CN117710508A - Near-surface temperature inversion method and device for generating countermeasure network based on improved condition - Google Patents

Abstract

The invention relates to a near-surface temperature inversion method and a near-surface temperature inversion device for generating an countermeasure network based on improved conditions, wherein a generator for constructing the near-surface temperature inversion device is based on a U-net frame, a multi-scale residual error module is introduced between an encoder and a decoder, and key features of original input data are better extracted; at the same time, attention mechanisms are introduced in the decoder, which is beneficial to guiding the generator to output data which is closer to the real temperature image. The wind and cloud satellite FY-4A temperature image is introduced into the generator and the discriminator as condition information, and the discriminator discriminates the generated image and the real image on multiple scales, so that the discriminator can capture global and local information of the image and evaluate the details of the generated image more accurately. The method and the device provided by the invention can accurately estimate the near-surface temperature, complete model training in a self-supervision mode, automatically output high-resolution near-surface air temperature data, and effectively solve the problem of near-surface air temperature data value loss caused by station point loss.

Description

Near-surface temperature inversion method and device for generating countermeasure network based on improved condition

Technical Field

The invention relates to the field of meteorological data processing, in particular to a near-surface temperature inversion method and device for generating an countermeasure network based on improved conditions.

Background

The air temperature is one of the most basic observation items in meteorological observation data, wherein the air temperature near the earth surface is used as the comprehensive reflection of the radiation exchange and the heat balance of the underlying surface, controls most of land surface processes, characterizes the cold and hot degree of near-surface air, is closely related to animal and plant growth and development and human activities, and is an important research object for climate change. The near-surface air temperature measurement mainly comprises two types, namely a site observation method and a remote sensing inversion method. In the site observation method, near-surface air temperature data are basically collected by a meteorological observation station, however, due to the limitation of factors such as geographical environment and the like, the observation stations are unevenly distributed, and particularly in a sparse area of people, the meteorological observation station is sparse and even cannot observe. The air temperature data obtained by the traditional spatial interpolation method has low resolution and accuracy, and the subsequent data analysis and application are limited. In recent years, satellite remote sensing technology and service products are mature, satellite remote sensing observation continuity is good, space coverage is wide, the influence of ground is small, space resolution is high, the defect of an observation site can be overcome, and a feasible technical approach is provided for obtaining continuous near-surface temperature space-time distribution information. Therefore, the earth surface temperature is estimated by using the inversion of the remote sensing satellite, so that a new near-earth surface temperature estimation idea is formed.

Among the conventional site-dependent observation methods, spatial interpolation is the most commonly used method. Common spatial interpolation methods include kriging interpolation and inverse distance weighted interpolation. The kriging interpolation is a spatial interpolation technique based on a statistical method. It takes into account the spatial correlation between sample points and uses a half-variational function to describe the degree of spatial correlation between sample points. The kriging interpolation assumes a certain variability in the differences between sample points and estimates the values at spatially arbitrary locations by fitting a semi-varying function. In the temperature estimation, the Kriging interpolation method can predict the temperature value of an unobserved point according to the spatial distribution condition of the site observation point. The inverse distance weighted interpolation is a distance-based interpolation method. It is based on the assumption that an observation point closer to a location has a greater influence on the value of that location, while an observation point farther from that location has a smaller influence on the value of that location. According to this assumption, the inverse distance weighted interpolation method estimates the temperature value of an unobserved location by weighted averaging of known temperature observation points, where the weight is inversely proportional to the distance of the observation point from the location to be estimated.

In recent years, satellite remote sensing technology and business products are mature, and the estimation of near-surface temperature by using remote sensing satellites is a trend. The main categories are as follows: (1) The data statistics method mainly comprises the steps of establishing a unitary or multiple linear regression model between ground surface temperature data inverted by remote sensing data and weather site measured data, and further estimating near-surface temperature, and can be divided into a single-factor statistics model and a multi-factor statistics model. The single factor statistical model is based on the high correlation of the air temperature and the ground surface temperature, and the air temperature is obtained by establishing a linear regression relation between the air temperature and the ground surface temperature of remote sensing data. The multi-factor statistical model takes influence factors of a plurality of air temperatures into consideration to establish a linear or nonlinear model to solve the air temperature. (2) The temperature-vegetation index (TVX) method performs near-surface air temperature inversion based on the theory that the dense vegetation temperature and air temperature of the remote sensing inversion on kilometer-level spatial scales are approximately equal to each other in relation to a highly linear correlation between the vegetation index NDVI and the surface temperature. (3) The earth surface energy balance method is used for conducting near-earth surface air temperature remote sensing inversion research based on an earth surface energy balance principle, and establishing a linear relation model between earth surface temperature and near-earth surface air temperature for iterative estimation. (4) The atmospheric vertical profile method utilizes a physical model formula to calculate, and the calculation difficulty is high due to factors such as difficult determination of aerodynamic impedance.

