Rainfall field construction method based on point-line-surface multi-source feature fusion

Abstract

The invention discloses a rainfall field construction method based on point-line-plane multi-source feature fusion, and the method comprises the following steps: (1) obtaining the rainfall data of each rainfall station in a target drainage basin, the microwave link data of a microwave base station, the satellite rainfall data and the radar rainfall data, and carrying out the preprocessing of the microwave data; (2) constructing a dry-wet period discrimination model based on machine learning; (3) calculating microwave signal basic attenuation correction by taking a dry-wet period judgment result as an auxiliary element; (4) carrying out inversion calculation on the rainfall process after the dry and wet periods are distinguished by utilizing a rainfall attenuation inversion model; and (5) using a satellite rainfall field as an initial background field, fusing microwave link data inversion data and rainfall station data based on a multi-grid variation analysis method, and constructing a satellite-microwave-rainfall station three-source fusion rainfall field. Through the multi-grid variational analysis method, the advantages of the microwave rainfall measurement technology can be effectively utilized, accurate reconstruction of rainfall field space distribution is achieved, and the defect that an existing satellite rainfall monitoring product is poor in precision is effectively overcome.

Description

Technical Field

[0001] The invention relates to a rainfall field construction method by fusing multi-source features of points, lines and surfaces, and belongs to the field of ground meteorological monitoring. Background Art

[0002] Accurate rainfall observation data and its spatial distribution information are of great practical significance for water resource management, reservoir operation, agricultural production, and scientific disaster prevention and mitigation. Rainfall observation is the primary prerequisite for obtaining precipitation data, and includes methods such as rain gauge observations, weather radar monitoring, satellite remote sensing, and microwave link inversion. Rain gauges are an early form of rainfall observation equipment and are considered the most direct and accurate way to obtain near-surface precipitation information. They observe rainfall accumulation or intensity at a point scale, but their shortcomings include the long distances between adjacent rain gauges and low correlation between stations. They can only represent local precipitation conditions and cannot capture rainfall centers or spatial variability. Weather radar monitors rainfall based on the reflection of radio electromagnetic wave signals. By determining the position, trajectory, speed, and dynamic trend of radar echoes, it can infer the rainfall area, rainfall intensity, start and end times, and the direction of movement of the rain area. Precipitation radars can provide continuous quantitative precipitation measurements over large areas, providing precipitation data with high temporal and spatial resolution. However, radars are susceptible to interference from non-meteorological factors such as topography and buildings, and the spatial disparity between near-surface precipitation and high-altitude radar echoes is significant, resulting in reduced rainfall estimation accuracy. Compared to weather radars, remote sensing satellites can obtain precipitation information over larger areas. Essentially, they identify rainfall events by acquiring cloud information from the atmosphere around the clock and over a large area. They process cloud images and identify rainfall rates, making them an indirect means of observing rainfall. However, the accuracy of satellite-derived precipitation is limited by factors such as altitude, inversion algorithms, and cloud information misinterpretation, making them incapable of fully reflecting actual rainfall conditions on the ground. Wireless microwave links are a new type of environmental perception sensor widely used for rainfall estimation. The basic principle is that when microwave signals pass through a rainy area, raindrops cause loss and attenuation of the microwave signal. These links can be used as "line rain gauges" to extract comprehensive rainfall information along the path. By leveraging the unique relationship between rain-induced attenuation and rainfall rate, rainfall intensity along the path can be inferred. Microwave communication base stations are densely distributed, and rainfall detection can be carried out using existing base stations without the need for supervision and additional costs. The inverted rainfall data can more accurately express the actual rainfall conditions under complex underlying surface conditions, and the measurement results can maintain good consistency with the rain gauge.

