Sea surface temperature space statistical downscaling method based on U-shaped convolutional neural network

Abstract

The invention belongs to the technical field of ocean data reanalysis space forecasting and image super-resolution, and particularly relates to a sea surface temperature space statistical downscaling method based on a U-shaped convolutional neural network, which comprises the following steps of: 1, acquiring statistical data to be analyzed and performing normalization preprocessing; 2, constructing a data set of network training, and dividing a training set and a test set; 3, constructing a U-shaped convolutional neural network; training the U-shaped convolutional neural network through the training set and the test set, and outputting the trained U-shaped convolutional neural network; 4, judging the downscaling effect of the U-shaped convolutional neural network; and step 5, acquiring real-time low-resolution ocean temperature data, preprocessing the real-time low-resolution ocean temperature data, and outputting downscaled high-resolution ocean temperature data by using the trained U-shaped convolutional neural network. According to the method, details and nonlinear features in the spatial information can be well extracted and restored, and the precision of statistical downscaling can be improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of ocean data reanalysis spatial forecasting and image super-resolution, and specifically relates to a sea surface temperature spatial statistical downscaling method based on a U-shaped convolutional neural network. Background Art

[0002] In the field of ocean forecasting, ocean data is often output in the form of gridded information. The denser the distribution of coordinate points within a region of the ocean data matrix, the higher the accuracy of the ocean forecast for that area. Currently, downscaling is a common method for ocean data processing internationally. Downscaling ocean data can provide richer ocean forecast information within a specific ocean area, making the description of local ocean regions more accurate. Downscaling methods establish a relationship between large-scale and small-scale ocean information variables. These methods mainly include dynamical downscaling and statistical downscaling. Dynamical downscaling involves using a low-resolution ocean environment data model to provide initial and boundary conditions for a high-resolution regional ocean environment data model. Through the action of internal physical processes, the regional ocean environment data not only incorporates the large-scale features found in the low-resolution ocean environment data, but also simulates more regional ocean environment information features than the low-resolution data. Statistical downscaling, on the other hand, does not rely on the boundary initial conditions provided by the large-scale ocean environment information model. Instead, it is based on the statistical relationship between local observational information and large-scale ocean environment information. Compared to computationally intensive dynamical downscaling, statistical downscaling offers the advantages of high computational efficiency, minimal computational effort, simple model construction, and ease of implementation. It is suitable for describing high-resolution changes in elements within a target region. The primary purpose of applying statistical downscaling to ocean data is to downscale low-resolution ocean variables to obtain accurate, high-resolution ocean information, thereby capturing more detailed ocean features and providing more accurate and refined information for ocean data monitoring, forecasting, and related reanalysis.

[0003] The main process of statistical downscaling is: first, the statistical relationship between large-scale ocean environmental variables and small-scale ocean environmental variables is studied, and a fitting model of this statistical relationship is established. Then, this statistical relationship is applied to large-scale ocean environmental data, and finally, refined information about the small-scale ocean environment in the corresponding area can be output. Although linear methods can achieve good downscaling effects under normal climatic conditions, linear methods cannot fit nonlinear variables well when applied to some nonlinear variable relationships, resulting in large errors. For this reason, the present invention proposes a sea surface temperature spatial statistical downscaling method based on a U-shaped convolutional neural network. Summary of the Invention

[0004] The present invention aims to provide a sea surface temperature spatial statistical downscaling method based on a U-shaped convolutional neural network. When downscaling spatial feature information with extreme, discrete, and other nonlinear variations, the U-shaped convolutional neural network effectively addresses the problems of increased spatial downscaling bias and even reconstruction errors caused by linear regression model fitting of irregular and variable details, effectively fitting and restoring nonlinear features and detail variation features. During the U-Net spatial downscaling process, the overall network structure is an "encoder-decoder" structure. After network training, it can effectively extract spatial features, nonlinearly map feature relationships, and reconstruct feature functions. Furthermore, the network includes a "skip connection" structure that combines detailed features in shallow network feature maps with abstract information in deep network feature maps. This effectively addresses the shortcoming of deep convolutional neural networks that often miss shallow feature details during super-resolution reconstruction, thereby improving the accuracy of spatial downscaling. The method proposed in the present invention can effectively improve the prediction accuracy and efficiency of spatial downscaling. Furthermore, the network downscaling is highly applicable, capable of modeling and fitting different spatial information objects and adjusting network structure parameters in real time according to different downscaling needs, changing the downscaling range and accuracy.

