A meteorological data processing method and device, electronic equipment and storage medium

Abstract

The present disclosure provides a meteorological data processing method and device, electronic equipment and storage medium, a fusion feature map is constructed by fusing multi-source data features; low-level fusion features corresponding to the fusion feature map are extracted, down-sampling is performed on the fusion feature map, and then high-level fusion features corresponding to the fusion feature map are further extracted while maintaining the spatial resolution; the fusion feature map after extracting the high-level fusion features is up-sampled to restore the spatial resolution, and the fusion feature map after up-sampling is spliced with the fusion feature map after down-sampling to determine a spliced feature map; convolution operation is performed on the spliced feature map, and the spliced feature map is converted into small-scale meteorological data after fusing the low-level fusion features and the high-level fusion features. The down-scaling processing of meteorological data can be realized by using less computing resources, and the meteorological data from different data sources is effectively fused, overcoming the limitation problem in processing nonlinear and complex meteorological relationships in the prior art.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of meteorological data analysis, and in particular, to a meteorological data processing method and device, an electronic device, and a storage medium. BACKGROUND

[0002] The acquisition and processing of meteorological data play a crucial role in environmental science, meteorology, and climate change research. Meteorological models typically provide global or regional-scale meteorological information at large scales and low resolutions. These models can capture macro meteorological patterns and trends but lack in local details. To meet specific application needs, such as agricultural planning, urban management, and extreme weather warning, it is necessary to convert these large-scale, low-resolution data into small-scale, high-resolution data, i.e., downscaling processing.

[0003] Traditional downscaling methods can be divided into two categories: dynamic downscaling and statistical downscaling. Dynamic downscaling uses regional climate models (RCMs) to simulate higher resolution based on global climate models (GCMs). This method is computationally expensive and requires high computing resources. Statistical downscaling establishes statistical relationships between meteorological variables using historical observation data and applies these relationships to future meteorological scenarios. This method relies on high-quality historical data and performs poorly when facing nonlinear and complex meteorological relationships. Multi-source meteorological data (such as temperature, precipitation, humidity, wind speed, etc.) provide rich information that can be used to improve the effectiveness of downscaling through data fusion and machine learning methods. However, how to effectively fuse and process these multi-source data remains a problem to be solved. SUMMARY

[0004] The embodiments of the present disclosure provide at least a meteorological data processing method and device, an electronic device, and a storage medium, which have the beneficial effect of achieving downscaling processing of meteorological data with fewer computing resources, effectively fusing meteorological data from different data sources, and overcoming the limitations of processing nonlinear and complex meteorological relationships in the prior art.

[0005] The embodiments of the present disclosure provide a meteorological data processing method, which comprises:

[0006] Obtaining a plurality of types of large-scale meteorological data, extracting features corresponding to each type of large-scale meteorological data to construct a preliminary feature map, and constructing a fusion feature map according to the feature dependency relationship between the preliminary feature maps;

[0007] Extracting low-level fusion features corresponding to the fusion feature map, downsampling the fusion feature map to reduce the spatial resolution of the fusion feature map, and then further extracting high-level fusion features corresponding to the fusion feature map while maintaining the spatial resolution.

[0008] Upsample the fusion feature map after extracting the high-level fusion feature to restore the spatial resolution, and splice the upsampled fusion feature map with the downsampled fusion feature map to determine a spliced feature map;

[0009] Perform a convolution operation on the spliced feature map, and convert the spliced feature map into small-scale meteorological data after fusing the low-level fusion feature and the high-level fusion feature.

[0010] In an optional implementation, the feature dependency relationship is determined based on the following steps:

[0011] Align the preliminary feature maps to the same spatial scale, and adjust the preliminary feature maps to the same number of channels using linear transformation;

[0012] Divide the aligned preliminary feature maps into a plurality of feature blocks;

[0013] Perform linear transformation on each of the feature blocks to determine a corresponding query vector, a keyword vector, and a value vector;

[0014] Determine an attention weight according to the dot product between the query vector and the keyword vector, and use the attention weight as the feature dependency relationship.

[0015] In an optional implementation, the fusion feature map is constructed based on the following steps:

[0016] Input the feature blocks into a multi-layer feature fusion network, and in each layer, for each feature block, weight the value vector corresponding to the feature block and the value vectors corresponding to other feature blocks according to the attention weight to determine a hierarchical fusion feature vector and generate a corresponding hierarchical fusion feature map, wherein the hierarchical fusion feature vector generated by a lower layer is retained and passed to a higher layer through residual connection between layers;

[0017] Splice and fuse the hierarchical fusion feature map corresponding to each layer to generate the fusion feature map.