[0003] Existing precipitation monitoring methods and forecasting systems still have significant room for improvement in terms of precision and accuracy. Single-type data cannot meet the requirements for refined rainfall data and are often unsuitable as input variables for hydrological models, numerical simulations, and flood forecasting. Many studies interpolate data from discrete ground-based meteorological stations to construct two-dimensional rainfall spatial patterns. However, due to their low density and uneven distribution, these data lack spatial representativeness and fail to capture the actual surface rainfall. A few studies have employed precipitation data fusion algorithms to improve the quality of rainfall field reconstruction, but these fusion efforts primarily focus on optimizing rain gauges, radar, and satellite data. The use of wireless microwave data is rare, as the advantages of microwave links, such as their widespread distribution and high quality, are limited in their application to precipitation data fusion. Summary of the Invention

[0004] Purpose of the Invention: To overcome the shortcomings of existing technologies, this invention provides a method for constructing a rainfall field by fusing multi-source features, including points, lines, and surfaces. This method uses satellites as the background field, superimposed with microwave-inverted rainfall data and rain gauge data for three-source fusion, and validated using radar data. This method improves the precipitation fusion algorithm, aiming to integrate different resources, further exploring the potential for refining near-surface rainfall fields, and addressing the accuracy deficiencies of existing single-satellite rainfall products.

[0005] Technical solution: To solve the above technical problems, the present invention provides a method for constructing a rainfall field by fusing multi-source features of points, lines and surfaces, which is characterized by comprising the following steps:

[0006] Step S1: Obtain rainfall data from each rain gauge in the target basin, microwave link data from microwave base stations, satellite rainfall data, and radar rainfall data, and pre-process the microwave link data;

[0007] Step S2: constructing a dry and wet period discrimination model based on machine learning;

[0008] Step S3: using the dry and wet period discrimination results as auxiliary factors, calculating the basic attenuation correction of the microwave signal;

[0009] Step S4: using the rain attenuation inversion model to perform inversion calculation on the rainfall process after distinguishing between dry and wet periods;

[0010] Step S5: Using the satellite rainfall field as the initial background field, the microwave link data inversion data and the rain gauge data are fused based on the multigrid variational analysis method to construct a satellite-microwave-rain gauge three-source fusion rainfall field, and the radar rainfall field is used for verification.

[0011] Preferably, in step S1, the required microwave data types include microwave link signal attenuation data, location information, polarization mode, and signal frequency data, and the satellite rainfall data refers to IMERG satellite precipitation product data.

[0012] Preferably, in the step S1, the microwave link data preprocessing includes:

[0013] Step S1.1: Statistically analyze the attenuation processes of all microwave links, and use the consistency relationship of signal attenuation processes between nearby microwave links to identify the abnormal behaviors of microwave signals caused by burst elements, and identify and eliminate abnormal links;

[0014] Step S1.2: Use the Z-score normalization method to detect outliers in the microwave attenuation data, and its calculation formula is:

[0015]

[0016] In the formula, Z is the observed value, μ is the average value of the sample data set, and σ is the standard deviation of the sample data set;

[0017] Step S1.3: Perform a correlation analysis on the signal attenuation data with a resolution of 10 minutes and the rainfall intensity data of the rain gauge closest to the straight-line distance, and remove all links with a correlation coefficient less than 0.6.

[0018] Preferably, in the step S2, the dry-wet period discrimination model uses a random forest model equipped with the CART decision tree algorithm, and uses the characteristics of the microwave attenuation data itself as the basis for dry-wet classification. The CART decision tree algorithm uses the Gini coefficient instead of the information gain. The definition formula of the Gini coefficient is as follows:

[0019]

[0020] In the formula, Gini(A) is the Gini coefficient, p(x i ) represents the probability. CART divides the node into left and right child nodes by selecting the feature with the largest decrease in the Gini coefficient before and after splitting until the tree construction is completed. To accurately evaluate the comprehensive effect of the dry-wet period classification model, a confusion matrix containing four situations is defined: precipitation actually occurs, and the prediction result is also classified as precipitation, marked as Hits (H); precipitation actually occurs, but the prediction result is classified as non-precipitation, marked as Misses (M); precipitation does not actually occur, and the prediction result is misclassified as precipitation, marked as False alarms (F); precipitation does not actually occur, and the prediction result is correctly classified as non-precipitation, marked as Correct negatives (C). Calculate the hit rate, false alarm rate, miss rate, and accuracy rate according to the confusion matrix as the evaluation indicators of the dry-wet period discrimination model.

