Infrared instantaneous precipitation retrieval method and device based on operator learning and electronic equipment

Abstract

The application relates to an infrared instantaneous precipitation inversion method and device based on operator learning and electronic equipment. The method comprises the following steps: preprocessing acquired satellite infrared radiation data and ground automatic rain gauge observation data to obtain multi-channel satellite infrared radiation data and sampling point coordinates within a receptive field range, constructing an infrared precipitation estimation operator network, training the infrared precipitation estimation operator network, and obtaining a trained infrared precipitation estimation operator network; and inputting satellite infrared radiation data to be estimated into the trained infrared precipitation estimation operator network to obtain local instantaneous precipitation rate of the point to be estimated. The method adopts operator learning to construct direct mapping from an infrared radiation field to a precipitation field, generates a spatially continuous precipitation estimation field, can calculate precipitation values at any positions without interpolation, directly calculates model errors by using measured true values of stations, and completely eliminates uncertainty caused by interpolation.

Description

Technical Field

[0001] This invention belongs to the field of meteorological satellite remote sensing and precipitation inversion technology, and relates to an infrared instantaneous precipitation inversion method, device and electronic equipment based on operator learning. Background Technology

[0002] Currently, the mainstream methods for precipitation observation include ground-based rain gauges, weather radar, and satellite remote sensing. Among these, rain gauges can directly measure precipitation with high accuracy. However, the uneven distribution of rain gauge stations leads to insufficient spatial sampling density, especially in sparsely populated areas, making it difficult to effectively capture large-scale spatial variability of weather phenomena such as convective precipitation. Ground-based weather radar provides better spatial continuity for precipitation system observations compared to rain gauges, but in complex terrains such as mountains, it is prone to observational omissions or distortions due to terrain obstruction.

[0003] Infrared radiometers on geostationary Earth orbit (GEO) platforms provide wide-area, near-real-time satellite observation data, offering irreplaceable advantages in ocean and remote regions. This method primarily utilizes infrared bands (especially the infrared split-window channel) to comprehensively determine cloud top temperature and water vapor content, thereby inferring precipitation distribution. In contrast, microwave channels have stronger penetrating power, allowing satellites carrying microwave radiometers to acquire structural information within precipitation systems. Satellite missions such as the Tropical Precipitation Measurement Program (TPP) and the Global Precipitation Measurement Program (GPP) carry microwave radiometers and onboard rain-measuring radars, enabling more precise detection of the vertical distribution and intensity of precipitation. However, due to the limited detection swath of microwave sensors, satellites carrying such sensors typically operate in low Earth orbit (LEO). Therefore, microwave-based precipitation observations cannot meet the needs for continuous, all-weather monitoring of specific areas.

[0004] In recent years, with the widespread application of artificial intelligence (AI) technology in the meteorological field, satellite precipitation estimation algorithms based on deep learning have developed rapidly. These methods typically utilize convolutional neural networks (CNNs) and their derivative models to efficiently learn multi-scale spatial features in satellite infrared data, thereby enhancing the ability to represent complex nonlinear relationships. By fusing observational data from multiple sources and multiple bands (such as ground rain gauges, radar, and infrared / microwave satellite data), the accuracy of precipitation identification and quantitative estimation can be further improved.

[0005] However, existing deep learning methods still face numerous challenges. First, models perform poorly when handling localized and differentiated features influenced by complex terrain, resulting in estimation errors that are significantly correlated with terrain. Second, multi-source fusion methods are constrained by issues such as data availability and synchronization, making them difficult to apply directly to operational systems with high real-time requirements. Furthermore, different data sources exhibit significant differences in spatiotemporal resolution and error characteristics, posing significant challenges to the fusion process in areas such as calibration of massive datasets, accuracy verification, and model parameter optimization. More importantly, current mainstream methods largely rely on ground station observations or gridded precipitation products as ground truth, both of which require interpolation to match the spatiotemporal resolution of satellite input. This interpolation process introduces additional errors, leading to insufficient reliability of the evaluation results. Summary of the Invention

[0006] To address the problems existing in the above-mentioned traditional methods, this invention proposes an infrared instantaneous precipitation inversion method, apparatus, and electronic equipment based on operator learning.

[0007] To achieve the above objectives, the embodiments of the present invention adopt the following technical solutions: On the one hand, an infrared instantaneous precipitation inversion method based on operator learning is provided, including the following steps: The acquired satellite infrared radiation data and ground automatic rain gauge observation data are preprocessed to obtain multi-channel satellite infrared radiation data and the coordinates of its sampling points within the sensing field.

[0008] An infrared precipitation estimation operator network is constructed and trained to obtain a trained infrared precipitation estimation operator network. The infrared precipitation estimation operator network includes: an input module, an adaptive spatial feature encoder, a dynamic branch network, a backbone network, a fusion module, and a prediction module. The input module processes multi-channel satellite infrared radiation data to obtain a mask-constrained input feature tensor. The adaptive spatial feature encoder processes the input feature tensor using two-dimensional convolution and global average pooling operations to obtain an embedding vector. The dynamic branch network compresses the embedding vector into branch basis vectors for precipitation operators. The backbone network performs dimensionality scaling and nonlinear mapping on the coordinates of sampling points to obtain backbone basis vectors. The fusion module performs nonlinear mapping on the dimensionally concatenated branch basis vectors and backbone basis vectors to obtain the parameters of all Gaussian radial basis function components. The prediction module splits the parameters of all Gaussian radial basis function components along the channel dimension, applies meteorological and physical constraints to the splitting results, and calculates the predicted precipitation rate based on the constrained parameters.

