Area rainfall inversion method and system based on ER-UNet and random forest model
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHENGDU UNIV OF INFORMATION TECH
- Filing Date
- 2026-02-09
- Publication Date
- 2026-06-09
Smart Images
Figure CN122172197A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of meteorological remote sensing and hydro-meteorological monitoring technology, specifically a method and system for regional rainfall inversion based on ER-UNet and random forest model. Background Technology
[0002] High spatiotemporal resolution acquisition of precipitation data is an important foundation for weather forecasting and disaster prevention and mitigation. In practice, regional precipitation products are usually generated by combining observation methods such as geostationary meteorological satellites, ground-based weather radars, and ground rain gauges.
[0003] Geostationary meteorological satellites have wide coverage and high observation frequency, but their observations mainly consist of cloud top radiation information. The correspondence between cloud top radiation and surface precipitation is affected by cloud microphysical processes, the stage of convection development, and the underlying topography. When using brightness temperature thresholds or empirical statistical relationships for direct inversion, the accuracy is easily affected by regional differences. Ground-based weather radars can provide spatial structure information of precipitation echoes, but due to the curvature of the Earth and topographic obstruction, their coverage is often limited and there are radar blind spots. To enhance the ability to characterize the spatial structure of precipitation, traditional methods use satellite multi-channel observations to train deep learning networks to reconstruct radar composite reflectivity, and then use it to generate precipitation-related products. These methods often use a combination of general networks and general channels, and use a large-scale radar mosaic as the training target, which has constraints in terms of local adaptation and geographical representativeness of training data.
[0004] In plateau regions with complex terrain, such as Xining (the ground-based radar used in this invention is mainly the Xining radar), radar blind spots and obstructions are more prominent. Schemes based on radar as a necessary input or heavily reliant on radar coverage have limited applicability under blind spot conditions. Furthermore, precipitation inversion relying solely on satellite cloud top radiation characteristics lacks stability under local terrain modulation and cloud evolution conditions. Overall, existing technologies still struggle to consistently output accurate, reliable, and spatially continuous gridded ground precipitation results in areas with limited radar coverage. Summary of the Invention
[0005] The purpose of this invention is to provide a method and system for retrieving regional rainfall based on ER-UNet and a random forest model, so as to solve the problems mentioned in the background art.
[0006] To achieve the above objectives, the present invention provides the following technical solution: a regional rainfall inversion method based on ER-UNet and a random forest model. This method uses the radar combined reflectivity calculated from ground-based weather radar volume scan data after data quality control as a supervisory basis. First, it establishes the correspondence between satellite observations and the spatial distribution characteristics of radar combined reflectivity. Then, it combines topography and underlying surface conditions to quantitatively regress rainfall, so that the inversion results can take into account both the spatial structure of precipitation and the ground value calibration requirements, and is applicable to application scenarios where radar coverage is limited or radar data is unavailable.
[0007] During the data preparation phase, multi-channel satellite observation data from Fengyun-4B satellite were acquired. This data underwent radiometric calibration, geolocation, resampling, and standardization to form a multi-channel satellite data cube. Simultaneously, regional ground-based weather radar volume scan data was acquired. This radar volume scan data underwent ground clutter suppression, precipitation attenuation correction, and velocity de-blurring. The radar combined reflectivity was then calculated from the processed radar volume scan data. Subsequently, based on the geolocation results of the satellite observation data, the radar combined reflectivity was spatially registered. The radar combined reflectivity generated from the radar volume scan results corresponding to the satellite observation time was selected to form time-corresponding sample pairs for training.
[0008] In the reflectivity inversion stage, the error residual type U-shaped convolutional neural network model of the encoder-decoder structure is trained using the sample pairs, enabling the model to generate satellite-inverted radar combined reflectivity based on the multi-channel satellite data cube. Since the radar combined reflectivity generated from radar volume scan base data after data quality control is used as the supervision basis in this stage, the consistency of the supervision labels can be improved, and the model can learn a more stable spatial distribution mapping relationship of precipitation.
[0009] In the rainfall regression phase, elevation data and land surface type data corresponding to the area where the ground-based radar is located are acquired, and the rainfall observed by the ground rain gauge is analyzed. Using the satellite-inverted radar combined reflectivity as the main characterization quantity, combined with the elevation data and the land surface type data as covariates, and the rainfall observed by the ground rain gauge as the supervised target variable, a random forest regression model is trained to output the ground rainfall results corresponding to a unified spatial grid.
[0010] In the application phase, the Fengyun-4B satellite multi-channel observation data at the time to be inverted undergoes the same preprocessing. First, the error residual type U-shaped convolutional neural network model generates the satellite-inverted radar combined reflectivity at the time to be inverted. Then, the satellite-inverted radar combined reflectivity, along with the elevation data and the surface type data, is input into the random forest regression model to obtain the regional gridded ground rainfall results. In this application phase, it does not rely on the ground-based weather radar observation data corresponding to the time to be inverted.
[0011] In a preferred embodiment, to improve the characterization of precipitation clouds by multi-channel satellite observation data from Fengyun-4B, the multi-channel satellite observation data is constructed into a satellite feature set before entering the error residual U-shaped convolutional neural network model. The satellite feature set includes at least the brightness temperature features of the tenth, thirteenth, and fourteenth channels of Fengyun-4B, the brightness temperature difference features formed by the brightness temperatures of the thirteenth and fourteenth channels, and the brightness temperature difference features formed by the brightness temperatures of the thirteenth and tenth channels. It further includes the local spatial gradient features of the thirteenth channel and the deep convection index. By combining features such as brightness temperature, brightness temperature difference, and spatial gradient, the satellite feature set can simultaneously reflect physical quantities such as cloud top thermal state, cloud boundary changes, instability, and convective potential, thereby providing a clearer feature basis for the mapping learning between satellite observations and radar combined reflectivity.
[0012] In another preferred embodiment, the ground-based weather radar volume scan data undergoes data quality control before calculating the radar combined reflectivity. The data quality control includes ground clutter suppression, precipitation attenuation correction, and velocity de-ambiguity. The radar combined reflectivity is calculated from the radar volume scan data after the data quality control is completed and serves as the supervision label for the error residual U-shaped convolutional neural network model, making the spatial distribution of the supervision label more consistent, thereby reducing the impact of supervision bias on the first stage of training.
