A method and apparatus for radar reflectivity inversion based on meteorological satellite data

Abstract

This invention provides a radar reflectivity inversion method and apparatus based on meteorological satellite data, relating to the field of data processing. The method includes: stitching together multi-time data along the channel dimension based on target satellite data and target radar combined reflectivity data in a sample set to obtain satellite feature data and time-series radar combined reflectivity data; generating an initial radar combined reflectivity inversion model by combining several axial attention modules, several global attention modules, and two feature change modules; discretizing the time-series radar combined reflectivity data; classifying the echoes at each point based on the discretization results; using the difference between the echo intensity at each point and the minimum value of its category as offset data; and training the initial radar combined reflectivity inversion model using echo category data, offset data, and input feature data to obtain the final radar combined reflectivity inversion model.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method and apparatus for radar reflectivity inversion based on meteorological satellite data. Background Technology

[0002] Currently, due to factors such as cost, environment, and terrain, radar echoes cannot cover all areas, and environmental factors also contribute to noise in radar echoes. Radar echo inversion algorithms play a crucial role in supplementing echo information in missing areas and filtering out abnormal echoes. These algorithms aim to invert radar echo data using meteorological satellite data combined with deep learning methods. Existing technologies typically employ deep learning methods for radar echo inversion, often using MAE or MSE loss mechanisms. However, this approach can lead to blurred and smoothed echo results, resulting in lower echo values ​​and ultimately lower accuracy. Summary of the Invention

[0003] The purpose of this invention is to provide a radar reflectivity inversion method and apparatus based on meteorological satellite data, so as to solve the technical problem of low accuracy of radar reflectivity inversion results.

[0004] In a first aspect, this application provides a radar reflectivity inversion method based on meteorological satellite data, the method comprising:

[0005] A sample set is obtained based on the collected digital terrain data, land use type data, satellite data, and radar combined reflectivity data of the target area;

[0006] Based on the target satellite data and target radar combined reflectivity data in the sample set, the multi-time data are spliced ​​together in the channel dimension to obtain satellite feature data and time-series radar combined reflectivity data.

[0007] The input feature data is obtained by summing the feature vectors corresponding to the satellite feature data, the feature vectors corresponding to the target land use type data in the sample set, the feature vectors corresponding to the target digital terrain data in the sample set, and the absolute position encoding feature vectors.

[0008] An axial attention module is constructed based on a two-layer Transformer model encoder layer, and a global attention module is constructed based on the single-layer encoder layer. The axial attention module is used to perform attention calculation and feature mining on the input feature data in the latitude and longitude directions respectively through the two encoder layers to obtain the module output data. The global attention module is used to perform attention calculation and feature mining on the module output data in a two-dimensional region.

[0009] An initial radar combined reflectivity inversion model is generated by combining several axial attention modules, several global attention modules, and two feature change modules.

[0010] The time-series radar combined reflectivity data is discretized, and the echoes at each point are classified according to the discretization results. The difference between the echo intensity of each point and the minimum value of its category is used as offset data. The initial radar combined reflectivity inversion model is trained using the echo category data, the offset data, and the input feature data to obtain the final radar combined reflectivity inversion model.

[0011] In one possible implementation, the sample set obtained based on the collected digital terrain data, land use type data, satellite data, and radar combined reflectivity data of the target area includes:

[0012] High-resolution digital terrain data, land use type data, satellite data, and radar combined reflectivity data of the target area are collected to obtain the first data; wherein, the high-resolution digital terrain data is digital terrain data with a resolution higher than a specified resolution;

[0013] The first data is transformed into the second data in the WGS84 coordinate system using a projection transformation algorithm, and the resolution of the second data is uniformly transformed to the specified grid resolution using an interpolation algorithm to obtain the third data.

[0014] The third data is filtered based on the combined reflectivity data from the weather radar to obtain a sample set.

[0015] In one possible implementation, the target satellite data and target radar combined reflectivity data based on the sample set are stitched together at multiple time intervals along the channel dimension to obtain satellite feature data and time-series radar combined reflectivity data, including:

[0016] Based on the satellite data in the sample set, the target satellite data within a specified first time period at a specified time are spliced ​​together in the channel dimension to form satellite feature data;

[0017] Based on the radar combined reflectivity data in the sample set, the target radar combined reflectivity data within a consecutive second specified time period at the specified time are stitched together in the channel dimension to form time-series radar combined reflectivity data; the first specified time period is longer than the second specified time period.

