Precipitation inversion method based on multi-branch fusion of satellite data and geographic information correction

Abstract

The invention discloses a precipitation inversion method based on multi-branch fusion of satellite data and geographic information correction, comprising the following steps: S1, obtaining observation data and preprocessing to obtain multi-modal precipitation inversion data with unified temporal and spatial resolution, wherein the observation data includes six types of image data, namely infrared, passive microwave, radar reflectivity factor, elevation, longitude and latitude; S2, establishing and training a dual-link precipitation inversion model, wherein the dual-link precipitation inversion model includes a regression link and a classification link, wherein both the regression link and the classification link include a three-branch fusion module and a geographic information correction module; S3, inputting the multi-modal precipitation inversion data into the trained dual-link precipitation inversion model for precipitation inversion, and obtaining regression results and classification results as precipitation inversion results. The multi-branch fusion and geographic information correction solve the problems of insufficient satellite data fusion and interaction in the prior art and the problem of lack of consideration of geographic factors, thereby improving the accuracy of precipitation inversion results.

Description

Technical Field

[0001] The invention belongs to the technical field of meteorological forecasting, and in particular relates to a precipitation inversion method based on multi-branch fusion of satellite data and geographic information correction. Background Art

[0002] Precipitation is a key component of the Earth's water cycle. Accurate precipitation measurement and estimation are of great significance to the fields of meteorology and hydrology, such as weather forecasting, flood and drought warning, and long-term climate change research. Conventional precipitation observations are achieved through the use of rain gauges and ground-based radars. Their deployment density and observation range are limited by poor geographical conditions, while satellite observations have a wide range and are not restricted by geographical conditions. The complementary information of multiple satellite data can observe precipitation more accurately. Therefore, it is very necessary to use multi-satellite data for precipitation inversion.

[0003] Generally speaking, traditional multi-satellite precipitation inversion methods are based on statistical or physical methods, but they require a large number of computational variables and storage requirements, as well as high consumption of computing resources. Therefore, with the continuous development of artificial intelligence technology, many studies have proposed various deep learning methods to invert precipitation. Although most of the current deep learning-based methods have achieved multi-satellite precipitation inversion, these methods only splice multi-source data in the channel dimension and send them into the deep learning network. The data fusion and interaction are insufficient, which will cause feature conflicts caused by modal heterogeneity of multi-satellite data. In addition, existing methods often ignore the impact of geographical factors on precipitation, such as complex terrain (such as mountains or plateaus) that easily affects water vapor transmission, and different climate zones represented by different longitudes and latitudes have different precipitation patterns. Therefore, the present invention proposes a precipitation inversion method based on multi-branch fusion of satellite data and geographic information correction. Summary of the invention

[0004] The purpose of the present invention is to solve the problems of insufficient satellite data fusion and interaction and lack of consideration of geographical factors in the prior art through multi-branch fusion and geographic information correction, thereby improving the accuracy of precipitation inversion results.

[0005] To achieve the above object, the technical solution adopted by the present invention is:

[0006] The precipitation inversion method based on multi-branch fusion of satellite data and geographic information correction proposed in the present invention comprises the following steps:

[0007] S1. Obtain observation data and preprocess them to obtain multi-modal precipitation inversion data with uniform temporal and spatial resolution. The observation data include six types of image data: infrared, passive microwave, radar reflectivity factor, elevation, longitude, and latitude;

[0008] S2. Establish a dual-link precipitation inversion model and perform training. The dual-link precipitation inversion model includes a regression link and a classification link. Both the regression link and the classification link include a three-branch fusion module and a geographic information correction module, wherein:

[0009] The three-branch fusion module includes three parallel encoders, an aggregator and a decoder. The input features of the three encoders of the three-branch fusion module are multi-modal precipitation inversion data corresponding to infrared, passive microwave and radar reflectivity factors. The output features of each encoder are aggregated by the aggregator and then input into the decoder to obtain the joint features of multi-satellite precipitation inversion.

[0010] The geographic information correction module includes a double convolution layer, a channel splicing module, three series-connected double convolution layers and an output layer. The input feature of the first double convolution layer is the multimodal precipitation inversion data corresponding to the elevation, longitude and latitude, and the output feature and the multi-satellite precipitation inversion joint feature of the corresponding link are spliced ​​in the channel dimension by the channel splicing module and then sequentially passed through three series-connected double convolution layers to obtain the optimized joint feature. The optimized joint feature of the regression link passes through the corresponding output layer to obtain the precipitation value as the regression result, and the optimized joint feature of the classification link passes through the corresponding output layer to obtain an array consisting of the probability of rain and the probability of no rain as the classification result;

[0011] S3. Input the multimodal precipitation inversion data into the trained dual-link precipitation inversion model for precipitation inversion, and obtain the regression result and classification result as the precipitation inversion result.

