A method for constructing historical precipitation datasets based on the synergistic fusion of accuracy and rainfall.
By combining categorical and continuous variables with a random forest model, the spatial coverage and representativeness issues of historical precipitation datasets were resolved. A high-precision, spatiotemporally continuous precipitation dataset was constructed, reducing false alarms for rainless locations and providing uncertainty analysis of precipitation.
Patent Information
- Application Number
- CN202611602003.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-16
- Publication Date
- 2026-11-17
AI Technical Summary
Existing historical precipitation datasets have limited spatial coverage and representativeness. The accuracy of multi-source fusion data is affected by the density and number of ground stations. Traditional methods are difficult to obtain objective accuracy indicators directly in areas without station grids. Furthermore, multi-source fusion cannot fully utilize the classification accuracy and quantitative accuracy characteristics of different precipitation products.
We used a triple combination method for categorical variables and a triple combination method for continuous variables to obtain the initial accuracy index at the grid. We then used a random forest fusion model to combine meteorological station observations, geographical variables, and spatial autocorrelation variables to construct a historical precipitation dataset based on the synergistic fusion of accuracy and rainfall. We processed the rainy and dry states and precipitation through random forest classification and regression models to form a spatiotemporally continuous fusion result.
It achieves the construction of a high-precision, spatiotemporally continuous historical precipitation dataset, reduces the problem of false precipitation in rainless locations, avoids the interference of zero precipitation on non-zero precipitation, and provides confidence intervals for precipitation, thus improving the uncertainty analysis of the fusion results.
Smart Images

Figure CN123116027A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of precipitation product accuracy evaluation and multi-source precipitation information fusion, and in particular to a method for constructing a historical precipitation dataset based on the synergistic fusion of accuracy and rainfall. Background Technology
[0002] Reliable historical precipitation datasets are essential for evaluating the accuracy of precipitation forecasts from numerical model ensembles and the effectiveness of statistical post-processing. Existing historical precipitation datasets mainly include precipitation data from discrete ground stations, spatially interpolated precipitation data from ground station networks, precipitation data retrieved from satellite remote sensing, reanalysis precipitation data, and multi-source fused precipitation data. Ground station data suffers from limited spatial coverage and representativeness due to the sparse and uneven distribution of the station network; while satellite remote sensing inversion and reanalysis precipitation data are spatially continuous, they have significant inherent errors; and the accuracy of existing multi-source fused precipitation data is further affected by the density and number of local ground stations.
[0003] Multi-source precipitation information fusion is an important approach to improving the spatiotemporal accuracy of historical precipitation datasets, typically involving two stages: precipitation product accuracy evaluation and precipitation information fusion. For precipitation product accuracy evaluation, traditional methods require ground-observed precipitation as the ground truth, making it difficult to directly obtain objective accuracy indicators at locations without station grids. The triple-combination method does not require ground-observed precipitation and can estimate classification or quantitative accuracy indicators based on three types of precipitation products with independent errors; however, its mathematical independence assumption is difficult to fully satisfy, potentially leading to biased accuracy indicators. For precipitation information fusion, random forests can fuse grid precipitation products, geographic variables, meteorological variables, and spatial autocorrelation variables; however, direct fusion fails to fully utilize the complementary characteristics of different precipitation products in terms of classification accuracy, quantitative accuracy, and different spatiotemporal locations. Summary of the Invention
[0004] Purpose of the invention: The purpose of this invention is to provide a method for constructing a historical precipitation dataset based on the synergistic fusion of accuracy and rainfall, which can output a high-precision, spatiotemporally continuous historical precipitation dataset.
[0005] Technical solution: The present invention provides a method for constructing a historical precipitation dataset based on the synergistic fusion of accuracy and rainfall, comprising:
[0006] S1: Obtain meteorological station observation precipitation and three precipitation products from the study watershed. The three precipitation products include satellite remote sensing inversion precipitation products, precipitation products estimated based on soil water information, and land surface reanalysis precipitation products. Unify the spatiotemporal resolution of each precipitation product, construct a triple combination that satisfies the error independence assumption, and divide all grids in the study watershed into grids with stations and grids without stations. The grids with stations are further divided into modeling grids and verification grids.
