Rainfall intelligent correction method, device and equipment and storage medium

Abstract

This invention discloses a method, apparatus, device, and storage medium for intelligent rainfall correction. The method includes: spatiotemporal matching and normalization of X-band phased array radar reflectivity data and meteorological grid real-time data to construct sample pairs; using a dual-branch network to extract features from the two types of data respectively, and strengthening key areas by introducing an attention mechanism; using a downsampling attention module to extract multi-scale spatial features, and then using a cross-modal residual embedding module to achieve deep fusion of local texture features of radar data and global physical features of meteorological data; using an enhanced upsampling channel attention module to progressively upsample the fused features to restore resolution, and combining skip connections and feature alignment, using channel attention to adaptively adjust information at each scale to optimize detail reconstruction. Finally, the constructed samples are used to train the network to obtain an intelligent rainfall data correction model, thereby improving the accuracy and reliability of correcting meteorological grid real-time data.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method, apparatus, device, and storage medium for intelligent rainfall correction. Background Technology

[0002] Rainfall is a key meteorological element affecting human production, daily life, and the evolution of the natural environment. Rainfall monitoring is also of great significance in meteorological and hydrological fields, disaster prevention and mitigation, and water resource management. High-resolution rainfall data is the foundation and essential support for meteorological research, and it is also necessary data for monitoring and evaluating the accuracy of rainfall forecasts. Against the backdrop of global climate change, extreme precipitation events are becoming more frequent and localized, posing severe challenges to flood control and socio-economic development. Therefore, obtaining high-precision and high-resolution rainfall data has become one of the core tasks of meteorological and hydrological research.

[0003] Traditional precipitation data acquisition primarily relies on meteorological station observations and conventional weather radar. While meteorological stations provide high-precision ground-based data, their sparse spatial distribution makes it difficult to reflect the spatial differences in local precipitation. Conventional weather radar can observe precipitation systems on a large scale, but its limited scanning speed makes it difficult to capture rapidly evolving severe convective weather processes in a timely manner. Single data sources often cannot meet the demand for high-precision precipitation data; in practical applications, it is often necessary to fuse observational data from different sources to improve the resolution and accuracy of precipitation products. Multi-source data fusion can fully leverage the complementary advantages of different observation methods to improve data quality and reliability, and has made significant progress in many fields such as meteorology, intelligent transportation, and medical imaging. However, most existing multi-source data fusion correction technologies still lack effective mining of the complementarity of different source data at different scales, making it difficult to achieve deep coupling of multi-source data.

[0004] X-band phased array radar, as a novel meteorological observation device, boasts advantages such as short wavelength, high spatial resolution, and rapid scanning. It can acquire high-precision three-dimensional reflectivity fields in an extremely short time, exhibiting stronger detection and capture capabilities for strong convective cells, fine-scale precipitation structures, and locally generated short-duration heavy precipitation. Currently, X-band phased array radar reflectivity data has been widely applied in areas such as refined urban monitoring, short-term forecasting, and quantitative precipitation estimation. However, X-band radar is highly sensitive to the attenuation of precipitation particles, and its reflectivity is easily affected by factors such as attenuation from heavy precipitation, the wet antenna effect, and ground clutter. Furthermore, the dual polarization parameters also exhibit certain uncertainties in quantitative inversion, leading to deviations in precipitation inversion. Single-radar observation information cannot fully reflect complex meteorological and physical processes, while real-time meteorological data (such as temperature, humidity, air pressure, and wind speed) can provide supplementary information reflecting the atmospheric background field. Summary of the Invention

[0005] The purpose of this invention is to provide a rainfall intelligent correction method, device, equipment and storage medium, which aims to solve the problems of insufficient modeling capability of single data source and insufficient utilization of features at different scales in existing rainfall inversion methods.

[0006] In a first aspect, embodiments of the present invention provide a rainfall intelligent correction method, comprising: Acquire reflectivity data from an X-band phased array radar and physical field data from a weather model, and preprocess the reflectivity data and physical field data to obtain sample pairs; A dual-branch network structure is used to extract features from the sample pairs to obtain phased array radar data features and meteorological grid real-time data features, respectively. The phased array radar data features and the meteorological grid real-time data features are input into the downsampling attention module for multi-scale spatial feature extraction, resulting in multi-scale phased array radar data features and multi-scale meteorological grid real-time data features, respectively. Based on the multi-scale phased array radar data features and the multi-scale meteorological grid real-time data features, deep and shallow fusion is performed at different levels through a cross-modal residual embedding module to obtain multi-scale fusion features and shallow fusion features respectively. The multi-scale fusion features are upsampled step by step by enhancing the attention module of the upsampling channel, and skip connections are made between each level and the corresponding shallow fusion features to obtain the rainfall correction result. Acquire measured rainfall data, and iteratively train the dual-branch network structure, the downsampling attention module, the cross-modal residual embedding module, and the enhanced upsampling channel attention module based on the measured rainfall data and the rainfall correction results to obtain an intelligent rainfall data correction model; The reflectivity data of the X-band phased array radar to be corrected and the physical field data of the meteorological model are obtained, and the reflectivity data of the X-band phased array radar to be corrected and the physical field data of the meteorological model are input into the intelligent correction model of the rainfall data to obtain the corrected rainfall data.

[0007] Secondly, embodiments of the present invention provide a rainfall intelligent correction device, comprising: The preprocessing module is used to acquire reflectivity data of X-band phased array radar and physical field data of meteorological models, and to preprocess the reflectivity data and physical field data to obtain sample pairs. The extraction module is used to extract features from the sample pairs using a dual-branch network structure, thereby obtaining phased array radar data features and meteorological grid real-time data features respectively. The input module is used to input the phased array radar data features and the meteorological grid real-time data features into the downsampling attention module for multi-scale spatial feature extraction, so as to obtain multi-scale phased array radar data features and multi-scale meteorological grid real-time data features respectively. The fusion module is used to perform deep and shallow fusion at different levels based on the multi-scale phased array radar data features and the multi-scale meteorological grid real-time data features, through the cross-modal residual embedding module, to obtain multi-scale fusion features and shallow fusion features respectively. The connection module is used to upsample the multi-scale fusion features step by step through the enhanced upsampling channel attention module, and to make skip connections with the corresponding shallow fusion features at each level to obtain the rainfall correction result. The training module is used to acquire measured rainfall data and iteratively train the dual-branch network structure, the downsampling attention module, the cross-modal residual embedding module, and the enhanced upsampling channel attention module based on the measured rainfall data and the rainfall correction results to obtain an intelligent rainfall data correction model. The acquisition module is used to acquire the X-band phased array radar reflectivity data to be corrected and the physical field data of the meteorological model, and input the X-band phased array radar reflectivity data to be corrected and the physical field data of the meteorological model into the intelligent correction model of the rainfall data to obtain the corrected rainfall data.

[0008] Thirdly, embodiments of the present invention provide a computer device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the rainfall intelligent correction method described in the first aspect.

[0009] Fourthly, embodiments of the present invention also provide a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program that, when executed by a processor, implements the rainfall intelligent correction method described in the first aspect.

