Multi-source rainfall data precision correction and downscaling method based on causal relationship constraint

Abstract

The invention provides a multi-source rainfall data precision correction and downscaling method based on a causal relationship constraint, and relates to the field of meteorology and hydrology, and the method comprises the steps: building a feature causal relationship structure tree for multi-source rainfall data, and adding the constraint of a feature causal relationship on the basis of an LSTM deep learning correction model; a CLSTM rainfall data precision correction model under the causal relationship constraint is established, a GRASPS rainfall data downscaling method is fused, a rainfall correction and downscaling model Casual-GRASPS of the causal relationship constraint is generated, and a high-precision, high-resolution and high-time-space continuous fusion rainfall data set conforming to the statistical causal relationship is generated. The method has the advantages that a reliable method is provided for downscaling of meteorological data such as rainfall, and a reliable data basis is provided for fine-scale rainfall research of drainage basins and urban areas.

Description

Technical Field

[0001] The present application relates to the field of meteorology and hydrology, and in particular to a method for precision correction and downscaling of multi-source precipitation data based on causal constraints. Background Art

[0002] Precipitation is an important component of the Earth's climate system, directly affecting agriculture, water resources, natural disasters, and other aspects. However, there are many problems in precipitation observation and prediction. For example, traditional observation methods have low spatial and temporal resolution and local sampling bias, which increase the uncertainty of precipitation data quality. Therefore, how to accurately and reliably obtain and analyze precipitation data has become a hot topic and difficulty in current climate change and extreme hydrological research.

[0003] Traditional precipitation observation methods include satellite measurement, ground radar measurement, meteorological data reanalysis and ground meteorological station monitoring, which have their own advantages and disadvantages. Satellite precipitation measurement has a wide range and high spatiotemporal continuity, but its accuracy is easily affected by factors such as climate region, seasonal changes, altitude and precipitation intensity; ground radar precipitation measurement accuracy is higher than satellite, but its measurement range is limited, and its accuracy is greatly affected by environmental and climatic conditions; meteorological reanalysis data has a wide spatial range and long time series, but there is a big difference with the real precipitation field; ground meteorological station data observation accuracy is the highest, but it can only measure single-point precipitation, and due to economic, topographic and other factors, the spatial distribution is uneven, and it is difficult to obtain spatially continuous precipitation data.

[0004] In order to combine the advantages of multi-source precipitation data and obtain more accurate, reliable and high-quality precipitation data, precipitation data correction and downscaling based on multi-source precipitation data fusion methods have become an important way to solve the problem of precipitation data quality. The current fusion methods can be divided into two categories, namely traditional mathematical methods and machine learning methods. Traditional mathematical methods mainly rely on interpolation and regression to fuse precipitation data. In recent years, deep learning methods in machine learning have also been widely used in data fusion. Since the current precipitation data has the characteristics of wide sources and easy to obtain, it is consistent with the requirement of deep learning that requires a large amount of data training. Therefore, combining multi-source precipitation data and fusing precipitation data through deep learning methods can better make up for the defects of various data, improve the quality and accuracy of precipitation data, and thus provide reliable data support for research in related fields. However, the internal structure and parameters of the deep learning model are similar to black boxes. The learning model is obtained by directly fitting a large amount of known data. Its internal mechanism is difficult to understand, the interaction between features is unclear, and the generalization performance of the model depends on a large amount of training data support.

[0005] Based on this, studying the statistical causal relationship between multi-source precipitation data and related factors, and exploring precipitation data correction and downscaling methods suitable for integrating different data sources can provide a more accurate and reliable data basis for precipitation research, and provide strong support and assistance for fields such as meteorological forecasting, natural disaster monitoring and water resources management. Summary of the invention

[0006] The purpose of the present invention is to provide a multi-source precipitation data accuracy correction and downscaling method based on causal constraints in order to solve the technical problems that traditional machine learning methods have internal mechanisms that are difficult to understand, interactions between features are unclear, and model generalization performance relies on the support of a large amount of training data.

[0007] The above-mentioned purpose of the present application is achieved through the following technical solutions: S1: Collect multi-source precipitation data and precipitation auxiliary data and pre-process them; multi-source precipitation data include: ground meteorological station precipitation data, satellite precipitation data and meteorological reanalysis data; S2: Construct a GCA-based characteristic causal relationship tree for each ground meteorological station in the study area; S3: Construct LSTM precipitation data correction model; S4: The LSTM precipitation data correction model is improved through the causal relationship information of the feature causal relationship structure tree to obtain the CLSTM precipitation data accuracy correction model; the CLSTM precipitation data accuracy correction model includes: the linear CLSTM-Linar precipitation data correction model and the nonlinear CLSTM-Nolinar precipitation data correction model; S5: According to the Thiessen polygon theory, the entire study area is divided into a preset number of sub-areas; each sub-area contains at least one ground meteorological station; the CLSTM precipitation data accuracy correction model in the sub-area is used to perform precipitation correction on the multi-source precipitation data of each sub-area; S6: Based on the GRASPS method, multi-source precipitation data are downscaled to obtain downscaled precipitation data; S7: Using the CLSTM-Nolinar precipitation data correction model and downscaled precipitation data, the precipitation correction and downscaling model Casual-GRASPS is constructed; S8: Through the precipitation correction and downscaling model Casual-GRASPS, a high-precision, high-resolution, and high-temporal and spatial continuity fused precipitation dataset is obtained.