[0009] The satellite infrared radiation data to be estimated is fed into a trained infrared precipitation estimation operator network to obtain the local instantaneous precipitation rate at the point to be estimated.

[0010] On the other hand, an infrared instantaneous precipitation inversion device based on operator learning is also provided, comprising: The data preprocessing unit is used to preprocess the acquired satellite infrared radiation data and ground automatic rain gauge observation data to obtain multi-channel satellite infrared radiation data and the coordinates of its sampling points within the sensing field. An infrared precipitation estimation operator network construction and training unit is used to construct and train an infrared precipitation estimation operator network to obtain a trained infrared precipitation estimation operator network. The infrared precipitation estimation operator network includes: an input module, an adaptive spatial feature encoder, a dynamic branch network, a backbone network, a fusion module, and a prediction module. The input module processes multi-channel satellite infrared radiation data to obtain a mask-constrained input feature tensor. The adaptive spatial feature encoder processes the input feature tensor using two-dimensional convolution and global average pooling operations to obtain an embedding vector. The dynamic branch network compresses the embedding vector into branch basis vectors for precipitation operators. The backbone network performs dimensionality scaling and nonlinear mapping on the sampling point coordinates to obtain backbone basis vectors. The fusion module performs nonlinear mapping on the dimensionally concatenated branch basis vectors and backbone basis vectors to obtain the parameters of all Gaussian radial basis function components. The prediction module splits the parameters of all Gaussian radial basis function components along the channel dimension, applies meteorological and physical constraints to the splitting results, and calculates the predicted precipitation rate based on the constrained parameters. The precipitation rate inversion calculation unit is used to input the satellite infrared radiation data to be estimated into a trained infrared precipitation estimation operator network to obtain the local instantaneous precipitation rate of the point to be estimated.

[0011] On another front, an electronic device is also provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of any of the above-mentioned operator-based infrared instantaneous precipitation inversion methods.

[0012] One of the above technical solutions has the following advantages and beneficial effects: The aforementioned method, apparatus, and electronic equipment for infrared instantaneous precipitation retrieval based on operator learning include the following steps: preprocessing acquired satellite infrared radiation data and ground-based automatic rain gauge observation data to obtain multi-channel satellite infrared radiation data and their sampling point coordinates within the sensing field; constructing and training an infrared precipitation estimation operator network to obtain a trained infrared precipitation estimation operator network; and inputting the satellite infrared radiation data to be estimated into the trained infrared precipitation estimation operator network to obtain the local instantaneous precipitation rate at the point to be estimated. This method uses operator learning to construct a direct mapping from the infrared radiation field to the precipitation field, generating a spatially continuous precipitation estimation field. Precipitation values ​​at any location can be calculated without interpolation, and model errors are directly calculated using the measured true values ​​from the stations, completely eliminating the uncertainties introduced by interpolation. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of this application or the conventional technology, the drawings used in the description of the embodiments or the conventional technology will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 This is a flowchart illustrating an operator-learning-based infrared instantaneous precipitation inversion method in one embodiment. Figure 2 This is a schematic diagram of an infrared precipitation estimation operator network in one embodiment; Figure 3 This is a schematic diagram of the training process of the infrared precipitation estimation operator network in one embodiment. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0016] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the application.

[0017] It should be noted that, in this document, the reference to "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The presentation of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. Those skilled in the art will understand that the embodiments described herein can be combined with other embodiments. The term "and / or" as used herein refers to any combination of one or more of the associated listed items, and all possible combinations, including such combinations.

[0018] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0019] In one embodiment, such as Figure 1 As shown, an infrared instantaneous precipitation inversion method based on operator learning is provided, which may include the following processing steps 1 to 3: Step 1: Preprocess the acquired satellite infrared radiation data and ground automatic rain gauge observation data to obtain multi-channel satellite infrared radiation data and the coordinates of its sampling points within the sensing field.

[0020] Specifically, read multi-channel infrared radiation data from the FY-4B satellite and extract... , and Channel infrared data is used to provide spatial distribution information of water vapor and extract... Infrared data from the channel is used to reflect surface temperature information. The readout time resolution is... Quantization resolution is Ground-based automatic rain gauge observation data.

[0021] The data from the ground-based automatic rain gauge observations were time-aligned with the extracted infrared radiation data from the four channels of the FY-4B satellite, and outlier cleaning was performed. Then, local spatial matching based on the receptive field was adopted to obtain the matching sample space. Each sample includes multi-channel satellite infrared radiation data within the receptive field and the coordinates of its sampling points.

[0022] Step 2: Construct and train an infrared precipitation estimation operator network to obtain a trained infrared precipitation estimation operator network. The infrared precipitation estimation operator network includes: an input module, an adaptive spatial feature encoder, a dynamic branch network, a backbone network, a fusion module, and a prediction module. The input module processes multi-channel satellite infrared radiation data to obtain a mask-constrained input feature tensor. The adaptive spatial feature encoder processes the input feature tensor using two-dimensional convolution and global average pooling operations to obtain an embedding vector. The dynamic branch network compresses the embedding vector into branch basis vectors for precipitation operators. The backbone network performs dimensionality scaling and nonlinear mapping on the sampling point coordinates to obtain backbone basis vectors. The fusion module performs nonlinear mapping on the dimensionally concatenated branch basis vectors and backbone basis vectors to obtain the parameters of all Gaussian radial basis function components. The prediction module splits the parameters of all Gaussian radial basis function components along the channel dimension, applies meteorological and physical constraints to the splitting results, and calculates the predicted precipitation rate based on the constrained parameters.

