Atmospheric temperature and humidity profile repairing method integrating statistical method and deep learning technology

Abstract

The invention provides an atmospheric temperature and humidity profile repairing method integrating a statistical method and a deep learning technology, belongs to the technical field of meteorological observation, and aims to solve the problem of large-area data missing caused by cloud pollution by taking an FY-4B satellite temperature and humidity profile product as a processing object. Data reconstruction is realized through a three-stage architecture of random cloud mask generation, statistical filling and pre-restoration, and deep learning and fine correction; according to the method, the problem of data truth value missing is solved, meanwhile, the advantages of a statistical method and a deep learning technology are combined, the repair precision in an extreme missing scene is remarkably improved while physical reasonability is guaranteed, and meanwhile the method has the advantages of being few in needed computing resources, high in computing speed and the like.

Description

Technical Field

[0001] The present invention belongs to the technical field of meteorological observation. Specifically, it relates to a method for repairing atmospheric temperature and humidity profiles by integrating statistical methods and deep learning techniques. Background Art

[0002] As a key parameter characterizing the three-dimensional thermal structure of the atmosphere, the integrity of atmospheric temperature and humidity profile data directly restricts the accuracy of numerical weather prediction and the reliability of climate models. When the existing satellite remote sensing detection (such as the FY-4B satellite) retrieves the temperature and humidity profiles through the infrared radiation detection data of the infrared hyperspectral atmospheric vertical detector, cloud contamination will cause large-area continuous loss of the whole-layer atmospheric parameters. That is, the radiation shielding effect of the cloud on the surface and the lower atmosphere makes the detector only able to obtain the atmospheric information above the cloud top, and the data below the cloud area is completely unobservable. Such a situation where the missing area accounts for more than 50% of the single-scene image area frequently occurs in weather processes such as typhoons and frontal systems, and traditional repair methods face severe challenges.

[0003] Currently, the methods for filling in the large-area continuous missing temperature and humidity profiles can be divided into four technical routes. One is the statistical regression method, which estimates the atmospheric temperature and humidity profiles based on known meteorological observation data and statistical relationships. Although it can maintain the continuity of the basic physical field, it cannot reflect the abnormal fluctuations caused by sudden weather processes, resulting in an overly smooth repair result and losing the fine structural features of weather systems. The second is the physical inversion method, which iteratively optimizes by coupling the radiation transfer equation and the atmospheric motion equation. Although it has a clear physical meaning, the calculation is time-consuming and sensitive to the initial value, and parameter divergence is likely to occur when the cloud optical thickness changes violently. The third is the interpolation method. Among them, time-dimensional interpolation is effective for slow-varying processes such as diurnal cycle changes, but in rapidly evolving scenarios such as severe convective weather, lag errors will occur due to the inability to capture the sudden changes in temperature and humidity; spatial-dimensional interpolation depends on neighboring valid data points. When encountering large-area continuous missing, the interpolation result will deviate seriously from the true value due to insufficient valid samples. The fourth is the deep learning method: using a deep neural network for end-to-end repair. Although it has the advantage of non-linear mapping, it has two inherent defects: 1) training data dependence, which requires complete true-value data for supervised training, while in practical applications, there is no effective observation data available for verification in the missing area, and there are significant morphological differences between the traditional random rectangular mask simulation and the spatial distribution characteristics of real cloud contamination; 2) extreme missing failure, when the proportion of the missing area exceeds 70%, the end-to-end network will generate non-physical oscillations due to the lack of effective context constraints, such as abnormal inversion layers and sudden changes in humidity.

[0004] In summary, the prior art has the following core defects: First, the repair accuracy is insufficient. The repair results of the statistical method are overly smoothed and it is difficult to capture the rapid evolution characteristics of the weather system. The deep learning method produces non-physical solutions in extreme missing scenarios. Second, the training paradigm is limited. Traditional supervised learning relies on the assumption of complete ground truth data, which conflicts with the fundamental contradiction of "no ground truth in the missing area" in real application scenarios. Summary of the Invention

[0005] Aiming at the deficiencies of the prior art, the present invention proposes an atmospheric temperature and humidity profile repair method that combines statistical methods and deep learning techniques. This method takes the FY-4B satellite temperature and humidity profile product as the processing object, and aims at the problem of large-area data missing caused by cloud pollution. Data reconstruction is realized through a three-stage architecture of random cloud mask generation, statistical filling pre-repair, and deep learning fine correction.

