Fusion framework of grid telemetry rainfall and point-measured rainfall based on virtual stations
By dividing molecular areas in areas with low rainfall site density and recommending virtual sites, combining actual and virtual sites for spatial interpolation, the problem of accurate estimation of rainfall fields in the existing technology is solved, and high-precision rainfall fields under any density conditions are achieved estimate.
Patent Information
- Application Number
- CN202211004273.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-08-22
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2042-08-22
AI Technical Summary
The prior art is difficult to accurately capture the spatial distribution of precipitation in areas with sparse rainfall sites. The method of integrating lattice-like telemetry and point-like actual rainfall measurement is reduced under high-density conditions, which cannot effectively solve the problem of accurate estimation of rainfall sites.
Using a fusion framework based on virtual sites, by dividing the area to be estimated into sub-regions, using multi-source grid-shaped telemetry rainfall products to recommend virtual sites, combining actual sites and virtual sites for spatial interpolation, establishing a transformation model to estimate rainfall, weakening grid-shaped Uncertainty in telemetry products.
Improve the estimation accuracy of rainfall fields under any rainfall site density conditions, enhance the supplementary spatial distribution information and the fusion efficiency of multi-source products, and provide higher spatial distribution accuracy of precipitation.
Smart Images

Figure CN115421218B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of rainfall estimation in areas with little data, and relates to a fusion framework of grid telemetry rainfall and point measured rainfall based on virtual sites. Background Art
[0002] Precipitation is crucial in hydrology, agriculture, meteorology and other fields, but the strong spatial heterogeneity of rainfall fields makes its accurate estimation a challenge. At present, precipitation monitoring mainly includes two methods: one is the point-based direct measurement method based on ground rain gauges, and the other is the grid-based indirect measurement method based on radar, satellites and numerical meteorological models. The direct measurement method is accurate at the site location, but it is limited by investment funds, terrain and climate, and equipment maintenance. In areas with sparse site spatial coverage, it is difficult for rain gauges to capture all spatial distribution information of precipitation, such as missing rainfall centers. Thanks to the rapid development of remote sensing technology, the spatial continuous estimation of indirect monitoring methods can reflect the spatial distribution of rainfall fields, especially for capturing the spatial maximum rainfall; however, the numerical estimation of rainfall by indirect measurement methods has errors and more uncertainties, especially for near-real-time or real-time precipitation products. Therefore, it is urgent to invent a rainfall estimation method that integrates grid-based remote sensing rainfall and point-based measured rainfall to more accurately estimate the rainfall field.
[0003] According to the characteristics of the two measurement methods, the fusion ideas are mainly divided into two categories. One is to use point-measured rainfall to correct grid telemetered rainfall; the other is to use grid telemetered rainfall to supplement the spatial distribution information of rainfall that is missing from point-measured rainfall. Many studies have made attempts to fuse grid telemetered rainfall and point-measured rainfall. The fusion methods mainly include three categories. The first category uses statistical analysis models to correct grid telemetered rainfall based on point-measured rainfall; the second category uses multi-source rainfall estimates to fuse multi-source rainfall estimates with different weights; the third category uses machine learning methods to fuse multi-source rainfall estimates. The steps of the above fusion methods can be summarized as follows: first, at the grid points with rainfall stations, different transformation models are trained or calibrated using different regression analysis methods based on the measured rainfall of ground stations; second, at the grid points without rainfall stations, rainfall is estimated using the trained transformation model. The above methods use a grid-by-grid fusion method without considering the spatial relationship between grids, and can be regarded as non-spatial fusion methods. Therefore, these three methods only follow the first fusion idea in the fusion process and ignore the second fusion idea.