[0018] In one possible implementation, the step of summing the feature vectors corresponding to the satellite feature data, the feature vectors corresponding to the target land use type data in the sample set, the feature vectors corresponding to the target digital terrain data in the sample set, and the absolute position encoded feature vectors to obtain the input feature data includes:

[0019] The satellite feature data is converted into a representation of satellite feature vectors;

[0020] The input feature data is obtained by summing the satellite feature vector, the feature vector corresponding to the target land use type data in the sample set, the feature vector corresponding to the target digital terrain data in the sample set, and the absolute location encoded feature vector; wherein, the absolute location encoded feature vector is used to provide location information for the model and to convey the implicit feature information of different geographical locations to the model.

[0021] In one possible implementation, the axial attention module includes a first encoder layer and a second encoder layer. The first encoder layer is used to perform attention calculation and feature mining on the input feature data in the longitude direction to generate intermediate variables. The second encoder layer is used to perform attention calculation and feature mining on the intermediate variables in the latitude direction to generate the final output data of the module.

[0022] The global attention module is used to convert the two-dimensional grid data of the final output data of the module into one-dimensional sequence data, perform attention calculation and feature mining on the one-dimensional sequence data to obtain the feature-transformed and mined one-dimensional sequence data, and restore the feature-transformed and mined one-dimensional sequence data to the dimension shape corresponding to the final output data of the module.

[0023] In one possible implementation, the discretization of the time-series radar combined reflectivity data is performed, and the echo at each point is classified based on the discretization result. The difference between the echo intensity at each point and the minimum value of its category is used as offset data, including:

[0024] The time-series radar combined reflectivity data is discretized to obtain the discretization result, and the echo of each point is classified based on the discretization result to obtain interval and category information.

[0025] Based on the interval and category information, the temporal radar combined reflectivity data is subtracted from the minimum value of the corresponding category of the radar data to obtain the offset data.

[0026] In one possible implementation, training the initial radar combined reflectivity inversion model using echo category data, the offset data, and the input feature data to obtain the final radar combined reflectivity inversion model includes:

[0027] The input feature data is input into the initial radar combined reflectivity inversion model, and the input feature data is then passed through several axial attention modules and input into several global attention modules to generate the final feature data.

[0028] The final feature data is input into the two feature transformation modules respectively to generate initial prediction data and offset values, and the final model prediction data is obtained based on the initial prediction data and the offset values; wherein, the initial prediction data is the prediction data of echo type in the target area, and the echo type is the type of each data point; the offset value represents the offset from the minimum value of the type it represents;

[0029] Based on the final model prediction data and the radar data corresponding to the categories, the classification loss function is calculated using cross-entropy loss.

[0030] The regression loss function is calculated using the mean squared error based on the offset values ​​in the offset data and the final model prediction data.

[0031] The final loss function is determined based on the classification loss function, the regression loss function, and the adjustment factors for the regression loss and classification loss;

[0032] The parameters of the radar combined reflectivity inversion model are adjusted and iterated using the final loss function to obtain the final trained radar combined reflectivity inversion model.

[0033] Secondly, this application provides a radar reflectivity inversion device based on meteorological satellite data, comprising:

[0034] The acquisition module is used to obtain a sample set based on the acquired digital terrain data, land use type data, satellite data, and radar combined reflectivity data of the target area;

[0035] The stitching module is used to stitch together multi-time data in the channel dimension based on the target satellite data and target radar combined reflectivity data in the sample set to obtain satellite feature data and time-series radar combined reflectivity data.

[0036] The summation module is used to sum the feature vectors corresponding to the satellite feature data, the feature vectors corresponding to the target land use type data in the sample set, the feature vectors corresponding to the target digital terrain data in the sample set, and the absolute position encoding feature vectors to obtain the input feature data.

[0037] The feature mining module is used to construct an axial attention module based on a two-layer Transformer model encoder layer and a global attention module based on a single-layer encoder layer. The axial attention module is used to perform attention calculation and feature mining on the input feature data in the latitude and longitude directions through the two encoder layers to obtain the module output data. The global attention module is used to perform attention calculation and feature mining on the module output data in a two-dimensional region.

[0038] The generation module is used to generate an initial radar combined reflectivity inversion model by combining several axial attention modules, several global attention modules and two feature change modules.

[0039] The training module is used to discretize the time-series radar combined reflectivity data, classify the echoes of each point according to the discretization results, use the difference between the echo intensity of each point and the minimum value of its category as offset data, and train the initial radar combined reflectivity inversion model using the echo category data, the offset data and the input feature data to obtain the final radar combined reflectivity inversion model.

[0040] Thirdly, this application also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program that can run on the processor, and the processor executes the computer program to implement the method described in the first aspect above.

[0041] Fourthly, this application also provides a computer-readable storage medium storing computer-executable instructions that, when invoked and executed by a processor, cause the processor to perform the method described in the first aspect above.