[0013] Furthermore, to ensure a clear correspondence between satellite observations and supervisory labels, the training sample pairs are formed through spatial registration and temporal correspondence. Spatial registration uses the geographic positioning results of the multi-channel satellite observation data of Fengyun-4B as a benchmark, mapping the combined radar reflectivity to a spatial position consistent with the satellite observations. Temporal correspondence is achieved by calculating the combined radar reflectivity from the radar volume scan results corresponding to the satellite observation time, and using this combined radar reflectivity to form the supervisory label. Through the above spatial registration and temporal correspondence, the satellite feature set and supervisory label in the same training sample point to the same precipitation process, reducing training errors caused by sample mismatch.
[0014] In terms of network structure, the error residual U-shaped convolutional neural network model adopts a five-layer encoder-decoder cascade structure. At least one preset convolutional layer in the encoder or decoder uses a first convolutional kernel size, and the remaining convolutional layers use a second convolutional kernel size. The first convolutional kernel size is larger than the second convolutional kernel size. By using a larger convolutional kernel size in the preset convolutional layers, the model can obtain a wider range of spatial correlation information while maintaining the ability to express local changes, so as to support the reconstruction of the spatial distribution of radar combined reflectivity.
[0015] In terms of training strategy, to improve the stability of the training process, L2 regularization and random deactivation regularization are introduced into the error residual type U-shaped convolutional neural network model. At the same time, the training process adopts a composite loss function consisting of mean square error loss, multi-scale structural similarity loss and gradient loss weighted together. The mean square error loss is used to constrain the numerical deviation of reflectivity, the multi-scale structural similarity loss is used to constrain the structural consistency of the reflectivity field, and the gradient loss is used to constrain the boundary consistency of the spatial variation of reflectivity, so that the satellite-inverted radar composite reflectivity output by the model is consistent with the supervision label in both numerical and spatial structure aspects.
[0016] In the second stage of training the random forest regression model, the rainfall observed by the ground rain gauge and the satellite-inverted radar combined reflectivity output in the first stage are spatially registered and temporally correlated to form a regression training sample. The random forest regression model uses the satellite-inverted radar combined reflectivity as the main characterization quantity, and uses elevation data and land surface type data as covariates to participate in the regression modeling. This allows the regression relationship to reflect the influence of topographic relief and underlying surface differences on rainfall distribution, thereby improving the applicability of gridded ground rainfall results in complex terrain areas.
[0017] This invention also provides a regional rainfall inversion system based on ER-UNet and a random forest model. This system uniformly accesses and processes multi-channel satellite observation data from Fengyun-4B, ground-based weather radar volume scan data, ground-based rain gauge observation data, as well as elevation and surface type data. Under the same data processing caliber, it completes training sample construction, two-stage model training, and application-stage inference output. During the training stage, the radar combined reflectivity calculated from radar volume scan data after data quality control is used as a supervision label, enabling the satellite-side model to learn the correspondence between satellite observations and the spatial distribution characteristics of precipitation. In the second stage, elevation and surface type data are introduced as covariates to regress and calibrate the rain gauge values, allowing the output results to adapt to the complex terrain and underlying surface differences of the region. In the application stage, the system can generate gridded ground rainfall results without relying on the ground-based weather radar observation data corresponding to the time of inversion, thus maintaining continuous output of rainfall products even when radar data is limited.
[0018] The system includes a data interface, a memory, and a processor. The data interface is used to receive multi-channel satellite observation data from Fengyun-4B, volume scan data from regional ground-based weather radar, observation data from ground rain gauges, elevation data, and land surface type data. The memory is used to store computer programs and data required for model training and inference. The processor is connected to the memory and is used to execute the computer program and complete the following processing flow: First, the multi-channel satellite observation data of Fengyun-4B satellite is preprocessed to form a satellite modeling dataset. Data quality control is performed on the volume scan data of ground-based weather radar, and the combined radar reflectivity is calculated. Based on this, the satellite modeling dataset and the combined radar reflectivity are spatially registered and temporally correlated to form the first-stage training samples. Based on this, an error residual U-shaped convolutional neural network model is trained, enabling the model to generate satellite-inverted combined radar reflectivity from the satellite modeling dataset.
[0019] Secondly, based on the aforementioned satellite-inverted radar combined reflectivity and fused elevation data and surface type data, the rainfall observed by ground rain gauges is used as the supervised target variable to complete the training of the random forest regression model, so that the random forest regression model can output the ground rainfall results formed on a unified spatial grid.
[0020] During the application phase, the Fengyun-4B satellite multi-channel observation data for the time to be inverted undergoes the same preprocessing as during the training phase. First, the satellite-inverted radar combined reflectivity for the corresponding time is generated. Then, the satellite-inverted radar combined reflectivity, along with elevation data and surface type data, is fed into a random forest regression model to obtain the regional gridded ground rainfall results. During this application phase, ground-based weather radar observation data corresponding to the time to be inverted are not input.
[0021] In a preferred embodiment, when generating satellite data for modeling, the processor constructs a satellite feature set from the multi-channel satellite observation data of Fengyun-4B satellite. The satellite feature set includes at least the brightness temperature features of the tenth, thirteenth, and fourteenth channels of Fengyun-4B satellite, the brightness temperature difference feature formed by the brightness temperatures of the thirteenth and fourteenth channels, and the brightness temperature difference feature formed by the brightness temperatures of the thirteenth and tenth channels. It further includes the local spatial gradient feature and deep convection index of the thirteenth channel. Through the combination of the above features, the satellite feature set can simultaneously reflect the thermal state of cloud tops, changes in cloud boundaries, and instability in the lower and middle troposphere, thereby providing a clearer feature basis for the error residual U-shaped convolutional neural network model to learn the mapping relationship between the spatial distribution of satellite observation and radar combined reflectivity.
