Extreme weather events and ecological risk prediction methods using generative adversarial networks

Abstract

This application belongs to the field of ecological risk prediction technology, and involves a method for predicting extreme weather events and ecological risks using a generative adversarial network. The meteorological generation branch uses a 3D residual network and a Transformer to synchronously extract the spatiotemporal characteristics of the meteorological field and predict future meteorological conditions. The ecological generation branch dynamically generates an ecological risk map based on meteorological forecast results and terrain data through a dilated convolution U‑Net, avoiding the resolution and dimensional mismatch caused by cross-model interpolation. The discriminator introduces a physical conservation verification module and an ecological association module to jointly optimize the physical rationality loss and ecological logic consistency loss of the generator in adversarial training. In addition, the dynamic feedback mechanism corrects the generation error through real-time observation data, and uses spectral normalization to constrain model parameters to further improve the accuracy of the coupled prediction of extreme events and ecological risks. Therefore, the present invention realizes seamless spatiotemporal coupling of meteorological and ecological fields, and the generated results have both physical conservation and ecological correlation.

Description

Technical Field

[0001] The present application belongs to the technical field of ecological risk prediction, and more specifically, relates to a method for predicting extreme weather events and ecological risks using a generative adversarial network. Background Art

[0002] Intensifying global climate change has led to frequent extreme weather events (such as hurricanes, heat waves, and heavy rains). The resulting ecological risks (such as vegetation degradation and species habitat loss) exhibit complex temporal and spatial correlations. Traditional meteorological forecasting and ecological impact assessment are often conducted in isolation: meteorological models focus on simulating physical processes, while ecological models rely on static thresholds or delayed response analysis. This disconnect leads to two key flaws:

[0003] Lack of physical plausibility: Weather forecasts do not reflect the dynamic response of ecosystems in real time and may violate energy conservation or atmospheric motion constraints;

[0004] Delayed ecological response: Risk assessment relies on historical experience and cannot capture the nonlinear coupling relationship between extreme events and ecological degradation.

[0005] The current mainstream method uses the numerical weather prediction model (NWP) cascaded with the ecological statistical model:

[0006] Weather forecasting stage: Simulate atmospheric motion based on physical equations (such as the Navier-Stokes equations) and output high-resolution weather fields;

[0007] Ecological assessment stage: input meteorological forecast results into ecological models (such as species distribution models) and calculate risk probabilities through regression analysis.

[0008] However, the above method has the following problems: the meteorological model and the ecological model run independently, and data transmission relies on spatiotemporal interpolation, resulting in a mismatch between the spatiotemporal resolution and physical dimensions of the meteorological field and the ecological field (such as the spatial misalignment of wind speed gradient and vegetation stress index). Summary of the Invention

[0009] The present invention provides a method for predicting extreme weather events and ecological risks using a generative adversarial network, aiming to solve the technical problem that meteorological models and ecological models run independently, data transmission relies on spatiotemporal interpolation, resulting in a mismatch in the spatiotemporal resolution and physical dimensions of the meteorological and ecological fields.

[0010] The extreme weather event and ecological risk prediction method using a generative adversarial network includes the following steps:

[0011] Data collection: Collect original meteorological data, remote sensing data, geographic data and ecological baseline data;

[0012] Data processing: The collected raw meteorological data, remote sensing data, geographic data, and ecological baseline data are aligned in time and space, and then integrated to construct a three-dimensional feature cube;

[0013] Calculation of meteorological dynamic characteristics: Based on the constructed three-dimensional characteristic cube, the quasi-geostrophic potential vortex is calculated, the water vapor input flux is extracted, and the physical diagnostic field is obtained;

[0014] Ecological response modeling: construct vegetation stress indices and generate species habitat suitability maps;

[0015] Generative adversarial joint training: A two-branch generative adversarial network is constructed, in which the generator includes a meteorological generation branch and an ecological generation branch. The meteorological generation branch extracts spatial features based on a 3D residual network, combines it with a Transformer to capture temporal dependencies, and outputs the meteorological field at future moments. The ecological generation branch uses a dilated convolutional U-Net model to generate dynamic risk probability maps based on predicted meteorological fields and terrain data. The discriminator is divided into a physical conservation verification module and an ecological correlation module. The physical conservation verification module is used to verify the rationality of the generated meteorological field. The ecological correlation module verifies the logical consistency of the ecological response by analyzing the statistical correlation between vegetation indices and meteorological variables. During the training process, the generator simultaneously optimizes the adversarial loss, physical conservation loss, and ecological correlation loss, ultimately generating a result that seamlessly couples meteorological forecasts with ecological risks.

[0016] Dynamic feedback optimization: The gradient information of the discriminator is used to locate the key areas of generation error, and the generator input layer is linearly corrected in combination with real-time observation data from satellites and buoy stations. Spectral normalization is then used to limit the parameter range of the generator.

