Multi-source precipitation product fusion method considering self-correlation information and spatial distribution influence

Abstract

The invention discloses a multi-source precipitation product fusion method considering self-correlation information and spatial distribution influence. The method comprises the following steps: obtaining precipitation data of a to-be-detected grid and a plurality of precipitation products of related grids of the to-be-detected grid; the related grid is a grid which shares a vertex and an edge with the to-be-detected grid; inputting the rainfall data into a preset rainfall intensity classification sub-model, and obtaining rainfall classification output by the rainfall intensity classification sub-model; and inputting the rainfall data into the rainfall segmentation regression sub-model corresponding to the rainfall classification, and receiving data output by the rainfall segmentation regression sub-model as the fused rainfall of the to-be-detected grid. According to the method, the precipitation intensity classification sub-model is constructed, precipitation events are accurately classified based on the precipitation intensity, and the problem that the capture capacity of a traditional RF model is unbalanced under different precipitation intensities is solved; and in combination with the rainfall segmentation regression sub-model, regression analysis is carried out on rainfall events with different intensities, so that refinement processing of rainfall data is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent processing of precipitation data, and in particular to a multi-source precipitation product fusion method that considers the influence of autocorrelated information and spatial distribution. Background Art

[0002] Satellite precipitation is easily accessible and covers a wide range, attracting widespread attention and application in fields such as meteorology, hydrology, agriculture, and disaster management. Since the launch of the Tropical Rainfall Measuring Mission (TRMM), the combined development of multiple technologies, including satellite inversion, data fusion, and numerical simulation, has led to the emergence of a number of high-resolution satellite precipitation products, including GSMaP (Global Satellite Mapping of Precipitation), TMPA (TRMM Multi-satellite Precipitation Analysis), CMORPH (Climate Prediction Center Morphing technique), GPM (Global Precipitation Measurement project), and PERSIANN (Precipitation from Remote Sensed Imagery Using Artificial Neural Network). These products have made significant contributions to global water cycle changes, water resources management, and water disaster monitoring.

[0003] Due to differences in satellite detectors and inversion algorithms, each precipitation product has its own strengths and weaknesses, and often performs differently across different spatial and temporal scales. The commonly used satellite product PERSIANN overestimates precipitation in high-altitude regions such as South Africa, the Indian subcontinent, and Uganda, while significantly underestimating precipitation in Ethiopia. IMERG (Integrated Multi-satellite Retrievals for Global Precipitation Measurement) outperforms the SM2RAIN-ASCAT product in humid regions, while the opposite is true in semi-arid regions. Therefore, the fusion of multi-source precipitation products can fully leverage the strengths of different data products and effectively compensate for the shortcomings of single or independent data sources.

[0004] Precipitation, a key indicator in dynamic flood risk assessment, is a climatic factor with significant spatial autocorrelation and distinct distribution characteristics. Therefore, the spatial autocorrelation of precipitation products and the influence of total precipitation distribution cannot be ignored. Furthermore, for dynamic flood risk assessment, the precipitation fusion method used must not only accurately determine whether a precipitation event has occurred but also precisely distinguish between precipitation events of varying intensities, particularly for torrential rain events that are prone to flooding. However, existing precipitation product fusion models fail to fully consider these key requirements during their construction. Summary of the Invention

[0005] In order to at least overcome the above-mentioned deficiencies in the prior art, the purpose of this application is to provide a multi-source precipitation product fusion method that takes into account the influence of autocorrelation information and spatial distribution.

[0006] In a first aspect, embodiments of the present application provide a multi-source precipitation product fusion method that considers the influence of autocorrelation information and spatial distribution, including:

[0007] Obtaining precipitation data of a plurality of precipitation products of a grid to be detected and grids related to the grid to be detected, wherein the related grids are grids that share vertices and edges with the grid to be detected;

[0008] Inputting the precipitation data into a preset precipitation intensity classification sub-model, and obtaining the precipitation classification output by the precipitation intensity classification sub-model;

[0009] The precipitation data is input into a precipitation segmentation regression sub-model corresponding to the precipitation classification, and data output by the precipitation segmentation regression sub-model is received as the fused precipitation amount of the grid to be detected.

[0010] In a possible implementation, generating the precipitation intensity classification sub-model and the precipitation segmentation regression sub-model includes:

[0011] Acquire grids where multiple precipitation monitoring stations in multiple river basins are located as sample grids, and generate measured precipitation data of the sample grids based on monitoring data of the precipitation monitoring stations;

[0012] Acquire precipitation data of a plurality of precipitation products of the sample grid and corresponding related grids as satellite precipitation data;

[0013] generating a database based on the measured precipitation data and the satellite precipitation data;

[0014] The precipitation intensity classification sub-model and the precipitation segmentation regression sub-model are trained according to the database.