[0021] Preferably, in the step S3, in the process of calculating the basic attenuation correction of the microwave signal, using the dry-wet period discrimination result as an auxiliary factor and considering the influence of wet antenna attenuation, the total transmission attenuation of the microwave is generalized as:

[0022] AT = A r + A B + n

[0023] In the formula, A T is the total microwave transmission attenuation on the path, A r is the attenuation caused by precipitation, that is, rain-induced attenuation, A B is the signal attenuation caused by other factors except rainfall, called basic attenuation, and n is the attenuation caused by the wetting of the antenna. In this paper, the wet antenna effect of a single precipitation event is set to a fixed value. After subtracting the basic attenuation and wet antenna attenuation from the total microwave attenuation time series, the remaining attenuation is the rain-induced attenuation process:

[0024]

[0025] In the formula, t j represents the time point j, that is, the attenuation value A is calculated at a specific moment r (t j ).

[0026] Preferably, in the step S4, the rainfall process inversion calculation includes:

[0027] Step S4.1: Calculate the rainfall intensity on the corresponding microwave link by using the ITU-R rain attenuation inversion model proposed by the International Telecommunication Union. The calculation formula is as follows:

[0028]

[0029] Among them, R is the rainfall intensity, with the unit of mm / h; A is the path integral attenuation of the link, with the unit of dB; L is the link length, with the unit of km, and a and b are power-law coefficients related to the microwave polarization mode and frequency. For microwave links with different frequencies and polarization modes, the calculation method is as follows:

[0030]

[0031] Step S4.2: Calculate the correlation coefficient, mean absolute error, root mean square error, and Nash efficiency coefficient respectively to quantitatively evaluate the error between the rainfall intensity obtained from the wireless microwave attenuation data and the rainfall intensity actually measured by the rain gauge station, and characterize the degree of agreement between the two.

[0032] Preferably, the step S5 includes the following steps:

[0033] Step S5.1: Use the Kriging interpolation method to interpolate the rain gauge station data and microwave data to obtain a two-dimensional rainfall field that can be used for fusion. The basic principle of the Kriging interpolation method is as follows:

[0034] Assume the research area D, the variable Z(x) ∈ D within the area, x1, x2,..., xn For n observation points within the area, Z(x1), Z(x2), …, Z(x n ) are the corresponding observed values. Assume that the estimated value of the point to be measured x0 is Z * (x0), which can be estimated using the following formula:

[0035]

[0036] In the formula, λ i is the weight of the i-th sampling point. The weight λ needs to satisfy unbiasedness and the minimum estimation variance, that is:

[0037]

[0038] In the formula, C(x i , x j ) is the covariance function of Z(x i ) and Z(x j ). μ is the Lagrange multiplier. λ i depends on the calculation result of the variogram. The empirical semivariogram γ(h) provides the spatial autocorrelation information of the sampling data set, and its estimated value γ * (h) is calculated by the following formula:

[0039]

[0040] In the formula, h is the lag distance, and N(h) is the number of observation points at a distance of h;

[0041] Step S5.2: Respectively use the IMERG satellite precipitation products with 1h resolution and 1d resolution as the initial background fields, and fuse the microwave link data inversion data and the rain gauge data based on the multi-grid variational analysis method to construct a satellite-microwave-rain gauge three-source fusion rainfall field. The calculation steps of the multi-grid variational analysis method are divided into two stages: coarse grid solution and fine grid solution. After obtaining an approximate solution on the coarse grid scale, it is then transferred to a finer grid through an iterative method;

[0042] The objective functional of the multi-grid variational analysis method is as follows:

[0043]

[0044] Among them, m = 1, 2, …, M is the number of grid layers, X (m) is the analysis increment of the m-th grid, o (m) is the covariance matrix of the observation error, H (m) is the interpolation operator from the grid to the observation space, Y (m) is the newly added information amount, and the calculation formula is as follows:

[0045]

[0046] Among them, Y obs is the observation field vector, and X b is the original background field. By superimposing the background field and the increment field, the final analysis field is obtained as follows:

[0047]

[0048] Step S5.3: Convert the radar echo intensity data into actual rainfall intensity according to the Z-R relationship. The calculation formula is as follows:

[0049] dBZ = 10 × log 10 (Z)

[0050] Z = αR k

[0051] In the formula, Z is the radar reflectivity factor, with the unit of mm 6 / m 3 ;

[0052] Step S5.4: Construct radar rainfall distribution maps with 1h and 1d resolutions, and use Python language programming to extract and convert longitude and latitude information to obtain a radar background field that can be used for verification, and verify the three-source fusion result.

