiency is much higher than traditional downscaling methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the specific implementation of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for the specific implementation or the prior art description. Obviously, the drawings described below are some implementations of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. In the drawings: Figure 1 A flow chart of a method for downscaling sea surface height anomaly data based on deep learning of the present invention is shown.

[0040] Figure 2 A low-resolution SSHA is shown.

[0041] Figure 3High-resolution SSHA is shown.

[0042] Figure 4 The SSHA after downscaling is shown. DETAILED DESCRIPTION

[0043] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0044] like Figure 1 A method for downscaling sea surface height anomaly data based on deep learning is shown, which specifically includes the following steps: S1, preprocess the data, including AVISO data, SWOT data and key auxiliary variables, including sea surface temperature SST, sea surface wind speed, pressure field, seabed topography, tidal data, spatial location information and seasonal characteristics. Obtain AVISO, key auxiliary variable data and SWOT, and ensure the accuracy and consistency of data preparation through time and space alignment, data cleaning and normalization.

[0045] S2, builds an enhanced super-resolution generative adversarial network ESRGAN to generate high-resolution feature maps ,ESRGAN includes: a generative network and a discriminative network.,The enhanced super-resolution generative adversarial network is used to combine low-resolution sea surface height anomaly data and auxiliary variables to generate high-resolution sea surface height anomaly.

[0046] S3 uses a combination of multiple loss functions to optimize the performance of the generation network and the discriminator network, including adversarial loss, perceptual loss, and pixel loss.

[0047] In the pre-training stage, the generative network is trained to achieve results consistent with the real high-resolution data by optimizing pixel loss and perceptual loss.

[0048] In the joint training stage, the generative network and the discriminative network are optimized through adversarial learning, so that the generative network can generate more realistic, high-quality high-resolution data while improving spatial details.

[0049] The generation network and the discriminant network are optimized through adversarial loss, perceptual loss and pixel loss to improve the authenticity and detail expression of the data.

[0050] S4, the pixel-level error and spatial detail consistency of ESRGAN output are evaluated using the root mean square error MSE and the structural similarity index SSIM, respectively.

[0051] This paper proposes a downscaling method for sea surface height anomaly (SSHA) data based on deep learning. By combining AVISO and SWOT data, the deep learning model is used to effectively overcome the limitations of traditional downscaling methods in capturing nonlinear characteristics and multi-scale spatiotemporal dynamic relationships. This method not only improves the spatial resolution and spatiotemporal continuity of SSHA data, but also takes into account the global coverage of AVISO and the high-precision characteristics of SWOT, providing an efficient, low-cost solution with strong generalization capabilities, and providing a new technical path for the study of small and medium-scale ocean dynamic processes and global ocean monitoring.

[0052] Specifically, step S1 includes the following steps: S1.1, AVISO data (low-resolution SSHA) is used as model input, and SWOT data (high-resolution SSHA) is used as model labels.

[0053] S1.2, time-match the data.

[0054] S1.3, spatially match and align the data.

[0055] S1.4, clean the data and process missing values. Check whether there are outliers or missing values ​​in the clipped data (such as areas near the coastline), and use interpolation to repair invalid data.

[0056] S1.5, normalize the data.

[0057] S1.6, divide the data into training set, validation set and test set according to time. The ratio is 80% training, 10% validation and 10% test to ensure the independence of training and validation / test data in time.

[0058] Specifically, step S1.2 is as follows: According to the observation time of SWOT, AVISO, sea surface temperature SST, sea surface wind speed, pressure field, seabed topography, tidal data, spatial location information and seasonal characteristics of the corresponding time point are selected. For example, the strip data obtained by SWOT from 3:08 to 3:39 on March 29, 2023 should match the AVISO data on the same day (March 29, 2023).

[0059] Sea Surface Temperature (SST): Obtained from reanalysis data (ERA5) and used to describe the thermal state of the ocean surface.