[0023] Specifically, this application treats satellite infrared radiation data and ground precipitation measurements (ground-based automatic rain gauge observation data) as discrete samples of two function fields within the same spatiotemporal domain. Based on the universal approximation theorem for continuous operators (Chen & Chen, 1995), neural networks (such as DeepONet and FNO) can be designed to approximate nonlinear continuous operators. Therefore, this method employs operator learning to construct a direct mapping from the infrared radiation field to the precipitation field, generating a spatially continuous precipitation estimation field. The key advantage of this method is that it can calculate precipitation values ​​at any location (including observation stations) without interpolation. This allows for direct calculation of model errors using the actual measured values ​​from the stations, completely eliminating the uncertainty introduced by interpolation. Clearly, the denser the station network, the more pronounced the advantages of this method.

[0024] In the infrared precipitation estimation operator network, an adaptive spatial feature encoder is used to extract spatial features, a dynamic branch network is used to encode the local infrared field, a backbone network is used to encode the target location, and a fusion module is used to extract feature parameters through nonlinear mapping.

[0025] Infrared precipitation estimation operator network, such as Figure 2 As shown.

[0026] Step 3: Input the satellite infrared radiation data to be estimated into the trained infrared precipitation estimation operator network to obtain the local instantaneous precipitation rate of the point to be estimated.

[0027] The aforementioned operator-based infrared instantaneous precipitation retrieval method includes: preprocessing acquired satellite infrared radiation data and ground-based automatic rain gauge observation data to obtain multi-channel satellite infrared radiation data and their sampling point coordinates within the receptive field; constructing and training an infrared precipitation estimation operator network to obtain a trained infrared precipitation estimation operator network; and feeding the satellite infrared radiation data to be estimated into the trained infrared precipitation estimation operator network to obtain the local instantaneous precipitation rate at the point to be estimated. This method uses operator learning to construct a direct mapping from the infrared radiation field to the precipitation field, generating a spatially continuous precipitation estimation field. It can calculate precipitation values ​​at any location without interpolation, and directly calculates model errors using the measured true values ​​from stations, completely eliminating the uncertainties introduced by interpolation.

[0028] In one embodiment, step 1 includes: reading multi-channel infrared radiation data from the FY-4B satellite and observation data from ground-based automatic rain gauges; extracting satellite infrared radiation data from four preset channels from the multi-channel infrared radiation data; performing time alignment and outlier cleaning on the ground-based automatic rain gauge observation data and the extracted four-channel satellite infrared radiation data; and then using local spatial matching based on the receptive field to obtain a matching sample space, where each sample includes: all satellite infrared radiation data within the receptive field and their coordinates; wherein the process of local spatial matching based on the receptive field includes: matching the geographic coordinates of each valid ground-based automatic rain gauge that has undergone quality control. As the target estimated location, the geographic coordinates of each ground automatic rain gauge are used. A circular region with a preset radius is defined as the receptive field of the model; all satellite infrared radiation data and their coordinates within the receptive field are extracted.

[0029] Specifically, the process of time-aligning the ground-based automatic rain gauge observation data and the extracted four-channel satellite infrared radiation data includes: using the actual observation time of the satellite data as the reference point for time matching, extracting the continuous observation sequence of the corresponding ground-based automatic rain gauge from 5 minutes before to 5 minutes after that time, averaging these 10 minutes of data, multiplying by 60, and converting them into precipitation intensity. The conversion formula is as follows: ; in, This represents the converted precipitation intensity, serving as the actual precipitation label for that satellite observation time. Number the ground automatic rain gauge station. Indicates the time of satellite data observation. Indicates automatic rain gauge station The intensity of precipitation as observed at any given time.

[0030] The specific process of outlier cleaning includes: using physical extreme value checks to remove negative values ​​caused by mechanical tipping errors and unrealistic extreme precipitation rates; then using time consistency checks to eliminate obviously isolated abnormal peak values; and finally using matched satellite data to cross-validate and identify and eliminate environmental interference.

[0031] The specific process of local spatial matching based on the receptive field includes: matching the geographic coordinates of each valid automatic ground rain gauge that has undergone quality control. The coordinates of the rain gauge are used as the target estimated location. A circular region is defined as the model's receptive field, centered on the model. All satellite infrared radiation data and their coordinates within this receptive field are extracted. The coordinate constraint formula is: ; in The radius of the receptive field.

[0032] In one embodiment, step 2 includes: in the input module: estimating the coordinates of the target point. Define a radius for the center of the circle. The receptive field is defined, and the relative coordinates of satellite infrared radiation data sampling points relative to the target estimation points within the receptive field are calculated; where... To determine the radius of the receptive field; and to extract the area centered at the target estimation point with a side length of... The local square satellite infrared radiation data matrix is ​​obtained, and a spatial mask matrix is ​​constructed using this receptive field range: ; in, For spatial mask matrix, The relative coordinates of the satellite infrared radiation data sampling points relative to the target estimation points.

[0033] The two relative coordinate channels and the four channels of satellite infrared radiation data are concatenated along the channel dimension and multiplied spatially with the spatial mask matrix to obtain the initialized input feature tensor.