[0007] S2: A categorical variable triple combination method is used to obtain the classification accuracy index of each precipitation product at all grids. Based on the classification accuracy index, the equilibrium accuracy of each precipitation product is obtained. A continuous variable triple combination method is used to obtain the quantitative accuracy index of each precipitation product at all grids, including the root mean square error. The traditional accuracy index corresponding to the precipitation observed by the meteorological station based on the modeled grid is calculated. For the equilibrium accuracy and root mean square error, a random forest fusion model of accuracy index is established respectively. Each random forest fusion model of accuracy index takes geographical variables, spatial autocorrelation variables, and the corresponding accuracy index at the modeled grid as inputs and the traditional accuracy index as output. The random forest fusion model of accuracy index is applied to the test grid to evaluate the accuracy index deviation, and then applied to all grids to obtain a spatially continuous fused accuracy index.
[0008] S3: Based on the balance accuracy of each precipitation product after fusion, determine the corresponding fusion weight for rainy and rainless states, perform weighted fusion of the rainy and rainless states of the three precipitation products, and obtain the preliminary fusion result of rainy and rainless states through the sign function; based on the root mean square error of each precipitation product after fusion, determine the corresponding precipitation amount fusion weight, perform weighted fusion of the precipitation amounts of the three precipitation products, and obtain the preliminary fusion result of precipitation amounts.
[0009] S4: Establish a random forest classification model and a random forest regression model. The random forest classification model takes the preliminary fusion result of rainy / dry state, geographic variables, spatial autocorrelation variables, and meteorological variables as input, and the actual rainy / dry state at the modeled grid as output. The random forest regression model takes the preliminary fusion result of precipitation, geographic variables, spatial autocorrelation variables, and meteorological variables as input, and the measured precipitation at the modeled grid as output. Apply the random forest classification model and the random forest regression model to all grids to obtain the rainy / dry state fusion result and the precipitation fusion result, respectively. Combine the rainy / dry state fusion result and the precipitation fusion result according to the corresponding grid and time period, and repeat the process for each time period to obtain a spatiotemporally continuous historical precipitation dataset.
[0010] Furthermore, the land surface reanalysis precipitation product and the precipitation product estimated based on soil water information are used as fixed members of the triple combination, and the satellite remote sensing inverted precipitation product is used as a variable member of the triple combination.
[0011] Furthermore, the unified spatiotemporal resolution includes: unifying each precipitation product to a spatial resolution of 0.1° and a temporal resolution of 3h; performing bilinear interpolation on precipitation products with different spatial resolutions, performing temporal accumulation on satellite remote sensing inversion precipitation products with different temporal resolutions, calculating the intra-day precipitation distribution ratio based on the satellite remote sensing inversion precipitation data, and decomposing the precipitation products estimated based on soil water information to a 3h temporal resolution using the intra-day precipitation distribution ratio.
[0012] Furthermore, the method for tripartite matching of categorical variables includes:
[0013] Set a precipitation threshold; +1 indicates rain, and -1 indicates no rain.
[0014] Under the condition that the classification and identification errors of the three precipitation products for the rainy and rainless states are independent, a classification accuracy index corresponding to the balance accuracy is obtained based on the covariance between the rainy and rainless states of the three precipitation products, and the balance accuracy of each precipitation product is obtained according to the classification accuracy index.
[0015] No. Balance accuracy of various precipitation products satisfy:
[0016]
[0017] in, For the first Estimated values of the rain-affected and rainless states of various precipitation products; This represents the actual state of whether it is raining or not. This represents conditional probability.
[0018] Furthermore, the continuous variable triple combination method includes:
[0019] Under the conditions that the expected precipitation estimation error of each precipitation product is zero, the precipitation estimation errors of different precipitation products are independent of each other, and the precipitation estimation errors of each precipitation product are independent of the actual precipitation, the root mean square error of each precipitation product is obtained based on the variance and pairwise covariance of the precipitation of the three precipitation products.
[0020] Furthermore, the first The fusion weights for the rainy / rain-free status of various precipitation products are represented as follows:
[0021]
[0022] in, Represents the th random forest after merging The balance accuracy of various precipitation products; It is a non-negative real power exponent; The non-negative real power exponent configuration is designed to ensure that the balance accuracy of the preliminary fusion results of rainy and rainless states is not lower than that of the best single precipitation product.
