Satellite-borne discrete multi-beam radar precipitation data reconstruction method, product and equipment

Abstract

The application provides a spaceborne discrete multi-beam radar precipitation data reconstruction method, product and equipment. The method comprises the following steps: obtaining a measured reflectivity factor through a spaceborne discrete multi-beam radar; obtaining brightness temperature data of different channels through a spaceborne microwave radiometer; inputting the reflectivity factor and the brightness temperature data into a pre-trained precipitation data reconstruction model to obtain a complete reflectivity factor profile, and the precipitation data reconstruction model is used for fusing the precipitation characteristics of the brightness temperature data on the basis of the reflectivity factor, reconstructing the precipitation data of the discrete beams and outputting the complete reflectivity factor profile, so as to complete the discrete beam observation. Through the method, the spatial vacancy of the discrete multi-beam observation can be effectively filled without increasing the hardware cost, the observation width of the precipitation profile is significantly improved, the physical continuity and the reconstruction precision of the precipitation structure are ensured, and reliable algorithm support is provided for a spaceborne low-cost high-temporal-and-spatial-resolution precipitation observation system.

Description

Technical Field

[0001] This invention relates to the field of meteorological monitoring technology, and in particular to a method, product, and equipment for reconstructing precipitation data from a spaceborne discrete multibeam radar. Background Technology

[0002] Currently, global-scale precipitation observations are generally only possible through satellite platforms. However, spaceborne visible light and infrared sensors can only detect clouds, not precipitation itself; spaceborne microwave radiometers cannot acquire the three-dimensional structure of precipitation systems, and detecting precipitation is extremely difficult under terrestrial conditions with high brightness temperature backgrounds and complex spatial variations; spaceborne radar can provide large-scale three-dimensional precipitation structure information, but its observation revisit cycle is long and its temporal resolution is low due to limitations in observation swath width and the number of devices. Therefore, there is an urgent need for a low-cost, multi-satellite network observation method to achieve high spatiotemporal resolution for global three-dimensional precipitation observation. Summary of the Invention

[0003] In view of the above problems, the present invention proposes a method, product and equipment for reconstructing precipitation data from spaceborne discrete multibeam radar that overcomes or at least partially solves the above problems.

[0004] One objective of this invention is to reconstruct three-dimensional precipitation data from discrete multibeam radar by combining active and passive microwave methods.

[0005] A further objective of this invention is to address the high cost problem of phased array radar and the limited swath width problem of single-beam radar.

[0006] Specifically, this invention provides a method for reconstructing precipitation data from a spaceborne discrete multibeam radar, comprising: The measured reflectivity factor was obtained using a spaceborne discrete multibeam radar. Brightness temperature data for different channels were obtained using a spaceborne microwave radiometer; Reflectivity factor and brightness temperature data are input into a pre-trained precipitation data reconstruction model to reconstruct a complete reflectivity factor profile. The precipitation data reconstruction model is used to reconstruct precipitation data for discrete beams based on reflectivity factor and fuse precipitation characteristics of brightness temperature data, and output a complete reflectivity factor profile, thereby supplementing discrete beam observations.

[0007] Optionally, the training steps for the precipitation data reconstruction model include: Acquire reflectivity factor data and brightness temperature data for training; Spatiotemporal matching of reflectivity factor data and brightness temperature data is performed to obtain reflectivity factor training data and brightness temperature training data. Three-dimensional tensor training sample data were constructed based on reflectivity factor training data and brightness temperature training data. Determine the basic architecture and loss function of the deep learning model, and introduce an attention mechanism; The deep learning model is trained based on the preset training strategy and three-dimensional tensor training sample data to obtain a precipitation data reconstruction model. The reconstruction results of the precipitation data reconstruction model were validated and evaluated, and channel contribution analysis was performed.

[0008] Optionally, the step of constructing three-dimensional tensor training sample data based on reflectivity factor training data and brightness temperature training data includes: Preprocess the training data for reflectivity factor and brightness temperature. The reflectivity factor training data and brightness temperature training data were standardized respectively to unify the data scale of different observation channels; The standardized reflectivity factor training data and brightness temperature training data are combined to obtain three-dimensional tensor sample data, which consists of three dimensions: number of channels, number of vertical layers, and number of cross-track beams. Using the center beam of the nadir point as a reference, a mask is constructed on the reflectivity factor training data channel according to a preset interval coefficient. The mask is used to simulate the observation geometry of discrete multibeam precipitation radar. Based on the mask, the standardized reflectivity factor training data is hollowed out to construct discrete beam observation data. The three-dimensional tensor sample data, mask, and discrete beam observation data are combined as the three-dimensional tensor training sample data.

[0009] Optionally, the preprocessing steps for the reflectivity factor training data and the brightness temperature training data include: Based on precipitation indicators and surface type, effective data screening is performed on the spatiotemporally matched reflectivity factor training data. Then, based on the sensitivity threshold and ground clutter height, the screened reflectivity factor training data is masked to retain effective reflectivity factor training data. The brightness temperature training data after spatiotemporal matching is extended along the vertical height, so that the extended brightness temperature training data has the same dimension as the reflectivity factor training data.

[0010] Optionally, the formula for standardizing the reflectivity factor training data is: in, This represents the standardized reflectivity factor training data. This represents the original reflectivity factor training data. This represents the minimum value in the original reflectivity factor training data. This represents the maximum value in the original reflectivity factor training data; The formula for standardizing the brightness temperature training data is: in, This represents the standardized brightness temperature training data. This represents the original brightness temperature training data. This represents the arithmetic mean of the brightness temperature training data. This represents the standard deviation of the brightness temperature training data; The three-dimensional tensor sample data is as follows: Where C represents the number of channels, H represents the number of vertical layers, and W represents the number of cross-track beams. The three-dimensional tensor sample data conforms to the structure of number of channels × number of vertical layers × number of beams. The mask construction formula is: in, This represents the value of the mask at the j-th cross-track beam position. This indicates the index of the center beam at the nadir point, and n represents the spacing coefficient of the discrete beam. The formula for constructing discrete beam observation data is: in, This represents the discrete beam observation data after the cut-out and filling process. This represents the training data for the reflectivity factor. Indicates the mask. This represents the fill value for the cut-out area. This represents the inverse matrix of the mask.

[0011] Optionally, the steps of determining the infrastructure and loss function of the deep learning model and introducing an attention mechanism include: The basic architecture was determined to be a convolutional neural network architecture; Set the input to the convolutional neural network architecture as: in To input 3D tensor sample data for a convolutional neural network architecture, For discrete beam observation data, For the mask, The brightness temperature training data is after standardization. The output of the convolutional neural network architecture is set as follows: in, This is the reconstructed complete reflectivity factor profile. For mapping functions of deep learning models; The loss function is: in, This is the total loss value. This is the global error value. and The preset weighting coefficients, This represents the weighted error value for the precipitation area. For cross-track gradient constraints; The formula for calculating the global error value is: The predicted values ​​of the complete reflectance factor profile output by the model are used to reconstruct precipitation data. This represents the true value of the complete reflectivity factor profile. This represents the average absolute error between the calculated true value and the predicted value. The formula for calculating the weighted error value of precipitation areas is: in, This indicates that only the absolute errors between data points in the precipitation region are summed. This represents the weighted average absolute error in calculating precipitation over a given area. The formula for calculating the trans-track gradient constraint is: in, This is a gradient operator along the trans-track direction, used to calculate the rate of change of the numerical value along the trans-track beam direction. This represents the prediction gradient of the predicted value along the trans-track direction. This represents the true gradient of the true value along the trans-track direction. This represents the summation of the absolute gradient errors between all predicted gradients and the true gradients; The attention mechanism is a channel attention mechanism, used to adaptively allocate channel weights to the intermediate feature tensors generated by the intermediate layers of a deep learning model. The intermediate feature tensors are... in, Indicates the number of intermediate feature channels. and These represent the feature dimensions of the intermediate feature tensor in the vertical direction and the cross-track direction, respectively; Global average pooling is performed on the c-th feature channel to obtain the channel descriptor. The formula for calculating the channel descriptor is: Channel attention weights are generated based on channel descriptors. The formula for calculating channel attention weights is as follows: Where s is a vector composed of the descriptors of each channel. and For learnable parameters, Represents a non-linear activation function. This represents the Sigmoid activation function, where w is the channel attention weight vector; We obtain the following by weighting the c-th feature channel: in, This represents the two-dimensional spatial feature map of the c-th feature channel. This represents the attention weight of the c-th feature channel. This represents the feature map of the c-th feature channel after attention weighting. This indicates element-wise multiplication or channel-based broadcast multiplication.