[0022] In another preferred embodiment, the processor introduces L2 regularization and employs random deactivation regularization when training the error residual type U-shaped convolutional neural network model. Simultaneously, it optimizes the model using a composite loss function weighted by mean squared error loss, multi-scale structural similarity loss, and gradient loss. The mean squared error loss constrains numerical deviations in reflectivity, the multi-scale structural similarity loss constrains the structural consistency of the reflectivity field, and the gradient loss constrains the boundary consistency of spatial variations in reflectivity. By simultaneously applying numerical and structural constraints during training, the satellite-inverted radar composite reflectivity output by the model is made consistent with the supervision label in spatial distribution and local boundaries, improving the stability of the first-stage reflectivity inversion results and providing a more consistent input representation for the subsequent random forest regression stage.
[0023] The beneficial effects of this invention are as follows: 1. This invention uses radar combined reflectivity calculated from ground-based weather radar volume scan data after data quality control as the first-stage supervision basis. By constructing training samples through spatial registration and observation time matching between satellite observation data and radar combined reflectivity, the echo reconstruction U-shaped convolutional neural network model is trained to output the satellite inversion radar combined reflectivity. This allows spatial structure reconstruction and quantitative regression of rainfall to be implemented in stages. The inversion process has intermediate representations with consistent physical meaning and reduces the structural bias caused by directly regressing rainfall from cloud top radiation, thereby providing an aligned spatial structure basis for subsequent quantitative estimation.
[0024] 2. This invention constructs a specialized satellite feature set based on the characteristics of regional precipitation cloud systems. By selecting specific channel brightness temperature, brightness temperature difference, local spatial gradient, and deep convection index, it characterizes convection development, boundary abrupt changes, and convective instability. In the echo reconstruction U-shaped convolutional neural network model, a five-layer encoding and decoding structure and differential convolution kernel configuration are set up. At the same time, L2 regularization and random deactivation regularization are introduced, and a composite loss constraint consisting of mean square error loss, multi-scale structural similarity loss, and gradient loss is used to keep the model training process stable and enhance the fitting of strong echo boundaries and fine structures, thereby improving the structural consistency and usability of satellite-derived radar combined reflectivity.
[0025] 3. This invention uses rainfall observed by ground rain gauges as the supervision target during the rainfall mapping stage. By integrating satellite-inverted radar combined reflectivity, elevation data, and surface type data through a random forest regression model, the rainfall estimation can reflect the modulation of precipitation distribution by topographic relief and underlying surface differences, thereby improving the spatial continuity and quantitative consistency of complex terrain areas. Furthermore, the inversion output does not require input of ground-based weather radar observation data at the time of inversion, enabling the output of gridded ground rainfall products even in radar blind spots or under limited coverage conditions, thus supporting continuous operational applications. Attached Figure Description
[0026] Figure 1 This is a flowchart illustrating the overall technical process of the present invention. Figure 2 This is a flowchart of the first stage of ER-UNet model training and reflectivity inversion in this invention; Figure 3 This is a flowchart of the second stage of random forest model training and rainfall inversion in this invention; Figure 4 This is the ground-based radar reflectivity map for Case 1; Figure 5 The first-stage model of Case 1 predicts the reflectance map; Figure 6 This is a map showing the predicted precipitation from the second-stage model in Case Study 1. Figure 7 This is the ground-based radar reflectivity map for Case 2; Figure 8 The first-stage model predicts the reflectance map for Case 2. Figure 9 This is a map showing the predicted precipitation from the second-stage model in Case Study 2. Detailed Implementation
[0027] 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0028] like Figures 1 to 9 As shown, this embodiment of the invention provides a regional rainfall inversion method based on ER-UNet and a random forest model. The method adopts a two-stage modeling process: the first stage uses an error residual U-shaped convolutional neural network model to generate satellite-inverted radar combined reflectivity based on the satellite modeling dataset; the second stage uses a random forest regression model to regress the satellite-inverted radar combined reflectivity into ground rainfall results formed on a unified spatial grid based on the fusion of elevation data and surface type data.
[0029] Data Acquisition and Preprocessing Multi-channel satellite observation data from Fengyun-4B was acquired. Radiometric calibration, geolocation, resampling, and standardization were sequentially performed on the multi-channel satellite observation data to obtain a satellite modeling dataset. Simultaneously, volume scan data from regional ground-based weather radar was acquired. Data quality control was performed on the radar volume scan data, including ground clutter suppression, precipitation attenuation correction, and velocity deblurring. After completing data quality control, the reflectivity fields at each elevation angle were mapped to a spatial grid consistent with the satellite modeling dataset. The maximum reflectivity at each elevation angle was then combined on the spatial grid to obtain the combined radar reflectivity. Furthermore, rainfall data observed by ground rain gauges was acquired, along with elevation and surface type data corresponding to the region. The elevation and surface type data were also mapped to the spatial grid.
[0030] Phase 1: Construction of Training Samples and Model Training Based on the spatial location relationships determined by the geolocation results of the satellite modeling dataset, the combined radar reflectivity is spatially registered so that the combined radar reflectivity corresponds to the satellite modeling dataset grid by grid. Based on the satellite observation time, the radar volume scan result with the smallest time difference within a preset time tolerance is selected to generate the combined radar reflectivity, and the time correspondence is completed accordingly. This creates a one-to-one correspondence between the satellite modeling dataset and the combined radar reflectivity, which serves as the first-stage training sample. Based on the first-stage training sample, an error residual U-shaped convolutional neural network model is trained, enabling the model to generate the satellite-inverted combined radar reflectivity corresponding to the combined radar reflectivity from the satellite modeling dataset.
[0031] The first stage of reasoning generates satellite-inverted radar combined reflectivity. For the multi-channel satellite observation data of Fengyun-4B satellite at the time to be inverted, the same preprocessing as in the training phase is performed to obtain the satellite modeling dataset at the time to be inverted. The satellite modeling dataset is then input into the error residual type U-shaped convolutional neural network model trained to obtain the satellite inversion radar combined reflectivity at the time to be inverted.
[0032] Second-stage model training and rainfall inversion In the second phase of training, the combined radar reflectivity retrieved from satellites was used as the main characteristic of the regression model, and elevation data and land surface type data were used as covariates. Rainfall observed by ground rain gauges was used as the supervised target variable. The random forest regression model was trained to establish a mapping relationship from the combined radar reflectivity retrieved from satellites and covariates to the ground rainfall. In order to construct the regression training samples, the rainfall observed by ground rain gauges was mapped to the grid cell of the station in the spatial grid according to its station location, and the observation time period was mapped to the time period of the sample to be trained.
