Agricultural meteorological early warning method based on regional microclimate data interpolation and correction

Abstract

The application relates to the field of agricultural meteorological early warning methods, in particular to an agricultural meteorological early warning method based on regional microclimate data interpolation and correction. A physical constraint loss function and an uncertainty quantification mechanism are introduced to construct a generative adversarial network model to generate a high-resolution, physically reliable static meteorological background field, so that the generated meteorological background field is not only rich in spatial details, but also strictly follows basic physical laws such as temperature vertical decrease rate, and can self-evaluate the uncertainty of each grid point as prior information of a Bayesian hierarchical dynamic correction model. A spatial covariance function fusing geographical distance and topographic distance is constructed in the Bayesian hierarchical model, so that a high spatial resolution and high timeliness meteorological analysis field can be obtained in a complex terrain area with sparse meteorological stations, thereby significantly improving the early warning accuracy and timeliness for local sudden agricultural meteorological disasters such as frost and dry hot wind.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of agricultural meteorological warning methods, in particular to an agricultural meteorological warning method based on regional microclimate data interpolation and correction. BACKGROUND

[0002] In areas such as Xinjiang where weather stations are sparse and the terrain is relatively complex, the regional microclimate characteristics shaped by the local underlying surface and the dynamic thermal effects of the terrain are significant. At present, the realization of accurate agricultural meteorological warning mainly relies on two technical paths: one is to generate high-resolution meteorological grid data based on spatial interpolation methods, such as the collaborative kriging method; the other is to use numerical prediction products or climate model outputs for warning, such as bias correction of regional climate model products.

[0003] However, the first technical path can generate detailed background fields, but it is essentially static and difficult to effectively integrate real-time observation data within the warning time limit, resulting in a lag in the warning of sudden microclimates such as frost and dry hot winds. The second technical path is dynamic, but its initial background field has a coarse spatial resolution, leading to distortion in the correction of key microclimate characteristics such as mountain-valley inversion and slope temperature difference. In such a biased coarse field, even if real-time observation data is assimilated, the analysis results are likely to be distorted, making it difficult to accurately capture key microclimate characteristics that are critical to agriculture.

[0004] Therefore, the existing technology cannot meet the real-time prediction of microclimate data in complex terrain areas, resulting in the problem of lagging agricultural warning. SUMMARY

[0005] In view of the deficiencies of the prior art, the present application provides an agricultural meteorological warning method based on regional microclimate data interpolation and correction, which can overcome the technical problems of the prior art that cannot meet the real-time prediction of microclimate data in complex terrain areas, resulting in lagging agricultural warning.

[0006] To achieve the above purpose, the present application provides the following technical solution: an agricultural meteorological warning method based on regional microclimate data interpolation and correction, which specifically comprises the following steps:

[0007] S1, obtaining historical meteorological observation data and digital elevation model data of a target region to construct a training data set;

[0008] S2, constructing a generative adversarial network model that integrates terrain physical constraints and training the model with the training data set, inputting the digital elevation model data of the target region and the meteorological observation data of the sparse stations into the trained generative adversarial network model to generate a meteorological element background field of the target region and its corresponding grid uncertainty estimation field; wherein the meteorological elements at least include air temperature;

[0009] S3, constructing a Bayesian hierarchical dynamic correction model comprising an observation model layer, a process model layer and a prior model layer, the input of the observation model layer being real-time meteorological observation data of a target area, the process model layer being used to represent the correlation structure of meteorological elements in space, and the prior model layer being constructed by a meteorological element background field and a corresponding grid uncertainty estimation field thereof;

[0010] S4, running the Bayesian hierarchical dynamic correction model, fusing the data of the observation model layer and the prior model layer under the constraint of the process model layer through Bayesian inference, and calculating to obtain a meteorological element analysis field and a posteriori uncertainty thereof;

[0011] S5, based on the meteorological element analysis field, combining a preset disaster judgment threshold related to a key growth period of a crop, generating and outputting agricultural meteorological early warning information facing specific farmland units.

[0012] As preferred, the generative adversarial network model comprises a generator G taking a random noise vector z and deterministic conditional information as input, and a discriminator D;

[0013] The generator G adopts a U-Net network with skip connection, and the encoder of the U-Net network comprises a first feature extraction branch and a second feature extraction branch in parallel;

[0014] The first feature extraction branch is used to down-sample and extract high-level features of a terrain feature map , which is derived based on digital elevation model data and at least includes normalized data of three channels of elevation, slope and aspect; The second feature extraction branch is used to down-sample and extract high-level features of a coarse weather field

[0015] , which is obtained by inverse distance weighted interpolation based on the historical meteorological observation data;

[0016] The high-level features extracted by the first feature extraction branch and the second feature extraction branch are taken as the deterministic conditional information;

[0017] The encoder is used to fuse the deterministic conditional information with the random noise vector z to obtain fused features;

[0018] The decoder of the U-Net network is used to up-sample the fused features, fuse low-level detailed features in the encoder through skip connection, and output a meteorological element background field consistent with the spatial resolution of the terrain feature map;

[0019] ​The discriminator D adopts a PatchGAN structure, and in the training, a real weather field sample is superimposed with a corresponding terrain feature map in the channel dimension to obtain a first sample, and a weather element background field output by the generator is superimposed with a corresponding terrain feature map in the channel dimension to obtain a second sample, and the first sample or the second sample is input into the discriminator D to obtain a two-dimensional matrix, and the value of each position in the two-dimensional matrix represents the probability that the local receptive field area corresponding to the input image conforms to the distribution rule of the real weather field.

[0020] In the model training stage, the random noise vector z is sampled from a preset prior distribution , and in the model application stage, for generating a weather element background field, the random noise vector z is a preset fixed value.