[0012] Optionally, the steps of training the deep learning model according to a preset training strategy and three-dimensional tensor training sample data to obtain a precipitation data reconstruction model include: The 3D tensor training sample data is divided into a training set and a validation set. Supervised learning is employed to iteratively train the deep learning model based on the training set; An adaptive learning rate optimization algorithm is used to iteratively update the model parameters, and a learning rate decay strategy is used for training. Determine the batch size; During training, the loss value or evaluation index of the deep learning model on the validation set is monitored, and the model parameters corresponding to the optimal evaluation on the validation set are determined as the parameters of the precipitation data reconstruction model.

[0013] Optionally, the steps for validating and evaluating the reconstruction results of the precipitation data reconstruction model and conducting channel contribution analysis include: Based on preset statistical indicators, the accuracy of the filled areas and the overall profile of the precipitation data reconstruction model is quantitatively evaluated. Based on the vertical profile error distribution and filling area error characteristics of the reconstruction results, the relationship between error and ground clutter height, surface type and precipitation structure complexity is determined. The channel configuration of the precipitation data reconstruction model input was adjusted, and multiple sets of control experiments were designed to quantify the contribution of each channel to the reconstruction performance. The experimental results based on the control experiment explain the mechanism by which different observational information reconstructs precipitation data from a physical perspective; The discrete multibeam radar feed array configuration scheme and the passive microwave frequency band selection scheme are evaluated based on the test results.

[0014] According to another aspect of the present invention, a computer program product is also provided, comprising a computer program that, when executed by a processor, implements the steps of any of the above-described methods for reconstructing satellite-borne discrete multibeam radar precipitation data.

[0015] According to another aspect of the present invention, a computer device is also provided, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the steps of any of the above-described methods for reconstructing precipitation data from a spaceborne discrete multibeam radar.

[0016] The spaceborne discrete multibeam radar precipitation data reconstruction method of this invention obtains measured reflectivity factors from a spaceborne discrete multibeam radar and brightness temperature data from different channels from a spaceborne microwave radiometer. The reflectivity factors and brightness temperature data are then input into a pre-trained precipitation data reconstruction model to reconstruct a complete reflectivity factor profile. This precipitation data reconstruction model uses the reflectivity factor as a benchmark, integrates precipitation characteristics from brightness temperature data, reconstructs precipitation data from discrete beams, and outputs a complete reflectivity factor profile, thereby supplementing discrete beam observations. This method can effectively fill spatial gaps in discrete multibeam observations without increasing hardware costs, significantly improving the observation swath of precipitation profiles. Simultaneously, relying on combined active and passive observations and the precipitation data reconstruction model, it ensures the physical continuity of precipitation structures and reconstruction accuracy, providing reliable algorithmic support for a low-cost, high spatiotemporal resolution spaceborne precipitation observation system.

[0017] The above and other objects, advantages and features of the present invention will become more apparent to those skilled in the art from the following detailed description of specific embodiments of the invention in conjunction with the accompanying drawings. Attached Figure Description

[0018] The following sections will describe some specific embodiments of the invention in detail by way of example and not limitation, with reference to the accompanying drawings. The same reference numerals in the drawings denote the same or similar parts or portions. Those skilled in the art should understand that these drawings are not necessarily drawn to scale. In the drawings: Figure 1 This is a flowchart illustrating a method for reconstructing precipitation data from a spaceborne discrete multibeam radar according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the construction process of three-dimensional tensor training sample data in a spaceborne discrete multibeam radar precipitation data reconstruction method according to an embodiment of the present invention; Figure 3This is a schematic diagram of the training process of the precipitation data reconstruction model in a spaceborne discrete multibeam radar precipitation data reconstruction method according to an embodiment of the present invention. Figure 4A -C are schematic diagrams of the vertical distribution of reconstruction error corresponding to different discrete beam spacing coefficients of the spaceborne discrete multibeam radar precipitation data reconstruction method according to an embodiment of the present invention. Figure 5A -E are schematic diagrams of the vertical profiles of the reconstructed reflectivity factors corresponding to different discrete beam spacing coefficients of the spaceborne discrete multibeam radar precipitation data reconstruction method according to an embodiment of the present invention. Figure 5F This is a schematic diagram of the true and complete reflectivity factor vertical profile of a method for reconstructing precipitation data from a spaceborne discrete multibeam radar according to an embodiment of the present invention. Figure 6A -E are schematic diagrams of channel contribution under different discrete beam intervals in a satellite-borne discrete multibeam radar precipitation data reconstruction method according to an embodiment of the present invention. Figure 7 This is a schematic diagram of a computer program product according to an embodiment of the present invention; Figure 8 This is a schematic diagram of a computer-readable storage medium according to an embodiment of the present invention; and Figure 9 This is a schematic diagram of a computer device according to an embodiment of the present invention. Detailed Implementation

[0019] Those skilled in the art should understand that the embodiments described below are merely a part of the embodiments of the present invention, and not all of the embodiments of the present invention. These partial embodiments are intended to explain the technical principles of the present invention and are not intended to limit the scope of protection of the present invention. Based on the embodiments provided by the present invention, all other embodiments obtained by those skilled in the art without creative effort should still fall within the scope of protection of the present invention.

[0020] It should be noted that the logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus or device (such as a computer-based system, a processor-included system or other system that can fetch and execute instructions from, an instruction execution system, apparatus or device).

[0021] Global-scale precipitation observations can only be achieved through satellite platforms. However, spaceborne visible light and infrared sensors can only detect clouds, not precipitation itself; spaceborne microwave radiometers cannot acquire the three-dimensional structure of precipitation systems, and detecting precipitation is extremely difficult under terrestrial conditions with high brightness temperature backgrounds and complex spatial variations; spaceborne radar can provide large-scale three-dimensional precipitation structure information, but its observation revisit cycle is long and its temporal resolution is low due to limitations in observation swath width and the number of devices. Developing spaceborne discrete multibeam precipitation radar enables low-cost, multi-satellite network observations, achieving high spatiotemporal resolution global three-dimensional precipitation observations.

[0022] Figure 1 This is a flowchart illustrating a method for reconstructing precipitation data from a spaceborne discrete multibeam radar according to an embodiment of the present invention, as shown below. Figure 1 As shown, the satellite-borne discrete multibeam radar precipitation data reconstruction method includes at least the following steps S101 to S103.

[0023] Step S101: Obtain the measured reflectivity factor using a spaceborne discrete multibeam radar. Here, a spaceborne discrete multibeam precipitation radar refers to a spaceborne radar that uses a feed array to form multiple discrete beams for cross-orbit scanning. Compared to traditional phased array radars, this type of radar can significantly reduce hardware costs; however, due to the limitations of the discrete beam observation mode, there are still gaps in inter-beam observations in the cross-orbit direction.

[0024] The reflectivity factor obtained in this step is generally three-dimensional data, containing information in three dimensions: cross-track beam direction, along-track scanning direction, and vertical height direction. However, due to limitations of discrete multibeam observation modes, there are observation gaps between beams in the cross-track direction. The reflectivity factor can provide high-precision, high-vertical-resolution measured precipitation benchmark data, providing reliable constraints for subsequent precipitation data reconstruction.

[0025] Step S102 involves acquiring brightness temperature data from different channels using a spaceborne microwave radiometer. A spaceborne microwave radiometer is a sensor that passively receives microwave radiation emitted from the Earth's surface and atmosphere to detect atmospheric parameters. It cannot acquire the three-dimensional structure of precipitation systems, and detecting precipitation is extremely difficult under terrestrial conditions with high brightness temperature backgrounds and complex spatial variations. Brightness temperature data from different frequency channels can reflect atmospheric water vapor and precipitation information at different altitudes. This data can supplement precipitation information in gaps between radar beams, providing continuous global precipitation characteristics and thus offering multi-dimensional observational data support for reconstructing combined active and passive precipitation data.

