Multi-scale cross-latitude atmospheric waveguide height forecasting method based on global space-time segmentation

Abstract

The invention discloses a multi-scale cross-latitude atmospheric waveguide height forecasting method based on global space-time segmentation. The method comprises the following steps: step 1, calculating atmospheric waveguide height data; step 2, partitioning global atmospheric waveguide height data according to latitudes; and step 3, constructing the multi-scale cross-latitude space-time forecasting method based on deep learning. According to the method disclosed by the invention, the advanced deep learning global space-time segmentation multi-scale space-time forecasting model is utilized, and the global reanalysis data set of the European mid-term weather forecast center ERA5 is combined, so that the accurate forecasting of the atmospheric waveguide height in the global range is realized; and important technical support and decision basis are provided for the fields of radio communication, radar detection and the like.

Description

Technical Field

[0001] The invention belongs to the technical field of radio meteorological environment forecasting and data preprocessing, and particularly relates to a multi-scale cross-latitude atmospheric duct height forecasting method based on global time-space segmentation in this field. Background Art

[0002] In the field of atmospheric science, atmospheric ducts, as a special phenomenon of electromagnetic wave propagation, have a profound impact on many fields such as radar detection, radio communications, and remote sensing observations. Atmospheric ducts can guide electromagnetic waves to propagate along specific paths, achieving beyond-line-of-sight communication and detection, which is of great significance for improving the efficiency and accuracy of military reconnaissance, civil aviation navigation, ocean monitoring and other activities. However, the formation, development and disappearance of atmospheric ducts are affected by many factors, including meteorological and natural environmental conditions, geographical topography, and seasonal changes throughout the year, making the prediction of atmospheric duct height and duct characteristics an extremely challenging task.

[0003] In the field of atmospheric duct height prediction, we are currently facing a series of complex and urgent technical problems. First, the atmospheric duct phenomenon is characterized by high uncertainty and non-uniformity. Its formation, development and disappearance are affected by multiple factors such as various meteorological conditions, topography and seasonal changes, making it difficult for traditional forecasting methods to accurately capture its changing patterns. Secondly, there are significant geographical differences in the characteristics of atmospheric ducts around the world, and the atmospheric environment in different latitudes has different effects on duct phenomena, which requires that the forecasting method must have the ability to perform cross-latitude analysis. In addition, with the rapid development of science and technology, the accuracy and real-time requirements for atmospheric duct height prediction in various fields are increasing, and existing methods often find it difficult to meet these requirements at the same time. Therefore, how to accurately predict the atmospheric duct height on a global scale has become a technical problem that needs to be solved urgently.

[0004] At present, the prediction methods of atmospheric duct height mainly include empirical estimation, model statistics and numerical prediction models. The empirical model relies on a large amount of observation data and obtains the empirical relationship between duct height and meteorological elements through statistical analysis, but it is difficult to cope with the complex and changeable atmospheric environment. The statistical model is based on statistical principles and makes predictions by establishing a statistical relationship between the prediction variables and the duct height, but it is also limited by the amount of data and data quality. Although the numerical model can simulate the atmospheric physical process more accurately, it has a large amount of calculations, takes a long time, and has high requirements for the accuracy of initial conditions and boundary conditions, which makes it difficult to meet the needs of real-time prediction. More importantly, most of the existing prediction methods are based on a single global site or local area, and lack the ability to predict and analyze global atmospheric ducts at multiple scales and across latitudes. This limitation leads to the forecast results often failing to fully reflect the actual changes in the atmospheric duct phenomenon, thus affecting the forecast accuracy and the application effect of duct height forecast. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a multi-scale cross-latitude spatiotemporal forecasting method for global atmospheric duct height based on ERA5 reanalysis data in view of the importance of atmospheric duct height forecasting and the limitations of existing methods. The method makes full use of the high-resolution and high-precision global atmospheric parameter data provided by the ERA5 dataset, combines deep learning multi-scale analysis and cross-latitude forecasting technology, and realizes a comprehensive and accurate forecast of atmospheric duct height.

