Precipitation quantitative retrieval method based on multi-modal remote sensing data and artificial intelligence

Abstract

This invention discloses a quantitative precipitation inversion method based on multimodal remote sensing data and artificial intelligence, belonging to the field of precipitation inversion technology. The method includes the following steps: Step 1: Acquire multimodal remote sensing data within a specified historical period and perform preprocessing; Step 2: Construct a dataset and partition the dataset; Step 3: Extract high-dimensional spatial features from each multimodal remote sensing data in the dataset and complete spatiotemporal matching; Step 4: Construct a cloud type hybrid expert module and output precipitation features corresponding to the cloud type; Step 5: Construct a latent space adaptive neural representation module and output multi-scale location coding features and time offsets of each multimodal remote sensing data; Step 6: Construct a precipitation estimation and fusion module and output quantitative precipitation estimation and fusion products. Therefore, this invention can achieve quantitative precipitation estimation and fusion at any spatial resolution, improving the accuracy of quantitative precipitation estimation and fusion.

Description

Technical Field

[0001] This invention relates to a quantitative precipitation inversion method based on multimodal remote sensing data and artificial intelligence, belonging to the technical field of precipitation prediction. Background Technology

[0002] Precipitation is a fundamental component of the global water cycle and a key hydrological variable in hydrology, climatology, and meteorology. Accurately estimating precipitation and its regional and global distribution has long been a challenging scientific goal. As part of the water cycle, precipitation, through its life cycle, continuously transports water from the ocean to the land, ensuring the renewal of water resources, promoting the exchange of matter and energy between regions, and playing a crucial role in shaping the diversity of landforms and biodiversity. Simultaneously, precipitation is accompanied by the absorption and release of latent heat, which also has a vital impact on energy transport between regions and across different atmospheric layers. Furthermore, accurately obtaining regional precipitation distribution is of great significance for watershed hydrological analysis, water resource planning and management, water conservancy project design and scheduling, flood and drought monitoring, and geological disaster early warning.

[0003] Quantitative precipitation estimation provides accurate and timely information on precipitation rate, amount, and distribution, which is crucial for monitoring, forecasting, and early warning of severe convective weather, as well as water resource management. Weather radar, geostationary satellites, and ground-based rain gauges are the primary means of monitoring and measuring precipitation intensity. Weather radar and geostationary satellites offer higher resolution observations with greater spatial coverage and temporal continuity, and are used in many countries for monitoring and forecasting severe convective weather. Ground-based rain gauges, due to their more accurate precipitation measurements, are commonly used for precipitation intensity retrieval from weather radar and geostationary satellites, as well as for evaluating and correcting biases in numerical models and artificial intelligence-based precipitation forecasts. Typically, quantitative precipitation estimation utilizes the relationship between weather radar reflectivity factor (Z) and precipitation intensity (R). With the development of polarization radar, numerous studies have developed quantitative precipitation estimation methods based on polarization radar observations, including variables such as radar reflectivity factor, differential reflectivity, and differential phase shift rate. These polarization parameters provide physical information about the shape and size of precipitation particles, offering better estimation of precipitation intensity compared to the conventional ZR relationship.

[0004] Both ZR relationship and polarization parameter-based precipitation estimation are based on nonlinear regressions of the relationship between variables such as radar reflectivity factor (Z), satellite brightness temperature, and precipitation intensity. These statistical regression methods are regionally limited, requiring adjustments to coefficients in different regions. They struggle to capture the dynamic evolution of precipitation processes and have poor generalization ability, leading to significant discrepancies between the retrieved precipitation and the actual precipitation. Some researchers have shown that quantitative precipitation estimation based on these statistical methods has substantial errors, especially the ZR relationship-based method.

[0005] In recent years, artificial intelligence (AI) technology has begun to be applied to quantitative precipitation estimation, mainly using machine learning methods to learn the nonlinear mapping between variables observed by weather radar and geostationary satellites and precipitation observed by ground rain gauges. However, these studies primarily use weather radar reflectivity factors and geostationary satellite brightness temperature observations as input variables when building models, and the labels are usually rain gauge precipitation interpolated to a grid. By learning the indirect relationship between variables such as radar reflectivity factor (Z), satellite brightness temperature, and precipitation intensity, quantitative estimation and fusion of precipitation are achieved, but the influence of surrounding rain gauge observations is not considered in the input. This leads to significant differences between the inverted precipitation and the actual precipitation, especially in areas with complex terrain and few rain gauges. In addition, interpolating rain gauges to a grid as labels introduces additional errors, reducing the ability of AI models to learn the mapping between weather radar / geostationary satellite observations and precipitation. Summary of the Invention

[0006] The purpose of this invention is to provide a quantitative precipitation inversion method based on multimodal remote sensing data and artificial intelligence. Based on observations from dual-polarization weather radar, geostationary satellites, and ground rain gauges, artificial intelligence technology is used to construct a quantitative precipitation estimation and fusion model to generate gridded quantitative precipitation estimation and fusion products with high temporal and spatial resolution, providing support for the monitoring and early warning of extreme precipitation.