[0033] During the rainfall inversion stage, the satellite-inverted radar combined reflectivity, elevation data, and surface type data at the time to be inverted are input into the trained random forest regression model, and the gridded ground rainfall results for the region are output. During this rainfall inversion stage, ground-based weather radar observation data corresponding to the time to be inverted are not input.
[0034] In this embodiment, a method for retrieving regional rainfall based on multi-channel satellite observation data from Fengyun-4B is provided. This method employs a two-stage modeling process: the first stage trains an error residual U-shaped convolutional neural network model, which generates satellite-retrieved radar combined reflectivity from satellite modeling data; the second stage trains a random forest regression model, which, under the condition of fusing topographic and underlying surface information, regresses the satellite-retrieved radar combined reflectivity to the ground rainfall result formed on a unified spatial grid. During the training stage, supervisory labels are constructed using ground-based weather radar volume scan data; during the rainfall retrieval stage, ground-based weather radar observation data corresponding to the time to be retrieval is not used, thereby ensuring that rainfall retrieval can still be completed when radar data is limited.
[0035] Data acquisition, unified grid and satellite modeling data construction The system acquires multi-channel satellite observation data from the Fengyun-4B satellite in the selected area where the ground-based radar is located. The multi-channel satellite observation data is then subjected to radiometric calibration, geolocation, and resampling in sequence to ensure that each channel observation corresponds to a pixel under a unified geographic coordinate grid. Missing pixels are marked with a mask. Subsequently, the brightness temperature characteristics of each channel are standardized to obtain satellite modeling data.
[0036] In this embodiment, the satellite modeling data includes the following features, all of which are calculated pixel-by-pixel on a unified grid: Brightness temperature characteristics of channels 10, 13, and 14; The brightness temperature difference between the thirteenth and fourteenth channels is calculated per pixel as the brightness temperature of the thirteenth channel minus the brightness temperature of the fourteenth channel. The brightness temperature difference between the thirteenth and tenth channels is calculated per pixel as the brightness temperature of the thirteenth channel minus the brightness temperature of the tenth channel. Local spatial gradient characteristics of the thirteenth channel; deep convection index characteristics.
[0037] Among them, the local spatial gradient features of the thirteenth channel are calculated by two-dimensional finite difference. For any pixel on the unified grid, the brightness temperature difference in the row direction and column direction is calculated respectively, and the difference amplitude is used as the gradient value of the pixel. Boundary pixels are calculated by one-sided difference. The difference results involving missing pixels are processed as missing and marked by a mask.
[0038] The deep convection index (DCI) is characterized by near-surface air temperature (T). 2m and T d2mThe structure is based on the temperature (T) at 2m above the ground, as monitored by local surface meteorological stations, and the lifting index (LI). 2m ) and dew point temperature (T) d2m The sum of the values of the deep convection index (LI) obtained from radiosonde data is then subtracted from the pixel value of the deep convection index, which is used to supplement the characterization of the development of deep convection clouds.
[0039] Radar volume scan data quality control and combined reflectivity supervision tag generation Acquire volumetric scanning data from regional ground-based weather radar; perform data quality control on the volumetric scanning data for the complex terrain. Data quality control includes suppression of ground clutter (isolated points and lines, ground features in specific orientations), suppression of critical refraction and super-refraction clutter (suppression based on special echo morphology and velocity field characteristics), precipitation attenuation correction, and velocity de-blurring processing to reduce the impact of non-precipitation echoes and measurement errors on the reflectivity field.
[0040] After completing data quality control, a radar combined reflectivity is generated on a unified grid as a supervision label. In this embodiment, the radar combined reflectivity is calculated as follows: after mapping the reflectivity fields obtained from volume scanning at each elevation angle to the same horizontal grid, the maximum value of the reflectivity at each elevation angle at the same grid position is taken as the combined reflectivity value at that grid position, thereby forming a two-dimensional radar combined reflectivity map.
[0041] Spatial registration and temporal correspondence between satellites and radars, and construction of the first-stage training samples. To construct the first-stage training samples, a spatial registration and temporal correspondence relationship was established between the satellite modeling data and the radar combined reflectivity.
[0042] Spatial registration uses the geographic positioning results observed by Fengyun-4B satellite as a benchmark, mapping the combined radar reflectivity to a spatial positional relationship consistent with the unified satellite grid, ensuring that the same grid index corresponds to the same geographic location.
[0043] The time correspondence is based on the satellite observation time. Among the radar volume scan data acquired before and after the satellite observation time, the volume scan result with the smallest time difference is selected, and the radar combined reflectivity is calculated from the volume scan result as the corresponding supervision label. This forms a one-to-one corresponding sample pair. The input is the satellite modeling data, and the supervision label is the radar combined reflectivity map, which serves as the first stage training sample.
[0044] Structural constraints and training process of error residual U-shaped convolutional neural network model The first stage uses an error residual type U-shaped convolutional neural network model to map the input satellite modeling data into a predicted combined reflectance map. This model has a symmetrical structure of encoder and decoder. The encoder and decoder fuse multi-scale spatial information through cross-layer connections. In this embodiment, the number of network layers is set to five, which means that the number of downsampling levels at the encoder end and the number of upsampling levels at the decoder end are symmetrically corresponding.
[0045] To enhance the ability to extract large-scale spatial structures while preserving local detail, this embodiment uses the first convolutional module at the encoding end and the last convolutional module at the decoding end as preset convolutional layers. The preset convolutional layers use a first convolutional kernel size, while the remaining convolutional layers use a second convolutional kernel size. The first convolutional kernel size is larger than the second convolutional kernel size. In this embodiment, the first convolutional kernel size is set to 5x5, and the second convolutional kernel size is set to 3x3, and this is kept consistent during the training and inference phases.
[0046] During training, L2 regularization and random deactivation regularization are introduced. L2 regularization is added to the optimization objective in the form of weight decay. Random deactivation regularization is applied to the feature output after the preset convolutional layer. During the training phase, it is randomly set to zero at a fixed dropout rate, and random deactivation is turned off during the inference phase.