[0004] Compared with spatial interpolation of rainfall field estimation using only point-like measured rainfall, the above fusion method can improve the rainfall estimation accuracy in areas where rainfall stations are extremely sparsely distributed. However, due to the lack of the second type of fusion idea, as the density of rainfall stations in the target area of the estimated rainfall field increases, this improvement in accuracy becomes weaker and weaker. When the station density exceeds a certain threshold, the performance of fusing the two types of rainfall will be inferior to spatial interpolation using only point-like measured rainfall. In the study of Wu et al., this threshold is about 60 stations / 10 6 km 2 , and the World Meteorological Organization recommends no less than 1000 stations / 10 6 km 2 In comparison, the stations are still extremely sparse, making it difficult to accurately capture the spatial distribution of precipitation. Therefore, it is urgent to invent a method for fusing grid telemetry rainfall and point-based measured rainfall that is applicable to any rainfall station density condition. Summary of the invention
[0005] In view of the problems existing in the prior art, the present invention provides a fusion framework of grid telemetry rainfall and point measured rainfall based on virtual sites.
[0006] The technical solution of the present invention is as follows:
[0007] A fusion framework of grid telemetry rainfall and point measured rainfall based on virtual sites includes the following steps:
[0008] In the first step, the target area for the rainfall field to be estimated is divided into multiple sub-areas according to the rainfall station set G = [G1, G2, ...].
[0009] First, the target area is divided into grids according to the selected resolution; second, the grids controlled by the same group of rainfall stations are divided into a sub-area, where the controllable distance of the rainfall stations is selected in the range of 10 to 50 km. Figure 1 As shown, taking any two rainfall stations G1 and G2 in space as examples, d c is the controllable distance of the rainfall station. Different sub-areas are distinguished by different colors in the figure. X Represents the rain gauges that control the same sub-area, which is a subset of all rain gauges.
[0010] For a sub-area, the estimated rainfall at any grid point x The theoretical value range of is:
[0011]
[0012] in,
[0013]
[0014]
[0015] in, is the actual rainfall at station G i The measured rainfall value, G i is the actual rainfall station within the control distance of grid point x.
[0016] In the second step, for each sub-region, the number and location of the required virtual stations are recommended. The number and location of the virtual stations are recommended after spatial comparison between multi-source grid telemetry rainfall products and point-based measured rainfall.
[0017] Grid points whose actual rainfall exceeds the theoretical value range described in the first step can be selected as virtual sites to be supplemented. The theoretical value range of P is selected, and the two grid points with the maximum and minimum rainfall that exceed the theoretical value range are selected as the virtual sites of the sub-area. It is worth noting that the possible situations in a sub-area include: the actual rainfall of some grid points is less than P min , and at the same time, the actual rainfall at some grid points is greater than P max , then the sub-area needs to add virtual stations at the two grid points where the maximum rainfall and the minimum rainfall are respectively located; there is no grid point where the actual rainfall is less than P min , there are some grid points where the actual rainfall is greater than P max , then the sub-area needs to add a virtual station at the grid point where the maximum rainfall is located; if the actual rainfall at the grid point is less than P min , there is no grid point where the actual rainfall is greater than P max , then the sub-area needs to add a virtual station at the grid point where the minimum rainfall is located; the actual rainfall of all grid points is within [P min ,P max ], then the sub-area does not need to add virtual stations. Therefore, the number of virtual stations that need to be added in each sub-area does not exceed two.
[0018] Among them, because the actual rainfall is unknown, the comparison between the actual rainfall and the boundary of the theoretical value range needs to be based on multi-source grid telemetry rainfall products. The specific steps are as follows:
[0019] The rainfall field estimated by each grid telemetry rainfall product at a single time step is normalized from 0 to 1 to obtain the normalized rainfall field Then at point x:
[0020]
[0021] In the formula, is the rainfall field estimated by product S in a single time step, is the rainfall estimated by product S at grid point x in a single time step.
[0022] For a sub-region, if
[0023]
[0024] So, The location of is the location of the maximum rainfall virtual site recommended by product S. Conversely, product S does not recommend the maximum rainfall virtual site.
[0025] In the formula, is a sub-region location X=[x1,x2,…,x m ] are the grid points contained in the sub-region, is the site location that controls the sub-area a max It is a parameter that enhances the significance of inequality, and the recommended value is between 0.1 and 0.2.
