Resunet quantitative precipitation estimation method and system fusing station precipitation and radar echo

Abstract

A ResUNet quantitative precipitation estimation method and system fusing station precipitation and radar echo, the quantitative precipitation estimation method comprising: acquiring radar live data and precipitation observation data of ground stations; interpolating the precipitation observation data of the ground stations to generate original data of a full-area grid; cutting the original data of the full-area grid to generate an original data set, and obtaining a model data set based on balanced samples of the original data set; dividing the model data set into a model training data set and a model validation data set according to a first proportion; training a ResUNet model, training the ResUNet model based on the loss of comparison between model training output data and ground station precipitation observation values, and fusing station observation precipitation into the model; saving the trained ResUNet model, processing the output result of the ResUNet model, and splicing the processed output result according to the full-area grid to generate a full-area grid quantitative precipitation estimation result. The application realizes reliable quantitative precipitation live with an hourly update and a 1km spatial resolution.

Description

Technical Field

[0001] This invention belongs to the technical field of radar inversion quantitative precipitation estimation, and in particular relates to a ResUNet quantitative precipitation estimation method and system that integrates station precipitation and radar echo. Background Technology

[0002] Quantitative precipitation data provides crucial information for hydrology and disaster prevention and mitigation, offering strong data support for decision-making in flood control and drought relief, and providing historical data for more accurate numerical model simulations and weather forecasts. Therefore, quantitative precipitation observation is an important means of disaster prevention and mitigation, enabling improvements in resource coordination and utilization efficiency, reducing casualties and property losses caused by floods, landslides, droughts, and other disasters, while also providing richer data for research on the global water cycle and the Earth's climate system.

[0003] Currently, quantitative precipitation observations mainly come from three sources: ground-based station observations, precipitation radar observations, and meteorological satellite remote sensing observations. While station-based rain gauge observations have high accuracy, they are concentrated at the station location, making it difficult to provide large-scale, dense observations. Furthermore, single-point observations cannot represent the precipitation situation in the surrounding area. Compared to station observations, precipitation radar can provide precipitation observations over a wider area with higher accuracy. However, radar's direct observation data is reflectivity, which needs to be converted into precipitation. Traditional radar reflectivity-to-precipitation conversion mainly relies on the ZR relationship between rain-measuring radar reflectivity (Z) and rainfall intensity (R). However, the coefficients of the ZR relationship are affected by raindrop spectra and vary significantly due to factors such as region and season. Each region and different season has different coefficients to adapt to this. Currently, three fixed ZR relationships are generally used, dynamically optimized over time for nationwide application. While this can alleviate errors in fixed ZR relationships across different seasons to some extent, it is difficult to reduce errors caused by regional variations.

[0004] Meteorological satellites can provide more comprehensive precipitation observations, but they cannot directly obtain precipitation data. They need to establish a mapping relationship with ground stations or radar observations to retrieve precipitation data, which reduces the accuracy of rainfall prediction. Furthermore, the precipitation data retrieved by multiple satellites may be based on the observations of payloads on multiple satellites, which limits the spatial resolution, generally to about 10km or less.

[0005] Currently, domestic quantitative precipitation estimation (QPE) products include two-source fusion products that combine US CMORPH (CPC Morphiing Technique) precipitation estimates with station observations, and three-source fusion products that further combine radar observations. CMORPH precipitation estimates primarily rely on satellite inversion, resulting in relatively low accuracy in China. Radar precipitation estimates still utilize a fixed Zr relationship, and their accuracy varies significantly across different regions. Fusion methods first correct for systematic errors in radar and satellite-derived precipitation using probability matching based on station observations. Then, Bayesian model averaging is used to fuse satellite and radar precipitation. Finally, optimal interpolation is used to integrate station-observed precipitation into the product. However, due to the spatially systematic bias in the fused products, optimal interpolation often produces elliptical, "bull's-eye" spurious precipitation data.

[0006] Therefore, it is necessary to propose a ResUNet quantitative precipitation estimation method that integrates station precipitation and radar echo to solve the problem of inaccurate precipitation prediction caused by the ZR relationship transformation and optimal interpolation. Summary of the Invention

[0007] The purpose of this invention is to provide a ResUNet quantitative precipitation estimation method that integrates station precipitation and radar echo data to solve the problem of inaccurate precipitation estimation caused by existing inherent ZR relationship transformation and optimal interpolation. This application utilizes the ResUNet model, incorporating ground station precipitation observation data as priors into a full-area grid. It matches radar reflectivity, the distance of grid points from the radar center, and the grid point latitude, longitude, and surface elevation data to achieve hourly updates and reliable quantitative precipitation data with a 1km spatial resolution. This provides high-quality hourly precipitation data support for disaster prevention and mitigation, decision-making, and climate system research.

[0008] The objective of this invention and the technical solution for achieving it are as follows.

[0009] This invention provides a ResUNet quantitative precipitation estimation method that integrates station precipitation and radar echo data. The quantitative precipitation estimation method includes:

[0010] Acquire real-time radar data and precipitation observation data from ground stations; interpolate the precipitation observation data from ground stations to generate raw data for the entire region's grid;

[0011] The original data of the entire grid is segmented to generate an original dataset, and a model dataset is obtained based on the balanced sample distribution of the original dataset; the model dataset is then divided into a model training dataset and a model validation dataset according to a first ratio.

[0012] Training the ResUNet model includes training the ResUNet model using a loss function that compares the model's training output data with precipitation observations from ground stations.

[0013] Save the trained ResUNet model, process the output of the ResUNet model, and stitch the processed output into a full-region grid to generate a full-region grid quantitative precipitation estimation result.