[0007] To achieve the above-mentioned technical objectives, the present invention will adopt the following technical solution:

[0008] A quantitative precipitation inversion method based on multimodal remote sensing data and artificial intelligence includes the following steps:

[0009] Acquire and preprocess multimodal remote sensing data within a specified historical time period; the multimodal remote sensing data includes weather radar observations, geostationary satellite observations, and ground rain gauge observations; construct a dataset based on the preprocessed multimodal remote sensing data; for each sample in the dataset, the label data is the ground rain gauge corresponding to the given latitude and longitude coordinates at the current time t. i The data is obtained from ground-based rain gauge observations, with the input data being the current time t extracted within a specified radius centered on given latitude and longitude coordinates. i and the previous time t i-1 Weather radar observations, geostationary satellite observations, and the previous time t i-1The dataset includes observations from all other ground rain gauge stations; high-dimensional spatial features of each multimodal remote sensing data in the dataset are extracted, and before extracting the high-dimensional spatial features of the ground rain gauge station observations, the density weights and station influence radii of the ground rain gauge station observations in sparse and dense regions are adaptively adjusted; the temporal information and latitude / longitude coordinate information of each multimodal remote sensing data in the dataset are respectively encoded with temporal information and geographic information, so that the multimodal remote sensing data in the dataset achieves spatiotemporal matching; based on the extracted high-dimensional spatial features of each multimodal remote sensing data, the cloud type hybrid expert module is used to output the high-dimensional spatial features of three precipitation-related cloud types: convective clouds, stratiform clouds, and mixed clouds; based on the extracted high-dimensional spatial features of each multimodal remote sensing data and their matched temporal information and geographic information encoding, the latent space adaptive neural representation module is used to output the multimodal remote sensing data relative to the surrounding environment at any given latitude / longitude coordinates. The multi-scale location coding features between grid points and the time offset of each multimodal remote sensing data relative to weather radar observations and geostationary satellite observations at a given time are used to train the precipitation estimation and fusion module. The outputs of the cloud type hybrid expert module and the latent space adaptive neural representation module are respectively input into the precipitation estimation and fusion module built based on the SE channel attention mechanism. This allows the input high-dimensional spatial features to be adaptively weighted, and higher weights are given to high-dimensional spatial features that are more correlated with precipitation. This outputs the mean and variance of the precipitation estimate, and finally generates a quantitative precipitation estimate and fusion product. The trained precipitation estimation and fusion module outputs the quantitative precipitation estimate and fusion product of the target ground rain gauge at the current time. The inputs are weather radar observations and geostationary satellite observations that are within the preset range of the target ground rain gauge at the current time, as well as all other ground rain gauge observations that were within the preset range of the target ground rain gauge at the previous time.

[0010] As a further optimization of the present invention, the loss function used during the training of the precipitation estimation and fusion module is... Represented as:

[0011] ;

[0012] In the formula: For precipitation estimation and fusion modules based on input Estimated precipitation results; For the corresponding precipitation observations; Indicates uncertainty-weighted loss; The weighting coefficients represent the uncertainty-weighted loss. and The precipitation estimation and fusion modules are based on the input. Estimated mean and standard deviation of precipitation; Indicates the estimated precipitation result Corresponding precipitation observations The average of the absolute differences between them is denoted as loss; express The weighting coefficients of the loss; Indicates weighted quantile loss; The weighting coefficients represent the weighted quantile loss.

[0013] As a further optimization of the present invention, each multimodal remote sensing data in the dataset selects a corresponding feature encoder to extract its own high-dimensional spatial features based on its different temporal and spatial resolutions. Among them, weather radar observations include radar reflectivity factor Z and differential phase shift rate K. DP and differential reflectivity Z DR Geostationary satellite observations include visible light brightness temperature data from the visible light channel, near-infrared brightness temperature data from the near-infrared channel, and infrared brightness temperature data from the infrared channel; ground rain gauge observations include precipitation observations from ground rain gauges; the radar reflectivity factor Z is extracted using a first feature encoder to capture its own high-dimensional spatial features; differential phase shift rate K DP A second feature encoder is used to extract its own high-dimensional spatial features; differential reflectance Z DR The high-dimensional spatial features of the data are extracted using a third feature encoder; the high-dimensional spatial features of the visible light brightness temperature data are extracted using a fourth feature encoder; the high-dimensional spatial features of the near-infrared brightness temperature data and the infrared brightness temperature data are extracted using a fifth feature encoder; the high-dimensional spatial features of precipitation observations are extracted using a discrete station gridding module based on a graph neural network; the discrete station gridding module can adaptively adjust the density weight and station influence radius of ground rain gauge observations located in sparse and dense regions respectively; the geographic and temporal information of each multimodal remote sensing data in the dataset are processed by the geographic and temporal encoding modules to achieve spatiotemporal matching of the geographic and temporal information of each multimodal remote sensing data.