[0012] Preferably, the encoder includes a double convolution layer and three downsampling layers connected in series, the aggregator includes a channel splicing module, a downsampling layer and a transposed convolution connected in series, the decoder includes three upsampling layers and a double convolution layer connected in series, the downsampling layer includes a double convolution layer and a maximum pooling layer connected in series, and the upsampling layer includes a transposed convolution and a double convolution layer connected in series.

[0013] Preferably, the dual convolution layer includes two convolution modules connected in series, and the convolution module includes a 2D convolution, a BN normalization and a ReLU activation function connected in series.

[0014] Preferably, the output layer of the regression link includes a series of 2D convolution and softmax operations, and the output layer of the classification link includes a 2D convolution.

[0015] Preferably, the convolution kernel size, convolution step size and padding width of the 2D convolution in the three-branch fusion module are (3, 1, 1), and the number of output channels of the double convolution layer of the encoder is C, the number of output channels of the three downsampling layers of the encoder are 2C, 4C, and 8C in the order of propagation, the number of output channels of the aggregator is 24C, the number of output channels of the three upsampling layers of the decoder are 24C, 12C, and 6C in the order of propagation, and the number of output channels of the double convolution layer of the decoder is 3C; the number of output channels of the first double convolution layer in the geographic information correction module is C, the number of output channels of the channel splicing module is 4C, and the number of output channels of the last three double convolution layers are 2C, C, and C / 2 in the order of propagation, and C=32;

[0016] The convolution kernel size, convolution step size, and padding width of the 2D convolution of the output layer of the regression link and the classification link are all (1, 1, 0), and the number of output channels of the 2D convolution of the output layer of the regression link is 2, and the number of output channels of the 2D convolution of the output layer of the classification link is 1.

[0017] Preferably, the precipitation inversion method based on multi-branch fusion of satellite data and geographic information correction also includes the following steps:

[0018] S4. Perform comprehensive calculation on the classification results and regression results to obtain the final precipitation inversion results. The comprehensive calculation is as follows:

[0019] S41, converting the classification result into an optimized array containing 1 and 0, that is, if the probability of rain corresponding to the pixel is greater than the probability of no rain, the value of the pixel is set to 1, indicating rain, otherwise it is set to 0, indicating no rain;

[0020] S42, multiplying the precipitation values ​​in the optimization array and the regression result pixel by pixel to obtain the final precipitation inversion result.

[0021] Preferably, the preprocessing is to perform time matching, space matching and normalization on the observed data in sequence.

[0022] Preferably, time matching is to determine the starting time when the radar reflectivity factor passes through the preset area, and to match the five image data of infrared, passive microwave, elevation, longitude and latitude with the time nearest neighbor at this starting time for time alignment;

[0023] Spatial matching is to determine the pixel latitude and longitude grid of the radar reflectivity factor, and the image data that has been time-aligned with the radar reflectivity factor is linearly interpolated to the pixel latitude and longitude grid for spatial alignment;

[0024] Normalized, the formula is as follows:

[0025] ;

[0026] in, Represents the original value of the corresponding pixel, is the normalized value of the corresponding pixel, is the maximum value of the corresponding pixel, is the minimum value of the corresponding pixel.

[0027] Preferably, when training the dual-link precipitation inversion model, CLDAS-V2.0 precipitation product data that are sequentially time-matched and spatially-matched are used as labels, and when the CLDAS-V2.0 precipitation product data are time-matched, the starting time when the radar reflectivity factor passes through the preset area is matched with the CLDAS-V2.0 precipitation product data at the end of the hour for time alignment.

[0028] Preferably, when training the dual-link precipitation inversion model, the loss value of the classification link is calculated using the cross entropy loss function, and the loss value of the regression link is calculated using the RMSE loss function.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] The image data acquired by the method includes three types of satellite data and three types of geographic information. Multimodal precipitation inversion data with unified temporal and spatial resolution are obtained through preprocessing. A dual-link precipitation inversion model is established and the multimodal precipitation inversion data is input into the dual-link precipitation inversion model. The two links use a three-branch fusion module to extract the features of the three satellite data and fuse and interact them to obtain the multi-satellite precipitation inversion joint features of each link. The geographic information correction module is used to correct the multi-satellite precipitation inversion joint features in combination with elevation and longitude and latitude data to obtain the optimized joint features of each link. The optimized joint features of the two links are input into the output layers of their respective links to obtain the rain classification results and precipitation value regression results of the pixels, and the final precipitation inversion results are obtained through comprehensive calculation. The present invention can extract features from multi-modal satellite data for fusion and interaction, and then optimize the features in combination with geographic information to give precipitation inversion results. Compared with the prior art, the present invention can better fuse multi-modal satellite data and correct the influence of geographic information on precipitation inversion, thereby achieving efficient inversion of precipitation, solving the problems of insufficient satellite data fusion and interaction and lack of consideration of geographical factors in the prior art, and improving the accuracy of precipitation inversion results. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a flow chart of the precipitation inversion method based on multi-branch fusion of satellite data and geographic information correction of the present invention;