[0021] As a preferred, the specific steps of generating the grid point uncertainty estimation field are:

[0022] After the training of the generative adversarial network model is completed, the parameters of the generator G and the discriminator D are fixed, and the fixing of the random noise vector by the generator G in the application stage is temporarily released;

[0023] Based on the same set of deterministic condition information, different random noise vectors z are injected into the generator G, and independent forward propagation is performed through different random noise vectors z to obtain a plurality of generated weather element background fields;

[0024] Each generated weather element background field is superimposed with a terrain feature map in the channel, and is input into the discriminator D to obtain a corresponding two-dimensional matrix, and is marked as a discrimination confidence atlas;

[0025] The sample variance of the weather element value of each generated weather element background field at each spatial grid point is calculated to obtain a generated variance field ;

[0026] The arithmetic mean of each discrimination confidence atlas is calculated to obtain an average discrimination confidence field;

[0027] The generated variance field and the average discrimination confidence field are fused to obtain a grid point uncertainty estimation field.

[0028] As a preferred, the encoder of the U-Net network further includes a parallel third feature extraction branch, the third feature extraction branch is used for downsampling and extracting high-level features of preprocessed satellite remote sensing inversion data which are spatiotemporally aligned with the terrain feature map, the satellite remote sensing inversion data at least include land surface temperature and normalized vegetation index data, and are marked as deterministic condition information.

[0029] As a preferred, the specific steps of fusing the features are:

[0030] The several deterministic condition information are spliced in the channel dimension to obtain preliminary fusion features;

[0031] The preliminary fusion features are globally averaged pooled along the spatial dimension to obtain a channel description vector A;

[0032] The nonlinear relationship between the channels is learned through several layers of fully connected networks with a bottleneck structure to obtain an attention weight vector w;

[0033] The attention weight vector w is multiplied with the preliminary fusion features channel by channel to obtain the final fusion features after recalibration.

[0034] As a preferred, the total loss function of the generative adversarial network model is a weighted sum of an adversarial loss, a reconstruction loss and a physical constraint loss; the adversarial loss is a least square loss, the reconstruction loss is a mean square error loss, and the physical constraint loss is used to limit the change rule of the temperature by taking the vertical temperature decreasing rate as a soft constraint.

[0035] The physical constraint loss is used to limit the change rule of the temperature by taking the vertical temperature decreasing rate as a soft constraint.

[0036] As a preferred, in step S3, the process model layer is used to construct the covariance matrix of the layer by a spatial covariance function to define the correlation structure of the meteorological elements in space, and the spatial covariance function is constructed to be monotonically decaying with the increase of the geographical distance and the terrain distance between any two spatial positions in the target region, and the terrain distance is the elevation difference between the two spatial positions.

[0037] As a preferred, in step S3, the specific steps for constructing the prior model layer by the meteorological element background field and the corresponding grid point uncertainty estimation field are as follows:

[0038] The prior model layer defines the prior distribution of the meteorological element true value field X as a multivariate Gaussian distribution;

[0039] The prior covariance matrix of the prior model layer is the sum of a diagonal matrix and the covariance matrix of the process model layer, and the diagonal elements in the diagonal matrix are obtained from the corresponding grid points in the grid point uncertainty estimation field and mapped by a pre-set monotonically increasing function.

[0040] As a preferred, in step S3, the observation model layer of the Bayesian hierarchical dynamic correction model is defined as: ; in the above formula, represents a real-time meteorological observation data vector obtained from a plurality of devices, a meteorological element true value field, an observation operator matrix for mapping the meteorological element true value field to observation point locations, is an observation error covariance matrix;

[0041] the real-time meteorological observation data vector at least contains observation data from fixed meteorological standard stations within the target area and observation data from Internet of Things mobile observation devices deployed in the farmland area;

[0042] The observation error covariance matrix is a block-diagonal matrix for characterizing the uncertainty differences of different types of observation data, and its expression is: In the above formula, and respectively represent the observation error covariance sub-matrices of the observation data of the fixed meteorological standard stations and the Internet of Things mobile observation devices, and are diagonal matrices, and the diagonal elements of the two and are respectively set according to the calibration accuracy parameters or historical observation error statistical information of the corresponding type of observation device, and satisfy , that is, the observation error variance of the fixed meteorological standard station is smaller than the observation error variance of the Internet of Things mobile observation device.

[0043] As preferred, in step S4, the specific step of the Bayesian inference is to fuse the prior model layer and the observation model layer to obtain a meteorological element analysis field and its posterior uncertainty, and the calculation formula is: ; ;; ; In the above formula, and respectively are the mean vector and the covariance matrix of the posterior distribution of the meteorological element true value field X, that is, the meteorological element analysis field and the posterior uncertainty, is an identity matrix, is a Kalman gain matrix, represents a meteorological element background field vector, represents a real-time meteorological observation data vector obtained by the observation model layer, represents an observation operator matrix, represents a prior covariance matrix of the prior model layer, represents an observation error covariance matrix of the observation model layer.

[0044] Compared with the prior art, the present application provides an agricultural meteorological early warning method based on regional microclimate data interpolation and correction, which has the following beneficial effects:

[0045] 1. The application realizes high spatial resolution and high timeliness of the meteorological analysis field in the complex terrain area with sparse meteorological stations by generating high-resolution, physically reliable static meteorological background field by the generative adversarial network as the prior information of the Bayesian hierarchical dynamic correction model, thereby significantly improving the warning accuracy and timeliness for local sudden agricultural meteorological disasters such as frost and dry hot wind.

[0046] 2. The generative adversarial network model of the application introduces terrain physical constraint loss function and grid point uncertainty estimation field to ensure that the finally generated meteorological element background field is not only rich in spatial details, but also strictly follows the basic physical law of temperature vertical decreasing rate, and can self-evaluate the uncertainty of each grid point, overcoming the defects of physical distortion and difficulty in error evaluation of traditional interpolation method.