[0015] In one possible implementation, obtaining the measured precipitation data includes:

[0016] The multiple basins are divided into semi-arid, semi-humid and humid areas, and a preset proportion of precipitation monitoring stations are selected from the semi-arid, semi-humid and humid areas respectively as selected stations by random sampling;

[0017] The grid where the selected site is located is used as the sample grid;

[0018] When there is only one selected site in the sample grid, the monitoring data of the selected site is used as the measured precipitation data of the sample grid;

[0019] When there are at least two selected sites in the sample grid, the average value of the monitoring data of all the selected sites in the sample grid is calculated as the measured precipitation data of the sample grid.

[0020] In one possible implementation, generating a database based on the measured precipitation data and the satellite precipitation data includes:

[0021] Acquire measured precipitation data of the sample grid along the time series, and arrange the measured precipitation data longitudinally along the time series to form an output matrix;

[0022] Obtain the satellite precipitation data of the sample grid and the corresponding related grids at the same time for all precipitation products, and arrange them horizontally to form a row vector;

[0023] Arrange the row vectors at multiple moments in the time sequence longitudinally along the time sequence, and align them with the time sequence of the output matrix to form an input matrix;

[0024] The database is formed according to the input matrix and the output matrix.

[0025] In a possible implementation, generating the precipitation intensity classification sub-model includes:

[0026] Converting the measured precipitation data in the output matrix into precipitation classifications to form a classification output matrix; the precipitation classifications are categories divided according to precipitation intensity;

[0027] The input matrix is used as input data, and the classification output matrix is used as output data to train a random forest classification model to form the precipitation intensity classification sub-model.

[0028] In a possible implementation, generating the precipitation segmented regression sub-model includes:

[0029] Splitting the input matrix into a plurality of regression input matrices according to the precipitation classification, and splitting the output matrix into a plurality of regression output matrices according to the precipitation classification; each regression input matrix corresponds to a precipitation classification; each regression output matrix corresponds to a precipitation classification; the precipitation classification includes a trace precipitation classification; the trace precipitation classification is a classification with a precipitation amount of less than 0.1 mm / d;

[0030] The regression input matrix corresponding to the non-trace precipitation classification is used as input data, and the regression output matrix corresponding to the same precipitation classification is used as output data to train the random forest regression model to form multiple precipitation segmented regression sub-models corresponding to different precipitation classifications.

[0031] In a possible implementation, inputting the precipitation data into a precipitation segmented regression sub-model corresponding to the precipitation classification includes:

[0032] When the precipitation data corresponds to the precipitation classification of the trace precipitation classification, the fused precipitation amount corresponding to the precipitation data is defined as 0.

[0033] In a second aspect, the present application provides a multi-source precipitation product fusion system that considers the influence of autocorrelation information and spatial distribution, including:

[0034] an acquisition unit configured to acquire precipitation data of a plurality of precipitation products of a grid to be detected and grids related to the grid to be detected; the related grids are grids that share vertices and edges with the grid to be detected;

[0035] a classification unit configured to input the precipitation data into a preset precipitation intensity classification sub-model and obtain a precipitation classification output by the precipitation intensity classification sub-model;

[0036] The fusion unit is configured to input the precipitation data into a precipitation segmentation regression sub-model corresponding to the precipitation classification, and receive the data output by the precipitation segmentation regression sub-model as the fused precipitation amount of the grid to be detected.

[0037] In a possible implementation, the system further includes a database unit; the database unit is configured to:

[0038] Acquire grids where multiple precipitation monitoring stations in multiple river basins are located as sample grids, and generate measured precipitation data of the sample grids based on monitoring data of the precipitation monitoring stations;

[0039] Acquire precipitation data of a plurality of precipitation products of the sample grid and corresponding related grids as satellite precipitation data;

[0040] generating a database based on the measured precipitation data and the satellite precipitation data;

[0041] The precipitation intensity classification sub-model and the precipitation segmentation regression sub-model are trained according to the database.

[0042] In a possible implementation, the database unit is further configured to:

[0043] The multiple basins are divided into semi-arid, semi-humid and humid areas, and a preset proportion of precipitation monitoring stations are selected from the semi-arid, semi-humid and humid areas respectively as selected stations by random sampling;

[0044] The grid where the selected site is located is used as the sample grid;

[0045] When there is only one selected site in the sample grid, the monitoring data of the selected site is used as the measured precipitation data of the sample grid;

[0046] When there are at least two selected sites in the sample grid, the average value of the monitoring data of all the selected sites in the sample grid is calculated as the measured precipitation data of the sample grid.