[0008] An electronic device comprises a processor, a memory, a user interface and a network interface, wherein the memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory so that the electronic device executes a multi-source precipitation data accuracy correction and downscaling method based on causal constraints.

[0009] A computer-readable storage medium stores instructions. When the instructions are executed, a method for multi-source precipitation data accuracy correction and downscaling based on causal relationship constraints is executed.

[0010] The beneficial effects of the technical solution provided by this application are: A causal relationship tree between precipitation and its auxiliary data features is established, and a CLSTM precipitation data accuracy correction model corresponding to causal constraints is constructed. The GRASPS method is integrated to generate a causal constraint multi-source precipitation data accuracy correction and downscaling method Casual-GRASPS, which provides a reliable method for downscaling precipitation and other meteorological data. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] The present application will be further described below with reference to the accompanying drawings and embodiments, in which: Figure 1 is an implementation flow chart of the multi-source precipitation data accuracy correction and downscaling method based on causal relationship constraints in an embodiment of the present application; Figure 2 is a weather station location map in an embodiment of the present application; Figure 3 It is a graph of the Pearson correlation coefficient analysis results in the embodiment of the present application; Figure 4 is a comparison chart of the characteristic causal relationship analysis capabilities in the embodiments of the present application; Figure 5 is a comparative analysis diagram of multi-source precipitation data correction results in an embodiment of the present application; Figure 6 is a comparative analysis diagram of performance results in the embodiments of the present application; Figure 7 is a comparative analysis chart of the precipitation accuracy performance results in the embodiment of the present application and the original satellite precipitation data; Figure 8 is a comparison diagram of the spatial distribution effects in the embodiments of the present application; Fig. 9 It is a comparative analysis diagram of the Casual-GRASPS, GRASPS, LSTM precipitation fusion downscaling results and the original satellite precipitation and rain gauge measured precipitation in the four seasons in the embodiment of the present application; Fig.10It is a schematic diagram of the structure of an electronic device in an embodiment of the present application. DETAILED DESCRIPTION

[0012] In order to have a clearer understanding of the technical features, purposes and effects of the present application, the specific implementation methods of the present application are now described in detail with reference to the accompanying drawings.

[0013] The embodiments of the present application provide a method for precision correction and downscaling of multi-source precipitation data based on causal constraints.

[0014] Please refer to Figure 1 , Figure 1 : is a step diagram of a multi-source precipitation data accuracy correction and downscaling method based on causal relationship constraints in an embodiment of the present application, including: S1: Collect multi-source precipitation data and precipitation auxiliary data and pre-process them; multi-source precipitation data include: ground meteorological station precipitation data, satellite precipitation data and meteorological reanalysis data; Step S1 includes: The satellite precipitation data include: IMERG satellite precipitation data with a spatial resolution of 0.1°, CHIRPS satellite precipitation data with a spatial resolution of 0.05°, and PERSIANN-CDR satellite precipitation data with a spatial resolution of 0.25°; the temporal resolution of satellite precipitation data is daily, and the spatial resolutions are 0.1°, 0.05°, and 0.25° respectively; Specifically, the use of IMERG satellite precipitation data, CHIRPS satellite precipitation data and PERSIANN-CDR satellite precipitation data can retain their respective advantages to make up for the shortcomings of existing precipitation products.

[0015] The meteorological reanalysis data include: cloud base height, predicted albedo, surface pressure, two-meter temperature, total cloud cover, 100-meter horizontal wind speed and 100-meter vertical wind speed; the time resolution of the meteorological reanalysis data is hourly and the spatial resolution is 0.25°; Specifically, the meteorological reanalysis data include cloud base height CBH (Cloud Base Height), forecast albedo FAL (Forecast Albedo), surface pressure SP (Surface Pressure), two-meter temperature T2M (2m Temperature), total cloud cover TCC (Total Cloud Cover), 100-meter horizontal wind speed U100 (100m U-component of wind) and 100-meter vertical wind speed V100 (100m V-component of wind) from ERA5, with a temporal resolution of every hour and a spatial resolution of 0.25°. The precipitation auxiliary data include: normalized difference vegetation index, digital elevation model and land use data.

[0016] Specifically, the precipitation data of ground meteorological stations come from the precipitation values ​​measured by rain gauges at 81 meteorological stations that have passed quality control screening, with time intervals of 12 hours and 24 hours.

[0017] Step S1 also includes: The pre-processing comprises: Perform formatting, unit conversion, anomaly processing and missing value processing on precipitation data from ground meteorological stations; Unify the spatial resolution of satellite precipitation data to 0.1°; Unify the temporal and spatial resolutions of meteorological reanalysis data, and perform anomaly and missing value processing.

[0018] S2: Construct a GCA-based characteristic causal relationship tree for each ground meteorological station in the study area; Step S2 includes: S21: Linear and nonlinear Granger causality tests are performed on multi-source precipitation data and precipitation auxiliary data to obtain the causal relationship between multi-source precipitation data and precipitation auxiliary data; S22: The differences in precipitation characteristics were obtained by analyzing the Pearson correlation coefficient between multi-source precipitation data and precipitation auxiliary data; S23: Construct a characteristic causal relationship structure tree for each ground meteorological station based on the causal relationship and the differences in precipitation characteristics.