[0018] In an optional implementation, the low-level fusion feature corresponding to the fusion feature map is extracted, and the fusion feature map is downsampled to reduce the spatial resolution of the fusion feature map, specifically including:

[0019] Input the fusion feature map into an encoder of a preset downscaling model, wherein the encoder includes a convolution layer, a residual layer, and a max-pooling layer;

[0020] The low-level fusion feature is determined by performing a convolution operation on the fusion feature map through the convolution layer, and the low-level fusion feature is transmitted to the residual layer;

[0021] The fusion feature map and the low-level fusion feature are superimposed through a residual connection in the residual layer, and are transmitted to the max-pooling layer;

[0022] The feature map with reduced spatial resolution is stored in the skip connection through downsampling performed by the max-pooling layer.

[0023] In an optional implementation, the high-level fusion feature corresponding to the fusion feature map is extracted while maintaining the spatial resolution, specifically including:

[0024] The downsampled feature map is input to the bottleneck layer of the preset reduction model, and the high-level fusion feature is extracted through convolution operation and residual connection;

[0025] The bottleneck layer does not perform a pooling operation to maintain the spatial resolution.

[0026] In an optional implementation, the fusion feature map after the high-level fusion feature is extracted is upsampled to restore the spatial resolution, and the upsampled fusion feature map and the downsampled fusion feature map are spliced to determine a spliced feature map, specifically including:

[0027] The feature map output by the encoder is input to the decoder of the preset reduction model, where the decoder includes a deconvolution layer and a feature fusion layer;

[0028] The spatial resolution is restored by upsampling the input feature map through the deconvolution layer, and the upsampled feature map is input to the skip connection;

[0029] In the skip connection, the stored feature map with reduced spatial resolution is spliced with the upsampled feature map to determine the spliced feature map.

[0030] In an optional implementation, the low-level fusion feature and the high-level fusion feature are fused, specifically including:

[0031] The spliced feature map is input to the feature fusion layer, and the low-level fusion feature and the high-level fusion feature are fused by performing a convolution operation on the spliced feature map.

[0032] The present disclosure also provides a meteorological data processing apparatus, including:

[0033] The feature fusion module is configured to acquire multiple types of large-scale meteorological data, extract a preliminary feature map corresponding to each type of the large-scale meteorological data, and construct a fused feature map according to a feature dependency relationship between the preliminary feature maps.

[0034] The downscaling module is configured to extract low-level fused features corresponding to the fused feature map, downsample the fused feature map to reduce a spatial resolution of the fused feature map, and then further extract high-level fused features corresponding to the fused feature map while maintaining the spatial resolution.

[0035] The resolution enhancement module is configured to upsample the fused feature map after the high-level fused features are extracted to restore the spatial resolution, splice the upsampled fused feature map and the downsampled fused feature map to determine a spliced feature map.

[0036] The data conversion module is configured to perform a convolution operation on the spliced feature map, and convert the spliced feature map into small-scale meteorological data after the low-level fused features and the high-level fused features are fused.

[0037] The embodiments of the present disclosure further provide an electronic device, including a processor, a memory and a bus, the memory stores machine readable instructions executable by the processor, when the electronic device is running, the processor and the memory communicate through the bus, and the machine readable instructions are executed by the processor to perform the above-mentioned meteorological data processing method or the steps in any possible implementation manner of the above-mentioned meteorological data processing method.

[0038] The embodiments of the present disclosure further provide a computer readable storage medium, the computer readable storage medium stores a computer program, and the computer program is executed by the processor to perform the above-mentioned meteorological data processing method or the steps in any possible implementation manner of the above-mentioned meteorological data processing method.

[0039] The embodiments of the present disclosure further provide a computer program product, including a computer program / instruction, the computer program / instruction is executed by the processor to implement the above-mentioned meteorological data processing method or the steps in any possible implementation manner of the above-mentioned meteorological data processing method.

[0040] The method, device, electronic device and storage medium provided by the embodiments of the present disclosure can realize the downscaling processing of meteorological data by using less computing resources, effectively fuse meteorological data from different data sources, and overcome the limitation problem in processing nonlinear and complex meteorological relationships in the prior art.

[0041] In order to make the above objectives, characteristics and advantages of the present disclosure more apparent, clear and easy to understand, the following will specifically describe the preferred embodiments of the present disclosure, and the accompanying drawings will be described in detail as follows. BRIEF DESCRIPTION OF DRAWINGS

[0042] In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the following will briefly introduce the drawings needed to be used in the embodiments. The drawings herein are incorporated into the specification and form a part of the specification, which illustrate the embodiments consistent with the present disclosure, and are used to explain the technical solutions of the present disclosure together with the specification. It should be understood that the following drawings only show some embodiments of the present disclosure, and therefore should not be regarded as a limitation to the scope, and other related drawings can also be obtained by those skilled in the art without any creative effort on the basis of these drawings.