[0032] Figure 2 A schematic diagram of the process of performing temporal and spatial matching of the observation data, CLDAS-V2.0 precipitation product data and GPM 2B-CMB precipitation product data of the present invention;

[0033] Figure 3 It is a schematic diagram of the structure of the dual-link precipitation inversion model of the present invention;

[0034] Figure 4 It is a schematic diagram of the structure of the regression link or the classification link of the present invention;

[0035] Figure 5 Schematic diagram of the correspondence between the final precipitation inversion results and elevations in different inversion regions of the present invention, wherein (a) is the final precipitation inversion results in different inversion regions, and (b) is the elevations in different inversion regions;

[0036] Figure 6 Schematic diagram for comparing the true value, 2B-CMB, the method of the present invention (without GIC) and the final precipitation inversion result of the method of the present invention, wherein (a) is the true value, (b) is 2B-CMB, (c) is the final precipitation inversion result of the method of the present invention (without GIC), and (d) is the final precipitation inversion result of the method of the present invention. DETAILED DESCRIPTION

[0037] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0038] It should be noted that, unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art in the technical field of this application. The terms used herein in the specification of this application are only for the purpose of describing specific embodiments and are not intended to limit this application.

[0039] like Figure 1-Figure 6 As shown in FIG. 1 , the precipitation inversion method based on multi-branch fusion of satellite data and geographic information correction includes the following steps:

[0040] S1. Obtain observation data and preprocess them to obtain multimodal precipitation inversion data with unified temporal and spatial resolution. The observation data include six types of image data: infrared, passive microwave, radar reflectivity factor, elevation, longitude, and latitude.

[0041] In one embodiment, the preprocessing is to perform time matching, space matching and normalization on the observed data in sequence.

[0042] In one embodiment, time matching is to determine the starting time when the radar reflectivity factor passes through the preset area, and to match the five image data of infrared, passive microwave, elevation, longitude, and latitude with the time nearest neighbor at this starting time for time alignment;

[0043] Spatial matching is to determine the pixel latitude and longitude grid of the radar reflectivity factor, and the image data that has been time-aligned with the radar reflectivity factor is linearly interpolated to the pixel latitude and longitude grid for spatial alignment;

[0044] Normalized, the formula is as follows:

[0045] ;

[0046] in, Represents the original value of the corresponding pixel, is the normalized value of the corresponding pixel, is the maximum value of the corresponding pixel, is the minimum value of the corresponding pixel.

[0047] Specifically, the observation data includes three types of satellite data: infrared, passive microwave, and radar reflectivity factor, as well as three types of geographic information: elevation, longitude, and latitude. Specifically, the infrared data of the FY-4A satellite is obtained, such as the Advanced Geostationary Radiometric Imager (AGRI) carried by the FY-4A, which has a total of 14 channels, and the infrared data of channels 7-14 are obtained; the passive microwave data of all 13 channels of the GPM (Global Precipitation Observation Program) microwave imager is obtained; the radar reflectivity factor data of the GPM dual-frequency rain radar is obtained, including the data of the 176 layers of Ku band and the 176 layers of Ka band, the GPM polar-orbiting satellite carries the GPM microwave imager and the GPM dual-frequency rain radar; the geographic information from the GPM data platform (URL link: https: / / gpm.nasa.gov / data / directory) is obtained, including DEM elevation data (elevation) and latitude and longitude data (longitude, latitude).

[0048] Preprocessing is to perform time matching, space matching and normalization in sequence. For example, time matching first determines the starting time when the GPM dual-frequency rainfall radar passes through the preset area (i.e. Figure 2 The start time of the GPM polar-orbiting satellite passing through the preset area) is used to match all other observation data of the nearest neighbors in time to ensure the time alignment of the data; after time alignment, spatial matching first determines the pixel longitude and latitude grid of the radar reflectivity factor, copies the radar reflectivity factor to the pixel longitude and latitude grid, and then linearly interpolates the remaining data except the radar reflectivity factor to this longitude and latitude grid to ensure the spatial alignment of the data; normalization uses the maximum and minimum values ​​of each data to scale the data after time and space alignment to the range of 0 to 1.