[0006] The present invention adopts the following technical solution:

[0007] A method for predicting the height of an atmospheric duct based on global time-space segmentation and multi-scale cross-latitude, the improvement of which is that it comprises the following steps:

[0008] Step 1, calculation of atmospheric duct height data:

[0009] Preprocess ERA5 data, calculate atmospheric refractivity and duct height, and obtain global atmospheric duct height data;

[0010] Step 2: partition the global atmospheric duct height data according to latitude:

[0011] Divide the world into three regions: low latitudes, mid-latitudes, and high latitudes;

[0012] Step 3: Construct a multi-scale cross-latitude spatiotemporal forecasting method based on deep learning:

[0013] Step 31, constructing a time feature prediction module:

[0014] It consists of three time parsers, each of which combines convolution operations, normalization, and activation functions, and adopts the U-Net encoder-decoder structure and its skip connection characteristics;

[0015] Step 32, construct a spatial feature prediction module:

[0016] The spatial feature prediction module has three branches. Branch 1 is the "local detail path". The operations in this path include more than three layers of 3x3 convolutions, each layer followed by batch normalization and activation functions; Branch 2 is the "medium-scale feature path". The operations in this path include more than three layers of 5x5 convolutions, each layer followed by batch normalization and activation functions; Branch 3 is the "global context path". The operations in this path include pooling operations and 7x7 convolutions, each layer contains batch normalization and activation functions;

[0017] Step 33, the spatiotemporal feature fusion module integrates the features learned by the temporal feature prediction module and the spatial feature prediction module to obtain the spatiotemporal prediction results of the global atmospheric duct height.

[0018] Furthermore, in step 1, the atmospheric refractive index N is quantitatively expressed as:

[0019]

[0020] In the above formula, T is the atmospheric temperature, P is the pressure, and e is the water vapor pressure;

[0021] The water vapor pressure e is related to the relative humidity RH and the actual temperature T by the following formula c Related:

[0022]

[0023] The calculation formula for correcting the atmospheric refractive index M is:

[0024] M=N+0.157h

[0025] In the above formula, h is the height from the ground. By taking the derivative of this formula, we can get the atmospheric correction refractive index gradient:

[0026]

[0027] When the atmospheric correction refractive index gradient When it is less than 0.079, an atmospheric duct phenomenon occurs.

[0028] Furthermore, in step 2, the low-latitude region refers to the region between 30° north latitude and 30° south latitude, the mid-latitude region refers to the region between 30° north latitude and 60° north latitude, and between 30° south latitude and 60° south latitude; the high-latitude region refers to the region between 60° north latitude and 90° north latitude, and between 60° south latitude and 90° south latitude.

[0029] Further, in step 31, the workflow of the time feature prediction module includes feature extraction and feature mapping:

[0030]

[0031] ZencfEnc(X)andzdec=fDec(zenc)

[0032] In the above formula, Y represents the output time characteristic forecast result, n represents the length of spatiotemporal data, and f T emporal(X) represents the input-output mapping of the overall model, where f T emporal is a sequence model composed of multiple U_Inception modules; the second formula represents the calculation process of the encoder and decoder inside the U_Inception module, zen c =f Enc(X) represents the calculation process of the encoder inside the U_Inception module. c =f Dec (zen c ) represents the calculation process of the decoder inside the U_Inception module, which completes the conversion from input to output through feature extraction and mapping.

[0033] Furthermore, in the "local detail path" of the spatial feature prediction module in step 32, the activation function is defined as f(x) = max(0, x), where x represents the linear combination of the convolutional layer outputs.

[0034] The beneficial effects of the present invention are:

[0035] The method disclosed in the present invention utilizes an advanced deep learning global space-time segmentation multi-scale space-time forecast model, combined with the global reanalysis data set of ERA5 of the European Centre for Medium-Range Weather Forecasts, to achieve accurate forecasts of atmospheric duct heights on a global scale, providing important technical support and decision-making basis for fields such as radio communications and radar detection.

[0036] The multi-scale cross-latitude spatiotemporal forecasting method for global atmospheric duct height based on ERA5 reanalysis data disclosed in the present invention achieves a significant improvement in the forecast accuracy of atmospheric duct height by integrating high-precision ERA5 data sets with advanced multi-scale cross-latitude forecasting technology, while increasing the stability and robustness of the forecast results. This method not only captures the complex changes of atmospheric duct phenomena at different time and space scales, but also fully considers the impact of geographic latitude on duct characteristics, provides continuous and dynamic forecast results, and exhibits a high degree of diversity and adaptability. In addition, the versatility and scalability of the method of the present invention enable it to cross disciplinary boundaries, provide strong technical support for multiple fields such as atmospheric science research, radar engineering, and communication technology, promote the deep integration of scientific research and engineering applications, and promote technological progress and innovative development in related fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a schematic diagram of the global atmospheric duct height data divided by latitude;