[0047] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0048] This invention fully considers the impact of spatial autocorrelation of precipitation products and spatial distribution of precipitation, and conducts in-depth innovative transformation of the traditional RF model. By constructing a precipitation intensity classification sub-model, precipitation events are accurately classified based on precipitation intensity, overcoming the problem of uneven capture ability of the traditional RF model under different precipitation intensities; combined with the precipitation segmented regression sub-model, regression analysis is performed separately for precipitation events of different intensities to achieve refined processing of precipitation data. This innovative near-real-time satellite precipitation product fusion model fundamentally solves the unstable phenomenon of overestimation of weak precipitation and underestimation of strong precipitation in precipitation fusion in the traditional RF model, significantly improving the accuracy of the original precipitation product and the ability to capture precipitation events. Whether in the detection of non-precipitation events, light to moderate precipitation events or extreme precipitation events, it shows excellent performance, providing high-quality data support for downstream applications such as flood inundation simulation and dynamic flood risk assessment, and strongly promoting the development of related research and applications in the field of meteorology and hydrology. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] The drawings described herein are used to provide a further understanding of the embodiments of the present invention, constitute a part of this application, and do not constitute a limitation of the embodiments of the present invention. In the drawings:

[0050] Figure 1 This is a schematic diagram of the steps of the method according to the embodiment of the present application;

[0051] Figure 2This is a schematic diagram of the input matrix structure of the embodiment of the present application;

[0052] Figure 3 The local Moran coefficient scatter points of the satellite precipitation product corresponding to the flood season precipitation in the embodiment of this application;

[0053] Figure 4 This is a box plot of the evaluation indicators of the embodiment of the present application;

[0054] Figure 5 Schematic diagram of the performance of the traditional random forest model RF and the fusion model P-RF of the present application at the watershed scale in an embodiment of the present application;

[0055] Figure 6 This is a schematic diagram of the probability density distribution of different precipitation data under different precipitation intensities according to an embodiment of the present application;

[0056] Figure 7 Schematic diagram of the training process of the precipitation intensity classification sub-model and the precipitation segmentation regression sub-model in an embodiment of the present application. DETAILED DESCRIPTION

[0057] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. It should be understood that the drawings in the present application only serve the purpose of illustration and description and are not used to limit the scope of protection of the present application. In addition, it should be understood that the schematic drawings are not drawn to scale. The flowcharts used in this application illustrate the operations implemented according to some embodiments of the embodiments of the present application. It should be understood that the operations of the flowcharts can be implemented out of sequence, and steps without logical context can be reversed or implemented simultaneously. In addition, those skilled in the art, under the guidance of the contents of this application, can add one or more other operations to the flowchart, or remove one or more operations from the flowchart.

[0058] In addition, the described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments. The components of the embodiments of the present application generally described and shown in the drawings here can be arranged and designed in various configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed application, but merely represents the selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of the present application.

[0059] Please refer to Figure 1, which is a flow chart of a multi-source precipitation product fusion method considering the influence of autocorrelation information and spatial distribution provided in an embodiment of the present invention. Furthermore, the multi-source precipitation product fusion method considering the influence of autocorrelation information and spatial distribution can specifically include the contents described in the following steps S1 to S3.

[0060] S1: Obtain precipitation data of a plurality of precipitation products of a grid to be detected and grids related to the grid to be detected; the related grids are grids that share vertices and edges with the grid to be detected;

[0061] S2: inputting the precipitation data into a preset precipitation intensity classification sub-model, and obtaining the precipitation classification output by the precipitation intensity classification sub-model;

[0062] S3: Inputting the precipitation data into a precipitation segmentation regression sub-model corresponding to the precipitation classification, and receiving data output by the precipitation segmentation regression sub-model as the fused precipitation amount of the grid to be detected.

[0063] During the implementation of the present invention, the inventors discovered that for daily rainfall monitoring during the flood season, the rainfall in a single grid does not exist in isolation, and the rainfall generated by adjacent grids is highly correlated with it. Furthermore, due to the uncertainty of the rainfall direction, in the present invention, all grids that share vertices and edges with the grid to be detected are considered as related grids to characterize the possible impact of rainfall changes in various directions on the rainfall in the current grid. Figure 2 , shows the state of a grid and the corresponding related grids. The current grid is i, and there are 8 related grids around it. In this way, no matter what the specific precipitation direction is, the precipitation data of the current grid can be characterized by the precipitation data of the related grids.