[0006] The present invention is realized through the following technical solutions: An atmospheric temperature and humidity profile repair method that combines statistical methods and deep learning techniques: The method specifically includes the following steps:

[0007] Step 1, generate a random cloud mask; simulate the cloud pollution area through the cube cloud mask or the historical cloud mask generation method;

[0008] Step 2, statistical filling pre-repair; adopt a multi-stage progressive filling strategy to preliminarily repair the missing area according to historical data;

[0009] Step 3, deep learning fine correction, realize data reconstruction of the atmospheric temperature and humidity profile through an improved 3D U-Net architecture.

[0010] Further, in Step 1,

[0011] Simulate the vertical occlusion characteristics of clouds through the cube cloud mask generation method, and randomly generate a three-dimensional mask with physical rationality: the horizontal dimension is a rectangular area with a side length not exceeding H / 2 and W / 2, and the vertical dimension continuously covers all pressure layers downward from the random starting pressure layer, ensuring that the proportion of effective pixel points covered by the mask is higher than 30%;

[0012] Extract the real cloud pollution area from the data at other times through the historical cloud mask generation method, and also perform masking on the basis of ensuring that the proportion of effective pixel points masked is higher than 30%.

[0013] Further, three types of mask matrices with clear physical meanings are finally generated, namely:

[0014] The historical data mask marks the effective historical data area that can be used for statistical filling;

[0015] The current observation mask identifies the missing area contaminated by simulated clouds at the current moment;

[0016] The true value verification mask is used for true value retention, retaining the true valid value area not covered by the mask;

[0017] Three types of mask matrices and the original data are used to construct training samples through tensor dot product operations.

[0018] Furthermore, in step 2,

[0019] A multi-stage progressive repair strategy is adopted to implement differential filling for the characteristics of temperature and humidity parameters:

[0020] For the temperature parameter, since the temperature changes slowly and the values are similar at the same time in different dates in the same region, a dynamic weighted filling model based on the spatio-temporal correlation of historical data is constructed;

[0021] For the specific humidity parameter, since its change is intense, but the humidity values in the adjacent space at the same time are similar, a mechanism based on the current spatio-temporal data is designed.

[0022] Furthermore, for the temperature parameter, first load the three-dimensional temperature and humidity profile data of the target area at the current moment, identify the coordinate index of the missing area, and implement fourth-order progressive filling for the temperature parameter:

[0023] Stage 1, spatio-temporal weighted filling: Retrieve the historical data sets at the same time in the previous N days, establish a linear regression weight model with a dynamic time window, and calculate the spatio-temporal weighted mean; when it is detected that the historical data of the current grid point is missing, gradually expand the time window until the effective sample number threshold is met;

[0024] Stage 2, time dimension interpolation: For the pressure layer with missing values in the whole layer, search forward / backward on the time axis for the nearest valid moment and perform linear interpolation in the time dimension;

[0025] Stage 3, spatial filling: For isolated missing points, implement spatial filling with an adaptive radius;

[0026] Initially, a grid window is constructed with the missing point as the center. If the effective data in the window is insufficient, gradually expand the grid range and fill the missing point with the average value of the valid values in the window. During this process, data in the same latitude band is preferentially considered to maintain zonal continuity;

[0027] Stage 4, statistical mean filling: For extreme missing scenarios, use the statistical mean of the pressure layers in the whole region at the current moment for final filling.

[0028] Furthermore, for the specific humidity parameter, focus on the data at the same time, and only retain the last three stages of the four stages of temperature processing.