[0047] 3. The application constructs a spatial covariance function that fuses geographical distance and terrain distance in the Bayesian hierarchical model, and constructs a prior covariance matrix using the grid point uncertainty estimation field output by the generative adversarial network, so that the Bayesian hierarchical model can more truly depict the correlation structure of meteorological elements under complex terrain, thereby realizing more accurate and reasonable multi-source data fusion.

[0048] 4. The observation model layer of the application constructs a block diagonal structure observation error covariance matrix for multi-source observation data such as fixed meteorological stations and farmland Internet of Things devices, realizes adaptive weighted fusion of different precision observation data, fully utilizes the spatial coverage advantage of high-density Internet of Things data, and effectively suppresses the noise influence, thereby greatly improving the utilization efficiency of observation data while ensuring the stability and reliability of the fusion result.

[0049] 5. The application converts the Bayesian posterior inference into solving a symmetric positive definite linear system, and uses efficient iterative algorithms such as conjugate gradient method for solving, avoiding the direct inversion of super-large scale dense matrix in the traditional method, significantly reducing the computational complexity and memory consumption under the premise of ensuring the theoretical optimality, so that the high-precision method can be efficiently applied to large-scale, high-resolution practical business scenarios, and has good engineering landing property. BRIEF DESCRIPTION OF DRAWINGS

[0050] The drawings described herein are used to provide further understanding of the present application, and constitute a part of the present application, the illustrative embodiments of the present application and the description thereof are used to explain the present application, and do not constitute improper limitation on the present application. In the drawings:

[0051] Figure 1 The flowchart of the agricultural meteorological warning method based on regional microclimate data interpolation and correction of the application. DETAILED DESCRIPTION

[0052] In order to make the above objectives, features and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments. Thus, the implementation process of how to apply technical means to solve technical problems and achieve technical effects can be fully understood and implemented.

[0053] Those of ordinary skill in the art can understand that all or part of the steps in the following embodiment methods can be completed by programs instructing relevant hardware, therefore, the present application can adopt a completely hardware embodiment, a completely software embodiment, or an embodiment in the form of combining software and hardware aspects. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer usable storage media (including but not limited to disk memory, CD-ROM, optical memory, etc.) containing computer usable program codes.

[0054] In order to solve the problem that the prior art is difficult to meet the real-time prediction of microclimate data in complex terrain areas, resulting in lag of agricultural early warning, the present application provides an agricultural meteorological early warning method based on regional microclimate data interpolation and correction, which combines a generative adversarial network with a Bayesian hierarchical model to construct a technical scheme that can generate high-precision and physically credible meteorological element background fields, dynamically fuse real-time observation data, realize interpolation and correction of regional microclimate data, and realize agricultural early warning, as shown in Figure 1 The method specifically includes the following steps:

[0055] S1, first, a certain agricultural area in Xinjiang can be taken as a target region, and all available meteorological stations (including national benchmark stations, basic stations and regional automatic stations, etc.) in the target region are collected for collecting hourly or daily historical meteorological observation data in the same period (such as the past 10 years) in history. Since air temperature significantly affects the growth of crops, the meteorological elements in the historical meteorological observation data at least include air temperature. In addition, terrain is a key static factor affecting the formation of microclimate such as mountain-valley inversion and slope temperature difference, and is the basis for generating terrain features. Digital elevation model data (DEM) of the target region is also needed, and then the historical meteorological observation data generally needs to be strictly quality controlled, such as removing outliers and filling short-time missing data, and uniformly spatially interpolated to a regular grid with a relatively coarse resolution, for example, a 5km x 5km grid, to construct a training data set.

[0056] S2, then in order to realize that under the given input data, can generate and match the real weather field in line with the historical distribution rule, build a topographic physical constraint of the generative adversarial network model and through the training data set model training, the target area of digital elevation model data and sparse site meteorological observation data input trained generative adversarial network model, to generate the target area of meteorological element background field and its corresponding grid grid uncertainty estimation field, wherein the meteorological element at least includes temperature, below begins to introduce the structure of generative adversarial network model GAN in detail, which includes generator G and discriminator D with random noise vector z and deterministic condition information as input. Wherein, the generator G adopts U-Net network with skip connection, the encoder of U-Net network includes parallel first feature extraction branch and second feature extraction branch, wherein each feature extraction branch of U-Net encoder can contain 4 down-sampling stages, each stage consists of two convolutional layers (the parameters are: convolution kernel size 3x3, step 1, padding 1) and a maximum pooling layer (the parameters are: pooling kernel 2x2, step 2), the initial channel number can be set to 64, and the channel number is doubled after each down-sampling stage, the decoder is symmetrical to it, and the transposed convolution is used for up-sampling.

[0057] According to DEM, the terrain feature map is derived Then input it into the first feature extraction branch for feature extraction, through a series of convolution and down-sampling layers, the high-level features of the terrain are extracted, such as mountain trend and basin contour, and the terrain feature map needs to include at least three channels of elevation, slope and slope direction. Because the U-Net and other convolutional neural networks are sensitive to the scale and distribution of input data, it is necessary to normalize the data of the three channels so that the value range is between [0, 1] or [-1, 1], thereby accelerating the convergence of the generative adversarial network model and improving its stability. The second feature extraction branch is similar to the first feature extraction branch, which also extracts the high-level features of the rough weather field through a series of convolution and down-sampling layers. The rough weather field Based on the inverse distance weighted interpolation of historical meteorological observation data. Herein view of the different data characteristics of the terrain feature map and the meteorological observation data, the first feature extraction branch and the second feature extraction branch are set to enable the generative adversarial network to learn the patterns of terrain morphology and meteorological distribution respectively, and then fuse them at the high-level semantic layer. Compared with simple splicing, it can learn more complex nonlinear relationships between the two, such as the nonlinear relationship between the temperature change rate difference on different slopes and the elevation.