[0023] Furthermore, the geographical variables include longitude, latitude, and elevation; the spatial autocorrelation variables are the Euclidean distance matrix between each grid cell in the study basin and each of the modeling grid cells; the meteorological correlation variables include air temperature at 2m, zonal wind speed at 10m above the ground, meridional wind speed at 10m above the ground, actual evapotranspiration, surface air pressure, total cloud cover, total column water volume, and total column water vapor.
[0024] This invention uses the Euclidean distance matrix from the modeling grid to represent spatial autocorrelation, avoiding the repeated use of spatial interpolation data of station observation precipitation at both the input and output ends of the model, and keeping the relationship between the modeling data clear.
[0025] Furthermore, the number of random features in each leaf node of each random forest model is one-third of the total number of features, and Bayesian optimization is used to fine-tune the number of decision trees and the minimum number of samples contained in the leaf nodes.
[0026] Furthermore, when obtaining the historical precipitation dataset, the fusion result of rainy and rainless states is represented as a state variable with a value of 1 for rainy and a value of 0 for rainless, and multiplied with the precipitation fusion result of the same grid and the same time period to correct the precipitation fusion result at the rainless location.
[0027] Existing technologies often treat multi-source precipitation fusion as a single regression task for precipitation estimation, which can easily lead to false precipitation reports in rainless locations or cause the estimation of heavy precipitation to be affected by a large number of zero-value precipitation. This invention processes rainy / rainless states and precipitation amounts separately, and uses the rainy / rainless states at the end to correct the precipitation fusion results. This can reduce the problem of false precipitation reports in rainless locations, while avoiding interference from a large number of zero-value precipitation reports on the fusion of non-zero precipitation amounts.
[0028] Furthermore, the method for constructing a historical precipitation dataset based on the synergistic fusion of accuracy and rainfall also includes establishing a quantile regression forest based on the input and output data of the random forest regression model trained at the modeling grid location, establishing a precipitation prediction probability distribution from the quantile regression forest, and calculating the predicted mean and 5%, 25%, 75%, and 95% quantiles of the precipitation fusion result. The 5% and 95% quantiles are used as the lower and upper limits of the 90% confidence interval, and the 25% and 75% quantiles are used as the lower and upper limits of the 50% confidence interval. Then, the upper and lower limits of each confidence interval are multiplied by the fusion result of the rainy and rainless states to obtain the confidence interval of the historical precipitation dataset.
[0029] Existing studies on multi-source precipitation fusion have relatively insufficient analysis of the uncertainties in the fusion results, making it difficult to simultaneously provide precipitation estimates and their confidence intervals. This invention uses quantile regression forests to provide quantiles and confidence intervals for precipitation fusion results, and synthesizes the confidence intervals with the fusion results of rainy and rainless states, so that historical precipitation data not only includes precipitation estimates, but also the uncertainty range corresponding to those estimates.
[0030] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:
[0031] (1) The present invention first uses the triple combination method of categorical variables and the triple combination method of continuous variables to obtain the initial accuracy index of all grid positions, and then uses the traditional accuracy index of the grid with station to train the accuracy index random forest fusion model. Therefore, the information of the grid with station can be used to correct the triple combination evaluation bias, and the correction relationship can be extended to the grid without station to form a spatially continuous accuracy index distribution.
[0032] (2) This invention integrates the accuracy index fusion and precipitation information fusion in series: the classification and quantitative fusion accuracy index output by the previous stage is directly used for the fusion weight calculation of the next stage, so that the complementary advantages of different precipitation products in classification identification, quantitative error and different time and space locations can enter the precipitation fusion process.
[0033] (3) Based on the preliminary fusion results of the triple combination, the random forest classification model and random forest regression model of the present invention further integrate meteorological station observation precipitation, geographical variables, spatial autocorrelation variables and meteorological related variables, and can simultaneously utilize multi-source precipitation information and multivariate interpretation information. Attached Figure Description
[0034] Figure 1 This is a flowchart of a method for constructing a historical precipitation dataset based on the synergistic fusion of accuracy and rainfall, provided by an embodiment of the present invention. Detailed Implementation
[0035] The invention will now be further described with reference to the accompanying drawings.
[0036] Definitions of abbreviations:
[0037] CTC is an abbreviation for Categorical Triple Collocation, which refers to the method of triplicity collocation of categorical variables.
