Machine learning multi-source rainfall data fusion method for mining multi-dimensional space information

Abstract

The invention relates to a machine learning multi-source rainfall data fusion method for mining multi-dimensional space information. Comprising the following steps: acquiring actually measured daily rainfall data, multi-source grid rainfall data, meteorological information, geographic information and a normalized vegetation index of an observation station in a target area; screening an environment prediction factor based on a geographic detector, and extracting a sample environment prediction factor; extracting multi-dimensional grid rainfall data by taking an observation station as a center, and performing normalization processing on the data and the sample environment prediction factors to obtain a sample data set; training a convolutional neural network by using the data set to obtain a precipitation event classification model and a precipitation prediction model, and obtaining multi-dimensional spatial features in prediction factors; and predicting a grid precipitation event and a wet day grid precipitation amount through the model to obtain a multi-source precipitation data fusion data set. Therefore, the problem that multi-dimensional space information is not sufficiently considered in the prior art is solved, and effective information in various environment variables and multi-dimensional data in the multi-source rainfall data fusion process is fully mined.

Description

Technical Field

[0001] This application relates to the field of data fusion technology, and in particular to a machine learning method for fusing multi-source precipitation data to mine multi-dimensional spatial information. Background Technology

[0002] Precipitation is a crucial component of the water cycle and a key link between atmospheric and hydrological processes. High spatiotemporal resolution, long-term precipitation data are indispensable for hydrological simulation and research on the hydrological response to climate change. Due to the influence of climate and topography, precipitation observation stations are often spatially unevenly distributed. In China, this is mainly manifested in a high density of stations in the east and a sparse distribution in the northwest and the Qinghai-Tibet Plateau region, making it difficult for station observation data to reflect the true spatiotemporal distribution characteristics of precipitation.

[0003] Satellite remote sensing precipitation data, providing high spatiotemporal resolution and continuous real-time precipitation information, has been widely used. However, the accuracy of satellite remote sensing data is susceptible to factors such as inversion algorithms and sensor limitations. Especially in high-altitude and cold regions, where solid precipitation accounts for a large proportion of total precipitation, infrared remote sensing technology struggles to detect it, often leading to an underestimation of precipitation by satellite data. Global reanalysis data offers another effective alternative to station-based observation data, providing long-term, multivariate information. Through data assimilation techniques, global reanalysis data integrates information from multiple sources, such as ground observations and satellite remote sensing data, to improve accuracy. However, global reanalysis data typically has a coarse resolution, making it difficult to reflect local or station-based precipitation characteristics. Regional reanalysis data, using global reanalysis data as initial and lateral boundary conditions and generated through dynamic downscaling using regional climate models, offers the advantage of high spatiotemporal resolution, providing more detailed spatial information and capturing, to some extent, the topographic effects of precipitation.

[0004] In general, precipitation data from different sources each have their own advantages and disadvantages. A single precipitation product cannot accurately and comprehensively reflect the spatiotemporal distribution characteristics of precipitation. Data fusion, the process of integrating multi-source precipitation data with environmental information, can absorb the advantages of various precipitation data and is an effective way to reduce the uncertainty of multi-source precipitation data. Currently, domestic and international scholars mainly use statistical methods, such as bias correction, spatial interpolation, and regression methods, along with machine learning methods for data fusion. Statistical methods have the advantages of being computationally simple and easy to interpret, but they also have certain limitations. For example, bias correction and regression methods are based on mathematical equations and certain assumptions, which are not always correct in reality; the effectiveness of spatial interpolation methods depends on station density. Interpolating sparse and unevenly distributed station observation data from high-altitude and cold regions into a grid makes it difficult to comprehensively characterize the spatiotemporal distribution characteristics of precipitation and may introduce significant uncertainty. Compared with statistical methods, machine learning methods have advantages such as strong generalization ability, effective handling of complex nonlinear relationships, capture of multi-layer features in high-dimensional data, and no need for strict statistical assumptions, and have been widely used in data fusion.

