A mask-enhanced intelligent spatial downscaling method for ocean feature fields

Abstract

The present invention relates to a mask-enhanced intelligent spatial downscaling method for marine element fields, comprising the following steps: marine data processing and sample set construction; model construction; result verification. After marine data processing and sample set construction, an intelligent spatial downscaling model MED-Net is constructed, and the model is enhanced by introducing sea and land masks and partial convolution. At the same time, a branch network is used to extract the features of high-resolution climate data and input them into the backbone network. The objective function combines the ocean loss and the sea-land boundary loss to improve the downscaling performance and accuracy of the model. The intelligent spatial downscaling model is tested using a test set, and the performance of the model is evaluated by indicators such as video memory occupancy, parameter quantity, and time consumption, and the accuracy of the model is evaluated. It has efficient computing capabilities while maintaining good accuracy, generating more refined high-resolution marine element fields, and providing new methods and new ideas for the research and development of marine environmental spatial downscaling technology.

Description

Technical Field

[0001] The present invention relates to the field of marine science and machine learning technology, and in particular to a mask-enhanced intelligent spatial downscaling method for marine element fields. Background Art

[0002] The marine environment plays a vital role in the Earth's ecosystem, has a profound impact on climate change, and is closely related to human production and life. To carry out marine scientific research and further understand and model the evolution of the marine environment, high-resolution marine element data is an important prerequisite and necessary condition.

[0003] However, the high computational cost of ocean numerical models limits the availability of fine-scale ocean forecasts, causing certain difficulties for the timeliness of ocean models and scientific research. The balance between computational efficiency and timeliness has become a key and challenging issue in the pursuit of high-resolution modeling and forecasting. Therefore, downscaling technology has become a research hotspot in the field of atmosphere and ocean.

[0004] Downscaling is the process of converting large-scale, low-resolution information into small-scale, high-resolution information. There are currently two commonly used downscaling methods: dynamical downscaling and statistical downscaling. The dynamical downscaling method has good physical significance, but requires a lot of computational cost; while the traditional statistical downscaling method cannot accurately establish the mapping relationship between low-resolution data and high-resolution data, resulting in poor simulation and prediction accuracy and low reliability.

[0005] With the development of deep learning, researchers have found that neural networks can be used to capture the nonlinear relationship between low-resolution and high-resolution data. Image super-resolution is an important application to improve the resolution of images and videos. There have been some application results in the field of atmosphere and ocean, but they still have limitations, including poor spatial generalization ability and difficulty in capturing important features. Summary of the invention

[0006] In view of the above-mentioned shortcomings of the prior art, an object of the present invention is to provide a mask-enhanced intelligent spatial downscaling method for ocean feature fields to solve one or more problems in the prior art.

[0007] To achieve the above object, the technical solution of the present invention is as follows:

[0008] A mask-enhanced intelligent spatial downscaling method for ocean feature fields comprises the following steps:

[0009] Ocean data processing and sample set construction;

[0010] Model building includes the following steps:

[0011] Introduce training validation set samples and partial convolution and build an intelligent spatial downscaling model, where the partial convolution includes the following steps:

[0012] Split the input ocean feature field into several small blocks of the same size.

[0013] Apply a standard convolution operation to each small block, that is, multiply the convolution kernel by the feature value of the small block and sum them to get the convolution result, as shown in the following formula:

[0014]

[0015] Where x′ is the value of the center position of the convolution kernel, W is the convolution kernel, X is the ocean feature field patch, sum(1) refers to a matrix of the same size as the convolution kernel with all elements set to 1, M is the input land and sea mask, where the ocean part is 1 and the land part is 0, and b is the bias of the convolution operation.

[0016] Combine the partial convolution results of all small blocks to obtain partial convolution features.

[0017] The intelligent spatial downscaling model includes a backbone network f D and a branch network f C , the construction steps include the following:

[0018] Branch Network C The high-resolution climate state data HRCM is convolved to obtain the low-resolution climate state.

[0019] It is concatenated with the low-resolution ocean data LROD and input into the backbone network f D ,

[0020] Adjust the size of high-resolution climate data HRCM,

[0021] From the backbone network D Extract high-dimensional features and combine them with processed high-resolution climate data HRCM.

[0022] Optimize the objective function;

[0023] Result verification.

[0024] Furthermore, the ocean data processing and sample set construction includes the following steps:

[0025] Perform quality control on raw high-resolution ocean data HROD;

[0026] Calculate high-resolution climate data HRCM;

[0027] The downsampling process obtains low-resolution ocean data LROD;

[0028] Normalize the maximum and minimum values ​​of all data and create training validation sets and test sets.