[0058] Subsequently, the high-level features extracted by the first feature extraction branch and the second feature extraction branch are used as deterministic condition information, and the encoder is used to fuse the deterministic condition information with the random noise vector z to obtain the fusion feature.

[0059] The decoder of the U-Net network is used for up-sampling the fused features and gradually recovering spatial details by fusing low-level detail features of the corresponding layers in the encoder through a skip connection, and finally outputs a meteorological element background field with a spatial resolution consistent with the terrain feature map .

[0060] The discriminator D adopts a PatchGAN structure of a full convolutional network. In training, a real meteorological field sample is stacked with the corresponding terrain feature map in the channel dimension to obtain a first sample, and a meteorological element background field output by the generator is stacked with the corresponding terrain feature map in the channel dimension to obtain a second sample. The first sample or the second sample is input into the discriminator D to obtain a two-dimensional matrix, and the value of each position in the two-dimensional matrix represents the probability that the local receptive field area corresponding to the input image conforms to the distribution rule of the real meteorological field. The reason for adopting the PatchGAN structure instead of the traditional discriminator is that the PatchGAN structure can force the generator to focus on the authenticity of the local area, ensure that the generated meteorological field is physically consistent with the terrain conditions in each small area (such as a hillside and a river valley, etc.), and avoid the problem of local distortion while only satisfying the global statistical characteristics.

[0061] In the model training phase, the training data set constructed by step S1 is used to conduct adversarial training on the generative adversarial network model GAN. The final goal of the generator G is to generate a meteorological element background field as realistic as possible , and the goal of the discriminator D is to accurately distinguish between real samples and samples generated by the generator G. Through this game training, the generator G can finally learn to generate a more realistic and reasonable meteorological element background field under any given terrain feature map . .

[0062] In the model training phase, a random noise vector z is sampled from a preset prior distribution to introduce diversity. In the model application phase, for generating a meteorological element background field, the random noise vector z is a preset fixed value, such as a zero vector, so as to produce a deterministic output.

[0063] In addition, in order to utilize the trained GAN to generate a grid uncertainty estimation field for uncertainty quantification, the internal randomness of the generator and the confidence of the discriminator on the result are systematically fused to obtain a more comprehensive spatial uncertainty evaluation. The specific steps are as follows:

[0064] After the training of the generative adversarial network model is completed, the parameters of the generator G and the discriminator D are fixed, in order to obtain uncertainty, temporarily release the generator G in the application stage to the random noise vector, based on the same set of deterministic condition information, inject random noise vectors z with different values into the generator G, and perform independent forward propagation through different random noise vectors z to obtain several generated meteorological element background fields, superimpose each generated meteorological element background field and the terrain feature map in the channel, and input the discriminator D to obtain the corresponding two-dimensional matrix, and mark it as a discriminant confidence atlas.

[0065] The sample variance of the meteorological element value of each generated meteorological element background field at each spatial grid point is calculated to obtain a generated variance field The calculation formula of the sample variance at any spatial grid point in the generated variance field can be expressed as: In the above formula, represents the sample variance of the generated variance field at the spatial grid point , represents the number of generated meteorological element background fields, represents the arithmetic mean field of all generated meteorological element background fields, represents the sample variance of the spatial grid point in the arithmetic mean field, represents the arithmetic mean value of the meteorological element value at the spatial grid point in all generated meteorological element background fields, represents the meteorological element value of the nth generated meteorological element background field at any spatial grid point , here is used instead of as the denominator, which is called unbiased estimation in statistics, which can more accurately estimate the overall variance of all possible outputs of the generator through limited samples , and avoid systematic underestimation of uncertainty.

[0066] Then, the arithmetic mean of each discriminant confidence atlas is calculated to obtain an average discriminant confidence field, and the calculation formula of the value of the average discriminant confidence field at any spatial grid point can be expressed as: In the above formula, represents the parameter value of the average discriminant confidence field at any spatial grid point , which represents the probability that the discriminator considers the local area to be real data, represents the parameter value of the nth discriminant confidence atlas at any spatial grid point .

[0067] The final grid point uncertainty estimation field is obtained by fusing the generated variance field and the average discriminative confidence field, and the calculation formula can be expressed as: ; in the above formula, represents the grid point uncertainty estimation field, is a predetermined normal number, wherein is a matrix with the same dimension as the average discriminative confidence field , and each element value is 1, is used to measure the output fluctuation caused by the input noise of the generator, that is, the cognitive uncertainty, is used to measure the degree of deviation of the generated result from the true data distribution as a whole, that is, the distribution uncertainty, and the product means that in the area where the generated result itself is unstable (that is, the variance is large) and is generally considered by the discriminator to be not like the true value (that is, the confidence is low), the comprehensive uncertainty value is given the highest, that is, the grid point uncertainty estimation field of the corresponding area in the above formula (that is, the comprehensive uncertainty value) is given the highest. In addition, the grid point uncertainty estimation field is calculated based on the same set of deterministic condition information mentioned above, and when the random noise vector z is a predetermined fixed value when the meteorological element background field is generated as input, a deterministic output is generated, that is, a deterministic meteorological element background field, and the meteorological element background field at this time can be marked as , so that after a set of deterministic condition information of a target area is input into the generative adversarial network model, the meteorological element background field of the target area and the corresponding grid point uncertainty estimation field can be obtained.