[0005] However, existing machine learning-based precipitation data fusion methods often lack the mining of multidimensional spatial proximity information that has an important impact on precipitation. This is mainly reflected in two aspects: (1) The input of the machine learning model lacks a description of multidimensional spatial information around the target location. Specifically, during the model training process, only different sources of precipitation at the location of the station are extracted to construct the training set and establish the fusion relationship between station observations and precipitation from different sources. However, the influence of spatial proximity information on precipitation is not considered. The fusion is carried out independently on each grid cell, which results in precipitation in space being discrete rather than continuous. In fact, precipitation not only has geographical (e.g., latitude and longitude, altitude, etc.) dependence, but is also affected by other meteorological factors (e.g., dew point temperature, wind speed, etc.). It is continuous in space, and the precipitation at the target location is also affected by the surrounding grid. (2) The fusion source data used has poor representativeness for local-scale precipitation. Previous studies have focused on fusion of satellite remote sensing data and station observation data. The data sources used are not representative enough for topographic precipitation and seasonal precipitation. Summary of the Invention

[0006] This application provides a machine learning method for fusing multi-source precipitation data to mine multi-dimensional spatial information, in order to solve the problem of insufficient consideration of multi-dimensional spatial information in related technologies, and to fully mine the effective information in various environmental variables and multi-dimensional data during the fusion of multi-source precipitation data.

[0007] To achieve the above objectives, the first aspect of this application proposes a machine learning method for fusing multi-source precipitation data to mine multi-dimensional spatial information, comprising the following steps: Acquire measured daily precipitation data, multi-source grid precipitation data, meteorological information, geographic information, and normalized vegetation index from observation stations within the target area; Based on a preset geographic detector, the explanatory power of the meteorological information, the geographic information, and the normalized vegetation index on precipitation is quantified, and target environmental prediction factors are screened based on the quantification results. The multi-source grid precipitation data is preprocessed, and grid precipitation data centered on the observation station or prediction grid and sample environmental prediction factors are extracted from the preprocessed multi-source grid precipitation data and the target environmental prediction factors. The grid precipitation data and the sample environmental prediction factors are then normalized to obtain a sample dataset. Convolutional neural networks are constructed, and the sample dataset is divided into a first dataset and a second dataset. The convolutional neural network is trained using the first dataset and the second dataset respectively to obtain a precipitation event classification model and a precipitation prediction model. The precipitation event classification model predicts grid precipitation events, and the precipitation amount prediction model predicts wet day grid precipitation. Based on the predicted grid precipitation events and the wet day grid precipitation, a multi-source precipitation data fusion dataset is obtained.

[0008] According to one embodiment of this application, the step of quantifying the explanatory power of the meteorological information, the geographical information, and the normalized vegetation index on precipitation based on a preset geographic detector, and screening target environmental prediction factors based on the quantification results, includes: Based on the meteorological information, the geographical information, and the normalized vegetation index, initial environmental prediction factors are determined, and the initial environmental prediction factors are discretized and graded to obtain discretized environmental prediction factors. Based on the preset formula for calculating the degree of precipitation interpretation, the degree of interpretation of the discrete environmental prediction factors on precipitation is calculated, and qualified environmental prediction factors that meet the preset conditions are selected according to the calculation results. The precipitation explanation degree of each pair of qualified environmental predictors is calculated, and the target environmental predictors are obtained based on the precipitation explanation degree of each pair of factors and the precipitation explanation degree of each factor in the qualified environmental predictors.

[0009] According to one embodiment of this application, determining the initial environmental prediction factors based on the meteorological information, the geographical information, and the normalized vegetation index includes: Based on the meteorological information, the geographical information, and the normalized vegetation index, the spatial environment prediction factors of the observation station are determined according to a preset interpolation strategy. The spatial environment prediction factors are unified into a daily scale according to a preset time scale, and the seasons in which the station observes precipitation are located are divided to obtain the temporal environment prediction factors. The spatial environment prediction factors and the temporal environment prediction factors are then fused to obtain the initial environment prediction factors.

[0010] According to one embodiment of this application, the preset formula for calculating the degree of precipitation interpretation is as follows: ; in, L For the number of categories of environmental predictors, N h and N These represent the number of units for the h-th type of environmental predictor and the number of units for the global environmental factor, respectively. σ 2 and σ These represent the precipitation variance of the h-th type environmental predictor and the precipitation variance of the global environmental predictor, respectively.

[0011] According to one embodiment of this application, training the convolutional neural network using the first dataset to obtain a precipitation event classification model includes: Based on a preset partitioning ratio, the first dataset is divided into a first training set and a first test set, and the first validation set is determined from the first training set using the K-fold cross-validation method. The first training set is input into the convolutional neural network for training to obtain the first initial model parameters; Based on the first initial model parameters, the first validation set is input into the convolutional neural network for performance evaluation, and the first model parameters are determined based on the first performance evaluation result. Based on the first model parameters, the first test set is input into the convolutional neural network for model testing, and the precipitation event classification model is obtained when the first test result meets the preset requirements.