The prior art has the following defects:

1. the geospatial interpolation method, such as inverse distance weight interpolation method and Kerling interpolation method, which rely on traditional site data is used for obtaining continuous near-surface air temperature space-time distribution information through interpolation of existing meteorological site data. But is limited by site uneven distribution, especially in complicated terrain areas with sparse human smoke, interpolation methods are difficult to obtain to meet the requirement of regional scale research, and the accuracy of near-surface temperature data which can be obtained is limited, meanwhile, the interpolation methods only consider the autocorrelation of the near-surface temperature, and the influence of other atmospheric physical factors on the near-surface temperature is ignored, so that the method has certain limitation.

2. Based on the data statistics method, a large amount of data input is needed, the model depends on the time and place acquired by modeling data, the model precision is limited by the number and representativeness of data samples, and the stability and portability of the model are poor.

3. TVX is not suitable for air temperature inversion of areas with low vegetation coverage or non-vegetation coverage, and is also limited by cloud occlusion in the image and mixed coverage of multiple vegetation. Meanwhile, the inversion accuracy is affected by the vegetation saturation value, the neighborhood window size, the variation range of the NDVI in the neighborhood window and the like.

4. The surface energy balance method requires a large amount of parameter input, and some parameters (such as dynamic impedance, surface roughness, wind speed and the like) cannot be directly obtained through remote sensing means, and other measurement means or model estimation are needed to be relied on. This increases the complexity of model building and application while reducing the feasibility of real-time applications.

5. The accuracy of the atmospheric profile extrapolation method air temperature estimation is affected by the seasonal climate change conditions in different regions, e.g. the characteristics of the air temperature vertical profile may vary significantly in different seasons or different regions. In addition, in the presence of cloud coverage, the accuracy of the atmospheric profile extrapolation method may be low, as the cloud may interfere with the observed data, making the profile extrapolation result inaccurate.

Disclosure of Invention

Aiming at the defects of the prior art, the invention provides a near-surface temperature inversion method for generating an countermeasure network based on improved conditions, which is characterized in that the inversion network constructed by the inversion method introduces a generator of an attention mechanism and a multi-scale cascade discriminator, takes a wind cloud satellite image as condition information, guides the generator to generate a temperature image with higher quality, discriminates on different scales and is beneficial to capturing global and local information of the temperature image, and the near-surface temperature inversion method comprises the following steps:

step 1: collecting a finishing data set, and adopting wind-cloud satellite data and temperature data of a range of two meters near the surface of the earth in analysis data ERA 5;

step 2: preprocessing a data set to obtain a training set and a verification set, filling missing values in wind-cloud satellite data and analysis data ERA5 by adopting space-time adjacent interpolation, processing abnormal values, and dividing the preprocessed data set into the training set and the verification set according to a stipulated proportion;

step 3: constructing a near-surface temperature inversion network, wherein the inversion network comprises a generator based on residual connection and an attention mechanism and a discriminator based on multi-scale cascading, and the inversion network is trained by inputting the training set, and specifically comprises the following steps:

step 31: inputting a training set into the generator, wherein the generator is based on a U-net structure and comprises an encoder module, a decoder module and a multi-scale residual error module arranged between the encoder module and the decoder module, and the specific processing comprises the following steps:

step 311: the encoder module comprises a first encoder, a second encoder, a third encoder, a fourth encoder and a fifth encoder, wherein the input near-surface temperature image size is as follows: 401×401×4, performing convolution and pooling operations on the input near-surface temperature image by the encoder, and extracting a high-level abstract feature map to obtain a fifth feature map;

step 312: inputting the fifth characteristic diagram into four multi-scale residual error modules which are sequentially connected, wherein each multi-scale residual error module comprises two convolution layers, each convolution layer comprises a two-dimensional deconvolution layer, a batch normalization layer and a ReLU layer, the design of residual error connection is adopted, the input and the output of the multi-scale residual error module are added to be used as the input of the next multi-scale residual error module, and the residual error characteristic diagram with the size of 25 multiplied by 1024 is finally output;

step 313: inputting the residual characteristic map into a decoder module, wherein the decoder module comprises a first decoder, a second decoder, a third decoder and a fourth decoder, sequentially decoding the input residual characteristic map, recovering the residual characteristic map to a size of 401×401×64, and then obtaining a predicted near-surface temperature image with a size of 401×401×4;

step 32: the predicted near-surface temperature image and the real near-surface temperature image output by the decoder are input into a multi-scale cascade discriminator for discrimination, and the discriminator comprises three sub-discriminators:

step 321: downsampling the predicted near-surface temperature image and the real near-surface temperature image twice, wherein the input data size is 401×401×5, and the downsampled data images with the sizes of 200×200×5 and 100×100×5 are respectively obtained;

step 322: inputting the three paired predicted near-surface temperature images and the real near-surface temperature images with different sizes in the step 321 into corresponding sub-discriminators, discriminating on three scales, and averaging or weighting the three discrimination results to obtain a final discrimination result;

step 33: the discriminator feeds back to the generator according to the discrimination results of the three sub-discriminators, the generator adjusts the parameters of the generator to continue training, and the steps 31 to 33 are repeated until the discriminator discriminates the predicted near-surface temperature data graph as true, and the training is finished and the parameters of the generator are fixed.

According to a preferred embodiment, each decoder module has built-in attention gating units which fuse intermediate feature maps of corresponding sizes in the encoder module as input for the jump connection into the corresponding decoder.

According to a preferred embodiment, the loss function of the near-surface temperature inversion method is the sum of the contrast loss and the pixel loss, balancing the fidelity of the generated image and the structural details of the maintained image by balancing the two loss terms, wherein,

the countermeasures include generator measures and arbiter measures, the generator measures being used to minimize a distribution difference between the predicted near-surface temperature image and the real near-surface temperature image;

the discriminator loss is used for maximizing the probability that the discriminator correctly distinguishes the predicted near-surface temperature image and the real near-surface temperature image to push the generator to generate an image approaching the real near-surface temperature, the discriminator loss specifically comprises three scale sub-discriminator losses, and then the sub-discriminator output of each scale is calculated according to average or weight weighting to obtain a final discrimination result;

the pixel loss is used to predict a pixel level difference between the near-surface temperature image and the real near-surface temperature image.

The near-surface temperature inversion device for generating the countermeasure network based on the improved conditions is characterized by comprising a preprocessing module, a generator module and a discriminator module, and is particularly characterized in that:

the preprocessing module carries out interpolation processing on the missing values in the wind-cloud satellite 4A data and the analysis data ERA5, and simultaneously carries out processing on the missing values;

the generator module includes an encoder module, a multi-scale residual module, and a decoder module, wherein,

the encoder module is used for extracting high-level semantic features from the input near-surface temperature data map through convolution and downsampling operations;

the multi-scale residual error module is arranged behind the encoder module and is used for further transforming the advanced semantic features extracted by the encoder module to better extract key features of the original input data;

the decoder module is used for receiving the output characteristic diagram of the multi-scale residual error module, recovering the characteristic diagram through convolution and up-sampling operation, and further comprises an attention gating unit, wherein the attention gating unit is used for taking the intermediate characteristic diagram with the corresponding size of the encoder as jump connection input, learning the relation between the characteristics more flexibly, and finally outputting a predicted near-surface temperature inversion data diagram;

the discriminator module is used for discriminating the predicted near-surface temperature inversion image and the real near-surface temperature image, and specifically, the discriminator is a multi-scale cascade discriminator and comprises a first sub-discriminator, a second sub-discriminator and a third sub-discriminator, and after the predicted near-surface temperature inversion data diagram is subjected to downsampling twice, the predicted near-surface temperature inversion data diagram is respectively input into the corresponding sub-discriminators for discrimination, and a multi-scale discrimination output discrimination result is obtained.

Compared with the prior art, the invention has the beneficial effects that:

1. according to the invention, the near-surface temperature is accurately estimated by utilizing the image generation advantage of the countermeasure generation network (GAN), the model training is completed in a self-supervision mode, the near-surface air temperature data with high resolution is automatically output, and the problem of near-surface air temperature data value deletion caused by station point deletion is effectively solved.

2. The invention provides a near-surface air temperature estimation method based on an improved condition generation countermeasure network (c-GAN), which utilizes the special advantage that the condition information is introduced into the condition generation countermeasure network (c-GAN) to control the image generation process, and introduces FY4A images as the condition information, on one hand, guides the control temperature image generation to generate samples which are closer to the actual surface temperature, improves the quality and diversity of the generated samples, and on the other hand, provides additional constraint and guidance and improves the stability of countermeasure training.

3. The generator designed by the invention combines the multistage residual error connection and the attention mechanism on the basis of the U-Net framework respectively at the encoder and decoder parts, can pertinently extract the characteristics with important guiding function on the temperature data generation task, enhances the function of effective characteristics and suppresses redundant information, simultaneously reduces the training difficulty of the network by utilizing the residual error network structure, and effectively avoids the gradient vanishing phenomenon.