[0019] Specifically, linear and nonlinear Granger causality tests were used to conduct in-depth analysis of the causal relationship between multi-source precipitation data and precipitation auxiliary data from multiple angles; by analyzing the Pearson correlation coefficient between each input feature and precipitation, the differences in precipitation characteristics between different stations were revealed, and a characteristic causal relationship structure tree for each ground meteorological station was constructed. The characteristic causal relationship structure trees constructed by linear and nonlinear GCA (Granger ausality analysis) were analyzed and compared, indicating that the linear causal relationship structure tree is more stable.

[0020] S3: Construct LSTM precipitation data correction model; Specifically, precipitation is a long-term changing process, and the LSTM model uses three different gates to transfer long-term and short-term memory to solve the long-term dependency problem. It can preserve long-term information and is suitable for long-term series tasks such as precipitation.

[0021] S4: The LSTM precipitation data correction model is improved through the causal relationship information of the feature causal relationship structure tree to obtain the CLSTM precipitation data accuracy correction model; the CLSTM precipitation data accuracy correction model includes: the linear CLSTM-Linar precipitation data correction model and the nonlinear CLSTM-Nolinar precipitation data correction model; Specifically, the LSTM model precipitation correction is based on causal constraints. The linear and nonlinear GCA causal structures are added to the constructed LSTM precipitation data correction model to improve it, forming the linear and nonlinear CLSTM-Linar and CLSTM-Nolinar precipitation data correction models. For each node of the feature causal structure tree Introducing a causal state , used to construct the nodes in the structure tree based on the characteristic causal relationship The parent node of ,The causal state of the parent node is further applied to the fully connected layer to perform precipitation accuracy correction.

[0022] S5: According to the Thiessen polygon theory, the entire study area is divided into a preset number of sub-areas; each sub-area contains at least one ground meteorological station; the CLSTM precipitation data accuracy correction model in the sub-area is used to perform precipitation correction on the multi-source precipitation data of each sub-area; Specifically, the precipitation correction was performed for each sub-region using the model of the stations within the region, overcoming the problem that rain gauges can only measure precipitation at a single point and cannot seamlessly cover the entire study area.

[0023] S6: Based on the GRASPS method, multi-source precipitation data are downscaled to obtain downscaled precipitation data; Specifically, the GRASPS method is a method that combines point-surface fusion and deep learning to generate high-precision, high-temporal-resolution, and spatiotemporal-continuous soil moisture data. The GRASPS downscaling method integrates multi-source precipitation datasets, including remote sensing products, model-based products, and the diverse perception capabilities of ground measurements. Its effectiveness has been well verified in the downscaling of variables such as soil moisture and temperature. In addition, the method also adds three high-resolution data, DEM, land use, and NDVI, as downscaling auxiliary variables, further improving the effect of data downscaling.

[0024] S7: Using the CLSTM-Nolinar precipitation data correction model and downscaled precipitation data, the precipitation correction and downscaling model Casual-GRASPS is constructed; Step S7 includes: S71: Pixel segmentation of the output image of the CLSTM-Nolinar precipitation data correction model is performed using a nearest neighbor interpolation method; Specifically, we first need to divide the original data range into grids according to the target resolution, and determine the corresponding coordinates of each pixel of the new image in the original image according to the correspondence between the divided grids and the original data pixels. According to the distance between the coordinates of the new pixel and the original pixel, the value of the original pixel is assigned to all new pixels in the corresponding area. In this way, the resolution of CLSTM can be improved while retaining its accuracy.

[0025] S72: Based on the multivariate linear regression model, secondary data fusion is performed on the output image after pixel segmentation and multi-source precipitation data; S73: Construct the precipitation correction and downscaling model Casual-GRASPS by minimizing the residual sum of squares.

[0026] Specifically, CLSTM pixel segmentation based on nearest neighbor interpolation: nearest neighbor interpolation pixel segmentation is a commonly used method in image processing, especially in the process of image scaling, rotation or transformation. The basic idea of ​​this method is to find the pixel closest to the pixel position of the target image from the original image and assign its value to the corresponding pixel of the target image; secondary data fusion based on multiple linear regression: multiple linear regression is a statistical method used to analyze the relationship between multiple independent variables and a dependent variable. It is a generalization of simple linear regression, which can consider the influence of multiple independent variables on the dependent variable.

[0027] Specifically, downscaled data accuracy verification: when the data set is segmented, 8 ground stations are reserved and not involved in the training. Here, the data sets of these 8 stations are used to perform multi-spatiotemporal scale analysis on the secondary fused data. The generated fused precipitation data set is subsequently called the Casual-GRASPS data set, which is a high-precision, high-resolution, and high-spatiotemporal continuity fused precipitation data set.

[0028] Step S72 includes: The basic form of the multiple linear regression model is as follows:

[0029] in, represents the dependent variable, which is the precipitation after secondary fusion, represents the downscaled precipitation data generated by the GRASPS method, Represents high-precision precipitation correction data generated using the nonlinear CLSTM-Nolinar precipitation data correction model; , and is the regression coefficient, represents the error term.