[0005] The technical solutions adopted by the present invention are as follows:

[0006] A sea surface temperature spatial statistical downscaling method based on a U-shaped convolutional neural network specifically includes the following steps:

[0007] Step 1: Obtain statistical data to be analyzed and perform normalization preprocessing;

[0008] Step 2: Construct a data set for network training and divide it into training set and test set;

[0009] Step 3: Construct a U-shaped convolutional neural network and set the optimizer parameters; train the U-shaped convolutional neural network using the training set and the test set, and output the trained U-shaped convolutional neural network;

[0010] Step 4: Evaluate the downscaling effect of the U-shaped convolutional neural network and output the U-shaped convolutional neural network;

[0011] Step 5: Obtain real-time low-resolution ocean temperature data, and after preprocessing through the above step 1, use the trained U-shaped convolutional neural network to output the downscaled high-resolution ocean temperature data.

[0012] Preferably, the data normalization preprocessing process in step 1 adopts the maximum and minimum normalization method, and the formula is as follows:

[0013]

[0014] In the formula, t and are the temperature data before and after normalization, t min is the minimum value of the temperature data, t max is the maximum value of the temperature data; first find the maximum value of the temperature data t from the image data set min and minimum value t max , and then normalize the grid temperature data on each image according to the maximum and minimum normalization formula in turn, and finally output the normalized data set.

[0015] Preferably, in step 2, historical ocean temperature data are first selected and divided into a training set and a test set according to the ratio of "training set: test set = 5:1"; the low-resolution images of the training set are used as input data for network training, and the high-resolution images are used to calculate the training loss during network training; the low-resolution images of the test set are used as input data for network testing, and the high-resolution images are used to calculate the test loss during network testing to adjust the network learning rate.

[0016] Preferably, in step 3, a neural network downscaling training model with a U-Net structure as the core performs downscaling learning on the low-resolution image data of sea surface temperature in the ocean temperature historical data, and continuously learns and adjusts the network parameters by dynamically adjusting the learning rate during training, and finally outputs the trained high-resolution image data; the specific steps of U-type convolutional neural network training are as follows:

[0017] Step 301: Initialize network-related parameters: initial learning rate, number of training times, loss function, learning rate reduction coefficient, and optimizer;

[0018] Step 302: The training set enters the U-Net network for training and the training error is calculated;

[0019] Step 303: After each training round, the U-Net network is tested with the test set to calculate the test error.

[0020] Step 304: Adjust the learning rate according to the change of the test error;

[0021] Step 305: Enter the network training loop again, re-normalize and pre-process the data, and divide the data into training set and test set, and repeat steps 301 to 305 until the training times are completed; save the trained network parameters.

[0022] Preferably, in step 4, the downscaling effect of the U-shaped convolutional neural network is firstly judged by comparing it with the traditional interpolation method in an intuitive visual way; in order to more accurately and specifically determine the downscaling effect of the U-shaped convolutional neural network, the root mean square error (RMSE), peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) are then selected as evaluation indicators to perform a quantitative comparison with the traditional interpolation downscaling; the downscaling effect of the U-shaped convolutional neural network is judged, and finally the kernel density estimation method (KDE) is used to evaluate the distribution of the forecast errors of different downscaling schemes, and the forecast effects of the downscaling schemes are compared and verified.

[0023] Preferably, when preliminarily evaluating the downscaling effect of the U-shaped convolutional neural network, the image features are compared with the traditional interpolation method. In order to more intuitively compare the performance gap between the neural network downscaling model and the bilinear interpolation downscaling model, the forecast images output by each downscaling model can be directly compared.