[0026] Step S103: The reflectivity factor and brightness temperature data are input into a pre-trained precipitation data reconstruction model to reconstruct a complete reflectivity factor profile. The precipitation data reconstruction model is used to reconstruct precipitation data for discrete beams based on the reflectivity factor and fuse the precipitation characteristics of brightness temperature data, thereby completing the discrete beam observations.

[0027] This method effectively overcomes the high cost of phased array radar and the limited swath width of single-beam radar by relying on spaceborne discrete multibeam precipitation radar. At the same time, it significantly improves the revisit period through network application, realizing high spatiotemporal resolution three-dimensional global precipitation observation. On the other hand, the method of this invention drives the precipitation data reconstruction model through active and passive microwave joint observation, which can realize the reconstruction of three-dimensional precipitation data by discrete multibeam radar with high accuracy and high reliability, improve the quality of observation data, and lay a solid foundation for subsequent precipitation inversion and application.

[0028] Figure 2 This is a schematic diagram illustrating the construction process of three-dimensional tensor training sample data in a spaceborne discrete multibeam radar precipitation data reconstruction method according to an embodiment of the present invention. The quality of the training samples directly determines the accuracy and generalization ability of the precipitation data reconstruction model. This process constructs high-quality training samples that meet the requirements of deep learning models by preprocessing and integrating active and passive observation data, laying a solid foundation for subsequent model training. Figure 2 As shown, the process includes at least the following steps S201 to S210.

[0029] Step S201: Obtain reflectivity factor data and brightness temperature data used for training.

[0030] Step S202 involves spatiotemporal matching of reflectivity factor data and brightness temperature data to obtain reflectivity factor training data and brightness temperature training data. Spatiotemporal matching refers to establishing a one-to-one correspondence between observation data acquired by the spaceborne discrete multibeam radar and the spaceborne microwave radiometer at the same time and in the same geographical area, based on observation time and geographic location information. This ensures that each set of reflectivity factor data and brightness temperature data corresponds to the same observation area of ​​the same precipitation system. This step eliminates spatiotemporal biases between active and passive observation data, guarantees the accuracy of data fusion, and avoids training sample distortion caused by mismatches in observation time or location.

[0031] Step S203: Based on precipitation markers and land surface types, the spatiotemporally matched reflectivity factor training data is filtered for effective data. Precipitation markers are label fields in radar data products used to indicate the presence or absence of precipitation in the observation area; invalid data from areas without precipitation can be removed. Land surface types include oceans, land, deserts, and snow, etc. Different land surface types have significantly different microwave radiation characteristics, which can cause varying degrees of interference to radar observations. Therefore, data corresponding to land surface types suitable for precipitation observation can be selected for subsequent training as needed.

[0032] The filtering operation is generally performed based on cross-track profiles to ensure the spatial continuity of the data. This step can remove invalid data that is irrelevant to the reconstruction of precipitation data, improve the purity of the training samples, and reduce the interference of irrelevant information on the model learning process.

[0033] Step S204 involves masking the filtered reflectivity factor training data based on a sensitivity threshold and ground clutter height. The sensitivity threshold generally refers to the lowest reflectivity factor value that the radar can accurately detect precipitation; data below this threshold with a low signal-to-noise ratio is considered unreliable. Ground clutter height refers to the maximum vertical height at which ground-reflected clutter can affect radar signals; radar signals below this height are severely contaminated by ground clutter and cannot be used for precipitation observation. The masking operation marks the values ​​in the unreliable regions as invalid, excluding them from loss calculations during model training. This operation removes weak signal interference and ground clutter contamination, retaining only the effective precipitation data that the radar can accurately observe, further ensuring the reliability of the training samples.

[0034] Step S205 involves expanding the spatiotemporally matched brightness temperature training data along the vertical height. Since the brightness temperature data acquired by the spaceborne microwave radiometer is generally two-dimensional planar data, containing only spatial information in the trans-orbit and along-orbit directions, it lacks a vertical height dimension. In contrast, the reflectivity factor data acquired by the spaceborne precipitation radar is three-dimensional data, including a vertical height dimension. To ensure dimensional consistency between the two types of data, this step expands the two-dimensional brightness temperature data along the vertical height direction. This ensures that the expanded brightness temperature data has the same value at each vertical height level as the original two-dimensional brightness temperature data, thus forming three-dimensional brightness temperature data that perfectly matches the reflectivity factor data dimension. This unifies the dimensional structure of the active and passive observation data, providing a prerequisite for the subsequent joint construction of three-dimensional tensor samples.

[0035] Step S206: Standardize the reflectivity factor training data and brightness temperature training data respectively. Since the physical meaning, units, and numerical ranges of reflectivity factor and brightness temperature data differ significantly, directly inputting them into the model would lead to unsatisfactory training results. Therefore, the method of this invention selects to standardize the two types of data separately, mapping them to similar numerical ranges.

[0036] In some optional embodiments, the formula for standardizing the reflectivity factor training data is shown in equation (1): Equation (1) in, This represents the standardized reflectivity factor training data. This represents the original reflectivity factor training data. This represents the minimum value in the original reflectivity factor training data. This represents the maximum value in the original reflectance factor training data. This formula allows for a linear mapping of the reflectance factor data to the [-1, 1] interval, preserving the relative distribution characteristics of the reflectance factor training data.

[0037] Optionally, the formula for standardizing the brightness temperature training data is shown in equation (2): Equation (2) in, This represents the standardized brightness temperature training data. This represents the original brightness temperature training data. This represents the arithmetic mean of the brightness temperature training data. This represents the standard deviation of the brightness temperature training data. This formula can convert the brightness temperature data into data that conforms to a standard normal distribution, thereby eliminating the influence of differences in brightness temperature value ranges between different channels.

[0038] Step S207 involves combining the standardized brightness temperature training data with the reflectivity factor training data to obtain three-dimensional tensor sample data. The three-dimensional tensor is obtained by stitching together different types of observation data along the channel dimension, and is an optional input data format for precipitation data reconstruction models.

[0039] Optionally, the three-dimensional tensor sample data is shown in equation (3): Equation (3) The three-dimensional tensor sample data conforms to the structure of number of channels × number of vertical layers × number of beams. C represents the number of channels, including reflectivity factor channels and brightness temperature channels. It may also include optional polarization difference channels, which are the differences between vertical polarization brightness temperature and horizontal polarization brightness temperature at the same frequency. H is the number of vertical layers, corresponding to different height layers observed by the radar. W is the number of cross-track beams, corresponding to different beam positions of radar cross-track scanning.

[0040] This step integrates multi-source heterogeneous active and passive observation data into unified structured data, which facilitates the extraction of multi-dimensional and multi-scale precipitation features for training in precipitation data reconstruction models.

[0041] Step S208: Using the nadir center beam as a reference, a mask is constructed on the reflectivity factor training data channel according to a preset interval coefficient. In some optional embodiments, the mask can first be constructed as a one-dimensional vector in the cross-track direction. Then broadcast vertically to form The mask matrix; in batch sample or along-track sample processing, it can be further broadcast along the sample dimension. The mask value is 0 or 1, where 1 indicates that the position is an effective observation beam, and 0 indicates that the position is a missing beam to be filled.

[0042] Alternatively, the mask construction formula is shown in equation (4): Equation (4) in, This represents the value of the mask at the j-th cross-track beam position. This indicates the index of the center beam at the nadir point, and n represents the spacing coefficient of the discrete beam.

[0043] Equation (4) means that when hour, A value of 1 indicates that the beam position retains the actual radar observation data (effective beam); other positions that do not meet the above conditions... A value of 0 indicates that the actual data at that beam position has been removed, and model reconstruction is required to complete it. The mask generation strategy is to remove the mask at intervals of n from the center beam of the nadir point to both sides, and retain one beam after every n beams.

[0044] Step S209 involves hollowing out the standardized reflectivity factor training data based on a mask to construct discrete beam observation data. Hollowing out and filling the data involves performing element-wise operations on the complete reflectivity factor data using a mask matrix, retaining the original reflectivity factor values ​​for valid observation beam positions, and replacing the values ​​for missing beam positions with preset fill values ​​(e.g., 0 or background reflectivity values).