[0042] This application brings the following beneficial effects:

[0043] This application provides a radar reflectivity inversion method and apparatus based on meteorological satellite data. It can obtain a sample set based on collected digital terrain data, land use type data, satellite data, and combined radar reflectivity data of a target area. Based on the target satellite data and target combined radar reflectivity data in the sample set, multi-time data are concatenated along the channel dimension to obtain satellite feature data and time-series combined radar reflectivity data. The feature vectors corresponding to the satellite feature data, the target land use type data, and the target digital terrain data in the sample set, along with the absolute position encoded feature vector, are summed to obtain input feature data. An axial attention module is constructed based on a two-layer Transformer model encoder layer, and a global attention module is constructed based on the single-layer encoder layer. The axial attention module is used to perform attention calculation and feature mining on the input feature data in the latitude and longitude directions respectively through the two encoder layers to obtain module output data. The global attention module is used to perform attention calculation and feature mining on the input feature data in the latitude and longitude directions respectively. Attention calculation and feature mining are performed on the output data of the module in the dimensional region; an initial radar combined reflectivity inversion model is generated by combining several axial attention modules, several global attention modules, and two feature change modules; the time-series radar combined reflectivity data is discretized, and the echoes at each point are classified according to the discretization results. The difference between the echo intensity of each point and the minimum value of its category is used as offset data. The initial radar combined reflectivity inversion model is trained using echo category data, the offset data, and the input feature data to obtain the final radar combined reflectivity inversion model. In this scheme, by adopting the method of separately inverting the echo category and the offset of the inverted echo distance from the minimum value of that category, and finally forming the inverted echo prediction value, the problem of the inability to effectively invert large-value echoes due to the fuzzy prediction results caused by regression loss can be effectively solved, thereby improving the accuracy of radar reflectivity inversion results and solving the technical problem of low accuracy of radar reflectivity inversion results.

[0044] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the specific embodiments of this application or the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0046] Figure 1 A flowchart illustrating the radar reflectivity inversion method based on meteorological satellite data provided in this application embodiment;

[0047] Figure 2 Another schematic diagram of the radar reflectivity inversion method based on meteorological satellite data provided in the embodiments of this application;

[0048] Figure 3 Another schematic diagram of the radar reflectivity inversion method based on meteorological satellite data provided in the embodiments of this application;

[0049] Figure 4 A schematic diagram of a radar reflectivity inversion device based on meteorological satellite data provided in this application embodiment;

[0050] Figure 5 A schematic diagram of the structure of an electronic device provided in an embodiment of this application is shown. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0052] The terms "comprising" and "having," and any variations thereof, used in the embodiments of this application, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the steps or units listed, but may optionally include other steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0053] Currently, existing radar reflectivity inversion methods suffer from the following problems: Traditional methods require parameter adjustments based on echo characteristics of different regions and environments, resulting in a large workload and complexity; moreover, common deep learning methods employ MAE or MSE loss, which can easily lead to blurred and smoothed inversion echo results, resulting in lower echo values ​​and affecting accuracy and visual appeal; furthermore, radar echo coverage is incomplete and noise is present; additionally, common methods only use land use type and topographic data to represent the influence of geographical location, without considering the impact of information implicit in geographical location on echo type; in addition, directly inverting the radar echo data of the most recent half hour using satellite observation data from the most recent hour can solve both the time alignment problem between radar echo data and satellite observation data and the problem of lack of continuity and inconsistency in radar echo estimation.

[0054] Based on this, this application provides a radar reflectivity inversion method and apparatus based on meteorological satellite data, which can solve the technical problem of low accuracy of radar reflectivity inversion results.

[0055] The embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0056] Figure 1 This is a flowchart illustrating a radar reflectivity inversion method based on meteorological satellite data, provided as an embodiment of this application. Figure 1 As shown, the method includes:

[0057] Step S110: A sample set is obtained based on the collected digital terrain data, land use type data, satellite data, and radar combined reflectivity data of the target area.

[0058] As one possible implementation, the above-mentioned sample set obtained based on the collected digital terrain data, land use type data, satellite data, and radar combined reflectivity data of the target area may specifically include the following steps:

[0059] High-resolution digital terrain data, land use type data, satellite data, and radar combined reflectivity data of the target area are collected to obtain the first data; wherein, the high-resolution digital terrain data is digital terrain data with a resolution higher than the specified resolution;

[0060] The first data is transformed into the second data in the WGS84 coordinate system using a projection transformation algorithm, and the resolution of all the second data is uniformly transformed to the specified grid resolution using an interpolation algorithm to obtain the third data. The third data is then filtered based on the combined reflectivity data of the weather radar to obtain the sample set.