[0010] This invention discloses a method, apparatus, device, and storage medium for intelligent rainfall correction. The method includes: acquiring reflectivity data from an X-band phased array radar and physical field data from a meteorological model; preprocessing the reflectivity data and the physical field data to obtain sample pairs; extracting features from the sample pairs using a dual-branch network structure to obtain phased array radar data features and meteorological grid real-time data features; inputting the phased array radar data features and the meteorological grid real-time data features into a downsampling attention module for multi-scale spatial feature extraction to obtain multi-scale phased array radar data features and multi-scale meteorological grid real-time data features; and based on the multi-scale phased array radar data features and multi-scale meteorological grid real-time data features, performing deep and shallow embedding at different levels through a cross-modal residual embedding module. The system performs a multi-scale fusion and a shallow fusion, respectively. The multi-scale fusion features are then upsampled level by level using an enhanced upsampling channel attention module, with skip connections made between each level and the corresponding shallow fusion features to obtain rainfall correction results. Measured rainfall data is acquired, and the dual-branch network structure, the downsampling attention module, the cross-modal residual embedding module, and the enhanced upsampling channel attention module are iteratively trained based on the measured rainfall data and the rainfall correction results to obtain an intelligent rainfall data correction model. Finally, the X-band phased array radar reflectivity data to be corrected and the physical field data of the meteorological model are acquired and input into the intelligent rainfall data correction model to obtain corrected rainfall data. This invention constructs a dual-branch network to extract local high-resolution texture features from phased array radar and global physical features from meteorological grid data. It utilizes a downsampling attention module to extract multi-scale spatial structure and employs a cross-modal residual embedding module to achieve deep fusion and dynamic weighting of the two types of data at different levels, effectively overcoming the problems of insufficient information from a single data source and difficulty in deep coupling of multi-source data. Simultaneously, an enhanced upsampling channel attention module, combined with a skip connection mechanism, reconstructs a high-resolution precipitation field, suppressing attenuation and clutter interference from X-band radar. This significantly improves the spatial resolution and correction accuracy of precipitation data, resulting in more accurate, stable, and generalized rainfall estimation results. This invention also provides a rainfall intelligent correction device, a computer-readable storage medium, and a computer device, all possessing the aforementioned beneficial effects, which will not be elaborated further here. Attached Figure Description

[0011] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 A flowchart illustrating the intelligent rainfall correction method; Figure 2 A schematic diagram of the structure of an intelligent correction model for rainfall data; Figure 3 A schematic diagram of the sub-branch network structure of a dual-branch network structure; Figure 4 This is a schematic diagram of the downsampling attention module; Figure 5 This is a schematic diagram of the cross-modal residual embedding module; Figure 6 A schematic diagram of the structure of the attention module for enhanced upsampling channels; Figure 7 A schematic block diagram of a rainfall intelligent correction device. Detailed Implementation

[0013] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0014] It should be understood that, when used in this specification and the appended claims, the terms “comprising” and “including” indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more of its features, integrals, steps, operations, elements, components and / or collections thereof.

[0015] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0016] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the relevant listed items and all possible combinations, and includes such combinations.

[0017] Please see Figure 1 and Figure 2 This embodiment provides a rainfall intelligent correction method, including: S101: Acquire reflectivity data of X-band phased array radar and physical field data of meteorological model, and preprocess the reflectivity data and physical field data to obtain sample pairs; Specifically, the X-band phased array radar acquires the three-dimensional reflectivity factor field using a high spatiotemporal resolution scanning method, covering the target area with a temporal resolution on the order of minutes and a spatial resolution on the order of tens to hundreds of meters. The physical field data of the meteorological model comes from numerical weather prediction models or reanalysis data, including variables such as temperature, humidity, air pressure, and wind field. Its spatial resolution is relatively coarse, covering a large scale area within the same region. To construct sample pairs for model training, both types of data are preprocessed. The first step is spatiotemporal matching: using the time of the meteorological grid's real-time data as a reference, the closest radar scan time within the time window is selected, and the radar data is interpolated vertically to match the vertical layer number of the meteorological data; simultaneously, the radar reflectivity data is resampled to the horizontal grid points of the meteorological data using nearest neighbor or bilinear interpolation methods. The second step is spatial interpolation: for missing grid points in the meteorological physical field, inverse range weighted or kriging interpolation is used to fill in the gaps, ensuring data integrity. The third step is to perform normalization: the radar reflectivity value is normalized to the [0, 1] interval through maximum and minimum value normalization, and meteorological variables are standardized or Z-score normalized according to their respective physical ranges to eliminate dimensional differences.

[0018] Next, while acquiring the X-band phased array radar reflectivity data, the radar's dual polarization parameters are simultaneously acquired, including at least the differential phase ΦDP. The differential phase ΦDP and reflectivity data are spatiotemporally matched to ensure alignment within the same time window and spatial grid. The path integral attenuation is estimated using the differential phase ΦDP. Specifically, along the radar radial direction, the differential phase change rate (i.e., the differential phase KDP) between adjacent range bins is calculated: KDP ≈ ΔΦDP / (2·Δr), where Δr is the range bin spacing. Then, using the empirical relationship between KDP and the attenuation coefficient, the attenuation on each path segment is inverted, and the cumulative attenuation is obtained by radial integration. This cumulative attenuation is applied to the corresponding reflectivity factor Zh_measured for attenuation correction, using the formula Zh_corrected = Zh_measured + attenuation correction value. After preliminary attenuation correction, the corrected reflectivity data Zh_corrected is obtained. This corrected reflectivity data is then used to construct the sample pairs.

[0019] After the above processing is completed, radar reflectivity data and meteorological physical field data at the same time and spatial location are paired to form sample pairs for subsequent network training. Each sample pair contains one phased array radar data and one meteorological grid real-time data, which serve as the synchronous input to the dual-branch network.

[0020] S102: A dual-branch network structure is used to extract features from the sample pairs to obtain phased array radar data features and meteorological grid real-time data features, respectively. In this embodiment, please refer to Figure 3 A dual-branch network structure was used to extract features from sample pairs, resulting in the following features: features of phased array radar data and features of meteorological grid real-time data. A primary spatial structure representation pair is constructed by progressively abstracting sample pairs using a three-layer convolutional structure. Dimensionality reduction of the primary spatial structure representation pairs is performed using 1×1 convolution to obtain compressed feature pairs; The compressed feature pairs are adaptively evaluated using a channel attention mechanism to obtain the evaluated feature pairs; Feature extraction was performed on the evaluation feature pairs using 3×3 convolution and downsampling to obtain the features of phased array radar data and meteorological grid real-time data, respectively.

[0021] This embodiment employs a three-layer convolutional structure to progressively abstract sample pairs, extracting hierarchical spatial features from the original input, ranging from bottom-level edge textures to mid-level local structures, thus enhancing the representation of multi-scale morphology of precipitation systems. Secondly, 1×1 convolutions are introduced for dimensionality reduction, effectively compressing the number of feature channels while preserving key information, reducing subsequent computational overhead, and enabling linear recombination of cross-channel information, thereby enhancing feature expressiveness. Thirdly, a channel attention mechanism is embedded, adaptively evaluating the importance of each channel through global pooling and weight mapping, resulting in higher responses for feature channels with significant physical meaning in precipitation inversion and suppressing interference from redundant or noisy channels. Finally, 3×3 convolutions combined with downsampling operations expand the receptive field while extracting deep semantic features. A dual-branch structure is used to independently process phased array radar data and meteorological grid data, avoiding feature aliasing between heterogeneous data sources, preserving the unique attributes of each modality, and providing cleaner and more discriminative feature inputs for subsequent cross-modal fusion.

[0022] Specifically, a dual-branch network structure is used to extract features from sample pairs. The two branches have the same structure but independent parameters, processing phased array radar data and meteorological grid data respectively. For each branch, the preprocessed sample pairs are input into a three-layer convolutional structure for progressive abstraction. The first convolutional layer uses a 3×3 kernel with a stride of 1 and padding of 1, and the number of output channels is set to the middle number of channels. It performs preliminary local feature extraction on the input feature map, capturing subtle changes in the edges, textures, and reflectivity gradients of the precipitation system, and outputs the first-level feature map. This is followed by a batch normalization layer and a ReLU activation function to stabilize the training process and introduce nonlinearity. The second convolutional layer also uses a 3×3 kernel, but doubles the number of output channels. Based on the first-level feature map, it further expands the receptive field, aggregates contextual information from neighboring regions, and begins to form preliminary representations of mesoscale structures in precipitation, such as convective cells and rainband outlines. This layer is also followed by batch normalization and ReLU activation. The third convolutional layer maintains a 3×3 kernel, and the number of output channels is either doubled again or remains the same as the second layer. This further abstracts a larger-scale spatial organization pattern of precipitation, such as the distinction between stratiform clouds and convective regions, and the location and morphology of the core area of ​​heavy precipitation, thus completing the construction of the primary spatial structure representation. Each of the three convolutional layers enhances feature representation by progressively increasing the number of channels, while using small-sized convolutional kernels to preserve spatial details. The final output primary spatial structure representation retains the fine spatial location information of the original input and incorporates local to mesoscale structural priors, providing a rich spatial feature foundation for subsequent dimensionality reduction and attention enhancement.