[0017] In the present invention, the spatiotemporal consistency of the input data is ensured by aligning multi-source data in time and space and fusing them into a three-dimensional feature cube; subsequently, the meteorological generation branch uses a 3D residual network and Transformer to synchronously extract the spatiotemporal characteristics of the meteorological field and predict future meteorological conditions, while the ecological generation branch dynamically generates an ecological risk map through a dilated convolutional U-Net based on meteorological forecast results and terrain data, avoiding the resolution and dimensional mismatch caused by cross-model interpolation; the discriminator introduces a physical conservation verification module and an ecological association module, and jointly optimizes the physical rationality loss and ecological logic consistency loss of the generator in adversarial training to ensure that the generation results conform to both the laws of atmospheric dynamics and the ecological response mechanism; in addition, the dynamic feedback mechanism corrects the generation error through real-time observation data, and uses spectral normalization to constrain model parameters, further improving the accuracy of the coupled prediction of extreme events and ecological risks; therefore, the present invention realizes seamless spatiotemporal coupling of meteorological and ecological fields, and the generated results have both physical conservation and ecological correlation, and can dynamically adapt to real-time observation data, providing reliable technical support for high-precision collaborative early warning of extreme weather and its ecological impacts.

[0018] Preferably, the data processing comprises the following steps:

[0019] The meteorological data space is downscaled through cubic spline interpolation, and the temporal resolution of remote sensing data is increased to the hourly level using bilinear interpolation to ensure that all data are in a unified spatiotemporal coordinate system. Then, invalid areas are eliminated through dynamic masking technology, and finally data of different dimensions are fused into a multidimensional spatiotemporal dataset to obtain standardized spatiotemporal grid data.

[0020] Preferably, the calculation of the quasi-geostrophic potential vorticity comprises the following steps:

[0021] Relative vorticity calculation: Calculate the spatial gradients of zonal and meridional winds to obtain the rate of change of the wind field on the horizontal plane, and then calculate the relative vorticity based on the second-order central difference method;

[0022] Calculate Coriolis parameters: Calculate the Coriolis parameters of each grid point using the standard formula for Coriolis parameters calculation;

[0023] Calculate potential temperature gradient: Calculate the potential temperature value based on air pressure and temperature, and then use the central difference method to calculate the gradient of potential temperature in the vertical direction;

[0024] Calculation of quasi-geostrophic potential vorticity: The quasi-geostrophic potential vorticity of each grid point is calculated by combining the relative vorticity, Coriolis parameter and potential temperature gradient, and then the calculated quasi-geostrophic potential vorticity is normalized.

[0025] Preferably, the specific steps of extracting the water vapor input flux are as follows:

[0026] Vertical integration: Based on the 37-layer air pressure data provided by ERA5, the trapezoidal integration method is used to calculate the contribution of each layer to the total water vapor flux;

[0027] Statistical water vapor flux components: The zonal flux is obtained by integrating the product of the zonal wind speed and the specific humidity; the meridional flux is obtained by integrating the product of the meridional wind speed and the specific humidity; at each level, the interval between pressure layers is used to weight the contribution of each layer;

[0028] Composite water vapor flux: The modulus of the total water vapor flux is obtained by calculating the sum of the squares of the latitudinal flux and the meridional flux, and the direction angle of the water vapor transport flux is obtained by calculating the ratio of the meridional flux to the latitudinal flux.

[0029] Preferably, the constructing of the vegetation stress index comprises the following steps:

[0030] Calculation of heat stress component: For each pixel, the difference between its daily maximum temperature and the species tolerance threshold is calculated, and the heat stress component is determined using the logistic function;

[0031] Water stress component calculation: The severity of water stress is measured based on the ratio of surface soil moisture to critical moisture. A terrain adjustment factor is introduced to adjust the water stress component. The terrain adjustment factor is calculated based on the terrain slope; the greater the slope, the smaller the terrain adjustment factor.

[0032] Calculation of comprehensive stress index: The heat stress component and water stress component are comprehensively considered, and the stress is corrected in combination with the NDVI value of vegetation to obtain the comprehensive stress index.

[0033] Preferably, the specific steps for generating a species habitat fitness map are as follows:

[0034] Feature engineering: Moran's I index was used to screen environmental variables significantly associated with species distribution and to obtain the selected features.

[0035] MaxEnt model optimization: Based on the selected features and species distribution data, the MaxEnt model is trained to output the suitability score for each habitat;

[0036] Phenological adjustment: Based on vegetation index data, the phenological phase angle of the species is calculated, and the phenological adjustment coefficient is calculated based on the phenological phase angle. The phenological adjustment coefficient is applied to the suitability value output by the MaxEnt model to obtain the seasonally adjusted habitat suitability score; based on the suitability value of each pixel, a species habitat suitability map is generated.

[0037] Preferably, the physical conservation verification module is used to check whether the generated meteorological data complies with the physical conservation principle, and the specific steps are as follows:

[0038] Mass conservation residual : According to the conservation equations of fluid mechanics, calculate the divergence of the fluid velocity field:

[0039] ;

[0040] Where: represents the air density, which is calculated by the ideal gas state equation; A three-dimensional vector representing the wind speed field; represents the divergence of the fluid, which is calculated by the central difference method;

[0041] Energy conservation residual verify:

[0042] ;

[0043] Where: represents the kinetic energy density of the gas; represents the specific heat at constant volume; represents the temperature field; represents the time derivative;

[0044] The mass conservation residual and energy conservation residual calculated above are compared with the threshold to determine whether the generated meteorological data follows the physical conservation principle.