[0061] For example, such as Figure 2As shown, firstly, high-resolution digital terrain data, land use type data, Himawari-8 and Himawari-9 satellite data, and combined weather radar reflectivity data of the target area are collected. A projection transformation algorithm is then used to convert all data to the WGS84 coordinate system, and an interpolation algorithm is used to unify the resolution of all data to a grid resolution of 0.01° × 0.01°. Next, the data is filtered based on the combined weather radar reflectivity data. Specifically, if no echo greater than 15 dBZ is observed within a continuous hour, the samples for that period are marked as invalid. After processing all samples, the invalid samples are filtered out, forming a sample set. The sample set is sorted chronologically, and the first 70% of the data is used as the training set, while the last 30% is used as the test set. This sample partitioning method, compared to random sampling, avoids the information leakage problem caused by similar features of samples from similar time periods, i.e., the test set cannot effectively evaluate the model's generalization ability.

[0062] Step S120: Based on the target satellite data and target radar combined reflectivity data in the sample set, the multi-time data are stitched together in the channel dimension to obtain satellite feature data and time-series radar combined reflectivity data.

[0063] In one possible implementation, the target satellite data and target radar combined reflectivity data based on the sample set are stitched together at multiple time intervals along the channel dimension to obtain satellite feature data and time-series radar combined reflectivity data. Specifically, this may include the following steps:

[0064] Satellite feature data is formed by stitching together target satellite data within a specified first time period at a specified time in the channel dimension based on satellite data in the sample set; time-series radar combined reflectivity data is formed by stitching together target radar combined reflectivity data within a specified second time period at a specified time in the channel dimension based on radar combined reflectivity data in the sample set; the first specified time period is longer than the second specified time period.

[0065] For example, let a certain time be T, and let T have a possible value of {0, 30} in the minute place. This selection takes into account two points: first, radar data and satellite data are aligned in time; second, data from similar times have similar characteristics, and a time step that is too fine has little effect on the model's training performance. Let high-resolution digital terrain data be... Land use type data Satellite data is recorded as... Satellite data for one consecutive hour at a given moment { , , , , , Satellite feature data is formed by stitching together data along the channel dimension. The combined reflectivity data of the weather radar is denoted as Rad. The combined radar reflectivity data for a continuous half-hour period at a given moment { , , , , The time-series radar composite reflectivity data is stitched together along the channel dimension. .in and This indicates the size of the study area, and P represents the number of satellite data channels. This represents the total number of channels after stitching together one hour of satellite data along the channel dimension. This indicates the number of time steps for half an hour of combined radar reflectivity data.

[0066] By using one hour of continuous satellite data as model input and half an hour of continuous radar echo data as model learning target, this approach allows the model to acquire more trend and implicit information. It also enables the model to complete the alignment task of two types of data with different time resolutions, which is more advantageous than algorithms that directly use rules for data time alignment.

[0067] Step S130: Summing the feature vectors corresponding to satellite feature data, the feature vectors corresponding to target land use type data in the sample set, the feature vectors corresponding to target digital terrain data in the sample set, and the absolute position encoding feature vectors to obtain the input feature data.

[0068] As an optional implementation, the above-mentioned summation of the feature vectors corresponding to satellite feature data, the feature vectors corresponding to target land use type data in the sample set, the feature vectors corresponding to target digital terrain data in the sample set, and the absolute position encoded feature vectors to obtain the input feature data may specifically include the following steps:

[0069] The satellite feature data is converted into a satellite feature vector representation. The satellite feature vector, the feature vector corresponding to the target land use type data in the sample set, the feature vector corresponding to the target digital terrain data in the sample set, and the absolute position encoded feature vector are summed to obtain the input feature data. Among them, the absolute position encoded feature vector is used to provide location information for the model and to convey the implicit feature information of different geographical locations to the model.

[0070] like Figure 2 As shown, land use type data After passing through the Embedding layer, the land type category data is converted into an embedded vector representation, denoted as... High-resolution digital terrain data The data is evenly divided into 512 parts based on the maximum and minimum values, transforming it into categorical data with 512 classes. After passing through an embedding layer, the digital terrain data is converted into an embedded vector representation. Design the FeaNet satellite feature extraction network to extract satellite data. Convert to satellite feature vectors, denoted as .remember Encoding feature vectors for absolute locations provides location information to the model and conveys implicit feature information from different geographical locations. This refers to the feature dimension size of the model's latent variables. Converting land use type data into vector representation allows the model to better understand and utilize category data. The above transformation converts the feature dimensions of digital elevation data, land use type data, and satellite data to a size consistent with the feature dimensions of the model's latent variables.

[0071] Then, satellite feature data Digital terrain feature data Land use type characteristic data Absolute position coding features Summing yields the input feature data .Right now = + + + .

[0072] The embedding layer transforms land use types and discretized digital terrain data into embedded vector representations. This approach allows for more efficient use of categorical data in Transformer-like models, while also enabling the model to find feature representations more suitable for the learning task and uncover more hidden feature information from both types of data.