[0038] TC is an abbreviation for Triple Collocation, which indicates a method of triple collocation of continuous variables.
[0039] RF is an abbreviation for Random Forest.
[0040] This invention integrates the triple-combination evaluation results that do not require ground observations with the traditional evaluation results that have station grids. Then, it uses the fused accuracy index to drive the separate fusion of rainy / rainy status and precipitation. Furthermore, it combines ground observations, geographic meteorological information, and spatial autocorrelation information to construct a spatially continuous historical precipitation dataset with uncertain quantitative results.
[0041] like Figure 1 As shown in the figure, this invention provides a method for constructing a historical precipitation dataset based on the synergistic fusion of accuracy and rainfall, including the following steps:
[0042] S1: Obtain meteorological station observation precipitation and three precipitation products from the study watershed. The three precipitation products include satellite remote sensing inversion precipitation products, precipitation products estimated based on soil water information, and land surface reanalysis precipitation products. Unify the spatiotemporal resolution of each precipitation product, construct a triple combination that satisfies the error independence assumption, and divide all grids in the study watershed into grids with stations and grids without stations. Grids with stations are further divided into modeling grids and verification grids.
[0043] Satellite remote sensing precipitation products include IMERG Final, CMORPH RAW, and GsMAP_MVK; precipitation products estimated based on soil water information are SM2RAIN-ASCAT; and land surface reanalysis precipitation products are ERA5-Land. ERA5-Land and SM2RAIN-ASCAT are used as fixed members in the triplet combination, and are combined with IMERG Final, CMORPH RAW, and GsMAP_MVK respectively to form three triplet combinations:
[0044] Triple combination A: ERA5-Land, SM2RAIN-ASCAT, and IMERG Final;
[0045] Triple combination B: ERA5-Land, SM2RAIN-ASCAT, and CMORPH RAW;
[0046] Triple combination C: ERA5-Land, SM2RAIN-ASCAT, and GsMAP_MVK.
[0047] The three precipitation products in the triple combination satisfy the conditional independence assumption, that is, given the actual rainy or rainless conditions, the classification and identification errors of each precipitation product are independent of each other.
[0048] All precipitation products were standardized to a spatial resolution of 0.1° and a temporal resolution of 3 hours. Bilinear interpolation was performed on precipitation products with different spatial resolutions (CMORPH RAW) to obtain 0.1° resolution data. Temporal accumulation was performed on satellite remote sensing inversion precipitation products with different temporal resolutions (IMERG Final, CMORPH RAW, and GsMAP_MVK) to obtain 3-hour precipitation data. The intra-day distribution ratio of precipitation was calculated based on the intra-day distribution ratio of precipitation, and the precipitation product estimated based on soil water information (SM2RAIN-ASCAT) was decomposed to a 3-hour temporal resolution.
[0049] The entire 0.1°×0.1° grid within the study basin was divided into grids with and without meteorological stations, with each grid containing one meteorological station. Using K-means clustering and random sampling, the grids with stations were further divided into modeling grids and validation grids. The modeling grids were used to subsequently establish accuracy indices for the random forest fusion model, random forest classification model, and random forest regression model; the validation grids were used to subsequently evaluate the bias of the accuracy indices of each precipitation product in the model output triplet and the accuracy of the precipitation fusion data.
[0050] S2: A triple combination method for categorical variables is used to obtain the classification accuracy index of each precipitation product at all grids, and the equilibrium accuracy of each precipitation product is obtained based on the classification accuracy index. A triple combination method for continuous variables is used to obtain the quantitative accuracy index of each precipitation product at all grids, including the root mean square error. Traditional accuracy indexes are calculated based on the meteorological station observation precipitation of the modeled grids. Stochastic forest fusion models for accuracy indices are established for equilibrium accuracy and root mean square error respectively. Each stochastic forest fusion model for accuracy indices takes geographical variables, spatial autocorrelation variables, and the corresponding accuracy indexes at the modeled grids as inputs and traditional accuracy indices as outputs. The stochastic forest fusion models for accuracy indices are applied to the test grids to evaluate the accuracy index bias, and then applied to all grids to obtain spatially continuous fused accuracy indices.