[0068] In addition, considering that it is difficult for the generative adversarial network model to accurately capture the microclimate effects caused by different land cover types such as oases, deserts, and ice and snow, such as oasis cold islands and desert dry heat, only by using terrain and sparse meteorological data, in order to further improve the authenticity of the meteorological element background field generated by the generative adversarial network model, the generator G of the generative adversarial network model is further enhanced in this embodiment, a third feature extraction branch is added in parallel with the first feature extraction branch and the second feature extraction branch in the encoder of the U-Net network as the generator G, satellite remote sensing inversion data spatiotemporally aligned with the terrain feature map are input into the third feature extraction branch, then high-level features of the satellite remote sensing inversion data can also be extracted through a series of convolution and down-sampling layers, and the extracted results are also marked as deterministic condition information, and then the encoder fuses a set of deterministic condition information with the random noise vector z to obtain the fusion features.

[0069] Since the surface temperature directly reflects the energy balance state of the surface, which is a key factor affecting the near-surface air temperature, and the normalized vegetation index is closely related to vegetation transpiration and soil moisture, which can indirectly indicate local air humidity and temperature regulation capacity, therefore, the satellite remote sensing inversion data at least includes the surface temperature and the normalized vegetation index data, in addition, the satellite data also needs to be pre-processed before use, such as cloud detection, missing value filling (such as using spatial and temporal interpolation) and normalization, to ensure the data quality and match the input scale of the generative adversarial network model, satellite remote sensing provides continuous spatial coverage of the underlying surface state information, which is input into the generative adversarial network model, so that the generator can better learn the complex mapping relationship between the surface properties and the near-surface meteorological conditions, for example, the generative adversarial network model can learn that in the same slope direction, the area with vegetation coverage has lower air temperature and higher humidity than the bare soil area. Through the high-level features extracted by the three branches in total, the generator G can generate a more realistic meteorological element background field.

[0070] In the process of fusing the deterministic condition information and the random noise vector z, information from different data sources is involved. In order to let the model automatically learn the importance weight of different data sources and different feature channels, and realize dynamic and context-dependent feature selection instead of simple weighted average, the following fusion steps are further designed:

[0071] Firstly, at a certain depth of the encoder, several deterministic condition information are spliced in the channel dimension to obtain preliminary fusion features, then the preliminary fusion features are globally averaged pooled along the spatial dimension (i.e. width and height) to obtain a channel description vector A, which aggregates the global information of each feature channel, and then a few layers of fully connected networks with bottleneck structure are used to learn the nonlinear relationship between channels to obtain an attention weight vector w. Taking two layers of fully connected networks as an example, its expression can be represented as: In the above formula, and are the learnable weight matrices corresponding to the second layer and the first layer of fully connected networks respectively, is the ReLU activation function, is the Sigmoid activation function. This structure first compresses the channel number to reduce the calculation amount and aggregate the information, and then restores the channel number and outputs the weight of each channel.

[0072] Finally, the attention weight vector w is multiplied with the preliminary fusion features channel by channel to obtain the final fusion features after recalibration, and its expression can be represented as: In the above formula, represents the final fusion features, represents the preliminary fusion features, i.e. the result of splicing the deterministic condition information output by several first layer convolutional layers in the channel dimension.

[0073] This fusion process allows the network to adaptively emphasize information-rich feature channels and suppress channels with little or no information. For example, in mountainous areas, the network may give higher weights to channels related to slope and elevation, while in flat desert areas, it may give higher weights to channels related to surface temperature.

[0074] To train a meteorological background field that generates realistic details and conforms to physical laws, this embodiment also designs a total loss function consisting of three weighted components. Specifically, the total loss function of the generative adversarial network model is a weighted sum of the adversarial loss, reconstruction loss, and physical constraint loss. The details of each loss are as follows:

[0075] Combat losses The least-squares loss drives the generator to produce indistinguishable outputs to deceive the discriminator; its expression can be given as: In the above formula, E represents the mathematical expectation operator, indicating that the expression within the square brackets is averaged over all samples, and x is the value derived from the true data distribution. The real meteorological field samples obtained from the sampling are from the training dataset. This refers to the probability distribution that a real weather field sample follows. Let z be the discriminator function, representing the scalar obtained by global average pooling of the output two-dimensional matrix after inputting samples and given deterministic conditional information C. z represents the value derived from the prior distribution. The random noise vector obtained by sampling in the middle, Let z represent the prior distribution of the random noise z, which is usually defined as the standard normal distribution. Let be the generator function, representing the meteorological background field obtained after processing a random noise vector z and deterministic conditional information C as input. .

[0076] Reconstruction loss Background field of meteorological elements used to characterize the generator output The mean absolute error between the data from known meteorological observation points and the actual observed values ​​is used to constrain the accuracy of the generated results at local points. This loss is a strong constraint, ensuring that the generator's output at locations with observed data must be as close as possible to the true value, guaranteeing the local fidelity of the generated results. This is crucial for injecting station data information into the model, and its calculation formula can be expressed as follows: In the above formula, Represents the background field of meteorological elements The total number of known meteorological observation points in China. and These represent the background fields of meteorological elements. The data and actual observed values ​​at the known observation point s in the meteorological observation data;

[0077] Physical constraint loss This method uses the vertical temperature lapse rate as a soft constraint to limit temperature variation. It constrains the relationship between temperature and geographical location; other constraints can be further set as needed, which will not be listed here. (Physical constraint loss) The expression can be represented as: In the above formula, This represents the total number of grid points in the target area. This represents the background field of meteorological elements at grid point i. The local temperature lapse rate relative to elevation h is typically calculated in practice by performing linear regression on the meteorological and elevation values ​​of the grid point and its surrounding neighboring grid points (e.g., if a 3x3 or 5x5 window is established with the grid point as the center, then other grid points within that window are considered its neighboring grid points). This slope is then obtained, and it represents the local temperature lapse rate. Represents the background field of meteorological elements Temperature data from meteorological element values ​​at grid point i. This represents the elevation value at grid point i. The loss is a pre-defined theoretical or statistical vertical temperature lapse rate that aligns with the climate characteristics of the target region; for example, it might be -0.65℃ / 100m for arid regions like Xinjiang. This loss does not rely on any additional observational data. Instead, it directly embeds the scientific knowledge that temperature decreases with altitude as a soft constraint into the model optimization process. By penalizing parts of the generated results that violate this fundamental physical law, it guides the generator to perform physically reasonable interpolation and extrapolation in areas without observations, greatly improving the scientific credibility of the generated results and avoiding the physically absurd outputs that might arise from a purely data-driven model. Ultimately, the training objective of the generator G is to minimize the total loss.