[0030] S8: Through the precipitation correction and downscaling model Casual-GRASPS, a high-precision, high-resolution, and high-temporal and spatial continuity fused precipitation dataset is obtained.

[0031] Specifically, Figure 1 It is a flowchart for the implementation of the multi-source precipitation data accuracy correction and downscaling method based on causal relationship constraints; Figure 2 It is the location, elevation, land use classification and location of meteorological stations used in the study in Hubei Province; Figure 3 is the result of Pearson correlation coefficient analysis between input characteristics and precipitation at each station; Figure 4 It is a comparison of the feature causal relationship analysis capabilities of three models: random forest (RF), decision tree (DT) and support vector machine (SVM); Figure 5 It is a comparative analysis of the precipitation data correction results of the three models: LSTM, CLSTM-Linear and CLSTM-Nolinear; Figure 6 It is a comparative analysis of the performance results of the three models, LSTM, CLSTM-Linear, and CLSTM-Nolinear, at different spatial locations; Figure 7 It is a comparative analysis of the performance results of the three datasets, Casual-GRASPS, GRASPS and CLSTM, in terms of precipitation accuracy with the original satellite precipitation data; Figure 8 It is a comparison of the spatial distribution effects of the Casual-GRASPS, CLSTM, and GRASPS precipitation fusion downscaling results and the original satellite precipitation data during a typical precipitation period; Fig. 9 It is a comparative analysis of the Casual-GRASPS, GRASPS, LSTM precipitation fusion downscaling results and the original satellite precipitation and raingauge measured precipitation under four seasons.

[0032] The implementation flow chart of the multi-source precipitation data accuracy correction and downscaling method based on causal relationship constraints of the method of the present invention is as follows: Figure 1 As shown, the specific steps are as follows: Embodiment (1): Collection of basic data; In this embodiment, four types of precipitation-related data were collected: the precipitation data of ground meteorological stations came from the precipitation values ​​measured by the rain gauges at the ground meteorological stations of the National Meteorological Information Center of the China Meteorological Administration. Taking into account the integrity of the data set, this study selected daily-scale precipitation measurements from January 1, 2017 to December 31, 2019. The time intervals for obtaining rain gauge data are 12 hours and 24 hours. Precipitation data is obtained once every day at 8:00 and 20:00 Beijing time (corresponding to 0:00 and 12:00 UTC time), which is the sum of the precipitation at the observation points within 12 hours. The two are added together to obtain the total precipitation for 24 hours. The three precipitation data have all undergone quality control. After screening, a total of 81 meteorological stations in the study area can be used for this study. See the specific distribution of ground meteorological stations for details. Figure 2 , Table 1 is the detailed information: Table 1 Ground meteorological station information

[0033] The satellite precipitation data include three types of satellites: IMERG, CHIRPS and PERSIANN-CDR. The spatial resolution of MERG data is 0.1° and the temporal resolution is daily. The spatial resolution of CHIRPS satellite data is 0.05° and the temporal resolution is daily. The spatial resolution of PERSIANN-CDR data is 0.25° and the temporal resolution is daily. The time range is from January 1, 2017 to December 31, 2019. The meteorological reanalysis data include cloud base height CBH, predicted albedo FAL, surface pressure SP, two-meter temperature T2M, total cloud cover TCC, 100-meter horizontal wind speed U100 and 100-meter vertical wind speed V100 from ERA5. The data time range is from January 1, 2017 to December 31, 2019, the temporal resolution is hourly, and the spatial resolution is 0.25°. Precipitation auxiliary data include normalized difference vegetation index NDVI, digital elevation model DEM and land use data. NDVI data comes from the vegetation condition index MODIS Terra product MOD13Q1 of the Moderate Resolution Imaging Spectroradiometer (MODIS), which has a spatial resolution of 250 m and a temporal resolution of 16 days; in this embodiment, the 2019 DEM data is selected, which comes from the United States Geological Survey (USGS), with a spatial resolution of 30 meters; this embodiment uses the 2020 land use data, which comes from the National Basic Geographic Information Center, with a spatial resolution of 30 meters. The original land use data is divided into ten categories in detail, including cultivated land, forest land, grassland, shrub land, wetland, water body, tundra, artificial surface, bare land, glacier and permanent snow. In order to simplify the analysis process and highlight the main land use types, these categories were reclassified. After reclassification, the land use types are integrated into five main categories: cultivated land, forest land, grassland, water body and building. The area and proportion of different land types within the scope of this embodiment are shown in Table 2: Table 2 Proportion of land use types

[0034] Embodiment (2): Data preprocessing; In this embodiment, the meteorological station data is converted in format and unit, the TXT format is converted into the CSV format, and then the data is further converted into the SHP format according to the latitude and longitude information of each station. The latitude and longitude fields of the original station data are "degrees and minutes", and the latitude and longitude are converted into the floating point "degree" format. The original precipitation value unit is 0.1mm, which is converted into mm. Individual stations with serious data missing are eliminated, and for individual stations with missing data on a certain day, the linear interpolation method is used to supplement this part of the missing data. The spatial resolution of the three satellite precipitation data of CHIRPS, IMERG and PESIANN-CDR is resampled to 0.1° using bilinear interpolation. The ERA5 meteorological reanalysis data is divided into 24 frames according to UTC 0 o'clock every day, and the mean is calculated to obtain the daily resolution. The spatial resolution is resampled to 0.1° using bilinear interpolation, and the null value representing the "cloudless" situation is replaced with 0.