[0051] The triple-combination method for categorical variables is used to evaluate the classification accuracy of precipitation products, including: setting precipitation thresholds, such as a threshold for the boundary between rainy and rainless conditions. When precipitation exceeds the threshold for distinguishing between rainy and dry weather The time is determined to be rainy. In the triple combination calculation stage, +1 indicates a rainy state and -1 indicates a rainless state.
[0052] Based on the covariance between the rainy and dry states of the three precipitation products, the classification accuracy index corresponding to the balance accuracy is calculated, and the balance accuracy of each precipitation product is obtained based on the classification accuracy index.
[0053] No. Balance accuracy of various precipitation products satisfy:
[0054]
[0055] in, For the first Estimated values of the rain-affected and rainless states of various precipitation products; This represents the actual state of whether it is raining or not. This represents conditional probability. It is the arithmetic mean of conditional probabilities and is a dimensionless quantity. Using balanced precision can mitigate the overestimation or underestimation of classification evaluation caused by the imbalance between the number of rainy and rainless samples.
[0056] In addition to the threshold for distinguishing between rainy and dry weather, the precipitation threshold can also be the 95th percentile determined based on observed precipitation, which can be used as the threshold for heavy precipitation to evaluate the ability of each precipitation product to classify and identify heavy precipitation events.
[0057] The triple combination method for continuous variables is used to evaluate the quantitative accuracy of non-zero precipitation. This method assumes that the expected value of the precipitation estimation error for the three precipitation products is zero, that the estimation errors for each product are independent of each other, and that the estimation errors for each product are independent of the actual precipitation. The method is based on the variance of the precipitation values from the three precipitation products. , , and pairwise covariance , , , No. The error variance of this precipitation product can be expressed as:
[0058]
[0059] Thus obtaining the first Root mean square error of precipitation products:
[0060]
[0061] in, The variance of precipitation for the three precipitation products is given. , , Let be the pairwise covariance of precipitation for the three precipitation products. and The units of measurement are all precipitation squared. The dimension of precipitation is the square of the amount of precipitation. It has the same dimensions as precipitation. RMSEr addresses the precipitation error after linear fitting of precipitation products and is not used interchangeably with the classic RMSE, which directly compares precipitation products and observed precipitation.
[0062] Geographic variables include longitude, latitude, and elevation; spatial autocorrelation variables are the Euclidean distance matrices between each grid cell and each modeling grid cell within the study watershed. The distance matrix is used to characterize spatial autocorrelation. Spatial interpolation data of station-observed precipitation are not used as spatial autocorrelation variables, thus avoiding the simultaneous use of observed precipitation at both the model input and output.
[0063] First, the accuracy index random forest fusion model is applied to the test grid, and the traditional accuracy index of the test grid is used as a benchmark to calculate the deviation. The accuracy index random forest fusion model is then used to reduce the deviation of the triple-combination accuracy index. Next, the accuracy index random forest fusion model is applied to all grids to obtain a spatially continuous fusion accuracy index. The fusion accuracy index is used to analyze the spatiotemporal accuracy differentiation of various precipitation products, and also serves as the weighting basis for the subsequent preliminary fusion of rainy / dry conditions and precipitation amounts.
[0064] S3: Based on the balance accuracy of each precipitation product after fusion, determine the corresponding fusion weight for rainy and rainless states, perform weighted fusion of the rainy and rainless states of the three precipitation products, and obtain the preliminary fusion result of rainy and rainless states through the sign function; based on the root mean square error of each precipitation product after fusion, determine the corresponding precipitation amount fusion weight, perform weighted fusion of the precipitation amounts of the three precipitation products, and obtain the preliminary fusion result of precipitation amounts.
[0065] The fusion weights for rainy and rainless states are calculated based on the balance accuracy of each precipitation product after random forest fusion. The fusion weights for the rainy / rain-free status of various precipitation products are represented as follows:
[0066]
[0067] in, Represents the th random forest after merging The balance accuracy of various precipitation products; It is a non-negative real power exponent; .
[0068] The fusion weight of rainy and rainless states is positively correlated with the classification and identification ability of the corresponding precipitation products. By optimizing the configuration of non-negative real power exponents, the balance accuracy of the preliminary fusion results of rainy and rainless states is no less than that of the best single precipitation product.