[0014] As a further optimization of the present invention, the first to fourth feature encoders all adopt a 3-layer convolution-batch normalization-activation function module; the fifth feature encoder adopts a 5-layer convolution-batch normalization-activation function module; the discrete station gridding module includes multiple graph neural network convolutional layers and graph neural network fully connected layers. The graph neural network convolutional layers are used for encoding the latitude and longitude coordinates of ground rain gauge observations, and the graph neural network fully connected layers are used for encoding the precipitation values ​​of ground rain gauge observations. Learnable parameters are initialized as the density weights of the corresponding ground rain gauges. At the same time, a learnable parameter is assigned to each ground rain gauge as the station influence radius of the corresponding ground rain gauge. The importance of ground rain gauges corresponding to sparse and dense areas is adaptively adjusted through the learnable density weights and station influence radii to output the gridded precipitation features.

[0015] As a further optimization of the present invention, when constructing a dataset based on preprocessed multimodal remote sensing data, the given latitude and longitude coordinates are obtained in the following way: precipitation events are identified one by one by observing the ground rain gauges in each location; the given latitude and longitude coordinates are the latitude and longitude coordinates of the corresponding ground rain gauges for the identified precipitation events.

[0016] As a further optimization of the present invention, the weather radar observations are from dual-polarization Doppler weather radar; the geostationary satellite observations are from the Himawari 8 / 9 geostationary satellite; the weather radar observations have a temporal resolution of 6 minutes and a spatial resolution of 1 km; the geostationary satellite observations have a temporal resolution of 10 minutes, the visible light brightness temperature data have a spatial resolution of 500 m, and the near-infrared brightness temperature data and infrared brightness temperature data have a spatial resolution of 2 km; the ground rain gauge observations have a temporal resolution of 10 minutes; the constructed dataset includes input data including weather radar observations at the current time and the past 6 minutes, geostationary satellite observations at the current time and the past 10 minutes, and ground rain gauge observations at the past 10 minutes, all extracted with a radius of 9 km centered on the given latitude and longitude coordinates; the label data is the ground rain gauge observation at the current time corresponding to the given latitude and longitude coordinates.

[0017] As a further optimization of the present invention, the preprocessing includes quality control, continuity check and normalization processing; when normalizing the multimodal remote sensing data, the weather radar observation adopts the maximum and minimum value normalization processing method to normalize the weather radar observation to between 0 and 1; the geostationary satellite observation adopts the mean-standard deviation normalization processing method; and the ground rain gauge observation adopts the standard deviation normalization method.

[0018] As a further optimization of the present invention, the cloud type hybrid expert module is constructed based on the sparse hybrid expert model, which includes three sparse MoE layers and distinguishes the precipitation characteristics of convective clouds, stratiform clouds and mixed cloud regions through a gating network learning mechanism.

[0019] As a further optimization of the present invention, the latent space adaptive neural representation module is composed of multiple multilayer sensing mechanisms.

[0020] Another technical objective of this invention is to provide an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the computer program executes the above-described method for quantitative precipitation inversion based on multimodal remote sensing data and artificial intelligence.

[0021] Based on the above-mentioned technical objectives, the present invention has the following advantages compared with the prior art:

[0022] 1. The precipitation quantitative inversion method based on multimodal remote sensing data and artificial intelligence described in this invention, when constructing the dataset, considers the spatial discreteness of ground rain gauge observations and fully considers the influence of ground rain gauge observations from surrounding areas and previous / near times on the quantitative estimation and fusion of precipitation at the target ground rain gauge. Other ground rain gauge observations from the previous time in the surrounding area are selected as one input for each sample in the dataset. A discrete site gridding module, a geographic information encoding module, and a time encoding module based on graph neural networks are introduced to achieve high-dimensional feature extraction of ground rain gauges and rain gauge observations from previous / near times. Simultaneously, an adaptive neural representation module for latent space is introduced to learn arbitrary neural representations in the latent space. The invention employs multi-scale location encoding features of latitude and longitude coordinates relative to surrounding grid points, as well as time offsets relative to weather radar and geostationary satellite input observations. This differs from existing methods that directly interpolate discrete rain gauges. Furthermore, considering the precipitation estimation bias problem of different cloud types, the invention uses a cloud type hybrid expert module to learn precipitation variation characteristics under different cloud types. It also introduces a precipitation quantitative estimation and fusion module based on channel attention mechanism to adaptively weight the output results of the latent space adaptive neural representation module and the cloud type hybrid expert module, thereby achieving precipitation quantitative estimation and fusion at arbitrary spatial resolution and improving the accuracy of precipitation quantitative estimation and fusion.

[0023] 2. This invention addresses the radar reflectivity factor and polarization parameters (differential phase shift rate K) of dual-polarization Doppler weather radar. DP and differential reflectivity Z DR For the brightness temperature observations of visible light, near-infrared and infrared channels of geostationary satellites, different coding modules are designed to extract features, thereby addressing the inconsistency between the temporal and spatial resolution of the above observations.