[0035] Example (3): Construction of causal structure of precipitation and its auxiliary data features based on causal relationship; In this embodiment, the statistical causal relationship between multi-source precipitation and its related data and the advantages of machine learning are utilized, and the pre-processed data obtained in embodiment (2) are combined to analyze the Pearson correlation coefficient between each input feature and the precipitation at each station, such as Figure 3 As shown, at sites 57251, 57475 and 57583, IMERG is obviously more relevant than other sites, and at sites 57435, 57445, 57447 ​​and 57543, U100 is more relevant, which reveals that there are differences in precipitation characteristics between different sites, and it is necessary to further construct a characteristic causal relationship structure tree for each site. Since the precipitation process is a highly nonlinear physical process, the interaction between various precipitation-related data does not follow a simple linear law, so the nonlinear causal structure tree is more stable. Therefore, three models, random forest (RF), decision tree (DT, Decision Tree) and support vector machine (SVM, Support Vector Machine), were selected to analyze and compare the causal structure tree constructed by nonlinear GCA. The characteristic causal relationship analysis capabilities of different models are compared. Figure 4 ,From the comparison of the three models, RF and DT are significantly ,outperformed SVM in the four feature analyses.

[0036] Embodiment (4): Precipitation correction based on LSTM model; In this embodiment, a precipitation data correction model is constructed based on the LSTM model. Since precipitation is a long-term changing process, the LSTM model uses three different gates to transmit long-term and short-term memory to solve the long-term dependency problem, can save long-term information, and is suitable for long-term series tasks such as precipitation.

[0037] Example (5): LSTM model precipitation correction based on causal constraints; In this embodiment, the LSTM precipitation data correction model of embodiment (4) is used as a basis, and is improved by adding linear and nonlinear GCA causal structures to form linear and nonlinear CLSTM-Linar and CLSTM-Nolinar precipitation data correction models. In order to introduce causal relationship information into LSTM, a new state is introduced for each node i, namely, the causal relationship state , to constrain the hidden state according to the parent node of node i in the causal structure (denoted as ). For fusion, the causal state of the root node (i.e., precipitation in this study) is further applied to the fully connected layer to perform precipitation accuracy correction. Since the causal state contains both temporal interdependence information and causal information, the CLSTM model is able to learn the temporal dependency and causal information of each node in the causal structure.

[0038] Embodiment (6): Regional precipitation correction based on station location; In this embodiment, based on the LSTM precipitation correction model obtained in Embodiment (4) and Embodiment (5) and the linear and nonlinear CLSTM-Linar and CLSTM-Nolinar precipitation correction models constrained by causality, the study area is divided into multiple regions according to the Thiessen polygon theory. Each region is calibrated for precipitation data based on three correction models trained at regional internal stations. The correction results of the three models are compared and analyzed based on the four evaluation indicators of PCC, RMSE, MAE and Bias, and a line graph is drawn for comparison ( Figure 5 ), the specific results are shown in Table 3: Table 3 Detailed information of model indicators

[0039] Compared with LSTM and CLSTM-Linear, CLSTM-Nolinear has better performance. First, from the perspective of PCC, the average value of CLSTM-Nolinear reached 0.55, far exceeding LSTM (0.43) and CLSTM-Linear (0.26). This shows that CLSTM-Nolinear fits the real precipitation more significantly and has a stronger linear relationship. Among the three error indicators, in terms of RMSE, the average value of CLSTM-Nolinear is only 7.47 mm, which is significantly lower than LSTM (8.27 mm) and CLSTM-Linear (8.76 mm). Further observing MAE, CLSTM-Nolinear presents the smallest mean absolute error, which is only 3.47 mm. In comparison, LSTM and CLSTM-Linear are 4.28 mm and 4.72 mm, respectively. Finally, in terms of Bias, the average bias of CLSTM-Nolinear is close to zero (-0.05 mm), and the correction bias of CLSTM-Nolinear is lower than that of LSTM (0.28 mm) and CLSTM-Linear (0.95 mm). This shows that CLSTM-Nolinear has a smaller correction error and is more accurate in correcting real precipitation data. It is worth noting that CLSTM-Linear performs poorly in all indicators, indicating that the linear causal analysis of the input features is difficult to capture the causal relationship between features, and the resulting causal structure is too simple, losing a lot of useful information, so that its performance is even lower than that of LSTM. The characteristic of LSTM that directly inputs all features can capture more precipitation information than CLSTM-Linear. In general, the CLSTM-Nolinear model obtained by adding nonlinear causal constraints to the LSTM structure can more accurately capture various types of precipitation data and the potential connection between related data, and its correction results are closer to actual precipitation.