4. The arbiter part adopted by the invention is cascaded with three sub-arbiter networks with different scales, and each sub-arbiter judges the generated image and the real image on different scales by setting different scale factors, so that the arbiter can capture the global and local information of the image and evaluate the details of the generated image more accurately.

5. According to the invention, the advantages of wide swath, full coverage, high timeliness and the like of the wind-cloud meteorological satellites in China are utilized, meanwhile, two channel data which can fully reflect near-surface temperature (cloud and surface air temperature) are introduced into the wind-cloud meteorological satellites in China, longitude and latitude information is added as auxiliary information, ERA5 grid data is used as constraint, so that data redundancy information is effectively increased, model overfitting is reduced, and near-surface temperature estimation accuracy is improved.

Drawings

FIG. 1 is a schematic diagram of the network of the near-surface temperature inversion method of the present invention;

FIG. 2 is a schematic diagram of the generator architecture of the present invention based on residual and attention gating cells;

FIG. 3 is a schematic diagram of a multi-scale residual module according to the present invention;

FIG. 4 is a schematic diagram of a multi-scale cascade arbiter according to the present invention;

FIG. 5 is a schematic diagram of the sub-discriminant of the present invention;

FIG. 6 is a schematic diagram of the structure of the near-surface temperature inversion apparatus of the present invention;

FIG. 7 is an experimental result of the temperature inversion method of the present invention.

Detailed Description

The objects, technical solutions and advantages of the present invention will become more apparent by the following detailed description of the present invention with reference to the accompanying drawings. It should be understood that the description is only illustrative and is not intended to limit the scope of the invention. In addition, in the following description, descriptions of well-known structures and techniques are omitted so as not to unnecessarily obscure the present invention.

The following detailed description refers to the accompanying drawings.

The invention mainly utilizes the remote sensing image of the wind-cloud meteorological satellite in China, and adopts an improved condition generation countermeasure network (c-GAN, conditional Generative Adversarial Network) method to solve the problem of estimating the near-earth surface temperature. The commonly used near-surface temperature is mostly collected by a meteorological observation station, however, the meteorological observation station is unevenly distributed and is very sparse in a part of special areas, so that observation is limited. The cloud satellite is a 'national weight' in China, the remote sensing image has wide space coverage, the acquired data is real-time, the high-resolution data can be provided, and meanwhile, a static orbit radiation imager (AGRI) carried by the cloud satellite 4A can acquire static multispectral image information and provide the earth surface temperature distribution and dynamic information.

The condition generation countermeasure network (c-GAN) generator and the arbiter add additional information as conditions, and the condition information is introduced as a part of the input layer to control the image generation process in a guiding way by conveying the additional information to the discrimination model and the generation model, so that the image generation can be performed in a specified direction instead of random, thereby realizing the expansion of the original GAN, and being widely applied in the field of image generation. Therefore, the invention utilizes the characteristics of a condition generation countermeasure network, takes a wind cloud satellite data real sample and an analysis data (ERA 5) real tag as information to be input into a generator and a discriminator of the c-GAN, and utilizes the mutual game between the generator and the discriminator to search the relationship between the near-surface temperature and the satellite remote sensing thermal infrared detection air temperature so as to realize accurate estimation of the near-surface temperature.

Aiming at the defects existing in the prior art, the invention provides a near-surface air temperature estimation method based on an improved condition GAN. FIG. 1A is a schematic diagram of a network structure of the near-surface temperature inversion method of the invention, as shown in FIG. 1, the method utilizes the characteristics of a condition generation countermeasure network, and FY4A images are introduced as condition information to guide and control the generation of the near-surface temperature, so that the training stability of the countermeasure network is improved. Meanwhile, a real sample of wind cloud satellite data and an analysis data (ERA 5) real label are used as information to be input into a generator and a discriminator of the c-GAN, and the relationship between the near-surface temperature and the satellite remote sensing thermal infrared detection air temperature is searched by utilizing the mutual game between the generator and the discriminator, so that accurate estimation of the near-surface temperature is realized. The method comprises 4 steps of data set construction and preprocessing, generating a predicted near-surface temperature image through a generator based on residual connection and an attention mechanism, discriminating a predicted image and a real image through a multi-scale cascade discriminator, and optimizing through a loss function.

Step 1: the collated data set is collected using the cloud satellite data and temperature data in the range of two meters near the surface of the analysis data ERA5.

The invention collects and collates temperature data in the near-surface two-meter range in 2020, 2021, 2022 cloud satellite 4A (FY-4A) data and analysis data (ERA 5). The wind cloud satellite FY-4A data uses 12 and 13 channels (cloud and earth surface temperature) data acquired by a carried stationary orbit radiation imager (AGRI), the spatial resolution is 4km, the time resolution is 15 minutes, the analysis data ERA5 spatial resolution is 0.25 degrees, and the time resolution is 1 hour.