[0029] Furthermore, the objective function calculation formula is as follows:

[0030]

[0031] In the formula: and β are the set coefficients, and Represent ocean losses and sea-land boundary losses respectively.

[0032] Furthermore, the optimization method of the optimization objective function includes the following contents:

[0033] Introduce ocean loss to remove the error effect of land;

[0034] The land-sea boundary loss is introduced to solve the influence of gradient changes at the land-sea boundary on the downscaling error.

[0035] Furthermore, the ocean loss calculation formula is as follows:

[0036]

[0037] Where: represents ocean loss, L1 represents L1 loss, M O Represents the ocean part mask, where the ocean grid points are 1 and the land grid points are 0. The operator Indicates that the model prediction result is multiplied by the mask before calculating the L1 loss.

[0038] Furthermore, the calculation formula for the sea-land boundary loss is as follows:

[0039]

[0040] Where: represents ocean loss, L1 represents L1 loss, M B Represents the sea-land boundary mask, where the ocean grid points adjacent to the land are 1 and the land grid points are 0. The operator Indicates that the model prediction result is multiplied by the mask before calculating the L1 loss.

[0041] Further, the result verification comprises the following steps:

[0042] Calculate performance indicators;

[0043] Compute the accuracy index.

[0044] Compared with the prior art, the beneficial technical effects of the present invention are as follows:

[0045] (1) The present invention constructs an intelligent spatial downscaling model, introduces land and sea masks and partial convolution into the model to enhance the downscaling performance of the model in the ocean and at the land-sea boundary, and compensates for the lack of high-frequency and small-scale information by adding high-resolution climate state features, thereby generating a more refined high-resolution ocean element field.

[0046] (ii) By combining ocean loss and sea-land boundary loss in the objective function, the accuracy of the model at the sea-land boundary is further improved. It can be used for intelligent spatial downscaling of various marine element fields, with efficient computing capabilities while maintaining good accuracy, providing new methods and new ideas for the research and development of marine environmental spatial downscaling technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 A schematic flow chart of a mask-enhanced intelligent spatial downscaling method for ocean feature fields according to an embodiment of the present invention is shown.

[0048] Figure 2 A schematic diagram of the MED-Net network model structure of a mask-enhanced ocean feature field intelligent spatial downscaling method according to an embodiment of the present invention is shown.

[0049] Figure 3 A partial convolution conceptual diagram of a mask-enhanced ocean feature field intelligent spatial downscaling method according to an embodiment of the present invention is shown.

[0050] Figure 4 A comparison diagram of error profiles of a mask-enhanced intelligent spatial downscaling method for ocean feature fields according to an embodiment of the present invention and a traditional cubic spline interpolation method in a three-dimensional sea temperature downscaling task is shown. DETAILED DESCRIPTION

[0051] In order to make the purpose, technical scheme and advantages of the present invention clearer, the following is a further detailed description of a mask-enhanced intelligent spatial downscaling method for ocean element fields proposed by the present invention in combination with the accompanying drawings and specific embodiments. According to the following description, the advantages and features of the present invention will be clearer. It should be noted that the accompanying drawings are in a very simplified form and use non-precise proportions, which are only used to conveniently and clearly assist in explaining the purpose of the embodiments of the present invention. In order to make the purpose, features and advantages of the present invention more obvious and easy to understand, please refer to the accompanying drawings. It should be noted that the structure, proportion, size, etc. illustrated in the drawings of this specification are only used to match the content disclosed in the specification for people familiar with this technology to understand and read, and are not used to limit the limiting conditions for the implementation of the present invention, so they have no technical substantive significance. Any modification of the structure, change of the proportional relationship or adjustment of the size, without affecting the effect that the present invention can produce and the purpose that can be achieved, should still fall within the scope of the technical content disclosed by the present invention.

[0052] See also Figure 1 , a mask-enhanced intelligent spatial downscaling method for ocean feature fields, comprising the following steps:

[0053] Step 1: Ocean data processing and sample set construction.

[0054] Step 1.1: Perform quality control on the original high-resolution ocean data HROD, and query, remove, correct and fill in outliers and missing values.

[0055] Step 1.2: Calculate high-resolution climate state data HRCM. Based on the quality-controlled HROD, calculate the multi-year average ocean element field on a monthly basis to obtain the monthly climate state data HRCM, providing a priori high-frequency information for the intelligent model.

[0056] Step 1.3: Downsampling to obtain low-resolution ocean data LROD. Downsample HROD using the nearest neighbor method to obtain low-resolution data LROD, and associate low-resolution ocean data LROD, high-resolution climate data HRCM and high-resolution ocean data HROD one by one. Low-resolution ocean data LROD and high-resolution climate data HRCM are used as model inputs, and high-resolution ocean data HROD is used as a model label.