[0053] Beneficial effects: Compared with the prior art, the method for constructing a rainfall field by fusing multi-source features of points, lines and surfaces of the present invention has the following beneficial effects:

[0054] 1. By fusing satellite rainfall fields, microwave inversion interpolation rainfall fields and rain gauge interpolation rainfall fields, the present invention realizes the complementary advantages of three-source data by using the multi-grid variational analysis method, greatly improves the spatial resolution and accuracy of rainfall field construction, and makes up for the shortcoming of limited accuracy of existing satellite products.

[0055] 2. The present invention inverses the rainfall intensity process with 1h and 30min resolutions, verifies the advantages of microwave link rainfall monitoring, can realize rainfall field construction at different time scales, and meets various application requirements from short-term near-term rainfall forecasting to historical analysis.

[0056] 3. By combining microwave link signal attenuation data, satellite rainfall data and traditional rain gauge rainfall data, the potential of various data resources is maximally exploited, and the limitations of a single data source are overcome. Brief description of the drawings

[0057] Figure 1 is the technical roadmap of the present invention;

[0058] Figure 2 is the schematic diagram of the multi-grid variational analysis method;

[0059] Figure 3 Scatter plot comparison of rainfall amounts from the IMERG satellite and the combined rainfall from three sources at 13:00 on July 25th with a 1-hour resolution for the target river basin. Detailed implementation manner

[0060] The present invention will be further described in conjunction with the accompanying drawings.

[0061] The present invention provides a method for constructing a three-source fusion rainfall field using a multi-grid variational analysis method, and the steps are as follows:

[0062] 1. Obtain the signal attenuation data, position information, polarization mode, and signal frequency data of each microwave link through microwave measurement stations within a certain river basin of the Three Gorges. Obtain the actual rainfall and meteorological radar data for the corresponding time period through rain gauges within the river basin, and obtain satellite rainfall data within the river basin through IMERG satellite precipitation products;

[0063] 2. Preprocess the obtained rain gauges, microwave links, meteorological radar, and satellite data. The main work includes the identification and elimination of abnormal links, the processing of unreasonable attenuation signals, and the optimization of microwave link length and frequency. The specific steps are as follows:

[0064] (1) Conduct a statistical analysis of the attenuation processes of all links. Utilize the consistency relationship between the signal attenuation processes of nearby microwave links to identify and eliminate abnormal behaviors of microwave signals caused by sudden elements.

[0065] (2) Use the Z-score normalization method to detect outliers in the microwave attenuation data. The calculation formula is:

[0066]

[0067] In the formula, Z is the observed value, μ is the average value of the sample data set, and σ is the standard deviation of the sample data set.

[0068] (3) Conduct a correlation analysis between the signal attenuation data with a 10-minute resolution and the rainfall intensity data of the rain gauge closest to the straight-line distance, and remove all links with a correlation coefficient less than 0.6.

[0069] 3. Construct two dry-wet period discrimination models using a random forest method based on microwave statistical characteristics. The specific steps are as follows:

[0070] (1)Construct a random forest dry and wet period discrimination model based on microwave statistical features. It is an algorithm constructed by the bootstrap aggregation integration method with decision trees as the base classifiers, which combines the advantages of decision trees and Bagging ensemble learning. A decision tree is a hierarchical tree-like decision structure. Its principle is to make judgments based on the conditions of decision points. After meeting the conditions, it enters the next decision point and continues to search until the leaf node of the decision tree ends. There are three steps in the learning process of a decision tree, namely feature selection, decision tree generation, and decision tree pruning. In this paper, the CART algorithm is selected for decision-making. The CART decision tree algorithm uses the Gini coefficient instead of the information gain. The definition formula of the Gini coefficient is as follows:

[0071]

[0072] In the formula, Gini(A) is the Gini coefficient, and p(x i ) represents the probability. CART divides the node into two child nodes, the left and the right, by selecting the feature with the largest decrease in the Gini coefficient before and after splitting until the tree construction is completed.