[0034] Specifically, the coordinates of the target estimated point Define a radius for the center of the circle. The receptive field. Calculate the relative coordinates of satellite infrared radiation data sampling points relative to the target estimation points. : ; ; in, These are the coordinates of the satellite infrared radiation data sampling points.

[0035] And extract with Centered on, with side length as A local square satellite infrared radiation data matrix was obtained, and a spatial mask matrix was constructed using this receptive field. : ; Two relative coordinate channels are concatenated with four infrared channels along the channel dimension, and spatially aligned with a spatial mask matrix. Multiplying them yields the initialized input feature tensor with dimension . ,in For batch size, This represents the six channels after being spliced ​​together.

[0036] In one embodiment, the adaptive spatial feature encoder in step 2 includes: two Convolutional module and a global average pooling layer The The convolution module includes: Convolutional layers, batch normalization, and the GELU activation function.

[0037] Specifically, adaptive spatial feature encoder Including: two Convolutional layer and a global average pooling layer This is used to extract spatial features from the input feature tensor of the mask constraint (i.e., the local cloud map of the mask constraint). The feature extraction layer of the adaptive spatial feature encoder. correspond Layer convolution kernel size Step length ,filling Spatial downsampling layer of adaptive spatial feature encoder correspond Layer convolution kernel size Step length ,filling Reduce spatial dimensions to The output dimension is obtained by applying global average pooling. The embedding vector, denoted by its dimension. .

[0038] In one embodiment, the dynamic branching network in step 2 includes: a first Fully connected layer, first GELU activation function, second Fully connected layer.

[0039] Specifically, dynamic branching networks The high-dimensional embedding vector is compressed into the basis vectors of the precipitation operator, and the dimension of the target basis vector is set to be... ,Require Dynamic branching network Including two Fully connected layer. Input dimension is The embedding vector, first Fully connected layer (transition hidden layer) The number of hidden nodes is set to Output feature dimension Following LayerNorm and the first GELU activation function, the second... Fully connected layer (output layer) The number of nodes is The output dimension is branch basis vectors .

[0040] In one embodiment, the backbone network in step 2 includes: a third Fully connected layer, second GELU activation function, fourth Fully connected layer, third GELU activation function, fifth Fully connected layer.

[0041] Specifically, independent backbone network Map low-dimensional geographic coordinates to spatial background basis vectors, and set the dimension of the target basis vectors to be [value missing]. Unlike branch basis vectors, independent backbone networks include: three Fully connected layer. The input layer has 2 nodes, and the third... Fully connected layer (transition hidden layer) The structure is ,fourth Fully connected layer (transition hidden layer) The structure is The number of nodes are respectively and Then followed by the third GELU activation function, and the fifth... Fully connected layer (output layer) The number of nodes is The output dimension is Main basis vectors .

[0042] In one embodiment, the fusion module in step 2 includes: the sixth Fully connected layer and the fourth GELU activation function, the seventh Fully connected layer.

[0043] Specifically, the fusion module Used for nonlinear combination of branch basis vectors and trunk basis vectors, including: two Fully connected layer. The input branch basis vectors... and the main basis vectors Directly concatenate the feature channels to form a dimension of The asymmetric fusion feature, the sixth Fully connected layer (fused hidden layer) Input feature dimension ,sixth Fully connected layer output feature dimension The fourth GRLU activation function follows; the seventh Fully connected layer (parameter mapping layer) The structure is The output length is The feature vectors are divided into 4 segments of length based on their physical properties. The subvectors are used as feature parameters.

[0044] In one embodiment, a symmetric spatial feature decoder is introduced during the training of the infrared precipitation estimation operator network; the symmetric spatial feature decoder includes: a fully connected layer, batch normalization processing and a fifth GELU activation function, dimension reshaping operation and a first... Convolutional layers and the second Convolutional layer; the symmetric spatial feature decoder is used to decode the embedding vector to obtain the reconstructed infrared tensor; the loss function construction process used in the training process of the infrared precipitation estimation operator network: Based on the reconstructed infrared tensor and the corresponding input feature tensor, the mean square error loss of infrared reconstruction in the masked region is calculated; the expression for the mean square error loss of infrared reconstruction in the masked region is: ; in, The mean square error loss for infrared reconstruction in the mask-constrained region. For mask Total number of valid pixels For the first The input feature tensor of each sample, This is an element-wise multiplication operation. For batch size, To reconstruct the infrared tensor; Based on the predicted precipitation rate and the corresponding actual precipitation rate, the spatial gradient weighted estimation loss is calculated; the expression for the spatial gradient weighted estimation loss is: ; in, Loss is estimated using spatial gradient weighting. For the first The true precipitation rate of each sample These are weighting coefficients, which are dynamically adjusted according to the local spatial gradient of the precipitation field. To predict precipitation rate; The total loss is determined based on the infrared reconstruction mean square error loss of the masked region and the spatial gradient weighted estimation loss; the expression for the total loss is: ; in, For the total loss, For receptive field radius The infrared reconstruction loss weighting coefficient is dynamically and adaptively adjusted.

[0045] Specifically, symmetric space feature decoder To reconstruct the embedded vectors into a two-dimensional true infrared radiation field, a fully connected layer and two two-dimensional transposed convolutional layers are used. Feature reshaping layer. Through the fully connected layer Perform mapping, and for Spatial tensor; Spatial upsampling layer correspond convolution kernel Step length ,filling Spatial upsampling layer correspond convolution kernel Step length ,filling The output dimension is limited to .