[0012] According to one embodiment of this application, training the convolutional neural network using the second dataset to obtain a precipitation prediction model includes: Based on a preset partitioning ratio, the second dataset is divided into a second training set and a second test set, and the second validation set is determined from the second training set using the K-fold cross-validation method. The second training set is input into the convolutional neural network to obtain the second initial model parameters; Based on the second initial model parameters, the second validation set is input into the convolutional neural network for performance evaluation, and the second target model parameters are determined based on the second performance evaluation results; Based on the second model parameters, the second test set is input into the convolutional neural network for model testing, and the precipitation prediction model is obtained when the second test result meets the preset requirements.

[0013] According to one embodiment of this application, obtaining a multi-source precipitation data fusion dataset based on the predicted grid precipitation events and the wet day grid precipitation includes: Determine whether the predicted precipitation event in the grid is a rainless day; If the predicted precipitation event is a rainless day, the daily precipitation is recorded as 0; otherwise, the daily precipitation is recorded as the wet day precipitation. A multi-source precipitation data fusion dataset was obtained based on multiple daily precipitation data.

[0014] The machine learning-based multi-source precipitation data fusion method proposed in this application, which mines multi-dimensional spatial information, acquires measured daily precipitation, multi-source grid precipitation, meteorological, geographical, and normalized vegetation indices. It uses a geographic detector to determine and extract sample environmental prediction factors, preprocesses multi-source grid precipitation data, extracts grid precipitation data and sample environmental prediction factors centered on observation stations, normalizes the data to obtain a sample dataset, and uses it to train a convolutional neural network to obtain a precipitation event classification model and a precipitation prediction model. The model then predicts grid precipitation events and wet day precipitation, forming a multi-source precipitation data fusion dataset. This method solves the problem of insufficient consideration of multi-dimensional spatial information in related technologies, fully mining the effective information from various environmental variables and multi-dimensional data during the multi-source precipitation data fusion process.

[0015] To achieve the above objectives, a second aspect of this application proposes a machine learning multi-source precipitation data fusion device for mining multi-dimensional spatial information, comprising: The acquisition module acquires measured daily precipitation data, multi-source grid precipitation data, meteorological information, geographic information, and normalized vegetation index from observation stations within the target area. The determination module, based on a preset geographic detector, quantifies the explanatory power of the meteorological information, the geographic information, and the normalized vegetation index on precipitation, and selects target environmental prediction factors based on the quantification results. The processing module preprocesses the multi-source grid precipitation data, extracts grid precipitation data centered on the observation station or prediction grid and sample environment prediction factors from the preprocessed multi-source grid precipitation data and the target environment prediction factors, and normalizes the grid precipitation data and the sample environment prediction factors to obtain a sample dataset. The module constructs a convolutional neural network and divides the sample dataset into a first dataset and a second dataset. The convolutional neural network is then trained using the first dataset and the second dataset to obtain a precipitation event classification model and a precipitation prediction model. The prediction module predicts grid precipitation events using the precipitation event classification model and predicts wet day grid precipitation using the precipitation amount prediction model. Based on the predicted grid precipitation events and the wet day grid precipitation, a multi-source precipitation data fusion dataset is obtained.

[0016] According to one embodiment of this application, the determining module is specifically used for: Based on the meteorological information, the geographical information, and the normalized vegetation index, initial environmental prediction factors are determined, and the initial environmental prediction factors are discretized and graded to obtain discretized environmental prediction factors. Based on the preset formula for calculating the degree of precipitation interpretation, the degree of interpretation of the discrete environmental prediction factors on precipitation is calculated, and qualified environmental prediction factors that meet the preset conditions are selected according to the calculation results. The precipitation explanation degree of each pair of qualified environmental predictors is calculated, and the target environmental predictors are obtained based on the precipitation explanation degree of each pair of factors and the precipitation explanation degree of each factor in the qualified environmental predictors.

[0017] According to one embodiment of this application, the determining module is specifically used for: Based on the meteorological information, the geographical information, and the normalized vegetation index, the spatial environment prediction factors of the observation station are determined according to a preset interpolation strategy. The spatial environment prediction factors are unified into a daily scale according to a preset time scale, and the seasons in which the station observes precipitation are located are divided to obtain the temporal environment prediction factors. The spatial environment prediction factors and the temporal environment prediction factors are then fused to obtain the initial environment prediction factors.

[0018] According to one embodiment of this application, the preset formula for calculating the degree of precipitation interpretation is as follows: ; in, L For the number of categories of environmental predictors, N h and N These represent the number of units for the h-th type of environmental predictor and the number of units for the global environmental factor, respectively. σ 2 and σ These represent the precipitation variance of the h-th type environmental predictor and the precipitation variance of the global environmental predictor, respectively.