[0060] Sea surface wind speed: ERA5 wind field reanalysis data is used to supplement the dynamic driving information of SSHA changes.

[0061] Pressure field (SLP): The ERA5 pressure field reanalysis data captures the direct impact of pressure changes on SSHA.

[0062] Seafloor topography: Seafloor topography information is extracted from ETOPO 2022 data to reflect the modulation effect of topography on ocean dynamic processes.

[0063] Tidal data: The TPXO tidal model is used to generate local tidal amplitude and phase data as the time period characteristic input of the model.

[0064] Spatial location information: The latitude and longitude information reflects the geographical location characteristics, helping the model understand the regional SSHA distribution differences.

[0065] Seasonal characteristics: Seasonal information is introduced through monthly time coding to capture the seasonal variation of SSHA.

[0066] Specifically, step S1.3 includes the following steps: S1.3.1, cut out regular rectangular areas from the SWOT strip-shaped distribution data, and extract the AVISO data of the corresponding areas to generate paired data blocks.

[0067] S1.3.2, align the grid coordinates of the cropped regular rectangular area to ensure that AVISO and SWOT are at the same spatial resolution and coordinate system.

[0068] S1.3.3, interpolate all key auxiliary data to the spatial resolution of a regular rectangular area to maintain spatial consistency; the spatial resolution of the seabed topography data is high, and the seabed topography within the regular rectangular area is intercepted and downsampled to the strip resolution using the resampling method.

[0069] S1.3.4, interpolate strips that cross boundaries to avoid discontinuities or data null problems.

[0070] Specifically, each variable is normalized to the range of [0, 1] or [-1, 1] to ensure that different variables of the input data are on the same scale to avoid the adverse effects of numerical range differences on model training. The calculation formula for normalization in step S1.5 is as follows: ; in, It is the original data; is the mean of the original data, that is ; is the standard deviation of the data, that is ; is the total number of data, For the data, is the normalized data value. In addition, sine and cosine encoding is used for tidal phase variables with periodic changes.

[0071] In order to achieve downscaling inference from low-resolution AVISO SSHA data to high-resolution SWOT SSHA data, this paper adopts a model architecture based on enhanced super-resolution generative adversarial network (ESRGAN). ESRGAN has the advantages of retaining detail information and perceptual quality when generating high-resolution data, and combines auxiliary variables (SST, wind field, etc.) and deep learning characteristics to effectively solve the shortcomings of traditional methods in nonlinear modeling capabilities and spatial detail reconstruction capabilities.

[0072] ESRGAN is an improved model based on the traditional super-resolution GAN. By introducing residual dense blocks and adversarial learning strategies, the performance of the generative network and adversarial network is optimized, and it can generate higher quality and more detailed high-resolution data. In ESRGAN, the generative network and the discriminative network are a pair of adversarial networks, which optimize each other's performance through the process of mutual game.

[0073] Specifically, the construction of the generation network in step S2 includes the following steps: S2.1, the main task of the generation network is to convert the input AVISO data and key auxiliary variable features into high-resolution SSHA data. The generation network includes: input layer, residual dense block, upsampling module and output layer. The input layer includes: main input features, i.e. low-resolution AVISO SSHA data; key auxiliary variables, as supplementary feature information, reflect the multi-dimensional influence of ocean dynamics and environmental factors. AVISO and key auxiliary variables are extracted through convolutional networks respectively, and the extracted features are spliced ​​in the channel dimension. The spliced ​​features are used as input features, and the channel attention mechanism is introduced. The mathematical expression is: ; in, is the weighted feature, that is, the feature representation adjusted by the attention mechanism; is the input feature, and is the attention weight matrix, is the Sigmoid activation function, is the activation function.