[0026] Similarly, if,
[0027]
[0028] So, The location is the location of the minimum rainfall virtual site recommended by product S. Otherwise, product S does not recommend the minimum rainfall virtual site.
[0029] In the formula, a min It is a parameter that enhances the significance of inequality, and the recommended value is between 0 and 0.1.
[0030] After all precipitation products have completed the steps of recommending virtual sites, if the virtual site locations recommended by more than half of the products are the same or adjacent, then this location is the final virtual site location; otherwise, no additional virtual sites are needed in this sub-area.
[0031] The third step is to estimate the rainfall at the virtual site using the existing non-spatial fusion method. The idea of the non-spatial fusion method can be summarized as follows:
[0032] 2.1) Training the model
[0033] Based on various regression analyses, a transformation model of dependent and independent variables is established, in which the dependent variable is the measured rainfall at the station, and the independent variables are the rainfall estimates of multi-source rainfall products at the actual station location and other related covariates.
[0034]
[0035] Where f(·) is the trained transformation model, P G is the measured rainfall at the actual station G, is the rainfall product S i The estimated rainfall at G, A Gis a set of covariates related to G, such as terrain information, and ε is the model error.
[0036] 2.2) Estimated rainfall
[0037] Estimate rainfall at virtual sites based on the trained transformation model The dependent variable is the rainfall at the virtual site, and the independent variables are the rainfall estimates of the multi-source rainfall products at the virtual site locations and other relevant covariates.
[0038]
[0039] In the formula, Rainfall Product S i The estimated rainfall at the virtual site V, A V is a covariate associated with V.
[0040] The fourth step is to determine whether the distances of all grid points in the target area from the nearest rainfall station are less than the controllable distance of the rainfall station selected in the first step, where the rainfall station is the total set of actual stations and virtual stations G = [G1, G2, ..., V1, V2, ...]. If so, go to the fifth step; if not, take the total set of stations as the updated actual station set and go to the first step, such as Figure 3 shown.
[0041] The fifth step is to perform spatial interpolation based on the existing spatial interpolation method using actual sites and virtual sites to obtain the rainfall field.
[0042] The effects and benefits of the present invention are:
[0043] The present invention uses multi-source grid telemetry rainfall products to infer the number, location and rainfall of virtual stations, and then integrates actual stations and virtual stations for rainfall spatial interpolation, thereby supplementing the rainfall spatial distribution information that is missing from rainfall stations and weakening the interference caused by the uncertainty of grid telemetry products. The present invention combines the advantages of multi-source product fusion and station spatial interpolation, providing a new idea for obtaining higher-precision precipitation spatial distribution. BRIEF DESCRIPTION OF THE DRAWINGS
[0044] Figure 1 This is a sub-area division flow chart involved in the present invention, and the meanings of the letters in the figure are the same as those of the letters involved in the formula.
[0045] Figure 2 This is a process of inferring the location of a virtual site involved in the present invention.
[0046] Figure 3 It is the operating framework of the present invention. DETAILED DESCRIPTION
[0047] Based on the fusion of multi-source precipitation products and site spatial interpolation, the present invention proposes a fusion framework of grid telemetered rainfall and point measured rainfall based on virtual sites.
[0048] The present invention will be further described below by way of examples.
[0049] A region in northern China is selected as the application area, with an area of 31352km 2 , with an altitude of 3-2400 meters. The region has a typical temperate monsoon climate, with dry and cold winters, rainy and hot summers, and short spring and autumn. The annual precipitation is about 1100mm, with large interannual variations and uneven distribution within the year, with precipitation concentrated from May to September. There are dense rainfall stations in the region, with a total of 624 rainfall stations, and there is at least one rainfall station in almost every 0.1° grid. Therefore, it can provide the true value of the daily rainfall field as the basis for accuracy evaluation.
[0050] The specific steps are as follows:
[0051] In the first step, the target area for the estimated rainfall field is divided into multiple sub-areas according to the rainfall stations.