Step 2: preprocessing a data set to obtain a training set and a verification set, filling missing values in wind cloud satellite data and analysis data ERA5 by adopting space-time adjacent interpolation, processing abnormal values, and dividing the preprocessed data set into the training set and the verification set according to a stipulated proportion.

The time resolution after pretreatment was 1 hour and the spatial resolution was 0.25 degrees. Finally, selecting data in a longitude and latitude range of 50 DEG S-50 DEG N and 40 DEG E-140 DEG E, wherein the image size is 401 pixels by 401 pixels. Data set partitioning concrete bit: the ratio of the training set to the testing set is approximately set to be 5:1, and 16979 FY-4A cloud satellite data and analysis data ERA5 in 1-12 months of 2020 and 2021 are selected as the training set. In order to verify the robustness of the network structure, 2022 years 1-12 months are selected, 850 pieces are randomly extracted in each quarter, and 3400 pieces of FY4A and ERA5 are taken as a test set.

Step 3: the method comprises the steps of constructing a near-surface temperature inversion network, wherein the inversion network comprises a generator based on residual connection and attention mechanism and a discriminator based on multi-scale cascading, inputting the training set to train the inversion network, and fig. 2 is a schematic diagram of the structure of the generator based on residual connection and attention mechanism. The training process specifically comprises the following steps:

step 31: inputting a training set into the generator, wherein the generator is based on a U-net structure and comprises an encoder module, a decoder module and a multi-scale residual error module arranged between the encoder module and the decoder module:

step 311: the encoder module comprises a first encoder, a second encoder, a third encoder, a fourth encoder and a fifth encoder, and is used for carrying out rolling and pooling operations on input meteorological data to extract a high-level abstract feature map:

specifically, the system comprises FY4A data of 12 and 13 channels and longitude and latitude information of the two channels, wherein the total of four channels is 401 x 4. The multi-channel input enables the model to simultaneously consider a plurality of meteorological factors, improves the acquisition capacity of comprehensive information, and provides more meaningful input for subsequent tasks.

The training set inputs the first encoder module and then outputs a first characteristic diagram with the size of 401 multiplied by 64;

performing initial downsampling on the first characteristic diagram to obtain a 200×200×64 characteristic diagram, inputting the characteristic diagram into a second encoder, and outputting a 200×200×128 second characteristic diagram;

performing second downsampling on the second feature map to obtain a feature map of 100×100×128, inputting the feature map into a third encoder, and outputting a third feature map of 100×100×256;

performing third downsampling on the third feature map to obtain a 50×50×256 feature map, inputting the feature map into a fourth encoder, and outputting a 50×50×512 fourth feature map;

downsampling the fourth feature map for the fourth time to obtain a feature map of 25×25×512, inputting the feature map to a fifth encoder, and outputting a fifth feature map of 25×25×1024;

after four downsampling operations, the output of the encoder is a 25×25×1024 feature map, which contains 1024 channels;

after each downsampling step, jumping connection is carried out, and the output characteristic diagram of each encoder module is input to a decoder with a corresponding scale for characteristic recovery. The jump connection is beneficial to maintaining the consistency of resolution and information, improving the learning and restoring capability of the network to details, effectively slowing down the information loss and being beneficial to better keeping the context information.

Step 312: and inputting the fifth characteristic diagram into four multi-scale residual error modules which are connected in sequence. Fig. 3 is a schematic structural diagram of a multi-scale residual module according to the present invention, as shown in fig. 3, where the multi-scale residual module includes two convolution layers, each convolution layer includes a two-dimensional deconvolution layer, a batch normalization layer and a ReLU layer, and a residual connection design is adopted, and an input and an output of the multi-scale residual module are added to be used as an input of a next multi-scale residual module, and finally a residual feature map with a size of 25×25×1024 is output.

Up-sampling the feature map by deconvolution operation, aiming at partially restoring the spatial resolution of the feature map; batch normalization operation, normalization of data distribution, acceleration of convergence and improvement of model stability; the activation function (typically using ReLU) introduces a nonlinear transformation, then performs a two-dimensional convolution operation, the number of channels remains unchanged, further extracts features, and finally adds the input to the convolved output (Identity Mapping).

The introduction of four multi-scale residual modules increases the depth of a network, improves the representation capability of a model, transforms advanced features through a convolution layer, helps the model to learn and restore key features in input data better, and residual connection also reserves original input information through jump connection, so that the gradient vanishing problem is effectively relieved, the training of a deeper network is facilitated, the adaptability of the network is improved, the model is facilitated to adapt to the change of different meteorological data better, and the robustness of the model is improved.