[0074] S2.2, RRDB and upsampling modules work together in the generative network. The input data first passes through multiple RRDBs to extract deep multi-scale features to form deep features These features are then gradually enlarged and mapped to the high-resolution space through the upsampling module, and finally the target high-resolution image is generated. The input data, i.e. the weighted features, first undergo an initial convolution operation, followed by multiple layers of RRDB to extract deep multi-scale features. Then, the multi-scale features are gradually enlarged and mapped to the high-resolution space through the upsampling module, and finally the target high-resolution image is generated. , the mathematical expression is: ; in represents the initial convolution operation, express The stacking operation of the layer RRDB, Represents an upsampling module.

[0075] Specifically, the residual dense block (RRDB) is the core module of ESRGAN, which is used to efficiently extract multi-scale features and achieve stable training in deep networks. RRDB combines the advantages of dense connections and residual connections. Through the reuse of multi-layer features and cross-layer connections, it can significantly improve the detail performance and perceptual quality of the model when reconstructing high-resolution data. The interior of RRDB is composed of multiple dense blocks, each of which captures features under different receptive fields through dense connections. Each dense block includes multiple interconnected convolutional layers. In the dense connection, the output of each layer of convolution is used as the input of all subsequent layers, and the output of each dense block is the input of the subsequent dense block, ensuring that the features can be fully reused. The residual dense block in step S2.2 includes multiple interconnected dense blocks, and the output of each dense block is the input of the subsequent dense block: ; The output features of each dense block are compressed by convolution and then superimposed with the input features through residual connections to form a local residual structure. In RRDB, multiple dense blocks are further nested, and the overall feature integration of the module is achieved through global residual connections. The output of RRDB is expressed by the following formula: ; in, For the The output of a dense block; It is a combination of convolution and activation function, usually the leaky rectified linear unit Leaky ReLU; is the combined feature of the internal dense blocks in RRDB, for AVISO and key auxiliary variables, is the residual scaling factor (usually set to 0.2), which is used to control the residual amplitude to avoid excessive feature updates affecting training stability. is the output of the residual dense block RRDB.

[0076] Specifically, the low-resolution features extracted by RRDB It needs to be mapped to the target high-resolution space through an upsampling module. The upsampling module of ESRGAN is usually implemented using sub-pixel convolution. The core idea is to expand the information of the channel dimension to the spatial dimension through convolution and pixel rearrangement operations, thereby generating a high-resolution feature map. Specifically, assuming that the input feature map size is , the upsampling ratio is , first generate a convolution with a dimension of The feature map of: ; in, is the convolution output feature map, and are the convolution weights and biases, respectively. Low-resolution features.

[0077] Then, a pixel rearrangement operation is used to transform the channel dimension of the feature map Convert to spatial dimensions , and obtain a high-resolution feature map: ; in, They are the output feature maps In the horizontal and vertical spatial position index, is the channel index, To round down, is a modulo operation, that is, calculating the sub-pixel position corresponding to the current pixel, is the upsampling ratio, is a high-resolution feature map of size ,in, , , represents the height of the feature map, represents the width of the feature map, Indicates the number of channels of the feature map, represents the height of the low-resolution feature map, Indicates the height of the high-resolution feature map.

[0078] This sub-pixel convolution method has a significant computational efficiency advantage over traditional interpolation methods. It directly generates a feature map of the target resolution in the feature space, avoiding redundant calculations that may be introduced in the interpolation operation. In addition, through end-to-end learning, sub-pixel convolution can effectively capture the nonlinear mapping relationship between input data and high-resolution output, thereby generating more realistic high-resolution images.

[0079] Specifically, the task of the discriminant network is to determine whether the input is real high-resolution data (SWOT data) or generated data from the generator network. The input of the discriminant network includes: the output of the generator network and the real high-resolution data. The discriminant network consists of a multi-layer convolutional network, which extracts multi-scale features through the convolutional layer and outputs the probability through the fully connected layer. , indicating whether the input data is real data. By confronting the generative network, the perceptual quality of the data can be improved.

[0080] The discriminant network extracts multi-scale features through multiple convolutional layers and outputs probabilities through fully connected layers. , indicating whether the input data is real SWOT data or generated data from the generation network.