[0064] In the embodiment of the present application, the inventor has conducted a feasibility analysis on the current expression of the relevant grid, see Figure 3 , shows the local Moran coefficient scatter plots of three different satellite precipitation products PERSIANN, GSMaP and IMERG in the relevant grids of this application. It can be seen from the figure that most of the grids of the three satellite precipitation products fall in the first and third quadrants, which shows that the flood season precipitation of the corresponding grids has obvious high-high aggregation properties and low-low aggregation properties with the relevant grids.

[0065] In the embodiments of the present application, in scientific practice, the inventors found that the traditional random forest model is prone to overestimating non-precipitation events (<0.1mm / d) and underestimating precipitation events above heavy rain (≥25mm / d) when fusing precipitation products within different precipitation intensity ranges; this leads to a significant increase in the number of small and moderate rain events (0.1-25mm / d); concentrated data will cause insufficient data diversity, making the random forest model overly dependent on concentrated data, and thus overly sensitive to the noise of these data, resulting in overfitting; therefore, in the embodiments of the present application, the precipitation data is first classified into precipitation scales through a pre-trained precipitation intensity classification sub-model, and then the precipitation data generated by the precipitation products are fused through the precipitation segment regression sub-model corresponding to different precipitation classifications, which can effectively improve the ability to capture precipitation events of different intensities and further improve the accuracy of the final fusion.

[0066] For example, the advantages of the technical solution of the present application can be significantly seen by comparing the data of the technical solution of the present application, the traditional random forest model technical solution and multiple satellite precipitation products; eight commonly used performance evaluation indicators are used for evaluation: the first category includes correlation coefficient (CC), Nash efficiency (NSE) and improved Kling-Gupta efficiency (KGE', Modified Kling-GuptaEfficiency), which are used to reflect the degree of consistency between precipitation products and station observations; the second category includes root mean square error (RMSE) and relative error (RBIAS, Relative Bias), which mainly reflect the overall degree of deviation between precipitation products and station observations; the third category includes detection rate (POD, Probability of Detection), false alarm rate (FAR, False Alarm Ratio) and critical success index (CSI, Critical Success Index), which are used to describe the ability of precipitation products to capture precipitation events.

[0067]

[0068] Where N is the total time step; and They refer to the station observation and satellite precipitation corresponding to time step t, respectively; and are the average values of site observation and satellite precipitation, respectively; σ s and σ oare the standard deviations of observed precipitation and satellite precipitation, respectively; H is the number of precipitation events captured simultaneously by the station and the satellite, that is, the number of hit events; M is the number of miss events; and F is the number of false alarm events, that is, the number of precipitation events captured by the precipitation product but not monitored by the ground station.

[0069] See also Figure 4 Figure 1 shows boxplots of the evaluation metrics for three precipitation products: PERSIANN, GSMaP, and IMERG, as well as the traditional random forest model RF and the proposed fusion model P-RF. During the training phase, the P-RF model improved CC, NSE, and KGE' to 0.94, 0.88, and 0.87, respectively. During the testing phase, the CC, NSE, and KGE' values of the fused precipitation obtained by the P-RF model far exceeded those of the original satellite precipitation and the RF fused precipitation, with medians of 0.88, 0.76, and 0.81, respectively, representing a significant improvement over the next-best product. Regarding the overall deviation index, the P-RF precipitation fusion product had the lowest RMSE value during both training and testing phases, while the median RBIAS value was the best during training. Although slightly higher than GSMaP in both temporal and spatial scales, the difference was minimal. In terms of precipitation event detection capability, while the median POD of the P-RF fused precipitation was lower than that of the RF fused precipitation during both training and testing phases, it significantly reduced the median FAR and achieved the highest CSI in both training and testing phases. In addition, except for POD, the boxes of the corresponding indicators of P-RF fusion precipitation are significantly shorter, indicating that the gap in its fusion performance at different stations is small and it has good stability in time and space.

[0070] See also Figure 5 , shows the performance of the traditional random forest model RF and the fusion model P-RF of this application at the watershed scale; the left side corresponds to the traditional RF model, and the right side corresponds to the P-RF model; it can be seen that the highest POD of the P-RF fused precipitation in the training stage is 0.87, and the highest in the test stage is 0.80, both from the Haihe River Basin. The lowest PODs are 0.80 (Pearl River Basin) and 0.74 (Yellow River Basin), which are acceptable in practical applications. In addition, the FAR and CSI of the P-RF fused precipitation are significantly better than those of the traditional RF. This shows that the P-RF precipitation fusion model has a certain degree of improvement in its ability to judge precipitation events in each watershed, effectively solving the problem of overestimation of the occurrence of precipitation events by the traditional RF model, which is consistent with the above-mentioned overall accuracy evaluation results.