[0078] The generative adversarial network model can be trained using the Adam optimizer, with an initial learning rate of 0.0002, a batch size of 4, and a training run of 200 epochs.

[0079] Meteorological background field generated in step S2 and its corresponding grid uncertainty estimation field It serves as the input for subsequently constructing the Bayesian hierarchical dynamic correction model, representing the background field of meteorological elements. It provides high-resolution spatial prior estimates, while grid-point uncertainty estimation fields This quantifies the reliability of prior estimates at different locations. The next step will be to construct a Bayesian hierarchical dynamic correction model, which uses the meteorological background field as a basis. and lattice uncertainty estimation field As prior knowledge, and through the process model layer to describe the spatial correlation structure of meteorological elements, ultimately through the fusion of real-time observation data, output a dynamic update and more accurate meteorological element analysis field.

[0080] S3, construct a Bayesian hierarchical dynamic correction model containing an observation model layer, a process model layer and a prior model layer, the input of the observation model layer is the real-time meteorological observation data of the target area, the process model layer is used to represent the spatial correlation structure of meteorological elements, and the prior model layer is constructed by the background field of meteorological elements and its corresponding lattice uncertainty estimation field, the core of the Bayesian hierarchical model is to estimate the unknown meteorological element true value field X, the process model layer is used to associate the information of the prior model layer and the observation model layer, and the most possible state of the meteorological element true value field X is calculated by using the Bayesian theorem after considering all information (posterior distribution), the process model layer is embedded in the prior, which defines the spatial structure of the meteorological element true value field X, and the process model layer of the embodiment is used to construct the covariance matrix of the layer through the spatial covariance function, so as to define the spatial correlation structure of meteorological elements, the spatial covariance function is constructed to be monotonically attenuated with the increase of the geographical distance and the terrain distance between any two spatial positions in the target area, the terrain distance is the elevation difference between two spatial positions, and its expression can be represented as: In the above formula, , the spatial covariance function is represented by and , respectively, the geographical distance and the terrain distance between the positions and in the target area, and , respectively, the elevation values of the positions and provided by the digital elevation model, the geographical distance can be obtained by calculating the plane Euclidean distance, or directly calculating the spherical shortest arc length according to the latitude and longitude of the two positions by using the Haversine formula, , represents the process variance parameter, which is used to describe the overall variability of meteorological elements in space, and , respectively, are range parameters for controlling the attenuation speed of geographical distance and terrain distance, the numerical value of the element in the i-th row and the j-th column of the covariance matrix of the process model layer is This function, by introducing terrain distance, clearly quantifies the decisive attenuation effect of elevation differences on the correlation of meteorological elements. For example, the temperature correlation between two points that are very close horizontally but have a large elevation difference will be calculated as very low by this function, while the correlation between two points that are slightly far horizontally but have similar elevations may be higher. This makes the model's characterization of spatial processes more consistent with the physical understanding of mountain meteorology, laying a correct spatial structure foundation for subsequent data fusion.

[0081] In addition, the prior model layer combines the output of step S2 with the structure of the process model layer to form a complete prior probability description of the true value field X of meteorological elements. The specific steps for its construction are as follows:

[0082] The prior model layer defines the prior distribution of the true value field X of meteorological elements as a multivariate Gaussian distribution, which can be expressed as: In the above formula, As the background field of meteorological elements, and serving as the prior mean vector, Let be the prior covariance matrix of the prior model layer. diagonal matrix and the covariance matrix of the process model layer The sum, its expression is: diagonal matrix diagonal elements in The value of the corresponding grid point in the field is estimated by grid uncertainty. And through a preset monotonically increasing function The mapping yields an expression that can be represented as: In the above formula, N is the total number of grid points in the target region, thus representing the generator G's assessment of its own output uncertainty (i.e., the grid uncertainty estimation field). Transforming into prior variance, grid-point uncertainty estimation field In areas with large median values ​​(such as complex terrain areas with no observations), the prior variance is... The larger the value, the lower the reliability of the prior estimate for that point, and the more likely it is to be adjusted based on the observed data in subsequent corrections.

[0083] diagonal matrix It provides the error magnitude of the point itself in a spatially heterogeneous environment, while the covariance matrix of the process model layer... It provides a spatial correlation structure between points. The sum of both. This means that the prior uncertainty of a point comes not only from its own uncertainty (i.e. Furthermore, it is also affected by the uncertainty of other related points through spatial structure (i.e. The influence transmitted from it. At the same time, The large variance in the middle will also be passed through Diffusion to the surrounding, so that the impact of the low-confidence area is statistically larger and more blurred, more in line with the intuitive understanding of the cognitive uncertainty of complex areas.

[0084] The observation model layer of the Bayesian hierarchical dynamic correction model observes real-time data The relationship between the true value field is defined as: ; in the above formula, represents the real-time meteorological observation data vector obtained from multiple devices, represents the true value field of the meteorological element, represents the observation operator matrix, which is used to map the true value field of the meteorological element to the observation point position. If the jth observation point is located in the ith grid, the corresponding element in the observation operator matrix is , otherwise the value of the element is 0, is the observation error covariance matrix.

[0085] The real-time meteorological observation data vector contains at least observation data from fixed meteorological standard stations within the target area and observation data from Internet of Things mobile observation devices deployed in farmland areas.