[0045] Preferably, the ecological correlation module verifies the logical consistency of the ecological response by analyzing the statistical correlation between vegetation indices and meteorological variables as follows:

[0046] Normalized Difference Vegetation Index:

[0047] ;

[0048] Where: represents the reflectivity in the near-infrared band; Indicates the reflectivity in the infrared band; represents the normalized difference vegetation index;

[0049] Temperature anomalies calculate:

[0050] ;

[0051] Where: Indicates the actual temperature at the current moment; represents the long-term average temperature;

[0052] Mutual information calculation: Based on mutual information, the correlation between NDVI and temperature anomaly is calculated:

[0053] ;

[0054] Where: Indicates NDVI value and temperature anomalies Probability of simultaneous occurrence; , They are the marginal distributions of NDVI and temperature anomalies; Indicates the number of bins for NDVI and temperature anomalies;

[0055] Risk map matching: The difference between the generated risk map and the true risk map is measured by KL divergence:

[0056] ;

[0057] Where: represents the probability distribution of the generated ecological risk map; represents the probability distribution of the real ecological risk map; N represents the number of pixels in the risk map; represents the probability value of the i-th pixel in the generated ecological risk map; represents the probability value of the ith pixel in the true ecological risk map; represents the matching degree of the risk graph;

[0058] When the difference between the calculated mutual information and the mutual information of the real data is less than the threshold, and the risk map matching degree is less than the threshold, the verification of the ecological association module is passed. If any of the above conditions is not met, the verification is not met.

[0059] The beneficial effects of the present invention include:

[0060] In the present invention, the spatiotemporal consistency of the input data is ensured by aligning multi-source data in time and space and fusing them into a three-dimensional feature cube; subsequently, the meteorological generation branch uses a 3D residual network and Transformer to synchronously extract the spatiotemporal characteristics of the meteorological field and predict future meteorological conditions, while the ecological generation branch dynamically generates an ecological risk map through a dilated convolutional U-Net based on meteorological forecast results and terrain data, avoiding the resolution and dimensional mismatch caused by cross-model interpolation; the discriminator introduces a physical conservation verification module and an ecological association module, and jointly optimizes the physical rationality loss and ecological logic consistency loss of the generator in adversarial training to ensure that the generation results conform to both the laws of atmospheric dynamics and the ecological response mechanism; in addition, the dynamic feedback mechanism corrects the generation error through real-time observation data, and uses spectral normalization to constrain model parameters, further improving the accuracy of the coupled prediction of extreme events and ecological risks; therefore, the present invention realizes seamless spatiotemporal coupling of meteorological and ecological fields, and the generated results have both physical conservation and ecological correlation, and can dynamically adapt to real-time observation data, providing reliable technical support for high-precision collaborative early warning of extreme weather and its ecological impacts. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the embodiments or descriptions of the prior art. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0062] Figure 1 This is a flowchart of the overall steps provided by an embodiment of the present invention.

[0063] Figure 2 An exemplary structural diagram of a generative adversarial network provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0064] In order to make the technical problems, technical solutions and beneficial effects to be solved by this application more clearly understood, this application is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0065] See also Figure 1 As shown, a method for optimizing energy efficiency of diving equipment based on environmental data prediction includes the following steps:

[0066] The extreme weather event and ecological risk prediction method using a generative adversarial network includes the following steps:

[0067] Data collection: Collect original meteorological data, remote sensing data, geographic data and ecological baseline data; meteorological data: including temperature, precipitation, wind speed, humidity, etc., usually from weather stations or satellite observations; remote sensing data: including vegetation index (such as NDVI), land cover data, etc., usually obtained through satellite remote sensing images; geographic data: including terrain height, land use type, slope, etc., usually from remote sensing data or geographic information system (GIS) data; ecological baseline data: including species distribution maps, habitat suitability, etc.

[0068] Data processing: The collected raw meteorological data, remote sensing data, geographic data, and ecological baseline data are aligned in time and space, and then integrated to construct a three-dimensional feature cube;

[0069] The data processing includes the following steps:

[0070] Meteorological data spatial downscaling: The original meteorological data usually has a low resolution, such as 0.25°, and needs to be downscaled to 1km in order to more accurately simulate the impact of meteorology on the ecosystem. In order to perform spatial downscaling, the cubic spline interpolation method is used to smoothly adjust the data to make it more accurate. For example, if the original meteorological data is , and its spatial resolution is , the target spatial resolution is , obtain the target resolution meteorological data through cubic spline interpolation :

[0071] ;

[0072] Where: represents the coefficients of a cubic polynomial; Indicates spatial location;

[0073] Remote sensing data time upscaling: Remote sensing data has a low temporal resolution and needs to be increased to the hourly level to match high-frequency meteorological data. Therefore, bilinear interpolation is used to interpolate the temporal resolution of remote sensing data. For example, the temporal resolution of the original remote sensing data is set to , the target time resolution is , calculate the target time point through bilinear interpolation Corresponding remote sensing data :

[0074] ;

[0075] Where: and Indicates a known point in time; and Represents the remote sensing data at the corresponding time point;

[0076] Data fusion and 3D feature cube construction: After data alignment, a 3D feature cube is constructed, where each dimension represents time, space (longitude and latitude) and different meteorological or ecological variables; all aligned data are sorted by time and space position to form a multidimensional dataset. For meteorological data, For remote sensing data, For geographic data, build a 3D cube:

[0077] ;

[0078] in: represents a three-dimensional feature cube, t represents time; x and y represent spatial coordinates;

[0079] Dynamic masking: For invalid areas (such as oceans, cities, etc.), dynamic masking technology is used to propose the impact of invalid areas. The land use / land cover map in the geographic information data and the high-resolution remote sensing image are used to generate a mask map. The invalid area is set to 0 and the valid area is set to 1. Finally, all data are multiplied by the mask map.