[0052] First, the object area is divided into 0.1° grids. The object area can be divided into 322 grids according to the 0.1° resolution. If multiple rain gauges are located in the same grid, the average of the measured values of multiple rain gauges is taken as the true rainfall value of the grid point. If there is no rain gauge in the grid, the average rainfall of the four nearest grids around the grid is taken as the true rainfall value of the grid point. If there is no rain gauge in the grid and there is no rain gauge in the nearest grid, the rainfall in the grid is recorded as missing. 10 grid points with a missing rate of less than 2% are selected in the object area and assumed to be the actual rain gauges in the area with little data. The values of other grid points are used to evaluate the estimation accuracy of the rainfall field. The set of actual rain gauge stations is recorded as G=[G1,G2,…G 10 ].
[0053] Secondly, the grids controlled by the same group of rainfall stations are divided into a sub-area, where the controllable distance of the rainfall stations is 50 km. Figure 1 As shown, taking any two rainfall stations G1 and G2 in space as examples, d c is the controllable distance of the rainfall station. Different sub-areas are distinguished by different colors in the figure. X Represents the rain gauges that control the same sub-area, which is a subset of all rain gauges.
[0054] For a sub-area, the estimated rainfall at any grid point x The theoretical value range of is:
[0055]
[0056] in,
[0057]
[0058]
[0059] in, is the actual rainfall at station G i The measured rainfall value, G i is the actual rainfall station within the control distance of grid point x.
[0060] In the second step, for each sub-region, the number and location of the required virtual stations are recommended. The number and location of the virtual stations are recommended after spatial comparison between multi-source grid telemetry rainfall products and point-based measured rainfall.
[0061] Five satellite-based precipitation products and three numerical meteorological model-based precipitation products are selected as example-driven models, namely GSMaP-V6-NRT (GSMaP-N), GSMaP-V6-Gauge-NRT (GSMaP-GN), MERG-V6-Early (IMERG-E), IMERG-V6-Late (IMERG-L), PERSIANN-CCS, ECMWF, JMA and NCEP. Table 1 shows the basic information of the selected precipitation products. The precipitation products of the embodiments of the present invention uniformly adopt a spatial resolution of 0.1°. For precipitation products with different initial spatial resolutions, their precision is converted to 0.1° by resampling.
[0062] Table 1 Basic information of selected gridded remote sensing precipitation products
[0063]
[0064]
[0065] The rainfall field estimated by each grid telemetry rainfall product at a single time step is normalized from 0 to 1 to obtain the normalized rainfall field Then at point x:
[0066]
[0067] In the formula, is the rainfall field estimated by product S in a single time step, is the rainfall estimated by product S at grid point x in a single time step.
[0068] For a sub-region, if
[0069]
[0070] So, The location of is the location of the maximum rainfall virtual site recommended by product S. Conversely, product S does not recommend the maximum rainfall virtual site.
[0071] In the formula, is a sub-region location X=[x1,x2,…,x m ] are the grid points contained in the sub-region, is the site location that controls the sub-area a max is a parameter that enhances the significance of the inequality and is set to 0.15 in this example.
[0072] Similarly, if,
[0073]
[0074] So, The location is the location of the minimum rainfall virtual site recommended by product S. Otherwise, product S does not recommend the minimum rainfall virtual site.
[0075] In the formula, a min is a parameter that enhances the significance of the inequality and is taken as 0.05 in this example.
[0076] After all precipitation products have completed the steps of recommending virtual sites, if more than half of the products recommend the same or adjacent virtual site locations, then this location is the final virtual site location. Otherwise, no additional virtual sites are needed in this sub-area. Figure 2 As shown, the rainfall distribution of the three products in the sub-areas shown in the figure shows that the white dots are The rainfall at this location is greater than the rainfall at the black spot where the actual rain gauge station is located, that is, Therefore, the white dots in the figure are recommended by multi-source precipitation products as virtual sites with maximum rainfall.
[0077] In the third step, the rainfall at the virtual station is estimated using random forest (RF) as a non-spatial class fusion method to determine the rainfall at the virtual station.
[0078] First, a random forest model f(·) is trained at the actual site locations.