[0069] The rain-or-nothing state variables of the three precipitation products are weighted using the rain-or-nothing state fusion weights, and the preliminary fusion results of the rain-or-nothing state are obtained through a sign function:
[0070]
[0071] in, This is the preliminary fusion result for rainy and dry conditions. Fusion weights for rainy and dry conditions. Estimated values for rainy and rainless conditions Both the balance accuracy and the balance precision are dimensionless quantities.
[0072] The precipitation fusion weights for each precipitation product are calculated based on the root mean square error of each product after random forest fusion, minimizing the variance of the initial weighted precipitation fusion result. By taking the partial derivatives of the fusion weights for each product, a unified weight calculation formula for each product can be derived as follows:
[0073]
[0074] in, This represents the th generation after random forest fusion. The root mean square error of precipitation products.
[0075] Preliminary precipitation fusion results were obtained using precipitation fusion weights:
[0076]
[0077] in, For the first Rainfall amount for various precipitation products This is the preliminary result of precipitation data aggregation. The precipitation aggregation weight is a dimensionless quantity. and The dimensions are consistent.
[0078] S4: Establish a random forest classification model and a random forest regression model. The random forest classification model takes the preliminary fusion results of rainy / dry states, geographic variables, spatial autocorrelation variables, and meteorological variables as inputs, and the actual rainy / dry state at the modeled grid as output. The random forest regression model takes the preliminary fusion results of precipitation, geographic variables, spatial autocorrelation variables, and meteorological variables as inputs, and the measured precipitation at the modeled grid as output. Apply the random forest classification model and the random forest regression model to all grids to obtain the rainy / dry state fusion results and precipitation fusion results, respectively. Combine the rainy / dry state fusion results and precipitation fusion results according to the corresponding grids and time periods, and repeat the process for each time period to obtain a spatiotemporally continuous historical precipitation dataset.
[0079] Meteorological variables include air temperature at 2m, zonal wind speed at 10m above ground, meridional wind speed at 10m above ground, actual evapotranspiration, surface air pressure, total cloud cover, total column water volume, and total column water vapor. Total column water volume and total column water vapor are used to characterize precipitable water.
[0080] When obtaining historical precipitation datasets, the fusion result of rainy and rainless states is represented as a state variable with a value of 1 for rainy and a value of 0 for rainless, and multiplied with the precipitation fusion result of the same grid and the same time period to correct the precipitation fusion result at rainless locations.
[0081]
[0082] in, Historical precipitation dataset; The result is a fusion of rainy and rainless conditions; This is the result of precipitation data aggregation. Since it is a dimensionless state variable, therefore and The dimensions are consistent. This processing makes the merged precipitation corresponding to the rainless state zero, avoiding the direct participation of the -1 state used in the triple combination calculation stage in the multiplication.
[0083] Each random forest model (including the accuracy-based random forest fusion model, random forest classification model, and random forest regression model) samples with replacement from the original dataset to form a training subset for each decision tree, and randomly selects some feature variables when splitting nodes. The number of random features in each leaf node is one-third of the total number of features, and Bayesian optimization is used to fine-tune the number of decision trees and the minimum number of samples contained in the leaf nodes.
[0084] Uncertainty analysis of fusion results: A quantile regression forest is established based on the input and output data of the random forest regression model trained at the modeling grid location. The quantile regression forest is used to establish the probability distribution of precipitation prediction and calculate the predicted mean and 5%, 25%, 75%, and 95% quantiles of the precipitation fusion results. The 5% and 95% quantiles are used as the lower and upper limits of the 90% confidence interval, and the 25% and 75% quantiles are used as the lower and upper limits of the 50% confidence interval. Since the quantile regression forest is only used for regression problems, the confidence intervals of the precipitation fusion results are obtained first. Then, the upper and lower limits of each confidence interval are multiplied by the fusion results of rainy and rainless states, respectively, to obtain the 90% and 50% confidence intervals of the historical precipitation dataset.
[0085] The following is an application example.
[0086] This example uses the Han River basin as the study area, covering the period from May to November each year from 2018 to 2022, with a target spatial resolution of 0.1° and a target temporal resolution of 3 hours. The Han River basin contains a total of 1504 evaluation grids, of which 59 are monitored by weather stations. Using K-means clustering and random sampling, 41 monitored grids were selected as modeling grids, and 18 monitored grids were selected as validation grids.