[0071] See also Figure 6, showing the probability density distribution for different precipitation intensities. On the horizontal axis, values <0.1 and 0.1–10 correspond to the left y-axis, while values 10–25, 25–50, and ≥50 correspond to the right y-axis. The P-RF model clearly outperforms. In capturing non-precipitation events, the P-RF fused precipitation differs from site observations by 1.26% (overestimation), demonstrating excellent performance. It performs most closely to ground observations for light rain events, underestimating them by only 2.41%. For moderate rain events, although its performance is only slightly better than that of the RF fused precipitation due to the intensity range, the P-RF fused precipitation probability density is closest to site measurements within the 0.1–25 mm / day range. For heavy rain and rainstorm events, the P-RF fused precipitation performance also meets practical application requirements. Within the ≥25 mm / day intensity range, its probability density compared to site measurements is second only to IMERG. In summary, it can be seen that the P-RF model used in the embodiments of the present application is effective in improving the consistency between precipitation products and site observations, reducing bias, and enhancing the ability to capture precipitation events. It effectively improves the instability of the traditional RF model in the precipitation fusion process, such as overestimation of weak precipitation and underestimation of strong precipitation, and exhibits good performance at different scales and precipitation intensities.

[0072] In a possible implementation, generating the precipitation intensity classification sub-model and the precipitation segmentation regression sub-model includes:

[0073] Acquire grids where multiple precipitation monitoring stations in multiple river basins are located as sample grids, and generate measured precipitation data of the sample grids based on monitoring data of the precipitation monitoring stations;

[0074] Acquire precipitation data of a plurality of precipitation products of the sample grid and corresponding related grids as satellite precipitation data;

[0075] generating a database based on the measured precipitation data and the satellite precipitation data;

[0076] The precipitation intensity classification sub-model and the precipitation segmentation regression sub-model are trained according to the database.

[0077] In one possible implementation, obtaining the measured precipitation data includes:

[0078] The multiple basins are divided into semi-arid, semi-humid and humid areas, and a preset proportion of precipitation monitoring stations are selected from the semi-arid, semi-humid and humid areas respectively as selected stations by random sampling;

[0079] The grid where the selected site is located is used as the sample grid;

[0080] When there is only one selected site in the sample grid, the monitoring data of the selected site is used as the measured precipitation data of the sample grid;

[0081] When there are at least two selected sites in the sample grid, the average value of the monitoring data of all the selected sites in the sample grid is calculated as the measured precipitation data of the sample grid.

[0082] When implementing the embodiments of the present application, it is necessary to first construct the database required for model training. Since the distribution of rain stations is extremely uneven, interpolation processing of station data often leads to errors and uncertainties to varying degrees. Therefore, this article evaluates the original and fused precipitation products at the station scale, first selecting the corresponding station, and then determining the sample grid through the station. At the same time, the stations are partitioned according to the total precipitation during the flood season, and several stations are selected in each partition for fusion model construction. The division of semi-arid areas, semi-humid areas and humid areas described in the embodiments of the present application can be obtained through correction; specifically, it includes: calculating the average of the ratio of the total precipitation during the flood season to the total precipitation for the whole year for all precipitation monitoring stations as the correction ratio; multiplying the boundary value of the dry and wet area division corresponding to the annual precipitation by the correction ratio to form a new division boundary; dividing the semi-arid area, semi-humid area and humid area according to the new division boundary. The main reason for the correction is that the data obtained in the embodiment of the present application are all flood season data for flood season precipitation detection, while the precipitation data used in the statistical caliber is the total precipitation for the whole year. This will produce a more obvious error in data sampling, so in the embodiment of the present application, it is necessary to first correct the division boundary to ensure the accuracy of subsequent data adoption. For example, the average proportion of the total precipitation of all stations during the flood season relative to the whole year is calculated, which is 86%, and the existing division boundaries are arid areas: <200mm; semi-arid areas: 200-400mm; semi-humid areas: 400-800mm and humid areas: ≥800mm; the division boundaries after correction are arid areas: <172mm; semi-arid areas: 172-344mm; semi-humid areas: 344-688mm and humid areas: ≥688mm.

[0083] In an embodiment of the present application, for each sub-region of the semi-arid area, the semi-humid area and the humid area, a preset proportion of precipitation monitoring stations needs to be randomly selected as selected stations to ensure the diversity of the training samples, wherein the preset proportion can be 70% to 80%; the specific measured precipitation data can be the daily precipitation data of the national meteorological observation stations of the Meteorological Bureau; and the satellite precipitation data can be the PERSIANN-CCS daily precipitation of the CHRS official website, the daily precipitation of the GSMaP official website and the IMERG daily precipitation of the NASA website.