[0080] Then, the masked data is standardized and mapped to a regular spatiotemporal grid to obtain the final spatiotemporal grid dataset. .

[0081] Calculation of meteorological dynamic characteristics: Based on the constructed three-dimensional characteristic cube, the quasi-geostrophic potential vortex is calculated, the water vapor input flux is extracted, and the physical diagnostic field is obtained;

[0082] The calculation of the quasi-geostrophic potential vorticity comprises the following steps:

[0083] Relative vorticity calculation: Calculate the spatial gradients of zonal and meridional winds to obtain the rate of change of the wind field on the horizontal plane, and then calculate the relative vorticity based on the second-order central difference method;

[0084] ;

[0085] Where: and are the latitudinal and longitudinal wind speeds respectively; x and y are the longitudinal and latitudinal coordinates;

[0086] Calculate the Coriolis parameters: Calculate the Coriolis parameters of each grid point using the standard formula for Coriolis parameters calculation; for example:

[0087] ;

[0088] Where: represents the Coriolis parameter; Indicates latitude; represents the angular velocity of the Earth's rotation;

[0089] Calculate potential temperature gradient: Calculate the potential temperature value based on air pressure and temperature, and then use the central difference method to calculate the gradient of potential temperature in the vertical direction;

[0090] ;

[0091] Where: represents the vertical gradient of potential temperature; k represents the vertical layer index, where ERA5 provides 37 layers of pressure data; Indicates the potential temperature of the upper layer; Indicates the potential temperature of the next layer; Indicates the air pressure of the upper layer; Indicates the air pressure of the next layer;

[0092] Calculate the quasi-geostrophic potential vorticity: Combine the relative vorticity, Coriolis parameter and potential temperature gradient to calculate the quasi-geostrophic potential vorticity of each grid point, and then normalize the calculated quasi-geostrophic potential vorticity;

[0093] ;

[0094] Where: represents the quasi-geostrophic potential vorticity; represents the gravitational acceleration; it is normalized using the historical climate mean and standard deviation based on the calculated quasi-geostrophic potential vorticity.

[0095] The specific steps of extracting the water vapor input flux are as follows:

[0096] From the surface pressure to the top of the atmosphere, the integration is performed according to the 37 pressure layers of ERA5. First, the flux component is calculated. The zonal flux component is obtained by integrating the product of the zonal wind speed and the specific humidity:

[0097] ;

[0098] Where: represents the zonal flux component; represents specific humidity; represents the zonal wind speed; represents the pressure difference; k represents the index of the pressure layer;

[0099] Then, the product of the meridional wind speed and the specific humidity is integrated to obtain the meridional flux component:

[0100] ;

[0101] Where: represents the meridional flux component; represents the meridional wind speed;

[0102] The contribution of each layer is calculated using the trapezoidal integration method:

[0103] ;

[0104] Where: Indicates the air pressure of the upper layer; Indicates the air pressure of the next layer;

[0105] Based on the latitudinal and longitudinal flux components, the flux mode length and direction angle are calculated:

[0106] ;

[0107] Where: represents the modulus of flux;

[0108] ;

[0109] Where: represents the flux direction angle.

[0110] By calculating the sum of the squares of the latitudinal flux and the meridional flux, the modulus of the total water vapor flux is obtained, and by calculating the ratio of the meridional flux to the latitudinal flux, the direction angle of the water vapor transport flux is obtained.

[0111] Ecological response modeling: construct vegetation stress indices and generate species habitat suitability maps;

[0112] As a possible implementation of this embodiment, constructing the vegetation stress index includes the following steps:

[0113] Calculation of heat stress component: For each pixel, the difference between its daily maximum temperature and the species tolerance threshold is calculated, and the heat stress component is determined using the logistic function;

[0114] ;

[0115] Where: Indicates the maximum daily temperature; represents the high temperature tolerance threshold of the species; represents the steepness coefficient of the Logistic curve; e represents the base of the natural logarithm;

[0116] Water stress component calculation: The severity of water stress is measured based on the ratio of surface soil moisture to critical moisture. A terrain adjustment factor is introduced to adjust the water stress component. The terrain adjustment factor is calculated based on the terrain slope; the greater the slope, the smaller the terrain adjustment factor.

[0117] ;

[0118] ;

[0119] Where: Indicates surface soil moisture; represents critical soil moisture; Indicates the water stress component; represents the terrain adjustment factor; Indicates the terrain slope;

[0120] Calculation of comprehensive stress index: The heat stress component and the water stress component are comprehensively considered, and the stress is corrected in combination with the NDVI value of the vegetation to obtain the comprehensive stress index;

[0121] ;

[0122] Where: Indicates the normalized vegetation index of the current pixel; Indicates the historical maximum value of the current pixel value.