[0049] In a specific embodiment, in terms of time matching, since the passive microwave, radar reflectivity factor, DEM elevation, longitude and latitude are all derived from the GPM data platform and have been aligned in time, it is only necessary to match the infrared data of the FY-4A satellite. First, the starting time of the GPM dual-frequency rain measuring radar passing through the preset area is determined, and then the infrared data of the nearest neighbor FY-4A satellite is matched at this time to achieve time unification; then, in terms of spatial matching, the pixel longitude and latitude grid with a spatial resolution of 5km of the radar reflectivity factor is used as a reference, and other data are linearly interpolated to the pixel longitude and latitude grid to achieve spatial unification; finally, in terms of normalization, the maximum and minimum values ​​of each data are set. , and then normalize the data. The maximum and minimum values ​​of each data are set as follows: the infrared data of the FY-4A satellite is [0,500], the passive microwave data of the GPM (Global Precipitation Observation Program) microwave imager is [50,350], the radar reflectivity factor data of the GPM dual-frequency rain radar is [0,100], the DEM elevation is [0,10000], the longitude is [0,360], and the latitude is [0,180]. The normalization formula is as above. For example, the original value of the corresponding pixel in the infrared and passive microwave data is the brightness temperature value, the radar reflectivity factor data is the power value, the elevation data is the height value, and the longitude and latitude data are the degree values.

[0050] S2. Establish a dual-link precipitation inversion model and perform training. The dual-link precipitation inversion model includes a regression link and a classification link. Both the regression link and the classification link include a three-branch fusion module and a geographic information correction module, wherein:

[0051] The three-branch fusion module includes three parallel encoders, an aggregator and a decoder. The input features of the three encoders of the three-branch fusion module are multi-modal precipitation inversion data corresponding to infrared, passive microwave and radar reflectivity factors. The output features of each encoder are aggregated by the aggregator and then input into the decoder to obtain the joint features of multi-satellite precipitation inversion.

[0052] The geographic information correction module includes a double convolution layer, a channel stitching module, three series-connected double convolution layers and an output layer. The input features of the first double convolution layer are multimodal precipitation inversion data corresponding to elevation, longitude and latitude, and the output features and the multi-satellite precipitation inversion joint features of the corresponding link are spliced ​​in the channel dimension by the channel stitching module (channel stitching) and then pass through three series-connected double convolution layers in sequence to obtain the optimized joint features. The optimized joint features of the regression link pass through the corresponding output layer to obtain the precipitation value as the regression result, and the optimized joint features of the classification link pass through the corresponding output layer to obtain an array consisting of the probability of rain and the probability of no rain as the classification result.

[0053] In one embodiment, the encoder includes a double convolution layer and three downsampling layers connected in series, the aggregator includes a channel splicing module, a downsampling layer and a transposed convolution connected in series, the decoder includes three upsampling layers and a double convolution layer connected in series, the downsampling layer includes a double convolution layer and a maximum pooling layer connected in series, and the upsampling layer includes a transposed convolution and a double convolution layer connected in series.

[0054] In one embodiment, the dual convolution layer includes two convolution modules connected in series, and the convolution module includes a 2D convolution, a BN normalization, and a ReLU activation function connected in series.

[0055] In one embodiment, the output layer of the regression link includes a series of 2D convolution and softmax operations, and the output layer of the classification link includes a 2D convolution.

[0056] In one embodiment, the convolution kernel size, convolution step size and padding width of the 2D convolution in the three-branch fusion module are (3, 1, 1), and the number of output channels of the double convolution layer of the encoder is C, the number of output channels of the three downsampling layers of the encoder are 2C, 4C, and 8C in the order of propagation, the number of output channels of the aggregator is 24C, the number of output channels of the three upsampling layers of the decoder are 24C, 12C, and 6C in the order of propagation, and the number of output channels of the double convolution layer of the decoder is 3C; the number of output channels of the first double convolution layer in the geographic information correction module is C, the number of output channels of the channel splicing module is 4C, and the number of output channels of the last three double convolution layers are 2C, C, and C / 2 in the order of propagation, and C=32;

[0057] The convolution kernel size, convolution step size, and padding width of the 2D convolution of the output layer of the regression link and the classification link are all (1, 1, 0), and the number of output channels of the 2D convolution of the output layer of the regression link is 2, and the number of output channels of the 2D convolution of the output layer of the classification link is 1.

[0058] Among them, the regression link and the classification link are responsible for the classification of whether the data pixel has precipitation and the regression of precipitation value respectively. Both links use the three-branch fusion module to extract the features of the three satellite data (independent precipitation features) and fuse and interact them to obtain the joint features of multi-satellite precipitation inversion of each link. In the corresponding link, the multi-modal precipitation inversion data corresponding to the infrared, passive microwave, and radar reflectivity factors are input into the encoder of the three-branch fusion module one by one to obtain the independent precipitation features corresponding to the three satellite data, and the three independent precipitation features are aggregated by the aggregator and input into the decoder to obtain the joint features of multi-satellite precipitation inversion.

[0059] Specifically, the dual-link precipitation inversion model constructed in this embodiment is as follows: Figure 3 , Figure 4As shown in the figure, the dual-link precipitation inversion model consists of a classification link responsible for determining whether a pixel has precipitation or not and a regression link responsible for the precipitation value of the pixel. The two links have the same structure except for the output layer.