[0079]
[0080] Where f(·) is the trained random forest model, P G is the measured rainfall at the actual station G, is the rainfall product S i The estimated rainfall at G. There are 8 products in this example. Gis a set of covariates related to G. In this example, the rainfall of the two actual stations closest to station G and the distances from these two actual rain stations to the target grid are selected. ε is the error of the random forest model.
[0081] Secondly, the rainfall at the virtual station V is estimated based on the trained f(·)
[0082]
[0083] In the formula, Rainfall Product S i The estimated rainfall value at V. There are 8 products in this example. V is a set of covariates related to V. In this example, the rainfall of the two actual stations closest to station V and the distances from these two actual rain stations to the target grid are selected.
[0084] The fourth step is to determine whether the distances of all grid points in the target area to the nearest rainfall station are less than the controllable distance of the rainfall station selected in the first step, where the rainfall station is the total set of actual stations and virtual stations G = [G1, G2, ..., G 10 ,v1,V2,…]. If yes, go to step 5; if no, take the total site set as the updated actual site set and go to step 1, e.g. Figure 3 shown.
[0085] In the fifth step, based on the inverse distance weighted (IDW) method, spatial interpolation is performed using actual stations and virtual stations to obtain the rainfall field.
[0086] Finally, the Pearson correlation coefficient was used and RMS error The accuracy of the estimated rainfall field is evaluated as follows:
[0087]
[0088]
[0089] in,
[0090]
[0091]
[0092] In the formula, T is the number of time steps involved in the evaluation, n is the number of grid points involved in the evaluation, Refers to the rainfall at grid point x.
[0093] The above are the implementation steps for the data-poor areas with 10 actual stations. 2, 3, 4, 6, 14, 22, and 30 grid point schemes with missing measurement rates less than 2% are selected in the study area and assumed to be the actual rainfall stations in the data-poor areas, and other grid point values are used to evaluate the estimation accuracy of the rainfall field. The present invention is driven based on the above selected 7 rainfall station schemes.
[0094] When the station density is different, the accuracy of the estimated rainfall field by the inverse distance weighted method (IDW), the random forest method (RF) and the present invention is ( and ) as shown in Table 2. It can be seen that when the stations are extremely sparse, the random forest method (RF) estimates the rainfall field better than the inverse distance weighting method (IDW); when the station density exceeds about 300 stations / 10 6 km 2 When the inverse distance weight method (IDW) estimates the rainfall field better than the random forest method (RF). However, no matter which site density scheme is used, the present invention has the highest accuracy in estimating the rainfall field.
[0095] Table 2 Comprehensive performance of different methods in estimating density rainfall fields at different stations
[0096]
[0097] The above-described embodiments merely express the implementation methods of the present invention, but they cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that for those skilled in the art, several modifications and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention.
Claims
1. A fusion framework of grid telemetry rainfall and point measured rainfall based on virtual sites, characterized in that: The following steps are involved: The first step is to divide the target area of the rainfall field to be estimated into multiple sub-areas according to the rainfall station set G = [G1, G2, ...]; First, the object area is divided into grids at the selected resolution; second, the grids controlled by the same set of rainfall stations are divided into a sub-area; d c is the controllable distance of the rainfall station, g X represents the rain gauges controlling the same sub-area, which is a subset of all rain gauges; For a sub-area, the estimated rainfall at any grid point x The theoretical value range of is: in, in, is the actual rainfall at station G i The measured rainfall value, G i is the actual rainfall station within the control distance of grid point x; The second step is to recommend the number and location of the required virtual stations for each sub-region. The number and location of the virtual stations are recommended after spatial comparison between multi-source grid telemetry rainfall products and point-based measured rainfall. The grid points where the actual rainfall exceeds the theoretical value range in the first step can be selected as the virtual sites to be supplemented. The two grid points where the maximum rainfall and the minimum rainfall exceed the theoretical value range are respectively selected as the virtual sites of the sub-area; if a sub-area has the following situation: there are some grid points with actual rainfall less than P min , and at the same time, the actual rainfall at some grid points