[0024] Preferably, when selecting the root mean square error (RMSE), peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) as evaluation indicators, in order to more comprehensively evaluate the downscaling effect of each model, the present invention selects three evaluation indicators commonly used to compare the degree of image fitting: root mean square error (RMSE), peak signal-to-noise ratio (PSNR) and structural similarity (SSIM).

[0025] The root mean square error (RMS) is used to measure the difference between the predicted value and the true value, which meets the evaluation requirement of downscaling low-resolution SST images to fit the true high-resolution images. The formula for the RMS error is shown below:

[0026]

[0027] Where x is the real data value, the real high-resolution sea temperature data;

[0028] is the forecast data value, downscaled forecast sea temperature data;

[0029] N is the data volume, the number of data points for each high-resolution image;

[0030] RMSE measures the accuracy of the downscaling model fitting. The smaller the RMSE value, the closer the downscaling model prediction effect is to the real data, and the higher the accuracy of the downscaling model.

[0031] The peak signal-to-noise ratio (PSNR) also evaluates the degree of image restoration based on the deviation between the predicted image and the true image. The formula for the PSNR is as follows:

[0032]

[0033] Where MSE and RMSE are the mean square error and root mean square error between the predicted image and the true image, respectively;

[0034] n is the number of bits per pixel, which is usually 8 for color images;

[0035] 2 n -1 is the maximum value of the pixel color, which is 255 when n is 8 bits;

[0036] The unit of PSNR is decibel (dB). From the changes in MSE and RMSE, we can see that the higher the PSNR value, the smaller the degree of deviation of the restored image from the real image, and the higher the accuracy of the downscaling model;

[0037] Structural similarity evaluates the similarity of images from three perspectives: brightness, contrast, and structure. The structural similarity formula is as follows:

[0038] SSIM(x,y)=[l(x,y)] α [c(x,y)] β [s(x,y)] γ

[0039] Where x and y are the two images to be compared; the predicted image and the real image;

[0040] α, β, and γ are weight adjustments for brightness, contrast, and structure information, respectively. α, β, and γ are all greater than 0;

[0041] l(x, y) is the brightness comparison function, c(x, y) is the contrast comparison function, and s(x, y) is the structure comparison function, which can be expressed as the following formulas respectively;

[0042]

[0043] In the above three formulas, μ x and μy are the mean brightness values of images x and y respectively, σ x and σ y are the standard deviations of images x and y, σ xy is the covariance of images x and y, c1, c2, c3 are constants;

[0044] In actual engineering calculations of SSIM, α, β, and γ are all set to 1, and c2 = c3 / 2. The SSIM formula can be simplified to the following:

[0045]

[0046] The SSIM function is symmetrical, and the SSIM value is between 0 and 1. The closer the SSIM value is to 1, the better the image quality is, the closer it is to the real image, and the more accurate the prediction effect is.

[0047] Preferably, when finally using the kernel density estimation method (KDE) to evaluate the distribution of forecast errors of different downscaling schemes, kernel density estimation is a non-parametric method for estimating the probability density function of the fitted sample data; the definition function of kernel density estimation is shown as follows:

[0048]

[0049] Where x i is the sample point, h is the bandwidth and h>0;

[0050] K(·) is a kernel function. The present invention adopts a Gaussian distribution kernel function, and the formula is shown below:

[0051]

[0052] The technical effects achieved by the present invention are:

[0053] In the downscaling of spatial feature information with extreme, discrete and other nonlinear changes, the present invention uses a U-shaped convolutional neural network to effectively solve the problems of increased spatial downscaling deviation and even reconstruction feature errors caused by irregular and variable details of linear regression models, and well fits and restores nonlinear features and detail change features. In the U-Net spatial downscaling process, the overall network structure is an "encoder-decoder" structure. After network training, it can effectively extract spatial features, nonlinearly map feature relationships and reconstruct feature functions. In addition, the network also includes a "skip connection" structure, which combines the detailed features in the shallow network feature map with the abstract information in the deep network feature map, and well solves the shortcoming of deep convolutional neural network super-resolution reconstruction that is prone to missing shallow feature details, thereby improving the accuracy of spatial downscaling. The method proposed in the present invention can effectively improve the prediction accuracy and efficiency of spatial downscaling. At the same time, the network downscaling has strong applicability. It can not only perform modeling and fitting for different spatial information objects, but also adjust the network structure parameters in real time according to different downscaling needs to change the downscaling range and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 This is a flow chart of a sea surface temperature spatial statistical downscaling method based on a U-shaped convolutional neural network according to the present invention;