[0060] 1) Multi-satellite data fusion:

[0061] In both links, a three-branch fusion module (MBF) is used for fusion and interaction to obtain the joint features of multi-satellite precipitation inversion in each link. Among them, the three-branch fusion module includes three parallel encoders, an aggregator and a decoder. The encoder includes a double convolution layer and three downsampling layers connected in series. The aggregator includes a channel splicing module, a downsampling layer and a transposed convolution connected in series. The decoder includes three upsampling layers and a double convolution layer connected in series. The double convolution layer is connected in series with two convolution modules (1 2D convolution, 1 BN normalization and 1 ReLU activation function); the downsampling layer is connected in series with 1 double convolution layer and 1 maximum pooling layer, and the upsampling layer is connected in series with 1 transposed convolution and 1 double convolution layer.

[0062] 2D convolution (kernel size, convolution step and padding width) are all set to (3, 1, 1); the input channels of the double convolution layer in the encoder correspond to the number of channels of different types of data, the number of output channels is set to C=32, the number of output channels of the three downsampling layers is set to 2C, 4C, and 8C respectively, the number of output channels of the aggregator is set to 24C, the number of output channels of the three upsampling layers in the encoder is set to 24C, 12C, and 6C respectively, and the number of output channels of the last double convolution layer in the decoder is set to 3C. In the corresponding link, the three types of data, infrared, passive microwave, and radar reflectivity factor, are input into three encoders to obtain independent precipitation features of three satellite data. The number of channels of the three independent precipitation features is 8C. After passing through one aggregator, the features with 24C channels are obtained, and then input into one decoder to obtain the joint features of multi-satellite precipitation inversion with 3C channels.

[0063] 2) Geographic information correction:

[0064] Both links use the geographic information correction module (GIC) to correct the joint features of multi-satellite precipitation inversion in combination with elevation and latitude and longitude data to obtain the optimized joint features of each link. The geographic information correction module connects 1 double convolution layer, 1 channel splicing module, 3 double convolution layers and 1 output layer in series. The output channel of the first double convolution layer is set to C=32. The channel splicing module is used to connect the output features of the first double convolution layer (channel number C) and the joint features of multi-satellite precipitation inversion of the corresponding link (channel number 3C) to form a feature with a channel number of 4C. The output channel numbers of the last three double convolution layers are set to 2C, C, and C / 2 respectively. For example, for a specific embodiment, in the corresponding link, the geographic information composed of elevation, longitude, and latitude is input into the first double convolutional layer to obtain a geographic feature with a channel number of C, the geographic feature with a channel number of C and the multi-satellite precipitation inversion joint feature with a channel number of 3C are connected in the channel dimension to obtain a feature with a channel number of 4C, and the feature with a channel number of 4C is input into the last three double convolutional layers to obtain an optimized joint feature with a channel number of C / 2. Figure 3 GIC 1 is the geographic information correction module of the regression link, and GIC 2 is the geographic information correction module of the classification link.

[0065] The optimized joint features of the two links are input into the output layers of their respective links to obtain the pixel rain classification results and precipitation value regression results as precipitation inversion results. Specifically, in the classification link, the output layer consists of a 2D convolution with the number of output channels set to 2 and a softmax operation, and the (convolution kernel size, convolution step size, and padding width) of the 2D convolution are set to (1, 1, 0); in the regression link, the output layer is a 2D convolution with the number of output channels set to 1, and the (convolution kernel size, convolution step size, and padding width) of the 2D convolution are set to (1, 1, 0).

[0066] In a specific embodiment, for the classification link optimization, after the joint features are input to the output layer, a two-layer two-dimensional array is obtained, and the values ​​of the first layer and the second layer are the probability of rain and the probability of no rain for each pixel, respectively. The sum of the probability of rain and the probability of no rain for any pixel is 1, and this precipitation probability array is used as the classification result; for the regression link optimization, after the joint features are input to the output layer, a one-layer two-dimensional array is obtained, and the values ​​in the array are the hourly precipitation values ​​of each pixel, and this hourly precipitation value array is used as the regression result.

[0067] In one embodiment, the dual-link precipitation inversion model is trained using CLDAS-V2.0 precipitation product data that are sequentially time-matched and spatially-matched as labels, and when the CLDAS-V2.0 precipitation product data is time-matched, the starting time when the radar reflectivity factor passes through a preset area is used to match the CLDAS-V2.0 precipitation product data at the end of the hour for time alignment.

[0068] Among them, the CLDAS-V2.0 precipitation product data is the precipitation product data of the land surface data assimilation system of the preset regional meteorological bureau, and the reference website is: https: / / data.cma.cn / data. When the CLDAS-V2.0 precipitation product data is time-matched, the starting time of the GPM dual-frequency rain radar passing through the preset area is first determined, and the starting time is matched with the CLDAS-V2.0 precipitation product data at the end of this hour and all other observation data of the nearest neighbors in time to ensure time alignment; after time alignment, the spatial matching first determines the pixel longitude and latitude grid of the radar reflectivity factor, copies the radar reflectivity factor to the pixel longitude and latitude grid, and then linearly interpolates the remaining data except the radar reflectivity factor to this longitude and latitude grid to ensure the spatial alignment of the data. It should be noted that the CLDAS-V2.0 precipitation product data is not normalized to ensure that the precipitation value is not distorted.