[0038] Figure 2 It is a model diagram of cross-latitude multi-scale spatiotemporal forecasting method;

[0039] Figure 3 is the experimental training loss and validation loss graph;

[0040] Figure 4 It is a comparison chart of the actual value and the predicted value in different periods;

[0041] Figure 5 It is a graph of the absolute value of the error between the true value and the predicted value in different time periods. DETAILED DESCRIPTION

[0042] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0043] Example 1: This example discloses a method for predicting the height of atmospheric ducts across latitudes based on global time-space segmentation and multi-scale. The core of this method is to make full use of the global high-resolution and high-precision atmospheric parameter information provided by the ERA5 dataset, and to combine multi-scale analysis and cross-latitude prediction technology to construct an efficient and accurate prediction method. Specifically, the following steps are included:

[0044] Step 1, calculation of atmospheric duct height data:

[0045] Preprocessing of ERA5 reanalysis data mainly involves data calculation and judgment. The atmospheric refractive index and duct height are calculated through the temperature, humidity, pressure and atmospheric height parameters in the original ERA5 data to obtain global atmospheric duct height data.

[0046] When deeply processing ERA5 reanalysis data, the first task is to carry out detailed data preprocessing, which includes using complex calculation and logical judgment methods based on core parameters such as temperature, humidity, pressure and atmospheric height in the original data set to accurately calculate the atmospheric refractive index and further derive atmospheric duct height data on a global scale.

[0047] In this study, the ERA5 dataset downloaded from the European Centre for Medium-Range Weather Forecasts (ECMWF) was used to derive atmospheric duct height data. The large troposphere is a layer of the atmosphere where radio waves are able to propagate outside the usual curvature of the Earth due to specific conditions such as temperature inversion. These conditions are captured in the ERA5 dataset through various parameters such as temperature, pressure and humidity.

[0048] According to Debye theory, the atmospheric refractive index N is quantitatively expressed as:

[0049]

[0050] In the above formula, T is the atmospheric temperature (unit: K), P is the pressure (unit: hPa), and e is the water vapor pressure (unit: hPa);

[0051] The water vapor pressure e is related to the relative humidity RH (unit: percentage) and the actual temperature T by the following formula c (Unit: Celsius) Related:

[0052]

[0053] In order to correct the influence of the earth's curvature, the atmospheric refractive index is usually corrected. The calculation formula for the corrected atmospheric refractive index M is:

[0054] M=N+0.157h

[0055] In the above formula, h is the height from the ground (unit: km). By taking the derivative of this formula, we can get the atmospheric correction refractive index gradient:

[0056]

[0057] When the atmospheric correction refractive index gradient When it is less than 0.079, it means that the refractive index decreases with increasing altitude. In this environment, the propagation path of electromagnetic waves will present a curvature smaller than the curvature of the earth's surface. Therefore, electromagnetic waves tend to propagate toward the earth's surface and experience multiple reflections between the upper and lower boundaries of this layer of atmosphere, forming an atmospheric waveguide phenomenon.

[0058] Step 2: partition the global atmospheric duct height data according to latitude:

[0059] Since the distribution of atmospheric ducts has latitude heterogeneity on a global scale, the closer to the equator, the higher the probability of atmospheric ducts; the closer to the poles, the lower the probability of atmospheric ducts. Therefore, the world is divided into three regions: low latitudes (near the equator), mid-latitudes (temperate regions), and high latitudes (near the poles). Each region uses customized forecast model parameters according to its distribution density characteristics;

[0060] Divide the logic into Figure 1 As shown, the low-latitude region refers to the area between 30° north latitude and 30° south latitude, the mid-latitude region refers to the area between 30° north latitude and 60° north latitude, and between 30° south latitude and 60° south latitude; the high-latitude region refers to the area between 60° north latitude and 90° north latitude, and between 60° south latitude and 90° south latitude.

[0061] Step 3: Construct a multi-scale cross-latitude spatiotemporal forecasting method based on deep learning:

[0062] The temporal feature prediction module is used to capture the long-term temporal dependence of sub-regions; the spatial feature prediction module is used to focus on the changes in spatial scale, especially for different regions after latitude partitioning, to analyze the spatial distribution and interaction of atmospheric parameters. Finally, the spatiotemporal feature fusion module fuses the temporal and spatial features to achieve global atmospheric duct height prediction.