[0046] (1) The preparation process of training set, validation set and test set before network training includes: 1) Read multi-channel infrared radiation data from FY-4B satellite and extract... , and Channel infrared data is used to provide spatial distribution information of water vapor and extract... Infrared data from the channel is used to reflect surface temperature information. The readout time resolution is... Quantization resolution is Ground-based automatic rain gauge observation data.

[0047] 2) Data preprocessing: The data from the ground automatic rain gauge observations and the infrared radiation data from the four channels of the FY-4B satellite were time-aligned and outlier cleaning was performed. Then, local spatial matching based on the receptive field was used to obtain the matching sample space.

[0048] 3) Dataset partitioning: For the matching space sample space, the training set and the remaining dataset are obtained by alternating sampling with a monthly time interval. In the remaining dataset, the validation set and the test set are further divided by alternating sampling with a shorter time scale, and the sample size ratio of the two is controlled to be about 1:2.

[0049] (2) Network Training: The training set is matched with the sample space to extract feature parameters through the infrared precipitation estimation operator network. The predicted precipitation rate is calculated through physical hard constraints, and the loss function is calculated to update the network parameters. The training process of the infrared precipitation estimation operator network is as follows: Figure 3 As shown. The specific training process includes: 1) Parameter initialization: Select the model The optimizer updates network parameters, initializes model hyperparameters, and sets optimizer momentum parameters. Numerical stability constant Weight decay coefficient Initial learning rate A cosine annealing decay strategy is adopted, and a minimum learning rate is set. The decay period training rounds are The number of input samples per batch during training Validation set single batch sample size Maximum number of training iterations At the same time, an early stop mechanism is set up so that when the precipitation estimation error of the validation set is continuous... Training will be terminated if there is no significant decline after a certain number of rounds.

[0050] 2) Feature parameter extraction: Let the initial tensor be... The input feature tensor is masked to obtain a dimension of tensor : ; in This is an element-wise multiplication operation. Let be the spatial mask matrix.

[0051] Will Input Adaptive Spatial Feature Encoder Output embedding vector As a dynamic branching network With symmetric space feature decoder The common input and the calculation process are as follows: ; ; ; in, This represents a two-way cross-correlation convolution operation. For adaptive spatial feature encoder The weights and biases of the first convolutional layer. For adaptive spatial feature encoder The weights and biases of the second convolutional layer, The first convolutional layer outputs features. This represents the number of output channels for the first layer. This is the output feature of the second convolutional layer. This represents the number of output channels for the second layer. This is a global average pooling operation.

[0052] embed vector Input symmetric spatial feature decoder Output the reconstructed infrared tensor The calculation process is as follows: ; ; ; in, For feature reshaping layer The weights and biases, Spatial upsampling layer The weights and biases, Spatial upsampling layer Weights and biases for Feature reshaping layer and feature reshaping layer The output, Dimensional reshaping operations.

[0053] embed vector Input dynamic branching network After dimensionality reduction mapping through two fully connected layers, the output branch basis vectors are obtained. , The dimension of the branch basis vectors is calculated as follows: ; ; in, Transition Hidden Layer for Dynamic Branch Networks The weights and biases, Output layer for dynamic branching network The weights and biases, For transition hidden layer Output characteristics.

[0054] coordinates of the target point Input independent backbone network After dimensionality scaling and nonlinear mapping through three fully connected layers, the backbone basis vectors are output. , The calculation process is as follows, taking the dimensions of the main basis vectors as the primary dimension: ; ; ; in, Transition Hidden Layer for Independent Backbone Network The weights and biases, Output layer of independent backbone network The weights and biases, For transition hidden layer Output characteristics.

[0055] Branch basis vectors With the main basis vector Concatenate along the feature dimension and input into the fusion module All Gaussian radial basis functions are obtained through nonlinear mapping. Parameters of the components , The total number of Gaussian radial basis function components is calculated as follows: ; ; in, For feature dimension concatenation operations, To merge hidden layers The weights and biases, For parameter mapping layer The weights and biases, To merge hidden layers Output characteristics.

[0056] 3) Calculation of predicted precipitation rate: The result obtained in step 2-4-1... Decomposed into peak weights along the channel dimension Spatial scale Peak x-axis Peak ordinate And impose meteorological and physical constraints: ; ; ; ; in, For sample index within a batch, for Component index, It is a very small constant. For activation function, The radius of the receptive field to be constructed.

[0057] Based on constraints Parameters, calculate the predicted precipitation rate : ; in, For sample index within a batch, for Component index, These are the coordinates of the target point.

[0058] 4) Loss function calculation: For the reconstructed infrared tensor The infrared reconstruction mean square error loss of the mask-constrained region is calculated using the aforementioned expression for the infrared reconstruction mean square error loss of the mask-constrained region. .

[0059] For predicting precipitation rate The spatial gradient weighted estimation loss is calculated using the above spatial gradient weighted estimation loss expression. Weighting coefficients The dynamic adjustment of the local spatial gradient of the precipitation field is defined as: ; in, For the first Each sample corresponds to the local spatial gradient of the precipitation field. For all observed precipitation gradients 0.99 quantiles after sorting in ascending order The range of values ​​is restricted to between.

[0060] Construct the total loss function as shown in the total loss function expression above. .

[0061] 5) Network parameter update: Complete forward propagation and total loss function. After calculation, the loss is solved by traversing backwards along the network computation graph to obtain the loss for the entire set of trainable parameters of the network. gradient : ; in, For supervised weighted gradient flow, For self-supervised gradient flow, For receptive field radius The infrared reconstruction loss weighting coefficient is dynamically and adaptively adjusted.