[0024] 3. In the quantitative precipitation estimation and fusion module, this invention uses mean-variance instead of direct precipitation estimation, thereby quantifying the uncertainty of precipitation estimation, improving the interpretability and generalization ability of the precipitation estimation and fusion model, as well as the physical consistency of the output. Attached Figure Description

[0025] Figure 1 This is a flowchart of the precipitation quantitative inversion method based on multimodal remote sensing data and artificial intelligence as described in this invention.

[0026] Figure 2 These are different observations of a typical strong convective process, where: (a) represents the radar reflectivity factor (REF); (b) represents the differential phase shift rate (K). DP (c) represents differential reflectance (abbreviated as Z). DR (d) represents K DP Column (KDPC for short); (e) represents Z DR The column (ZDRC) represents the precipitation at the corresponding rain gauge station, with larger dots indicating stronger precipitation intensity.

[0027] Figure 3 It is based on the ZR relationship to estimate the scatter distribution of precipitation and related parameters from observations;

[0028] Figure 4 The scatter distribution and related parameters of precipitation estimates (i.e., the vertical axis in the figure, which is obtained based on deep learning) and observations are obtained based on the quantitative precipitation inversion method proposed in this invention. Detailed Implementation

[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the present invention or its application or use. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention. Unless otherwise specifically stated, the relative arrangement, expressions, and values ​​of components and steps set forth in these embodiments do not limit the scope of the present invention. Techniques, methods, and devices known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and devices should be considered part of the specification. In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.

[0030] Example 1

[0031] like Figure 1 As shown in this embodiment, the quantitative precipitation inversion method based on multimodal remote sensing data and artificial intelligence includes the following steps:

[0032] Step 1: Acquire multimodal remote sensing data within a specified historical time period and perform preprocessing:

[0033] In this embodiment, the multimodal remote sensing data includes weather radar observations, geostationary satellite observations, and ground rain gauge observations, wherein:

[0034] Weather radar observation specifically refers to observational data obtained through weather radar, including radar reflectivity factor Z and differential phase shift rate K. DP Differential reflectivity Z DR This includes the corresponding geographic and temporal information. Geostationary satellite observations specifically refer to brightness temperature data obtained through geostationary satellite channels related to precipitation, along with their corresponding geographic and temporal information. Precipitation-related brightness temperature data includes visible light brightness temperature data from the visible light channel, near-infrared brightness temperature data from the near-infrared channel, and infrared brightness temperature data from the infrared channel. Ground-based rain gauge observations refer to observational data from ground-based rain gauges, including precipitation observations and their corresponding geographic and temporal information.

[0035] In this embodiment, in order to construct a dataset for training, testing, and validating the precipitation estimation and fusion module, the acquired multimodal remote sensing data includes observations from dual-polarization Doppler weather radar (6-minute temporal resolution, 1 km spatial resolution), Himawari 8 / 9 geostationary satellite (10-minute temporal resolution and 500m spatial resolution for the visible light channel; and 10-minute temporal resolution and 2 km spatial resolution for the near-infrared and infrared channels), and ground rain gauges from 2019 to 2024.

[0036] The acquired multimodal remote sensing data are preprocessed, including quality control, continuity checks, and normalization.

[0037] In this embodiment, quality control of multimodal remote sensing data mainly involves checking the integrity and consistency of the multimodal remote sensing data and removing abnormal observations.

[0038] The continuity check of multimodal remote sensing data mainly involves checking whether there are any missing observations in each channel of the geostationary satellite. If there are missing moments, the missing data is supplemented by observations at nearby moments.

[0039] When normalizing multimodal remote sensing data, this embodiment employs different normalization methods considering the differences between weather radar observations, geostationary satellite observations, and ground rain gauge observations. Specifically, weather radar observations use a maximum-minimum value normalization method to normalize them to the range of 0-1; geostationary satellite observations use a mean-standard deviation normalization method, which calculates the mean and standard deviation of each channel of the geostationary satellite data before normalization; and ground rain gauge observations use a standard deviation normalization method. This is because precipitation observations from ground rain gauges have a clear zero-value boundary but no clear upper limit. Using mean-standard deviation normalization would lead to a high probability of zero values, so standard deviation normalization is used, which simply involves dividing the precipitation observations from ground rain gauges by their standard deviation.

[0040] Step 2: Constructing and partitioning the dataset:

[0041] When constructing the dataset, based on the multimodal remote sensing data obtained in step one, precipitation events are first identified by observing ground rain gauges to filter out ground rain gauges with small precipitation values ​​(e.g., filtering out ground rain gauges with precipitation values ​​less than 1 mm / h). At the same time, the temporal continuity and spatial consistency of ground rain gauge observations are checked.