[0055] Figure 2 This is a flowchart of data set partitioning in a specific embodiment of the present invention;

[0056] Figure 3 It is a flow chart of the spatial downscaling training model of the U-shaped convolutional neural network of the present invention;

[0057] Figure 4This is a comparison chart of the 1 / 2° downscaling forecast effect in the present invention;

[0058] Figure 5 is a comparison curve of the root mean square error of downscaling prediction in the present invention;

[0059] Figure 6 It is a schematic diagram of the U-Net network structure in the present invention;

[0060] Figure 7 It is a bilinear interpolation downscaling flow chart in the prior art;

[0061] Figure 8 It is a visualization flow chart of the U-shaped convolutional neural network training results in the present invention;

[0062] Figure 9 This is an error distribution analysis diagram of the downscaling scheme on a certain day in the present invention. DETAILED DESCRIPTION

[0063] In order to make the purpose and advantages of the present invention more clearly understood, the present invention is described in detail below with reference to the following examples. It should be understood that the following text is only used to describe one or more specific embodiments of the present invention and does not strictly limit the scope of protection of the present invention.

[0064] like Figures 1-9 As shown in FIG, a spatial statistical downscaling method of sea surface temperature based on a U-shaped convolutional neural network includes the following steps:

[0065] Step 1: Obtain statistical data to be analyzed and perform normalization preprocessing;

[0066] Step 2: Construct a data set for network training and divide it into training set and test set;

[0067] Step 3: Construct a U-shaped convolutional neural network, such as Figure 6 As shown, the optimizer parameters are set; the U-shaped convolutional neural network is trained through the training set and the test set, and the trained U-shaped convolutional neural network is output; the U-Net network can also extract and restore image features well, and the image super-resolution performance is relatively excellent. When the U-Net network performs spatial downscaling, its "encoding-decoding" structure and "jump connection" structure can extract and combine the detailed features in the shallow network feature map and the abstract information in the deep network feature map, which well solves the shortcoming of the deep convolutional neural network that easily misses the details of shallow features during super-resolution reconstruction, thereby being able to more accurately restore image features and having relatively excellent spatial downscaling performance;

[0068] Step 4: Evaluate the downscaling effect of the U-shaped convolutional neural network and output the U-shaped convolutional neural network;

[0069] Step 5: Obtain real-time low-resolution ocean temperature data, and after preprocessing through step 1, use the trained U-shaped convolutional neural network to output the downscaled high-resolution ocean temperature data.

[0070] The present invention first applies a U-shaped convolutional neural network to statistical downscaling. This method can well extract and restore the details and nonlinear features in spatial information, effectively reduce the errors caused by nonlinear relationships, and improve the accuracy of statistical downscaling.

[0071] Preferably, the data normalization preprocessing process in step 1 adopts the maximum and minimum normalization method, and the formula is as follows:

[0072]

[0073] In the formula, t and are the temperature data before and after normalization, t min is the minimum value of the temperature data, t max is the maximum value of the temperature data; first find the maximum value of the temperature data t from the image data set min and minimum value t max , and then normalize the grid temperature data on each image according to the maximum and minimum normalization formula in turn, and finally output the normalized data set.

[0074] In this method, the original reanalysis data is first obtained by selecting the research spatial data elements, spatiotemporal scope, and initial scale of the raster data based on the research requirements of spatial downscaling. However, the original spatial raster reanalysis data may contain complex situations such as the intersection of data sampling monitoring areas and non-data sampling monitoring areas, requiring preliminary raster data screening and differentiation processing. In addition, in order to accelerate the convergence speed of neural network training and improve the accuracy of neural network training, the original reanalysis data needs to be normalized to eliminate the influence of different units between the variable data and suppress the adverse effects of outlier data points on network training.