[0043] Figure 1 A flowchart of a method for processing meteorological data is shown;

[0044] Figure 2 A flowchart of a method for constructing a fusion feature map is shown;

[0045] Figure 3 A schematic diagram of a device for processing meteorological data is shown;

[0046] Figure 4 A schematic diagram of an electronic device is shown. DETAILED DESCRIPTION

[0047] The technical solutions in the embodiments of the present disclosure will be described clearly and completely below with reference to the drawings in the embodiments of the present disclosure. Obviously, the described embodiments are only part of the embodiments of the present disclosure, rather than all the embodiments. The components of the embodiments of the present disclosure described and shown in the drawings can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present disclosure provided in the drawings is not intended to limit the scope of the claimed present disclosure, but only represents selected embodiments of the present disclosure. Based on the embodiments of the present disclosure, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present disclosure.

[0048] It should be noted that similar reference numerals and letters refer to similar items throughout the accompanying drawings, and therefore, once an item is defined in one drawing, it need not be further defined and explained in subsequent drawings.

[0049] The term "and / or" used herein only describes an associated relationship, which means that there can be three relationships, for example, A and / or B can represent three cases of A alone, A and B together, and B alone. In addition, the term "at least one" used herein means any one of a plurality or any combination of at least two of a plurality, for example, including at least one of A, B, and C can mean including any one or more elements selected from the set consisting of A, B, and C.

[0050] It is found through research that traditional downscaling methods can be divided into two categories: dynamic downscaling and statistical downscaling. Dynamic downscaling uses regional climate models (RCMs) to simulate higher resolution on the basis of global climate models (GCMs). This method has high computational cost and requires extremely high computing resources. Statistical downscaling uses historical observation data to establish statistical relationships between meteorological variables and applies these relationships to future meteorological scenarios. This method relies on high-quality historical data and performs poorly when faced with nonlinear and complex meteorological relationships. Multi-source meteorological data (such as temperature, precipitation, humidity, wind speed, etc.) provide rich information, which can improve the effect of downscaling through data fusion and machine learning methods. However, how to effectively fuse and process these multi-source data is still a problem to be solved.

[0051] Based on the above research, the present disclosure provides a meteorological data processing method and device, electronic equipment and storage medium. By obtaining multiple types of large-scale meteorological data, the corresponding features of each large-scale meteorological data are extracted to construct a preliminary feature map, and a fusion feature map is constructed according to the feature dependency relationship between the preliminary feature maps. The low-level fusion features corresponding to the fusion feature map are extracted, and the fusion feature map is down-sampled to reduce the spatial resolution of the fusion feature map. Then, the high-level fusion features corresponding to the fusion feature map are further extracted while maintaining the spatial resolution. The fusion feature map after extracting the high-level fusion features is up-sampled to restore the spatial resolution, and the up-sampled fusion feature map is spliced with the down-sampled fusion feature map to determine a spliced feature map. The convolution operation is performed on the spliced feature map, and after fusing the low-level fusion features and the high-level fusion features, the spliced feature map is converted into small-scale meteorological data. The downscaling processing of meteorological data can be realized with less computing resources, and the meteorological data from different data sources can be effectively fused to overcome the limitations of processing nonlinear and complex meteorological relationships in the prior art.

[0052] To facilitate the understanding of the present embodiment, first, a meteorological data processing method disclosed by the present embodiment is introduced in detail. The execution subject of the meteorological data processing method provided by the present embodiment is generally a computer device with certain computing power, which may, for example, include a terminal device or a server or other processing device. The terminal device may be a user equipment (UE), a mobile device, a user terminal, a terminal, a cellular phone, a cordless phone, a personal digital assistant (PDA), a handheld device, a computing device, a vehicle-mounted device, a wearable device, etc. In some possible implementation manners, the meteorological data processing method can be realized by a processor calling computer readable instructions stored in a memory.

[0053] Referring to Figure 1 The flowchart of the meteorological data processing method provided by the present embodiment is shown in FIG. 1. The method includes steps S101-S104, wherein:

[0054] S101, multiple types of large-scale meteorological data are obtained, the corresponding features of each large-scale meteorological data are extracted to construct a preliminary feature map, and a fusion feature map is constructed according to the feature dependency relationship between the preliminary feature maps.

[0055] In a specific implementation, first, a plurality of types of large-scale meteorological data are acquired from different data sources, and a feature extraction operation is performed for each type of large-scale meteorological data to respectively extract a preliminary feature map containing features corresponding to the type of large-scale meteorological data. Then, a feature fusion operation is performed on the preliminary feature maps corresponding to all types of large-scale meteorological data to fuse meteorological data from different data sources and generate a fused feature map with rich semantics and details.

[0056] Here, the large-scale meteorological data is large-scale and low-resolution meteorological data. Large-scale refers to meteorological data with a wide spatial coverage, such as global or continental level meteorological model output. Low-resolution refers to less spatial detail, usually with each data point covering a larger area, such as each data point representing an area of 100 kilometers or more.

[0057] For example, large-scale meteorological data can include terrain mask, FY4B cloud image, numerical prediction data (such as CMA-MESO 3km pressure), temperature, potential height, radial wind, zonal wind, vertical speed, specific humidity, precipitation, relative humidity, etc.

[0058] It should be noted that each data source provides at least one type of large-scale meteorological data.