[0062] The AdamW optimizer is used to update the parameters, and the current iteration step is denoted as . The parameters in the previous step were The updated parameters are The update process is as follows: ; ; ; ; ; in, Estimation of the first and second moments, respectively. This represents the deviation correction value for the first and second moments. The learning rate for the current iteration step. It is the numerical stability constant. This is the weight decay coefficient.

[0063] (3) Perform unbiased evaluation of the validation set at the end of each training round, and use cosine learning rate decay and early stopping mechanism to extract the optimal hyperparameters.

[0064] 1) Unbiased evaluation on the validation set: Switch the network model's running mode to evaluation mode, forcibly freeze the moving mean and variance updates of all batch normalized layers in the network, and disable all randomly deactivated layers. Traverse the complete validation set within the gradient blocking context. Calculate the core indices for global precipitation estimation on the validation set: ; ; in, Precipitation estimation error measures the overall deviation between the predicted precipitation rate and the actual precipitation rate. The critical success index measures the model's overall ability to detect precipitation events. The total number of samples in the validation set, The number of samples that were correctly predicted. The number of empty reported samples To account for the number of underreported samples, For the model to the first Precipitation rate prediction for each validation sample. For the first The actual precipitation rate observed by the rain gauges for each validation sample.

[0065] 2) Optimal hyperparameter selection: Setting an early stop counter Early cessation of patience threshold If the current round Validation set If the precipitation estimation error of the current round of validation set is the lowest value since the start of training, then the current network hyperparameters are saved as follows: and reset the early stop counter. If the current round of validation set ,but ,when If the loop terminates prematurely, the learning rate is used to schedule and update the next round. learning rate To achieve smooth decay of the learning rate, the cosine annealing learning rate update formula is: ; in, The initial learning rate, To minimize the learning rate, For cosine annealing cycle, This represents the current completed training round.

[0066] Training reaches the maximum number of iterations Or, when the early stopping mechanism is triggered, the optimal set of hyperparameters is obtained. Training complete. It should be understood that, although the above... Figure 1 The steps are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise explicitly stated in this document, there is no strict order in which these steps are executed; they can be performed in other orders. Furthermore, the above... Figure 1 At least some of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0067] In one embodiment, an infrared instantaneous precipitation inversion device based on operator learning is also provided, comprising: The data preprocessing unit is used to preprocess the acquired satellite infrared radiation data and ground automatic rain gauge observation data to obtain multi-channel satellite infrared radiation data and the coordinates of its sampling points within the sensing field. An infrared precipitation estimation operator network construction and training unit is used to construct and train an infrared precipitation estimation operator network to obtain a trained infrared precipitation estimation operator network. The infrared precipitation estimation operator network includes: an input module, an adaptive spatial feature encoder, a dynamic branch network, a backbone network, a fusion module, and a prediction module. The input module processes multi-channel satellite infrared radiation data to obtain a mask-constrained input feature tensor. The adaptive spatial feature encoder processes the input feature tensor using two-dimensional convolution and global average pooling operations to obtain an embedding vector. The dynamic branch network compresses the embedding vector into branch basis vectors for precipitation operators. The backbone network performs dimensionality scaling and nonlinear mapping on the sampling point coordinates to obtain backbone basis vectors. The fusion module performs nonlinear mapping on the dimensionally concatenated branch basis vectors and backbone basis vectors to obtain the parameters of all Gaussian radial basis function components. The prediction module splits the parameters of all Gaussian radial basis function components along the channel dimension, applies meteorological and physical constraints to the splitting results, and calculates the predicted precipitation rate based on the constrained parameters. The precipitation rate inversion calculation unit is used to input the satellite infrared radiation data to be estimated into a trained infrared precipitation estimation operator network to obtain the local instantaneous precipitation rate of the point to be estimated.

[0068] In one embodiment, the data preprocessing unit is further configured to read multi-channel infrared radiation data from the FY-4B satellite and ground-based automatic rain gauge observation data, extract satellite infrared radiation data from four preset channels from the multi-channel infrared radiation data; after performing time alignment and outlier cleaning on the ground-based automatic rain gauge observation data and the extracted four-channel satellite infrared radiation data, a matching sample space is obtained by employing local spatial matching based on the receptive field. Each sample includes: all satellite infrared radiation data within the receptive field and their coordinates; wherein the process of local spatial matching based on the receptive field includes: matching the geographic coordinates of each quality-controlled valid ground-based automatic rain gauge. As the target estimated location, the geographic coordinates of each ground automatic rain gauge are used. A circular region with a preset radius is defined as the receptive field of the model; all satellite infrared radiation data and their coordinates within the receptive field are extracted.

[0069] In one embodiment, the infrared precipitation estimation operator network construction and training unit is also used in the input module to: estimate the coordinates of the target points. Define a radius for the center of the circle. The receptive field is defined, and the relative coordinates of satellite infrared radiation data sampling points relative to the target estimation points within the receptive field are calculated; where... To determine the radius of the receptive field; and to extract the area centered at the target estimation point with a side length of... The local square satellite infrared radiation data matrix is ​​obtained, and a spatial mask matrix is ​​constructed using this receptive field range: ; in, For spatial mask matrix, The relative coordinates of the satellite infrared radiation data sampling points to the target estimation points; The two relative coordinate channels and the four channels of satellite infrared radiation data are concatenated along the channel dimension and multiplied spatially with the spatial mask matrix to obtain the initialized input feature tensor.