[0045] Alternatively, the formula for constructing discrete beam observation data is shown in equation (5): Equation (5) in, This represents the discrete beam observation data after the cut-out and filling process. This represents the training data for the reflectivity factor. Indicates the mask. This represents the fill value for the cut-out area. This represents the inverse matrix of the mask. Its core logic is: the first half retains only the true values ​​of the effective beam. The data is divided into two parts: the second half is filled with fixed values ​​only at the cut-out positions, and the two parts are added together to obtain the final discrete beam incomplete data.

[0046] Step S210: Combine the three-dimensional tensor sample data, mask, and discrete beam observation data as three-dimensional tensor training sample data.

[0047] Specifically, the final training sample data mainly consists of two parts: the model input part and the label part. The model input part is composed of discrete beam observation data, masks, and brightness temperature training data from the three-dimensional tensor sample data; the label part is the complete standardized reflectivity factor training data, that is, the complete three-dimensional reflectivity factor profile that the model needs to predict and output.

[0048] This method can construct complete supervised learning sample pairs, providing the model with accurate inputs and corresponding standard answers, enabling the model to learn the mapping relationship to recover the complete precipitation profile from discrete active and passive observation data.

[0049] Figure 3 This is a schematic diagram of the training process of the precipitation data reconstruction model in a spaceborne discrete multibeam radar precipitation data reconstruction method according to an embodiment of the present invention. The performance of the precipitation data reconstruction model directly determines the accuracy and physical rationality of filling in the missing regions of discrete beams. This process constructs a high-precision active-passive joint precipitation data reconstruction model through scientific model architecture design, multi-constraint loss function construction, introduction of attention mechanism, and systematic training and evaluation strategies. Figure 3 As shown, the process includes at least the following steps S301 to S306.

[0050] Step S301: Acquire reflectivity factor data and brightness temperature data for training. Training data typically originates from on-orbit observations of multiple satellite-borne precipitation radars and microwave radiometers, prioritizing active-passive data pairs from synchronous observations on the same platform. For example, synchronous observations from the Precipitation Measurement Radar (PMR) on FY-3G and the MicroWave Radiation Imager for the Rainfall Mission (MWRI-RM) on FY-3G are preferred. Training data can cover different precipitation types, land surface types, seasons, and geographical regions to ensure the model can learn diverse precipitation characteristics and improve generalization ability.

[0051] Optionally, polarization difference training data at specific frequencies can also be acquired simultaneously to enrich the feature dimensions of the training samples. This step provides sufficient, diverse, and representative basic observational data for model training, ensuring that the model can cover various precipitation scenarios in real-world applications.

[0052] Step S302: Spatiotemporal matching is performed on the reflectivity factor data and brightness temperature data to obtain reflectivity factor training data and brightness temperature training data.

[0053] Step S303: Training sample data is constructed based on the reflectivity factor training data and brightness temperature training data. One possible implementation of this step is as follows: Figure 2 As shown, through steps such as spatiotemporal matching, data filtering, masking, dimensional expansion, standardization, and discrete beam simulation, the original active and passive observation data are converted into supervised learning sample pairs that meet the input requirements of deep learning models. This generates high-quality, standardized training samples, providing a data foundation for the model to learn the mapping relationship between discrete observations and complete precipitation profiles.

[0054] Step S304: Determine the basic architecture and loss function of the deep learning model, and introduce an attention mechanism.

[0055] In some optional embodiments, the basic architecture of the precipitation data reconstruction model can generally be a convolutional neural network architecture, such as the U-Net architecture. The U-Net architecture has an encoder-decoder symmetric structure and a cross-layer skip connection mechanism. The encoder part extracts multi-scale precipitation features from the input data step by step through multi-layer convolution and pooling operations. The decoder part restores the spatial resolution of the data and generates the output results step by step through upsampling and convolution operations. The cross-layer skip connection fuses the shallow features of the encoder part with the deep features of the decoder part and other multi-scale features to achieve a joint representation of global features and local details, which can effectively preserve the fine features of the precipitation structure.

[0056] Alternatively, the input to the convolutional neural network architecture is shown in equation (6): Equation (6) in To input 3D tensor sample data for a convolutional neural network architecture, For discrete beam observation data, For the mask, The input tensor is the standardized brightness temperature training data. It integrates the discrete reference information measured by the radar, the beam position information, and the global continuous precipitation information from the microwave radiometer, providing multi-dimensional observation constraints for the model.

[0057] Alternatively, the output of the convolutional neural network architecture is shown in equation (7): Equation (7) in, This is the reconstructed complete reflectivity factor profile. This is the mapping function for the deep learning model.

[0058] To simultaneously ensure the overall accuracy of the reconstruction results, the accuracy of the precipitation core region, and the physical continuity, the method of this invention also constructs a loss function. An optional example of the loss function is shown in equation (8): Equation (8) in, This is the total loss value. This is the global error value. and The preset weighting coefficients, This represents the weighted error value for the precipitation area. For cross-track gradient constraints.

[0059] Optionally, the formula for calculating the global error value is shown in equation (9): Equation (9) The predicted values ​​of the complete reflectance factor profile output by the model are used to reconstruct precipitation data. This represents the true value of the complete reflectivity factor profile. It represents the average absolute error between the calculated true value and the predicted value.

[0060] Optionally, the formula for calculating the weighted error value of precipitation area is shown in equation (10): Equation (10) in, This indicates that only the absolute errors between data points in the precipitation region are summed. This represents the weighted average absolute error in calculating precipitation areas. This loss term strengthens the model's ability to reconstruct precipitation areas by assigning higher weights to the core precipitation areas, thereby preventing the precipitation intensity from being underestimated.

[0061] Optionally, the formula for calculating the cross-track gradient constraint is shown in equation (11): Equation (11) in, This is a gradient operator along the trans-track direction, used to calculate the rate of change of the numerical value along the trans-track beam direction. This represents the prediction gradient of the predicted value along the trans-track direction. This represents the true gradient of the true value along the trans-track direction. This represents the summation of the absolute gradient errors between all predicted and actual gradients. This loss term ensures the physical continuity between the reconstructed region and the adjacent measured region by constraining the gradient of the reconstructed result to be consistent with the gradient of the actual data, thus avoiding abrupt changes that do not conform to the physical laws of precipitation.

[0062] To further improve the model's efficiency in utilizing high-contribution channels, the method of this invention also introduces a channel attention mechanism to adaptively allocate channel weights to the intermediate feature tensors generated by the intermediate layers of the deep learning model. At the same time, a channel attention module can be introduced into the encoder output features or the skip connection fusion path to adaptively enhance the feature channels that contribute highly to the reconstruction of precipitation data and suppress the interference of noise features or redundant features.

[0063] In some alternative embodiments, the intermediate feature tensor is as shown in equation (12): Equation (12) in, Indicates the number of intermediate feature channels. and These represent the feature dimensions of the intermediate feature tensor in the vertical direction and the cross-track direction, respectively; Global average pooling is performed on the c-th feature channel to obtain the channel descriptor. The formula for calculating the channel descriptor is shown in equation (13): Equation (13) Channel attention weights are generated based on channel descriptors, and the calculation formula for channel attention weights is shown in equation (14): Equation (14) Where s is a vector composed of the descriptors of each channel. and For learnable parameters, Represents a non-linear activation function. This represents the Sigmoid activation function, where w is the channel attention weight vector; The formula for calculating the feature map by weighting the c-th feature channel is shown in equation (15): Equation (15) in, This represents the two-dimensional spatial feature map of the c-th feature channel. This represents the attention weight of the c-th feature channel. This represents the feature map of the c-th feature channel after attention weighting. This indicates element-wise multiplication or channel-based broadcast multiplication.

[0064] Step S305: Train the deep learning model according to the preset training strategy and training sample data to obtain the precipitation data reconstruction model.

[0065] In some optional embodiments, the steps of training a deep learning model according to a preset training strategy and training sample data to obtain a precipitation data reconstruction model generally include: dividing the training sample data into a training set and a validation set, wherein the training set is used for iterative updates of model parameters, and the validation set is used to monitor the generalization ability of the model and prevent overfitting; using supervised learning to iteratively train the deep learning model based on the training set; using an adaptive learning rate optimization algorithm to iteratively update the model parameters, and combining it with a learning rate decay strategy for training, that is, a larger learning rate can be used in the early stage of training to accelerate the convergence speed, and the learning rate is gradually reduced in the later stage of training to improve the accuracy of parameter updates; determining the batch size to ensure the stability of gradient estimation while taking into account training efficiency; monitoring the loss value or evaluation index of the deep learning model on the validation set during training, and determining the model parameters corresponding to the optimal evaluation on the validation set as the parameters of the precipitation data reconstruction model.