[0042] Specifically, when constructing the dataset, based on the multimodal remote sensing data obtained in step one, the current time t is determined... i Can ground-based rain gauge observations identify precipitation events? If the results indicate that the current time t... i If ground rain gauge observations can identify precipitation events, then the current time t is extracted within a specified radius, centered on the latitude and longitude coordinates of the ground rain gauge corresponding to the ground rain gauge observation that identified the precipitation event. i and the previous time t i-1 Weather radar observations, geostationary satellite observations, and the previous time t i-1 All other ground rain gauge observations were used as input data for the dataset, and the ground rain gauge corresponding to the ground rain gauge observation that identified the precipitation event was extracted at the current time t. i The data consists of ground-based rain gauge observations, which serve as the label data for the dataset. In other words, for each sample in the dataset, the input data is the current time *t* extracted within a specified radius centered on a given latitude and longitude coordinate (the given latitude and longitude coordinates are the latitude and longitude coordinates of the ground-based rain gauge observation corresponding to the identified precipitation event). i and the previous time t i-1 Weather radar observations, geostationary satellite observations, and the previous time t i-1 All other ground rain gauge observations, with the label data being the ground rain gauge station corresponding to the given latitude and longitude coordinates at the current time t.i Observations from ground-based rain gauge stations.

[0043] In this embodiment, the weather radar observations are from dual-polarization Doppler weather radar; the geostationary satellite observations are from the Himawari 8 / 9 geostationary satellite. Therefore, the weather radar observations have a temporal resolution of 6 minutes and a spatial resolution of 1 km; the geostationary satellite observations have a temporal resolution of 10 minutes, and the spatial resolution of the visible light brightness temperature data from the visible light channel of the geostationary satellite is 500 m, while the spatial resolution of the near-infrared brightness temperature data from the near-infrared channel and the infrared brightness temperature data from the infrared channel of the geostationary satellite is 2 km. Therefore, the constructed dataset includes input data comprising weather radar observations at the current time and the past 6 minutes, geostationary satellite observations at the current time and the past 10 minutes, and all other ground rain gauge observations from the past 10 minutes, all extracted within a 9 km radius centered on the given latitude and longitude coordinates; the label data consists of ground rain gauge observations from the ground rain gauge corresponding to the given latitude and longitude coordinates within the 10 minutes following the current time.

[0044] Step 3: Extract high-dimensional spatial features from each multimodal remote sensing data in the dataset and complete spatiotemporal matching:

[0045] Considering the inconsistency in temporal and spatial resolution of different observations (including weather radar observations, geostationary satellite observations, and ground rain gauge observations) in this embodiment, different feature encoders are used to extract their respective high-dimensional spatial features. Specifically:

[0046] The radar reflectivity factor Z is extracted using a first feature encoder to capture its own high-dimensional spatial features; the differential phase shift rate K DP A second feature encoder is used to extract its own high-dimensional spatial features; differential reflectance Z DR The dataset employs a third feature encoder to extract its own high-dimensional spatial features; a fourth feature encoder is used to extract the high-dimensional spatial features of visible light brightness temperature data from the visible light channel of geostationary satellites; a fifth feature encoder is used to extract the high-dimensional spatial features of near-infrared brightness temperature data from the near-infrared channel of geostationary satellites and infrared brightness temperature data from the infrared channel of geostationary satellites; precipitation observations from ground rain gauges are extracted using a discrete site gridding module based on a graph neural network; the discrete site gridding module can adaptively adjust the density weight and site influence radius of ground rain gauge observations located in sparse and dense regions respectively; the geographic and temporal information of each multimodal remote sensing data in the dataset is processed by a geographic and temporal encoding module to achieve spatiotemporal matching of the geographic and temporal information of each multimodal remote sensing data.

[0047] In this embodiment, the first to fourth feature encoders are all constructed using a 3-layer convolutional-batch normalization-activation function module (during execution, the data input to the first to fourth feature encoders first undergoes convolution processing through a convolutional layer, then batch normalization processing through a batch normalization layer, and finally activation processing through an activation function module, and the above steps are repeated 3 times to obtain the output). The fifth feature encoder is constructed using a 5-layer convolutional-batch normalization-activation function module (during execution, the data input to the fifth feature encoder first undergoes convolution processing through a convolutional layer, then batch normalization processing through a batch normalization layer, and finally activation processing through an activation function module, and the above steps are repeated 5 times to obtain the output). The discrete station gridding module includes multiple graph neural network convolutional layers and graph neural network fully connected layers. The graph neural network convolutional layers are used to encode the latitude and longitude coordinates of the ground rain gauge observations, and the graph neural network fully connected layers are used to encode the precipitation values ​​of the ground rain gauge observations. Learnable parameters are initialized as the density weights of the corresponding ground rain gauges. At the same time, a learnable parameter is assigned to each ground rain gauge as the station influence radius. The importance of rain gauges in sparse and dense areas is adaptively adjusted through the learnable density weights and station influence radii to output the gridded precipitation characteristics.

[0048] In this embodiment, the geographic and temporal information of each multimodal remote sensing data is processed by the geographic and temporal encoding modules. Specifically, the temporal information of each multimodal remote sensing data is input into the temporal encoding module to encode and align the temporal information. At the same time, the geographic information (latitude and longitude coordinates) of each multimodal remote sensing data is input into the relative position encoder module to extract the relative positional relationship between different grid observations and the center latitude and longitude coordinates, thereby achieving accurate spatial relationship mapping.