[0123] As a possible implementation of this embodiment, the specific steps of generating a species habitat fitness map are as follows:

[0124] Feature engineering: Moran's I index was used to screen environmental variables significantly associated with species distribution and to obtain the selected features.

[0125] ;

[0126] Where: represents the spatial weight matrix; and Respectively represent the values of environmental variables at different spatial locations; represents the mean of the environmental variables;

[0127] After filtering by the Moran's I index, the following environmental variables are retained for subsequent modeling: PV anomaly intensity, water vapor flux divergence, VSI, and DEM (terrain elevation);

[0128] MaxEnt model optimization: Based on the selected features and species distribution data, the MaxEnt model is trained to output a suitability score for each habitat. The MaxEnt model is a commonly used species distribution modeling method suitable for dealing with the relationship between physical distribution and environmental variables. Its goal is to train the model using known species distribution points and generate species suitability predictions based on the characteristics of environmental variables. The output of the MaxEnt model is the habitat suitability score for each pixel.

[0129] Phenological adjustment: The habitat suitability of species is not only affected by environmental conditions but also closely related to phenology (seasonal changes). Therefore, the suitability prediction results of the model need to be adjusted according to seasonal changes.

[0130] Seasonal adjustment: Based on MODIS EVI (vegetation index) data, the phenological phase angle of species is calculated, which reflects the growth and reproduction cycle of species in different seasons;

[0131] Phenological adjustment coefficient: Adjust habitat suitability through the phenological adjustment coefficient:

[0132] ;

[0133]

[0134] Where: represents the original fitness value given by the MaxEnt model; represents the adjustment coefficient calculated based on phenological data; A represents the amplitude coefficient; t represents the current time point index; Represents the phenological phase angle; N represents the total number of time points; based on this, we generate a habitat suitability map according to the suitability value of each pixel.

[0135] See also Figure 2As shown in the figure, generative adversarial joint training: a two-branch generative adversarial network is constructed, in which the generator contains a meteorological generation branch and an ecological generation branch. The meteorological generation branch extracts spatial features based on a 3D residual network, combines with the Transformer to capture temporal dependencies, and outputs the meteorological field at future moments; the ecological generation branch uses a dilated convolutional U-Net model to generate a dynamic risk probability map based on the predicted meteorological field and terrain data; the discriminator is divided into a physical conservation verification module and an ecological association module. The physical conservation verification module is used to verify the rationality of the generated meteorological field; the ecological association module verifies the logical consistency of the ecological response by analyzing the statistical correlation between vegetation indices and meteorological variables; during the training process, the generator simultaneously optimizes the adversarial loss, physical conservation loss, and ecological association loss, and finally generates a result of seamless coupling of meteorological forecasts and ecological risks;

[0136] Exemplarily, the weather generation branch is as follows:

[0137] The network structure includes a 3D ResNet-50 backbone network, a Transformer spatiotemporal attention module, and a decoding prediction layer;

[0138] The input of the 3D ResNet-50 backbone network and Splicing along the channel dimension (total number of channels );in Represents the meteorological field at time t (temperature, humidity, wind field, etc., ); represents the quasi-geostrophic potential vorticity;

[0139] Each residual block contains 3D convolution (kernel size 3×3×3, stride 1) + batch normalization + ReLU;

[0140] Output feature map after processing based on residual block ;

[0141] Transformer spatiotemporal attention module:

[0142] Spatial serialization: Reshape into The sequence is then multi-headed attention is used to calculate the spatiotemporal correlation weights, and the features are enhanced based on the calculated spatiotemporal correlation weights to obtain spatiotemporal enhanced features. ;

[0143] Decoding layer prediction: Use 3D deconvolution layer (kernel size 3×3×3, stride 1) to reduce the number of channels to , output weather generation results ;

[0144] In this embodiment, through the combination of the 3D ResNet-50 backbone network and the Transformer spatiotemporal attention module, the network can simultaneously capture the spatial characteristics and temporal correlation of meteorological data; 3D convolution can retain the spatiotemporal structure of the input data, while the Transformer's multi-head attention mechanism can effectively enhance the spatiotemporal dependency; comprehensive consideration of spatiotemporal information makes the predicted meteorological field more in line with the actual situation and more accurate.

[0145] The ecological generation branch network structure is as follows:

[0146] Dilated Convolutional Encoder: Averaging Along the Time Dimension ,get ,and After splicing, input the dilated convolution layer (expansion rate = 2, 4, 8) and output multi-scale features ;

[0147] U-Net decoder: The features of each layer of the encoder are spliced with the upsampling results of the decoder through skip connections; the number of channels is compressed to 1 through 1×1 convolution, and the ecological risk probability map is obtained through sigmoid activation:

[0148] ;

[0149] Where: Represents Sigmoid activation; W represents the convolution kernel parameter; represents the bias term; Represents the splicing result of the jump connection;

[0150] In this embodiment, through the dilated convolutional encoder, the network can extract ecological features at different scales, ensuring the comprehensiveness and accuracy of ecological risk assessment; the dilated convolution can expand the receptive field without increasing the computational cost, and is suitable for processing complex ecological data; the U-Net decoder combined with the jump connection can retain more information, making the generated ecological risk map more detailed and accurate.