[0023] Next, the two primary spatial structure representations output by the dual-branch network are input into a 1×1 convolutional layer for dimensionality reduction. This 1×1 convolutional layer has a 1×1 kernel size, a stride of 1, no padding, and its output channel count is set to half the number of channels in the primary spatial structure representation. When the feature map size of the primary spatial structure representation is H×W and the number of channels is C, the 1×1 convolution generates a new feature map with C / 2 channels by linearly weighting the C channels at each spatial location. This operation does not change the spatial resolution of the feature map but significantly reduces the channel dimension, thereby reducing subsequent computation and parameter requirements. Simultaneously, the 1×1 convolution enables cross-channel information exchange and feature recalibration, fusing correlated features from different channels and removing redundant information. During convolution computation, each output channel corresponds to a set of learnable weight coefficients, which are applied to various spatial locations of the input feature map using pointwise convolution. The compressed feature pairs obtained after dimensionality reduction retain the core spatial response patterns in the primary spatial structure representation pairs, while compressing the interference of non-critical channels, providing more compact and information-density feature inputs for the adaptive evaluation of subsequent channel attention mechanisms.

[0024] The channel attention module then performs global average pooling on the compressed feature map, compressing the spatial dimension features of each channel into a scalar value, thus obtaining a channel description vector of length C / 2. This vector represents the average activation intensity of the global spatial response on each channel. This description vector is then fed into a gating mechanism consisting of two fully connected layers. The first fully connected layer reduces the number of channels from C / 2 to C / 4, introducing nonlinearity using the ReLU activation function, achieving dimensionality reduction and information compression. The second fully connected layer restores the number of channels to the original C / 2, outputting a channel weight vector after nonlinear transformation. Finally, the Sigmoid activation function maps each element of the weight vector to between 0 and 1, generating the final channel attention weights. The original compressed feature map is multiplied channel-by-channel by the generated channel attention weights; that is, the feature map of each channel is multiplied by the corresponding weight coefficient, achieving recalibration of the importance of different channels. For the phased array radar data branch, this mechanism enhances channels sensitive to the core region of precipitation echoes and suppresses channels affected by attenuation or clutter. For the meteorological grid data branch, it highlights meteorological variable channels that are indicative of precipitation physical processes, such as vertical velocity and water vapor flux. The evaluation feature pairs output after channel attention weighting significantly improve the response intensity of key channels while maintaining the original spatial structure, providing more discriminative feature inputs for subsequent deep feature extraction.

[0025] Finally, feature extraction is performed on the evaluation feature pairs using 3×3 convolution and downsampling. A 3×3 convolution with a stride of 2 is used to convolve the evaluation feature map, expanding the receptive field while halving the spatial size, thus extracting deeper semantic features. This downsampling operation not only reduces the resolution of the feature map but also enhances the model's robustness to translation and deformation. After the above processing, the dual-branch network outputs phased array radar data features and meteorological grid real-time data features, respectively. Both have the same number of channels and spatial size, facilitating subsequent multi-scale feature extraction and cross-modal fusion.

[0026] S103: Input the phased array radar data features and the meteorological grid real-time data features into the downsampling attention module to perform multi-scale spatial feature extraction, and obtain multi-scale phased array radar data features and multi-scale meteorological grid real-time data features respectively. In this embodiment, please refer to Figure 4 The phased array radar data features and meteorological grid real-time data features are input into the downsampling attention module (EDSA) for multi-scale spatial feature extraction, resulting in the following multi-scale phased array radar data features and multi-scale meteorological grid real-time data features: Based on the characteristics of phased array radar data and meteorological grid real-time data, downsampling is performed through step convolution to obtain radar downsampling feature maps and meteorological downsampling feature maps, respectively. Channel max pooling was performed on the radar downsampled feature map and the meteorological downsampled feature map to obtain the radar spatial structure features and the meteorological spatial structure features, respectively. Based on radar spatial structure features and meteorological spatial structure features, radar spatial attention maps and meteorological spatial attention maps are generated using 3×3 convolution and Sigmoid activation, respectively. The radar spatial attention map and the radar downsampled feature map are weighted element-wise to obtain the multi-scale phased array radar data features. By weighting the meteorological spatial attention map and the meteorological downsampling feature map element by element, the features of real-time meteorological grid data at multiple scales are obtained.

[0027] In this embodiment, downsampling using stride convolution expands the receptive field while compressing the spatial dimension, enabling the model to capture multi-scale structural information of the precipitation system, from local cells to large-scale rainbands. Channel max pooling is applied to the downsampled feature map, preserving the strongest response value at each spatial location, effectively highlighting the core precipitation region and strong echo centers, and enhancing the model's sensitivity to areas with significant precipitation. A spatial attention map is generated using 3×3 convolution combined with sigmoid activation, allowing the network to adaptively learn the importance distribution of spatial locations, assigning higher weights to areas with high precipitation intensity and complex structures, while suppressing interference from non-precipitation areas such as ground clutter and attenuation artifacts. Finally, the spatial attention map and the downsampled feature map are weighted element-wise, achieving selective enhancement of key spatial regions and improving the quality of multi-scale feature representation. This module is independently applied to the phased array radar branch and the meteorological data branch, ensuring that the spatial structural features of each type of data are optimized in a targeted manner, providing more discriminative multi-scale feature inputs for subsequent cross-modal deep fusion.

[0028] Specifically, the phased array radar data features and meteorological grid real-time data features output from the dual-branch network are input into the downsampling attention module for multi-scale spatial feature extraction. For each branch, the input feature is defined as: ; Where X represents the input feature, R is the real number field, C is the number of input channels, and H and W are the input space dimensions.

[0029] Next, downsampling is performed using stride convolution, employing a convolution operation with a kernel size of k×k and a stride of s, to map the input feature X to a downsampled feature map l. t The specific calculation method is as follows: ; Among them, l t This is the downsampled feature map, where c is the channel number; W is the output spatial location.c Here, u and v are the convolution kernels for the corresponding channels; k is the kernel size, s is the stride, and b is the convolution kernel index. c This is the channel bias term.

[0030] After this stride convolution, the spatial dimension is compressed to 1 / s of the original, resulting in radar downsampled feature maps and meteorological downsampled feature maps.

[0031] Next, channel max pooling is performed on the downsampled feature map to extract spatial structure information. For each spatial location (i, j), the maximum value among all channels is taken, i.e. ; Where Max and max are channel max pooling.

[0032] Max pooling is used to obtain single-channel spatial structure feature maps, referred to as radar spatial structure features and meteorological spatial structure features, respectively. This operation preserves the strongest response at each location, highlighting the core precipitation region.

[0033] Then, based on the aforementioned spatial structure features, a spatial attention map is generated using 3×3 convolution and sigmoid activation. Specifically: ; Among them, W s This refers to the 3×3 convolution kernel in the spatial attention mechanism; The q is the Sigmoid activation function; sp This is a spatial attention map.

[0034] The generated spatial attention map is then element-wise weighted with the downsampled feature map to obtain multi-scale feature output: ; Among them, f EDSA · represents the final output of the EDSA module; · represents element-wise multiplication.

[0035] For the phased array radar branch, the multi-scale phased array radar data features are obtained by weighting the radar spatial attention map and the radar downsampled feature map; for the meteorological branch, the multi-scale meteorological grid real-time data features are obtained by weighting the meteorological spatial attention map and the meteorological downsampled feature map. Through the above process, both branches obtain multi-scale feature representations with spatial dimension compression and significant regional enhancement.