is greater than P max , then the sub-area needs to add virtual stations at the two grid points where the maximum rainfall and the minimum rainfall are respectively located; there is no grid point where the actual rainfall is less than P min , there are some grid points where the actual rainfall is greater than P max , then the sub-area needs to add a virtual station at the grid point where the maximum rainfall is located; if the actual rainfall at the grid point is less than P min , there is no grid point where the actual rainfall is greater than P max , then the sub-area needs to add a virtual station at the grid point where the minimum rainfall is located; the actual rainfall of all grid points is within [P min , P max ], then the sub-area does not need to add virtual stations; therefore, the number of virtual stations that need to be added in each sub-area does not exceed two; After all precipitation products have completed the steps of recommending virtual sites, if the virtual site locations recommended by more than half of the products are the same or adjacent, then this location is the final virtual site location. Otherwise, no additional virtual sites are needed in this sub-area. The third step is to estimate the rainfall at the virtual site using the existing non-spatial fusion method. The idea of the non-spatial fusion method can be summarized as follows: 2.1) Training the model Based on various regression analyses, a transformation model of dependent and independent variables is established, where the dependent variable is the measured rainfall at the station, and the independent variables are the rainfall estimates of multi-source rainfall products at the actual station location and other relevant covariates. Where f(·) is the trained transformation model, P G is the measured rainfall at the actual station G, is the rainfall product S i The estimated rainfall at G, A G is a set of covariates related to G, such as terrain information, and ε is the model error; 2.2) Estimated rainfall Estimation of rainfall at virtual sites based on the trained transformation model The dependent variable is the rainfall at the virtual site, and the independent variables are the rainfall estimates of the multi-source rainfall products at the virtual site locations and other relevant covariates; In the formula, Rainfall Product S i The estimated rainfall at the virtual site V, A V is a covariate associated with V; The fourth step is to determine whether the distance between all grid points in the target area and the nearest rainfall station is less than the controllable distance d of the rainfall station selected in the first step. c , where the rainfall site is the total set of actual sites and virtual sites G = [G1, G2, ..., V1, V2, ...]; if yes, go to step 5; if no, take the total set of sites as the updated actual site set and go to step 1; The fifth step is to perform spatial interpolation based on the existing spatial interpolation method using actual sites and virtual sites to obtain the rainfall field.
2. The fusion framework of grid-shaped remotely sensed rainfall and point-shaped measured rainfall based on virtual sites according to claim 1 is characterized in that: In the first step, the controllable distance d of the rainfall station c Select between 10 and 50 km.
3. The fusion framework of grid-shaped remotely sensed rainfall and point-shaped measured rainfall based on virtual sites according to claim 1 is characterized in that: In the second step, since the actual rainfall is unknown, the comparison between the actual rainfall and the boundary of the theoretical value range needs to be made with the help of multi-source grid telemetry rainfall products. The specific steps are as follows: The rainfall field estimated by each grid telemetry rainfall product at a single time step is normalized to 0-1 to obtain the normalized rainfall field Then at point x: In the formula, is the rainfall field estimated by product S in a single time step, is the rainfall estimated by product S at grid point x in a single time step; For a sub-region, if So, The location of is the location of the maximum rainfall virtual site recommended by product S; otherwise, product S does not recommend the maximum rainfall virtual site; where, is a sub-region location X=[x1,x2,…,x m ] are the grid points contained in the sub-region, is the site location that controls the sub-area a max is a parameter that enhances the significance of the inequality; Similarly, if, So, The location of is the location of the minimum rainfall virtual site recommended by product S; otherwise, product S does not recommend the minimum rainfall virtual site; where a min is a parameter that enhances the significance of the inequality.
4. The fusion framework of grid-shaped remotely sensed rainfall and point-shaped measured rainfall based on virtual sites according to claim 3 is characterized in that: The a max The value is between 0.1-0.
2.
5. The fusion framework of grid-shaped remotely sensed rainfall and point-shaped measured rainfall based on virtual sites according to claim 3 is characterized in that: The a min The value is between 0-0.1.