[0014] Optionally, the raw data for generating the full-area grid by interpolating precipitation observation data from ground stations includes:

[0015] Precipitation observation data from ground stations are interpolated in a first-resolution full-area grid using nearest-neighbor methods to generate the original full-area grid data.

[0016] Optionally, the raw data for generating the full-area grid by interpolating precipitation observation data from ground stations in a nearest-neighbor manner within a first-resolution full-area grid includes:

[0017] Precipitation observation data from ground stations are interpolated into a full-area grid with a resolution of 1 (km) * 1 (km) using the nearest neighbor method;

[0018] A mask is obtained from the precipitation observation data of the ground stations. The mask is then matched with the latitude and longitude, distance of grid points from the radar center, and elevation data of the entire region to generate the original data of the entire region grid.

[0019] Optionally, the step of cutting the original data of the entire region grid to generate the original dataset includes:

[0020] The original data map, cut into a full-area grid, generates multiple original data maps of local regions. These multiple original data maps of local regions constitute the original dataset, wherein:

[0021] The original data maps of the multiple local regions are matched in size with the data map input to the ResUNet model, and the position of the original data map of each local region in the original data map of the whole region grid is recorded.

[0022] Optionally, generating multiple local region original data maps from the original data map by cutting the entire area grid includes:

[0023] The original data map of the whole region grid is slidably cut by using the method of overlapping areas between the original data maps of adjacent local regions to generate original data maps of multiple local regions.

[0024] Optionally, obtaining the model dataset based on balancing the sample distribution of the original dataset, and dividing the model dataset into a model training dataset and a model validation dataset according to a first ratio, includes:

[0025] The model dataset is generated by removing observation data samples from the original dataset where the number of ground station precipitation points within a preset first threshold range is less than a preset second threshold.

[0026] The model dataset is divided into a model training dataset and a model validation dataset according to the first ratio.

[0027] Optionally, after obtaining the model dataset based on the balanced sample distribution of the original dataset, and before training the ResUNet model, the following steps are also included:

[0028] The precipitation data of ground stations in the original dataset are added to the network in a priori manner, encoded using different network links, and finally combined with other encoded features.

[0029] Optionally, the ResUNet model can be trained based on the AdamW optimizer.

[0030] Optionally, training the ResUNet model using a loss function that compares the model's training output data with precipitation observations from ground stations includes:

[0031] The model training output data and precipitation observations from ground stations are compared and loss is calculated; the comparison and loss calculation are based on the weighted mean square error (MSE) as the loss function.

[0032] The evaluation of the ResUNet model based on the mean absolute error (MAE) includes: calculating the mean absolute error (MAE) for different precipitation levels; scoring the result maps for different precipitation levels based on the texture clarity of the result maps of the ResUNet model at different precipitation levels; the horizontal axis of the result map represents the methods and features of different ResUNet models, and the vertical axis represents the mean absolute error (MAE); the scoring is based on the principle that the lower the mean absolute error (MAE), the higher the score.

[0033] The method and features of the ResUNet model used for quantitative precipitation estimation are determined based on the highest score.

[0034] This invention also provides a ResUNet quantitative precipitation estimation system that integrates station precipitation and radar echo data, for implementing the above-described method, characterized in that it includes:

[0035] The data acquisition and interpolation module is used to acquire real-time radar data and precipitation observation data from ground stations; it interpolates the precipitation observation data from ground stations to generate raw data for the entire region grid.

[0036] The dataset generation and partitioning module is used to cut the original data of the entire region grid to generate the original dataset, and obtain the model dataset based on the balanced sample distribution of the original dataset; the model dataset is divided into the model training dataset and the model validation dataset according to a first ratio;

[0037] The model training module is used to train the ResUNet model, including training the ResUNet model with a loss based on the comparison between the model training output data and precipitation observations from ground stations.

[0038] The quantitative precipitation estimation module is used to save the trained ResUNet model, process the output of the ResUNet model, and stitch the processed output into a full-region grid to generate a full-region grid quantitative precipitation estimation result.

[0039] This invention has significant advantages and beneficial effects compared with the prior art. Through the above technical solution, this invention has at least one of the following advantages and beneficial effects:

[0040] I. The ResUNet quantitative precipitation estimation method merging station precipitation and radar echo provided by this invention involves: acquiring radar real-time data and precipitation observation data from ground stations; interpolating the precipitation observation data from ground stations to generate raw data for a full-area grid; segmenting the raw data of the full-area grid to generate a raw dataset; obtaining a model dataset based on balancing the sample distribution of the raw dataset; dividing the model dataset into a model training dataset and a model validation dataset according to a first ratio; training the ResUNet model, including training the ResUNet model using a loss function based on comparing the model training output data and precipitation observation values ​​from ground stations; saving the trained ResUNet model; processing the output results of the ResUNet model; and stitching the processed output results together according to the full-area grid to generate a quantitative precipitation estimation result for the full-area grid. This application utilizes the ResUNet model to incorporate precipitation observation data from ground stations as priors into a full-area grid. By matching radar reflectivity, the distance of grid points from the radar center, and the latitude, longitude, and elevation data of grid points, it achieves reliable quantitative precipitation data with hourly updates and 1km spatial resolution. This provides high-quality hourly precipitation data to support disaster prevention and mitigation, decision-making, and climate system research.