[0036] S104: Based on the multi-scale phased array radar data features and the multi-scale meteorological grid real-time data features, deep and shallow fusion is performed at different levels through the cross-modal residual embedding module to obtain multi-scale fusion features and shallow fusion features respectively. In this embodiment, please refer to Figure 5Based on the characteristics of multi-scale phased array radar data and multi-scale meteorological grid real-time data, deep and shallow fusion is performed at different levels through a cross-modal residual embedding module (CRE) to obtain multi-scale fusion features and shallow fusion features, including: The features of multi-scale phased array radar data and the features of multi-scale meteorological grid real-time data are spliced ​​together along the channel dimension to obtain joint features; The joint features are processed once by two cascaded channel fusion submodules to obtain updated fused features. The updated fusion features are divided into phased array radar data feature information and meteorological grid real-time data feature information according to channels, and deep fusion dynamic weight matrix and shallow fusion dynamic weight matrix are generated through activation function. Based on the deep fusion dynamic weight matrix, the multi-scale phased array radar data features and the multi-scale meteorological grid real-time data features are weighted and superimposed point by point to obtain multi-scale fused features. Based on the shallow fusion dynamic weight matrix, the features of multi-scale phased array radar data and the features of multi-scale meteorological grid real-time data are weighted and superimposed point by point to obtain shallow fusion features.

[0037] This embodiment preserves the integrity and spatial correspondence of dual-modal information by stitching multi-scale features from phased array radar and meteorological grid data along the channel dimension, providing a unified feature space for cross-modal interaction. Two cascaded channel fusion sub-modules are used to progressively strengthen the channel-level correlation between radar and meteorological features. The compression-then-recovery structure effectively reduces redundant features and improves fusion efficiency. The updated fused features are divided into two groups according to channels, and dynamic weight matrices are generated for each group, enabling adaptive evaluation of the contribution of the two modes and avoiding fusion bias caused by fixed weights. Point-by-point weighted superposition based on the deep fusion dynamic weight matrix deeply couples the local fine precipitation structure with the large-scale environmental physical field, significantly improving the expressive power of the fused features. Simultaneously, shallow fused features are generated based on the shallow fusion dynamic weight matrix and passed to the upsampling stage through skip connections, preserving high-resolution spatial details and effectively assisting in deep feature reconstruction. This module simultaneously outputs deep and shallow fused features within the same framework, achieving multi-scale and multi-level information complementarity.

[0038] Specifically, let the characteristics of multi-scale meteorological grid real-time data be... Multi-scale phased array radar data characteristics are .

[0039] Next, the two are concatenated along the channel dimension to obtain the joint feature: ; Among them, f concatThis represents the bimodal features after channel concatenation; Concat indicates the channel dimension splicing operation.

[0040] The spliced ​​joint features are then sequentially processed through two cascaded channel fusion submodules (CF). Each submodule first compresses the number of channels to reduce redundant features, then filters effective channels through nonlinear activation, and finally reconstructs the feature dimensions through channel restoration operations, thereby gradually strengthening the channel-level correlation between radar and meteorological features.

[0041] The CF module structure is represented as follows: ; Where W1 and W2 are 1×1 convolution kernels used for channel compression or restoration; b1 and b2 are bias terms. is the ReLU activation function; 'a' is the input variable of the ReLU activation function; Y represents the output of the dual CF module; Y is the final output feature of the CF module.

[0042] More specifically, the joint features are processed once by two cascaded channel fusion submodules to obtain updated fused features, including: Within the first channel fusion submodule, the joint features are compressed through 1×1 convolution to obtain the first channel compressed features. The first activated feature is obtained by performing a nonlinear transformation on the compressed feature of the first channel using a nonlinear activation function. Channel recovery features are obtained by performing channel recovery on the first activation feature through 1×1 convolution. Within the second channel fusion submodule, the channel recovery features are compressed using a 1×1 convolution to obtain the second channel compressed features. The second channel compressed feature is obtained by performing a nonlinear transformation on the nonlinear activation function; The second activation feature is restored by channel recovery through 1×1 convolution to obtain the updated fused feature.

[0043] Next, the fusion feature f will be updated. cf2 Divide the data into two parts according to the channel dimension. For example, the updated fused feature f after two channel interactions is... cf2 When the number of channels in the channel dimension is C, the first C / 2 channels are used as the feature information of the phased array radar data, and the corresponding feature is represented as follows: The last C / 2 channels are used as the feature information of the meteorological grid real-time data, and the corresponding features are represented as follows: .Right now .

[0044] Based on respectively and The Sigmoid activation function is used to generate deep fusion dynamic weight matrices and shallow fusion dynamic weight matrices. Both deep fusion dynamic weight matrices contain corresponding weights for multi-scale phased array radar data features and corresponding weights for multi-scale meteorological grid real-time data features. The difference between the two is that the weight values ​​are different. , ; in, w1 and w2 are the Sigmoid activation function and the Sigmoid dynamic channel weights.

[0045] In this embodiment, based on a deep fusion dynamic weight matrix, the multi-scale phased array radar data features and the multi-scale meteorological grid real-time data features are weighted and superimposed point by point to obtain the multi-scale fused features, including: The deep fusion dynamic weight matrix is ​​multiplied element-wise with the corresponding multi-scale phased array radar data features and multi-scale meteorological grid real-time data features, and then added to obtain the weighted fusion output. Neutral mean baseline characteristics for calculating the features of multi-scale phased array radar data and multi-scale meteorological grid real-time data; The weighted fusion output is weighted and summed with the neutral average baseline features to obtain the multi-scale fusion features.

[0046] The deep fusion dynamic weight matrix achieves adaptive channel-level weighted fusion by multiplying and summing bimodal features element-wise. This allows the model to flexibly adjust the contribution ratio of local phased array radar texture and global meteorological physical field according to the current precipitation scenario, significantly improving the discriminative ability of the fused features. Introducing neutral average baseline features as residuals and weighting the weighted fusion output with the baseline features effectively mitigates the feature distribution shift that dynamic weights may cause, preserving the original average information of the bimodals as a stable reference and enhancing the robustness of the fusion process. Learnable weighted summation parameters further endow the model with the ability to seek the optimal balance between maintaining information integrity and highlighting key features, avoiding the loss of effective information caused by extreme weights. This design combines adaptability, stability, and information preservation, and compared to simple splicing or fixed-weight fusion, it can output more expressive multi-scale fused features.

[0047] Specifically, for deep fusion, the dynamic weights w1 and w2 mentioned above are directly used to perform point-by-point weighted superposition of the original pre-splitting features X1 and X2, while a neutral average term is added as the residual baseline to obtain the multi-scale fused feature p. The calculation process is as follows: ; Where ⊙ represents element-wise multiplication; f fused The output after weighted fusion; fbase β represents the neutral average baseline feature of the two modalities; β is the learnable residual weight parameter; and p is the output feature after cross-modal residual fusion.

[0048] For shallow fusion, a similar but independent set of dynamic weight matrices is used (in actual implementation, shallow fusion can reuse the same set of weights or generate them through additional lightweight branches; here, another set of shallow dynamic weights w1' and w2' is used). X1 and X2 are weighted and superimposed without adding a residual baseline to obtain the shallow fusion feature p. shallow =X1⊙w1'+X2⊙w2'. The shallow fusion features maintain high spatial resolution and are used for skip connections in the subsequent upsampling stage. Thus, the cross-modal residual embedding module simultaneously outputs multi-scale fusion features and shallow fusion features.

[0049] In some embodiments, it also includes: In the dual-branch feature extraction process, a scene classification branch is added in parallel. This branch receives the feature map output from the second convolutional layer of the dual-branch network structure as input, performs global average pooling on the feature map to obtain a global feature vector, and then passes the global feature vector through two fully connected layers. The first fully connected layer compresses the feature dimension to one-quarter of the original dimension and activates it with ReLU. The second fully connected layer outputs a two-dimensional scene category score, which is then normalized by the Softmax function to obtain the probability values ​​of layered cloud rainfall scene and convective cloud rainfall scene. The category with the larger probability value is selected as the field of the current input sample. The scene classification results are output along with scene category weight coefficients. Among them, the weight enhancement coefficient for the stratiform cloud rainfall scene corresponding to the meteorological grid real-time data is a1, and the weight reduction coefficient for the phased array radar data is a2. The weight enhancement coefficient for the convective cloud rainfall scene corresponding to the phased array radar data is a3, and the weight reduction coefficient for the meteorological grid real-time data is a4. a1, a2, a3, and a4 are preset hyperparameters. Then, the scene category weight coefficients are passed to the cross-modal residual embedding module. That is, in the point-by-point weighted superposition operation, the scene category weight coefficients are multiplied by the features of the corresponding modality and then added together to obtain the scene adaptive multi-scale fusion features.