[0049] Step 4: Construct a cloud type hybrid expert module and output the precipitation characteristics of the corresponding cloud type:

[0050] The high-dimensional spatial features of each multimodal remote sensing data extracted in step three are input into the cloud type hybrid expert module. The cloud type hybrid expert module described in this embodiment is constructed based on the sparse hybrid expert model and includes three sparse MoE layers. It distinguishes the precipitation characteristics of convective clouds, stratiform clouds and mixed cloud regions through a gating network learning mechanism.

[0051] Step 5: Construct a latent space adaptive neural representation module and output the multi-scale location coding features and temporal offsets of each multimodal remote sensing data:

[0052] Based on the high-dimensional spatial features of each multimodal remote sensing data extracted in step three, and their matching temporal and geographic information codes, a latent space adaptive neural representation module is used to output the multi-scale positional coding features of each multimodal remote sensing data relative to surrounding grid points at given arbitrary latitude and longitude coordinates, as well as the time offset of each multimodal remote sensing data relative to weather radar observations and geostationary satellite observations at a given time. In this embodiment, the latent space adaptive neural representation module is composed of multiple multilayer sensing mechanisms. It can learn the multi-scale positional coding features of different observations (including weather radar observations, geostationary satellite observations, and ground rain gauge observations) relative to surrounding grid points at given arbitrary latitude and longitude coordinates, as well as the time offset of different observations relative to weather radar observations and geostationary satellite observations at a given time. That is, through the latent space adaptive neural representation module, the multi-scale positional coding features of different observations relative to surrounding grid points at given arbitrary latitude and longitude coordinates, as well as the time offset of different observations relative to weather radar observations and geostationary satellite observations at a given time, can be obtained.

[0053] Step 6: Construct a precipitation estimation and fusion module and output quantitative precipitation estimation and fusion products:

[0054] A precipitation estimation and fusion module is constructed based on the SE channel attention mechanism. The outputs of the cloud type hybrid expert module and the latent space adaptive neural representation module are input to train the precipitation estimation and fusion module. This enables adaptive weighting of each high-dimensional spatial feature, and gives higher weights to high-dimensional spatial features that are more correlated with precipitation. The mean and variance of the precipitation estimate are then output, thus obtaining a quantitative precipitation estimation and fusion product.

[0055] The loss function L used in training the precipitation estimation and fusion module is expressed as:

[0056] ;

[0057] In the formula: For input-based Estimated precipitation results; For the corresponding precipitation observations; This represents the uncertainty-weighted loss. The weighting coefficients represent the uncertainty-weighted loss. and Based on input Estimated mean and standard deviation of precipitation; Indicates the precipitation estimate results Corresponding precipitation observations The average of the absolute differences between them is denoted as loss; express The weighting coefficients of the loss; Indicates weighted quantile loss; The weighting coefficients represent the weighted quantile loss.

[0058] Obtaining uncertainty-weighted loss At that time, by input At that time, the precipitation estimation and fusion module outputs the mean and variance of precipitation, and calculates the corresponding probability based on the obtained mean and variance of precipitation. This allows us to obtain the estimated precipitation probability distribution. .

[0059] Weighted quantile loss Represented as:

[0060] ;

[0061] In the formula: Indicates the number of samples; This represents the precipitation observation corresponding to the i-th sample; This represents the precipitation result estimated based on the i-th sample; The weighting coefficients are learnable parameters; different weighting coefficients are set for precipitation estimates in different intervals. .

[0062] Uncertainty-weighted loss during training Primarily used to estimate the uncertainty of quantitative precipitation, MAE loss focuses on estimating light to moderate rainfall, while weighted quantile loss... The focus is on estimating extreme precipitation. The weights of each loss term are automatically optimized using learnable parameters to calculate the final loss. The learning rate is set to 1e-3, and the learning rate adjustment scheme is as follows: if the loss does not decrease for more than 10 epochs, the learning rate is reduced by a factor of 10.

[0063] This example adjusts the model's hyperparameters through multiple experiments based on the above process until the model achieves its optimal performance. The experimental environment configuration used during training is shown in Table 1.

[0064] Table 1 Experimental Environment Information

[0065] Configuration CPU: Intel 6226R * 8 Memory: 768GB Test system Operating system: CentOS 7.6 GPU processor NVIDIA RTX A6000 * 4 Software environment Pytorch2.4, Python3.12, Cuda 12.8

[0066] The output of the trained precipitation estimation and fusion module is a quantitative precipitation estimate and fusion product for any location of the target ground rain gauge. The input consists of weather radar observations and geostationary satellite observations within a preset range (a range centered on the latitude and longitude coordinates of the target ground rain gauge with a specified radius, such as 9km) at the current time, as well as observations from all other ground rain gauges within the preset range of the target ground rain gauge (a range centered on the latitude and longitude coordinates of the target ground rain gauge with a specified radius, such as 9km) at the previous time. Here, the current time is the time corresponding to the precipitation estimate.