[0063] Step 31, constructing a time feature prediction module:

[0064] This module is specially designed to process time series data with non-uniform characteristics, such as atmospheric waveguide data. It consists of three time resolvers (UTCN, namely Unified Temporal Convolutional Network), each of which integrates convolution operations, normalization and activation functions, and adopts a similar U-Net encoder-decoder structure and its jump connection characteristics; through the multi-scale time resolver UTCN, combined with the encoder-decoder structure and jump connection of U-Net, it effectively processes non-uniform time series data and improves the accuracy and robustness of time forecasting. The model adaptively adjusts the convolution kernel size according to different regions and scales of the data to better capture local details and global patterns, providing a new solution for time series forecasting.

[0065] like Figure 2 As shown in area (a), the input data size is reshaped to (B, T*C, Lat, Lon), and the multi-scale is divided into low scale, medium scale and high scale. Different convolution kernel sizes correspond to different scales, such as 3*3, 5*5, 7*7. The convolution kernel sizes used are also different for different sub-regions. For low-latitude regions, the probability of atmospheric ducts is high, the distribution is wide, and the data is dense. Usually, rich local features can be captured. When selecting convolution kernels, larger convolution kernels tend to be selected; on the contrary, for high-latitude regions, smaller convolution kernels are preferred to capture local features. Therefore, whether for the same region or different regions, multi-scale fusion is used for time series forecasting. The main calculation steps of the model include feature extraction and feature mapping. These steps are expressed as a set of formulas, feature extraction and mapping formulas:

[0066]

[0067] The core calculation of U_Inception:

[0068] z enc =f Enc (X) and z dec =f Dec (z enc )

[0069] In the above formula, Y represents the output time characteristic forecast result, n represents the length of spatiotemporal data, and f Temporal (X) represents the input-output mapping of the overall model, where f Temporal is a sequence model composed of multiple U_Inception modules; the second formula represents the calculation process of the encoder and decoder inside the U_Inception module, z enc =f Enc(X) represents the calculation process of the encoder inside the U_Inception module, z dec =f Dec (z enc ) represents the calculation process of the decoder inside the U_Inception module, which completes the conversion from input to output through feature extraction and mapping.

[0070] Step 32, construct a spatial feature prediction module:

[0071] like Figure 2 As shown in the (b) area in the figure, the spatial feature prediction module has three branches. Branch one is the "local detail path", which uses a smaller convolution kernel (3x3) to capture local details and short-term spatial changes in the data, which is crucial for understanding subtle changes in spatial data. The operations in this path include multiple layers of 3x3 convolutions, each followed by batch normalization and activation functions (ReLU) to ensure nonlinearity and improved training dynamics; Branch two is the "medium-scale feature path", which uses a slightly larger convolution kernel (5x5) and combines a deeper network structure. The choice of a 5x5 kernel allows the model to integrate information from a wider neighborhood, enabling it to detect medium-scale spatial patterns that may be missed by smaller kernels. The deeper network structure in this branch enhances its ability to learn more complex representations and helps extract medium-scale spatial features. The operations in this path include more than three layers of 5x5 convolutions, each followed by batch normalization and activation functions to maintain stable gradients and introduce nonlinearity; Branch three is the "global context path", which aims to capture the overall spatial context and long-distance dependencies in the data. This path does this by reducing the spatial resolution through pooling layers, which aggregate information from larger regions. After the pooling operation, this path applies larger convolution kernels (e.g. 7x7) to the reduced resolution feature maps. Larger convolution kernels are able to cover a wider range of spatial context, allowing the model to integrate information from distant regions. The global context path is particularly valuable for understanding broad spatial trends and long-range interactions. Similar to the other paths, each layer in this branch contains batch normalization and ReLU activation functions to enhance the learning process;

[0072] Local detail pathway: captures local details and short-term spatial variations in the data. Use multiple 3×3 convolution kernels. This small size of convolution kernel helps the model focus on local areas, thereby capturing subtle spatial variations. Batch normalization is applied after each convolution layer to speed up the training process and reduce internal covariate shift. The ReLU (Rectified Linear Unit) activation function is used to introduce nonlinearity to help the model learn complex feature representations. The ReLU activation function is defined as f(x) = max(0,x), where x represents the linear combination of the output of the convolution layer, that is, the input signal of each layer in the network.

[0073] Medium-scale feature pathway: Detects medium-scale spatial patterns and integrates information from a wider neighborhood. Use a 5×5 convolution kernel. This size of convolution kernel can cover a larger area than 3×3, thus capturing a wider range of spatial features. Due to the need to process more complex spatial patterns, the medium-scale feature pathway usually contains a deeper network structure than the local detail pathway. Similarly, batch normalization and ReLU activation function are applied after each convolution layer.