[0050] This embodiment utilizes the mid-layer feature map output from the second convolutional layer for scene classification. This layer's features contain structural information without being overly abstract, effectively distinguishing between stratiform cloud precipitation and large-scale uniform texture features, as well as the localized, intense structure of convective cloud precipitation. Global average pooling compresses spatial information into a global feature vector, reducing computational cost. Two fully connected layers and a Softmax layer then output the probabilities of stratiform and convective clouds, achieving automatic identification of precipitation types in the input samples. Different modal weighting coefficients are preset for different precipitation types. In the stratiform cloud scenario, the weight of meteorological grid data is increased while the weight of phased array radar data is decreased, leveraging the global physical field advantage of meteorological data. In the convective cloud scenario, the weight of phased array radar data is increased while the weight of meteorological grid data is decreased, utilizing the radar data's ability to capture localized fine structures. The weight coefficients generated from the scene classification results are passed to the cross-modal residual embedding module to replace the original dynamic weight matrix. In the point-by-point weighted superposition operation, the features of the two modes are subjected to scene-adaptive weight modulation, which enables the model to dynamically adjust the fusion strategy according to the current precipitation type, significantly improving the adaptability and generalization performance of different weather types.

[0051] Specifically, in the dual-branch feature extraction process, a scene classification branch is added in parallel. This branch receives the feature map output from the second convolutional layer of the dual-branch network structure as input. This feature map simultaneously contains mid-level features from both the phased array radar data branch and the meteorological grid real-time data branch. Specifically, the feature maps output from the second convolutional layer of the two branches are concatenated along the channel dimension and then fed into the scene classification branch. The scene classification branch first performs global average pooling on the input feature map, compressing the spatial dimension of each channel into a scalar to obtain a global feature vector. The length of this vector is equal to the number of channels after concatenation. This global feature vector is then passed through two fully connected layers: the first fully connected layer compresses the feature dimension to one-quarter of the original dimension and performs a non-linear transformation using the ReLU activation function to output intermediate features; the second fully connected layer maps the intermediate features into a two-dimensional score vector, corresponding to the unnormalized scores of the stratiform cloud rainfall scene and the convective cloud rainfall scene, respectively. Subsequently, the two-dimensional score vector is normalized to the probability values ​​of the two scenes using the Softmax function, representing the probability that the current input sample belongs to the stratiform cloud rainfall scene and the probability that it belongs to the convective cloud rainfall scene, respectively. The category with the higher probability value is selected as the scene classification result for the current input sample. Based on the classification result, scene category weight coefficients are output: if the scene is classified as stratiform cloud precipitation, the weight enhancement coefficient for meteorological grid real-time data is set to a1, and the weight reduction coefficient for phased array radar data is set to a2; if the scene is classified as convective cloud precipitation, the weight enhancement coefficient for phased array radar data is set to a3, and the weight reduction coefficient for meteorological grid real-time data is set to a4. Here, a1, a2, a3, and a4 are pre-defined hyperparameters, satisfying a1 > 1, a2 < 1, a3 > 1, and a4 < 1, to reflect the bias towards different modalities under different scenarios.

[0052] Next, the scene category weight coefficients output by the scene classification branch are passed to the cross-modal residual embedding module. In the point-by-point weighted superposition operation of the cross-modal residual embedding module, specifically, for the layered cloud precipitation scene, the meteorological grid real-time data features are multiplied by an enhancement coefficient a1, and the phased array radar data features are multiplied by a reduction coefficient a2, then element-by-element summation is performed, i.e., f fused = a1·X1+ a2·X2; For convective cloud precipitation scenarios, f fused = a3·X2 + a4·X1. Simultaneously, the neutral mean baseline term is retained as a residual, ultimately yielding a scene-adaptive multi-scale fusion feature. This fusion feature reflects the optimal modal contribution ratio under the current precipitation type while maintaining numerical stability, and is output to step five for subsequent upsampling processing.

[0053] S105: The multi-scale fusion features are upsampled step by step by the enhanced upsampling channel attention module, and skip connections are made between each level and the corresponding shallow fusion features to obtain the rainfall correction result. In this embodiment, please refer to Figure 6 By using the Enhanced Upsampling Channel Attention Module (EUCA) to progressively upsample multi-scale fused features and performing skip connections between each level and the corresponding shallow fused features, the rainfall correction results include: The spatial resolution of the multi-scale fusion features is restored step by step by the deconvolution structure to obtain upsampled features. During the upsampling process, the global average pooling features and the channel weights generated by the multilayer perceptron are calculated. The channel weights are multiplied element by element with the upsampled features to obtain high-resolution features. The reconstructed features are obtained by reconstructing the local spatial structure of high-resolution features and shallow fusion features of corresponding layers through the convolution thinning module. An upsampling operation is performed on the reconstructed features to obtain rainfall correction results.

[0054] This embodiment restores spatial resolution stepwise through a deconvolutional structure. By combining global average pooling with channel weights generated by a multilayer perceptron, it achieves adaptive enhancement of important channels in high-resolution features, effectively improving the reconstruction quality of precipitation field details. The enhanced high-resolution features are then linked to shallow fused features at corresponding levels via a skip connection, and a convolutional thinning module is used for local structural reconstruction. This preserves shallow spatial details while incorporating deep semantic information, suppressing noise propagation and information loss during upsampling. Finally, the reconstructed features are upsampled to output precipitation correction results, achieving an end-to-end mapping from multi-scale fused features to a high-precision precipitation field.

[0055] Specifically, the multi-scale fused features output from the cross-modal residual embedding module are input to the enhanced upsampling channel attention module for progressive upsampling. Let the current level's input feature be X, and the upsampling factor be s. Spatial resolution is restored through a deconvolution structure. The deconvolution operation uses a transposed convolution kernel, and the output features... ; Among them, H t W is the upsampled feature tensor. deconv X represents the transposed convolution weights; X represents the input features. For the upsampling operator, the spatial resolution is magnified by a factor of s.

[0056] Obtain upsampled features H t Then, the global average pooling feature is calculated along the channel dimension. For the c-th channel, the formula for calculating the global average pooling feature is: ; Where, q c is the globally averaged feature corresponding to the c-th channel; sH and sW are the height and width of the feature map after upsampling. For upsampling features in The value at the specified location; b is the batch index.

[0057] Then all channels' q c Concatenate them into vector q; ; Where T is the vector transpose.

[0058] The vector q is then fed into a multilayer perceptron (MLP). First, the number of channels is compressed to 1 / r of the original number (where r is the compression ratio) through the first fully connected layer. Then, the hidden features are obtained by ReLU activation. Where δ is the ReLU activation function; then, the number of channels is restored to the original number through a second fully connected layer, and channel attention weights are generated by Sigmoid activation. The channel weights α and the upsampled features H are then compared. t Element-wise multiplication yields channel-enhanced high-resolution features. This process enables adaptive enhancement of important channels in high-resolution feature maps.

[0059] in, is the Sigmoid activation function; W3 and W4 are the weight matrices of the fully connected layer in the MLP; b3 and b4 are the bias terms. This represents the final output feature of the EUCA module; · indicates element-wise multiplication.

[0060] Before performing local spatial structure reconstruction, the high-resolution features of the corresponding level output by the enhanced upsampling channel attention module and the shallow fusion features of the same resolution level generated by the cross-modal residual embedding module are first aligned to unify the spatial grid size and channel dimension of the two types of features, ensuring that the features are completely matched in spatial location and number of channels.