[0081] Specifically, the adversarial loss in step S3 is: ; in, To combat losses; Low-resolution input, namely: AVISO and key auxiliary variables; For a true SWOT; and Respectively express and expectations; represents the logarithmic function; is the discriminant network, that is, the probability of whether the output data is real. Adversarial loss is used to improve the authenticity of generated data.

[0082] In ESRGAN, the Visual Geometry Group (VGG) network is used as a pre-trained feature extractor to calculate the perceptual loss. The role of the perceptual loss is to compare the similarity between the generated image and the real image in the high-level feature space. The expression of the perceptual loss is: ; in, For perceived loss; is the feature map of a layer in the VGG network; It is expectation; Represents the square of the Euclidean distance.

[0083] Pixel loss uses mean square error to measure the difference between the generated image and the real image at the pixel level. The expression of pixel loss is: ; in, is the pixel loss.

[0084] Total loss The expression is: ; in, , and is a weight parameter used to balance the contribution of each loss to optimization.

[0085] The entire training is divided into two main stages: the pre-training stage of the generator network and the joint training stage of the generator network and the discriminator network. This phased training strategy aims to gradually optimize the downscaling ability of the generator network while improving the perceptual quality and detail restoration ability of the generated data.

[0086] In the first stage, only the generator network is trained to generate high-resolution SSHA data that is consistent with the true label at the pixel and feature level. In this stage, the loss function optimized by the model includes pixel loss and perceived loss The pixel loss measures the difference between the generated data and the real data in terms of value, while the perceptual loss extracts high-level features through the VGG network to ensure that the generated data is close to the real data in terms of visual quality and structural characteristics. The network parameters are optimized using the Adam optimizer, and the initial learning rate is set to 10 -4 , and gradually decays during the training process until the error of the generated network on the validation set converges.

[0087] In the second stage, the generator and discriminator networks enter joint training to further improve the generator's capabilities through adversarial learning. The goal of this stage is to enable the generator network to not only generate pixel-consistent data, but also improve the perceived quality and detail performance through adversarial strategies. In this stage, the total loss function of the joint training combines pixel loss, perceptual loss, and adversarial loss. The adversarial loss LGAN of the generator network aims to make it impossible for the discriminator network to distinguish between generated data and real data, while the goal of the discriminator network is to maximize the ability to distinguish between real data and generated data. By alternately updating the parameters of the generator and discriminator networks, the generator network's ability to generate high-resolution data is gradually enhanced.

[0088] During the entire training process, the generator network is usually updated more often than the discriminator network to ensure that the generator network is prioritized. The training is completed when the generator network achieves the optimal loss function value on the validation set and the discriminator network cannot effectively distinguish between the generated data and the real data.

[0089] Specifically, in step S4 It is the average square difference between the predicted value and the true value, which is mainly used to evaluate numerical accuracy. The calculation formula is: ; in, is the root mean square error to be obtained; For the The true value of samples; For the The predicted value of a sample.

[0090] SSIM is used to evaluate the spatial detail restoration capability of high-resolution SSHA fields, structural similarity index The calculation formula is: ; in, and are the reference image and the image to be compared respectively; and Respectively and The mean of , ; and They are and The variance of , ; for and The covariance of ; and is a constant, usually taken , , is the dynamic range of the image. Since the normalized SSHA range is [-1, 1], the dynamic range is 2. The value range of SSIM is [0, 1], where 1 means that the two images are exactly the same and 0 means that the two images are completely different.