[0074] Global context path: Captures the overall spatial context and long-range dependencies in the data. A 7×7 (or larger) convolution kernel is applied. This large kernel size is able to cover a wider range of spatial context. Again, each convolution layer is followed by batch normalization and ReLU activation function.

[0075] Step 33, the spatiotemporal feature fusion module integrates the features learned by the temporal feature prediction module and the spatial feature prediction module to obtain the spatiotemporal prediction results of the global atmospheric duct height.

[0076] The model of the method is trained and verified, and the forecast performance is evaluated, including indicators such as accuracy, stability and real-time performance. Based on the evaluation results, the model is iteratively optimized, and the model structure and parameter settings are adjusted to meet the needs of different application scenarios.

[0077] In order to ensure the effectiveness and robustness of the multi-scale cross-latitude spatiotemporal forecasting method, this application designs a scientific and reasonable method model training and verification process. The whole process includes key steps such as data preparation, model training, model verification and performance evaluation.

[0078] In order to demonstrate the performance of the spatiotemporal forecasting method of this application, it is run on a single NVIDIA4090 GPU. During training, the mean square error (MSE) as the loss function and the adaptive moment estimation (Adam) as the optimizer are applied, the learning rate is set to 0.001, the batch_size is set to 2, and the model is trained by 100 iterations. For both pairs of samples, 80% of the data are selected as training samples, and the remaining 20% ​​of the data are used as test data. In addition, the training and test samples are selected in chronological order rather than randomly. For data mapping, the test samples are samples of the atmospheric duct height corresponding to the same time point as the training samples; for spatiotemporal forecasts, the test samples are future samples relative to the training samples.

[0079] The data selected are ERA5 model reanalysis data from January 1, 2016 to March 31, 2016, a total of 90 days. The original data is constructed in the following format: [2160,3,38,721,1440], where the first number is the number of ERA5 samples, the next number represents the temperature, humidity and pressure attributes in the ERA5 original data, the third number is the data of the 100th to 137th layers, a total of 38 layers, and the last two numbers are the latitude Lat and longitude Lon of the input data. After calculation, the output is the atmospheric waveguide height, without stratification, and the output data format is: [2160,1,721,1440], which is also the format of the input data of the spatiotemporal forecast. The input data format is [2160, 1, 721, 1440], where the first number is the number of ERA5 samples, and [1, 721, 1440] represents the size of a sample global atmospheric duct height data. It is input into the model and passed through the time feature prediction module and the spatial feature prediction module respectively. Finally, the two results are fused to obtain the prediction result. The experimental results are as follows Figure 3 — Figure 5 As shown, different comparison results are shown in Table 1 and Table 2 below.

[0080] Table 1: Comparison of global data with different segmentation methods and different forecast step lengths

[0081]

[0082] Table 2: Comparison of global atmospheric duct height prediction results from different prediction models

[0083]

[0084] In summary, based on an in-depth exploration of the complexity and dynamics of the global climate system, this application has significantly improved the accuracy and reliability of climate forecasts through a series of carefully designed strategies. First, taking into account the significant differences in climate characteristics, ecosystem distribution, and the impact of human activities in different latitudes on the earth's surface, this application adopts a global regional subdivision method based on latitude differences, dividing the world into three unique latitude bands: low latitudes (near the equator), mid-latitudes (temperate regions), and high latitudes (near the poles). This subdivision strategy not only helps the model better understand and simulate the climate characteristics of different regions, but also fully captures the spatial heterogeneity caused by latitude changes, providing a more refined data foundation for subsequent deep learning and feature extraction.

[0085] In order to further explore the spatiotemporal characteristics of climate data, this application designs two core modules: the time multi-scale perception module and the space multi-scale perception module. The time multi-scale perception module aims to enhance the model's ability to capture time correlations. By introducing a multi-level temporal convolutional network structure and attention mechanism, the module can automatically learn and adapt to climate fluctuations and trends on different time scales. Whether it is short-term weather changes or long-term climate evolution, it can provide accurate and comprehensive temporal feature representation, providing strong temporal dimension support for model forecasting.