[0073] In this embodiment, as described above, one hour of continuous satellite data is used as input and half an hour of continuous radar data is used as output. This provides the model with trend change information and more hidden information, effectively solving the data time alignment problem caused by the inconsistency in the time resolution of radar and satellite data. Simultaneously, inverting multiple time-series radar data at once can, to some extent, solve the problem of inconsistent echoes between different time periods during radar echo estimation and inversion. By simultaneously inverting multiple time-series outputs from multiple time-series inputs, reasonable data alignment across different time resolutions can be achieved.

[0074] Step S140: Construct an axial attention module based on a two-layer Transformer model encoder layer, and construct a global attention module based on a single-layer encoder layer.

[0075] The axial attention module is used to perform attention calculation and feature mining on the input feature data in the latitude and longitude directions through two encoder layers to obtain the module output data; the global attention module is used to perform attention calculation and feature mining on the module output data in a two-dimensional region.

[0076] In one optional implementation, the axial attention module includes a first encoder layer and a second encoder layer. The first encoder layer is used to perform attention calculation and feature mining on the input feature data in the longitude direction to generate intermediate variables. The second encoder layer is used to perform attention calculation and feature mining on the intermediate variables in the latitude direction to generate the final output data of the module. The global attention module is used to convert the two-dimensional grid data of the final output data of the module into one-dimensional sequence data, perform attention calculation and feature mining on the one-dimensional sequence data to obtain the feature-transformed and mined one-dimensional sequence data, and restore the feature-transformed and mined one-dimensional sequence data to the dimension shape corresponding to the final output data of the module.

[0077] For example, designing an axial attention module is denoted as... Specifically, it consists of two Transformer Encoder Layers, denoted as... and .in, This module is responsible for performing attention calculations and feature mining on the input features in the precision direction. The sequence length used for attention calculation is [length missing]. . This module is responsible for performing attention calculations and feature mining on the input features along the dimensional direction. The sequence length used for attention calculation is [length missing]. Input features are processed After module processing, intermediate variables are generated, and then the intermediate variables are processed... The module processes the data to form the final output. ,Right now = Then, design the global attention module, denoted as... G, specifically, consists of a single-layer Transformer Encoder Layer. The module is in... Attention calculation and feature mining are performed on a two-dimensional region, that is, the two-dimensional grid data is unfolded into one-dimensional sequence data, and after feature transformation and mining, the data is restored to the original dimensional shape of the input. .

[0078] Step S150: Based on the combination of several axial attention modules, several global attention modules and two feature change modules, an initial radar combined reflectivity inversion model is generated.

[0079] For example, the design of a radar combined reflectivity inversion model is denoted as... Its cause indivual Module, indivual The G module, and two feature change modules and Composition. Input features go through indivual Input after module indivual The G module ultimately generates feature data. .Will Enter them separately and Module, generation and .in For E time intervals, the echo type is predicted for each point within the study area. A value greater than 0 represents the distance to each point within the study area over E time periods. The value represents the offset of the minimum value of each point type within the study area across E time periods. The final model prediction is:

[0080] .

[0081] In this embodiment, shallow features are extracted and mined using an axial attention module with relatively low computational cost, while deep features are mined using a global attention module with strong feature mining capabilities. Since low-level features focus more on mining local features, while deep-level features focus more on analyzing and utilizing global information, this combination can significantly reduce the computational cost of the model while having little impact on the model's inversion performance.

[0082] Step S160: Discretize the time-series radar combined reflectivity data, classify the echoes of each point based on the discretization results, use the difference between the echo intensity of each point and the minimum value of its category as offset data, and train the initial radar combined reflectivity inversion model using echo category data, offset data and input feature data to obtain the final radar combined reflectivity inversion model.

[0083] like Figure 3As shown, to construct a labeling system suitable for supervised learning, radar data from multiple time points are first concatenated along the channel dimension to form a unified radar feature vector. Then, each data point is discretized and grouped according to its echo intensity value ([0, 20), [20, 30), [30, 40), [40, 50), [50, +∞)), assigning a unique class label to each data point. Based on this, the difference between each data point and the minimum echo intensity value within its class is calculated, serving as the offset label for that point. Finally, the class label and offset label together constitute a dual-task supervision signal, guiding the model to simultaneously learn the classification structure of echo intensity and the regression characteristics of relative intensity within each class, thereby improving the modeling ability for the spatiotemporal evolution characteristics of radar echoes.

[0084] As an optional implementation, the above-mentioned discretization processing based on time-series radar combined reflectivity data, the classification of the echo at each point based on the discretization result, and the difference between the echo intensity at each point and the minimum value of its category are used as offset data. Specifically, this may include the following steps:

[0085] The time-series radar combined reflectivity data is discretized to obtain the discretization result. Based on the discretization result, the echo of each point is classified to obtain interval and category information. Based on the interval and category information, the time-series radar combined reflectivity data is subtracted from the minimum value of the corresponding category of the radar data to obtain the offset data.