[0069] In one embodiment, when the dual-link precipitation inversion model is trained, the loss value of the classification link is calculated using the cross entropy loss function, and the loss value of the regression link is calculated using the RMSE loss function.

[0070] S3. Input the multimodal precipitation inversion data into the trained dual-link precipitation inversion model to obtain the regression result and the classification result as the precipitation inversion result.

[0071] In one embodiment, the precipitation inversion method based on multi-branch fusion of satellite data and geographic information correction also includes the following steps:

[0072] S4. Perform comprehensive calculation on the classification results and regression results to obtain the final precipitation inversion results. The comprehensive calculation is as follows:

[0073] S41, converting the classification result into an optimized array containing 1 and 0, that is, if the probability of rain corresponding to the pixel is greater than the probability of no rain, the value of the pixel is set to 1, indicating rain, otherwise it is set to 0, indicating no rain;

[0074] S42, multiplying the precipitation values ​​in the optimization array and the regression result pixel by pixel to obtain the final precipitation inversion result.

[0075] Specifically, the two-layer probability array in the transformation classification result (the first layer is the probability of rain, and the second layer is the probability of no rain) is transformed into a layer of rain or no rain array containing only 1 and 0. If the probability of rain in the corresponding pixel is greater than the probability of no rain, the value of the pixel in the rain or no rain array is set to 1, otherwise it is set to 0; then the rain or no rain array and the hourly precipitation value array of the regression result are multiplied pixel by pixel to obtain the final precipitation inversion result.

[0076] The present invention also trains the dual-link precipitation inversion model. The training of the regression link and the classification link models is the same except that the loss function used is different. The training process is as follows:

[0077] First, the dataset is divided into three parts: training set, validation set and test set. The dataset collects 3886 sets of data from 2021 to 2022. Each set of data consists of infrared data from the FY-4A satellite, passive microwave data from the GPM microwave imager, radar reflectivity factor data from the GPM dual-frequency rain radar, DEM elevation data, longitude and latitude data, and CLDAS-V2.0 precipitation product data (target data: CLDAS-V2.0). The 1935 groups in 2021 are divided into training and validation sets at a ratio of 8:2, and all 1951 groups in 2022 are used as test sets. At the same time, in order to batch training, the training set is cropped with a sliding window (each sample can be cropped multiple times), the sliding window size is 49×256, the step size is 128, and samples without precipitation are removed, resulting in 2281 training samples. It should be noted that the validation set and test set are not cropped. The final training set, validation set, and test set contain 2281, 387, and 1951 data respectively.

[0078] During training: the parameters of each convolutional layer in the dual-link precipitation inversion model are initialized with random weights, the learning rate of the model optimizer RMSProp is set to 0.00001, the momentum is 0.4, and a total of 200 complete trainings are performed on the training set. After each training, the output result of the model is compared with the CLDAS-V2.0 precipitation product as the true value to obtain the loss value. The loss value of the classification link is calculated using the cross entropy loss function, and the loss value of the regression link is calculated using the RMSE loss function. The model optimizer is used to minimize the loss, and finally the trained dual-link precipitation inversion model is obtained. The trained dual-link precipitation inversion model is tested on the test set to obtain the classification and regression results. The classification and regression results can be further used to comprehensively calculate and filter out false positives to obtain the final precipitation inversion results. The final precipitation inversion results are tested against the corresponding CLDAS-V2.0 precipitation product data as the true value mainly through three binary indicators and two non-binary indicators. The binary indicators will compare the two according to the set rainfall intensity threshold, and the non-binary indicators will evaluate the similarity between the two.

[0079] The binarization indicators used are probability of detection (POD), false alarm rate (FAR), and critical success index (CSI). The threshold used is 0.1mm / h. The binarization indicators are mainly calculated by the following four indicators: true positive rate (TP), true negative rate (TN), false positive rate (FP), and false negative rate (FN). The calculation formulas for POD, FAR, and CSI are as follows:

[0080]

[0081]

[0082]

[0083] The non-binarization indicators used are root mean square error (RMSE) and Pearson correlation coefficient (CC), and the calculation formula is as follows:

[0084]

[0085]

[0086] in, is the model inversion value, is the precipitation value of the CLDAS-V2.0 precipitation product data as the true value, is the mean of the model inversion values, is the mean precipitation value of CLDAS-V2.0 precipitation product data, is the variance of the model inversion value, is the variance of precipitation values ​​of CLADS-V2.0 precipitation product data, i=1~n, n is the total number of pixels.