[0073] The specific steps of the random forest algorithm are as follows: Assume that the number of samples in the initial training set is N, and the number of feature attributes is M. Use the Bootstrap sampling method to extract n samples from the initial training set to form a training subset; randomly select m features from the M feature attributes as alternative features, and select the optimal attribute for splitting at each node of the decision tree according to the corresponding rules until all the training samples at this node belong to the same class, and no pruning is performed during the process of complete splitting; repeat the above steps k times to construct k decision trees and generate a random forest; use the random forest for decision-making, summarize the classification results of each decision tree, and according to the majority voting principle, the class with the most votes is the final classification result. Majority voting principle:

[0074] H(x) = {Major}{h i (x)|i = 1, 2, …, m}

[0075] In the formula, H(x) is the random forest model, and h i (x) is a single decision tree.

[0076] (2) Define a confusion matrix containing four cases: actual precipitation occurs, and the prediction result is also classified as precipitation, marked as Hits (H); actual precipitation occurs, but the prediction result is classified as non-precipitation, marked as Misses (M); actual precipitation does not occur, but the prediction result is incorrectly classified as precipitation, marked as False alarms (F); actual precipitation does not occur, but the prediction result is correctly classified as non-precipitation, marked as Correct negatives (C). The hit rate, false alarm rate, missed alarm rate, and accuracy rate are calculated as evaluation indicators to measure the microwave link's ability to detect rainfall events. The optimal model at 1-h resolution is selected and used in subsequent inversion applications.

[0077] Table 1 Confusion matrix of dry and wet season classification

[0078]

[0079] The probability of detection (POD) represents the percentage of actual precipitation periods that the microwave link correctly inverts. This value reflects the microwave link's ability to detect precipitation. The larger the value, the stronger the microwave's ability to detect precipitation. The formula is as follows:

[0080]

[0081] The false alarm ratio (FAR) is the percentage of false detections among all precipitation periods detected by microwaves. This metric reflects the degree of false alarms of precipitation events by the microwave link. The smaller the value, the more accurate the microwave's precipitation detection. The calculation formula is as follows:

[0082]

[0083] The Missing Alarm Rate (MAR) refers to the ratio of the actual rainfall time to the total rainfall time that the microwave link missed reporting. The smaller the value, the fewer rainfall moments the microwave missed reporting. The calculation formula is:

[0084]

[0085] Accuracy Rate (ACC) reflects the ability of the classification model to correctly judge the overall sample. The higher the accuracy, the better the effect of the dry and wet classification model. The calculation formula is:

[0086]

[0087] In the above formula, H is the total time when precipitation is observed at both the rain gauge station and the microwave link, M is the total time when precipitation is observed at the rain gauge station but not recognized as precipitation by the microwave link, F is the total time when no precipitation is observed at the rain gauge but the microwave link recognizes it as precipitation, and C is the duration when no precipitation is observed at both the rain gauge station and the microwave link.

[0088] 4. Using the dry-wet period discrimination result as an auxiliary factor, all the signal attenuation amounts during the non-rainy period are classified as the basic attenuation value to calculate the basic attenuation correction of the microwave signal. This process takes into account the influence of wet antenna attenuation. The total transmission attenuation of the microwave can be generalized as:

[0089] A T =A r +A B +m

[0090] In the formula, A T is the total microwave transmission attenuation on the path, A r is the attenuation caused by precipitation, that is, rain attenuation, A B is the signal attenuation caused by other factors except rainfall, called basic attenuation, and m is the attenuation amount caused by antenna wetting. In this paper, the wet antenna effect of a precipitation event is set as a fixed value. After subtracting the basic attenuation and wet antenna attenuation from the total microwave attenuation time series, the remaining attenuation amount is the rain attenuation process.

[0091]

[0092] In the formula, t j represents the time point j, that is, calculating the attenuation value A r (t j ) at a specific moment.