[0059] Specifically, preliminary feature extraction can be performed for each type of large-scale meteorological data. The feature extraction process can be implemented using a convolutional neural network (CNN) or a shallow Transformer module to obtain a preliminary feature map corresponding to each type of large-scale meteorological data.

[0060] Further, as a possible implementation of the construction process of the fused feature map, please refer to Figure 2 The figure shows a flowchart of a method for constructing a fused feature map provided by an embodiment of the present disclosure, which includes steps S1011-S1015, wherein:

[0061] S1011, align the preliminary feature maps to the same spatial scale, and use linear transformation to adjust the preliminary feature maps to the same number of channels.

[0062] S1012, divide the aligned preliminary feature maps into a plurality of feature blocks; perform linear transformation on each feature block to determine the corresponding query vector, keyword vector, and value vector.

[0063] S1013, determine the attention weight according to the dot product between the query vector and the keyword vector, and use the attention weight as the feature dependency relationship.

[0064] S1014, input the feature blocks into a multi-layer feature fusion network, in each layer, for each of the feature blocks, the value vector corresponding to the feature block is weighted with the value vectors corresponding to the other feature blocks according to the attention weights, a hierarchical fusion feature vector is determined and a corresponding hierarchical fusion feature map is generated.

[0065] S1015, the hierarchical fusion feature maps corresponding to each layer are spliced and fused to generate the fusion feature map.

[0066] In a specific implementation, the feature fusion process between multi-source meteorological data can be implemented by using a multi-layer feature fusion network, preferably a Swin Transformer model. In the processing process, since the preliminary feature maps corresponding to large-scale meteorological data of different data sources may have different spatial resolutions or perspectives, it is necessary to first align the feature maps to the same spatial scale by interpolation, translation or other alignment methods. At the same time, the preliminary feature maps of different data sources can have different channel numbers, which are adjusted to the same channel number by linear transformation or other techniques.

[0067] Here, the relationship and dependence between the large-scale meteorological data corresponding to different data sources can be captured by referring to the multi-head self-attention mechanism, to establish a connection between different feature maps and weight them according to their relevance.

[0068] Specifically, first, the aligned preliminary feature map is divided into fixed-size feature blocks, which are used as inputs to the multi-layer feature fusion network. Then, at each level, the multi-head self-attention mechanism is used to linearly transform each feature block to obtain query vectors, key vectors and value vectors. The attention weights are obtained by calculating the dot product of the query vectors and the key vectors, and the corresponding value vectors are weighted by these weights to obtain the fused fusion feature vectors.

[0069] As a possible implementation, in order to reduce the computational complexity, the preliminary feature map can be divided into multiple local windows. Self-attention calculation is performed within each window to capture the feature relationship in the local area. In order to maintain the flow of global information, window sliding or other methods can also be used to enable information exchange and fusion between different windows.

[0070] It should be noted that since the multi-layer feature fusion network adopts a hierarchical architecture, each layer performs self-attention calculation and fusion operation on the feature map. After each layer is processed, the output feature map is used as the input of the next layer. At the same time, the hierarchical fusion feature vector generated by the bottom layer is retained and transmitted to the higher layer through residual connection.

[0071] Further, through layer-by-layer processing and cross-layer connection, the feature maps of each level are gradually fused, and the processed feature maps are upsampled to the target resolution to form the final fused feature map.

[0072] In S102, low-level fused features corresponding to the fused feature map are extracted, the fused feature map is downsampled to reduce the spatial resolution of the fused feature map, and then high-level fused features corresponding to the fused feature map are further extracted while maintaining the spatial resolution.

[0073] In a specific implementation, after the meteorological data of different data sources are fused, the fused feature map is input into a preset downscaling model, and a downsampling operation is performed in the preset downscaling model to reduce the spatial resolution of the fused feature map, i.e., to reduce the scale of the meteorological data.

[0074] Here, the preset downscaling model can be implemented by using a pix2pixGAN model, which includes a generator and a discriminator. The generator is responsible for converting the input fused feature map into small-scale, high-resolution meteorological data, and the discriminator is used to make the results generated by the generator more approximate to higher-resolution data.

[0075] The generator can adopt a ResUNet architecture, including the structures of an encoder, a bottleneck layer, a decoder, and an output layer. The encoder extracts multi-level features and gradually reduces the spatial resolution of the feature map while retaining important information; the bottleneck layer further processes the feature map to extract high-level semantic features while maintaining the size of the feature map; the decoder restores the spatial resolution of the feature map to generate high-resolution meteorological data, and simultaneously fuses low-level detailed information by using a skip connection; and the output layer converts the feature map of the decoder part into the final small-scale, high-resolution meteorological data.

[0076] Specifically, the downsampling process can be implemented by using the encoder in the generator, including the following steps 1-4:

[0077] Step 1, input the fused feature map into the encoder of the preset downscaling model, wherein the encoder includes a convolution layer, a residual layer, and a max-pooling layer.