[0075] Preferably, in step 2, historical ocean temperature data are first selected and divided into a training set and a test set according to the ratio of "training set: test set = 5:1"; the low-resolution images of the training set are used as input data for network training, and the high-resolution images are used to calculate the training loss during network training; the low-resolution images of the test set are used as input data for network testing, and the high-resolution images are used to calculate the test loss during network testing to adjust the network learning rate.

[0076] The data set division process in the present invention is not only to facilitate the training and testing of the neural network, but also to facilitate the subsequent visualization comparison of the neural network downscaling effect with the traditional interpolation downscaling effect, thereby dividing the data set into a training set and a test set. The training set is the sea surface temperature image data from 2015 to 2018, and the test set is the sea surface temperature image data from 2019. For the training network part, the low-resolution images of the training set are used as input data for network training, and the high-resolution images are used to calculate the training loss during network training; the low-resolution images of the test set are used as input data for network testing, and the high-resolution images are used to calculate the test loss during network testing to adjust the network learning rate. For the visualization part of the network training effect and the interpolation effect, in order to facilitate the subsequent unified comparative analysis, the low-resolution images of the test set are used as the visualization input, so that high-resolution images of the downscaling forecast can be obtained, and the high-resolution images are used as the benchmark for evaluating the downscaling effect. The data set division process is as follows: Figure 2 shown.

[0077] Preferably, in step 3, a neural network downscaling training model with a U-Net structure as the core performs downscaling learning on the low-resolution image data of sea surface temperature in the ocean temperature historical data, and continuously learns and adjusts the network parameters by dynamically adjusting the learning rate during training, and finally outputs the trained high-resolution image data; the specific steps of U-shaped convolutional neural network training are as follows:

[0078] Step 301: Initialize network-related parameters: initial learning rate, number of training times, loss function, learning rate reduction coefficient, and optimizer;

[0079] Step 302: The training set enters the U-Net network for training and the training error is calculated;

[0080] Step 303: After each training round, the U-Net network is tested with the test set to calculate the test error.

[0081] Step 304: Adjust the learning rate according to the change of the test error;

[0082] Step 305: Enter the network training loop again, re-normalize and pre-process the data, and divide the data into training set and test set, and repeat steps 301 to 305 until the training times are completed; save the trained network parameters.

[0083] According to the specific steps of the neural network downscaling training model, the overall process of establishing the neural network downscaling training model is drawn as follows Figure 3 As shown;

[0084] The U-shaped convolutional neural network spatial downscaling training model process was implemented in the PyTorch framework. The main idea of the program is to first establish and run the sample dataset partitioning and normalization program, then establish the U-Net network model program, and then establish the network loop training model program. Parameters such as the learning rate, number of training times, loss function, learning rate reduction coefficient, and optimizer are set. Finally, the training and test sets are put into the training model for training and testing, and the training and test loss curves are observed.

[0085] Preferably, in step 4, the downscaling effect of the U-shaped convolutional neural network is firstly judged by comparing it with the traditional interpolation method in an intuitive visual way; in order to more accurately and specifically judge the downscaling effect of the U-shaped convolutional neural network, the root mean square error (RMSE), peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) are then selected as evaluation indicators to perform quantitative comparison with the traditional interpolation downscaling; the downscaling effect of the U-shaped convolutional neural network is judged, and finally the kernel density estimation method (KDE) is used to evaluate the distribution of the forecast errors of different downscaling schemes, and the forecast effects of the downscaling schemes are compared and verified.

[0086] To more accurately and comprehensively evaluate the downscaling performance of various models, this paper selected three commonly used evaluation metrics for comparing image fit: root mean square error, peak signal-to-noise ratio, and structural similarity. These three metrics can assess the deviation between the downscaled image and the true image, quantitatively describing the quality of the restored image and serving as a basis for scientifically evaluating the spatial downscaling performance of various downscaling methods.