[0029] An atmospheric temperature and humidity profile repair system integrating statistical methods and deep learning technologies:

[0030] The system includes a random cloud mask generation module, a statistical filling pre-repair module, and a deep learning refinement correction module;

[0031] The random cloud mask generation module simulates cloud polluted areas through a cube cloud mask or a historical cloud mask generation method;

[0032] The statistical filling pre-repair module uses a multi-stage progressive filling strategy to preliminarily repair missing areas according to historical data;

[0033] The deep learning refinement correction module realizes data reconstruction of atmospheric temperature and humidity profiles through an improved 3D U-Net architecture.

[0034] Furthermore, in the deep learning refinement correction module,

[0035] The statistically filled data and the original mask are used as inputs, and an improved 3D U-Net architecture is adopted, in which all encoder-decoder structures use three-dimensional convolution operations

[0036] In the encoder part, a 2-level downsampling path is adopted. Each level contains two 3×3×3 convolutional layers and a max pooling layer. After the convolutional layer, batch normalization and LeakyReLU activation functions are connected. At the same time, residual connections are added after each downsampling stage to retain the gradient change characteristics between different pressure layers;

[0037] In the decoder part, transposed convolution is used to perform upsampling. After each level of upsampling, skip connections are made with the feature maps of the corresponding encoder layers to fuse multi-scale information; finally, the repaired temperature and humidity profile data is obtained.

[0038] An electronic device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps of the above method are implemented.

[0039] A computer-readable storage medium is used to store computer instructions, and when the computer instructions are executed by a processor, the steps of the above method are implemented.

[0040] Advantages of the present invention

[0041] Compared with the prior art, the advantages of the present invention are as follows:

[0042] 1. Through the dual-channel generation strategy of cube masks and historical masks, the vertical occlusion characteristics and spatio-temporal distribution laws of cloud pollution are accurately simulated. Compared with traditional random rectangular masks, the generated mask matrix retains a multi-scale cluster structure in the horizontal dimension and restricts the number of consecutive missing layers in the vertical dimension

[0043] 2. Design independent filling paths according to the physical property differences of temperature and humidity parameters, pay attention to the continuous change of temperature profiles over time in the same space, and pay attention to the spatial continuity of humidity profiles at the same time.

[0044] 3. Through the "statistical filling - deep learning correction" architecture, first, the calculation efficiency is improved; second, the situation where the deep learning model fails in the case of extremely missing values is avoided; third, a reliable starting point for deep learning correction is established through statistical filling.

[0045] 4. Compared with the traditional single - stage repair scheme, while solving the problem of missing true data values, this method combines the advantages of statistical methods and deep learning techniques, significantly improves the repair accuracy in extremely missing scenarios while ensuring physical rationality, and also has the advantages of less required computing resources and fast computing speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is a schematic diagram of the atmospheric temperature and humidity profile repair model that combines the statistical method and deep learning technology of the present invention.

[0047] Figure 2 It is the random cloud mask generation module of the present invention.

[0048] Figure 3 It is the deep learning module of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0049] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0050] The experimental methods used in the following embodiments are all conventional methods unless otherwise specified. The materials, reagents, methods, and instruments used, unless otherwise specified, are all conventional materials, reagents, methods, and instruments in the art, and those skilled in the art can obtain them through commercial channels.

[0051] The present invention proposes an atmospheric temperature and humidity profile repair method that combines statistical methods and deep learning techniques, and its overall process is as Figure 1 shown: The method specifically includes the following steps:

[0052] Step 1, generate a random cloud mask; simulate the cloud - contaminated area through the cube cloud mask or historical cloud mask generation method;

[0053] Simulate the vertical occlusion characteristics of clouds through the above-mentioned cube cloud mask generation method, and randomly generate a three-dimensional mask with physical rationality: the horizontal dimension is a rectangular area with a side length not exceeding H / 2 and W / 2, and the vertical dimension continuously covers all pressure layers downward from a randomly selected starting pressure layer, ensuring that the proportion of valid pixel points covered by the mask is higher than 30%;

[0054] Extract the real cloud contamination area from the data at other times through the above-mentioned historical cloud mask generation method, and also perform masking on the basis of ensuring that the proportion of valid pixel points masked is higher than 30%.