[0093] 5. Using the ITU-R rain attenuation inversion model to perform inversion calculations on the rainfall process after distinguishing the dry-wet period, including the following steps:

[0094] (1) Adopting the ITU-R rain attenuation inversion model proposed by the International Telecommunication Union, the calculation formula is as follows:

[0095]

[0096] Among them, R is the rainfall intensity, with the unit of mm / h; A is the path integral attenuation of the link, with the unit of dB; L is the link length, with the unit of km. a and b are power-law coefficients related to the microwave polarization mode and frequency. For microwave links with different frequencies and polarization modes, their calculation methods are as follows:

[0097]

[0098] Based on the relationship between rainfall intensity and microwave rain attenuation, the International Telecommunication Union gives references for the values of a and b in horizontal and vertical polarization modes, as shown in Table 2:

[0099] Table 2 ITU's Suggestions on the Values of a and b in Horizontal and Vertical Polarization Modes

[0100]

[0101] (2) Calculate the rainfall intensity on the corresponding microwave link according to the ITU-R rain attenuation inversion model. At the same time, for the same rainfall event, various rainfall inversion indicators within the time period are analyzed. The five selected indicators are as follows:

[0102] Correlation Coefficient (CC):

[0103]

[0104] Mean Absolute Error (MAE):

[0105]

[0106] Root Mean Square Error (RMSE):

[0107]

[0108] Relative Error (RE):

[0109]

[0110] Nash-Sutcliffe Efficiency (NSE):

[0111]

[0112] In the above formulas, p i represents the observed rainfall intensity value measured at the rain gauge station, is the average value of the measured rainfall intensity, r i is the rainfall intensity inversion value obtained using the rain attenuation model, is the average value of the inverted rainfall intensity.

[0113] 6. Respectively use the satellite rainfall fields with 1h resolution and 1d resolution as the initial background fields, fuse the microwave link data inversion data and the rain gauge station data based on the multi-grid variational analysis method, construct a satellite-microwave-rain gauge station three-source fusion rainfall field, and use the radar rainfall field for verification, including the following steps:

[0114] (1) Use the Kriging interpolation method to interpolate the rain gauge data and microwave data to obtain a two-dimensional rainfall field that can be used for fusion. The basic principle of the Kriging interpolation method is as follows:

[0115] Assume a study area D, where the variable Z(x) ∈ D, and x1, x2, …, x n are n observation points in the area, and Z(x1), Z(x2), …, Z(x n ) are the corresponding observed values. Assume that the estimated value of the point to be measured x0 is Z * (x0), which can be estimated by the following formula:

[0116]

[0117] where λ i is the weight of the i-th sampling point. The weight λ needs to satisfy unbiasedness and minimum estimation variance, that is:

[0118]

[0119] where C(x i , x j ) is the covariance function of Z(x i ) and Z(x j ), μ is the Lagrange multiplier, and λ i depends on the calculation result of the variogram. The empirical semivariogram γ(h) provides the spatial autocorrelation information of the sampling data set, and the estimated value γ * (h) is calculated by the following formula:

[0120]

[0121] where h is the lag distance and N(h) is the number of observation points at a distance of h.

[0122] (2) Use the IMERG satellite precipitation products with 1h resolution and 1d resolution as the initial background fields respectively. Based on the multi-grid variational analysis method, fuse the microwave link data inversion data and the rain gauge data to construct a satellite-microwave-rain gauge three-source fusion rainfall field. The calculation steps of the multi-grid variational analysis method are divided into two stages: coarse grid solution and fine grid solution. After obtaining an approximate solution on the coarse grid scale, it is then transferred to a finer grid through an iterative method.

[0123] The objective functional of the multi-grid variational analysis method is as follows:

[0124]

[0125] where m = 1, 2, …, M is the number of grid layers, and X (m) is the analysis increment of the m-th grid, and O(m) is the covariance matrix of the observation error, and H (m) is the interpolation operator from the grid to the observation space, and Y (m) is the newly added information. The calculation formula is as follows:

[0126]

[0127] where Y obs is the observation field vector, and X b is the original background field. Adding the background field and the increment field gives the final analysis field as:

[0128]

[0129] (3) Convert the radar echo intensity data to the actual rainfall intensity according to the Z-R relationship, and its calculation formula is as follows:

[0130] dBZ = 10 × log 10 (Z)

[0131] Z = αR k

[0132] where Z is the radar reflectivity factor, with the unit of mm 6 / m 3 .