[0041] II. This invention interpolates precipitation observation data from ground stations into a 1km*1km resolution full-area grid using a nearest-neighbor method to generate the original data for the full-area grid. By interpolating the precipitation observation data from ground stations into a 1km*1km resolution full-area grid using a nearest-neighbor method, and by obtaining a mask for the precipitation observation data from the ground stations, the original data for the full-area grid is generated by matching the mask with the latitude, longitude, distance from the grid point to the radar center, and elevation data of the entire area. This application ensures the comprehensive generalization and accuracy of the generated original data for the full-area grid by interpolating the first-resolution full-area grid into a 1km*1km resolution full-area grid using a nearest-neighbor method. By matching the mask with the latitude, longitude, and elevation data of the entire area to generate the original data for the full-area grid, the generalization of the model across different terrains is enhanced, achieving hourly updates of quantitative precipitation data at 1km resolution in different regions and seasons.

[0042] Third, this invention uses overlapping areas between the original data maps of adjacent local regions to slide and cut the original data map of the whole region grid, thereby generating multiple original data maps of local regions. This ensures the continuity of the quantitative precipitation estimation results generated by stitching the processed output results together according to the whole region grid. By using the method of cross-cutting large maps, the spatial continuity between the original data maps of adjacent local regions is preserved, making the results obtained by the model more continuous in the whole region grid space, and improving the accuracy of the whole region result estimation.

[0043] Fourth, the present invention generates the model dataset by removing observation data samples from the original dataset where the number of ground station precipitation points within a preset first threshold range is less than a preset second threshold; and divides the model dataset into a model training dataset and a model validation dataset according to the first ratio, thereby ensuring the accuracy of the final output of the full-area result estimate.

[0044] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention, it can be implemented according to the contents of the specification. In order to make the above and other objects, features and advantages of the present invention more concise and easy to understand, preferred embodiments are given below, and detailed descriptions are provided in conjunction with the accompanying drawings. Attached Figure Description

[0045] Figure 1 This is a flowchart illustrating a ResUNet quantitative precipitation estimation method that integrates station precipitation and radar echoes according to an embodiment of the present invention.

[0046] Figure 2This is a flowchart illustrating another embodiment of the ResUNet quantitative precipitation estimation method that integrates station precipitation and radar echoes.

[0047] Figure 3a This is a schematic diagram of radar data used for training a ResUNet model according to an embodiment of the present invention;

[0048] Figure 3b This is a schematic diagram of a ground station corresponding to a network dataset according to an embodiment of the present invention;

[0049] Figure 4a This is a schematic diagram of the mean absolute error (MAE) score results for precipitation levels of 0 to 8 mm according to an embodiment of the present invention.

[0050] Figure 4b This is a schematic diagram of the mean absolute error (MAE) score results for precipitation levels of 8 to 20 mm according to an embodiment of the present invention.

[0051] Figure 4c This is a schematic diagram of the mean absolute error (MAE) score results for precipitation levels of 20 to 50 mm according to an embodiment of the present invention.

[0052] Figure 4d This is a schematic diagram of the mean absolute error (MAE) score results for precipitation levels greater than 50 mm according to an embodiment of the present invention.

[0053] Figure 5a This is a schematic diagram of ground station observation results according to an embodiment of the present invention;

[0054] Figure 5b This is a schematic diagram of the estimation result of the ResUNet model according to an embodiment of the present invention;

[0055] Figure 6 This is a schematic diagram of the structure of a ResUNet quantitative precipitation estimation system that integrates station precipitation and radar echo according to an embodiment of the present invention;

[0056] Figure 7 This is a schematic diagram illustrating the operational deployment of quantitative precipitation estimation based on fusion site observations and radar observations, according to another embodiment of the present invention. Detailed Implementation

[0057] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the specific implementation methods, structures, features and effects of the present invention will be described in detail below with reference to the accompanying drawings and preferred embodiments.

[0058] Precipitation shapes the living environment of life on Earth, influences daily human life, and is a crucial component of the water cycle and energy exchange within the Earth's climate system. Timely and adequate precipitation ensures normal human production activities and provides essential water for plant and animal growth, while extreme water scarcity and excessive precipitation can easily lead to disasters such as droughts, floods, or landslides. Quantitative precipitation data provides vital information for hydrology and disaster prevention and mitigation, offering strong data support for flood and drought control decisions, and providing historical data for more accurate numerical model simulations and weather forecasts. Therefore, quantitative precipitation observation is an important means of disaster prevention and mitigation, enabling improvements in resource coordination and utilization efficiency, reducing casualties and property losses caused by floods, landslides, and droughts, and providing richer data for research on the global water cycle and the Earth's climate system.

[0059] This invention provides a ResUNet quantitative precipitation estimation method that integrates station precipitation and radar echo data to address the inaccurate precipitation estimation caused by existing inherent ZR relationship transformations and optimal interpolation. This application utilizes the ResUNet model, incorporating ground station precipitation observation data as priors into a full-region grid. The raw data for the full-region grid is generated by matching radar reflectivity, the distance of grid points from the radar center, and the latitude, longitude, and elevation data of the grid points. This achieves reliable, quantitative precipitation updates at the hourly level with a 1km spatial resolution, providing high-quality hourly precipitation data support for disaster prevention and mitigation, decision-making, and climate system research.

[0060] An embodiment of the present invention provides a ResUNet quantitative precipitation estimation method that integrates station precipitation and radar echo intensity, as shown in the appendix. Figure 1 As shown, the ResUNet quantitative precipitation estimation method includes:

[0061] S1: Acquire real-time radar data and precipitation observation data from ground stations; interpolate the precipitation observation data from ground stations to generate raw data for the entire region grid;

[0062] S2, the original data of the entire grid is cut to generate the original dataset, and the model dataset is obtained based on the balanced sample distribution of the original dataset; the model dataset is divided into the model training dataset and the model validation dataset according to the first ratio;

[0063] S3, Training the ResUNet model, including training the ResUNet model with a loss based on the comparison between the model's training output data and precipitation observations from ground stations;

[0064] S4. Save the trained ResUNet model, process the output of the ResUNet model, and stitch the processed output into a full-region grid to generate a full-region grid quantitative precipitation estimation result.