[0066] This step ensures the stability and efficiency of the model training process, avoids overfitting, and results in a precipitation data reconstruction model with strong generalization ability and high reconstruction accuracy.

[0067] Step S306 involves verifying and evaluating the reconstruction results of the precipitation data reconstruction model, as well as performing channel contribution analysis. This step is used to comprehensively evaluate the actual performance of the model and provide quantitative support for the hardware design of discrete multibeam precipitation radar.

[0068] In some optional embodiments, the steps of verifying and evaluating the reconstruction results of the precipitation data reconstruction model and analyzing channel contributions generally include: quantitatively evaluating the accuracy of the filled area and the overall profile of the precipitation data reconstruction model based on preset statistical indicators, where statistical indicators generally include mean absolute deviation (MAE), root mean square error (RMSE), and standard deviation (STD), etc., where MAE reflects the average deviation of the reconstruction results, RMSE reflects the influence of large errors, and STD reflects the dispersion of errors; determining the relationship between errors and ground clutter height, surface type, and precipitation structure complexity based on the vertical profile error distribution and filled area error characteristics of the reconstruction results; adjusting the channel configuration input to the precipitation data reconstruction model to design multiple sets of control experiments to quantify the contribution of each channel to the reconstruction performance; explaining the mechanism of action of different observation information on precipitation data reconstruction from a physical perspective based on the experimental results of the control experiments; evaluating the discrete multibeam radar feed array configuration scheme and passive microwave frequency band selection scheme based on the experimental results, thereby providing theoretical basis and algorithmic support for the hardware design and engineering implementation of the discrete multibeam precipitation radar small satellite system.

[0069] This method effectively overcomes the high cost of phased array radar and the limited swath width of single-beam radar by relying on spaceborne discrete multibeam precipitation radar. Furthermore, network application significantly improves the revisit period, enabling high spatiotemporal resolution three-dimensional global precipitation observation. On the other hand, the method of this invention drives a precipitation data reconstruction model through joint active and passive microwave observations, achieving high-precision and high-reliability reconstruction of three-dimensional precipitation data from discrete multibeam radar, improving the quality of observation data and laying a solid foundation for subsequent precipitation inversion and applications. In addition, based on the sensitivity test results of different channel configurations, the configuration scheme of the multibeam radar feed array and the passive microwave frequency band selection scheme can be evaluated, providing effective support for the hardware design and engineering implementation of discrete multibeam precipitation radar small satellite systems.

[0070] In some optional embodiments, the method of the present invention mainly includes four core steps: spatiotemporal matching of active and passive microwave observations, processing of matching data pairs, model architecture and training path, and verification, evaluation and channel contribution analysis.

[0071] I. Spatiotemporal Matching of Active and Passive Microwave Observations This step is fundamental to the entire technical solution, addressing the spatiotemporal mismatch between spaceborne precipitation radar and microwave radiometer data caused by differences in their observation modes. Because the precipitation radar uses a discrete multi-beam scanning mode while the microwave radiometer uses a wide-swath pushbroom mode, their beam footprint size, scanning frequency, and observation time differ. Direct fusion would result in severe spatiotemporal misalignment errors. Therefore, spatiotemporal matching of reflectivity factor training data and brightness temperature training data is chosen, mapping reflectivity factor data and brightness temperature data from the same time and geographical area to form a one-to-one matched active and passive observation data pair. The technical effect of this step is to ensure the spatiotemporal consistency of active and passive observation data from the source, eliminate fusion errors caused by observation time or location deviations, and provide reliable basic data for subsequent data processing and model training.

[0072] II. Data Pair Processing This step performs parallel preprocessing and standardization on the spatiotemporally matched active and passive observation data, ultimately generating three-dimensional tensor samples that meet the input requirements of deep learning models. Data processing is generally divided into two independent processing chains: precipitation radar data (Zm) and microwave radiometer data (Tb). Precipitation radar data processing chain: First, precipitation / surface screening is performed, using precipitation markers to remove invalid data from areas with no precipitation, and filtering data corresponding to strongly interfering surfaces according to surface type; then, sensitivity / clutter masking is performed, marking weak signal data below the radar sensitivity threshold and contaminated data below the clutter height as invalid values; finally, valid data is standardized and mapped to a unified numerical range.

[0073] Microwave radiometer data processing link: First, the two-dimensional brightness temperature data is copied and extended along the vertical height layer of the precipitation radar so that the extended brightness temperature data has a three-dimensional structure that is completely consistent with the reflectivity factor data; then, the extended three-dimensional brightness temperature data is standardized to eliminate the numerical scale differences between different microwave channels.

[0074] After the two links are processed, the standardized three-dimensional reflectivity factor data and the three-dimensional brightness temperature data are stitched together according to the channel dimension to generate three-dimensional tensor sample data with the dimension of number of channels C × number of vertical layers H × number of cross-track beams W.

[0075] This step involves systematic data cleaning, dimensional unification, and standardization to transform multi-source heterogeneous active and passive observation data into structured and standardized training samples. This effectively improves the purity and usability of the training samples, laying the foundation for the model to learn stable precipitation characteristics.

[0076] III. Model Architecture and Training Path This section constructs a U-Net deep learning model with an attention mechanism to achieve end-to-end mapping from discrete active and passive observation data to a complete three-dimensional reflectivity factor profile. The model input is typically a C×H×W three-dimensional tensor sample, specifically including discrete beam observation data, beammask matrix, and standardized brightness temperature data. The model output is a complete H×W-dimensional trans-track-vertical reflectivity factor profile. The outputs of multiple trans-track samples, arranged along the trans-track direction, form a complete three-dimensional reflectivity factor structure. The model generally adopts the U-Net encoder-decoder symmetric architecture. The encoder part progressively extracts multi-scale precipitation features through multi-layer convolution and pooling operations, while the decoder part progressively restores spatial resolution through upsampling and convolution operations. The encoder and decoder parts are fused through cross-layer skip connections, fusing shallow detail features with deep semantic features, effectively preserving the fine information of the precipitation structure. Simultaneously, an attention mechanism is introduced into the model's skip connection paths to enhance channel features that contribute significantly to precipitation data reconstruction and suppress interference from noise and irrelevant features.

[0077] This step fully leverages the advantages of radar's high-precision vertical structure and microwave radiometer's global continuous observation capabilities, enabling high-precision completion of gaps between discrete beams. The reconstructed results retain the true physical characteristics of precipitation while also exhibiting good spatial continuity.

[0078] IV. Validation Evaluation and Channel Contribution Analysis This step is used to comprehensively quantify the reconstruction performance of the model and provide quantitative support for the hardware design of discrete multibeam precipitation radar. This step generally includes two parts: Validation and evaluation: Statistical indicators such as mean absolute deviation (MAE), root mean square error (RMSE), and standard deviation (STD) were used to quantitatively evaluate the accuracy of the filled area and the overall profile in the model reconstruction results. By drawing vertical profile error distribution maps, the variation law of error with height was analyzed, and the correlation between error and ground clutter height, surface type, and precipitation structure complexity was explored to clarify the applicable boundaries of the model and the sources of error.

[0079] Channel contribution analysis: By adjusting the channel configuration of the model input, multiple sets of channel sensitivity experiments and channel ablation experiments were designed to quantify the contribution of different radar beam configurations and different microwave frequency band channels to the reconstruction performance. Channel importance histograms were plotted to visually demonstrate the differences in contribution of each observation channel, and the mechanism by which different observation information affects precipitation data reconstruction was explained from a physical perspective. Based on the analysis results, the feed array configuration scheme and passive microwave frequency band selection scheme of discrete multibeam radar can be optimized and evaluated.

[0080] This step not only comprehensively and objectively verifies the model's reconstruction performance, but also enables collaborative optimization of algorithm design and hardware design, providing direct theoretical basis and data support for the engineering implementation of discrete multibeam precipitation radar small satellite systems.