[0084] In one possible implementation, generating a database based on the measured precipitation data and the satellite precipitation data includes:

[0085] Acquire measured precipitation data of the sample grid along the time series, and arrange the measured precipitation data longitudinally along the time series to form an output matrix;

[0086] Obtain the satellite precipitation data of the sample grid and the corresponding related grids at the same time for all precipitation products, and arrange them horizontally to form a row vector;

[0087] Arrange the row vectors at multiple moments in the time sequence longitudinally along the time sequence, and align them with the time sequence of the output matrix to form an input matrix;

[0088] The database is formed according to the input matrix and the output matrix.

[0089] When implementing the embodiment of the present application, in order to realize the training of the subsequent model, it is necessary to integrate the data in the database accordingly. The measured precipitation data needs to be arranged longitudinally along the time series to form a column vector, which is the output matrix; and when integrating the precipitation data, since the embodiment of the present application also provides multiple related grids for the same sample grid, when these data are used as input data for random forest model training, each data can be provided to the model as a feature, thereby improving the feature dimension of the random forest model and further improving the accuracy of the final model operation. For an example, please refer to Figure 2 , a specific input matrix generation method is given, where the grid corresponding to the measured site is the sample grid, and the surrounding 8 grids are used as the correlation network. The same precipitation product generates nine precipitation data at a time. At this time, arranging these data horizontally can form a 9-dimensional row vector. In this example, three precipitation products are used, so the row vector formed after splicing is 27-dimensional. Figure 2 The interval corresponding to each moment in is generally selected as 24 hours. After arranging these row vectors vertically in time sequence, the input matrix can be formed.

[0090] In a possible implementation, generating the precipitation intensity classification sub-model includes:

[0091] Converting the measured precipitation data in the output matrix into precipitation classifications to form a classification output matrix; the precipitation classifications are categories divided according to precipitation intensity;

[0092] The input matrix is used as input data, and the classification output matrix is used as output data to train a random forest classification model to form the precipitation intensity classification sub-model.

[0093] When implementing the embodiments of this application, please refer to Figure 7, shows the training process of the precipitation intensity classification sub-model and the precipitation segmented regression sub-model. The purpose of the precipitation intensity classification sub-model is to more accurately describe the precipitation classification of the corresponding grid based on the data provided by the precipitation product. For example, precipitation events are divided into three categories: non-precipitation events (<0.1mm / d), light rain events (0.1-25mm / d), and heavy rain and above events (≥25mm / d). During training, the following formula can be used to evaluate the accuracy of the classification results:

[0094]

[0095] Where: n correct and n total =(N) represents the number of accurately classified samples and the total number of samples, respectively; N represents the total number of non-precipitation events; H represents the number of precipitation events accurately classified by both the site and the forecast; M represents the number of events where the forecast classification disagrees with the site classification; and F represents the number of precipitation events classified in the forecast but not observed by the ground station. Accurately classifying precipitation events allows the calculation of actual precipitation using models corresponding to different classifications.

[0096] In a possible implementation, generating the precipitation segmented regression sub-model includes:

[0097] Splitting the input matrix into a plurality of regression input matrices according to the precipitation classification, and splitting the output matrix into a plurality of regression output matrices according to the precipitation classification; each regression input matrix corresponds to a precipitation classification; each regression output matrix corresponds to a precipitation classification; the precipitation classification includes a trace precipitation classification; the trace precipitation classification is a classification with a precipitation amount of less than 0.1 mm / d;

[0098] The regression input matrix corresponding to the non-trace precipitation classification is used as input data, and the regression output matrix corresponding to the same precipitation classification is used as output data to train the random forest regression model to form multiple precipitation segmented regression sub-models corresponding to different precipitation classifications.

[0099] When the embodiment of the present application is implemented, based on the rainfall situation during the flood season, the impact of trace precipitation on natural disasters is relatively small. At the same time, affected by the accuracy of precipitation monitoring, the data of precipitation products related to trace precipitation and the data of ground monitoring stations have large errors. Model training using these data will cause the model deviation to be too large, so directly treating this part of the data as no precipitation can effectively improve the accuracy of the precipitation data outputted in the end. When the input matrix is split to form multiple regression input matrices, since the row vectors corresponding to each moment of the input matrix have been marked by precipitation classification, it is only necessary to extract the row vectors corresponding to the marks and splice them to form multiple regression input matrices. RF regression models are constructed for various precipitation intensity events, and the fusion results within different intensity ranges are merged and rearranged in chronological order to obtain the final precipitation fusion product. For example, when precipitation types are classified into three categories: non-precipitation events (<0.1 mm / d), light to moderate rain events (0.1 to 25 mm / d), and heavy rain events (≥25 mm / d), after excluding non-precipitation events (<0.1 mm / d), the precipitation intensity classification sub-model and the precipitation segmented regression sub-model are constructed by calling the Random Forest Classifier and Random Forest-Regressor libraries in Python. The hyperparameters of each sub-model, including n_estimators, max_depth, min_samples_split, and max_feature, are shown in the table below:

[0100]

[0101] Among them, "sqrt" means that the maximum number of features max_feature is the square root of the total number of features; "None" means setting the maximum depth of the decision tree max_depth to the maximum value, which is the default value of max_depth in the Random Forest Regressor library.