[0055] Finally, generate three mask matrices with clear physical meanings, namely:

[0056] The historical data mask label can be used to count the effective historical data area filled;

[0057] The current observation mask identifies the missing area contaminated by the simulated clouds at the current moment;

[0058] The ground truth verification mask is used for ground truth retention, retaining the real valid value area not covered by the mask;

[0059] The three mask matrices and the original data are used to construct training samples through tensor dot multiplication operations.

[0060] Step 2, Statistical filling pre-repair; adopt a multi-stage progressive filling strategy to preliminarily repair the missing area according to historical data;

[0061] Adopt a multi-stage progressive repair strategy to implement differential filling according to the characteristics of temperature and humidity parameters:

[0062] For temperature parameters, since the temperature changes slowly and the values are similar at the same time in different dates in the same area, construct a dynamic weighted filling model based on the spatio-temporal correlation of historical data;

[0063] For specific humidity parameters, since they change violently, but the humidity values in adjacent spaces at the same time are similar, design based on the current spatio-temporal data mechanism.

[0064] For temperature parameters, first load the three-dimensional temperature and humidity profile data of the target area at the current moment, identify the coordinate index of the missing area, and implement fourth-order progressive filling for the temperature parameters:

[0065] Stage 1, Spatio-temporal weighted filling: Retrieve the historical data sets at the same time in the previous N days, establish a linear regression weight model with a dynamic time window, and calculate the spatio-temporal weighted mean; when it is detected that the historical data of the current grid point is missing, gradually expand the time window until the effective sample number threshold is met;

[0066] Phase II, Temporal Dimension Interpolation: For the pressure layers with entire missing layers, search forward / backward on the time axis for the nearest valid time and perform linear interpolation in the temporal dimension;

[0067] Phase III, Spatial Filling: For isolated missing points, perform spatial filling with an adaptive radius;

[0068] Initially, construct a grid window centered on the missing point. If the valid data within the window is insufficient, gradually expand the grid range and fill the missing point with the average value of the valid values in the window. During this process, prioritize data in the same latitude band to maintain zonal continuity;

[0069] Phase IV, Statistical Mean Filling: For extreme missing scenarios, use the statistical mean of the pressure layers in the entire region at the current moment for final filling.

[0070] For the specific humidity parameter, focus on the data at the same moment and only retain the last three stages of the four-stage temperature processing.

[0071] Step 3, Deep Learning Refinement: Implement data reconstruction of the atmospheric temperature and humidity profiles through an improved 3D U-Net architecture.

[0072] An Atmospheric Temperature and Humidity Profile Repair System Integrating Statistical Methods and Deep Learning Technologies

[0073] The system includes a random cloud mask generation module, a statistical filling pre-repair module, and a deep learning refinement module;

[0074] As Figure 1 shown, in the embodiment, the FY-4B satellite temperature and humidity profile product is used as the processing object,

[0075] The input length of the model is the current moment t and the missing atmospheric temperature and humidity profiles at the same moment in the previous 10 - 20 days Output the repaired complete atmospheric temperature and humidity profiles at the current moment t

[0076] First, obtain the training data through the random cloud mask generation module. This module randomly simulates and generates cloud masks according to the cloud contamination characteristics, and simultaneously solves two core problems: 1) The problem of supervised training under the condition that the true values in the missing areas cannot be obtained; 2) The morphological difference problem between traditional random masks and the spatial distribution characteristics of real cloud contamination. The specific architecture of the random mask module is as Figure 2 shown.

[0077] For the original temperature and humidity profile data at time t The random cloud mask generation module adopts the cube cloud mask generation or historical cloud mask generation method to obtain as Figure 2The three masks shown, where the dark part represents the pixels to be retained and the white part represents the masked pixels.