[0086] At the same time, the spatial multi-scale perception module focuses on the feature extraction of spatial variability. This module uses advanced convolutional neural network technology and pyramid spatial pooling strategies to capture local details and global patterns of climate data from multiple spatial scales. Through continuous iteration and optimization, it can generate highly representative spatial feature maps that not only reflect the spatial distribution characteristics of the climate system, but also reveal the interactions and impacts between different regions.

[0087] This application achieves deep integration and mining of spatiotemporal correlation by cleverly integrating the outputs of the two modules. This fusion process not only retains the fine information of temporal changes and spatial distribution, but also promotes the interaction and impact analysis between the two. By deeply mining spatiotemporal correlation, this application can more accurately predict future climate change trends and provide strong technical support for climate science research, disaster warning, and policy making.

[0088] In general, the submission of this application marks an important step forward in the field of global climate forecasting, providing a more scientific, accurate and effective solution to addressing the challenges of global climate change.

Claims

1. A multi-scale cross-latitude atmospheric duct height prediction method based on global space-time segmentation, characterized in that: The steps include: Step 1, calculation of atmospheric duct height data: Preprocess ERA5 data, calculate atmospheric refractivity and duct height, and obtain global atmospheric duct height data; Step 2: partition the global atmospheric duct height data according to latitude: Divide the world into three regions: low latitudes, mid-latitudes, and high latitudes; Step 3: Construct a multi-scale cross-latitude spatiotemporal forecasting method based on deep learning: Step 31, constructing a time feature prediction module: It consists of three time parsers, each of which combines convolution operations, normalization, and activation functions, and adopts the U-Net encoder-decoder structure and its skip connection characteristics; Step 32, construct a spatial feature prediction module: The spatial feature prediction module has three branches. Branch 1 is the "local detail path". The operations in this path include more than three layers of 3x3 convolutions, each followed by batch normalization and activation functions; Branch 2 is the "medium-scale feature path". The operations in this path include more than three layers of 5x5 convolutions, each followed by batch normalization and activation functions; Branch 3 is the "global context path". The operations in this path include pooling operations and 7x7 convolutions, each layer contains batch normalization and activation functions; Step 33, the spatiotemporal feature fusion module integrates the features learned by the temporal feature prediction module and the spatial feature prediction module to obtain the spatiotemporal prediction results of the global atmospheric duct height.

2. The method for predicting the atmospheric duct height based on global time-space segmentation and multi-scale cross-latitude according to claim 1 is characterized in that: In step 1, the atmospheric refractive index N is quantitatively expressed as: In the above formula, T is the atmospheric temperature, P is the pressure, and e is the water vapor pressure; The water vapor pressure e is related to the relative humidity RH and the actual temperature T by the following formula c Related: The calculation formula for correcting the atmospheric refractive index M is: M=N+0.157h In the above formula, h is the height from the ground. By taking the derivative of this formula, we can get the atmospheric correction refractive index gradient: When the atmospheric correction refractive index gradient When it is less than 0.079, an atmospheric duct phenomenon occurs.

3. The method for predicting the atmospheric duct height based on global time-space segmentation and multi-scale cross-latitude according to claim 1 is characterized in that: In step 2, the low-latitude region refers to the area between 30° north latitude and 30° south latitude, the mid-latitude region refers to the area between 30° north latitude and 60° north latitude, and between 30° south latitude and 60° south latitude; the high-latitude region refers to the area between 60° north latitude and 90° north latitude, and between 60° south latitude and 90° south latitude.

4. The method for predicting the atmospheric duct height based on global space-time segmentation and multi-scale cross-latitude according to claim 1 is characterized in that: In step 31, the workflow of the temporal feature prediction module includes feature extraction and feature mapping: z enc =f Enc (X)and z dec =f Dec (z enc ) In the above formula, Y represents the output time characteristic forecast result, n represents the length of spatiotemporal data, and f Temporal (X) represents the input-output mapping of the overall model, where f Temporal is a sequence model composed of multiple U_Inception modules; the second formula represents the calculation process of the encoder and decoder inside the U_Inception module, z enc =f Enc (X) represents the calculation process of the encoder inside the U_Inception module, z dec =f Dec (z enc ) represents the calculation process of the decoder inside the U_Inception module, which completes the conversion from input to output through feature extraction and mapping.

5. The method for predicting the atmospheric duct height based on global time-space segmentation and multi-scale cross-latitude according to claim 1 is characterized by: In the "local detail path" of the spatial feature prediction module in step 32, the activation function is defined as f(x) = max(0, x), where x represents the linear combination of the convolutional layer outputs.

Images