[0081] The method of this invention, through the organic combination of the above four major steps, constructs a complete system of spaceborne discrete multibeam radar precipitation data reconstruction technology. It can significantly improve the observation swath and temporal resolution of spaceborne precipitation radar without increasing hardware costs, providing a feasible solution for achieving high spatiotemporal resolution global three-dimensional precipitation observation.

[0082] Optionally, Figures 4 to 6 are schematic diagrams of the performance verification and channel contribution analysis results of the spaceborne discrete multibeam radar precipitation data reconstruction method according to an embodiment of the present invention. They are used to comprehensively verify the reconstruction accuracy, robustness and contribution of each observation channel of the method under different discrete beam sparsity from three dimensions: quantitative error statistics, qualitative profile restoration and channel sensitivity analysis, and support the feasibility verification and payload configuration optimization of the discrete multibeam radar system.

[0083] Figure 4A -C are schematic diagrams showing the vertical distribution of reconstruction errors corresponding to different discrete beam spacing coefficients in a spaceborne discrete multibeam radar precipitation data reconstruction method according to an embodiment of the present invention.

[0084] Specifically, Figure 4A This is a schematic diagram of the vertical distribution of the mean absolute deviation (MAE) when the discrete beam spacing coefficients n=1 to n=5. Figure 4AIn the graph, the vertical axis represents altitude in kilometers (km), corresponding to a vertical height of approximately 12km from the ground; the horizontal axis represents mean absolute deviation (MAE) in decibels (dBZ), reflecting the average degree of deviation between the reconstructed result and the true value. Figure 4A In the diagram, n=1 to n=5 represent different discrete beam spacing coefficients, with a larger n indicating a sparser observation beam. The curve distribution shows that, under all discrete beam spacings, the average absolute deviation is less than 2 dBZ in most altitude layers above ground clutter height. Even under the sparsest n=5 condition, the error in the middle and upper levels does not exceed 2 dBZ. The overall error distribution follows a pattern of "stable in the middle and lower levels, slightly increasing in the upper levels, and a small increase near the ground." The increase in upper-level error stems from the low concentration of precipitation particles and weak radar echo signals, while the increase in near-surface error is related to ground clutter interference, consistent with the physical laws of precipitation observation.

[0085] Figure 4B This is a schematic diagram illustrating the vertical distribution of the root mean square error (RMSE) for discrete beam spacing coefficients n=1 to n=5. Figure 4B In the graph, the vertical axis represents altitude (km), and the horizontal axis represents root mean square error (RMSE), expressed in decibels (dBZ), reflecting the impact of large errors on the reconstruction results. The curve trend and... Figure 4A Consistent, the root mean square error under different n values ​​is generally controlled within 3dBZ, and the error does not increase significantly with increasing n, proving that the method of the present invention can still maintain stable reconstruction accuracy under beam sparse conditions.

[0086] Figure 4C This is a schematic diagram of the vertical distribution of the standard deviation (STD) of the discrete beam spacing coefficients n=1 to n=5. Figure 4C In the figure, the vertical axis represents altitude (km), and the horizontal axis represents standard deviation (STD), with the unit being decibels (dBZ), reflecting the degree of error dispersion. The curve distribution shows that the error dispersion is low under different beam sparsity conditions, with no abnormal fluctuations, further verifying the robustness of the method of this invention. The results in this section demonstrate that, under discrete multibeam radar systems, the reconstruction error of the method of this invention is within the acceptable range for meteorological operations, providing accuracy support for low-cost, wide-swath spaceborne precipitation observation.

[0087] Figure 5A -E are schematic diagrams of the vertical profiles of the reconstructed reflectivity factors corresponding to different discrete beam spacing coefficients of the spaceborne discrete multibeam radar precipitation data reconstruction method according to an embodiment of the present invention. Figure 5FThis is a schematic diagram of the true and complete vertical profile of reflectivity factor in a spaceborne discrete multibeam radar precipitation data reconstruction method according to an embodiment of the present invention. Here, idx=56635 indicates that the index number of this precipitation profile is 56635. The horizontal axis represents the cross-track beam index (Cross-trackbeam(0..24)), corresponding to the 25 beam positions of the radar cross-track scan; the vertical axis represents the height index (Height index(0~299)), corresponding to the vertical height layer from the ground to high altitude; the color bars on the right represent the reflectivity factor values ​​in decibels (dBZ). A larger value indicates stronger precipitation intensity; purple corresponds to weak precipitation, and yellow corresponds to strong convective precipitation.

[0088] Specifically, Figure 5A This is a schematic diagram of the vertical profile of the reconstructed reflectivity factor when the discrete beam spacing coefficient n=1. Figure 5A In this figure, the reconstructed profile is shown under the condition of n=1. At this time, the beam spacing is the smallest, and the reconstruction result is highly consistent with the actual precipitation structure. The bright band of the zero-degree layer of the layered precipitation is clearly visible.

[0089] Figure 5B This is a schematic diagram of the vertical profile of the reconstructed reflectivity factor when the discrete beam spacing coefficient n=2. Figure 5B The reconstructed profile under the n=2 working condition still accurately restores the horizontal continuity and vertical structure evolution of precipitation. The gaps between beams are smoothly filled without any obvious abrupt changes.

[0090] Figure 5C This is a schematic diagram of the vertical profile of the reconstructed reflectivity factor when the discrete beam spacing coefficient n=3. Figure 5C Despite further improvements in beam sparsity, the reconstructed profile still retains key features of precipitation, and the intensity and distribution of convective precipitation do not show significant distortion.

[0091] Figure 5D This is a schematic diagram of the vertical profile of the reconstructed reflectivity factor when the discrete beam spacing coefficient n=4. Figure 5D In the reconstruction results, the horizontal continuity of the stratified precipitation is good, and the location and intensity of the high-value center of convective precipitation still match the true characteristics. There is no break in the precipitation structure caused by the sparse beams.

[0092] Figure 5E This is a schematic diagram of the vertical profile of the reconstructed reflectivity factor when the discrete beam spacing coefficient n=5. Figure 5E In the example shown, this is the sparsest beam configuration. The reconstructed profile still accurately restores the three-dimensional structure of precipitation, with clear bright bands in the zero-degree layer and no significant shift in the high-value center of convective precipitation, proving that the method of this invention still has good structure restoration capability under extremely sparse beam conditions.

[0093] Figure 5FThis is a schematic diagram of the vertical cross-section of the true and complete reflectivity factor. Figure 5F As a benchmark for comparison of the reconstruction results. By comparing with Figures 5A to 5E As can be seen from the comparison, the reconstructed profiles under different beam sparsity are highly consistent with the real profiles, and the key physical characteristics of precipitation (such as the zero-degree layer bright band and the high value center of convective precipitation) are effectively restored, which intuitively verifies the physical rationality of the method of the present invention.

[0094] Figure 6A -E are schematic diagrams illustrating the channel contribution under different discrete beam intervals in the spaceborne discrete multibeam radar precipitation data reconstruction method according to an embodiment of the present invention. These are bar charts illustrating the channel importance under different discrete beam intervals, used to quantify the contribution of each input channel to the reconstruction performance, providing a basis for optimizing the payload configuration of spaceborne radar and microwave radiometers. The horizontal axis represents the channel name, corresponding to different observation channels in the model input; the vertical axis represents the increase in root mean square error (RMSE) of the reconstruction result after removing the channel. The larger the increase, the higher the channel's contribution to the reconstruction performance.

[0095] Specifically, Figure 6A This is a schematic diagram illustrating the channel contribution when the discrete beam spacing coefficient n=1. Figure 6A In the n=1 case, the radar reflectivity factor channel has the highest contribution. Removing this channel significantly increases the error, while the passive microwave brightness temperature channel shows a smaller increase in error. In this case, the radar measured data provides the main constraints for reconstruction.

[0096] Figure 6B This is a schematic diagram illustrating the channel contribution when the discrete beam spacing coefficient n=2. Figure 6B In the meantime, as the beam sparsity increases, the contribution of the passive microwave brightness temperature channel rises slightly, indicating that brightness temperature data begins to play a role in supplementing the missing information in the radar beam.

[0097] Figure 6C This is a schematic diagram illustrating the channel contribution when the discrete beam spacing coefficient n=3. Figure 6C In the middle, the radar reflectivity factor channel still has the highest contribution, but the error of the passive microwave brightness temperature channel has increased further, and its impact on reconstruction performance has gradually become more prominent.