[0067] In practical applications, Figure 2 The radar reflectivity factor Z and differential phase shift rate K at the current moment are shown. DP Differential reflectivity Z DR The brightness temperatures of the visible light channel, near-infrared channel, and infrared channel, as well as the observations from surrounding ground rain gauges over the past 10 minutes, are input into the precipitation estimation and fusion module. The precipitation estimation and fusion module will output normalized quantitative precipitation estimates and fusion products with corresponding temporal and spatial resolutions. Then, the corresponding normalization parameters are used for inverse normalization to obtain the final quantitative precipitation estimates and fusion products.

[0068] After applying the quantitative precipitation inversion method described in this embodiment and the traditional ZR relationship method to a target area for quantitative precipitation estimation, it can be seen that the traditional ZR relationship (see Appendix) is superior. Figure 3 The precipitation estimates obtained from previous methods show a significant overestimation, while the quantitative precipitation inversion method proposed in this invention (see Appendix) is significantly more accurate. Figure 4 The method significantly reduced the mean absolute error (MAE) and root mean square error (RMSE). The RMSE decreased from 8.12 mm to 1.53 mm, and the MAE decreased from 4.76 mm to 0.78 mm. The correlation coefficient (CC) increased from 0.62 to 0.66. In summary, the quantitative precipitation inversion method proposed in this invention significantly improves the precipitation estimation results.

[0069] Example 2

[0070] The present invention also provides a storage medium, wherein the computer program stored in the storage medium executes the above-described method for quantitative precipitation inversion based on multimodal remote sensing data and artificial intelligence when running.

[0071] Example 3

[0072] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the above-described method for quantitative precipitation inversion based on multimodal remote sensing data and artificial intelligence through the computer program.

[0073] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0074] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0075] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0076] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0077] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0078] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0079] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A quantitative precipitation inversion method based on multimodal remote sensing data and artificial intelligence, characterized in that, Includes the following steps: Acquire and preprocess multimodal remote sensing data within a specified historical period; the multimodal remote sensing data includes weather radar observations, geostationary satellite observations, and ground rain gauge observations. A dataset is constructed based on preprocessed multimodal remote sensing data; for each sample in the dataset, the label data is the ground rainfall station at the current time t corresponding to the given latitude and longitude coordinates. i The data is obtained from ground-based rain gauge observations, with the input data being the current time t extracted within a specified radius centered on given latitude and longitude coordinates. i and the previous time t i-1 Weather radar observations, geostationary satellite observations, and the previous time t i-1 All other ground-based rain gauge observations; High-dimensional spatial features of each multimodal remote sensing data in the dataset are extracted separately. Before extracting the high-dimensional spatial features of ground rain gauge observations, the density weight and station influence radius of ground rain gauge observations in sparse and dense regions are adaptively adjusted. The time information and latitude and longitude coordinate information of each multimodal remote sensing data in the dataset are encoded with time information and geographic information respectively, so that the multimodal remote sensing data in the dataset can achieve spatiotemporal matching; Based on the high-dimensional spatial features of the extracted multimodal remote sensing data, the cloud type mixing expert module is used to output the high-dimensional spatial features of three precipitation-related cloud types: convective clouds, stratiform clouds, and mixed clouds. Based on the high-dimensional spatial features of each extracted multimodal remote sensing data and their matching temporal and geographic information codes, the latent space adaptive neural representation module outputs the multi-scale positional coding features of each multimodal remote sensing data relative to surrounding grid points at any given latitude and longitude coordinates, as well as the time offset of each multimodal remote sensing data relative to weather radar observations and geostationary satellite observations at a given time. The outputs of the cloud type hybrid expert module and the latent space adaptive neural representation module are respectively input into the precipitation estimation and fusion module built based on the SE channel attention mechanism to train the precipitation estimation and fusion module. This allows the input high-dimensional spatial features to be adaptively weighted, and higher weights to be given to high-dimensional spatial features that are more correlated with precipitation. This outputs the mean and variance of the precipitation estimate, and finally generates a quantitative precipitation estimation and fusion product. The trained precipitation estimation and fusion module outputs a quantitative precipitation estimate and fusion product from the target ground rain gauge station at the current moment, while the input consists of weather radar observations and geostationary satellite observations within the preset range of the target ground rain gauge station at the current moment, as well as observations from all other ground rain gauge stations within the preset range of the target ground rain gauge station at the previous moment.

2. The precipitation quantitative inversion method based on multimodal remote sensing data and artificial intelligence according to claim 1, characterized in that, The loss function used during the training of the precipitation estimation and fusion module. Represented as: ； In the formula: For precipitation estimation and fusion modules based on input Estimated precipitation results; For the corresponding precipitation observations; Indicates uncertainty-weighted loss; The weighting coefficients represent the uncertainty-weighted loss. and The precipitation estimation and fusion modules are based on the input. Estimated mean and standard deviation of precipitation; Indicates the estimated precipitation result Corresponding precipitation observations The average of the absolute differences between them is denoted as loss; express The weighting coefficients of the loss; Indicates weighted quantile loss; The weighting coefficients represent the weighted quantile loss.