[0040] Considering that this embodiment covers a wide area, the performance of the model in different spatial locations varies. In order to further explore the performance changes of the model in different spatial locations, the spatial distribution of the four evaluation indicators is further analyzed in detail. The results are as follows: Figure 6As shown. In the comparison of model performance at different locations, CLSTM showed the best performance. It is worth noting that among all the stations, 56 stations have CLSTM-Nolinear Bias between -1.00 ~ -0.50mm, indicating that the model slightly underestimates the actual precipitation. However, compared with the other two models, CLSTM-Nolinear is still the best model. In general, the CLSTM-Nolinear model still shows reliable precipitation correction performance under the influence of different spatial factors such as elevation and land type.

[0041] Example (7): Downscaling of precipitation data based on GRASPS; In this embodiment, the GRASPS method can complement the advantages of multi-source precipitation data, obtain the characteristics of high-resolution meteorological data, and use the GRASPS method to downscale the precipitation data. Combined with the CLSTM precipitation correction data obtained in Example (4) and Example (5), on the basis of the CLSTM, the GRASPS method also adds three high-resolution data of DEM, land use and NDVI as downscaling auxiliary variables. In terms of spatial scale, in order to meet the requirements of high spatial resolution of precipitation data, all auxiliary variables are resampled to 1 km. Then, the RF algorithm is used to perform pixel-by-pixel GRASPS fusion on the multi-source precipitation data products, and the average result algorithm is used to calculate the average value of all RF decision tree fusion results to obtain the final fusion result. Finally, under the GRASPS downscaling framework, a precipitation downscaling dataset with a spatial resolution of 1 km is generated.

[0042] Example (8): Downscaling of Casual-GRASPS precipitation data based on causal secondary fusion; In this embodiment, the CLSTM-Nolinear precipitation data correction model of embodiment (5) and the GRASPS precipitation data downscaling method of embodiment (7) are combined, and the optimal model that has been trained multiple times is used to fuse the GRASPS data and the CLSTM data to form the precipitation correction and downscaling model Casual-GRASPS. The four evaluation indicators of PCC, RMSE, MAE, and Bias are used to compare the performance of the three data sets of Casual-GRASPS, GRASPS, and CLSTM in terms of precipitation accuracy. Figure 7As shown. Compared with satellite precipitation data, GRASPS also has a significant improvement in accuracy. Compared with the three satellite precipitation data, the PCC of GRASPS increased by 65.71% to 93.33%. Secondly, as a correction model, the CLSTM model has higher data accuracy, and its PCC is 24.14% higher than that of GRASPS. From the three error indicators of RMSE and MAE, they were reduced by 17.75% and 39.11% respectively. Overall, GRASPS and CLSTM have overestimation and underestimation of precipitation. The Bias of GRASPS is 0.32 mm, indicating that it tends to overestimate precipitation, while the Bias of CLSTM is -0.11, and its underestimation is significantly less than that of GRASPS. These results fully demonstrate the excellent performance of CLSTM in improving the accuracy of precipitation data. Finally, Casual-GRASPS is a combination of the two models, CLSTM and GRASPS. Compared with the previous two models, Casual-GRASPS combines the advantages of CLSTM and GRASPS, and has a high spatial resolution of 1 km while greatly improving the data accuracy. Compared with CLSTM, the PCC of Casual-GRASPS is improved by about 5.56%. Since Casual-GRASPS is a fusion of CLSTM and GRASPS, while obtaining GRASPS high-resolution data, for those specific cases where GRASPS accuracy is better than CLSTM, Casual-GRASPS achieves a slight but further improvement. In short, by comparing and analyzing the precipitation accuracy of the three datasets of Casual-GRASPS, GRASPS and CLSTM, their advantages and characteristics in their respective fields are obtained. As a fusion of the two, Casual-GRASPS not only has a fine spatial resolution of 1 km, but also has a further improvement in accuracy.

[0043] In order to verify the downscaling effect of Casual-GRASPS, a typical precipitation period was selected and the downscaling results of the day were evaluated, such as Figure 8 shown. Figure 8 (a) Casual-GRASPS and Figure 8 (b) The spatial resolution of GRASPS is 1 km, the highest among the six figures. Figure 8 (c) CLSTM is based on IMERG, and its spatial resolution and Figure 8 (d) The IMERG is the same, both are 0.1°. Figure 8 (e) CHIRPS is the product with the highest resolution among the three satellite precipitation data, which can reach 0.05°. Figure 8 (f) PERSIANN-CDR is the product with the lowest resolution among the three satellite precipitation data, with a resolution of 0.25°.

[0044] In terms of resolution, Casual-GRASPS and GRASPS reached the highest resolution of 1 km, showing obvious advantages in the display of spatial details. In terms of the spatial distribution of precipitation, Casual-GRASPS, GRASPS and CLSTM all showed similar trends. In the precipitation distribution of the entire study area, they all showed a change pattern of first decreasing, then increasing and then decreasing from west to east. In particular, Casual-GRASPS and CLSTM are almost consistent in precipitation distribution. This is because the accuracy of Casual-GRASPS benefits from CLSTM, so the performance in precipitation correction is the most similar. At the same time, the three satellite precipitation data were also evaluated. The results show that IMERG underestimates the true precipitation in most areas, while CHIRPS and PERSIANN-CDR underestimate the true precipitation in the northwest of the study area and overestimate the true precipitation in the eastern region. In contrast, Casual-GRASPS effectively corrected these errors and more accurately reflected the actual precipitation situation. It is worth noting that while GRASPS corrected the satellite precipitation errors, it also successfully integrated the areas where satellite data monitoring was accurate. For example, in the southwest and central parts of China, GRASPS incorporates the high precipitation trends of CHIRPS and PERSIANN-CDR; in the north and northeast, it incorporates the high precipitation trends of IMERG. This fusion strategy allows GRASPS to improve the accuracy of precipitation data while maintaining high spatial resolution. Overall, Casual-GRASPS successfully combines the correct precipitation parts of the three satellite precipitation data and corrects the erroneous parts, thereby greatly improving the spatial resolution and significantly improving the accuracy of precipitation data.