Step 313: inputting the residual characteristic map into a decoder module, wherein the decoder module comprises a first decoder, a second decoder, a third decoder and a fourth decoder, sequentially decoding the input residual characteristic, and recovering the residual characteristic map to a size of 401×401×64 to finally obtain a predicted near-surface temperature image with a size of 401×401×4;

each decoder module is internally provided with an attention gating unit, and the attention gating unit takes an intermediate characteristic diagram with a corresponding size in the encoder module as jump connection input and fuses the intermediate characteristic diagram into a corresponding decoder; the model is allowed to dynamically adjust the weight of each pixel point, and the relation between the features is more flexibly learned.

The jump connection is beneficial to maintaining the consistency of resolution and information, improving the learning and restoring capability of the network to details, effectively slowing down the information loss and being beneficial to better keeping the context information. Meanwhile, the attention gating unit enables jump connection to be more adaptive, information extracted in the encoding process can be better utilized in the decoding process, and the performance of the whole model is improved.

Step 32: the predicted near-surface temperature image and the real near-surface temperature image output by the decoder module are input into a multi-scale cascade discriminator for discrimination, the discriminator comprises three sub-discriminators, and fig. 4 is a schematic structural diagram of the multi-scale cascade discriminator, and the multi-scale cascade discriminator specifically comprises:

step 321: downsampling the predicted near-surface temperature image and the real near-surface temperature image twice, wherein the input data size is 401×401×5, and the downsampled data images with the sizes of 200×200×5 and 100×100×5 are respectively obtained; the number of channels is 5, comprising an input image of 4 channels and an output image of 1 channel.

Step 322: the three pairs of predicted near-surface temperature images and real near-surface temperature images with different sizes in the step 321 are input into corresponding sub-discriminators, specifically, a feature map with the size of 401×401×5 is input into a first sub-discriminator, a feature map with the size of 200×200×5 is input into a second sub-discriminator, a feature map with the size of 100×100×5 is input into a third sub-discriminator, finally, output results with the sizes of 53×53×1, 28×28×1 and 16×16×1 are respectively obtained, and the three output results are averaged or final discrimination results are obtained according to weight adjustment.

As shown in fig. 4, this is a multi-scale arbiter similar to a pyramid structure, and 3 cascaded sub-discriminators are used to allow the sub-discriminators to discriminate the input image on different scales, so as to extract hierarchical features, enhance the perception of the detail and global structure by the network, facilitate the more comprehensive understanding of the input image, and improve the perception capability of the network.

Step 33: the discriminators feed back to the generator according to the output results of the three sub-discriminators, the generator adjusts model parameters to continue training, and the steps 31 to 33 are repeated until the discriminators discriminate that the predicted near-surface temperature data graph is true, and the training is finished, so that the parameters of the generator are fixed.

FIG. 5 is a schematic diagram of a sub-arbiter of the present invention, as shown in FIG. 5, the sub-arbiter employs a Patch GAN structure, the sub-arbiter includes five convolution layers, the activation parameters of the first four convolution layers are ReLU functions, and the fifth convolution layer employs Sigmoid functions.

The ReLU function is adopted to help introduce nonlinear transformation, the expression capacity of a model is improved, the activation function of the fifth-layer convolution is changed into Sigmoid, and the probability interval of mapping the output to (0, 1) is convenient to judge whether the input image is real or generated. The sub-discriminant structure is different from a common discriminant, and can classify each small area in an image instead of singly classifying the whole image, and the design structure of each layer highlights the design advantages of the sub-discriminant structure in the aspects of multi-level feature extraction and sensitivity to local information, so that the model is beneficial to capturing the local information of the image better.

The loss function of the near-surface temperature inversion method includes summing the anti-loss and pixel loss, balancing the fidelity of the generated image and preserving structural detail by balancing the different loss terms. Wherein,

the countermeasures include a generator penalty and a arbiter penalty, the generator penalty for minimizing a distribution difference between the generated image and the real image.

The discriminator loss is used for maximizing the probability that the discriminator correctly distinguishes the generated image and the real image to push the generator to generate the image approaching to the real near-surface temperature, the discriminator loss comprises the sub-discriminator loss of three scales, and then the sub-discriminator output of each scale is calculated according to average or weight weighting to obtain the final discrimination result. The sub-discriminant output of each scale is respectively assigned with weights through learning or manual setting, and a discrimination result comprehensively considering different scale information is finally obtained through the multi-scale discriminant, so that the performance and the generation effect of the network are improved.

The pixel loss is used to predict pixel level differences between the near-surface temperature data and the true near-surface temperature data map, the pixel loss ensuring that image retention structures and details are generated. The present invention uses the L1 penalty to calculate the absolute difference between the generated image and the real image.