[0102] In a possible implementation, inputting the precipitation data into a precipitation segmented regression sub-model corresponding to the precipitation classification includes:

[0103] When the precipitation data corresponds to the precipitation classification of the trace precipitation classification, the fused precipitation amount corresponding to the precipitation data is defined as 0.

[0104] In a second aspect, the present application provides a multi-source precipitation product fusion system that considers the influence of autocorrelation information and spatial distribution, including:

[0105] an acquisition unit configured to acquire precipitation data of a plurality of precipitation products of a grid to be detected and grids related to the grid to be detected; the related grids are grids that share vertices and edges with the grid to be detected;

[0106] a classification unit configured to input the precipitation data into a preset precipitation intensity classification sub-model and obtain a precipitation classification output by the precipitation intensity classification sub-model;

[0107] The fusion unit is configured to input the precipitation data into a precipitation segmentation regression sub-model corresponding to the precipitation classification, and receive the data output by the precipitation segmentation regression sub-model as the fused precipitation amount of the grid to be detected.

[0108] In a possible implementation, the system further includes a database unit; the database unit is configured to:

[0109] Acquire grids where multiple precipitation monitoring stations in multiple river basins are located as sample grids, and generate measured precipitation data of the sample grids based on monitoring data of the precipitation monitoring stations;

[0110] Acquire precipitation data of a plurality of precipitation products of the sample grid and corresponding related grids as satellite precipitation data;

[0111] generating a database based on the measured precipitation data and the satellite precipitation data;

[0112] The precipitation intensity classification sub-model and the precipitation segmentation regression sub-model are trained according to the database.

[0113] In a possible implementation, the database unit is further configured to:

[0114] The multiple basins are divided into semi-arid, semi-humid and humid areas, and a preset proportion of precipitation monitoring stations are selected from the semi-arid, semi-humid and humid areas respectively as selected stations by random sampling;

[0115] The grid where the selected site is located is used as the sample grid;

[0116] When there is only one selected site in the sample grid, the monitoring data of the selected site is used as the measured precipitation data of the sample grid;

[0117] When there are at least two selected sites in the sample grid, the average value of the monitoring data of all the selected sites in the sample grid is calculated as the measured precipitation data of the sample grid.

[0118] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the above description has generally described the composition and steps of each example according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0119] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or units, or can be electrical, mechanical or other forms of connection.

[0120] The units described as separate components may or may not be physically separated. As units, it is obvious that a person of ordinary skill in the art can realize that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in terms of function in the above description. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0121] In addition, the functional units in the various embodiments of the present invention may be integrated into a single processing unit, each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0122] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or grid device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program code, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0123] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A multi-source precipitation product fusion method that considers the influence of autocorrelation information and spatial distribution, characterized by: include: Obtaining precipitation data of a plurality of precipitation products of a grid to be detected and grids related to the grid to be detected; The relevant grid is a grid that shares vertices and edges with the grid to be detected; Inputting the precipitation data into a preset precipitation intensity classification sub-model, and obtaining the precipitation classification output by the precipitation intensity classification sub-model; The precipitation data is input into a precipitation segmentation regression sub-model corresponding to the precipitation classification, and data output by the precipitation segmentation regression sub-model is received as the fused precipitation amount of the grid to be detected.

2. The multi-source precipitation product fusion method considering the influence of autocorrelation information and spatial distribution according to claim 1 is characterized in that: The generation of the precipitation intensity classification sub-model and the precipitation segmentation regression sub-model includes: Acquire grids where multiple precipitation monitoring stations in multiple river basins are located as sample grids, and generate measured precipitation data of the sample grids based on monitoring data of the precipitation monitoring stations; Acquire precipitation data of a plurality of precipitation products of the sample grid and corresponding related grids as satellite precipitation data; generating a database based on the measured precipitation data and the satellite precipitation data; The precipitation intensity classification sub-model and the precipitation segmentation regression sub-model are trained according to the database.