[0045] The occurrence of precipitation has obvious seasonal characteristics, and the downscaling accuracy is closely related to the intensity of precipitation. Therefore, the performance of the fusion model may be different in different seasons. In order to further analyze the performance of the fusion model in different seasons, the data set is divided into four seasons: spring, summer, autumn and winter for comparative analysis. The seasonal division is based on the following standards: spring: March, April, and May, summer: June, July, and August, autumn: September, October, and November, and winter: December, January, and February. Here, 8 test sites are selected and the Taylor diagram is used for comparative analysis. The results are as follows: Fig. 9As shown in the figure. From the perspective of season, among the eight stations, all precipitation data show a trend that the observed precipitation values ​​in winter are closest to the origin. This is mainly because the study area has less precipitation in winter and the average daily precipitation value is low, so the range of precipitation values ​​is relatively small. On the contrary, in summer, except for station 57265, the precipitation data of the other seven stations are farthest from the origin. In summer, precipitation is concentrated, and the precipitation values ​​are generally high, but there are also cases of no precipitation, which leads to an increase in the range between the upper and lower limits of precipitation. Although the fusion effect of Casual-GRASPS in autumn is relatively poor, its PCC reaches a level close to 0.7 in all stations. This shows that despite certain errors, Casual-GRASPS still maintains a relatively high accuracy in downscaling precipitation data in autumn. From the perspective of stations, Casual-GRASPS has the best precipitation correction and downscaling results among all stations. In particular, station 57494 has the most significant effect, with PCC close to 0.9 in all four seasons and even reaching 0.91 in winter. This shows the stability of Casual-GRASPS in different seasons and highlights its excellent performance at specific sites. In general, Casual-GRASPS has achieved excellent correction and downscaling results in different seasons and at different sites, showing its superior performance in both spatial and temporal performance.

[0046] The Casual-GRASPS precipitation accuracy correction and downscaling model of the present invention are compared with the existing precipitation accuracy correction research and precipitation downscaling research, and the results are shown in Table 4. From the perspective of spatial resolution, the spatial resolution of existing precision correction methods is mostly between 0.05° and 0.5°, while existing downscaling methods and the Casual-GRASPS in this paper can downscale the data to 1 km; from the perspective of temporal resolution, existing downscaling methods are mainly concentrated on the monthly and annual scales, while existing precision correction methods and the Casual-GRASPS in this paper can reach the daily scale; from the perspective of the number of selected data sources, the Casual-GRASPS in this paper selected three types of satellite precipitation data for fusion, realizing complementary advantages, while most existing downscaling methods only correct one type of data; from the perspective of feature selection methods, most existing precipitation precision correction and downscaling studies use the method of direct feature input. This study analyzes the causal relationship between features, constructs a feature causal structure tree, and realizes feature input that conforms to the causal relationship; from the perspective of model generalization, the Casual-GRASPS method in this paper can accurately capture heavy precipitation, with the highest PCC reaching 0.91, while most existing studies have not conducted special research on heavy water; from the perspective of considered factors, Casual-GRASPS takes into account the differences in feature importance caused by different spatial locations of stations, thereby establishing a fusion model for each station. In general, compared with previous studies, the Casual-GRASPS proposed in this paper has made significant improvements in precipitation accuracy correction and downscaling, and innovatively introduced a characteristic causal structure into the model, which improved the generalization and interpretability of the model.

[0047] Table 4 Comparison of precipitation data accuracy correction and downscaling methods

[0048] The present application also discloses an electronic device. Fig.10 , Fig.10 The electronic device 500 may include: at least one processor 501 , at least one network interface 504 , a user interface 503 , a memory 505 , and at least one communication bus 502 .

[0049] The communication bus 502 is used to realize the connection and communication between these components.

[0050] The user interface 503 may include a display screen, and the optional user interface 503 may also include a standard wired interface or a wireless interface.

[0051] The network interface 504 may optionally include a standard wired interface or a wireless interface (such as a WI-FI interface).

[0052] The present application also discloses a computer-readable storage medium storing a plurality of instructions, which are suitable for a processor to load to execute the above-mentioned multi-source precipitation data accuracy correction and downscaling method based on causal relationship constraints.

[0053] The above are only exemplary embodiments of the present disclosure and cannot be used to limit the scope of the present disclosure. That is, any equivalent changes and modifications made according to the teachings of the present disclosure are still within the scope of the present disclosure. After considering the disclosure of the specification and the truth of practice, those skilled in the art will easily think of other embodiments of the present disclosure.