[0087] In the accuracy evaluation stage, the equilibrium accuracy (BA), equilibrium accuracy under the heavy precipitation threshold, and root mean square error (RMSEr) were calculated using the categorical variable triple combination method, the continuous variable triple combination method, the traditional method based on precipitation observations from meteorological stations, and the random forest fusion method of this invention. Verification grid results show that after random forest fusion, the deviations of the classification accuracy and quantitative accuracy of each precipitation product relative to the traditional method approach zero, and the spatial differences in accuracy deviations are reduced.
[0088] In the precipitation information fusion stage, three sets of triple combinations were subjected to continuous variable triple combination fusion, categorical variable and continuous variable triple combination fusion, and the CTC-TC-RF fusion of the present invention, respectively. Experimental results show that, compared with the CTC-TC method, the CTC-TC-RF method improves the critical success index and Kling-Gupta efficiency coefficient by more than 10% and 20%, respectively; the fusion results can better capture the spatiotemporal evolution of heavy precipitation areas, and the 50% confidence interval has high coverage of observed precipitation.
[0089] Among the three triplet combinations, the combined data obtained by the A-driven CTC-TC-RF method consisting of ERA5-Land, SM2RAIN-ASCAT, and IMERG Final shows slightly better overall performance. Therefore, the 0.1° and 3h precipitation fusion data output by this triplet combination can be used as a historical precipitation dataset for evaluating the accuracy of ensemble forecast precipitation and for statistical post-processing.
Claims
1. A method for constructing a historical precipitation dataset based on the synergistic fusion of accuracy and rainfall, characterized in that, include: S1: Obtain meteorological station observation precipitation and three precipitation products from the study watershed. The three precipitation products include satellite remote sensing inversion precipitation products, precipitation products estimated based on soil water information, and land surface reanalysis precipitation products. Unify the spatiotemporal resolution of each precipitation product, construct a triple combination that satisfies the error independence assumption, and divide all grids in the study watershed into grids with stations and grids without stations. The grids with stations are further divided into modeling grids and verification grids. S2: The classification accuracy index of each precipitation product at all grids is obtained by using the triple combination method of classification variables, and the balance accuracy of each precipitation product is obtained based on the classification accuracy index. A continuous variable triple combination method is used to obtain the quantitative accuracy index of each precipitation product at all grids, including the root mean square error. Traditional accuracy indices are calculated based on the meteorological station observations of the modeled grids. For the equilibrium accuracy and root mean square error, separate random forest fusion models for accuracy indices are established. Each random forest fusion model takes geographical variables, spatial autocorrelation variables, and the corresponding accuracy index at the modeled grid as inputs and the traditional accuracy index as output. The random forest fusion models for each accuracy index are applied to the test grid to evaluate the accuracy index deviation, and then applied to all grids to obtain spatially continuous fused accuracy indices. S3: Based on the balance accuracy of each precipitation product after fusion, determine the corresponding fusion weight for rainy and rainless states, perform weighted fusion of the rainy and rainless states of the three precipitation products, and obtain the preliminary fusion result of rainy and rainless states through the sign function; based on the root mean square error of each precipitation product after fusion, determine the corresponding precipitation amount fusion weight, perform weighted fusion of the precipitation amounts of the three precipitation products, and obtain the preliminary fusion result of precipitation amounts. S4: Establish a random forest classification model and a random forest regression model. The random forest classification model takes the preliminary fusion results of rainy / dry states, geographical variables, spatial autocorrelation variables, and meteorological variables as inputs, and the actual rainy / dry state at the modeling grid as output. The random forest regression model takes the preliminary fusion results of precipitation, geographical variables, spatial autocorrelation variables, and meteorological variables as inputs, and the measured precipitation at the modeling grid as output. Random forest classification model and random forest regression model were applied to all grids to obtain the fusion results of rainy and rainless states and precipitation fusion results, respectively. The results of merging rainy and dry conditions and precipitation data are combined according to the corresponding raster and time period, and repeated for each time period to obtain a spatiotemporally continuous historical precipitation dataset.
2. The method for constructing a historical precipitation dataset based on the synergistic fusion of accuracy and rainfall as described in claim 1, characterized in that, The land surface reanalysis precipitation product and the precipitation product estimated based on soil water information are used as fixed members of the triple combination, and the satellite remote sensing inverted precipitation product is used as a variable member of the triple combination.