[0047] The model training employs a composite loss function to jointly constrain the predicted combined reflectance map and the supervised label combined reflectance map. The composite loss consists of a weighted average of mean squared error loss, multi-scale structural similarity loss, and gradient loss: mean squared error loss is used to constrain pixel numerical deviation; multi-scale structural similarity loss is used to constrain structural consistency at different scales; and gradient loss uses the two-dimensional finite difference gradient difference between the predicted map and the label map on a unified grid as a constraint to strengthen the fitting of boundaries and fine structures. The weight coefficients of the three are fixed before training begins and satisfy non-negative and normalized constraints.
[0048] Sample construction, feature definition and training process of random forest regression model In the second stage, a random forest regression model is used to map the combined reflectance and topographic surface covariates to ground rainfall, obtain rainfall data observed by ground rain gauges, and place the rain gauge locations into a unified grid according to the latitude and longitude of the stations. The supervision target of the grid cell where the station is located is defined as the cumulative rainfall of the rain gauge in the corresponding time period. When the same grid cell contains multiple rain gauge stations, the supervision target of the grid cell is defined as the average of the cumulative rainfall of the multiple stations. The time correspondence is aligned with the rain gauge observation period and the satellite observation time. The combined radar reflectance retrieved from the satellite under the same correspondence is used as the main characterization of the sample.
[0049] Elevation data and land surface type data are acquired and mapped to a unified grid. For each grid cell, a random forest regression model is constructed with input feature vectors. The input features include: satellite-inverted radar combined reflectivity pixel values, elevation pixel values, and land surface type pixel values output from the first stage. When the land surface type is a categorical variable, this embodiment uses one-hot encoding to convert the land surface type into a numerical feature vector, and maintains consistent encoding rules during the training and inference stages. Based on the above samples, the random forest regression model is trained to establish a mapping relationship from input features to ground rainfall.
[0050] Input constraints and output generation in the rainfall inversion stage In the rainfall inversion stage, multi-channel satellite observation data of Fengyun-4B satellite at the time to be inverted is acquired. Satellite modeling data at the time to be inverted is generated according to the same preprocessing and feature construction method as described above, and input into the trained error residual type U-shaped convolutional neural network model to obtain the satellite inversion radar combined reflectivity at the time to be inverted. Subsequently, the satellite inversion radar combined reflectivity, together with the pre-prepared elevation data and surface type data that have been mapped to a unified grid, are input into the trained random forest regression model to output the ground rainfall results formed on the unified spatial grid for the region (the region where the ground-based radar is located).
[0051] In this embodiment, ground-based weather radar observation data corresponding to the time to be inverted is not input during the rainfall inversion stage; the ground-based weather radar volume scan data is only used to construct supervision labels and complete model training during the training stage.
[0052] This embodiment provides a regional rainfall retrieval system based on ER-UNet and a random forest model. The system includes a data interface, a memory, and a processor. The data interface is used to access multi-channel satellite observation data from Fengyun-4B, volumetric scanning data from ground-based weather radar in the selected region, ground rain gauge observation data, elevation data, and land surface type data. The memory is used to store computer programs, training samples, model files, preprocessing parameters, and output products. The processor is used to execute the computer program to complete the construction of training samples, two-stage model training, and inference inversion output.
[0053] To ensure that data from different sources can correspond to each grid in spatial location, this embodiment uses a unified spatial grid as the basic coordinate carrier within the system. In this embodiment, the resampled grid after satellite preprocessing is used as the unified spatial grid, and the same grid definition, the same geographic positioning caliber, and the same resampling strategy are used in both the training and inference phases to avoid data mismatch caused by spatial location drift.
[0054] Satellite modeling data generation After receiving multi-channel satellite observation data from Fengyun-4B, the processor sequentially performs radiometric calibration, geolocation, resampling, and standardization to obtain satellite modeling data. Resampling maps observations from different channels to a unified spatial grid, ensuring pixel-by-pixel correspondence between channels under the same grid index. For missing pixels, a missing pixel mask is generated and maintained consistently during subsequent training and inference. The parameters required for standardization are statistically generated during the training phase and stored along with the model file to ensure the same parameter caliber is used during inference.
[0055] In this preferred implementation, the satellite modeling data, in addition to including multi-channel brightness temperature features, further includes brightness temperature difference features, local spatial gradient features, and deep convection index features. All of these are calculated pixel by pixel on a unified spatial grid. The local spatial gradient features are calculated using two-dimensional finite difference to determine the difference between the brightness temperature in the row and column directions, and the magnitude of the difference is taken as the gradient value. The deep convection index features are constructed from the channel brightness temperature difference. In this embodiment, the difference between the sum of the near-surface air temperature and the dew point temperature and the lifting index is used as the pixel value of the deep convection index. The above feature construction remains consistent during the training and inference phases.
[0056] Radar combined reflectivity supervision tag generation After receiving the volumetric scanning data from the ground-based weather radar, the processor first performs data quality control and then generates the radar composite reflectivity as a supervisory label for the first stage of training. The data quality control includes at least ground clutter suppression, precipitation attenuation correction, and velocity deblurring to reduce the impact of non-precipitation echoes and measurement errors on the reflectivity field.
[0057] After completing data quality control, the reflectivity fields obtained from volume scanning at each elevation angle are mapped to a unified spatial grid. A two-dimensional radar combined reflectivity map is then generated on the unified spatial grid according to a fixed synthesis rule. In this embodiment, the maximum value synthesis rule is adopted: the maximum value of the reflectivity at each elevation angle at the same grid position is taken as the combined reflectivity value at that grid position, thereby forming a two-dimensional radar combined reflectivity map. For areas with missing measurements due to radar obstruction, blind spots, or no effective echo, a radar missing measurement mask is generated, and the missing measurement positions are skipped in the supervised calculation during the training phase to avoid inconsistent supervision labels.
[0058] Spatial registration and temporal correspondence (training sample construction) To construct the first-stage training samples, the processor establishes a spatial registration and temporal correspondence between the satellite modeling data and the radar combined reflectivity. The spatial registration is based on the satellite geolocation results, ensuring that the radar combined reflectivity and the satellite modeling data are on the same unified spatial grid, so that the same grid index corresponds to the same geographical location.