[0078] Step 2, perform convolution operation on the fused feature map by using the convolution layer to determine the low-level fused features, and pass the low-level fused features to the residual layer.

[0079] Step 3, superimpose the fused feature map and the low-level fused features by using residual connection in the residual layer, and pass them to the max-pooling layer.

[0080] Step 4, down-sampling by the max-pooling layer, and storing the feature map with reduced spatial resolution into the skip connection.

[0081] In a specific implementation, the fused feature map is first subjected to a preliminary convolution operation in the convolutional layers of the encoder to extract low-level fused features. For example, a 3x3 convolution kernel can be used with a stride of 1 and padding of 1, and the output feature map maintains the original size. Then, the low-level fused features are passed to the residual block, each residual layer of which contains multiple convolution blocks with residual connections. Preferably, each residual layer is composed of two 3x3 convolutions, batch normalization, and ReLU activation, and the residual connection ensures that the input features are superimposed with the features after the convolution operation to enhance gradient flow. Further, a max-pooling operation is performed after the residual layer to down-sample the feature map and reduce the spatial resolution. For example, a 2x2 pooling with a stride of 2 can be used to halve the size of the feature map.

[0082] In this way, the spatial dimensions of the feature map are continuously reduced through the down-sampling operation of the max-pooling layer to aggregate information of larger spatial regions by reducing the spatial resolution.

[0083] For example, assuming that the size of the input fused feature map is 256x256xC (C is the number of channels), after the convolutional layer operation, the size of the feature map can be 256x256xC, after the residual layer operation, the size of the feature map can be 256x256xC, and after the down-sampling of the max-pooling layer, the size of the feature map can be 128x128xC.

[0084] It should be noted that after each down-sampling step, the down-sampled feature map is stored through the skip connection for subsequent use by the decoder.

[0085] Here, after the down-sampling reduces the spatial resolution of the fused feature map, a convolution operation and a residual connection are performed through the bottleneck layer to further process the feature map and extract high-level semantic features while maintaining the size of the feature map. This can be achieved in the following way: the down-sampled feature map is input into the bottleneck layer of the pre-set down-scaling model, and the high-level fused features are extracted through the convolution operation and the residual connection.

[0086] Here, the bottleneck layer does not perform a pooling operation to maintain the spatial resolution. For example, the 128x128xC convolution operation can be continued.

[0087] In this way, the high-resolution fused feature map output after the feature fusion of large-scale, low-resolution meteorological data is down-sampled through convolution and max-pooling operations to extract multi-level features, achieve down-scaling, and further processed by the bottleneck layer to extract high-level semantic features, so as to extract multi-level features and gradually reduce the spatial resolution of the feature map while preserving important information.

[0088] S103, up-sampling the fusion feature map after extracting the high-level fusion feature to restore the spatial resolution, and splicing the up-sampled fusion feature map with the down-sampled fusion feature map to determine a spliced feature map.

[0089] In a specific implementation, after the encoder down-samples the fusion feature map to reduce the spatial resolution, the decoder of the preset down-scaling model is input, the spatial resolution of the feature map is gradually restored, high-resolution meteorological data is generated, and low-level detail information is fused by using a skip connection.

[0090] Here, the up-sampling process can include the following steps 1-3:

[0091] Step 1, input the feature map output by the encoder into the decoder of the preset down-scaling model, wherein the decoder includes a deconvolution layer and a feature fusion layer.

[0092] Step 2, up-sample the input feature map through the deconvolution layer to restore the spatial resolution, and input the up-sampled feature map to the skip connection.

[0093] Step 3, in the skip connection, splice the stored feature map with reduced spatial resolution with the up-sampled feature map to determine the spliced feature map.

[0094] In a specific implementation, the feature map is gradually up-sampled by a deconvolution or an up-sampling layer to restore the spatial resolution. For example, 2x2 deconvolution (or up-sampling) is used, the stride is 2, and the feature map size is doubled. Then, the skip connection is used to splice the feature map stored by the encoder part with the up-sampled feature map to determine the spliced feature map that retains the detail information.

[0095] For example, assuming that the size of the up-sampled feature map is 256x256xC, the size of the up-sampled feature map is 256x256xC, and then the size of the spliced feature map (obtained from the skip connection storage) of the corresponding layer feature map of the encoder is 256x256x2C.

[0096] It should be noted that although the decoder part gradually restores the spatial resolution of the feature map by up-sampling (deconvolution or other up-sampling operations), this process is different from the input data, and small-scale, high-resolution data is generated. This up-sampling process is actually to restore higher-resolution output data from the abstract features after down-scaling.

[0097] S104, performing convolution operation on the spliced feature map, converting the spliced feature map into small-scale meteorological data after fusing the low-level fusion feature and the high-level fusion feature.

[0098] In a specific implementation, the decoder performs convolution operation on the spliced feature map through a convolution layer and a residual block for the spliced feature map, fuses high-level semantic and low-level detail information, and then converts the spliced feature map output by the decoder into final small-scale and high-resolution meteorological data through an output layer.