[0133] (4) Construct radar rainfall distribution maps with 1h and 1d resolutions and use Python language programming to implement the extraction and conversion of longitude and latitude information to obtain a radar background field that can be used for verification, and verify the three-source fusion result. Figure 3 is the comparison of rainfall scatter plots of the 1h-resolution IMERG satellite and the three-source rainfall fusion at 13:00 on July 25th for the target basin. The results show that the three-source precipitation fusion result using the multi-grid variational analysis method has been significantly improved compared to the IMERG satellite precipitation product.

[0134] In summary, the method for constructing a three-source fusion rainfall field using the multi-grid variational analysis method invented in this paper has greatly improved the spatial resolution and accuracy of rainfall field construction, and made up for the shortcoming of the limited accuracy of existing satellite products.

[0135] The above is only the preferred embodiment of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. A method for constructing a rainfall field by fusing multi-source features of points, lines and planes, characterized in that, It includes the following steps: Step S1: Obtain the rainfall data of each rain gauge in the target basin, the microwave link data of microwave base stations, and the satellite rainfall data, and preprocess the microwave link data; Step S2: Construct a dry-wet period discrimination model based on machine learning; Step S3: Calculate the basic attenuation correction of microwave signals with the discrimination result of the dry-wet period as an auxiliary factor; Step S4: Perform inversion calculation on the rainfall process after distinguishing the dry-wet period using the rain attenuation inversion model; Step S5: Use the satellite rainfall field as the initial background field, and fuse the inversion data of microwave link data and the data of rain gauges based on the multi-grid variational analysis method to construct a satellite-microwave-rain gauge three-source fusion rainfall field.

2. The method for constructing a rainfall field by fusing multi-source features of points, lines and surfaces according to claim 1 is characterized in that: In the above step S1, the types of microwave data required include the signal attenuation data, location information, polarization mode, and signal frequency data of each microwave link. The satellite rainfall data refers to the IMERG satellite precipitation product data.

3. The method for constructing a rainfall field by fusing multi-source features of points, lines and surfaces according to claim 2, characterized in that, In the above step S1, the preprocessing of microwave link data includes: Step S1.1: Conduct statistical analysis on the attenuation processes of all microwave links, use the consistency relationship of signal attenuation processes between nearby microwave links to identify the abnormal behavior of microwave signals caused by sudden elements, and identify and eliminate abnormal links; Step S1.2: Adopt the Z-score normalization method to detect outliers in microwave attenuation data, and its calculation formula is: In the formula, Z is the observed value, μ is the average value of the sample data set, and σ is the standard deviation of the sample data set; Step S1.3: Conduct a correlation analysis on the signal attenuation data with a resolution of 10 minutes and the rainfall intensity data of the rain gauge closest to the straight-line distance, and remove all links with a correlation coefficient less than 0.

6.

4. A method for constructing a rainfall field by fusing multi-source features of points, lines, and planes according to claim 3, characterized in that In the above step S2, the dry-wet period discrimination model uses a random forest model equipped with the CART decision tree algorithm, and uses the characteristics of the microwave attenuation data itself as the basis for dry-wet classification. The CART decision tree algorithm uses the Gini coefficient instead of the information gain. The definition formula of the Gini coefficient is as follows: where Gini(A) is the Gini coefficient, and p(x i ) represents probability. CART divides the node into left and right child nodes by selecting the feature with the largest decrease in the Gini coefficient before and after splitting until the tree construction is completed. To accurately evaluate the comprehensive effect of the wet and dry period classification model, a confusion matrix containing four cases is defined: precipitation actually occurs and the prediction result is also classified as precipitation, marked as Hits; precipitation actually occurs, but the prediction result is classified as non-precipitation, marked as Misses; When precipitation actually does not occur and the prediction result is misclassified as precipitation, it is marked as False alarms; When there is actually no precipitation and the prediction result is correctly classified as non-precipitation, it is marked as Correct negatives. Calculate the hit rate, false alarm rate, miss rate, and accuracy rate according to the confusion matrix as the evaluation indicators of the dry-wet period discrimination model.