[0086] For example, combining time-series radar reflectivity data Discretize the data, assigning each point to [0, 20), [20, 30), [30, 40), [40, 50), [50, The data is divided into five categories, forming radar echo category label data, denoted as... .remember An array formed for each type of minimum value. The radar echo offset label data is obtained by subtracting the minimum value of its corresponding class:

[0087] .

[0088] In one alternative implementation, such as Figure 3 As shown, the above-mentioned training of the initial radar combined reflectivity inversion model using echo category data, offset data, and input feature data to obtain the final radar combined reflectivity inversion model can specifically include the following steps:

[0089] The input feature data is input into the initial radar combined reflectivity inversion model. The input feature data is then passed through several axial attention modules and input into several global attention modules to generate the final feature data.

[0090] The final feature data is input into two feature transformation modules to generate initial prediction data and offset values. The final model prediction data is obtained based on the initial prediction data and offset values. The initial prediction data is the prediction data of echo type within the target area, and the echo type is the type of each data point. The offset value represents the offset from the minimum value of the type it represents.

[0091] Based on the corresponding categories of the final model prediction data and radar data, the cross-entropy loss is used to calculate the classification loss function; based on the offset data and the offset values ​​in the final model prediction data, the mean squared error is used to calculate the regression loss function.

[0092] The final loss function is determined based on the classification loss function, the regression loss function, and the adjustment factors for the regression loss and the classification loss. The parameters of the radar combined reflectivity inversion model are adjusted and iterated using the final loss function to obtain the final radar combined reflectivity inversion model after training.

[0093] like Figure 3 As shown, the classification loss is calculated using cross-entropy loss. Calculate regression loss using mean squared error The final loss was... + .in This serves as an adjustment factor for both regression and classification losses. This loss function design effectively avoids the problem of overly smooth echo inversion results, which fail to reproduce large-value echoes. In this embodiment, model training is performed according to the above process.

[0094] By employing the method of separately inverting the echo category and the offset of the inverted echo distance from the minimum value of that category, and finally forming the predicted value of the inverted echo, the problem of the inability to effectively invert large-value echoes due to the ambiguity of the prediction results caused by regression loss can be effectively solved, thus improving the accuracy of radar reflectivity inversion results.

[0095] Figure 4 A schematic diagram of a radar reflectivity inversion device based on meteorological satellite data is provided. Figure 4 As shown, the radar reflectivity inversion device 400 based on meteorological satellite data includes:

[0096] The acquisition module 401 is used to obtain a sample set based on the acquired digital terrain data, land use type data, satellite data and radar combined reflectivity data of the target area;

[0097] The stitching module 402 is used to stitch together multi-time data in the channel dimension based on the target satellite data and target radar combined reflectivity data in the sample set to obtain satellite feature data and time-series radar combined reflectivity data.

[0098] The summation module 403 is used to sum the feature vector corresponding to the satellite feature data, the feature vector corresponding to the target land use type data in the sample set, the feature vector corresponding to the target digital terrain data in the sample set, and the absolute position encoding feature vector to obtain the input feature data.

[0099] The feature mining module 404 is used to construct an axial attention module based on a two-layer Transformer model encoder layer and a global attention module based on a single-layer encoder layer. The axial attention module is used to perform attention calculation and feature mining on the input feature data in the latitude and longitude directions through the two encoder layers to obtain module output data. The global attention module is used to perform attention calculation and feature mining on the module output data in a two-dimensional region.

[0100] The generation module 405 is used to generate an initial radar combined reflectivity inversion model by combining several axial attention modules, several global attention modules and two feature change modules.

[0101] The training module 406 is used to discretize the time-series radar combined reflectivity data, classify the echoes of each point according to the discretization results, use the difference between the echo intensity of each point and the minimum value of its category as offset data, and train the initial radar combined reflectivity inversion model using the echo category data, the offset data and the input feature data to obtain the final radar combined reflectivity inversion model.

[0102] The radar reflectivity inversion device based on meteorological satellite data provided in this application has the same technical features as the radar reflectivity inversion method based on meteorological satellite data provided in the above embodiments, so it can also solve the same technical problems and achieve the same technical effects.

[0103] An electronic device provided in this application embodiment, such as Figure 5 As shown, the electronic device 500 includes a processor 502 and a memory 501. The memory stores a computer program that can run on the processor. When the processor executes the computer program, it implements the steps of the method provided in the above embodiments.