[0099] Here, small-scale refers to meteorological data with a small spatial coverage (for example, regional-level meteorological model output); high-resolution refers to rich spatial details, with each data point covering a small area (for example, each data point representing an area of 1 kilometer or less).

[0100] Specifically, the spliced feature map is input to the feature fusion layer of the decoder, and convolution operation is performed on the spliced feature map to fuse low-level fusion features and high-level fusion features.

[0101] Further, the generator output layer finally converts the feature map output by the decoder into small-scale and high-resolution meteorological data through a convolution layer. For example, a 1x1 convolution kernel is used, and the number of output channels is equal to the number of channels of the small-scale and high-resolution meteorological data.

[0102] As a possible implementation, the loss function of the preset downscaling model includes a generator loss and a discriminator loss. The generator loss can include an L1 loss (or an L2 loss) and an adversarial loss. The L1 loss is used to measure the pixel-level difference between the generated meteorological data and the real high-resolution meteorological data, and the adversarial loss is used to improve the authenticity of the generated data. The discriminator loss is used to optimize the discriminator so that it can effectively distinguish between the generated high-resolution meteorological data and the real data.

[0103] Here, the generator and the discriminator can be alternately trained during the training process: the generator generates high-resolution meteorological data, and the discriminator judges its authenticity, and the two networks are continuously optimized through mutual confrontation.

[0104] The method for processing meteorological data provided in the embodiments of the present disclosure comprises the following steps: acquiring large-scale meteorological data of multiple types, extracting a preliminary feature map corresponding to each type of the large-scale meteorological data, and constructing a fusion feature map according to the feature dependency relationship between the preliminary feature maps; extracting low-level fusion features corresponding to the fusion feature map, performing down-sampling on the fusion feature map to reduce the spatial resolution of the fusion feature map, and then further extracting high-level fusion features corresponding to the fusion feature map while maintaining the spatial resolution; performing up-sampling on the fusion feature map after the high-level fusion features are extracted to restore the spatial resolution, splicing the fusion feature map after the up-sampling and the fusion feature map after the down-sampling to determine a spliced feature map; and performing convolution operation on the spliced feature map, converting the spliced feature map into small-scale meteorological data after fusing the low-level fusion features and the high-level fusion features. The method can realize the downscaling processing of meteorological data by using fewer computing resources, effectively fuse meteorological data from different data sources, and overcome the limitation problem in processing nonlinear and complex meteorological relationships in the prior art.

[0105] Those skilled in the art can understand that, in the above method of the specific implementation, the writing order of each step does not mean a strict execution order and does not constitute any limitation on the implementation process, and the specific execution order of each step should be determined according to its function and possible internal logic.

[0106] Based on the same inventive concept, the embodiments of the present disclosure also provide a device for processing meteorological data corresponding to the method for processing meteorological data. Since the principle of solving problems in the device of the embodiments of the present disclosure is similar to the above-mentioned method for processing meteorological data, the implementation of the device can be referred to the implementation of the method, and the repeated parts will not be described herein.

[0107] Please refer to Figure 3 , Figure 3 The device for processing meteorological data provided in the embodiments of the present disclosure is shown in the following figure. Figure 3 The device for processing meteorological data 300 provided in the embodiments of the present disclosure comprises the following parts:

[0108] The feature fusion module 310 is configured to acquire large-scale meteorological data of multiple types, extract a preliminary feature map corresponding to each type of the large-scale meteorological data, and construct a fusion feature map according to the feature dependency relationship between the preliminary feature maps.

[0109] The downscaling module 320 is configured to extract low-level fusion features corresponding to the fusion feature map, perform down-sampling on the fusion feature map to reduce the spatial resolution of the fusion feature map, and then further extract high-level fusion features corresponding to the fusion feature map while maintaining the spatial resolution.

[0110] The resolution enhancement module 330 is used to upsample the fusion feature map after extracting the high-level fusion features to restore the spatial resolution, and to stitch the upsampled fusion feature map with the downsampled fusion feature map to determine the stitched feature map.

[0111] The data conversion module 340 is used to perform convolution operations on the spliced ​​feature map, and after fusing the low-level fusion features and the high-level fusion features, convert the spliced ​​feature map into small-scale meteorological data.

[0112] The processing flow of each module in the device and the interaction flow between each module can be referred to the relevant descriptions in the above method embodiments, and will not be detailed here.