[0065] It should be noted that one embodiment of the present invention provides a ResUNet quantitative precipitation estimation method that integrates station precipitation and radar echo intensity. This method uses precipitation data from ground stations as the target and radar echo intensity from radar observations as a feature. Loss calculations and ResUNet model training and learning are performed at the precipitation observation points of the ground stations, as shown in the attached figure. Figure 2 The following is a flowchart illustrating a ResUNet quantitative precipitation estimation method that integrates station precipitation and radar echoes, according to another embodiment of the present invention. The specific implementation steps are as follows.

[0066] (1) Obtain real-time data of radar echo intensity of Caiyun weather and precipitation observation data of ground stations in a certain country, such as China. The time span of the precipitation observation data is within 2020. Since precipitation in China is mainly concentrated in June to August each year, only the precipitation observation data from June to August is selected.

[0067] (2) Remove outliers from the precipitation observation data of the selected ground stations within China and interpolate to a 1km resolution grid using the nearest neighbor method. Obtain the mask of the precipitation observation data of the selected ground stations and match it with the national radar echo intensity data, longitude and latitude coordinate data and the elevation data corresponding to the longitude and latitude coordinates, such as (digital elevation model DEM data).

[0068] (3) To create the original dataset, since the 1km resolution grid data of the entire China region is large (the data matrix of the entire region is 6000x7500 pixels in size. If the data in the data matrix corresponds to the actual distance of the entire region, then one pixel represents 1km*1km. The data matrix of the entire region is represented by a size of 6000x7500 pixels), the grid data of the large map of the entire China region is first cut, and the small maps of multiple local regions generated after the cutting (such as 512x512 pixels in size) are used to form the original dataset.

[0069] (4) Balancing the original dataset samples: Due to the uncertainty of precipitation across China and the prevalence of small precipitation in local precipitation observation data from ground stations, the original dataset exhibits an imbalance in data distribution, with fewer precipitation points than no precipitation and fewer large precipitation points than small precipitation points. Therefore, the original dataset is balanced by removing samples from the ground station precipitation observation data where the number of observed precipitation points is less than a preset threshold. This yields the final dataset (e.g., 100 precipitation observation data points from 100 ground stations correspond to 100 pixels; removing the number of observed precipitation points corresponding to a precipitation observation data point less than a preset precipitation observation threshold or the preset threshold number is used as an example). To verify the performance and generalization ability of the ResUNet model for quantitative precipitation estimation, the final dataset is divided into a model training dataset and a model validation dataset according to a preset ratio (e.g., a 4:1 ratio).

[0070] (5) Ground station observations are added to the China-wide data network in a priori manner. When the quantitative precipitation estimation ResUNet model is being learned and trained, the AdamW optimizer is used for learning and training. The loss function can be either weighted mean square error (MSE) or mean absolute error (MAE) to evaluate the performance of the ResUNet model in quantitative precipitation estimation.

[0071] (6) Train the ResUNet model. The ResUNet model is trained and learned. The loss is calculated by comparing the training result data output by the ResUNet model network with the real values ​​observed by the ground station. The ResUNet model network is trained based on the loss calculated by comparison.

[0072] (7) When the loss calculated by the network comparison of the ResUNet model meets the preset loss requirements, save the trained ResUNet model to obtain the ResUNet quantitative precipitation estimation model that integrates the precipitation observed by the site and the radar echo intensity.

[0073] (8) Post-process the results output by the ResUNet quantitative precipitation estimation model, and combine the small images of the local areas after post-processing (e.g., 512x512 pixels) into a result dataset and stitch it back to the original full-area large image size according to the original cut mask and longitude and latitude coordinate data.

[0074] (9) To examine the performance and generalization of the ResUNet quantitative precipitation estimation model, the mean absolute error (MAE) at different magnitudes was calculated, and the texture clarity of the resulting map of the original full-area map size was evaluated. The final feature combination method with the lowest mean absolute error (MAE) and the best texture clarity effect used in this invention can be selected through the evaluation.

[0075] This invention primarily utilizes nationwide radar mosaic and ground station rainfall observations, combined with regional surface elevation data, grid distance data from the radar center, and longitude and latitude data, to learn hourly rainfall estimates for a preset time period using the ResUNet model. This ResUNet model differs from traditional fixed ZR relationships used for radar-based rainfall estimation and station interpolation methods. It can directly utilize hourly radar echo intensity observations and ground station rainfall data collection to simultaneously estimate and interpolate quantitative rainfall across the entire region. This solves the problem that the same ZR relationship cannot be universally applied across different regions and seasons, and also addresses the limitations of optimal interpolation algorithms, such as limited estimation scope and susceptibility to "bull's-eye" false rainfall estimates. Furthermore, by incorporating surface elevation data and grid distance data from the radar center, this application enhances the ResUNet model's generalization ability across different terrains, achieving accurate and reliable hourly quantitative rainfall updates with 1km resolution across various regions and seasons.

[0076] Optionally, the process of interpolating precipitation observation data from ground stations to generate raw data for the full-area grid includes: interpolating precipitation observation data from ground stations in a nearest-neighbor manner within a first-resolution full-area grid to generate raw data for the full-area grid.