[0086] The observation error covariance matrix is a block diagonal matrix used to represent the uncertainty differences of different types of observation data, and its expression can be represented as: ; in the above formula, and respectively represent the observation error covariance sub-matrices and of the observation data of the fixed meteorological standard station and the Internet of Things mobile observation device, and are diagonal matrices, and the diagonal elements and of the two are set according to the calibration accuracy parameters or historical observation error statistical information of the corresponding type of observation device, and satisfy , that is, the observation error variance of the fixed meteorological standard station is smaller than that of the low-cost Internet of Things mobile observation device. The accuracy and reliability of different types of observation data are very different. Generally, the data of the fixed meteorological standard station is more reliable and should be given a higher weight; the Internet of Things data has a large noise and should be given a lower weight, which avoids the pollution of low-precision data to high-precision data, fully utilizes the spatial coverage advantage of high-density Internet of Things data, and at the same time suppresses the noise influence.

[0087] S4, run the Bayesian hierarchical dynamic correction model, fuse the data of the observation model layer and the prior model layer under the constraint of the process model layer through Bayesian inference, and calculate to obtain the meteorological element analysis field and its posterior uncertainty. Its calculation formula can be represented as: ; ; ; in the above formula, and respectively are the mean vector and the covariance matrix of the posterior distribution of the true field of meteorological elements X, that is, the analyzed field of meteorological elements and the posterior uncertainty, respectively, is a unit matrix, is a Kalman gain matrix, represents a background field vector of meteorological elements, represents a real-time meteorological observation data vector obtained by an observation model layer, represents an observation operator matrix, represents a prior covariance matrix of the prior model layer, represents an observation error covariance matrix of the observation model layer, in actual calculation, since the Kalman gain matrix involves an inverse operation of a large-scale dense matrix , to avoid the inversion of the large-scale dense matrix, an iterative optimization algorithm based on the conjugate gradient method can be used to directly solve the mean vector of the posterior distribution, that is, to solve the following linear system: , the coefficient matrix of the equation set usually has good properties (such as symmetric positive definite), and can be efficiently solved by using an iterative algorithm such as the conjugate gradient method, which avoids constructing and storing a large dense matrix, thereby ensuring the calculation feasibility of the method of the present application in a large range and high resolution application scenario.

[0088] S5, based on the analyzed field of meteorological elements, combined with a preset disaster judgment threshold related to a key growth period of a crop, agricultural meteorological early warning information facing specific farmland units is generated and output, the disaster judgment threshold is set based on research results of the relationship between crop physiological characteristics and meteorological stress, and has strong timeliness and pertinence, for example, for the cotton bud boll period, a specific low-temperature frost threshold is set, for the wheat filling period, a composite meteorological index threshold of dry hot wind is set, by comparison, the system automatically identifies which meteorological conditions of which geographical positions may reach a disaster standard within the early warning timeliness, so as to judge whether meteorological early warning needs to be made for each specific farmland unit.

[0089] The above embodiments have introduced the present application in detail, specific examples have been applied in this paper to explain the principles and embodiments of the present application, the above embodiment descriptions are only used to help understand the method of the present application and its core idea; at the same time, for those skilled in the art, according to the idea of the present application, the specific embodiments and application range will be changed, and the above description should not be understood as a limitation of the present application.

Claims

1. An agricultural meteorological early warning method based on regional microclimate data interpolation and correction, characterized in that, The method specifically comprises the following steps: S1, obtaining historical meteorological observation data and digital elevation model data of a target area to construct a training data set; S2, constructing a generative adversarial network model fusing terrain physical constraints and performing model training through the training data set, inputting the digital elevation model data of the target area and the meteorological observation data of sparse sites into the trained generative adversarial network model to generate a meteorological element background field of the target area and a corresponding grid point uncertainty estimation field thereof; wherein the meteorological element at least includes air temperature; S3, constructing a Bayesian hierarchical dynamic correction model comprising an observation model layer, a process model layer and a prior model layer, the input of the observation model layer being real-time meteorological observation data of the target area, the process model layer being used to represent the correlation structure of the meteorological element in space, and the prior model layer being constructed from the meteorological element background field and the corresponding grid point uncertainty estimation field thereof; S4, running the Bayesian hierarchical dynamic correction model, fusing the data of the observation model layer and the prior model layer under the constraint of the process model layer through Bayesian inference, and calculating to obtain a meteorological element analysis field and a posterior uncertainty thereof; S5, based on the meteorological element analysis field, combining a preset disaster judgment threshold related to a critical growth period of crops, generating and outputting agricultural meteorological early warning information facing specific farmland units.

2. The agrometeorological warning method according to claim 1, characterized in that, The generative adversarial network model comprises a generator G taking a random noise vector z and deterministic conditional information as input, and a discriminator D; The generator G adopts a U-Net network with skip connection, and the encoder of the U-Net network comprises a first feature extraction branch and a second feature extraction branch in parallel; The first feature extraction branch is configured to downsample and extract a terrain feature map of high-level features, the terrain feature map derived based on digital elevation model data, and at least including normalized data of three channels of elevation, slope, and slope direction; The second feature extraction branch is configured to down-sample and extract high-level features of a coarse weather field, and the coarse weather field is obtained based on the historical weather observation data after inverse distance weighted interpolation. is obtained based on the historical weather observation data after inverse distance weighted interpolation. The high-level features extracted by the first feature extraction branch and the second feature extraction branch are taken as the deterministic conditional information; The encoder is used to fuse the deterministic conditional information with the random noise vector z to obtain fused features; The decoder of the U-Net network is used for up-sampling the fused features, fusing low-level detailed features in the encoder through a skip connection, and outputting a meteorological element background field consistent with the spatial resolution of the terrain feature map ; The discriminator D adopts a PatchGAN structure, in training, a real meteorological field sample is superimposed with a corresponding terrain feature map in the channel dimension to obtain a first sample, a meteorological element background field output by the generator is superimposed with a corresponding terrain feature map in the channel dimension to obtain a second sample, the first sample or the second sample is input into the discriminator D to obtain a two-dimensional matrix, and the value of each position in the two-dimensional matrix represents the probability that the local receptive field area corresponding to the input image conforms to the real meteorological field distribution rule; In the model training stage, the random noise vector z is obtained from a preset prior distribution In the model application stage, for generating the meteorological element background field, the random noise vector z is a preset fixed value.