[0087] like Figure 7 as well as Figure 8 As shown, preferably, when preliminarily judging the downscaling effect of the U-type convolutional neural network, the image features are compared with the traditional interpolation method. In order to more intuitively compare the performance gap between the neural network downscaling model and the bilinear interpolation downscaling model, the forecast images output by each downscaling model can be directly compared.

[0088] To directly compare the performance gap between the neural network downscaling model and the bilinear interpolation downscaling model, we can directly compare the predicted images output by each downscaling model. By comparing the differences in the detailed features of the real high-resolution image and the output images of each spatial downscaling method, we can preliminarily judge the performance of the U-Net and traditional downscaling methods.

[0089] Preferably, when selecting the root mean square error (RMSE), peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) as evaluation indicators, in order to more comprehensively evaluate the downscaling effect of each model, the present invention selects three evaluation indicators commonly used to compare the degree of image fitting: root mean square error (RMSE), peak signal-to-noise ratio (PSNR) and structural similarity (SSIM).

[0090] The root mean square error (RMS) is used to measure the difference between the predicted value and the true value, which meets the evaluation requirement of downscaling low-resolution SST images to fit the true high-resolution images. The formula for the RMS error is shown below:

[0091]

[0092] Where x is the real data value, the real high-resolution sea temperature data;

[0093] is the forecast data value, downscaled forecast sea temperature data;

[0094] N is the data volume, the number of data points for each high-resolution image;

[0095] RMSE measures the accuracy of the downscaling model fitting. The smaller the RMSE value, the closer the downscaling model prediction effect is to the real data, and the higher the accuracy of the downscaling model.

[0096] The peak signal-to-noise ratio (PSNR) also evaluates the degree of image restoration based on the deviation between the predicted image and the true image. The formula for the PSNR is as follows:

[0097]

[0098] Where MSE and RMSE are the mean square error and root mean square error between the predicted image and the true image, respectively;

[0099] n is the number of bits per pixel, which is usually 8 for color images;

[0100] 2 n -1 is the maximum value of the pixel color, which is 255 when n is 8 bits;

[0101] The unit of PSNR is decibel (dB). From the changes in MSE and RMSE, we can see that the higher the PSNR value, the smaller the degree of deviation of the restored image from the real image, and the higher the accuracy of the downscaling model;

[0102] Structural similarity evaluates the similarity of images from three perspectives: brightness, contrast, and structure. The structural similarity formula is as follows:

[0103] SSIM(x,y)=[l(x,y)] α [c(x,y)] β [s(x,y)] γ

[0104] Where x and y are the two images to be compared; the predicted image and the real image;

[0105] α, β, and γ are weight adjustments for brightness, contrast, and structure information, respectively. α, β, and γ are all greater than 0;

[0106] l(x, y) is the brightness comparison function, c(x, y) is the contrast comparison function, and s(x, y) is the structure comparison function, which can be expressed as the following formulas respectively;

[0107]

[0108]

[0109] In the above three formulas, μ x and μ y are the mean brightness values of images x and y, σ x and σ y are the standard deviations of images x and y, σ xy is the covariance of images x and y, c1, c2, c3 are constants;

[0110] In actual engineering calculations of SSIM, α, β, and γ are all set to 1, and c2 = c3 / 2. The SSIM formula can be simplified to the following:

[0111]

[0112] The SSIM function is symmetrical, and the SSIM value is between 0 and 1. The closer the SSIM value is to 1, the better the image quality is, the closer it is to the real image, and the more accurate the prediction effect is.

[0113] Preferably, when finally using the kernel density estimation method (KDE) to evaluate the distribution of forecast errors of different downscaling schemes, kernel density estimation is a non-parametric method for estimating the probability density function of the fitted sample data; the definition function of kernel density estimation is shown as follows:

[0114]

[0115] Where x i is the sample point, h is the bandwidth, and h>0;

[0116] K(·) is a kernel function. The present invention adopts a Gaussian distribution kernel function, and the formula is shown below:

[0117]

[0118] In the present invention, the effect of downscaling forecast can be evaluated based on the shape of the probability density curve output by the kernel density estimation method, such as Figure 9 As shown in the figure, the bilinear interpolation downscaling probability density curve is "low and fat" and the distribution is relatively dispersed, which means that the bilinear interpolation downscaling prediction error is large and the prediction accuracy is poor; while the U-shaped convolutional neural network downscaling probability density curve is "tall and thin" and the distribution is very concentrated, which means that the neural network downscaling prediction error is generally small, most of the errors are distributed near 0, and the prediction accuracy is high.