[0098] Figure 6D This is a schematic diagram illustrating the channel contribution when the discrete beam spacing coefficient n=4. Figure 6DIn the case of high beam sparsity, the contribution of the passive microwave brightness temperature channel is significantly increased. After removing some brightness temperature channels, the error increases significantly, proving that brightness temperature data is crucial for filling gaps and maintaining the continuity of precipitation structure under beam sparsity conditions.

[0099] Figure 6E This is a schematic diagram illustrating the channel contribution when the discrete beam spacing coefficient n=5. Figure 6E The passive microwave brightness temperature channel contributed the most, and the error increased significantly after removing multiple brightness temperature channels, indicating that passive microwave data is a key supplementary information to ensure reconstruction accuracy under extremely sparse beam conditions.

[0100] The results in this section verify the necessity of the active-passive joint reconstruction scheme of this invention: the radar reflectivity factor channel provides the core benchmark constraint for reconstruction, while the passive microwave brightness temperature channel plays a key supplementary role under beam sparsity conditions. These results provide direct data support for the hardware configuration optimization of discrete multi-beam radars, allowing for the rational configuration of the frequency band and channel of the passive microwave radiometer based on beam sparsity, achieving a balance between cost and performance.

[0101] In summary, it can be seen that as the discrete beam spacing increases, the contribution of the passive microwave channel also increases. This demonstrates the role of the attention mechanism introduced in this invention in enhancing the channel contribution and illustrates the importance of combining active and passive methods for reconstructing three-dimensional precipitation data. Even with the largest discrete beam spacing (n=5), the reflectivity factor profile reconstructed by this invention can still accurately restore the three-dimensional structural characteristics of real precipitation: the zero-degree bright band of stratiform precipitation is clearly visible, and the location, intensity, and range of the high-value centers of convective precipitation are consistent with the actual profile height. The continuity of the horizontal distribution of precipitation and the evolution of its vertical structure are also well preserved. The discretely distributed high-value points in the figure represent the effective observation beams measured by radar. Based on these sparse measured reference data and combined with global continuous precipitation information provided by microwave radiometers, this invention can accurately fill in the gaps between beams and generate a complete three-dimensional reflectivity factor profile.

[0102] This embodiment comprehensively verifies the high accuracy and robustness of the method of the present invention by combining quantitative error statistics with qualitative profile visualization. This result fully demonstrates the feasibility of the design concept of a spaceborne discrete multibeam precipitation radar, indicating that the active-passive joint reconstruction algorithm of the present invention can significantly reduce radar hardware costs and significantly improve the observation swath while ensuring the quality of precipitation observations, providing solid technical support for achieving low-cost, high spatiotemporal resolution global three-dimensional precipitation observation.

[0103] In summary, the design concept of the spaceborne discrete multibeam precipitation radar observation system in the method of this invention is a new spaceborne precipitation observation technology route that breaks through the technical bottlenecks of traditional phased array radar and single-beam radar. Its core is to achieve global three-dimensional precipitation observation with wide swath and high spatiotemporal resolution by combining low-cost hardware discrete beam observation with high-precision algorithm completion and reconstruction. This significantly reduces hardware costs.

[0104] The hardware design concept replaces the expensive phased array radar with a low-cost feed array, abandoning the traditional phased array radar approach of achieving continuous cross-orbit observation through complex electronic scanning. Instead, it uses a low-cost feed array to form multiple spatially discrete observation beams, directly expanding the observation swath on the basis of a single-beam radar. The algorithm completion logic transforms the observation gaps into an image restoration problem. By combining passive observation data from the onboard microwave radiometer on the same platform, deep learning technology is used to reconstruct the precipitation information between beams in three dimensions, filling in the spatial gaps in the discrete beam observations and generating a complete three-dimensional reflectivity factor profile.

[0105] The flowchart provided in this embodiment is not intended to indicate that the operations of the method will be performed in any particular order, or that all operations of the method are included in every case. Furthermore, the method may include additional operations. Within the scope of the technical concept provided by the method in this embodiment, additional variations can be made to the above method.

[0106] It should be understood that in some embodiments, the components may be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods may be implemented using software or firmware stored in memory and executed by a suitable instruction execution system.

[0107] This embodiment also provides a computer program product 10, a computer-readable storage medium 20, and a computer device 30. Figure 7 This is a schematic diagram of a computer program product 10 according to an embodiment of the present invention. Figure 8 This is a schematic diagram of a computer-readable storage medium 20 according to an embodiment of the present invention. Figure 9 This is a schematic diagram of a computer device 30 according to an embodiment of the present invention. The computer program product 10 includes a computer program 11, which, when executed by the processor 32, implements the steps of any of the above-described methods for reconstructing spaceborne discrete multibeam radar precipitation data. A computer-readable storage medium 20 stores the computer program 11 thereon, which, when executed by the processor 32, implements the steps of any of the above-described methods for reconstructing spaceborne discrete multibeam radar precipitation data. The computer device 30 may include a memory 31, a processor 32, and the computer program 11 stored in the memory 31 and running on the processor 32.

[0108] The computer program 11 used to perform the operations of this invention may be assembly instructions, Instruction Set Architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state setting data, integrated circuit configuration data, or source code or object code written in any combination of one or more programming languages ​​and procedural programming languages. The computer program 11 may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the latter case, the remote computer may be connected to the user's computer via any type of network, including a Local Area Network (LAN) or Wide Area Network (WAN), or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, to perform aspects of this invention, electronic circuits, including, for example, programmable logic circuits, Field-Programmable Gate Arrays (FPGAs), or Programmable Logic Arrays (PLAs), may execute computer-readable program instructions to personalize the electronic circuits by utilizing state information from computer-readable program instructions.

[0109] For the purposes of this embodiment, computer program product 10 is a related product containing computer program 11. For the purposes of this embodiment, computer-readable storage medium 20 is a tangible device capable of holding and storing computer program 11, and can be any device capable of containing, storing, communicating, propagating, or transmitting program 11 for use by or in conjunction with an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable storage medium 20 include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable optical disc read-only memory (CD-ROM), digital versatile disc (DVD), memory stick, floppy disk, mechanical encoding device, and any suitable combination thereof.

[0110] Computer device 30 can be, for example, a server, desktop computer, laptop computer, tablet computer, or smartphone. In some examples, computer device 30 can be a cloud computing node. Computer device 30 can be described in the general context of computer system executable instructions (such as program modules) executed by a computer system. Typically, program modules can include routines, programs, object programs, components, logic, data structures, etc., that perform specific tasks or implement specific abstract data types. Computer device 30 can be implemented in a distributed cloud computing environment where tasks are performed by remote processing devices linked through a communication network. In a distributed cloud computing environment, program modules can reside on local or remote computing system storage media, including storage devices.

[0111] Computer device 30 may include a processor 32 adapted to execute stored instructions and a memory 31 that provides temporary storage space for the operation of said instructions during operation. The processor 32 may be a single-core processor, a multi-core processor, a computing cluster, or any other configuration. The memory 31 may include random access memory (RAM), read-only memory, flash memory, or any other suitable storage system.

[0112] Computer device 30 may also include a network adapter / interface and an input / output (I / O) interface. The I / O interface allows external devices that can be connected to the computer device to input and output data. The network adapter / interface provides communication between the computer device and a network, typically represented as a communication network.

[0113] Therefore, those skilled in the art should recognize that although numerous exemplary embodiments of the present invention have been shown and described in detail herein, many other variations or modifications conforming to the principles of the present invention can be directly determined or derived from the disclosure of the present invention without departing from the spirit and scope of the invention. Thus, the scope of the present invention should be understood and construed as covering all such other variations or modifications.

Claims

1. A method for reconstructing precipitation data from a spaceborne discrete multibeam radar, comprising: The measured reflectivity factor was obtained using the spaceborne discrete multibeam radar. Brightness temperature data for different channels were obtained using a spaceborne microwave radiometer; The reflectivity factor and the brightness temperature data are input into a pre-trained precipitation data reconstruction model to reconstruct a complete reflectivity factor profile. The precipitation data reconstruction model is used to reconstruct precipitation data for discrete beams based on the reflectivity factor and fuse the precipitation characteristics of the brightness temperature data, and output a complete reflectivity factor profile, thereby completing the discrete beam observation.