[0077] It should be noted that the precipitation observation data from ground stations can be interpolated using the nearest neighbor method within a full-area grid of a first resolution size (e.g., 6000x7500 pixels, where each pixel represents a 1km*1km area) to generate the original full-area grid data. The nearest neighbor method can include, but is not limited to, at least one of the following: left side, right side, top side, bottom side, top left, top right, bottom left, and bottom right side of the precipitation observation data from the ground stations. Alternatively, an inverse distance-weighted interpolation method can be used.

[0078] Optionally, the raw data for generating the full-area grid by interpolating precipitation observation data from ground stations in a nearest-neighbor manner within a first-resolution full-area grid includes:

[0079] Precipitation observation data from ground stations are interpolated into a full-area grid with a resolution of 1 (km) * 1 (km) using the nearest neighbor method;

[0080] A mask is obtained from the precipitation observation data of the ground stations. The mask is then matched with the latitude and longitude, distance of grid points from the radar center, and elevation data of the entire region to generate the original data of the entire region grid.

[0081] It should be noted that after interpolating the precipitation observation data from ground stations to a 1km*1km resolution full-area grid using the nearest neighbor method, the 1km*1km resolution full-area grid data corresponding to the precipitation observation locations of the ground stations is obtained. A mask for the precipitation observation data from the ground stations is also obtained. This mask data is then matched with the latitude and longitude coordinates of the actual radar echo intensity data for the entire area, as well as the elevation data corresponding to the latitude and longitude coordinates of the actual radar echo intensity data for the entire area, to generate the original data for the full-area grid. By matching the precipitation observation data from the ground stations with the actual radar echo intensity data, the latitude and longitude coordinates of the actual data, and the elevation data corresponding to the actual radar echo intensity data for the entire area through mask identification, the accurate correlation between the training results output by the ResUNet model and the true values ​​observed by the ground stations is ensured when calculating the loss.

[0082] In one embodiment of the present invention, radar echo intensity observation data from different sources or precipitation observation data from different ground stations can also be selected for interpolation in a nearest neighbor manner to generate the original data of the whole area grid.

[0083] Optionally, the step of cutting the original data of the entire region grid to generate the original dataset includes:

[0084] The original data map, cut into a full-area grid, generates multiple original data maps of local regions. These multiple original data maps of local regions constitute the original dataset, wherein:

[0085] The original data maps of the multiple local regions are matched in size with the data map input to the ResUNet model, and the position of the original data map of each local region in the original data map of the whole region grid is recorded.

[0086] It should be noted that the original data maps of the multiple local regions are matched in size with the data map input to the ResUNet model, ensuring the accuracy of each local region result data map generated by the ResUNet model after matching the latitude and longitude coordinates of the real-world radar echo intensity data for the entire region with the elevation data corresponding to the latitude and longitude coordinates of the real-world data. The position of the original data map of each local region in the original data map of the entire region grid is recorded. This is used to accurately stitch each local region result data map back to the corresponding position in the original data map of the entire region grid according to the recorded position, accurately stitching together to generate the result data map of the entire region grid, thus ensuring the accuracy of the result data map of the entire region grid.

[0087] Optionally, generating multiple local region original data maps from the original data map by cutting the entire area grid includes:

[0088] The original data map of the entire grid is cut by using overlapping areas between the original data maps of adjacent local regions to generate original data maps of multiple local regions.

[0089] It should be noted that segmenting the original data map of the full-area grid by using overlapping areas between adjacent local regions is not limited to segmenting with multiple different overlapping area sizes. For example, in the original data map of the full-area grid, a small overlapping area size can be used to segment where the elevation data of the Digital Elevation Model (DEM) is relatively flat. This reduces the number of original datasets generated from segmentation into multiple local regions, reducing the amount of data required for ResUNet model computation and improving its efficiency while ensuring the continuity of the resulting data maps of adjacent local regions. Conversely, in the original data map of the full-area grid, when the elevation data of the DEM changes significantly, exceeding a preset threshold, a large overlapping area size can be used to segment, ensuring the continuity of the resulting data maps of adjacent local regions when stitching them together.

[0090] In a preferred embodiment, the result data map of the overlapping area can be averaged to obtain the result data map of the overlapping area. Furthermore, the result data map of the overlapping area can be smoothed based on the size and elevation data changes of the overlapping area to ensure the continuity of the result data maps of adjacent local areas in the result data map of the overlapping area.

[0091] Optionally, obtaining the model dataset based on balancing the sample distribution of the original dataset, and dividing the model dataset into a model training dataset and a model validation dataset according to a first ratio, includes:

[0092] The model dataset is generated by removing observation data samples from the original dataset where the number of ground station precipitation points within a preset first threshold range is less than a preset second threshold.

[0093] The model dataset is divided into a model training dataset and a model validation dataset according to the first ratio.

[0094] It should be noted that ground station precipitation points in the original dataset with fewer than a preset threshold number can be removed. For example, in a 512x512 data matrix, each data point can represent the precipitation data of one ground station. Samples of precipitation points in the aforementioned region with precipitation exceeding a preset first threshold of 0 mm are filtered and summarized. If the number of filtered and summarized ground station precipitation points is less than a preset second threshold (not limited to any value between 150 and 70), then the observation data samples with fewer than the preset threshold in the original dataset are removed, and a model dataset is generated. The model dataset is divided into a model training dataset and a model validation dataset according to the first ratio. To ensure that the model training dataset meets the requirements for quantity and full-region characteristics, and to simultaneously ensure the accuracy of quantitative estimation of the model across the entire region after validation training, the ratio of model training dataset to model validation dataset extracted from the entire region is not limited to a range of 3:1 to 5:1.

[0095] Optionally, after obtaining the model dataset based on the balanced sample distribution of the original dataset, and before training the ResUNet model, the following steps are also included:

[0096] The precipitation data of ground stations in the original dataset are added to the network in a priori manner and encoded using different network links. The features of the network with added ground station precipitation and the encoded network links are then combined.