5. A method for constructing a rainfall field by fusing multi-source features of points, lines and surfaces according to claim 4, characterized in that In the above step S3, in the process of calculating the basic attenuation correction of microwave signals, with the discrimination result of the dry-wet period as an auxiliary factor, the influence of wet antenna attenuation is considered. The total transmission attenuation of microwave is generalized as: A T = A r + A B + m Where A T is the total microwave transmission attenuation on the path, A r is the attenuation caused by precipitation, i.e., rain-induced attenuation, A B is the signal attenuation caused by other factors except rainfall, called basic attenuation, m is the attenuation caused by antenna wetting. In this paper, the wet antenna effect of a single precipitation event is set to a fixed value. After subtracting the basic attenuation and wet antenna attenuation from the total microwave attenuation time series, the remaining attenuation is the rain-induced attenuation process: where t j represents the time point j, i.e., the decay value A is calculated at the specific moment j r (t j ).

6. The method for constructing a rainfall field by fusing multi-source features of points, lines and surfaces according to claim 5, characterized in that, In the above step S4, the inversion calculation of the rainfall process includes: Step S4.1: Calculate the rainfall intensity on the corresponding microwave link using the ITU-R rain attenuation inversion model proposed by the International Telecommunication Union. The calculation formula is as follows: Among them, R is the rainfall intensity, with the unit of mm / h; A is the path integral attenuation of the link, with the unit of dB; L is the link length, with the unit of km, and a and b are power-law coefficients related to the microwave polarization mode and frequency. For microwave links with different frequencies and polarization modes, their calculation methods are as follows: Step S4.2: Calculate the correlation coefficient, mean absolute error, root mean square error, and Nash efficiency coefficient respectively to quantitatively evaluate the error between the rainfall intensity obtained from the wireless microwave attenuation data and the rainfall intensity actually measured by the rain gauge station, and characterize the degree of coincidence between the two.

7. The method for constructing a rainfall field by fusing multi-source features of points, lines and surfaces according to claim 6, characterized in that: The following steps are included in the said Step S5: Step S5.1: Use the Kriging interpolation method to interpolate the rain gauge station data and the microwave data to obtain a two-dimensional rainfall field available for fusion. The basic principle of the Kriging interpolation method is as follows: Assume the research area D, the variable Z(x) ∈ D within the area, x1, x2, …, x n are n observation points within the area, Z(x1), Z(x2), …, Z(x n ) are the corresponding observed values. Assume the estimated value of the point to be measured x0 is Z * (x0), which can be estimated by the following formula: where λ i is the weight of the i-th sampling point, and the weight λ needs to satisfy unbiasedness and minimum estimation variance, that is: where C(x i ,x j ) is the covariance function of Z(x i ) and Z(x j ), μ is the Lagrange multiplier, λ i depends on the calculation result of the variogram. The empirical semivariogram γ(h) provides the spatial autocorrelation information of the sampling data set, and the calculation formula of its estimated value γ * (h) is: In the formula, h is the lag distance, and N(h) is the number of observation points at a distance of h; Step S5.2: Use the IMERG satellite precipitation products with 1h resolution and 1d resolution as the initial background fields respectively, and based on the multi-grid variational analysis method, fuse the microwave link data inversion data and the rain gauge station data to construct a satellite-microwave-rain gauge station three-source fusion rainfall field; The objective functional of the multi-grid variational analysis method is as follows: where \(m = 1, 2, \ldots, M\) is the number of grid layers, \(X\) (m) is the analysis increment of the \(m\)-th level grid, \(O\) (m) is the covariance matrix of the observation error, \(H\) (m) is the interpolation operator from the grid to the observation space, \(Y\) (m) is the newly added information, and the calculation formula is as follows: Among them, Y obs is the observation field vector, X b is the original background field. The background field and the incremental field are superimposed to obtain the final analysis field X a :

Images