[0119] To further verify and compare the performance differences between different downscaling schemes, this paper uses kernel density estimation (KDE) to evaluate the distribution of prediction errors for different downscaling schemes from a probabilistic statistical perspective. The KDE method can visualize the degree of error in image restoration and better reflect the accuracy of spatial downscaling, making it a suitable method for verifying and evaluating spatial downscaling performance.

[0120] The present invention adopts the U-shaped convolutional neural network method in the spatial downscaling forecast scheme. Compared with the traditional interpolation downscaling forecast, this method can effectively improve the downscaling forecast accuracy. The image and root mean square error of the U-shaped convolutional neural network downscaling forecast result are significantly better than the bilinear interpolation downscaling forecast method. The comparison of the forecast result image and root mean square error is as follows Figure 4 as well as Figure 5 shown.

[0121] Compared with bilinear interpolation downscaling, the U-shaped convolutional neural network downscaling prediction effect is better throughout the year, its application range is wider, and the forecast data is more accurate. Especially in the winter and spring when the image data differ greatly, its fitting advantage is more prominent.

[0122] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art may make various improvements and modifications without departing from the principles of the present invention, and such improvements and modifications are also within the scope of protection of the present invention. Structures, devices, and operating methods not specifically described or explained herein shall, unless otherwise specified or limited, be implemented in accordance with conventional means in the art.

Claims

1. A sea surface temperature spatial statistical downscaling method based on a U-shaped convolutional neural network, characterized by: The specific steps include: Step 1: Obtain statistical data to be analyzed and perform normalization preprocessing; Step 2: Construct a data set for network training and divide it into training set and test set; Step 3: Construct a U-shaped convolutional neural network and set the optimizer parameters; train the U-shaped convolutional neural network using the training set and the test set, and output the trained U-shaped convolutional neural network; Step 4: Evaluate the downscaling effect of the U-shaped convolutional neural network and output the U-shaped convolutional neural network; Step 5: Obtain real-time low-resolution ocean temperature data, and after preprocessing through the above step 1, use the trained U-shaped convolutional neural network to output the downscaled high-resolution ocean temperature data.

2. The method for spatial statistical downscaling of sea surface temperature based on a U-shaped convolutional neural network according to claim 1, characterized in that: The data normalization preprocessing process in step 1 adopts the maximum and minimum normalization method, and the formula is as follows: In the formula, t and are the temperature data before and after normalization, t min is the minimum value of the temperature data, t max is the maximum value of the temperature data; first find the maximum value of the temperature data t from the image data set min and minimum value t max , and then normalize the grid temperature data on each image according to the maximum and minimum normalization formula in turn, and finally output the normalized data set.

3. The method for spatial statistical downscaling of sea surface temperature based on a U-shaped convolutional neural network according to claim 2, characterized in that: In step 2, historical ocean temperature data is first selected and divided into a training set and a test set according to a ratio of "training set:test set = 5:1"; the low-resolution images in the training set are used as input data for network training, while the high-resolution images are used to calculate the training loss during network training; The low-resolution images of the test set are used as input data for network testing, while the high-resolution images are used to calculate the test loss during network testing to adjust the network learning rate.

4. The method for spatial statistical downscaling of sea surface temperature based on a U-shaped convolutional neural network according to claim 3, characterized in that: In step 3, a neural network downscaling training model with a U-Net structure as the core performs downscaling learning on the low-resolution sea surface temperature image data in the ocean temperature historical data, and continuously adjusts the network parameters by dynamically adjusting the learning rate during training, and finally outputs the trained high-resolution image data; the specific steps of the U-shaped convolutional neural network training are as follows: Step 301: Initialize network-related parameters: initial learning rate, number of training times, loss function, learning rate reduction coefficient, and optimizer; Step 302: The training set enters the U-Net network for training and the training error is calculated; Step 303: After each training round, the U-Net network is tested with the test set to calculate the test error. Step 304: Adjust the learning rate according to the change of the test error; Step 305: Enter the network training loop again, re-normalize and pre-process the data, and divide the data into training set and test set, and repeat steps 301 to 305 until the training times are completed; save the trained network parameters.