[0057] Step 1.4: Normalize the maximum and minimum values ​​of all data and create training validation sets and test sets. Randomly select some months as the test set and the remaining samples as the training validation set.

[0058] Further, step 2: model building.

[0059] Step 2.1: Introduce training and validation set samples and partial convolution and build an intelligent spatial downscaling model. Self-build an intelligent spatial downscaling model, the model name is MED-Net, which includes a backbone network f D and a branch network f C , where f D Contains 32 basic blocks with residual structures, f C It includes 2 batch normalization layers and 2 convolutional layers. All convolution kernels in the network structure use partial convolution. It is preferred that experiments in this embodiment prove that 32 residual blocks can bring the best spatial downscaling effect.

[0060] Furthermore, the MED-Net model structure is as follows Figure 2 shown.

[0061] Figure 2 In, f D represents the spatial downscaling backbone network, f Crepresents the climate state feature branch network, NormalizationLayer is the batch normalization layer, Residual Block represents the residual block, PartialConv2d is the partial convolution layer, ReLU is the activation function layer, ⊕ represents the splicing operation, Pixel Shuffle is the pixel reorganization layer, and Dimensionalization Layer is the dimensional conversion layer.

[0062] Furthermore, the steps of building the intelligent spatial downscaling model include the following:

[0063] Step 2.1.1: The first branch network C The high-resolution climate state data HRCM is convolved to obtain the low-resolution climate state.

[0064] Furthermore, a mask enhancement effect is achieved by partial convolution, and the steps of the partial convolution are as follows:

[0065] Step 2.1.1.1: Split the input ocean feature field into several small blocks of equal size.

[0066] Step 2.1.1.2: Apply a standard convolution operation to each small block, that is, multiply the convolution kernel by the element value of the small block and sum them to obtain the convolution result, as shown in Equation 1 below:

[0067]

[0068] Where: X is the input ocean feature field data, M is the sea and land mask, with a value of 0-1 distribution, where the ocean part is 1 and the land part is 0, W is the convolution kernel, b is the bias of the convolution operation, sum(1) refers to a matrix with the same size as the convolution kernel and all elements are 1, and the output x' is the value of the center position of the convolution kernel. Here we only discuss the pixels in the current operation area. For example, if the size of the W convolution kernel is 3×3, then the size of X and M is also 3×3.

[0069] Step 2.1.1.3: Combine the partial convolution results of all small blocks to obtain partial convolution features.

[0070] Furthermore, the concept of partial convolution is as follows Figure 3 As shown, this operation only works in the region where M is 1, and the result enhances times, so some convolutions in MED-Net are more sensitive to ocean areas.

[0071] Step 2.1.2: Then concatenate it with the low-resolution ocean data LROD and use it as the backbone network f D Input.

[0072] Step 2.1.3: The second time, the high-resolution climate data HRCM is resized.

[0073] Step 2.1.4: Then connect to the backbone network f D Concatenation of extracted high-dimensional features.

[0074] Step 2.2: Optimize the objective function.

[0075] The objective function calculation formula is shown in Formula 2 below:

[0076]

[0077] In the formula: and β are setting coefficients, and Respectively represent the ocean loss and the sea-land boundary loss. Preferably, in this embodiment, they are respectively set as and β = 0.1.

[0078] Furthermore, the optimization method of the optimization objective function includes the following contents:

[0079] The ocean loss is introduced to remove the error effect of land: the ocean loss calculation formula is shown in the following formula 3:

[0080]

[0081] Where: represents ocean loss, L1 represents L1 loss, M O Represents the ocean part mask, where the ocean grid points are 1 and the land grid points are 0. The operator Indicates that the model prediction result is multiplied by the mask before calculating the L1 loss. Ocean loss means that only the L1 loss of the ocean part is calculated.

[0082] The land-sea boundary loss is introduced to solve the influence of gradient change at the land-sea boundary on the downscaling error: The land-sea boundary loss calculation formula is shown in Formula 4 below:

[0083]

[0084] Where: represents ocean loss, L1 represents L1 loss, M O Represents the ocean part mask, where the ocean grid points are 1 and the land grid points are 0. The operator Indicates that the model prediction result is multiplied by the mask before calculating the L1 loss. The land-sea boundary loss means that only the L1 loss of a circle of ocean grid points on the edge of the land is calculated.

[0085] Further, step 3: result verification: the test set samples are combined with the intelligent spatial downscaling model built in step 2 to form a mask-enhanced spatial downscaling model and output the model downscaling results.