3. The quantitative precipitation inversion method based on multimodal remote sensing data and artificial intelligence according to claim 2, characterized in that, Each multimodal remote sensing data in the dataset selects a corresponding feature encoder based on its own temporal and spatial resolution to extract its own high-dimensional spatial features, where: Weather radar observations include radar reflectivity factor Z and differential phase shift rate K. DP and differential reflectivity Z DR Geostationary satellite observations include visible light brightness temperature data from the visible light channel of geostationary satellites, near-infrared brightness temperature data from the near-infrared channel of geostationary satellites, and infrared brightness temperature data from the infrared channel of geostationary satellites; ground rain gauge observations include precipitation observations from ground rain gauges. The radar reflectivity factor Z is extracted using the first feature encoder to extract its own high-dimensional spatial features; Differential phase shift rate K DP A second feature encoder is used to extract its own high-dimensional spatial features; Differential reflectivity Z DR A third feature encoder is used to extract its own high-dimensional spatial features; Visible light brightness temperature data are extracted using a fourth feature encoder to extract their own high-dimensional spatial features; Near-infrared brightness temperature data and infrared brightness temperature data were used to extract high-dimensional spatial features from both using the fifth feature encoder. Precipitation observations employ a discrete station gridding module based on graph neural networks to extract their own high-dimensional spatial features; the discrete station gridding module can adaptively adjust the density weight and station influence radius of ground rain gauge observations located in sparse and dense regions respectively; The geographic and temporal information of each multimodal remote sensing data in the dataset is processed by the geographic and temporal encoding modules to achieve spatiotemporal matching of the geographic and temporal information of each multimodal remote sensing data.

4. The quantitative precipitation inversion method based on multimodal remote sensing data and artificial intelligence according to claim 3, characterized in that, The first to fourth feature encoders all employ a 3-layer convolution-batch normalization-activation function module; The fifth feature encoder uses a 5-layer convolution-batch normalization-activation function module; The discrete station gridding module includes multiple graph neural network convolutional layers and graph neural network fully connected layers. The graph neural network convolutional layers are used to encode the latitude and longitude coordinates of the ground rain gauge observations, and the graph neural network fully connected layers are used to encode the precipitation values ​​of the ground rain gauge observations. Learnable parameters are initialized as the density weights of the corresponding ground rain gauges. At the same time, a learnable parameter is assigned to each ground rain gauge as the station influence radius. The importance of ground rain gauges corresponding to sparse and dense areas is adaptively adjusted through the learnable density weights and station influence radii to output the gridded precipitation characteristics.

5. The quantitative precipitation inversion method based on multimodal remote sensing data and artificial intelligence according to claim 3, characterized in that, When constructing a dataset based on preprocessed multimodal remote sensing data, the given latitude and longitude coordinates are obtained in the following way: Precipitation events are identified one by one using observations from local surface rain gauges; given latitude and longitude coordinates, these are the latitude and longitude coordinates of the corresponding surface rain gauges for the identified precipitation events.

6. The quantitative precipitation inversion method based on multimodal remote sensing data and artificial intelligence according to claim 5, characterized in that, Weather radar observations were obtained from dual-polarization Doppler weather radar data; geostationary satellite observations were obtained from the Himawari 8 / 9 geostationary satellites; the temporal resolution of weather radar observations was 6 minutes and the spatial resolution was 1 km; the temporal resolution of geostationary satellite observations was 10 minutes, the spatial resolution of visible light brightness temperature data was 500 m, and the spatial resolution of near-infrared brightness temperature data and infrared brightness temperature data was 2 km; the temporal resolution of ground rain gauge observations was 10 minutes. The constructed dataset includes input data such as weather radar observations of the current time and the past 6 minutes, geostationary satellite observations of the current time and the past 10 minutes, and ground rain gauge observations of the past 10 minutes, all extracted with a radius of 9km centered on the given latitude and longitude coordinates. The label data consists of ground rain gauge observations at the current time corresponding to the given latitude and longitude coordinates.

7. The quantitative precipitation inversion method based on multimodal remote sensing data and artificial intelligence according to claim 3, characterized in that, Preprocessing includes quality control, continuous inspection, and normalization. When normalizing the multimodal remote sensing data, the maximum and minimum value normalization method is used for weather radar observations to normalize the weather radar observations to the range of 0-1; the mean-standard deviation normalization method is used for geostationary satellite observations. For ground-based rain gauge observations, the standard deviation normalization method was used.

8. The quantitative precipitation inversion method based on multimodal remote sensing data and artificial intelligence according to claim 1, characterized in that, The cloud type hybrid expert module is built on a sparse hybrid expert model, which includes three sparse MoE layers. It distinguishes the precipitation characteristics of convective clouds, stratiform clouds and mixed cloud regions through a gating network learning mechanism.

9. The quantitative precipitation inversion method based on multimodal remote sensing data and artificial intelligence according to claim 1, characterized in that, The latent space adaptive neural representation module is composed of multiple multi-layer sensing mechanisms.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the computer program is run, it executes the precipitation quantitative inversion method based on multimodal remote sensing data and artificial intelligence as described in any one of claims 1 to 9.

Images