[0061] After alignment, the high-resolution features and the shallow fused features are concatenated along the channel dimension to form a joint feature tensor. This joint feature is then input into a convolutional thinning module, which consists of two convolutional layers. The first layer uses a 3×3 convolution with a stride of 1 and padding of 1, maintaining the same number of output channels. Following the convolution is a batch normalization layer and a ReLU activation function, used to initially fuse the two types of features and suppress noise. The second layer also uses a 3×3 convolution, with the number of output channels set to the preset number of precipitation field channels. Following the convolution is only a batch normalization layer, without adding an activation function to preserve the linear output range. After these two convolutional layers, the output reconstructed features have clear precipitation texture boundaries and gradient changes, achieving an organic fusion of deep semantic information and shallow spatial details.

[0062] Subsequently, a final upsampling operation is performed on the reconstructed features. A deconvolutional structure is used to progressively restore the spatial resolution, with a deconvolution kernel size of 4×4, a stride of 2, and padding of 1, enlarging the spatial size of the reconstructed features to a resolution completely consistent with the original input image. No additional attention mechanism is introduced during the upsampling process; the spatial dimension expansion is achieved solely through learnable transposed convolution parameters. The final output feature map is then passed through a 1×1 convolutional layer to map the number of channels to a single channel, yielding the rainfall correction result. The spatial scale of this result is the same as the input data in the original sample pair, and each pixel value represents the corrected rainfall intensity at that location.

[0063] S106: Obtain measured rainfall data, and iteratively train the dual-branch network structure, the downsampling attention module, the cross-modal residual embedding module, and the enhanced upsampling channel attention module based on the measured rainfall data and the rainfall correction result to obtain an intelligent rainfall data correction model; Specifically, the error between the rainfall correction result and the measured rainfall data is calculated, using the mean absolute error or mean square error as the loss function, with the specific expression being: ; Where L1 is the mean absolute error, also known as L1 loss; MSE is the mean squared error, also known as L2 loss; and n is the number of samples. y represents the model prediction for the i-th sample (i.e., the rainfall correction result); i This represents the actual observed value (i.e., measured rainfall data) for the i-th sample.

[0064] Subsequently, the backpropagation algorithm is used to calculate gradients for all learnable parameters in each module, including convolutional kernel weights, bias parameters, weight matrices in the attention mechanism, and learnable residual parameters, based on the loss function value. The calculated gradient values ​​are then backpropagated to the corresponding layers of each module using an optimization algorithm. All parameters are iteratively updated. After each parameter update, preprocessed radar and meteorological data samples are input into the updated model to regenerate the rainfall correction result and calculate the loss function value. This process is repeated. During iterative training, the trend of the loss function value is continuously monitored. When the loss function value stabilizes and shows no significant decrease after multiple iterations, it is determined that the loss function has converged, and the iterative training process is immediately terminated. At this point, the parameters of each module have been optimally optimized. The trained dual-branch network structure, downsampling attention module, cross-modal residual embedding module, and enhanced upsampling channel attention module are integrated to form a complete intelligent rainfall data correction model.

[0065] In some embodiments, the method further includes: obtaining rainfall correction results; calculating the optical flow field constraint loss between adjacent time points on the rainfall correction results in a time series to constrain the continuous evolution of rainfall intensity and prevent physically unreasonable abrupt changes, thereby obtaining a time continuity constraint loss value; obtaining the reflectivity factor in the multi-scale phased array radar data features; and constructing a regularization loss term based on the ZR relationship according to the reflectivity factor and the rainfall intensity in the rainfall correction results, wherein the ZR relationship is expressed as Z=a*R. b a and b are empirical parameters. The physical consistency loss value is obtained by calculating the difference between the rainfall intensity retrieved by the reflectivity factor and the rainfall correction result. Then, the time continuity constraint loss value and the physical consistency loss value are weighted and added together with the data-driven loss value based on L1 and MSE that has been constructed to obtain the total loss function, which is used for subsequent network parameter iterative updates.

[0066] This embodiment applies a continuity penalty to the rainfall correction results at adjacent time points in the time series through optical flow field constraint loss, forcing the rainfall field output by the model to satisfy the laws of fluid motion in the time dimension. This effectively suppresses inter-frame abrupt changes or non-physical flickering phenomena, improving the temporal smoothness and realism of the precipitation evolution process. A ZR relation regularization loss term is introduced to establish a physical constraint between the radar reflectivity factor and the rainfall intensity predicted by the model. This ensures that the network output does not statistically violate the classical precipitation inversion formula, avoiding inversion bias or "illusion" rainfall caused by purely data-driven approaches. The weighted summation of the above two physical constraints with the data-driven L1 and MSE losses forms the total loss function, achieving joint optimization between data fitting and physical consistency. This significantly enhances the prediction stability and reliability of the model in generalized scenarios such as extreme precipitation and complex terrain, while also making the training process more interpretable.

[0067] Specifically, the rainfall correction result output by the enhanced upsampling channel attention module is obtained, corresponding to the output sequence of multiple consecutive time steps. The optical flow field constraint loss between adjacent time steps is calculated on the rainfall correction result over time. Specifically, a gradient-based optical flow method is used to estimate the motion vector of the rainfall intensity field between two adjacent frames, and the difference between the predicted motion vector and the theoretical smoothness constraint is calculated to obtain the temporal continuity constraint loss value L. flow This loss forces the model to generate rainfall fields at adjacent time points that conform to the continuously evolving fluid motion laws in terms of spatial structure, effectively suppressing non-physical abrupt changes or flickering phenomena.

[0068] Secondly, the multi-scale phased array radar data features output by the downsampling attention module are obtained, and the original reflectivity factor Z is extracted from them. Based on the classical ZR relationship, Z = a*R b Where a and b are empirical parameters related to precipitation type, the reflectivity factor is inverted into a physical reference rainfall intensity R using this relationship.phys = (Z / a) 1 / b Simultaneously, the rainfall intensity R in the rainfall correction results is obtained. pred The mean square error between the two is calculated as the physical consistency loss value L. zr The loss term constrains the rainfall intensity output by the network to ensure that the physical mechanism does not violate the fundamental relationship between radar reflectivity and precipitation intensity, thus reducing non-physical inversion biases caused by purely data-driven approaches.

[0069] Finally, the time continuity constraint loss value L flow Loss value L of physical consistency zr Multiply by their respective weighting coefficients λ flow and λ zr Then, a weighted sum is performed, and then it is added to the previously constructed data-driven loss value L based on L1 and MSE. data Summing yields the total loss function L. total =L data +λ flow ·L flow +λ zr ·L zr The total loss function is used as the optimization objective for backpropagation, and is subsequently applied to iterative updates of network parameters. By jointly optimizing data fitting and physical consistency, the model can generate rainfall correction results that are reasonable and reliable in terms of both temporal evolution and radar inversion physical relationships while maintaining high accuracy.

[0070] S107: Obtain the X-band phased array radar reflectivity data to be corrected and the physical field data of the meteorological model, and input the X-band phased array radar reflectivity data to be corrected and the physical field data of the meteorological model into the rainfall data intelligent correction model to obtain the corrected rainfall data.

[0071] Specifically, the X-band phased array radar reflectivity data and meteorological model physical field data to be corrected are acquired. The data to be corrected comes from real-time observations or historical archives, covering the target area, with a temporal resolution of minutes and a spatial resolution consistent with the training data. The two types of data to be corrected undergo the same preprocessing operations as in the training phase: first, spatiotemporal matching is performed, using the time of the meteorological grid real-time data as the benchmark, selecting the closest radar scan time within the time window, and resampling the radar reflectivity data onto the horizontal grid points of the meteorological data using bilinear interpolation; then, normalization is performed, mapping the radar reflectivity values ​​to the [0, 1] interval according to the normalization parameters used during training, and standardizing the meteorological variables according to their respective physical ranges using Z-score to eliminate dimensional differences. After preprocessing, the processed radar reflectivity feature map and multi-channel meteorological physical field feature map are used as sample pairs and input into the trained intelligent correction model for rainfall data. This model consists of a dual-branch network structure, a downsampling attention module, a cross-modal residual embedding module, and an enhanced upsampling channel attention module connected in series. Data propagation in the model: a dual-branch network extracts features from radar and meteorological data respectively; a downsampling attention module extracts multi-scale spatial features; a cross-modal residual embedding module performs deep and shallow fusion; and an enhanced upsampling channel attention module upsamples level by level and reconstructs a high-resolution precipitation field. The model's final output is a single-channel precipitation intensity map with the same spatial size as the input, where each pixel value represents the corrected precipitation intensity at the corresponding location. The output is then denormalized to restore it to actual rainfall units, yielding the final corrected precipitation data. This data can be directly used for meteorological monitoring, short-term forecasting, hydrological simulation, or disaster prevention and mitigation decision support.