3. The method for constructing a historical precipitation dataset based on the synergistic fusion of accuracy and rainfall as described in claim 1, characterized in that, The unified spatiotemporal resolution includes: unifying all precipitation products to a spatial resolution of 0.1° and a temporal resolution of 3 hours; performing bilinear interpolation on precipitation products with different spatial resolutions; performing temporal accumulation on precipitation products retrieved from satellite remote sensing with different temporal resolutions; calculating the intra-day distribution ratio of precipitation based on the intra-day distribution ratio of precipitation; and decomposing the precipitation products estimated based on soil water information to a 3-hour temporal resolution.
4. The method for constructing a historical precipitation dataset based on the synergistic fusion of accuracy and rainfall as described in claim 1, characterized in that, The method for tripartite categorical variables includes: Set a precipitation threshold; +1 indicates rain, and -1 indicates no rain. Under the condition that the classification and identification errors of the three precipitation products for the rainy and rainless states are independent, a classification accuracy index corresponding to the balance accuracy is obtained based on the covariance between the rainy and rainless states of the three precipitation products, and the balance accuracy of each precipitation product is obtained according to the classification accuracy index. No. Balance accuracy of various precipitation products satisfy: in, For the first Estimated values of the rain-affected and rainless states of various precipitation products; This represents the actual state of whether it is raining or not. This represents conditional probability.
5. The method for constructing a historical precipitation dataset based on the synergistic fusion of accuracy and rainfall as described in claim 1, characterized in that, The continuous variable triple combination method includes: Under the conditions that the expected precipitation estimation error of each precipitation product is zero, the precipitation estimation errors of different precipitation products are independent of each other, and the precipitation estimation errors of each precipitation product are independent of the actual precipitation, the root mean square error of each precipitation product is obtained based on the variance and pairwise covariance of the precipitation of the three precipitation products.
6. The method for constructing a historical precipitation dataset based on the synergistic fusion of accuracy and rainfall as described in claim 1, characterized in that, No. The fusion weights for the rainy / rain-free status of various precipitation products are represented as follows: in, Represents the th random forest after merging The balance accuracy of various precipitation products; It is a non-negative real power exponent; The non-negative real power exponent configuration is designed to ensure that the balance accuracy of the preliminary fusion results of rainy and rainless states is not lower than that of the best single precipitation product.
7. The method for constructing a historical precipitation dataset based on the synergistic fusion of accuracy and rainfall as described in claim 1, characterized in that, The geographic variables include longitude, latitude, and elevation; the spatial autocorrelation variables are the Euclidean distance matrix between each grid cell in the study watershed and each of the modeling grid cells; the meteorological correlation variables include air temperature at 2m, zonal wind speed at 10m above the ground, meridional wind speed at 10m above the ground, actual evapotranspiration, surface air pressure, total cloud cover, total column water volume, and total column water vapor.
8. The method for constructing a historical precipitation dataset based on the synergistic fusion of accuracy and rainfall as described in claim 1, characterized in that, Each random forest model uses one-third of the total number of random features in each leaf node, and Bayesian optimization is used to fine-tune the number of decision trees and the minimum number of samples contained in each leaf node.
9. The method for constructing a historical precipitation dataset based on the synergistic fusion of accuracy and rainfall as described in claim 1, characterized in that, When the historical precipitation dataset is obtained, the fusion result of rainy and rainless states is represented as a state variable with a value of 1 for rainy and a value of 0 for rainless, and multiplied with the precipitation fusion result of the same grid and the same time period to correct the precipitation fusion result at the rainless location.
10. The method for constructing a historical precipitation dataset based on the synergistic fusion of accuracy and rainfall as described in claim 1, characterized in that, It also includes establishing a quantile regression forest based on the input and output data of the random forest regression model trained at the modeling grid location, establishing the precipitation prediction probability distribution from the quantile regression forest, and calculating the predicted mean and 5%, 25%, 75%, and 95% quantiles of the precipitation fusion result. The 5% and 95% quantiles are used as the lower and upper limits of the 90% confidence interval, and the 25% and 75% quantiles are used as the lower and upper limits of the 50% confidence interval. Then, the upper and lower limits of each confidence interval are multiplied by the fusion result of the rainy and rainless states to obtain the confidence interval of the historical precipitation dataset.