[0087] Comparing the GPM 2B-CMB precipitation product data produced by the method of the present invention with the existing method (URL link: https: / / doi.org / 10.5067 / GPM / DPRGMI / CMB / 2B / 07), it is easy to understand that when the GPM 2B-CMB precipitation product data is time-matched, the starting time when the GPM dual-frequency rainfall radar passes through the preset area is matched with the GPM2B-CMB precipitation product data at the end of the hour to ensure time alignment before spatial matching. In order to ensure that the precipitation value is not distorted, the GPM 2B-CMB precipitation product data is also not normalized.

[0088] Figure 2 and Figure 5 The horizontal coordinate x represents longitude, and the vertical coordinate y represents latitude. The comparison results are shown in Table 1, where ↑ means the larger the better, and ↓ means the smaller the better.

[0089] Table 1 Comparison results of performance indicators

[0090] RMSE(mm / h)↓ CC↑ POD↑ FAR↓ CSI↑ 2B-CMB 0.5563 0.5242 0.4501 0.4010 0.3459 Method of the present invention 0.3367 0.7918 0.5594 0.2400 0.4754

[0091] The comparison results in Table 1 show that the final precipitation inversion results of the method of the present invention are significantly better than the existing methods in any index, which shows that the performance of the method of the present invention is better than the existing technical solutions. Among them, 2B-CMB refers to the data produced by the method disclosed in the document Olson WS, Masunaga H, Gpm CRRAT. GPM combined radar-radiometer precipitation algorithm theoretical basis document (version 7)[J]. NASA: Washington, DC, USA, 2022., that is, the public data is aligned in time and space and the sliding window is clipped to form the GPM 2B-CMB precipitation product data.

[0092] The final precipitation inversion result of the method of the present invention is Figure 5 and Figure 6 In the display, Figure 5 Figure (a) shows the GPM polar satellite orbit and the final precipitation inversion results of the method of the present invention within the orbit, where different rows correspond to different inversion areas. Figure 5 The box in (a) represents Figure 5 Figure (b) in Figure 6 Figure (a) in Figure 6 Figure (b) in Figure 6 Figure (c) in Figure 6 The corresponding area in (d) is Figure 5 Figure (b) shows the elevation of the corresponding area. Figure 6 Figure (a) shows the CLDAS-V2.0 precipitation product data, i.e. the true value. Figure 6 Figure (b) shows the GPM2B-CMB precipitation product data, namely 2B-CMB. Figure 6 Figure (c) shows the final precipitation inversion result of the present invention (without GIC). Figure 6 Figure (d) shows the final precipitation inversion result of the present invention. It can be seen that the model has learned the precipitation area and precipitation value, and the final precipitation inversion result is closer to the true value in the classification of whether there is rain or not and the regression of precipitation value. At the same time, the result without adding geographic information correction module is also compared. The comparison shows that the addition of geographic information correction module can improve the inversion performance of the model for precipitation in complex terrain areas.

[0093] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0094] The above-described embodiments only express the more specific and detailed embodiments described in this application, but they cannot be understood as limiting the scope of the application. It should be pointed out that for ordinary technicians in this field, several modifications and improvements can be made without departing from the concept of this application, which all belong to the protection scope of this application. Therefore, the protection scope of this application shall be based on the attached claims.

Claims

1. A precipitation inversion method based on multi-branch fusion of satellite data and geographic information correction, characterized by: The precipitation inversion method based on multi-branch fusion of satellite data and geographic information correction comprises the following steps: S1. Obtain observation data and perform preprocessing to obtain multi-modal precipitation inversion data with uniform temporal and spatial resolution, wherein the observation data includes six types of image data: infrared, passive microwave, radar reflectivity factor, elevation, longitude, and latitude; S2. Establishing and training a dual-link precipitation inversion model, wherein the dual-link precipitation inversion model includes a regression link and a classification link, wherein both the regression link and the classification link include a three-branch fusion module and a geographic information correction module, wherein: The three-branch fusion module includes three parallel encoders, an aggregator and a decoder. The input features of the three encoders of the three-branch fusion module are multi-modal precipitation inversion data corresponding to infrared, passive microwave and radar reflectivity factors, respectively. The output features of each encoder are aggregated by the aggregator and then input into the decoder to obtain multi-satellite precipitation inversion joint features. The geographic information correction module includes a double convolution layer, a channel splicing module, three series-connected double convolution layers and an output layer, the input feature of the first double convolution layer is the multimodal precipitation inversion data corresponding to the elevation, longitude and latitude, and the output feature and the multi-satellite precipitation inversion joint feature of the corresponding link are spliced ​​in the channel dimension by the channel splicing module and then sequentially passed through three series-connected double convolution layers to obtain the optimized joint feature, then the optimized joint feature of the regression link passes through the corresponding output layer to obtain the precipitation value as the regression result, and the optimized joint feature of the classification link passes through the corresponding output layer to obtain an array composed of the probability of rain and the probability of no rain as the classification result; S3. Input the multimodal precipitation inversion data into the trained dual-link precipitation inversion model for precipitation inversion, and obtain the regression result and classification result as the precipitation inversion result.