[0097] It should be noted that the precipitation data from ground stations in the original dataset are added to the network in a priori manner and encoded using different network links. The features of the network with added ground station precipitation and the encoded network links are combined. This feature combination method enables deep fusion of the two data sources, station-observed precipitation and radar-estimated precipitation, to ensure the integrity and accuracy of the training data for the ResUNet model.

[0098] Optionally, the ResUNet model can be trained based on the AdamW optimizer.

[0099] It should be noted that the AdamW optimizer, in training the ResUNet model, comprehensively considers the first-moment estimate (mean of the gradient) and second-moment estimate (variance of the gradient) of the loss gradient by comparing the ResUNet model training output data with precipitation observations from ground stations, thus calculating a better new training step size. When calculating the exponential moving average of the comparative loss gradient, it also considers the gradient momentum of previous time steps; similarly, when calculating the exponential moving average of the squared comparative loss gradient, it considers the influence of the squared gradient of previous time steps. Therefore, the initialization of the comparative loss gradient is significantly biased towards zero, especially in the early stages of training. It can also correct the deviation of the mean of the comparative loss gradient and the mean of the squared gradient. In updating the training step size parameter, it can adaptively adjust from both the mean of the comparative loss gradient and the mean of the squared gradient, rather than being directly determined by the currently trained comparative loss gradient. This application optimizes the contrastive loss algorithm by employing the AdamW optimizer, thereby improving the efficiency of contrastive loss calculation and ResUNet model training. During the contrastive loss calculation process, by stabilizing and optimizing the contrastive loss and training the ResUNet model, it exhibits good generalization performance even when the data samples in the original dataset and the model training dataset are sparse, and can naturally optimize and adjust the learning rate in training the ResUNet model.

[0100] Optionally, training the ResUNet model using a loss function that compares the model's training output data with precipitation observations from ground stations includes:

[0101] The model training output data and precipitation observations from ground stations are compared and loss is calculated; the comparison and loss calculation are based on the weighted mean square error (MSE) as the loss function.

[0102] The evaluation of the ResUNet model based on the mean absolute error (MAE) includes: calculating the mean absolute error (MAE) for different precipitation levels; scoring the result maps for different precipitation levels based on the texture clarity of the result maps of the ResUNet model at different precipitation levels; the horizontal axis of the result map represents the methods and features of different ResUNet models, and the vertical axis represents the mean absolute error (MAE); the scoring is based on the principle that the lower the mean absolute error (MAE), the higher the score.

[0103] The method and features of the ResUNet model used for quantitative precipitation estimation are determined based on the highest score.

[0104] It should be noted that, as Figure 3a and 3b This is a schematic diagram for learning quantitative precipitation data. Figure 3a A schematic diagram of radar data used for training the ResUNet model; Figure 3bThis is a schematic diagram of the ground stations corresponding to the network dataset. The network dataset output by training the ResUNet model is compared with the precipitation observation values ​​of the corresponding ground station precipitation observation points through multiple comparisons, and the network grid of the ResUNet model is learned and trained.

[0105] The scoring results are as follows Figure 4a , 4b As shown in 4c and 4d, the horizontal axis represents different methods and features, and the vertical axis represents the mean absolute error (MAE). The values ​​are: radar_only, dist1_dem0, dist1_dem1, dist0_dem1, l on l at, prior, dist1_dem1_prior, dist1_dem1_l on l at, dist1_dem1_l on l at_prior, baseline, and zr. Here, radar_only means the network's input features consist only of radar observations; dist1_dem0, dist1_dem1, and dist0_dem1 respectively indicate that in addition to radar observations, the distance from the grid point to the radar center is added without elevation; both the distance from the grid point to the radar center and elevation are added; and only elevation is added without adding the distance from the grid point to the radar center; prior indicates that the input features are radar observations, and station-observed precipitation is incorporated in a priori manner; dist1_dem1_prior... "or" refers to the input features including radar observations, the distance and elevation of grid points from the radar center, and site-observed precipitation, all incorporated a priori. "dist1_dem1_l on l at" refers to the input features including radar observations, the distance and elevation of grid points from the radar center, and latitude and longitude. "dist1_dem1_l on l at_pr i or" refers to the input features including radar observations, the distance and elevation of grid points from the radar center, and latitude and longitude, all incorporated a priori. "base line" refers to the precipitation calculated using radar observations based on commonly used ZR relationships. "zr" refers to the precipitation calculated using radar observations based on ZR relationships after grid search. "dist1_dem1_l on l at_pr i or" represents the final method and features used in this invention. The scoring results show that for values ​​below 50mm, this ordinate represents the mean absolute error (MAE). The combined method of the final method and features used in "dist1_dem1_l on l at_pr i or" has the lowest MAE and the best texture clarity.

[0106] In one embodiment of the present invention, the example results are shown as follows: Figure 5a and 5b As shown, Figure 5aThis is a schematic diagram of ground station observation results according to an embodiment of the present invention; Figure 5b This is a schematic diagram of the ResUNet model estimation results according to an embodiment of the present invention. The morphology and range of precipitation distribution and the rainfall value in the schematic diagram of the ResUNet model estimation results are basically consistent with those in the schematic diagram of the verified ground station observation results. This verifies the accuracy of the ResUNet quantitative precipitation estimation method of the present invention, which integrates station precipitation and radar echo intensity. In addition, the schematic diagram of the ResUNet model estimation results can also accurately predict the morphology and range of precipitation distribution and the rainfall value in areas without ground station observation results, and it can match the actual rainfall situation.