3. The multi-source precipitation product fusion method considering the influence of autocorrelation information and spatial distribution according to claim 2 is characterized in that: The acquisition of the measured precipitation data includes: The multiple basins are divided into semi-arid, semi-humid and humid areas, and a preset proportion of precipitation monitoring stations are selected from the semi-arid, semi-humid and humid areas respectively as selected stations by random sampling; The grid where the selected site is located is used as the sample grid; When there is only one selected site in the sample grid, the monitoring data of the selected site is used as the measured precipitation data of the sample grid; When there are at least two selected sites in the sample grid, the average value of the monitoring data of all the selected sites in the sample grid is calculated as the measured precipitation data of the sample grid.

4. The multi-source precipitation product fusion method considering the influence of autocorrelation information and spatial distribution according to claim 2 is characterized in that: Generating a database according to the measured precipitation data and the satellite precipitation data includes: Acquire measured precipitation data of the sample grid along the time series, and arrange the measured precipitation data longitudinally along the time series to form an output matrix; Obtain the satellite precipitation data of the sample grid and the corresponding related grids at the same time for all precipitation products, and arrange them horizontally to form a row vector; Arrange the row vectors at multiple moments in the time sequence longitudinally along the time sequence, and align them with the time sequence of the output matrix to form an input matrix; The database is formed according to the input matrix and the output matrix.

5. The multi-source precipitation product fusion method considering the influence of autocorrelation information and spatial distribution according to claim 4 is characterized in that: The generation of the precipitation intensity classification sub-model includes: Converting the measured precipitation data in the output matrix into precipitation classifications to form a classification output matrix; the precipitation classifications are categories divided according to precipitation intensity; The input matrix is used as input data, and the classification output matrix is used as output data to train a random forest classification model to form the precipitation intensity classification sub-model.

6. The multi-source precipitation product fusion method considering the influence of autocorrelation information and spatial distribution according to claim 5 is characterized in that: The generation of the precipitation segmented regression sub-model includes: Splitting the input matrix into a plurality of regression input matrices according to the precipitation classification, and splitting the output matrix into a plurality of regression output matrices according to the precipitation classification; each regression input matrix corresponds to a precipitation classification; each regression output matrix corresponds to a precipitation classification; the precipitation classification includes a trace precipitation classification; the trace precipitation classification is a classification with a precipitation amount of less than 0.1 mm / d; The regression input matrix corresponding to the non-trace precipitation classification is used as input data, and the regression output matrix corresponding to the same precipitation classification is used as output data to train the random forest regression model to form multiple precipitation segmented regression sub-models corresponding to different precipitation classifications.

7. The multi-source precipitation product fusion method considering the influence of autocorrelation information and spatial distribution according to claim 6 is characterized in that: Inputting the precipitation data into the precipitation segmentation regression sub-model corresponding to the precipitation classification includes: When the precipitation data corresponds to the precipitation classification of the trace precipitation classification, the fused precipitation amount corresponding to the precipitation data is defined as 0.

8. A multi-source precipitation product fusion system that considers the influence of autocorrelation information and spatial distribution using the method according to any one of claims 1 to 7, characterized in that: include: an acquiring unit configured to acquire precipitation data of a plurality of precipitation products of a grid to be detected and grids related to the grid to be detected; The relevant grid is a grid that shares vertices and edges with the grid to be detected; a classification unit configured to input the precipitation data into a preset precipitation intensity classification sub-model and obtain a precipitation classification output by the precipitation intensity classification sub-model; The fusion unit is configured to input the precipitation data into a precipitation segmentation regression sub-model corresponding to the precipitation classification, and receive the data output by the precipitation segmentation regression sub-model as the fused precipitation amount of the grid to be detected.

9. The multi-source precipitation product fusion system considering the influence of autocorrelation information and spatial distribution according to claim 8 is characterized in that: Also included is a database unit; the database unit is configured to: Acquire grids where multiple precipitation monitoring stations in multiple river basins are located as sample grids, and generate measured precipitation data of the sample grids based on monitoring data of the precipitation monitoring stations; Acquire precipitation data of a plurality of precipitation products of the sample grid and corresponding related grids as satellite precipitation data; generating a database based on the measured precipitation data and the satellite precipitation data; The precipitation intensity classification sub-model and the precipitation segmentation regression sub-model are trained according to the database.

10. The multi-source precipitation product fusion system considering the influence of autocorrelation information and spatial distribution according to claim 9 is characterized in that: The database unit is further configured to: The multiple basins are divided into semi-arid, semi-humid and humid areas, and a preset proportion of precipitation monitoring stations are selected from the semi-arid, semi-humid and humid areas respectively as selected stations by random sampling; The grid where the selected site is located is used as the sample grid; When there is only one selected site in the sample grid, the monitoring data of the selected site is used as the measured precipitation data of the sample grid; When there are at least two selected sites in the sample grid, the average value of the monitoring data of all the selected sites in the sample grid is calculated as the measured precipitation data of the sample grid.

Images