[0078] Among them, 1) the cube cloud mask method simulates the vertical occlusion characteristics of clouds and randomly generates a three-dimensional mask with physical rationality: randomly determine the mask center coordinates (h c , w c ) in the horizontal dimension, generate a rectangular area with a side length not exceeding H / 2 and W / 2, randomly select the starting pressure layer d start ∈ [1, D] in the vertical dimension, continuously cover all pressure layers downward, and perform masking on the basis that the proportion of masked valid pixels is higher than 30%;

[0079] 2) The historical cloud mask generation method extracts the real cloud pollution area from the data at other times, and also performs masking on the basis that the proportion of masked valid pixels is higher than 30%. Finally, three mask matrices with clear physical meanings are generated: the historical data mask marks the valid historical data area that can be used for statistical filling, and the current observation mask identifies the missing area contaminated by the simulated cloud at the current moment truth verification mask for truth retention, that is

[0080]

[0081] Among them represents the valid value indication matrix of X t , 0 represents an invalid value, and 1 represents a valid value, that is

[0082]

[0083] The three types of masks and the original data are used to construct training samples through tensor dot multiplication, as Figure 2 shown.

[0084] After entering the statistical filling module, the module preliminarily repairs the missing area according to the historical data. The statistical filling module adopts a multi-stage progressive repair strategy and implements differential filling according to the characteristics of temperature and humidity parameters:

[0085] For the temperature parameter (t), since the temperature changes slowly and the values are similar at the same location and the same time on different dates, a dynamic weighted filling model based on the spatio-temporal correlation of historical data is constructed;

[0086] For the specific humidity parameter (q), since its change is intense, but the humidity values in the adjacent space at the same time are similar, a mechanism based on the current spatio-temporal data is designed.

[0087] For the temperature parameter, first load the three-dimensional temperature and humidity profile data at the current moment in the target area and identify the coordinate indices of the missing areas.

[0088] Implement a four-stage progressive filling for the temperature parameter:

[0089] 1) Retrieve the historical data sets {X^{t-k}|k∈[1,30]} at the same moment in the previous 10 - 30 days, establish a linear regression weight model with a dynamic time window, and calculate the spatio-temporal weighted mean:

[0090]

[0091] where the weight coefficient w k is obtained using a linear regression model based on the valid historical data. When it is detected that the historical data of the current grid point (d, h,, w) is missing, the time window is gradually expanded (initial window of 10 days, step size of 5 days) until the valid sample number threshold is met.

[0092] 2) For the pressure layer d where there are still missing data for the entire layer, activate the spatio-temporal compensation mechanism and search forward / backward on the time axis for the nearest valid moments t before 、t after , and perform linear interpolation in the time dimension:

[0093]

[0094] where the interpolation coefficient α = (t - t before ) / (t after - t before ), ensuring time continuity.

[0095] 3) For isolated missing points, implement spatial filling with an adaptive radius.

[0096] Initially, construct a 3×3 grid window centered on the missing point. If the valid data within the window is insufficient, gradually expand it to a 10×10 grid range, and fill the missing point with the average value of the valid values in the window. During this process, data in the same latitude band is preferentially considered to maintain zonal continuity.

[0097] 4) For extreme missing scenarios, use the statistical mean of the pressure layer d in the entire region at the current moment for final filling.

[0098] For the specific humidity parameter, focus on the data at the same moment and only retain the last three stages of the four-stage temperature processing.

[0099] Statistically fill the data and jointly input it with the original mask M into the deep learning correction module. This module uses an improved 3D U-Net architecture, and all its encoder-decoder structures use three-dimensional convolutional operations. The specific model architecture is as Figure 3 shown.

[0100] In the encoder part, according to the characteristic of small scale of atmospheric temperature and humidity profile data, only a 2-level downsampling path is adopted. Each level contains two 3×3×3 convolutional layers and one max pooling layer. After the convolutional layer, batch normalization and LeakyReLU activation function are connected. At the same time, residual connections are added after each downsampling stage to retain the gradient change characteristics between different pressure layers.

[0101] In the decoder part, transposed convolution is used to implement upsampling. After each level of upsampling, skip connections are made with the feature maps of the corresponding encoder layer to fuse multi-scale information.

[0102] In the training stage, the loss is calculated only on the masked valid element values to ensure that the numerical distribution learned by the model is consistent with the true numerical distribution. That is, the loss calculation can be expressed as

[0103]

[0104] where represents the output of the deep learning correction module, X t is the true observed data, is the truth verification mask mentioned above, represents the L0 norm of the mask matrix, that is, the total number of missing points, and ⊙ represents the Hadamard product.