[0107] This invention also provides a ResUNet quantitative precipitation estimation system that integrates station precipitation and radar echo intensity, for implementing the methods described above, as shown in the appendix. Figure 6 As shown, the ResUNet quantitative precipitation estimation system 100 includes:

[0108] The data acquisition and interpolation module 101 is used to acquire radar real-time data and precipitation observation data from ground stations; and to interpolate the precipitation observation data from ground stations to generate raw data for the entire region grid.

[0109] The dataset generation and partitioning module 102 is used to cut the original data of the whole region grid to generate the original dataset, and obtain the model dataset based on the balanced sample distribution of the original dataset; the model dataset is divided into the model training dataset and the model validation dataset according to the first ratio;

[0110] Model training module 103 is used to train the ResUNet model, including training the ResUNet model with loss based on the comparison between the model training output data and precipitation observations at ground stations.

[0111] The quantitative precipitation estimation module 104 is used to save the trained ResUNet model, process the output of the ResUNet model, and stitch the processed output into a full-area grid to generate a full-area grid quantitative precipitation estimation result.

[0112] It should be noted that the ResUNet quantitative precipitation estimation system 100 mentioned above is the system corresponding to the ResUNet quantitative precipitation estimation method mentioned above. When the ResUNet quantitative precipitation estimation system 100 is running, it executes the ResUNet quantitative precipitation estimation method mentioned above. The ResUNet quantitative precipitation estimation model provided by this invention, which integrates station-observed precipitation and radar echo intensity, mainly includes the following features: (1) It integrates station precipitation observation and hourly radar echo intensity observation, and uses the ResUNet quantitative precipitation estimation system 100 to achieve a real-time precipitation estimation with 1km resolution and hourly update time; (2) This invention replaces the fixed ZR relationship between radar echo intensity and rainfall by adopting the ResUNet model, which is more efficient in the case of thunderstorms. (2) A more accurate quantitative estimation effect was achieved in estimating precipitation; (3) The present invention incorporates the ground station rainfall observation data into the ResUNet quantitative precipitation estimation model network in a priori form, which does not increase the model network parameters, and can achieve interpolation of high-resolution grids in the whole region while quantitatively estimating precipitation, thus obtaining efficient and accurate estimation results in both high rainfall and different levels of precipitation distribution in the whole region; (4) When learning and training the ResUNet model, the present invention considers the rich underlying surface information of the whole region, such as China, and the influence of the distance of different grid points from the radar center, and incorporates the surface elevation data of the whole region and the distance characteristics of different grid points from the radar center, thereby improving the generalization performance of the ResUNet model in different regions and different observation distances. The specific operation steps and processes are described in the above ResUNet quantitative precipitation estimation method, and will not be repeated here.

[0113] In one embodiment of the present invention, as shown in the appendix Figure 7 This is a schematic diagram illustrating the operational deployment of quantitative precipitation estimation based on fusion of site-observed precipitation and radar observations, according to another embodiment of the present invention. The overall implementation process of the present invention in the ResUNet model training and testing phases (e.g.) Figure 2 After the ResUNet model is trained, the business-oriented real-time execution process is deployed on the ResUNet model (e.g., ...). Figure 7 The technical solution adopted in this embodiment includes the following specific steps:

[0114] (1) Obtain the most recent ground station rainfall observation data based on the current time, remove outliers from the ground station rainfall observation data, interpolate to a 1(km)*1(km) resolution full-area grid using the nearest neighbor method, and obtain the mask of the ground station rainfall observation location. For example, ground station rainfall observation data may have missing measurements or abnormal rainfall measurement values ​​due to actual rainfall measurement failures, etc. Therefore, before mapping the ground station rainfall observation data to the ResUNet model grid, outlier detection and removal of the ground station rainfall observation data are performed to ensure the normality of the final ResUNet model grid training and output results.

[0115] (2) Based on the time of the rainfall observation files of ground stations, obtain radar echo intensity observation data for a pre-set period such as 1 hour and a national radar echo intensity mosaic after noise reduction.

[0116] (3) Obtain and match the surface elevation data of the entire region, the distance of grid points from the radar center and the longitude and latitude (l on l at) coordinate data.

[0117] (4) Cut the large grid data map of the whole region, so that the multiple local grid data small maps after cutting match the input size of the ResUNet model, and record the position of each local grid data small map in the large grid data map of the whole region.

[0118] (5) Combine the ground station rainfall observation data, radar echo intensity observation data, and the whole region surface elevation data into the trained ResUNet quantitative precipitation estimation model to perform quantitative precipitation estimation, and output the quantitative precipitation estimation results of each local area.

[0119] (6) Reassemble the quantitative precipitation estimation results of each local area into a large map of quantitative precipitation estimation results of the whole region according to its position in the whole region.

[0120] (7) Save the quantitative precipitation estimation results data and images for the whole region, and publish the quantitative precipitation estimation results images for the whole region to the webpage.

[0121] The ResUNet quantitative precipitation estimation method that integrates station precipitation and radar echo in this embodiment has the following features and functions: (1) It uses the ResUNet model to replace the ZR relationship for quantitative estimation of radar precipitation. Through training and learning the ResUNet model, a more suitable mapping relationship between radar echo intensity observation data and quantitative precipitation is obtained than the ZR relationship, and it has better generalization in different regions and seasons; (2) By adding precipitation data from ground station rainfall observation data to the ResUNet model network in a priori form for training, while estimating radar echo intensity precipitation, the same ResUNet model network is used to achieve high-resolution interpolation of ground stations. This not only simplifies the training of the ResUNet model network, but also makes the R... The ResUNet model performs better in quantitative estimation of high-level rainfall, and the estimation results are more accurate and reliable; (3) By adding the influence factor of the distance between the grid points and the radar center, the ResUNet model can learn the differences in the characteristics and precipitation mapping relationship brought about by the radar observation distance, which enhances the generalization of quantitative precipitation estimation of the ResUNet model; (4) By adding surface elevation data and longitude and latitude coordinate data, the applicability of the ResUNet model in different regions is further enhanced; (5) By using the method of cross-overlapping the large map of the whole region grid, the spatial continuity between two adjacent local area grid small maps is ensured, making the result data or result map obtained by the ResUNet model more continuous in the whole region space, and improving the accuracy of the result data.