[0104] See Figure 5The electronic device also includes a bus 503 and a communication interface 504. The processor 502, the communication interface 504 and the memory 501 are connected through the bus 503. The processor 502 is used to execute executable modules, such as computer programs, stored in the memory 501.

[0105] The memory 501 may include high-speed random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Communication between this system network element and at least one other network element is achieved through at least one communication interface 504 (which can be wired or wireless), such as the Internet, wide area network, local area network, or metropolitan area network.

[0106] Bus 503 can be an ISA bus, PCI bus, or EISA bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 5 The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.

[0107] The memory 501 is used to store programs. After receiving an execution instruction, the processor 502 executes the program. The method executed by the apparatus defined by the process disclosed in any of the preceding embodiments of this application can be applied to the processor 502 or implemented by the processor 502.

[0108] Processor 502 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of processor 502 or by instructions in software form. The processor 502 can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this application can be directly manifested as execution by a hardware decoding processor, or execution by a combination of hardware and software modules in the decoding processor. The software module can reside in a mature storage medium in the field, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory 501, and processor 502 reads the information from memory 501 and, in conjunction with its hardware, completes the steps of the above method.

[0109] Corresponding to the above-described radar reflectivity inversion method based on meteorological satellite data, this application embodiment also provides a computer-readable storage medium storing computer-executable instructions. When the computer-executable instructions are called and run by a processor, the computer-executable instructions cause the processor to perform the steps of the above-described radar reflectivity inversion method based on meteorological satellite data.

[0110] The radar reflectivity inversion device based on meteorological satellite data provided in this application embodiment can be specific hardware on the device or software or firmware installed on the device. The implementation principle and technical effects of the device provided in this application embodiment are the same as those in the foregoing method embodiments. For the sake of brevity, any parts not mentioned in the device embodiment can be referred to the corresponding content in the foregoing method embodiments. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can all be referred to the corresponding processes in the above method embodiments, and will not be repeated here.

[0111] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the displayed or discussed mutual couplings, direct couplings, or communication connections may be through some communication interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0112] For example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0113] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0114] In addition, the functional units in the embodiments provided in this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0115] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the radar reflectivity inversion method based on meteorological satellite data described in various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0116] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. In addition, the terms "first", "second", "third", etc. are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0117] Finally, it should be noted that the above-described embodiments are merely specific implementations of this application, used to illustrate the technical solutions of this application, and not to limit them. The protection scope of this application is not limited thereto. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the scope of the technology disclosed in this application; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application. All should be covered within the protection scope of this application. Therefore, the protection scope of this application should be determined by the protection scope of the claims.

Claims

1. A method for retrieving radar reflectivity based on meteorological satellite data, characterized in that, The method comprises: obtaining a sample set based on collected digital terrain data, land use type data, satellite data and radar combined reflectivity data of a target region; splicing multi-time data in the channel dimension based on target satellite data and target radar combined reflectivity data in the sample set to obtain satellite feature data and time-series radar combined reflectivity data; summing corresponding feature vectors of the satellite feature data, corresponding feature vectors of target land use type data in the sample set, corresponding feature vectors of target digital terrain data in the sample set and absolute position encoding feature vectors to obtain input feature data; constructing an axial attention module based on two layers of encoder layers of a Transformer model, and constructing a global attention module based on a single layer of the encoder layers; the axial attention module is used for attention calculation and feature mining of the input feature data in the latitude and longitude directions respectively by two layers of the encoder to obtain module output data; the global attention module is used for attention calculation and feature mining of the module output data in a two-dimensional region; combining a plurality of the axial attention modules, a plurality of the global attention modules and two feature change modules to generate an initial radar combined reflectivity inversion model; performing discretization processing based on the time-series radar combined reflectivity data, classifying the echoes of each point based on the discretization processing result, taking the difference between the echo intensity of each point and the minimum value of the category to which the point belongs as offset data, and training the initial radar combined reflectivity inversion model by using the echo category data, the offset data and the input feature data to obtain a final radar combined reflectivity inversion model.

2. The method of claim 1, wherein, The method comprises: obtaining a sample set based on collected digital terrain data, land use type data, satellite data and radar combined reflectivity data of a target region; obtaining first data by collecting high-resolution digital terrain data, land use type data, satellite data and radar combined reflectivity data of a target region; wherein the high-resolution digital terrain data is digital terrain data higher than a specified resolution; converting all the first data into second data in the WGS84 coordinate system by using a projection conversion algorithm, and converting the resolution of all the second data to a specified grid resolution by using an interpolation algorithm to obtain third data; 3. The method of claim 1, wherein, filtering the third data according to weather radar combined reflectivity data to obtain a sample set. The method comprises: splicing target satellite data in a specified time period in the channel dimension based on satellite data in the sample set to form satellite feature data; splicing target radar combined reflectivity data in a specified time period in the channel dimension based on radar combined reflectivity data in the sample set to form time-series radar combined reflectivity data; the first specified time period is longer than the second specified time period.