[0070] In one embodiment, the adaptive spatial feature encoder in the infrared precipitation estimation operator network construction and training unit includes: two Convolutional module and a global average pooling layer The The convolution module includes: Convolutional layers, batch normalization, and the GELU activation function.

[0071] In one embodiment, the dynamic branch network in the infrared precipitation estimation operator network construction and training unit includes: a first Fully connected layer, first GELU activation function, second Fully connected layer.

[0072] In one embodiment, the backbone network in the infrared precipitation estimation operator network construction and training unit includes: a third... Fully connected layer, second GELU activation function, fourth Fully connected layer, third GELU activation function, fifth Fully connected layer.

[0073] In one embodiment, the fusion module in the infrared precipitation estimation operator network construction and training unit includes: the sixth... Fully connected layer and the fourth GELU activation function, the seventh Fully connected layer.

[0074] In one embodiment, the infrared precipitation estimation operator network construction and training unit is further configured to introduce a symmetric spatial feature decoder during the training process of the infrared precipitation estimation operator network; the symmetric spatial feature decoder includes: a fully connected layer, batch normalization processing and a fifth GELU activation function, dimension reshaping operation and a first Convolutional layers and the second Convolutional layer; the symmetric spatial feature decoder is used to decode the embedding vector to obtain the reconstructed infrared tensor; The loss function construction process used in the training of the infrared precipitation estimation operator network: Based on the reconstructed infrared tensor and the corresponding input feature tensor, the mean square error loss of the infrared reconstruction in the masked constrained region is calculated as follows: ; in, The mean square error loss for infrared reconstruction in the mask-constrained region. For mask Total number of valid pixels For the first The input feature tensor of each sample, This is an element-wise multiplication operation. For batch size, To reconstruct the infrared tensor; Based on the predicted precipitation rate and the corresponding actual precipitation rate, the spatial gradient weighted estimation loss is calculated as follows: ; in, Loss is estimated using spatial gradient weighting. For the first The true precipitation rate of each sample These are weighting coefficients, which are dynamically adjusted according to the local spatial gradient of the precipitation field. To predict precipitation rate; Based on the infrared reconstruction mean square error loss of the masked constrained region and the spatial gradient weighted estimation loss, the total loss is determined as follows: ; in, For the total loss, For receptive field radius The infrared reconstruction loss weighting coefficient is dynamically and adaptively adjusted.

[0075] It is understood that for a detailed explanation of the operator-learning-based infrared instantaneous precipitation inversion device, please refer to the corresponding explanations of the various embodiments of the operator-learning-based infrared instantaneous precipitation inversion method above, and will not be repeated here. Each module in the above-described operator-learning-based infrared instantaneous precipitation inversion device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in hardware or independently of a device with data processing capabilities, or stored in software in the memory of the aforementioned device, so that the processor can call and execute the operations corresponding to each module. The aforementioned device can be, but is not limited to, various types of data processing computer equipment already existing in the art.

[0076] In one embodiment, a computer device is also provided, including a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps in the method embodiment.

[0077] It is understood that, in addition to the memory and processor mentioned above, the computer equipment described above also includes other hardware and software components not listed in this specification. The specific components can be determined according to the model of the image processing computer in different application scenarios, and will not be listed and described in detail in this specification.

[0078] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0079] The above embodiments are merely illustrative of several implementation methods of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of protection of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and all such modifications and improvements fall within the scope of protection of this application.

Claims

1. A method for inverting infrared instantaneous precipitation based on operator learning, characterized in that, Including the following steps: The acquired satellite infrared radiation data and ground automatic rain gauge observation data are preprocessed to obtain multi-channel satellite infrared radiation data and the coordinates of its sampling points within the sensing field. Construct an infrared precipitation estimation operator network and train it to obtain a well-trained infrared precipitation estimation operator network; The infrared precipitation estimation operator network includes: an input module, an adaptive spatial feature encoder, a dynamic branch network, a backbone network, a fusion module, and a prediction module. The input module processes multi-channel satellite infrared radiation data to obtain a mask-constrained input feature tensor. The adaptive spatial feature encoder processes the input feature tensor using two-dimensional convolution and global average pooling operations to obtain an embedding vector. The dynamic branch network compresses the embedding vector into branch basis vectors for the precipitation operator. The backbone network performs dimensionality scaling and nonlinear mapping on the sampling point coordinates to obtain the backbone basis vectors. The fusion module performs nonlinear mapping on the dimensionally concatenated branch basis vectors and backbone basis vectors to obtain the parameters of all Gaussian radial basis function components. The prediction module splits the parameters of all Gaussian radial basis function components along the channel dimension, applies meteorological and physical constraints to the splitting results, and calculates the predicted precipitation rate based on the constrained parameters. The satellite infrared radiation data to be estimated is fed into a trained infrared precipitation estimation operator network to obtain the local instantaneous precipitation rate at the point to be estimated. The acquired satellite infrared radiation data and ground-based automatic rain gauge observation data are preprocessed to obtain preprocessed results. These preprocessed results include: all satellite infrared radiation data within the receptive field and the coordinates of their sampling points, including: Read multi-channel infrared radiation data from FY-4B satellite and ground automatic rain gauge observation data, and extract satellite infrared radiation data from four preset channels from the multi-channel infrared radiation data; After time alignment and outlier cleaning of the ground automatic rain gauge observation data and the extracted four-channel satellite infrared radiation data, a matching sample space is obtained by using local spatial matching based on the receptive field. Each sample includes: all satellite infrared radiation data within the receptive field and their coordinates. The process of local spatial matching based on receptive fields includes: The geographic coordinates of each valid, quality-controlled automatic ground rain gauge As the target estimated location, the geographic coordinates of each ground automatic rain gauge are used. Centered on a predetermined radius, a circular region is defined as the receptive field of the model. Extract all satellite infrared radiation data and their coordinates within the receptive field; In this process, a symmetric spatial feature decoder is introduced during the training of the infrared precipitation estimation operator network. The symmetric spatial feature decoder includes: a fully connected layer, batch normalization, a fifth GELU activation function, dimension reshaping operations, and a first... Convolutional layers and the second Convolutional layer; the symmetric spatial feature decoder is used to decode the embedding vector to obtain the reconstructed infrared tensor; The loss function construction process used in the training of the infrared precipitation estimation operator network: Based on the reconstructed infrared tensor and the corresponding input feature tensor, the mean square error loss of the infrared reconstruction in the masked constrained region is calculated as follows: in, The mean square error loss for infrared reconstruction in the mask-constrained region. For mask Total number of valid pixels For the first The input feature tensor of each sample, This is an element-wise multiplication operation. For batch size, To reconstruct the infrared tensor; Based on the predicted precipitation rate and the corresponding actual precipitation rate, the spatial gradient weighted estimation loss is calculated as follows: in, Loss is estimated using spatial gradient weighting. For the first The true precipitation rate of each sample These are weighting coefficients, which are dynamically adjusted according to the local spatial gradient of the precipitation field. To predict precipitation rate; Based on the infrared reconstruction mean square error loss of the masked constrained region and the spatial gradient weighted estimation loss, the total loss is determined as follows: in, For the total loss, For receptive field radius The infrared reconstruction loss weighting coefficient is dynamically and adaptively adjusted.