[0091] In the data preparation stage, the present invention obtains the main input feature data, key auxiliary data and label data, the main input data is AVISO, and the label data is high-resolution SWOT SSHA. After the data is matched in time and space, the feature and label data sets required for model training are formed. In the model training stage, the ESRGAN model used is composed of a generation network and a discriminant network, wherein the generator includes an input layer, a residual dense block and an upsampling module. The input layer is mainly used for multi-input feature fusion. The main input feature AVISO and key auxiliary data (such as sea surface temperature, sea surface wind speed, etc.) are extracted through independent convolutional networks, and feature splicing is performed in the channel dimension. After feature fusion, the channel attention mechanism is introduced to dynamically adjust the weight of each input feature. The residual dense block consists of multiple dense blocks, each of which captures features under different receptive fields through dense connections. The output features of each dense block are channel compressed through convolution, and then superimposed with the input features through residual connections to form a local residual structure. The upsampling module is implemented using sub-pixel convolution. The core idea is to expand the information of the channel dimension to the spatial dimension through convolution and pixel rearrangement operations, thereby generating a high-resolution feature map. The discriminant network consists of a multi-layer convolutional network. The convolutional layer extracts multi-scale features and the fully connected layer outputs the probability , indicating whether the input data is real data. The total loss function of the model comprehensively considers perceptual loss, pixel loss and adversarial loss. In the model evaluation stage, the root mean square error and structural similarity index are used to evaluate the numerical accuracy and spatial detail restoration ability of the model output.

[0092] The present invention aims to solve the problems of insufficient resolution and coverage of current SSHA data. The SSHA data provided by AVISO have the advantages of global coverage and long-term observation, but the spatial resolution is low, which makes it difficult to meet the needs of refined research on small and medium-scale dynamic processes; SWOT data has high-resolution characteristics, but its observation period is long and its coverage is limited, making it difficult to achieve long-term global continuous monitoring. In order to bridge the gap between the two, the present invention uses a deep learning model to construct an efficient downscaling method by taking AVISO low-resolution SSHA data as feature input and SWOT high-resolution SSHA data as labels, thereby generating a high-resolution SSHA data field and improving the spatial refinement and spatiotemporal continuity of the data.

[0093] Figure 2-Figure 4 The figure shows the comparison of SSHA before and after downscaling. Based on the trained model, low-resolution sea surface height anomaly data and auxiliary data are input to finally obtain high-resolution sea surface height anomaly data.

[0094] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the essential scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A method for downscaling sea surface height anomaly data based on deep learning, characterized in that: The specific steps include: S1, preprocessing the data, including AVISO data, SWOT data and key auxiliary variables, including sea surface temperature SST, sea surface wind speed, pressure field, seabed topography, tidal data, spatial location information and seasonal characteristics; S2, builds an enhanced super-resolution generative adversarial network ESRGAN to generate high-resolution feature maps ,ESRGAN includes: a generating network and a discriminative network; S3, uses a combination of multiple loss functions to optimize the performance of the generation network and the discriminative network, including adversarial loss, perceptual loss, and pixel loss; S4, the pixel-level error and spatial detail consistency of ESRGAN output are evaluated using the root mean square error MSE and the structural similarity index SSIM, respectively.

2. The method for downscaling sea surface height anomaly data based on deep learning according to claim 1, characterized in that: Step S1 specifically includes the following steps: S1.1, AVISO data is used as model input and SWOT data is used as model label; S1.2, time matching of data; S1.3, spatially match and align the data; S1.4, clean the data and process missing values; S1.5, normalize the data; S1.6, divide the data into training set, validation set and test set according to time.

3. The method for downscaling sea surface height anomaly data based on deep learning according to claim 2, characterized in that: Step S1.2 is specifically as follows: According to the observation time of SWOT, AVISO, sea surface temperature SST, sea surface wind speed, air pressure field, seabed topography, tidal data, spatial location information and seasonal characteristics of the corresponding time point were screened.