The invention provides a near-surface temperature inversion device for generating an countermeasure network based on improved conditions, and fig. 6 is a schematic structural diagram of the inversion device. As shown in fig. 6, the near-surface temperature inversion device includes a preprocessing module, a generator module and a discriminator module, specifically:

the preprocessing module carries out interpolation processing on missing values in the wind cloud satellite 4A data and the analysis data ERA5, and simultaneously carries out processing on the missing values.

The generator module includes an encoder module, a multi-scale residual module, and a decoder module, wherein,

the encoder module is configured to extract advanced semantic features from an input near-surface temperature data map through convolution and downsampling operations.

The multi-scale residual error module is arranged behind the encoder module and is used for further transforming the advanced semantic features extracted by the encoder module to better extract key features of the original input data.

The decoder module is used for receiving the output characteristic diagram of the multi-scale residual error module, recovering the characteristic diagram through convolution and up-sampling operation, the decoder also comprises an attention gating unit, and the attention gating unit is used for taking the intermediate characteristic diagram with the corresponding size of the encoder as jump connection input, learning the relation between the characteristics more flexibly, and finally outputting a predicted near-surface temperature inversion data diagram.

The discriminator module is used for discriminating the predicted near-surface temperature inversion image and the real near-surface temperature image, and specifically, the discriminator is a multi-scale cascade discriminator and comprises a first sub-discriminator, a second sub-discriminator and a third sub-discriminator, and after the predicted near-surface temperature inversion data diagram is subjected to downsampling twice, the predicted near-surface temperature inversion data diagram is respectively input into the corresponding sub-discriminators for discrimination, and a multi-scale discrimination output discrimination result is obtained.

FIG. 7 is an experimental result of the temperature inversion method of the present invention. The most advanced analytical product ERA5 (near surface air temperature) was used as the true value. FIG. 7 (a) is a near-surface temperature image of the 12 th pass of a cloud satellite; FIG. 7 (b) is a near-surface temperature image of the 13 th pass of the cloud satellite; FIG. 7 (c) is an inverted near-surface temperature image of the present invention; fig. 7 (d) is a near-surface temperature image of ERA5. From fig. 7, it can be seen that the inversion effect of the present invention is almost close to the true value ERA5.

The invention compares the proposed method with the prior art method, and the adopted quantitative evaluation indexes comprise: root mean square error (root mean square error, RMSE), pearson correlation coefficient (pearson correlation coefficient, CC), peak signal to noise ratio (PSNR), and Structural Similarity (SSIM), the experimental results are shown in table 1. The project proposal algorithm is quantitatively compared with the prior deep learning method U-Net and the prior re-analysis product CFSv2 under the conditions of the same time space range, resolution and the like by taking the international most advanced re-analysis product ERA5 (near surface air temperature) as a true value. It can be observed from table 1 that the project proposal method achieves the best performance over the existing method. The above experimental data demonstrate the effectiveness and superiority of the project presentation method.

Table 1 quantitative comparison of Experimental results

It should be noted that the above-described embodiments are exemplary, and that a person skilled in the art, in light of the present disclosure, may devise various solutions that fall within the scope of the present disclosure and fall within the scope of the present disclosure. It should be understood by those skilled in the art that the present description and drawings are illustrative and not limiting to the claims. The scope of the invention is defined by the claims and their equivalents.

Claims (4)

1. The inversion method is characterized in that an inversion network constructed by the inversion method introduces a generator of an attention mechanism and a multi-scale cascade discriminator, takes a wind cloud satellite image as condition information, guides the generator to generate a temperature image with higher quality, discriminates on different scales, and is beneficial to capturing global and local information of the temperature image, and the inversion method comprises the following steps:

step 1: collecting a finishing data set, and adopting wind-cloud satellite data and temperature data of a range of two meters near the surface of the earth in analysis data ERA 5;

step 2: preprocessing a data set to obtain a training set and a verification set, filling missing values in wind-cloud satellite data and analysis data ERA5 by adopting space-time adjacent interpolation, processing abnormal values, and dividing the preprocessed data set into the training set and the verification set according to a stipulated proportion;

step 3: constructing a near-surface temperature inversion network, wherein the inversion network comprises a generator based on residual connection and an attention mechanism and a discriminator based on multi-scale cascading, and the inversion network is trained by inputting the training set, and specifically comprises the following steps:

step 31: inputting a training set into the generator, wherein the generator is based on a U-net structure and comprises an encoder module, a decoder module and a multi-scale residual error module arranged between the encoder module and the decoder module, and the specific processing comprises the following steps:

step 311: the encoder module comprises a first encoder, a second encoder, a third encoder, a fourth encoder and a fifth encoder, wherein the input near-surface temperature image size is as follows: 401×401×4, performing convolution and pooling operations on the input near-surface temperature image by the encoder, and extracting a high-level abstract feature map to obtain a fifth feature map;

step 312: inputting the fifth characteristic diagram into four multi-scale residual error modules which are sequentially connected, wherein each multi-scale residual error module comprises two convolution layers, each convolution layer comprises a two-dimensional deconvolution layer, a batch normalization layer and a ReLU layer, the design of residual error connection is adopted, the input and the output of the multi-scale residual error module are added to be used as the input of the next multi-scale residual error module, and the residual error characteristic diagram with the size of 25 multiplied by 1024 is finally output;

step 313: inputting the residual characteristic map into a decoder module, wherein the decoder module comprises a first decoder, a second decoder, a third decoder and a fourth decoder, sequentially decoding the input residual characteristic map, recovering the residual characteristic map to a size of 401×401×64, and then obtaining a predicted near-surface temperature image with a size of 401×401×4;

step 32: the predicted near-surface temperature image and the real near-surface temperature image output by the decoder are input into a multi-scale cascade discriminator for discrimination, and the discriminator comprises three sub-discriminators:

step 321: downsampling the predicted near-surface temperature image and the real near-surface temperature image twice, wherein the input data size is 401×401×5, and the downsampled data images with the sizes of 200×200×5 and 100×100×5 are respectively obtained;

step 322: inputting the three paired predicted near-surface temperature images and the real near-surface temperature images with different sizes in the step 321 into corresponding sub-discriminators, discriminating on three scales, and averaging or weighting the three discrimination results to obtain a final discrimination result;

step 33: the discriminator feeds back to the generator according to the discrimination results of the three sub-discriminators, the generator adjusts the parameters of the generator to continue training, and the steps 31 to 33 are repeated until the discriminator discriminates the predicted near-surface temperature data graph as true, and the training is finished and the parameters of the generator are fixed.

2. The near-surface temperature inversion method of claim 1 wherein each decoder module incorporates an attention gating unit that fuses intermediate feature maps of corresponding sizes in the encoder modules as input to the jump connection into the corresponding decoder.

3. The near-surface temperature inversion method of claim 2 wherein the loss function of the near-surface temperature inversion method is the sum of the contrast loss and the pixel loss, balancing the fidelity of the generated image and maintaining the structural details of the image by balancing the two loss terms, wherein,

the countermeasures include generator measures and arbiter measures, the generator measures being used to minimize a distribution difference between the predicted near-surface temperature image and the real near-surface temperature image;

the discriminator loss is used for maximizing the probability that the discriminator correctly distinguishes the predicted near-surface temperature image and the real near-surface temperature image to push the generator to generate an image approaching the real near-surface temperature, the discriminator loss specifically comprises three scale sub-discriminator losses, and then the sub-discriminator output of each scale is calculated according to average or weight weighting to obtain a final discrimination result;

the pixel loss is used to predict a pixel level difference between the near-surface temperature image and the real near-surface temperature image.

4. The near-surface temperature inversion device for generating the countermeasure network based on the improved conditions is characterized by comprising a preprocessing module, a generator module and a discriminator module, and is particularly characterized in that:

the preprocessing module carries out interpolation processing on the missing values in the wind-cloud satellite 4A data and the analysis data ERA5, and simultaneously carries out processing on the missing values;

the generator module includes an encoder module, a multi-scale residual module, and a decoder module, wherein,

the encoder module is used for extracting high-level semantic features from the input near-surface temperature data map through convolution and downsampling operations;

the multi-scale residual error module is arranged behind the encoder module and is used for further transforming the advanced semantic features extracted by the encoder module to better extract key features of the original input data;

the decoder module is used for receiving the output characteristic diagram of the multi-scale residual error module, recovering the characteristic diagram through convolution and up-sampling operation, and further comprises an attention gating unit, wherein the attention gating unit is used for taking the intermediate characteristic diagram with the corresponding size of the encoder as jump connection input, learning the relation between the characteristics more flexibly, and finally outputting a predicted near-surface temperature inversion data diagram;

the discriminator module is used for discriminating the predicted near-surface temperature inversion image and the real near-surface temperature image, and specifically, the discriminator is a multi-scale cascade discriminator and comprises a first sub-discriminator, a second sub-discriminator and a third sub-discriminator, and after the predicted near-surface temperature inversion data diagram is subjected to downsampling twice, the predicted near-surface temperature inversion data diagram is respectively input into the corresponding sub-discriminators for discrimination, and a multi-scale discrimination output discrimination result is obtained.