2. The infrared instantaneous precipitation inversion method based on operator learning according to claim 1, characterized in that, In the input module: Estimated coordinates of the target point Define a radius for the center of the circle. The receptive field is defined, and the relative coordinates of satellite infrared radiation data sampling points relative to the target estimation points within the receptive field are calculated; where... The radius of the receptive field; And extract the target estimated point coordinates as the center, with a side length of... The local square satellite infrared radiation data matrix is ​​obtained, and a spatial mask matrix is ​​constructed using this receptive field range: in, For spatial mask matrix, The relative coordinates of the satellite infrared radiation data sampling points to the target estimation points; The two relative coordinate channels and the four channels of satellite infrared radiation data are concatenated along the channel dimension and then multiplied in space with the spatial mask matrix to obtain the initialized input feature tensor.

3. The infrared instantaneous precipitation inversion method based on operator learning according to claim 1, characterized in that, The adaptive spatial feature encoder includes: two Convolutional module and a global average pooling layer The The convolution module includes: Convolutional layers, batch normalization, and the GELU activation function.

4. The infrared instantaneous precipitation inversion method based on operator learning according to claim 1, characterized in that, Dynamic branching networks include: First Fully connected layer, first GELU activation function, second Fully connected layer.

5. The infrared instantaneous precipitation inversion method based on operator learning according to claim 1, characterized in that, The backbone network includes: the third Fully connected layer, second GELU activation function, fourth Fully connected layer, third GELU activation function, fifth Fully connected layer.

6. The infrared instantaneous precipitation inversion method based on operator learning according to claim 1, characterized in that, The fusion module includes: the sixth Fully connected layer and the fourth GELU activation function, the seventh Fully connected layer.

7. An infrared instantaneous precipitation inversion device based on operator learning, characterized in that, The device employs the operator-based infrared instantaneous precipitation inversion method described in claim 1 to achieve infrared instantaneous precipitation inversion; comprising: The data preprocessing unit is used to preprocess the acquired satellite infrared radiation data and ground automatic rain gauge observation data to obtain multi-channel satellite infrared radiation data and the coordinates of its sampling points within the sensing field. An infrared precipitation estimation operator network construction and training unit is used to construct and train an infrared precipitation estimation operator network to obtain a trained infrared precipitation estimation operator network. The infrared precipitation estimation operator network includes: an input module, an adaptive spatial feature encoder, a dynamic branch network, a backbone network, a fusion module, and a prediction module. The input module processes multi-channel satellite infrared radiation data to obtain a mask-constrained input feature tensor. The adaptive spatial feature encoder processes the input feature tensor using two-dimensional convolution and global average pooling operations to obtain an embedding vector. The dynamic branch network compresses the embedding vector into branch basis vectors for precipitation operators. The backbone network performs dimensionality scaling and nonlinear mapping on the sampling point coordinates to obtain backbone basis vectors. The fusion module performs nonlinear mapping on the dimensionally concatenated branch basis vectors and backbone basis vectors to obtain the parameters of all Gaussian radial basis function components. The prediction module splits the parameters of all Gaussian radial basis function components along the channel dimension, applies meteorological and physical constraints to the splitting results, and calculates the predicted precipitation rate based on the constrained parameters. The precipitation rate inversion calculation unit is used to input the satellite infrared radiation data to be estimated into a trained infrared precipitation estimation operator network to obtain the local instantaneous precipitation rate of the point to be estimated.

8. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and when the processor executes the computer program, it implements the steps of the infrared instantaneous precipitation inversion method based on operator learning as described in any one of claims 1 to 6.

Images