4. The method of claim 1, wherein, The satellite feature data corresponding feature vector, the target land use type data corresponding feature vector in the sample set, the target digital terrain data corresponding feature vector in the sample set and the absolute position encoding feature vector are summed to obtain input feature data, including: Converting the satellite feature data into a satellite feature vector representation; The satellite feature vector, the target land use type data corresponding feature vector in the sample set, the target digital terrain data corresponding feature vector in the sample set and the absolute position encoding feature vector are summed to obtain input feature data; wherein the absolute position encoding feature vector is used to provide position information for the model and convey implicit feature information of different geographic locations for the model.

5. The method of claim 1, wherein, The axial attention module includes a first encoder layer and a second encoder layer, the first encoder layer is used for attention calculation and feature mining of the input feature data in the longitude direction to generate an intermediate variable; the second encoder layer is used for attention calculation and feature mining of the intermediate variable in the latitude direction to generate module final output data; The global attention module is used to convert the two-dimensional grid data of the module final output data into one-dimensional sequence data, perform attention calculation and feature mining on the one-dimensional sequence data to obtain one-dimensional sequence data after feature transformation and mining, and restore the one-dimensional sequence data after feature transformation and mining to the dimension shape corresponding to the module final output data.

6. The method of claim 1, wherein, The time series radar combined reflectivity data is discretized, the echoes of each point are classified based on the discretization result, the difference between the echo intensity of each point and the minimum value of the corresponding category is taken as the offset data, including: The time series radar combined reflectivity data is discretized to obtain a discretization result, and each point's echo is classified based on the discretization result to obtain interval and category information; The time series radar combined reflectivity data is subtracted from the minimum value of the corresponding category of the radar data based on the interval and category information to obtain offset data.

7. The method of claim 6, wherein, The initial radar combined reflectivity inversion model is trained using the echo category data, the offset data and the input feature data to obtain a final radar combined reflectivity inversion model, including: The input feature data is input into the initial radar combined reflectivity inversion model, and the input feature data is input into several global attention modules after passing through several axial attention modules to generate final feature data; The final feature data is input into the two feature change modules respectively to generate initial prediction data and offset value, and the final model prediction data is obtained based on the initial prediction data and the offset value; wherein the initial prediction data is the prediction data of the echo type in the target area, and the echo type is the type of each data point; the offset value represents the offset from the minimum value of the represented type; Based on the final model prediction data and the radar data corresponding to the category data, a cross-entropy loss is calculated to obtain a classification loss function; Based on the offset data and the offset value in the final model prediction data, a mean square error is calculated to obtain a regression loss function; A final loss function is determined according to the classification loss function, the regression loss function, and the adjustment factor of regression loss and classification loss; The parameters of the radar combined reflectivity retrieval model are adjusted and iterated using the final loss function to obtain a trained final radar combined reflectivity retrieval model.

8. A radar reflectivity retrieval apparatus based on meteorological satellite data, characterized by, It comprises: The acquisition module is used to obtain a sample set based on the collected digital terrain data, land use type data, satellite data and radar combined reflectivity data of the target area; The splicing module is used to splice the multi-time data in the channel dimension based on the target satellite data and the target radar combined reflectivity data in the sample set to obtain satellite feature data and time series radar combined reflectivity data; The summation module is used to sum the corresponding feature vectors of the satellite feature data, the target land use type data in the sample set, the target digital terrain data in the sample set, and the absolute position encoding feature vector to obtain input feature data; The feature mining module is used to construct an axial attention module based on two layers of encoder layers of a Transformer model, and construct a global attention module based on a single layer of the encoder layer; The axial attention module is used to perform attention calculation and feature mining on the input feature data in the latitude and longitude directions respectively through two layers of the encoder to obtain module output data; The global attention module is used to perform attention calculation and feature mining on the module output data in a two-dimensional region; The generation module is used to combine several axial attention modules, several global attention modules and two feature change modules to generate an initial radar combined reflectivity retrieval model; The training module is used to perform discretization processing based on the time series radar combined reflectivity data, classify the echoes of each point based on the discretization processing result, take the difference between the echo intensity of each point and the minimum value of the category to which it belongs as the offset data, and train the initial radar combined reflectivity retrieval model based on the echo category data, the offset data and the input feature data to obtain a final radar combined reflectivity retrieval model.

9. An electronic device comprising a memory, a processor, the memory having stored therein a computer program executable on the processor, characterized in that, The processor executes the computer program to realize the steps of the method of any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores computer executable instructions, and when the computer executable instructions are called and run by the processor, the computer executable instructions cause the processor to run the method of any one of claims 1 to 7.

Images