5. The method for spatial statistical downscaling of sea surface temperature based on a U-shaped convolutional neural network according to claim 1, characterized in that: In step 4, the downscaling effect of the U-shaped convolutional neural network is firstly evaluated by visual comparison with the traditional interpolation method; then, the root mean square error (RMSE), peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) are selected as evaluation indicators to perform quantitative comparison with the traditional interpolation downscaling; the downscaling effect of the U-shaped convolutional neural network is evaluated, and finally, the kernel density estimation method (KDE) is used to evaluate the distribution of the prediction errors of different downscaling schemes, and the prediction effects of the downscaling schemes are compared and verified.

6. The method for spatial statistical downscaling of sea surface temperature based on a U-shaped convolutional neural network according to claim 5, characterized in that: When preliminarily evaluating the downscaling effect of the U-shaped convolutional neural network, the image features are compared with the traditional interpolation method, and the forecast images output by each downscaling model are directly compared.

7. The method for spatial statistical downscaling of sea surface temperature based on a U-shaped convolutional neural network according to claim 5, characterized in that: When the root mean square error (RMSE), peak signal-to-noise ratio (PSNR), and structural similarity (SSIM) are selected as evaluation indicators, the formula for the root mean square error is as follows: Where x is the real data value, the real high-resolution sea temperature data; is the forecast data value, downscaled forecast sea temperature data; N is the data volume, the number of data points for each high-resolution image; RMSE measures the accuracy of the downscaling model fitting. The smaller the RMSE value, the closer the downscaling model prediction effect is to the real data, and the higher the accuracy of the downscaling model. The peak signal-to-noise ratio (PSNR) also evaluates the degree of image restoration based on the deviation between the predicted image and the true image. The formula for the PSNR is as follows: Where MSE and RMSE are the mean square error and root mean square error between the predicted image and the true image, respectively; n is the number of bits per pixel, 8 for color images; 2 n -1 is the maximum value of the pixel color, which is 255 when n is 8 bits; The unit of PSNR is decibel (dB). From the changes in MSE and RMSE, we can see that the higher the PSNR value, the smaller the degree of deviation of the restored image from the real image, and the higher the accuracy of the downscaling model; Structural similarity evaluates the similarity of images from three perspectives: brightness, contrast, and structure. The structural similarity formula is as follows: SSIM(x,y)=[l(x,y)] α [c(x,y)] β [s(x,y)] γ Where x and y are the two images to be compared; the predicted image and the real image; α, β, and γ are weight adjustments for brightness, contrast, and structure information, respectively. α, β, and γ are all greater than 0; l(x, y) is the brightness comparison function, c(x, y) is the contrast comparison function, and s(x, y) is the structure comparison function, which can be expressed as the following formulas respectively; In the above three formulas, μ x and μ y are the mean brightness values of images x and y, σ x and σ y are the standard deviations of images x and y, σ xy is the covariance of images x and y, c1, c2, c3 are constants; The SSIM function is symmetrical, and the SSIM value is between 0 and 1. The closer the SSIM value is to 1, the better the image quality is, the closer it is to the real image, and the more accurate the prediction effect is.

8. The method for spatial statistical downscaling of sea surface temperature based on a U-shaped convolutional neural network according to claim 5, characterized in that: Finally, the kernel density estimation method (KDE) is used to evaluate the distribution of forecast errors of different downscaling schemes. Kernel density estimation is a non-parametric method used to estimate the probability density function of the fitted sample data; The definition function of kernel density estimation is as follows: Where x i is the sample point, h is the bandwidth, and h>0; K(·) is a kernel function. The present invention adopts a Gaussian distribution kernel function, and the formula is shown below:

Images