[0105] Finally, according to the output of the deep learning module, combined with the current observation mask the original input data X t and the valid value indication matrix t of X the repaired temperature and humidity profile data is obtained

[0106]

[0107] An electronic device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps of the above method are implemented.

[0108] A computer-readable storage medium is used to store computer instructions. When the computer instructions are executed by a processor, the steps of the above method are implemented.

[0109] The memory in the embodiments of the present application may be a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memories. Among them, the non-volatile memory may be a read only memory (ROM), a programmable ROM (PROM), an erasable PROM (EPROM), an electrically erasable PROM (EEPROM), or a flash memory. The volatile memory may be a random access memory (RAM), which is used as an external cache. By way of example but not limitation, many forms of RAM are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchlink DRAM (SLDRAM), and direct rambus RAM (DR RAM). It should be noted that the memory of the method described in the present invention is intended to include but not limited to these and any other suitable types of memory.

[0110] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer instructions are loaded and executed on a computer, the processes or functions described in the embodiments of the present application are generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wire, such as coaxial cable, optical fiber, digital subscriber line (DSL), or wirelessly, such as infrared, wireless, microwave, etc. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or a data center that includes one or more integrated available media. The available medium can be a magnetic medium, such as a floppy disk, a hard disk, or a magnetic tape, an optical medium, such as a high-definition digital video disc (DVD), or a semiconductor medium, such as a solid state disc (SSD), etc.

[0111] In the implementation process, each step of the above method can be completed by the integrated logic circuit of the hardware in the processor or the instructions in the form of software. The steps of the method disclosed in combination with the embodiments of the present application can be directly embodied as being executed by the hardware processor or completed by the combination of the hardware and software modules in the processor. The software module can be located in a mature storage medium in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, or electrically erasable programmable memory, registers, etc. This storage medium is located in the memory, and the processor reads the information in the memory and combines its hardware to complete the steps of the above method. To avoid repetition, it will not be described in detail here.

[0112] It should be noted that the processor in the embodiments of the present application may be an integrated circuit chip with signal processing capabilities. In the implementation process, the steps of the above method embodiments can be completed by the integrated logic circuit in the hardware of the processor or instructions in software form. The above-mentioned processor may be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. It can implement or execute the various methods, steps and logic block diagrams disclosed in the embodiments of the present application. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc. The steps of the method disclosed in combination with the embodiments of the present application can be directly embodied as being executed and completed by a hardware decoding processor, or executed and completed by a combination of hardware and software modules in the decoding processor. The software module may be located in a mature storage medium in the art such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register, etc. This storage medium is located in the memory, and the processor reads the information in the memory and combines its hardware to complete the steps of the above method.

[0113] The above has introduced in detail a method for repairing the atmospheric temperature and humidity profile by integrating statistical methods and deep learning techniques proposed by the present invention, and has elaborated on the principle and implementation manner of the present invention. The description of the above embodiments is only used to help understand the method and its core idea of the present invention; at the same time, for those of ordinary skill in the art, according to the idea of the present invention, there will be changes in the specific implementation manner and application scope. In summary, the content of this specification should not be construed as a limitation to the present invention.

Claims

1. An atmospheric temperature and humidity profile restoration method that combines statistical methods and deep learning techniques, characterized in that: The method specifically includes the following steps: Step 1, generate a random cloud mask; simulate the cloud polluted area through the cube cloud mask or the historical cloud mask generation method; Step 2, statistical filling for pre-restoration; adopt a multi-stage progressive filling strategy to preliminarily restore the missing area according to historical data; Step 3, deep learning fine-tuning correction, realize data reconstruction of the atmospheric temperature and humidity profile through an improved 3D U-Net architecture.

2. The repair method according to claim 1, wherein: In step 1, Simulate the vertical occlusion characteristics of the cloud layer through the cube cloud mask generation method, and randomly generate a three-dimensional mask with physical rationality: the horizontal dimension is a rectangular area with a side length not exceeding H / 2 and W / 2, and the vertical dimension continuously covers all pressure layers downward from the randomly starting pressure layer, ensuring that the proportion of valid pixel points covered by the mask is higher than 30%; Extract the real cloud polluted area from the data at other times through the historical cloud mask generation method, and also perform masking on the basis of ensuring that the proportion of valid pixel points masked is higher than 30%.