[0151] Discriminator:

[0152] The physical conservation verification module is used to check whether the generated meteorological data complies with the physical conservation principle. The specific steps are as follows:

[0153] Mass conservation residual : According to the conservation equations of fluid mechanics, calculate the divergence of the fluid velocity field:

[0154] ;

[0155] Where: represents the air density, which is calculated by the ideal gas state equation; A three-dimensional vector representing the wind speed field; represents the divergence of the fluid, which is calculated by the central difference method;

[0156] Energy conservation residual verify:

[0157] ;

[0158] Where: represents the kinetic energy density of the gas; represents the specific heat at constant volume; represents the temperature field; represents the time derivative;

[0159] The mass conservation residual and energy conservation residual calculated above are compared with the threshold to determine whether the generated meteorological data follows the physical conservation principle.

[0160] In this embodiment, by calculating the residuals of mass conservation and energy conservation, it is ensured that the generated meteorological field data follows the basic laws of physics; it effectively reduces non-physically reasonable prediction results and improves the credibility of the model.

[0161] As a possible implementation of this embodiment, the ecological correlation module verifies the logical consistency of the ecological response by analyzing the statistical correlation between the vegetation index and the meteorological variables in the following steps:

[0162] Normalized Difference Vegetation Index:

[0163] ;

[0164] Where: represents the reflectivity in the near-infrared band; Indicates the reflectivity in the infrared band; represents the normalized difference vegetation index;

[0165] Temperature anomalies calculate:

[0166] ;

[0167] Where: Indicates the actual temperature at the current moment; represents the long-term average temperature;

[0168] Mutual information calculation: Based on mutual information, the correlation between NDVI and temperature anomaly is calculated:

[0169] ;

[0170] Where: Indicates NDVI value and temperature anomalies Probability of simultaneous occurrence; , They are the marginal distributions of NDVI and temperature anomalies; Indicates the number of bins for NDVI and temperature anomalies;

[0171] Risk map matching: The difference between the generated risk map and the true risk map is measured by KL divergence:

[0172] ;

[0173] Where: represents the probability distribution of the generated ecological risk map; represents the probability distribution of the real ecological risk map; N represents the number of pixels in the risk map; represents the probability value of the i-th pixel in the generated ecological risk map; represents the probability value of the ith pixel in the true ecological risk map; represents the matching degree of the risk graph;

[0174] When the difference between the calculated mutual information and the mutual information of the real data is less than the threshold, and the risk map matching degree is less than the threshold, the verification of the ecological association module is passed. If any of the above conditions is not met, the verification is not met.

[0175] In this embodiment, by calculating the mutual information between vegetation index and meteorological variables and the KL divergence of the risk map, it is ensured that the generated ecological risk map is consistent with the actual ecological response, thereby improving the reliability of the ecological risk assessment.

[0176] The generator's loss function consists of three parts: adversarial loss, physical conservation loss, and ecological correlation loss. The specific expression is as follows:

[0177] Fighting Loss:

[0178] ;

[0179] Where: represents the output of the discriminator; represents the output of the generator; represents the distribution of real data; x represents the real data sample; represents the logarithm of the probability of the discriminator for the real sample; z represents the sample in the latent space; represents the distribution of the latent space; represents the mapping of the generator to the latent vector z;

[0180] Physical conservation loss:

[0181] ;

[0182] Where: represents the weight of mass conservation loss; represents the weight of energy conservation loss;

[0183] Ecologically associated losses:

[0184] ;

[0185] Where: Represents the weight of ecological association loss.

[0186] In this embodiment, the adversarial loss function is used to train the generator so that it can generate realistic meteorological fields and ecological risk maps, while enabling the discriminator to correctly distinguish between real data and generated data; the physical conservation loss function is used to ensure that the generated meteorological field meets the physical constraints of mass conservation and energy conservation; the ecological correlation loss function is used to ensure that the generated dynamic risk map conforms to ecological principles, that is, the matching degree between the generated map and the real map.

[0187] In this embodiment, the meteorological field and ecological risk map are generated through the forward propagation generator, and then the adversarial loss, physical conservation loss and ecological correlation loss of the generator are calculated. The parameters of the generator and discriminator are updated according to the loss. The training is completed by repeated iteration until the loss converges.

[0188] Dynamic feedback optimization: The gradient information of the discriminator is used to locate the key areas of generation error, and the generator input layer is linearly corrected in combination with real-time observation data from satellites and buoy stations. Spectral normalization is then used to limit the parameter range of the generator.

[0189] In this embodiment, since the training process of the generator depends on the feedback information of the discriminator, the gradient information of the discriminator can help locate the areas where errors occur when the generator outputs the generated meteorological field and ecological risk map. Therefore, the discriminator's extraction is calculated until the generator adjusts the input data structure or network structure. Based on this, the gradient information of the discriminator's loss function with respect to the generator input is calculated using the chain rule:

[0190] ;

[0191] Where: represents the potential input to the generator; represents the output of the generator; represents the gradient of the discriminator loss with respect to the generator output; represents the derivative of the generator with respect to the latent input;

[0192] The gradient information provided by the discriminator can be used to make corrections at the input layer of the generator. The goal of the correction is to modify the generator's latent space or its input features, thereby improving the accuracy of the generated meteorological field and ecological risk map.