[0113] This disclosure provides a meteorological data processing apparatus that acquires various types of large-scale meteorological data, extracts features corresponding to each type of large-scale meteorological data to construct a preliminary feature map, and constructs a fused feature map based on the feature dependencies between the preliminary feature maps. It then extracts low-level fusion features corresponding to the fused feature map, downsamples the fused feature map to reduce its spatial resolution, and further extracts high-level fusion features while maintaining the spatial resolution. The fused feature map after extracting the high-level fusion features is upsampled to restore the spatial resolution, and the upsampled fused feature map is concatenated with the downsampled fused feature map to determine a concatenated feature map. Finally, a convolution operation is performed on the concatenated feature map, and after fusing the low-level and high-level fusion features, the concatenated feature map is converted into small-scale meteorological data. This method can achieve meteorological data downscaling with fewer computational resources, while effectively fusing meteorological data from different sources, overcoming the limitations of existing technologies in handling nonlinear and complex meteorological relationships.

[0114] Corresponding to Figure 1 In addition to the method for processing meteorological data, this disclosure also provides an electronic device 400, such as... Figure 4 The diagram shown is a structural schematic of an electronic device 400 provided in an embodiment of this disclosure, including:

[0115] The processor 41, the memory 42, and the bus 43; the memory 42 is used for storing execution instructions, including the internal memory 421 and the external memory 422; the internal memory 421 here is also called the internal memory, used for temporarily storing operation data in the processor 41 and data exchanged with the external memory 422 such as a hard disk, the processor 41 exchanges data with the external memory 422 through the internal memory 421, when the electronic device 400 is running, the processor 41 and the memory 42 communicate through the bus 43, so that the processor 41 executes Figure 1 the steps of the processing method of the meteorological data.

[0116] The embodiment of the present disclosure further provides a computer readable storage medium, which stores a computer program, and the computer program performs the steps of the processing method of the meteorological data when executed by a processor. The storage medium can be a volatile or non-volatile computer readable storage medium.

[0117] The embodiment of the present disclosure further provides a computer program product, which comprises computer instructions, and the computer instructions perform the steps of the processing method of the meteorological data when executed by a processor. For details, refer to the above method embodiments, which will not be repeated here.

[0118] The computer program product can be specifically implemented by hardware, software or a combination thereof. In an optional embodiment, the computer program product is specifically embodied as a computer storage medium, and in another optional embodiment, the computer program product is specifically embodied as a software product, such as a software development kit (Software Development Kit, SDK) and the like.

[0119] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the above-described device can refer to the corresponding process in the foregoing method embodiments, which will not be repeated here. In several embodiments provided by the present disclosure, it should be understood that the disclosed devices and methods can be implemented by other ways. The device embodiments described above are only schematic, for example, the division of the units is only a logical function division, and actual implementation can have another division manner, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the shown or discussed mutual elements can be indirect coupling or communication connection through some communication interface, device or unit, which can be electrical, mechanical or other forms.

[0120] The units described as separate components may or may not be physically separate, and the components displayed as units may or may not be physical units, i.e. may be located in one place, or may be distributed on multiple network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the embodiment scheme.

[0121] In addition, each functional unit in various embodiments of the present disclosure can be integrated into one processing unit, or each unit can exist physically, or two or more units can be integrated into one unit.

[0122] If the functions are realized in the form of software functional units and sold or used as independent products, they can be stored in a non-volatile computer readable storage medium executable by a processor. Based on this understanding, the technical solutions of the present disclosure essentially or the part of the prior art or the part of the technical solutions can be embodied in the form of a software product, which is stored in a storage medium and includes a number of instructions for making a computer device (which can be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the method described in various embodiments of the present disclosure. The foregoing storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), magnetic disk or optical disk, and various program code storage media.

[0123] Finally, it should be noted that: the above-described embodiments are only specific embodiments of the present disclosure, used to illustrate the technical solutions of the present disclosure, and not to limit it, the protection scope of the present disclosure is not limited to this, although the present disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand: any person skilled in the art in the technical range disclosed by the present disclosure, still can modify or easily think of changes to the technical solutions recorded in the foregoing embodiments, or make equivalent replacement to part of the technical features; and these modifications, changes or replacements do not make the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present disclosure, and all should be covered in the protection scope of the present disclosure. Therefore, the protection scope of the present disclosure should be subject to the protection scope of the claims.

Claims

1. A method of processing meteorological data, characterized by, The method comprises the following steps: acquiring a plurality of types of large-scale meteorological data, extracting a preliminary feature map corresponding to each of the large-scale meteorological data, and constructing a fusion feature map according to a feature dependency relationship between the preliminary feature maps; extracting low-level fusion features corresponding to the fusion feature map, downsampling the fusion feature map to reduce the spatial resolution of the fusion feature map, and then further extracting high-level fusion features corresponding to the fusion feature map while maintaining the spatial resolution; upsampling the fusion feature map after extracting the high-level fusion features to restore the spatial resolution, and splicing the upsampled fusion feature map with the downsampled fusion feature map to determine a spliced feature map; performing convolution operations on the spliced feature map, converting the spliced feature map into small-scale meteorological data after fusing the low-level fusion features and the high-level fusion features; determining the feature dependency relationship based on the following steps: aligning the preliminary feature maps to the same spatial scale, and adjusting the preliminary feature maps to the same number of channels using linear transformation; dividing the aligned preliminary feature maps into a plurality of feature blocks; performing linear transformation on each of the feature blocks to determine a corresponding query vector, a keyword vector, and a value vector; determining an attention weight based on the dot product between the query vector and the keyword vector, and taking the attention weight as the feature dependency relationship; wherein the preliminary feature map is divided into a plurality of local windows, and self-attention calculation is performed in each window to capture the feature relationship in the local area, and the information between different windows is transmitted and fused through window sliding.