[0122] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been described above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content as equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes, and alterations made to the above embodiments without departing from the scope of the present invention and based on the technical essence of the present invention shall still fall within the scope of the present invention.

Claims

1. A ResUNet quantitative precipitation estimation method that integrates station precipitation and radar echo, characterized in that, include: Acquire real-time radar data and precipitation observation data from ground stations, using the precipitation observation data from ground stations as prior data; By matching the mask of precipitation observation data from ground stations with radar echo data, geographic location data of the entire region, distance and elevation data of grid points from the radar center, the precipitation observation data from ground stations is interpolated to generate the original data of the entire region grid. The original data of the full-area grid is segmented to generate an original dataset. The segmented original data map of the full-area grid generates multiple original data maps of local areas. The original data maps of the multiple local areas constitute the original dataset. The original data map of the full-area grid is slidably segmented by having overlapping areas between the original data maps of adjacent local areas to generate multiple original data maps of local areas. The model dataset is obtained based on the balanced sample distribution of the original dataset. The model dataset is divided into a model training dataset and a model validation dataset according to a first ratio. Training the ResUNet model includes training the ResUNet model using a loss function that compares the model's training output data with precipitation observations from ground stations. The model training output data and precipitation observations from ground stations are compared and loss is calculated; the comparison and loss calculation are based on the weighted mean square error (MSE) as the loss function. The evaluation of the ResUNet model based on the mean absolute error (MAE) includes: calculating the mean absolute error (MAE) for different precipitation levels; scoring the result maps for different precipitation levels based on the texture clarity of the result maps of the ResUNet model at different precipitation levels; the horizontal axis of the result map represents the methods and features of different ResUNet models, and the vertical axis represents the mean absolute error (MAE); the scoring is based on the principle that the lower the mean absolute error (MAE), the higher the score. The method and features of the ResUNet model used for quantitative precipitation estimation are determined based on the highest score. Save the trained ResUNet model, process the output of the ResUNet model, and stitch the processed output into a full-region grid to generate a full-region grid quantitative precipitation estimation result.

2. The quantitative precipitation estimation method according to claim 1, characterized in that, The raw data for generating the full-area grid by interpolating precipitation observation data from ground stations includes: Precipitation observation data from ground stations are interpolated in a first-resolution full-area grid using nearest-neighbor methods to generate the original full-area grid data.

3. The quantitative precipitation estimation method according to claim 2, characterized in that, The raw data for generating the full-area grid by interpolating precipitation observation data from ground stations in a nearest-neighbor manner within a first-resolution full-area grid includes: Precipitation data from ground stations are interpolated to 1 km using the nearest neighbor method. In a full-area grid with a resolution of 1 (km); A mask is obtained from the precipitation observation data of the ground stations. The mask is then matched with the latitude and longitude, distance of grid points from the radar center, and elevation data of the entire region to generate the original data of the entire region grid.

4. The quantitative precipitation estimation method according to claim 3, characterized in that, The process of cutting the original data of the entire grid to generate the original dataset includes: The original data maps of the multiple local regions are matched in size with the data map input to the ResUNet model, and the position of the original data map of each local region in the original data map of the whole region grid is recorded.

5. The quantitative precipitation estimation method according to claim 1, characterized in that, The process of obtaining the model dataset based on the balanced sample distribution of the original dataset, and dividing the model dataset into a model training dataset and a model validation dataset according to a first ratio, includes: The model dataset is generated by removing observation data samples from the original dataset where the number of ground station precipitation points within a preset first threshold range is less than a preset second threshold. The model dataset is divided into a model training dataset and a model validation dataset according to the first ratio.

6. The quantitative precipitation estimation method according to claim 5, characterized in that, After obtaining the model dataset based on the balanced sample distribution of the original dataset, and before training the ResUNet model, the following steps are also included: Precipitation data from ground stations in the original dataset is added to the network in a priori manner and encoded using different network links. The features of the network with added precipitation data and the encoded network links are then combined.

7. The quantitative precipitation estimation method according to claim 1, characterized in that, The ResUNet model was trained using the AdamW optimizer.

8. A ResUNet quantitative precipitation estimation system that integrates station precipitation and radar echo, used to implement the method according to any one of claims 1 to 7, characterized in that, include: The data acquisition and interpolation module (101) is used to acquire radar real-time data and precipitation observation data from ground stations; Precipitation observation data from ground stations are interpolated to generate raw data for the entire region's grid. The dataset generation and partitioning module (102) is used to cut the original data of the whole region grid to generate the original dataset, and obtain the model dataset based on the balanced sample distribution of the original dataset; the model dataset is divided into the model training dataset and the model validation dataset according to the first ratio; The model training module (103) is used to train the ResUNet model, including training the ResUNet model with a loss based on the comparison between the model training output data and precipitation observations at ground stations. The quantitative precipitation estimation module (104) is used to save the trained ResUNet model, process the output of the ResUNet model, and stitch the processed output into a full-area grid to generate the quantitative precipitation estimation result of the full-area grid.

Images