[0059] The time correspondence is based on the satellite observation time. Among the radar volume scan data acquired before and after the satellite observation time, the volume scan result with the smallest time difference is selected, and the radar combined reflectivity is generated from this volume scan result as the corresponding supervision label. In order to ensure the reproducibility of the aperture, this embodiment sets the maximum allowable time difference as the system configuration parameter and determines it according to the radar volume scan interval. When the minimum time difference exceeds the configuration threshold, the sample pair does not participate in the training, thus forming a one-to-one corresponding training sample pair. The input is satellite modeling data, and the supervision label is the radar combined reflectivity map.
[0060] Phase 1 Model Training and Consolidation In training mode, the system trains an error residual U-shaped convolutional neural network model, enabling it to generate satellite-inverted radar combined reflectivity from satellite modeling data. The model adopts a symmetrical encoder and decoder structure and fuses multi-scale spatial information through cross-layer connections. In this embodiment, the number of network layers is fixed at five, which means that the number of downsampling levels at the encoding end and the number of upsampling levels at the decoding end correspond symmetrically.
[0061] To balance the extraction of large-scale cloud structures with the representation of local boundary details, this embodiment sets the first convolutional module of the encoding end and the last convolutional module of the decoding end as preset convolutional layers. The preset convolutional layer uses the first convolutional kernel size, and the remaining convolutional layers use the second convolutional kernel size. The first convolutional kernel size is larger than the second convolutional kernel size. In this embodiment, the first convolutional kernel size is fixed at 5x5 and the second convolutional kernel size is 3x3, and this is kept consistent during the training and inference phases.
[0062] During training, L2 regularization and random deactivation regularization are introduced. The training loss adopts a composite loss function, which is composed of mean square error loss, multi-scale structural similarity loss and gradient loss weighted together. All three types of losses take the predicted radar combined reflectivity map and the supervised tag radar combined reflectivity map as input. The gradient loss uses two-dimensional finite difference to calculate the spatial gradient difference, and skips the calculation at the mask position with a consistent caliber. The weight coefficients of the composite loss are fixed and recorded before the start of training to ensure consistency in caliber during repeated training and inference evaluation.
[0063] After training is completed, the processor will solidify the first-stage model into a model file and write it into the memory, and simultaneously record the corresponding preprocessing parameters, unified spatial grid definition, feature construction aperture and mask processing aperture to ensure that the same processing chain can be reproduced in the inference stage.
[0064] Phase Two: Training and Consolidation of the Random Forest Regression Model In training mode, the system trains a random forest regression model to map the combined reflectivity of satellite-inverted radar with topography and underlying surface covariates to ground rainfall. The processor maps elevation data and surface type data to a unified spatial grid, making them correspond to the combined reflectivity of satellite-inverted radar grid grid by grid.
[0065] Ground rain gauge observations are point-based observations. In this embodiment, a grid-based method is used to construct the monitoring target: the grid cell where the rain gauge station is located is determined according to its latitude and longitude, and the cumulative rainfall of the grid cell in the corresponding time period is used as the monitoring target; when the same grid cell contains multiple rain gauge stations, the average of the cumulative rainfall of the multiple stations is taken as the monitoring target of the grid cell. The time correspondence adopts the same principle as the first stage, so that the input features and the monitoring target point to the same precipitation process.
[0066] For each grid cell, an input feature vector for the random forest regression model is constructed. The input features include at least the satellite-inverted radar combined reflectivity pixel value, elevation pixel value, and land surface type pixel value. When the land surface type is a categorical variable, this embodiment uses one-hot encoding to convert it into a numerical vector, and the encoding rules and feature order are solidified and stored to avoid the drift of input meaning during the inference stage. After training is completed, the random forest regression model is solidified into a model file and written to the memory.
[0067] Inference Inversion Output and Input Constraints In the inference inversion mode, the system outputs the regional gridded ground rainfall results. The processor receives the multi-channel satellite observation data of Fengyun-4B satellite at the time to be inverted, generates the satellite modeling data at the time to be inverted according to the same preprocessing and feature construction method as the training stage, and inputs it into the first-stage model to obtain the satellite inversion radar combined reflectivity at the time to be inverted. Subsequently, the satellite inversion radar combined reflectivity, together with the elevation data and surface type data that have been mapped to a unified spatial grid, are input into the second-stage random forest regression model to output the ground rainfall results.
[0068] In this embodiment, the input in the inference inversion mode does not include ground-based weather radar observation data corresponding to the time to be inverted. The ground-based weather radar volume scan data is only used to construct supervision labels and complete the first stage of model training during the training phase. This allows the system to output continuous rainfall products by relying solely on satellite observations and static geographic covariates even when radar data is missing, coverage is limited, or blind spots exist.
[0069] In this embodiment, unified grid processing of satellite modeling data and construction of satellite feature sets are performed. In this implementation, the processor preprocesses the multi-channel satellite observation data of Fengyun-4B satellite to generate satellite modeling data, and uses the satellite modeling data as input to the echo reconstruction U-shaped convolutional neural network model.
[0070] To ensure that different channels correspond to each other pixel by pixel in spatial location, the processor uses a unified spatial grid as the resampling target. It performs radiometric calibration, geolocation and resampling on each channel observation in sequence, so that each channel corresponds to the same pixel coordinate under the unified spatial grid. The spatial resolution, projection parameters and grid range of the unified spatial grid are recorded by the memory and remain consistent in the sample construction, model training and inversion inference stages.
[0071] To address missing and invalid values from satellite observations, the processor generates a missing mask based on the quality identifier and valid range of the satellite data product. Subsequent feature calculations, sample construction, and model training are all processed using the same missing mask aperture to avoid introducing unrealistic brightness-temperature differences and spatial gradients at the missing locations.
[0072] After completing the unified grid processing described above, the processor calculates the satellite feature set pixel by pixel on the unified spatial grid. The satellite feature set consists of the following features, and each feature is obtained from satellite observation data at the same time after being processed with a consistent missing measurement mask: The brightness temperature characteristics of channels 10, 13, and 14 are obtained by converting the radiation amount of the channel after radiation calibration according to the brightness temperature inversion relationship of the corresponding channel, and then normalizing or standardizing according to the standardization calibrated in the training phase.
[0073] The brightness temperature difference feature between the thirteenth and fourteenth channels, whose pixel value is the difference between the brightness temperatures of the thirteenth and fourteenth channels, is used to characterize the different responses of different infrared channels to differences in cloud top microphysics and cloud phase.