3. The agrometeorological warning method according to claim 2, characterized in that, The specific steps of generating the grid point uncertainty estimation field are as follows: After the generative adversarial network model is trained, the parameters of the generator G and the discriminator D are fixed, and the generator G is temporarily released from the fixation of the random noise vector in the application stage; Based on the same set of deterministic conditional information, different random noise vectors z with different values are injected into the generator G, and independent forward propagation is performed through different random noise vectors z respectively to obtain several generated meteorological element background fields; Each generated meteorological element background field is superimposed with a terrain feature map in the channel, and is input into the discriminator D to obtain a corresponding two-dimensional matrix and is marked as a discrimination confidence atlas; calculating a sample variance of the values of the meteorological element at each spatial grid point of the respective generated meteorological element background field to obtain a generated variance field ; The arithmetic mean of each discrimination confidence atlas is calculated to obtain an average discrimination confidence field; The variance field and the average discriminant confidence field are fused to obtain a grid point uncertainty estimation field.

4. The agrometeorological warning method according to claim 2, characterized in that, The encoder of the U-Net network further comprises a third feature extraction branch in parallel, which is used for downsampling and extracting high-level features of the preprocessed satellite remote sensing inversion data which are spatio-temporally aligned with the terrain feature map, and marking the satellite remote sensing inversion data as deterministic condition information, wherein the satellite remote sensing inversion data at least include land surface temperature and normalized vegetation index data.

5. The agrometeorological warning method according to claim 2, characterized in that, The specific steps of obtaining the fused features are as follows: The deterministic condition information is spliced in the channel dimension to obtain preliminary fused features; The preliminary fused features are globally averaged pooled along the spatial dimension to obtain a channel description vector A; A nonlinear relationship between channels is learned through a plurality of fully connected networks with a bottleneck structure to obtain an attention weight vector w; The attention weight vector w is multiplied with the preliminary fused features channel by channel to obtain final fused features after recalibration.

6. The agrometeorological warning method according to claim 2, characterized in that, The total loss function of the generative adversarial network model is a weighted sum of an adversarial loss, a reconstruction loss and a physical constraint loss; the adversarial loss is a least square loss, and the reconstruction loss The meteorological element background field used for characterizing the generator output The average absolute error between the data on the observation point where the meteorological observation data is known and the actual observation value. the physical constraint loss The vertical temperature decrement rate is used as a soft constraint to limit the variation of the temperature.

7. The agrometeorological warning method according to claim 1, characterized in that, In step S3, the process model layer is used to construct a covariance matrix of the layer by a spatial covariance function, so as to define the correlation structure of the meteorological elements in space, wherein the spatial covariance function is constructed to monotonically decay with the increase of the geographical distance and the terrain distance between any two spatial positions in the target region, and the terrain distance is the elevation difference between the two spatial positions.

8. The agrometeorological alert method according to claim 1, characterized in that, In step S3, the specific steps for constructing the prior model layer by the meteorological element background field and the corresponding grid point uncertainty estimation field are as follows: The prior model layer defines the prior distribution of the meteorological element true value field X as a multivariate Gaussian distribution; The prior covariance matrix of the prior model layer is the sum of a diagonal matrix and the covariance matrix of the process model layer, wherein the diagonal elements of the diagonal matrix are obtained from the corresponding grid points in the grid point uncertainty estimation field and are mapped by a preset monotonic increasing function.

9. The agrometeorological alert method according to claim 1, characterized in that, In step S3, the observation model layer of the Bayesian hierarchical dynamic correction model is defined as: ; in the above formula, denotes a real-time meteorological observation data vector acquired from a plurality of devices, denotes a meteorological element true value field, denotes an observation operator matrix for mapping the meteorological element true value field to observation point positions, is an observation error covariance matrix; the real-time weather observation data vector at least contains observation data from fixed weather standard stations within the target area and observation data from Internet of Things mobile observation devices deployed in the farmland area; The observation error covariance matrix is a block diagonal matrix used to represent the uncertainty difference of different types of observation data, and its expression is as follows: ; in the above formula, and respectively represent the observation error covariance sub-matrix of the observation data of the fixed meteorological standard station and the Internet of Things mobile observation equipment, and are diagonal matrices, and the diagonal elements of the two are and respectively set according to the calibration accuracy parameters or historical observation error statistical information of the corresponding type of observation equipment, and satisfy , that is, the observation error variance of the fixed meteorological standard station is smaller than the observation error variance of the Internet of Things mobile observation equipment.

10. The agrometeorological alert method according to claim 1, characterized in that, In step S4, the specific steps of Bayesian inference are as follows: the prior model layer and the observation model layer are fused to obtain the meteorological element analysis field and its posterior uncertainty, and the calculation formula is as follows: ; ; ; in the above formula, and are the mean vector and the covariance matrix of the posterior distribution of the true field X of the meteorological element, that is, the analyzed field and the posterior uncertainty of the meteorological element, respectively, is a unit matrix, is a Kalman gain matrix, represents a background field vector of the meteorological element, represents a real-time meteorological observation data vector obtained by the observation model layer, represents an observation operator matrix, represents a prior covariance matrix of the prior model layer, represents an observation error covariance matrix of the observation model layer.

Images