3. The restoration method according to claim 2, characterized in that: Finally, three mask matrices with clear physical meanings are generated, namely: The historical data mask marks the valid historical data area that can be used for statistical filling; The current observation mask identifies the missing area polluted by the simulated cloud at the current moment; The true value verification mask is used for true value retention, and retains the real valid value area not covered by the mask; The three mask matrices and the original data construct training samples through tensor dot product operation.

4. The repair method according to claim 3, wherein: In step 2, Adopt a multi-stage progressive restoration strategy, and implement differential filling according to the characteristics of temperature and humidity parameters: For the temperature parameter, since the temperature changes slowly and the values are similar at the same time in different dates in the same region, construct a dynamic weighted filling model based on the spatio-temporal correlation of historical data; For the specific humidity parameter, since its change is intense, but the humidity values in the adjacent space at the same moment are similar, design a mechanism based on the current spatio-temporal data.

5. The restoration method according to claim 4, characterized in that: For the temperature parameter, first load the three-dimensional temperature and humidity profile data of the target area at the current moment, identify the coordinate index of the missing area, and implement four-order progressive filling for the temperature parameter: Stage 1, spatio-temporal weighted filling: retrieve the historical data sets at the same time in the previous N days, establish a linear regression weight model with a dynamic time window, and calculate the spatio-temporal weighted mean; when it is detected that the historical data of the current grid point is missing, gradually expand the time window until the effective sample number threshold is met; Stage 2, time dimension interpolation: for the pressure layer with missing data in the whole layer, search forward / backward on the time axis for the nearest valid moment, and perform linear interpolation in the time dimension; Stage 3, space filling: for isolated missing points, implement space filling with an adaptive radius; Initially construct a grid window centered on the missing point. If the valid data in the window is insufficient, gradually expand the grid range, and fill the missing point with the average value of the valid values in the window. During this process, give priority to considering the data in the same latitude band to maintain zonal continuity; Stage 4, statistical mean filling: For extreme missing scenarios, the statistical mean of the pressure layer in the entire region at the current moment is used for final filling.

6. The repair method according to claim 5, wherein: For the specific humidity parameter, data at the same moment is mainly considered, and only the last three stages of the four-stage temperature processing are retained.

7. A repair system for an atmospheric temperature and humidity profile repair method that combines a statistical method and deep learning technology according to any one of claims 1 to 6, wherein: The system includes a random cloud mask generation module, a statistical filling pre-repair module, and a deep learning refinement correction module; The random cloud mask generation module simulates the cloud pollution area through a cube cloud mask or a historical cloud mask generation method; The statistical filling pre-repair module uses a multi-stage progressive filling strategy to preliminarily repair the missing area according to historical data; The deep learning refinement correction module realizes data reconstruction of the atmospheric temperature and humidity profile through an improved 3D U-Net architecture.

8. The repair method according to claim 6, characterized in that: In the deep learning refinement correction module, The data after statistical filling and the original mask are used as inputs, and an improved 3D U-Net architecture is adopted, and all its encoder-decoder structures use three-dimensional convolution operations. In the encoder part, a 2-level downsampling path is adopted, each level includes two 3×3×3 convolutional layers and a max pooling layer. After the convolutional layer, batch normalization and LeakyReLU activation functions are connected. At the same time, residual connections are added after each downsampling stage to retain the gradient change characteristics between different pressure layers; In the decoder part, transposed convolution is used to perform upsampling. After each level of upsampling, skip connections are made with the feature maps of the corresponding encoder layers to fuse multi-scale information; finally, the repaired temperature and humidity profile data is obtained.

9. An electronic device, comprising a memory and a processor, the memory storing a computer program, characterized in that, The processor implements the steps of the method according to any one of claims 1 to 6 when executing the computer program.

10. A computer-readable storage medium for storing computer instructions, characterized in that, The computer instructions implement the steps of the method according to any one of claims 1 to 6 when executed by the processor.

Images