[0074] The brightness temperature difference feature between the thirteenth and tenth channels, whose pixel value is the difference between the brightness temperatures of the thirteenth and tenth channels, is used to characterize the convection development features corresponding to the differences in cloud top thermal state and channel response.
[0075] The local spatial gradient feature of the thirteenth channel is used to characterize the degree of spatial abrupt change in the brightness temperature of the thirteenth channel, so as to enhance the depiction of the boundary and texture changes of convective clouds. Specifically, the processor uses the brightness temperature field of the thirteenth channel as the calculation object, calculates the brightness temperature difference between each pixel and its neighboring pixels in the row direction and column direction for each pixel on the unified spatial grid, and uses the sum of the absolute values of the row direction difference and the absolute values of the column direction difference as the local spatial gradient value of the pixel; one-sided difference is used for boundary pixels; when the difference operation involves pixels marked by missing measurement masks, the local spatial gradient of the pixel is kept as missing measurement and marked by the missing measurement mask.
[0076] The deep convection index feature is used to supplement the indication information of the development degree of deep convection clouds. The processor constructs the deep convection index based on the preset deep convection index calculation rules, using the difference between the sum of the near-surface air temperature and dew point temperature and the lifting index, and obtains the deep convection index feature pixel by pixel on a unified spatial grid. The calculation rules of the deep convection index and the channel brightness temperature difference aperture adopted are recorded by the memory and kept consistent during the training and inference phases.
[0077] The satellite feature set formed in the above manner simultaneously contains cloud top thermal state information, inter-channel difference information, and spatial structure change information, providing a consistent input basis for subsequent mapping learning from satellite observation to radar combined reflectivity spatial distribution. The technical orientation of the above feature specialization is consistent with the disclosure document, which is different from the scheme path that adopts general channel combination and general training target.
[0078] Training Implementation Guidelines for Regularization and Composite Loss In this implementation, the processor trains an echo reconstruction U-shaped convolutional neural network model using the first-stage training samples, so that it outputs a radar combined reflectance map corresponding to the supervision label. The radar combined reflectance of the supervision label is calculated from the volume scan data of the selected area's ground-based weather radar after data quality control. The data quality control includes at least ground clutter suppression, precipitation attenuation correction, and velocity deblurring.
[0079] To improve training stability and suppress overfitting, the processor introduces L2 regularization and random deactivation regularization during training. L2 regularization applies to the network's trainable parameters in a weight decay manner; random deactivation regularization applies to the output features of the convolutional modules, setting some feature responses to zero at a fixed dropout rate during training iterations, and disabling random deactivation during the inference phase to ensure output determinism. The weight decay coefficient and dropout rate are recorded in memory as training configuration parameters and remain consistent during model retraining, evaluation, and inference phases.
[0080] Meanwhile, the processor uses a composite loss function consisting of a weighted average of mean squared error loss, multi-scale structural similarity loss, and gradient loss to optimize the model. The mean squared error loss is used to constrain pixel-by-pixel numerical deviation, the multi-scale structural similarity loss is used to constrain structural consistency at different scales, and the gradient loss is used to constrain the consistency of spatial boundary changes. The gradient calculation of the gradient loss uses a two-dimensional finite difference operator: the difference gradients in the row and column directions are calculated for the predicted radar combined reflectivity map and the supervised tag radar combined reflectivity map, respectively. Then, the difference between the two difference gradients is used to construct the gradient loss to enhance the fitting ability of the echo boundary and fine structure. The weighting coefficients of each loss component in the composite loss are non-negative weights, which are fixed before training begins and recorded by memory, and remain unchanged during training and evaluation.
[0081] For locations where the supervision label is missing or marked by the missing label mask, the processor skips the corresponding pixel when calculating the mean squared error loss, multi-scale structural similarity loss, and gradient loss to avoid interference from the missing supervision during the training process. After training is completed, the model parameters and training configuration parameters are fixed and stored together to ensure that the input feature construction, missing label processing, and loss configuration are strictly consistent with those in the training stage during the inference stage.
[0082] Case 1: Figure 4-6 As shown, the radar reflectivity predicted by the first-stage model in this invention is illustrated. Figure 5 ) and the ground-based radar reflectivity at the corresponding time ( Figure 4 ) and the precipitation predicted by the second-stage model ( Figure 6 The image, showing the time as 14:45 on August 6, 2018, with a resolution of 1km per pixel, indicates that the convection around Xining had reached its peak maturity. The predicted [data / information] can be seen. Figure 5 With the ground-based radar reflectivity at the same moment ( Figure 4 The main meteorological echoes are located in the areas of Haiyan County-Huangyuan County-Huangzhong County and Menyuan County-Datong County, and it is predicted that ( Figure 5 ) and the real-time echo of ground-based radar ( Figure 4 The shape and intensity distribution are basically consistent; the predicted precipitation distribution location ( Figure 6 ) and two banded strong echo regions ( Figure 4 and Figure 5 The results are basically consistent; the predictions are consistent. Figure 6 The area of strongest precipitation (red filled area, maximum precipitation of 58 mm) is close to the actual maximum hourly precipitation on the ground (52 mm), and the predicted location of the strongest precipitation is consistent with the actual location of the strongest precipitation on the ground. This indicates that the prediction results of this invention are relatively accurate (the specific station names and latitude and longitude in the figure are not specified in detail due to confidentiality requirements).
[0083] Case 2: Figure 7-9 As shown, at 12:00 on August 6, 2018 (the area within the green circle in Figure c is the center of the highest actual precipitation, with a maximum precipitation of 22 mm; due to confidentiality requirements, the specific station name and latitude and longitude are not indicated).