2. The method of claim 1, wherein, constructing the fusion feature map based on the following steps: inputting the feature blocks into a multi-layer feature fusion network, in each layer, for each of the feature blocks, weighting the value vector corresponding to the feature block and the value vectors corresponding to the other feature blocks according to the attention weight, determining a hierarchical fusion feature vector and generating a corresponding hierarchical fusion feature map, wherein the hierarchical fusion feature vector generated by the bottom layer is reserved and transmitted to the high layer through residual connection between layers; splicing and fusing each layer corresponding to the hierarchical fusion feature map to generate the fusion feature map.

3. The method of claim 1, wherein, extracting low-level fusion features corresponding to the fusion feature map, downsampling the fusion feature map to reduce the spatial resolution of the fusion feature map, specifically including: inputting the fusion feature map into an encoder of a preset downscaling model, wherein the encoder includes a convolution layer, a residual layer, and a max pooling layer; performing convolution operations on the fusion feature map through the convolution layer to determine the low-level fusion features, and transmitting the low-level fusion features to the residual layer; in the residual layer, superimposing the fusion feature map and the low-level fusion features through residual connection, and transmitting to the max pooling layer; performing downsampling through the max pooling layer, and storing the feature map with reduced spatial resolution into a skip connection.

4. The method of claim 3, wherein, extracting high-level fusion features corresponding to the fusion feature map while maintaining the spatial resolution, specifically including: The down-sampled feature map is input to a bottleneck layer of the preset down-scaling model, and the high-level fusion feature is extracted through convolution operation and residual connection; The bottleneck layer does not perform a pooling operation to maintain the spatial resolution.

5. The method of claim 3, wherein, The fusion feature map after extracting the high-level fusion feature is up-sampled to restore the spatial resolution, and the up-sampled fusion feature map is spliced with the down-sampled fusion feature map to determine a spliced feature map, and the spliced feature map specifically includes: The feature map output by the encoder is input to a decoder of the preset down-scaling model, and the decoder includes a deconvolution layer and a feature fusion layer; The deconvolution layer is used for up-sampling the input feature map to restore the spatial resolution, and the up-sampled feature map is input to the skip connection; In the skip connection, the stored feature map with reduced spatial resolution is spliced with the up-sampled feature map to determine the spliced feature map.

6. The method of claim 5, wherein, The low-level fusion feature and the high-level fusion feature are fused, and the fusion specifically includes: The spliced feature map is input to the feature fusion layer, and the low-level fusion feature and the high-level fusion feature are fused through convolution operation on the spliced feature map.

7. A weather data processing apparatus characterized by comprising: It includes: The feature fusion module is configured to obtain multiple types of large-scale meteorological data, extract features corresponding to each type of large-scale meteorological data to construct a preliminary feature map, and construct a fusion feature map according to a feature dependency relationship between the preliminary feature maps; The down-scaling module is configured to extract a low-level fusion feature corresponding to the fusion feature map, down-sample the fusion feature map to reduce the spatial resolution of the fusion feature map, and then further extract a high-level fusion feature corresponding to the fusion feature map while maintaining the spatial resolution; The resolution improvement module is configured to up-sample the fusion feature map after extracting the high-level fusion feature to restore the spatial resolution, and splice the up-sampled fusion feature map with the down-sampled fusion feature map to determine a spliced feature map; The data conversion module is configured to perform convolution operation on the spliced feature map, and convert the spliced feature map into small-scale meteorological data after fusing the low-level fusion feature and the high-level fusion feature. The feature fusion module is specifically configured to align the preliminary feature maps to the same spatial scale, and adjust the preliminary feature maps to the same number of channels by using linear transformation; divide the aligned preliminary feature maps into a plurality of feature blocks; perform linear transformation on each feature block to determine a corresponding query vector, a keyword vector, and a value vector; determine an attention weight according to a dot product between the query vector and the keyword vector, and use the attention weight as the feature dependency relationship; wherein, the preliminary feature maps are divided into a plurality of local windows, self-attention calculation is performed in each window to capture the feature relationship in the local area, and information between different windows is transmitted and fused through window sliding.

8. An electronic device, comprising: It includes: A processor, a memory, and a bus, the memory storing machine readable instructions executable by the processor, the processor in communication with the memory via the bus when the electronic device is running, the machine readable instructions, when executed by the processor, performing the steps of the method for processing meteorological data according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, the computer program, when executed by the processor, performing the steps of the method for processing meteorological data according to any one of claims 1 to 6.

Images