4. The method for downscaling sea surface height anomaly data based on deep learning according to claim 2, characterized in that: Step S1.3 specifically includes the following steps: S1.3.1, cut out a regular rectangular area from the SWOT strip-shaped distribution data, and extract the AVISO data of the corresponding area to generate a paired data block; S1.3.2, align the grid coordinates of the cropped regular rectangular area to ensure that AVISO and SWOT are at the same spatial resolution and coordinate system; S1.3.3, interpolate all key auxiliary data to the spatial resolution of a regular rectangular area to maintain spatial consistency; intercept the seafloor topography within the regular rectangular area and downsample to the strip resolution using the resampling method; S1.3.4, interpolate strips that cross boundaries.

5. The method for downscaling sea surface height anomaly data based on deep learning according to claim 1, characterized in that: The construction of the generation network in step S2 includes the following steps: S2.1, AVISO and key auxiliary variables are extracted through convolutional networks respectively, the extracted features are spliced ​​in the channel dimension, the spliced ​​features are used as input features, and the channel attention mechanism is introduced. The mathematical expression is: ； in, is the weighted feature, that is, the feature representation adjusted by the attention mechanism; is the input feature, and is the attention weight matrix, is the Sigmoid activation function, is the activation function; S2.2, the input data, i.e., the weighted features, first undergo an initial convolution operation, followed by multiple layers of RRDB to extract deep multi-scale features. Then, the multi-scale features are gradually enlarged and mapped to the high-resolution space through the upsampling module, and finally the target high-resolution image is generated. , the mathematical expression is: ； in represents the initial convolution operation, express The stacking operation of the layer RRDB, Represents an upsampling module.

6. The method for downscaling sea surface height anomaly data based on deep learning according to claim 5, characterized in that: The RRDB in step S2.2 includes multiple interconnected dense blocks, each of which includes multiple interconnected convolutional layers. The output of each dense block is the input of the subsequent dense block. The output features of each dense block are channel compressed by convolution and then superimposed with the input features through residual connection to form a local residual structure: ; ; in, For each dense block The output of the convolutional layer; is the convolution and activation function; is the combined feature of the internal dense blocks in RRDB, for AVISO and key auxiliary variables, is the residual scaling factor, is the output of the residual dense block RRDB.

7. The method for downscaling sea surface height anomaly data based on deep learning according to claim 5, characterized in that: In step S2.2, the low-resolution features extracted by RRDB are converted into Mapping to target high-resolution space: ; in, is the convolution output feature map, and are the convolution weights and biases, respectively. It is a low-resolution feature; Subsequently, a pixel rearrangement operation is used to convert the channel dimension of the feature map into a spatial dimension to obtain a high-resolution feature map: ； in, They are the output feature maps In the horizontal and vertical spatial position index, is the channel index, To round down, is a modulo operation, that is, calculating the sub-pixel position corresponding to the current pixel, is the upsampling ratio, is a high-resolution feature map.

8. The method for downscaling sea surface height anomaly data based on deep learning according to claim 1, characterized in that: The discriminant network extracts multi-scale features through multiple convolutional layers and outputs probabilities through fully connected layers. , indicating whether the input data is real SWOT data or generated data from the generation network.

9. The method for downscaling sea surface height anomaly data based on deep learning according to claim 1, characterized in that: The adversarial loss in step S3 is: ； in, To combat losses; for AVISO and key auxiliary variables; For a true SWOT; and Respectively express and expectations; represents the logarithmic function; For the discriminant network, that is, the probability of whether the output data is true; The expression of perceptual loss is: ； in, For perceived loss; is the feature map of a layer in the VGG network; It is expectation; represents the square of the Euclidean distance; The expression of pixel loss is: ； in, is the pixel loss; Total loss The expression is: ； in, , and is the weight parameter.

10. The method for downscaling sea surface height anomaly data based on deep learning according to claim 1, characterized in that: In step S4 is the average squared difference between the predicted value and the true value, calculated as: ； in, is the root mean square error to be obtained; For the The true value of samples; For the The predicted value of samples; Structural Similarity Index The calculation formula is: ; in, and are the reference image and the image to be compared respectively; and Respectively and The mean of and They are and The variance of for and The covariance of and is a constant.

Images