[0084] List the radar reflectivity predicted by the first-stage model in this invention ( Figure 8 ) and the ground-based radar reflectivity at the corresponding time ( Figure 7 ) and the precipitation predicted by the second-stage model ( Figure 9 The image, dated August 6, 2018 at 12:00 PM, has a resolution of 1 km per pixel. The predicted values can be seen. Figure 8 With the ground-based radar reflectivity at the same moment ( Figure 7 The main meteorological echoes are concentrated in the northern part of Huangyuan County, Huangzhong County, Xining main urban area, Ping'an County, and Guide County, and it is predicted that ( Figure 7 ) and the real-time echo of ground-based radar ( Figure 8 The shape and intensity distribution are basically consistent; the predicted values are... Figure 9 The area of strongest precipitation (red filled area, maximum precipitation of 27 mm) has a scale of less than 4 km for precipitation greater than 20 mm, which is close to the actual maximum hourly precipitation (22 mm). The predicted location of the strongest precipitation is consistent with the actual location of the strongest precipitation; and it is consistent with the distribution of the main echo areas ( Figure 7 and Figure 8 Compared to the predicted precipitation distribution, Figure 9 There was a large blank area in the central part of the forecast (predicted precipitation was between 0-0.2 mm), but no precipitation occurred in this area. Subsequently, convective precipitation moved northwestward and intensified along with the predicted main rainfall area (distributed on both sides of the central blank area). At 14:45, the actual precipitation was concentrated in two zones, corresponding to the predicted rainfall area at 14:45. Figure 6 The results are basically consistent. This indicates that the prediction results of this invention are relatively accurate (the specific station names and latitude and longitude in the figure are not specified in detail due to confidentiality requirements).
[0085] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.
[0086] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A regional rainfall inversion method based on ER-UNet and random forest model, characterized by: The specific steps include the following: Acquire multi-channel satellite observation data from FY-4B and preprocess it to obtain a satellite dataset for modeling; Acquire ground-based weather radar volume scan data, perform data quality control on the volume scan data and calculate the radar combined reflectivity; align the satellite dataset with the radar combined reflectivity in time and space to construct the first-stage training samples. The ER-UNet model is trained based on the training samples from the first stage to output the combined radar reflectivity retrieved from the satellite. Elevation data and land surface type data are acquired, and rainfall observed by ground rain gauges is used as the supervised target variable to train a random forest regression model. The random forest regression model maps the satellite-inverted radar combined reflectivity with the elevation data and the land surface type data into gridded rainfall. In the rainfall inversion stage, the FY-4B multi-channel satellite observation data at the time to be inverted is input into the ER-UNet model trained to obtain the satellite-inverted radar combined reflectivity at the time to be inverted. The satellite-inverted radar combined reflectivity, along with the elevation data and the land surface type data, is input into the random forest regression model trained to output gridded ground rainfall data. When outputting the gridded ground rainfall data, the ground-based weather radar observation data corresponding to the time to be inverted is not input.
2. The regional rainfall inversion method based on ER-UNet and random forest model according to claim 1, characterized in that: The satellite dataset consists of brightness temperature features of FY-4B channels 10, 13, and 14; brightness temperature difference features between channels 13 and 14; brightness temperature difference features between channels 13 and 10; local spatial gradient features of channel 13; and the deep convection index (DCI).
3. The regional rainfall inversion method based on ER-UNet and random forest model according to claim 2, characterized in that: The data quality control includes ground clutter suppression, precipitation attenuation correction, and velocity de-ambiguity, and the radar combined reflectivity is calculated from the volume scan base data that has undergone the data quality control.
4. The regional rainfall inversion method based on ER-UNet and random forest model according to claim 3, characterized in that: The spatiotemporal alignment includes: spatially registering the radar combined reflectivity based on the geographic positioning results of the FY-4B multi-channel satellite observation data, and selecting the radar volume scan data corresponding to the observation time of the FY-4B multi-channel satellite observation data to calculate the radar combined reflectivity as the supervision label of the first-stage training sample.
5. The regional rainfall inversion method based on ER-UNet and random forest model according to claim 4, characterized in that: When training the ER-UNet model, the ER-UNet model is structurally constrained. The structural constraints include setting the number of network layers to 5, using a first convolutional kernel size for the preset convolutional layers, and using a second convolutional kernel size for the remaining convolutional layers, wherein the first convolutional kernel size is larger than the second convolutional kernel size.
6. The regional rainfall inversion method based on ER-UNet and random forest model according to claim 5, characterized in that: When training the ER-UNet model, L2 regularization and Dropout regularization are introduced, and the ER-UNet model training adopts a composite loss function consisting of mean squared error loss, multi-scale structural similarity loss and gradient loss weighted together.
7. The regional rainfall inversion method based on ER-UNet and random forest model according to claim 6, characterized in that: When training the random forest regression model, the rainfall observed by the ground rain gauge is spatiotemporally aligned with the satellite-inverted radar combined reflectivity, elevation data, and land surface type data, and the satellite-inverted radar combined reflectivity, elevation data, and land surface type data are used as input features on the aligned samples.
8. A regional rainfall inversion system based on ER-UNet and a random forest model, characterized in that: The system includes the regional rainfall inversion method based on ER-UNet and random forest model as described in any one of claims 1 to 7, comprising: The data interface is used to receive FY-4B multi-channel satellite observation data, ground-based weather radar scan data, ground rain gauge observation data, elevation data, and land surface type data; Memory, used to store computer programs; A processor, connected to the memory, is used to execute the computer program to: preprocess FY-4B multi-channel satellite observation data to form a satellite dataset; Data quality control is performed on the radar volume scan data and the combined radar reflectivity is calculated. The data is then spatiotemporally aligned with the satellite dataset to construct the first-stage training samples and train the ER-UNet model. The ER-UNet model outputs satellite-inverted radar composite reflectivity and trains a random forest regression model by combining elevation data, land surface type data, and rainfall observed by rain gauges. In the rainfall inversion stage, gridded ground rainfall data is output, wherein ground-based weather radar observation data corresponding to the time to be inverted is not input when outputting the gridded ground rainfall data.
9. The regional rainfall inversion system based on ER-UNet and random forest model according to claim 8, characterized in that: The satellite dataset constructed by the processor includes brightness temperature features of FY-4B channels 10, 13 and 14, brightness temperature difference features between channels 13 and 14, brightness temperature difference features between channels 13 and 10, local spatial gradient features of channel 13, and deep convection index (DCI).
10. The regional rainfall inversion system based on ER-UNet and random forest model according to claim 9, characterized in that: When training the ER-UNet model, the processor introduces L2 regularization and uses Dropout regularization, and employs a composite loss function consisting of a weighted average of mean squared error loss, multi-scale structural similarity loss, and gradient loss.