[0086] Step 3.1: Calculate performance indicators. Analyze the performance of the model based on the memory usage, number of parameters, and time consumption of the spatial downscaling model during testing.

[0087] Step 3.2: Calculate accuracy indicators. According to the model downscaling results and the test sample labels, calculate error indicators such as mean absolute error, root mean square error, structural similarity, and peak signal-to-noise ratio to analyze the accuracy of the model.

[0088] Further, the process ends.

[0089] Specifically, the method of the present invention performs intelligent downscaling on the three-dimensional sea temperature in a certain sea area and compares it with the traditional cubic spline interpolation method to obtain the data shown in Table 1 and Figure 4 As shown:

[0090]

[0091] Table 1

[0092] Compared with the traditional cubic spline interpolation method, the MAE of the MED-Net intelligent spatial downscaling result is reduced by 62.9%, the RMSE is reduced by 64.11%, the SSIM is increased by 6.5%, and the PSNR is increased by 19.8%, indicating that the method described in the present invention is significantly superior to the traditional cubic spline interpolation method.

[0093] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0094] The above-mentioned embodiments only express several implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A mask-enhanced intelligent spatial downscaling method for ocean feature fields, characterized by The following steps are involved: Ocean data processing and sample set construction; Model building includes the following steps: Introduce training validation set samples and partial convolution and build an intelligent spatial downscaling model, where the partial convolution includes the following steps: Split the input ocean feature field into several small blocks of the same size. Apply a standard convolution operation to each small block, that is, multiply the convolution kernel by the feature value of the small block and sum them to get the convolution result, as shown in the following formula: Where x′ is the value of the center position of the convolution kernel, W is the convolution kernel, X is the ocean feature field patch, sum(1) refers to a matrix of the same size as the convolution kernel with all elements set to 1, M is the input land and sea mask, where the ocean part is 1 and the land part is 0, and b is the bias of the convolution operation. Combine the partial convolution results of all small blocks to obtain partial convolution features. The intelligent spatial downscaling model includes a backbone network f D and a branch network f C , the construction steps include the following: Branch Network C Convolve the high-resolution climate state data HRCM to obtain the low-resolution climate state. It is concatenated with the low-resolution ocean data LROD and input into the backbone network f D , Adjust the size of high-resolution climate data HRCM, From the backbone network D Extract high-dimensional features and combine them with processed high-resolution climate data HRCM to optimize the objective function; Result verification.

2. A mask-enhanced intelligent spatial downscaling method for ocean feature fields as claimed in claim 1, characterized in that The ocean data processing and sample set construction include the following steps: Perform quality control on raw high-resolution ocean data HROD; Calculate high-resolution climate data HRCM; The downsampling process obtains low-resolution ocean data LROD; Normalize the maximum and minimum values ​​of all data and create training validation sets and test sets.

3. A mask-enhanced intelligent spatial downscaling method for ocean feature fields as claimed in claim 2, characterized in that The calculation formula of the objective function L is as follows: In the formula: and β are the set coefficients, and Represent ocean losses and sea-land boundary losses respectively.

4. A mask-enhanced intelligent spatial downscaling method for ocean feature fields as claimed in claim 3, characterized in that The optimization method of the optimization objective function includes the following contents: Introduce ocean loss to remove the error effect of land; The land-sea boundary loss is introduced to solve the influence of gradient changes at the land-sea boundary on the downscaling error.

5. A mask-enhanced intelligent spatial downscaling method for ocean feature fields as claimed in claim 4, characterized in that The ocean loss calculation formula is as follows: Where: represents ocean loss, L1 represents L1 loss, which is a machine learning loss function that represents the total average of the absolute value of the difference between the predicted value and the true value, and M O Represents the ocean part mask, where the ocean grid points are 1 and the land grid points are 0. The operator Indicates that the model prediction result is multiplied by the mask before calculating the L1 loss.

6. A mask-enhanced intelligent spatial downscaling method for ocean feature fields as claimed in claim 5, characterized in that The calculation formula of the sea-land boundary loss is as follows: Where: represents ocean loss, L1 represents L1 loss, which is a machine learning loss function that represents the total average of the absolute value of the difference between the predicted value and the true value, and M B Represents the sea-land boundary mask, where the ocean grid points adjacent to the land are 1 and the land grid points are 0. The operator Indicates that the model prediction result is multiplied by the mask before calculating the L1 loss.

7. A mask-enhanced ocean feature field intelligent spatial downscaling method as claimed in claim 6, characterized in that The result verification comprises the following steps: Calculate performance indicators; Compute the accuracy index.

Images