[0054] This application is intended to cover any variation, use or adaptation of the present disclosure, which follows the general principles of the present disclosure and includes common knowledge or customary technical means in the art not described in the present disclosure. The description and examples are to be regarded as exemplary only, and the scope and spirit of the present disclosure are defined by the claims.

Claims

1. A method for precision correction and downscaling of multi-source precipitation data based on causal constraints, characterized in that: The method comprises the following steps: S1: Collect multi-source precipitation data and precipitation auxiliary data and pre-process them; multi-source precipitation data include: ground meteorological station precipitation data, satellite precipitation data and meteorological reanalysis data; S2: Construct a GCA-based characteristic causal relationship tree for each ground meteorological station in the study area; S3: Construct LSTM precipitation data correction model; S4: The LSTM precipitation data correction model is improved through the causal relationship information of the feature causal relationship structure tree to obtain the CLSTM precipitation data accuracy correction model; the CLSTM precipitation data accuracy correction model includes: the linear CLSTM-Linar precipitation data correction model and the nonlinear CLSTM-Nolinar precipitation data correction model; S5: According to the Thiessen polygon theory, the entire study area is divided into a preset number of sub-areas; each sub-area contains at least one ground meteorological station; the CLSTM precipitation data accuracy correction model in the sub-area is used to perform precipitation correction on the multi-source precipitation data of each sub-area; S6: Based on the GRASPS method, multi-source precipitation data are downscaled to obtain downscaled precipitation data; S7: Using the CLSTM-Nolinar precipitation data correction model and downscaled precipitation data, the precipitation correction and downscaling model Casual-GRASPS is constructed; S8: Through the precipitation correction and downscaling model Casual-GRASPS, a high-precision, high-resolution, and high-temporal and spatial continuity fused precipitation dataset is obtained.

2. The method for multi-source precipitation data accuracy correction and downscaling based on causal relationship constraints according to claim 1, characterized in that: Step S1 includes: The satellite precipitation data include: IMERG satellite precipitation data with a spatial resolution of 0.1°, CHIRPS satellite precipitation data with a spatial resolution of 0.05°, and PERSIANN-CDR satellite precipitation data with a spatial resolution of 0.25°; the temporal resolution of satellite precipitation data is daily, and the spatial resolutions are 0.1°, 0.05°, and 0.25° respectively; The meteorological reanalysis data include: cloud base height, predicted albedo, surface pressure, two-meter temperature, total cloud cover, 100-meter horizontal wind speed and 100-meter vertical wind speed; the time resolution of the meteorological reanalysis data is hourly and the spatial resolution is 0.25°; The precipitation auxiliary data include: normalized difference vegetation index, digital elevation model and land use data.

3. The method for multi-source precipitation data accuracy correction and downscaling based on causal relationship constraints according to claim 1, characterized in that: Step S1 also includes: The pre-processing comprises: Perform formatting, unit conversion, anomaly processing and missing value processing on precipitation data from ground meteorological stations; Unify the spatial resolution of satellite precipitation data to 0.1°; Unify the temporal and spatial resolutions of meteorological reanalysis data, and perform anomaly and missing value processing.

4. The method for multi-source precipitation data accuracy correction and downscaling based on causal relationship constraints according to claim 1, characterized in that: Step S2 includes: S21: Linear and nonlinear Granger causality tests are performed on multi-source precipitation data and precipitation auxiliary data to obtain the causal relationship between multi-source precipitation data and precipitation auxiliary data; S22: The differences in precipitation characteristics were obtained by analyzing the Pearson correlation coefficient between multi-source precipitation data and precipitation auxiliary data; S23: Construct a characteristic causal relationship structure tree for each ground meteorological station based on the causal relationship and the differences in precipitation characteristics.

5. The method for multi-source precipitation data accuracy correction and downscaling based on causal relationship constraints according to claim 1, characterized in that: Step S7 includes: S71: Pixel segmentation of the output image of the CLSTM-Nolinar precipitation data correction model is performed using a nearest neighbor interpolation method; S72: Based on the multivariate linear regression model, secondary data fusion is performed on the output image after pixel segmentation and multi-source precipitation data; S73: Construct the precipitation correction and downscaling model Casual-GRASPS by minimizing the residual sum of squares.

6. The method for multi-source precipitation data accuracy correction and downscaling based on causal relationship constraints according to claim 5, characterized in that: Step S72 includes: The basic form of the multiple linear regression model is as follows: in, represents the dependent variable, which is the precipitation after secondary fusion, represents the downscaled precipitation data generated by the GRASPS method, Represents high-precision precipitation correction data generated using the nonlinear CLSTM-Nolinar precipitation data correction model; , and is the regression coefficient, represents the error term.

7. An electronic device, characterized in that: The electronic device comprises a processor (501), a memory (505), a user interface (503) and a network interface (504), wherein the memory (505) is used to store instructions, the user interface (503) and the network interface (504) are used to communicate with other devices, and the processor (501) is used to execute the instructions stored in the memory (505) so that the electronic device executes the method according to any one of claims 1 to 6.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores instructions, and when the instructions are executed by a computer, the method according to any one of claims 1 to 6 is executed.

Images