2. The method for reconstructing precipitation data from a spaceborne discrete multibeam radar according to claim 1, wherein, The training steps for the precipitation data reconstruction model include: Acquire reflectivity factor data and brightness temperature data for training; Spatiotemporal matching is performed on the reflectivity factor data and the brightness temperature data to obtain reflectivity factor training data and brightness temperature training data; Three-dimensional tensor training sample data is constructed based on the reflectivity factor training data and the brightness temperature training data. Determine the basic architecture and loss function of the deep learning model, and introduce an attention mechanism; The deep learning model is trained according to the preset training strategy and the three-dimensional tensor training sample data to obtain the precipitation data reconstruction model; The reconstruction results of the precipitation data reconstruction model were verified and evaluated, and channel contribution analysis was performed.

3. The method for reconstructing precipitation data from a spaceborne discrete multibeam radar according to claim 2, wherein, The step of constructing three-dimensional tensor training sample data based on the reflectivity factor training data and the brightness temperature training data includes: Preprocess the reflectivity factor training data and the brightness temperature training data; The reflectivity factor training data and the brightness temperature training data are standardized respectively to unify the data scale of different observation channels; The standardized reflectivity factor training data and the brightness temperature training data are combined to obtain three-dimensional tensor sample data, which consists of three dimensions: number of channels, number of vertical layers, and number of cross-track beams. Using the center beam of the sub-satellite point as a reference, a mask is constructed on the reflectivity factor training data channel according to a preset interval coefficient. The mask is used to simulate the observation geometry of discrete multibeam precipitation radar. Based on the mask, the standardized reflectivity factor training data is hollowed out to construct discrete beam observation data. The three-dimensional tensor sample data, the mask, and the discrete beam observation data are combined as the three-dimensional tensor training sample data.

4. The method for reconstructing precipitation data from a spaceborne discrete multibeam radar according to claim 3, wherein, The preprocessing steps for the reflectivity factor training data and the brightness temperature training data include: Based on precipitation indicators and surface type, the spatiotemporally matched reflectivity factor training data is filtered for effective data, and the filtered reflectivity factor training data is masked based on sensitivity threshold and ground clutter height, thereby retaining the effective reflectivity factor training data. The brightness temperature training data after spatiotemporal matching is extended along the vertical height so that the extended brightness temperature training data has the same dimension as the reflectivity factor training data.

5. The method for reconstructing precipitation data from a spaceborne discrete multibeam radar according to claim 3, wherein, The formula for standardizing the training data of the reflectivity factor is as follows: Among them, the This represents the standardized training data for the reflectivity factor. This represents the original training data for the reflectivity factor. This represents the minimum value in the original training data for the reflectivity factor. This represents the maximum value in the original training data for the reflectivity factor; The formula for standardizing the brightness temperature training data is as follows: Among them, the The standardized brightness temperature training data, the This represents the original brightness temperature training data, the The brightness temperature training data represents the arithmetic mean. This represents the standard deviation of the brightness temperature training data; The three-dimensional tensor sample data is as follows: Wherein C represents the number of channels, H represents the number of vertical layers, W represents the number of cross-track beams, and the three-dimensional tensor sample data conforms to the structure of number of channels × number of vertical layers × number of beams; The formula for constructing the mask is: Among them, the This represents the value of the mask at the j-th cross-track beam position. The index of the center beam at the sub-satellite point is indicated, and n represents the spacing coefficient of the discrete beam. The formula for constructing the discrete beam observation data is: Among them, the This represents the discrete beam observation data after the cut-out and filling process. This represents the training data for the reflectivity factor. The mask represents the... The value representing the fill value of the hollowed-out area, the This represents the inverse matrix of the mask.

6. The method for reconstructing precipitation data from a spaceborne discrete multibeam radar according to claim 3, wherein, The steps of determining the basic architecture and loss function of the deep learning model and introducing an attention mechanism include: The infrastructure is determined to be a convolutional neural network architecture; The input to the convolutional neural network architecture is set as follows: The above To input the three-dimensional tensor sample data of the convolutional neural network architecture, the For the discrete beam observation data, the For the mask, the The brightness temperature training data after standardization; The output of the convolutional neural network architecture is set as follows: Among them, the For the reconstructed complete reflectivity factor profile, the This is the mapping function of the deep learning model; The loss function is: Among them, the The total loss value, the The global error value, the and stated The preset weighting coefficients, the The weighted error value for precipitation area, the For cross-track gradient constraints; The formula for calculating the global error value is as follows: The The predicted value of the complete reflectance factor profile output by the precipitation data reconstruction model, the The true value of the complete reflectivity factor profile, the This represents the average absolute error between the calculated true value and the predicted value; The formula for calculating the weighted error value of the precipitation area is as follows: Among them, the This indicates that only the absolute errors between data points in the precipitation area are accumulated. This represents the weighted average absolute error in calculating precipitation over a given area. The formula for calculating the trans-track gradient constraint is as follows: Among them, the The gradient operator along the trans-track direction is used to calculate the rate of change of the numerical value along the trans-track beam direction. This represents the prediction gradient of the predicted value along the trans-track direction. This represents the true gradient of the true value along the cross-track direction. This represents the summation of the absolute gradient errors between all predicted gradients and the true gradients; The attention mechanism is a channel attention mechanism, used to adaptively allocate channel weights to the intermediate feature tensors generated by the intermediate layers of the deep learning model. The intermediate feature tensors are... Among them, the Indicates the number of intermediate feature channels, the and stated These represent the feature dimensions of the intermediate feature tensor in the vertical direction and the cross-track direction, respectively; Global average pooling is performed on the c-th feature channel to obtain the channel descriptor. The formula for calculating the channel descriptor is as follows: Channel attention weights are generated based on the channel descriptors, and the formula for calculating the channel attention weights is as follows: Wherein, s is a vector composed of the descriptors of each of the channels, and the and stated As learnable parameters, the Represents a nonlinear activation function, the This represents the Sigmoid activation function, where w is the channel attention weight vector; We obtain the following by weighting the c-th feature channel: Among them, the The two-dimensional spatial feature map representing the c-th feature channel, wherein This represents the attention weight of the c-th feature channel. This represents the feature map of the c-th feature channel after attention weighting. This indicates element-wise multiplication or channel-based broadcast multiplication.

7. The method for reconstructing precipitation data from a spaceborne discrete multibeam radar according to claim 2, wherein, The step of training the deep learning model according to the preset training strategy and the three-dimensional tensor training sample data to obtain the precipitation data reconstruction model includes: The three-dimensional tensor training sample data is divided into a training set and a validation set; The deep learning model is trained iteratively using a supervised learning approach based on the training set. An adaptive learning rate optimization algorithm is used to iteratively update the model parameters, and a learning rate decay strategy is used for training. Determine the batch size; During training, the loss value or evaluation index of the deep learning model on the validation set is monitored, and the model parameters corresponding to the optimal evaluation on the validation set are determined as the parameters of the precipitation data reconstruction model.

8. The method for reconstructing precipitation data from a spaceborne discrete multibeam radar according to claim 2, wherein, The steps for verifying and evaluating the reconstruction results of the precipitation data reconstruction model and performing channel contribution analysis include: The accuracy of the filled area and the overall profile of the precipitation data reconstruction model is quantitatively evaluated based on preset statistical indicators. Based on the vertical profile error distribution and filling area error characteristics of the reconstruction results, the relationship between error and ground clutter height, surface type, and precipitation structure complexity is determined. The channel configuration of the precipitation data reconstruction model input was adjusted to design multiple sets of control experiments to quantify the contribution of each channel to the reconstruction performance. Based on the experimental results of the aforementioned control experiment, the mechanism by which different observational information reconstructs precipitation data is explained from a physical perspective; The discrete multibeam radar feed array configuration scheme and the passive microwave frequency band selection scheme are evaluated based on the test results.

9. A computer program product comprising a computer program that, when executed by a processor, implements the steps of the satellite-borne discrete multibeam radar precipitation data reconstruction method as described in any one of claims 1 to 8.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the steps of the spaceborne discrete multibeam radar precipitation data reconstruction method according to any one of claims 1 to 8.

Images