[0193] The real-time data provides actual measurements of meteorological fields and ecological risk maps, which can help locate the spatial and temporal errors of the generator. The accuracy of the generator output is evaluated by calculating the difference between the predicted value and the observed value:

[0194] ;

[0195] Where: It represents the real observation data provided by satellite or buoy station; Represents the predicted value of the generator output; Indicates the generation error;

[0196] Combining the gradient information of the discriminator and real-time observation data, the generated input is corrected:

[0197] ;

[0198] Where: represents the corrected generator input; Represents the learning rate, which controls the corrective effect of the discriminator gradient on the input; Represents the learning rate, which controls the correction effect between the observed data and the generated data error;

[0199] Through linear correction, the generator input will be adjusted so that the generated meteorological field and ecological risk map are more consistent with the real observation data, while satisfying physical conservation and ecological logic;

[0200] Furthermore, in this embodiment, in order to avoid the parameters of the generator network being too large or unstable, spectral normalization technology is used to constrain the weights of the generator, and the generator parameters are ensured to be within a stable range by limiting the maximum eigenvalue of each convolutional layer, thereby improving the stability of training.

[0201] In the present invention, the spatiotemporal consistency of the input data is ensured by aligning multi-source data in time and space and fusing them into a three-dimensional feature cube; subsequently, the meteorological generation branch uses a 3D residual network and Transformer to synchronously extract the spatiotemporal characteristics of the meteorological field and predict future meteorological conditions, while the ecological generation branch dynamically generates an ecological risk map through a dilated convolutional U-Net based on meteorological forecast results and terrain data, avoiding the resolution and dimensional mismatch caused by cross-model interpolation; the discriminator introduces a physical conservation verification module and an ecological association module, and jointly optimizes the physical rationality loss and ecological logic consistency loss of the generator in adversarial training to ensure that the generation results conform to both the laws of atmospheric dynamics and the ecological response mechanism; in addition, the dynamic feedback mechanism corrects the generation error through real-time observation data, and uses spectral normalization to constrain model parameters, further improving the accuracy of the coupled prediction of extreme events and ecological risks; therefore, the present invention realizes seamless spatiotemporal coupling of meteorological and ecological fields, and the generated results have both physical conservation and ecological correlation, and can dynamically adapt to real-time observation data, providing reliable technical support for high-precision collaborative early warning of extreme weather and its ecological impacts.

[0202] The above is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent replacements and improvements made within the spirit and principles of the present application should be included in the scope of protection of the present application.

Claims

1. The extreme weather event and ecological risk prediction method using generative adversarial networks is characterized by: The following steps are involved: Data collection: Collect original meteorological data, remote sensing data, geographic data and ecological baseline data; Data processing: The collected raw meteorological data, remote sensing data, geographic data, and ecological baseline data are aligned in time and space, and then integrated to construct a three-dimensional feature cube; Calculation of meteorological dynamic characteristics: Based on the constructed three-dimensional characteristic cube, the quasi-geostrophic potential vortex is calculated, the water vapor input flux is extracted, and the physical diagnostic field is obtained; Ecological response modeling: construct vegetation stress indices and generate species habitat suitability maps; Generative adversarial joint training: A two-branch generative adversarial network is constructed, in which the generator includes a meteorological generation branch and an ecological generation branch. The meteorological generation branch extracts spatial features based on a 3D residual network, combines it with a Transformer to capture temporal dependencies, and outputs the meteorological field at future moments. The ecological generation branch uses a dilated convolutional U-Net model to generate dynamic risk probability maps based on predicted meteorological fields and terrain data. The discriminator is divided into a physical conservation verification module and an ecological correlation module. The physical conservation verification module is used to verify the rationality of the generated meteorological field. The ecological correlation module verifies the logical consistency of the ecological response by analyzing the statistical correlation between vegetation indices and meteorological variables. During the training process, the generator simultaneously optimizes the adversarial loss, physical conservation loss, and ecological correlation loss, ultimately generating a result that seamlessly couples meteorological forecasts with ecological risks. Dynamic feedback optimization: The gradient information of the discriminator is used to locate the key areas of generation error, and the generator input layer is linearly corrected in combination with real-time observation data from satellites and buoy stations. Spectral normalization is then used to limit the parameter range of the generator.

2. The method for predicting extreme weather events and ecological risks using a generative adversarial network according to claim 1, characterized in that: The data processing includes the following steps: The meteorological data space is downscaled through cubic spline interpolation, and the temporal resolution of remote sensing data is increased to the hourly level using bilinear interpolation to ensure that all data are in a unified spatiotemporal coordinate system. Then, invalid areas are eliminated through dynamic masking technology, and finally data of different dimensions are fused into a multidimensional spatiotemporal dataset to obtain standardized spatiotemporal grid data.