[0072] This embodiment constructs a dual-branch network to extract local high-resolution texture features from phased array radar and global physical features from meteorological grid data. A downsampling attention module is used to extract multi-scale spatial structure, and a cross-modal residual embedding module is employed to achieve deep fusion and dynamic weighting of the two types of data at different levels. This effectively overcomes the problems of insufficient information from a single data source and the difficulty in deep coupling of multi-source data. Simultaneously, an enhanced upsampling channel attention module, combined with a skip connection mechanism, reconstructs a high-resolution precipitation field, suppressing attenuation and clutter interference from X-band radar. This significantly improves the spatial resolution and correction accuracy of precipitation data, resulting in more accurate, stable, and generalized precipitation estimation results.

[0073] Please see Figure 7 This embodiment provides a rainfall intelligent correction device 200, including: The preprocessing module 201 is used to acquire reflectivity data of X-band phased array radar and physical field data of meteorological models, and to preprocess the reflectivity data and physical field data to obtain sample pairs. Extraction module 202 is used to extract features from the sample pairs using a dual-branch network structure to obtain phased array radar data features and meteorological grid real-time data features, respectively. Input module 203 is used to input the phased array radar data features and the meteorological grid real-time data features into the downsampling attention module for multi-scale spatial feature extraction, so as to obtain multi-scale phased array radar data features and multi-scale meteorological grid real-time data features respectively; The fusion module 204 is used to perform deep and shallow fusion at different levels based on the multi-scale phased array radar data features and the multi-scale meteorological grid real-time data features, through the cross-modal residual embedding module, to obtain multi-scale fusion features and shallow fusion features respectively. Connection module 205 is used to upsample the multi-scale fusion features step by step through the enhanced upsampling channel attention module, and to make skip connections with the corresponding shallow fusion features at each level to obtain rainfall correction results; Training module 206 is used to acquire measured rainfall data and iteratively train the dual-branch network structure, the downsampling attention module, the cross-modal residual embedding module, and the enhanced upsampling channel attention module based on the measured rainfall data and the rainfall correction results to obtain an intelligent rainfall data correction model. The acquisition module 207 is used to acquire the X-band phased array radar reflectivity data to be corrected and the physical field data of the meteorological model, and input the X-band phased array radar reflectivity data to be corrected and the physical field data of the meteorological model into the rainfall data intelligent correction model to obtain the corrected rainfall data.

[0074] Furthermore, the extraction module 202 includes: The representation pair building unit is used to abstract the sample pair step by step through a three-layer convolutional structure to construct a primary spatial structure representation pair. The dimensionality reduction unit is used to perform dimensionality reduction processing on the primary spatial structure representation pair using 1×1 convolution to obtain compressed feature pairs; An evaluation unit is used to adaptively evaluate the compressed feature pairs using a channel attention mechanism to obtain evaluation feature pairs. The feature extraction unit is used to extract features from the evaluation feature pairs through 3×3 convolution and downsampling to obtain phased array radar data features and meteorological grid real-time data features, respectively.

[0075] Furthermore, the input module 203 includes: The downsampling unit is used to perform downsampling based on the phased array radar data features and the meteorological grid real-time data features, through stride convolution, to obtain radar downsampling feature maps and meteorological downsampling feature maps, respectively. A pooling unit is used to perform channel max pooling on the radar downsampled feature map and the meteorological downsampled feature map to obtain radar spatial structure features and meteorological spatial structure features, respectively. The activation unit is used to generate radar spatial attention map and meteorological spatial attention map respectively based on the radar spatial structure features and meteorological spatial structure features, using 3×3 convolution and Sigmoid activation. The first weighting unit is used to perform element-wise weighting of the radar spatial attention map and the radar downsampled feature map to obtain multi-scale phased array radar data features. The second weighting unit is used to perform element-wise weighting of the meteorological spatial attention map and the meteorological downsampling feature map to obtain multi-scale meteorological grid real-time data features.

[0076] Furthermore, the fusion module 204 includes: The splicing unit is used to splice the multi-scale phased array radar data features and the multi-scale meteorological grid real-time data features in the channel dimension to obtain joint features; The fusion processing unit is used to perform channel fusion processing on each of the joint features once through two cascaded channel fusion submodules to obtain updated fused features; The weight generation unit is used to divide the updated fusion features into phased array radar data feature information and meteorological grid real-time data feature information according to channels, and generate deep fusion dynamic weight matrix and shallow fusion dynamic weight matrix through activation function. The first weighted superposition unit is used to perform point-by-point weighted superposition of the multi-scale phased array radar data features and the multi-scale meteorological grid real-time data features based on the deep fusion dynamic weight matrix to obtain multi-scale fusion features. The second weighted overlay unit is used to perform point-by-point weighted overlay of the multi-scale phased array radar data features and the multi-scale meteorological grid real-time data features based on the shallow fusion dynamic weight matrix to obtain shallow fusion features.

[0077] Furthermore, the fusion processing unit includes: The first compression subunit is used to perform channel compression on the joint features through 1×1 convolution within the first channel fusion submodule to obtain the first channel compressed feature; The first transformation subunit is used to perform a nonlinear transformation on the compression feature of the first channel through a nonlinear activation function to obtain the first activation feature; The first recovery subunit is used to perform channel recovery on the first activation feature through a 1×1 convolution to obtain the channel recovery feature; The second compression subunit is used to perform channel compression on the channel recovery features through 1×1 convolution within the second channel fusion submodule to obtain the second channel compressed features; The second transformation subunit is used to perform a nonlinear transformation on the second channel compression feature through a nonlinear activation function to obtain the second activation feature; The second recovery subunit is used to perform channel recovery on the second activation feature through 1×1 convolution to obtain the updated fused feature.

[0078] Furthermore, the first weighted stacking unit includes: The addition subunit is used to multiply the deep fusion dynamic weight matrix element-wise with the corresponding multi-scale phased array radar data features and the multi-scale meteorological grid real-time data features, and then add them to obtain a weighted fusion output. The feature calculation subunit is used to calculate the neutral average baseline features of the multi-scale phased array radar data features and the multi-scale meteorological grid real-time data features. The weighted summation subunit is used to perform a weighted summation of the weighted fusion output and the neutral average baseline features to obtain multi-scale fusion features.

[0079] Furthermore, the connection module 205 includes: The average pooling unit is used to restore the spatial resolution of the multi-scale fusion features step by step through the deconvolution structure to obtain upsampled features. During the upsampling process, the global average pooling features and the channel weights generated by the multilayer perceptron are calculated. The channel weights are multiplied element by element with the upsampled features to obtain high-resolution features. The structure reconstruction unit is used to perform local spatial structure reconstruction on the high-resolution features and the corresponding shallow fusion features through the convolutional thinning module to obtain the reconstructed features; The result acquisition unit is used to perform an upsampling operation on the reconstructed features to obtain rainfall correction results.

[0080] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed, can implement the methods provided in the above embodiments. The storage medium may include various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0081] The present invention also provides a computer device, which may include a memory and a processor. The memory stores a computer program, and when the processor calls the computer program in the memory, it can implement the methods provided in the above embodiments. Of course, the computer device may also include various network interfaces, power supplies, and other components.

[0082] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section. It should be noted that those skilled in the art can make various improvements and modifications to this invention without departing from its principles, and these improvements and modifications also fall within the protection scope of the claims of this invention.

[0083] It should also be noted that, in this specification, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusivity.