2. The precipitation inversion method based on multi-branch fusion of satellite data and geographic information correction according to claim 1, characterized in that: The encoder includes a double convolution layer and three downsampling layers connected in series, the aggregator includes a channel splicing module, a downsampling layer and a transposed convolution connected in series, the decoder includes three upsampling layers and a double convolution layer connected in series, the downsampling layer includes a double convolution layer and a maximum pooling layer connected in series, and the upsampling layer includes a transposed convolution and a double convolution layer connected in series.

3. The precipitation inversion method based on multi-branch fusion of satellite data and geographic information correction as claimed in claim 1 or 2, characterized in that: The dual convolution layer includes two convolution modules connected in series, and the convolution module includes a 2D convolution, a BN normalization and a ReLU activation function connected in series.

4. The precipitation inversion method based on multi-branch fusion of satellite data and geographic information correction as claimed in claim 2, characterized in that: The output layer of the regression link includes a series of 2D convolution and softmax operations, and the output layer of the classification link includes a 2D convolution.

5. The precipitation inversion method based on multi-branch fusion of satellite data and geographic information correction as claimed in claim 4, characterized in that: The convolution kernel size, convolution step size and padding width of the 2D convolution in the three-branch fusion module are (3, 1, 1), and the number of output channels of the dual convolution layer of the encoder is C, the number of output channels of the three downsampling layers of the encoder are 2C, 4C, and 8C in the order of propagation, the number of output channels of the aggregator is 24C, the number of output channels of the three upsampling layers of the decoder are 24C, 12C, and 6C in the order of propagation, and the number of output channels of the dual convolution layer of the decoder is 3C; the number of output channels of the first dual convolution layer in the geographic information correction module is C, the number of output channels of the channel splicing module is 4C, and the number of output channels of the last three dual convolution layers are 2C, C, C / 2 in the order of propagation, C=32; the convolution kernel size, convolution step size and padding width of the 2D convolution of the output layer of the regression link and the classification link are all (1, 1, 0), and the number of output channels of the 2D convolution of the output layer of the regression link is 2, and the number of output channels of the 2D convolution of the output layer of the classification link is 1.

6. The precipitation inversion method based on multi-branch fusion of satellite data and geographic information correction according to claim 1, characterized in that: The precipitation inversion method based on multi-branch fusion of satellite data and geographic information correction also includes the following steps: S4. Perform comprehensive calculation on the classification results and regression results to obtain the final precipitation inversion result. The comprehensive calculation is as follows: S41, converting the classification result into an optimized array containing 1 and 0, that is, if the probability of rain corresponding to the pixel is greater than the probability of no rain, the value of the pixel is set to 1, indicating rain, otherwise it is set to 0, indicating no rain; S42, multiplying the precipitation values ​​in the optimization array and the regression result pixel by pixel to obtain the final precipitation inversion result.

7. The precipitation inversion method based on multi-branch fusion of satellite data and geographic information correction according to claim 1, characterized in that: The preprocessing is to perform time matching, space matching and normalization on the observed data in sequence.

8. The precipitation inversion method based on multi-branch fusion of satellite data and geographic information correction according to claim 7, characterized in that: The time matching is to determine the starting time when the radar reflectivity factor passes through the preset area, and to match the five image data of infrared, passive microwave, elevation, longitude and latitude with the nearest time neighbor at this starting time for time alignment; The spatial matching is to determine the pixel longitude and latitude grid of the radar reflectivity factor, and linearly interpolate the image data after time alignment with the radar reflectivity factor to the pixel longitude and latitude grid for spatial alignment; The normalization formula is as follows: ； in, Represents the original value of the corresponding pixel, is the normalized value of the corresponding pixel, is the maximum value of the corresponding pixel, is the minimum value of the corresponding pixel.

9. The precipitation inversion method based on multi-branch fusion of satellite data and geographic information correction according to claim 1, characterized in that: When training the dual-link precipitation inversion model, CLDAS-V2.0 precipitation product data that are sequentially time-matched and spatially-matched are used as labels, and when the CLDAS-V2.0 precipitation product data are time-matched, the starting time when the radar reflectivity factor passes through the preset area is matched with the CLDAS-V2.0 precipitation product data at the end of the hour for time alignment.

10. The precipitation inversion method based on multi-branch fusion of satellite data and geographic information correction according to claim 1, characterized in that: When the dual-link precipitation inversion model is trained, the loss value of the classification link is calculated using the cross entropy loss function, and the loss value of the regression link is calculated using the RMSE loss function.

Images