3. The method for predicting extreme weather events and ecological risks using a generative adversarial network according to claim 1, wherein: The calculation of the quasi-geostrophic potential vorticity comprises the following steps: Relative vorticity calculation: Calculate the spatial gradients of zonal and meridional winds to obtain the rate of change of the wind field on the horizontal plane, and then calculate the relative vorticity based on the second-order central difference method; Calculate Coriolis parameters: Calculate the Coriolis parameters of each grid point using the standard formula for Coriolis parameters calculation; Calculate potential temperature gradient: Calculate the potential temperature value based on air pressure and temperature, and then use the central difference method to calculate the gradient of potential temperature in the vertical direction; Calculation of quasi-geostrophic potential vorticity: The quasi-geostrophic potential vorticity of each grid point is calculated by combining the relative vorticity, Coriolis parameter and potential temperature gradient, and then the calculated quasi-geostrophic potential vorticity is normalized.

4. The method for predicting extreme weather events and ecological risks using a generative adversarial network according to claim 1, wherein: The specific steps of extracting the water vapor input flux are as follows: Vertical integration: Based on the 37-layer air pressure data provided by ERA5, the trapezoidal integration method is used to calculate the contribution of each layer to the total water vapor flux; Statistical water vapor flux components: The zonal flux is obtained by integrating the product of the zonal wind speed and the specific humidity; the meridional flux is obtained by integrating the product of the meridional wind speed and the specific humidity; at each level, the interval between pressure layers is used to weight the contribution of each layer; Composite water vapor flux: The modulus of the total water vapor flux is obtained by calculating the sum of the squares of the latitudinal flux and the meridional flux, and the direction angle of the water vapor transport flux is obtained by calculating the ratio of the meridional flux to the latitudinal flux.

5. The method for predicting extreme weather events and ecological risks using a generative adversarial network according to claim 1, wherein: Described constructing vegetation stress index comprises the following steps: Calculation of heat stress component: For each pixel, the difference between its daily maximum temperature and the species tolerance threshold is calculated, and the heat stress component is determined using the logistic function; Water stress component calculation: The severity of water stress is measured based on the ratio of surface soil moisture to critical moisture. A terrain adjustment factor is introduced to adjust the water stress component. The terrain adjustment factor is calculated based on the terrain slope; the greater the slope, the smaller the terrain adjustment factor. Calculation of comprehensive stress index: The heat stress component and water stress component are comprehensively considered, and the stress is corrected in combination with the NDVI value of vegetation to obtain the comprehensive stress index.

6. The method for predicting extreme weather events and ecological risks using a generative adversarial network according to claim 1, wherein: The specific steps to generate a species habitat fitness map are as follows: Feature engineering: Moran's I index was used to screen environmental variables significantly associated with species distribution and to obtain the selected features. MaxEnt model optimization: Based on the selected features and species distribution data, the MaxEnt model is trained to output the suitability score for each habitat; Phenological adjustment: Based on vegetation index data, the species' phenological phase angle is calculated, and the phenological adjustment coefficient is calculated based on the phenological phase angle. The phenological adjustment coefficient is applied to the suitability value output by the MaxEnt model to obtain the seasonally adjusted habitat suitability score; Generate a species habitat suitability map based on the suitability value of each pixel.

7. The method for predicting extreme weather events and ecological risks using a generative adversarial network according to claim 1, wherein: The physical conservation verification module is used to check whether the generated meteorological data complies with the physical conservation principle. The specific steps are as follows: Mass conservation residual : According to the conservation equations of fluid mechanics, calculate the divergence of the fluid velocity field: ； Where: represents the air density, which is calculated by the ideal gas state equation; A three-dimensional vector representing the wind speed field; represents the divergence of the fluid, which is calculated by the central difference method; Energy conservation residual verify: ； Where: represents the kinetic energy density of the gas; represents the specific heat at constant volume; represents the temperature field; represents the time derivative; The mass conservation residual and energy conservation residual calculated above are compared with the threshold to determine whether the generated meteorological data follows the physical conservation principle.

8. The method for predicting extreme weather events and ecological risks using a generative adversarial network according to claim 1, wherein: The ecological correlation module verifies the logical consistency of ecological responses by analyzing the statistical correlation between vegetation indices and meteorological variables. The steps are as follows: Normalized Difference Vegetation Index: ； Where: represents the reflectivity in the near-infrared band; Indicates the reflectivity in the infrared band; represents the normalized difference vegetation index; Temperature anomalies calculate: ； Where: Indicates the actual temperature at the current moment; represents the long-term average temperature; Mutual information calculation: Based on mutual information, the correlation between NDVI and temperature anomaly is calculated: ； Where: Indicates NDVI value and temperature anomalies Probability of simultaneous occurrence; , They are the marginal distributions of NDVI and temperature anomalies; Indicates the number of bins for NDVI and temperature anomalies; Risk map matching: The difference between the generated risk map and the true risk map is measured by KL divergence: ； Where: represents the probability distribution of the generated ecological risk map; represents the probability distribution of the real ecological risk map; N represents the number of pixels in the risk map; represents the probability value of the i-th pixel in the generated ecological risk map; represents the probability value of the ith pixel in the true ecological risk map; represents the matching degree of the risk graph; When the difference between the calculated mutual information and the mutual information of the real data is less than the threshold, and the risk map matching degree is less than the threshold, the verification of the ecological association module is passed. If any of the above conditions is not met, the verification is not met.

Images