[0084] The term "comprises" implies that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprises a..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

Claims

1. A rainfall intelligent correction method, characterized in that, include: Acquire reflectivity data from an X-band phased array radar and physical field data from a weather model, and preprocess the reflectivity data and physical field data to obtain sample pairs; A dual-branch network structure is used to extract features from the sample pairs to obtain phased array radar data features and meteorological grid real-time data features, respectively. The phased array radar data features and the meteorological grid real-time data features are input into the downsampling attention module for multi-scale spatial feature extraction, resulting in multi-scale phased array radar data features and multi-scale meteorological grid real-time data features, respectively. Based on the multi-scale phased array radar data features and the multi-scale meteorological grid real-time data features, deep and shallow fusion is performed at different levels through a cross-modal residual embedding module to obtain multi-scale fusion features and shallow fusion features respectively. The multi-scale fusion features are upsampled step by step by enhancing the attention module of the upsampling channel, and skip connections are made between each level and the corresponding shallow fusion features to obtain the rainfall correction result. Acquire measured rainfall data, and iteratively train the dual-branch network structure, the downsampling attention module, the cross-modal residual embedding module, and the enhanced upsampling channel attention module based on the measured rainfall data and the rainfall correction results to obtain an intelligent rainfall data correction model; The reflectivity data of the X-band phased array radar to be corrected and the physical field data of the meteorological model are obtained, and the reflectivity data of the X-band phased array radar to be corrected and the physical field data of the meteorological model are input into the intelligent correction model of the rainfall data to obtain the corrected rainfall data.

2. The intelligent rainfall correction method according to claim 1, characterized in that, The feature extraction of the sample pairs using a dual-branch network structure, resulting in phased array radar data features and meteorological grid real-time data features, includes: The sample pairs are abstracted step by step using a three-layer convolutional structure to construct primary spatial structure representation pairs. The primary spatial structure representation pairs are reduced in dimensionality using 1×1 convolution to obtain compressed feature pairs. The compressed feature pairs are adaptively evaluated using a channel attention mechanism to obtain evaluated feature pairs; Feature extraction was performed on the evaluation feature pairs using 3×3 convolution and downsampling to obtain phased array radar data features and meteorological grid real-time data features, respectively.

3. The intelligent rainfall correction method according to claim 1, characterized in that, The step of inputting the phased array radar data features and the meteorological grid real-time data features into the downsampling attention module for multi-scale spatial feature extraction, and obtaining the multi-scale phased array radar data features and multi-scale meteorological grid real-time data features respectively, includes: Based on the phased array radar data features and the meteorological grid real-time data features, downsampling is performed through step convolution to obtain radar downsampling feature maps and meteorological downsampling feature maps, respectively. Channel max pooling is performed on the radar downsampled feature map and the meteorological downsampled feature map to obtain the radar spatial structure features and the meteorological spatial structure features, respectively. Based on the aforementioned radar spatial structure features and meteorological spatial structure features, radar spatial attention maps and meteorological spatial attention maps are generated using 3×3 convolution and Sigmoid activation, respectively. The radar spatial attention map and the radar downsampled feature map are weighted element-wise to obtain multi-scale phased array radar data features. The meteorological spatial attention map and the meteorological downsampling feature map are weighted element by element to obtain the multi-scale meteorological grid real-time data features.

4. The intelligent rainfall correction method according to claim 1, characterized in that, The multi-scale phased array radar data features and the multi-scale meteorological grid real-time data features are used to perform deep and shallow fusion at different levels through a cross-modal residual embedding module to obtain multi-scale fusion features and shallow fusion features, respectively, including: The features of the multi-scale phased array radar data and the features of the multi-scale meteorological grid real-time data are concatenated along the channel dimension to obtain joint features; The joint features are processed once by two cascaded channel fusion submodules to obtain updated fused features; The updated fusion features are divided into phased array radar data feature information and meteorological grid real-time data feature information according to channels, and deep fusion dynamic weight matrix and shallow fusion dynamic weight matrix are generated by activation function. Based on the deep fusion dynamic weight matrix, the multi-scale phased array radar data features and the multi-scale meteorological grid real-time data features are weighted and superimposed point by point to obtain multi-scale fusion features. Based on the shallow fusion dynamic weight matrix, the multi-scale phased array radar data features and the multi-scale meteorological grid real-time data features are weighted and superimposed point by point to obtain shallow fusion features.

5. The intelligent rainfall correction method according to claim 4, characterized in that, The process of performing channel fusion processing on each of the joint features once through two cascaded channel fusion submodules to obtain updated fused features includes: Within the first channel fusion submodule, the joint features are compressed using a 1×1 convolution to obtain the first channel compressed features. The first activation feature is obtained by performing a nonlinear transformation on the compressed feature of the first channel using a nonlinear activation function. Channel recovery features are obtained by performing channel recovery on the first activation feature through 1×1 convolution; Within the second channel fusion submodule, the channel recovery features are compressed using a 1×1 convolution to obtain the second channel compressed features. The second channel compression feature is obtained by performing a nonlinear transformation on the nonlinear activation function; The second activation feature is channel restored by 1×1 convolution to obtain the updated fused feature.

6. The intelligent rainfall correction method according to claim 4, characterized in that, The multi-scale fused features are obtained by weighting and superimposing the multi-scale phased array radar data features and the multi-scale meteorological grid real-time data features point-by-point based on the deep fusion dynamic weight matrix, including: The deep fusion dynamic weight matrix is ​​multiplied element-wise with the corresponding multi-scale phased array radar data features and multi-scale meteorological grid real-time data features, and then added together to obtain the weighted fusion output; Calculate the neutral mean baseline characteristics of the multi-scale phased array radar data features and the multi-scale meteorological grid real-time data features; The weighted fusion output is weighted and summed with the neutral average baseline features to obtain the multi-scale fusion features.

7. The intelligent rainfall correction method according to claim 1, characterized in that, The step of upsampling the multi-scale fusion features through an enhanced upsampling channel attention module, and then performing skip connections between each level and the corresponding shallow fusion features to obtain the rainfall correction result includes: The spatial resolution of the multi-scale fusion features is restored step by step by a deconvolution structure to obtain upsampled features. During the upsampling process, the global average pooling features and the channel weights generated by the multilayer perceptron are calculated. The channel weights are then multiplied element by element by the upsampled features to obtain high-resolution features. The high-resolution features and the corresponding shallow fusion features are reconstructed locally using a convolutional thinning module to obtain the reconstructed features. An upsampling operation is performed on the reconstructed features to obtain rainfall correction results.

8. A rainfall intelligent correction device, characterized in that, include: The preprocessing module is used to acquire reflectivity data of X-band phased array radar and physical field data of meteorological models, and to preprocess the reflectivity data and physical field data to obtain sample pairs. The extraction module is used to extract features from the sample pairs using a dual-branch network structure, thereby obtaining phased array radar data features and meteorological grid real-time data features respectively. The input module is used to input the phased array radar data features and the meteorological grid real-time data features into the downsampling attention module for multi-scale spatial feature extraction, so as to obtain multi-scale phased array radar data features and multi-scale meteorological grid real-time data features respectively. The fusion module is used to perform deep and shallow fusion at different levels based on the multi-scale phased array radar data features and the multi-scale meteorological grid real-time data features, through the cross-modal residual embedding module, to obtain multi-scale fusion features and shallow fusion features respectively. The connection module is used to upsample the multi-scale fusion features step by step through the enhanced upsampling channel attention module, and to make skip connections with the corresponding shallow fusion features at each level to obtain the rainfall correction result. The training module is used to acquire measured rainfall data and iteratively train the dual-branch network structure, the downsampling attention module, the cross-modal residual embedding module, and the enhanced upsampling channel attention module based on the measured rainfall data and the rainfall correction results to obtain an intelligent rainfall data correction model. The acquisition module is used to acquire the X-band phased array radar reflectivity data to be corrected and the physical field data of the meteorological model, and input the X-band phased array radar reflectivity data to be corrected and the physical field data of the meteorological model into the intelligent correction model of the rainfall data to obtain the corrected rainfall data.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the intelligent rainfall correction method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, causes the processor to perform the intelligent rainfall correction method as described in any one of claims 1 to 7.

Images