[0019] According to one embodiment of this application, the construction module is specifically used for: Based on a preset partitioning ratio, the first dataset is divided into a first training set and a first test set, and the first validation set is determined from the first training set using the K-fold cross-validation method. The first training set is input into the convolutional neural network for training to obtain the first initial model parameters; Based on the first initial model parameters, the first validation set is input into the convolutional neural network for performance evaluation, and the first model parameters are determined based on the first performance evaluation result. Based on the first model parameters, the first test set is input into the convolutional neural network for model testing, and the precipitation event classification model is obtained when the first test result meets the preset requirements.

[0020] According to one embodiment of this application, the construction module is specifically used for: Based on a preset partitioning ratio, the second dataset is divided into a second training set and a second test set, and the second validation set is determined from the second training set using the K-fold cross-validation method. The second training set is input into the convolutional neural network to obtain the second initial model parameters; Based on the second initial model parameters, the second validation set is input into the convolutional neural network for performance evaluation, and the second target model parameters are determined based on the second performance evaluation results; Based on the second model parameters, the second test set is input into the convolutional neural network for model testing, and the precipitation prediction model is obtained when the second test result meets the preset requirements.

[0021] According to one embodiment of this application, the prediction module is specifically used for: Determine whether the predicted precipitation event in the grid is a rainless day; If the predicted precipitation event is a rainless day, the daily precipitation is recorded as 0; otherwise, the daily precipitation is recorded as the wet day precipitation. A multi-source precipitation data fusion dataset was obtained based on multiple daily precipitation data.

[0022] The machine learning-based multi-source precipitation data fusion device for mining multi-dimensional spatial information proposed in this application acquires measured daily precipitation, multi-source grid precipitation, meteorological, geographical, and normalized vegetation indices. It determines and extracts sample environmental prediction factors using a geographic detector, preprocesses the multi-source grid precipitation data, extracts grid precipitation data and sample environmental prediction factors centered on observation stations, normalizes the data to obtain a sample dataset, and uses it to train a convolutional neural network to obtain a precipitation event classification model and a precipitation prediction model. The model then predicts grid precipitation events and wet day precipitation, forming a multi-source precipitation data fusion dataset. This addresses the problem of insufficient consideration of multi-dimensional spatial information in related technologies, fully mining the effective information from various environmental variables and multi-dimensional data during the multi-source precipitation data fusion process.

[0023] To achieve the above objectives, a third aspect of this application provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the program to implement the machine learning multi-source precipitation data fusion method for mining multi-dimensional spatial information as described in the above embodiments.

[0024] To achieve the above objectives, a fourth aspect of this application provides a computer-readable storage medium storing a computer program that is executed by a processor to implement the machine learning multi-source precipitation data fusion method for mining multi-dimensional spatial information as described in the above embodiments.

[0025] To achieve the above objectives, a fifth aspect of this application provides a computer program product, which, when executed by a processor, implements the machine learning multi-source precipitation data fusion method for mining multi-dimensional spatial information as described in the above embodiments.

[0026] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0027] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart illustrating a machine learning method for fusing multi-source precipitation data to mine multi-dimensional spatial information, according to an embodiment of this application. Figure 2 This is a schematic diagram illustrating the screening of environmental prediction factors by a geographic detector according to an embodiment of this application; Figure 3 This is a schematic diagram of a multichannel convolutional neural network model provided according to an embodiment of this application; Figure 4This is a flowchart of a machine learning method for fusing multi-source precipitation data to mine multi-dimensional spatial information according to an embodiment of this application; Figure 5 This is a block diagram of a machine learning multi-source precipitation data fusion device for mining multi-dimensional spatial information according to an embodiment of this application; Figure 6 This is a schematic diagram of the structure of an electronic device provided according to an embodiment of this application. Detailed Implementation

[0028] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0029] The following describes, with reference to the accompanying drawings, a machine learning method for fusing multi-source precipitation data to mine multi-dimensional spatial information according to an embodiment of this application. First, the machine learning method for fusing multi-source precipitation data to mine multi-dimensional spatial information according to an embodiment of this application will be described with reference to the accompanying drawings.

[0030] Figure 1 This is a flowchart of a machine learning method for fusing multi-source precipitation data to mine multi-dimensional spatial information, according to an embodiment of this application.

[0031] like Figure 1 As shown, the machine learning-based multi-source precipitation data fusion method for mining multi-dimensional spatial information includes the following steps: In step S101, measured daily precipitation data, multi-source grid precipitation data, meteorological information, geographic information, and normalized vegetation index are obtained from observation stations within the target area.

[0032] Among them, measured daily precipitation data refers to precipitation data directly measured by ground observation equipment and statistically summarized on a daily basis. The Normalized Difference Vegetation Index (NDVI) is an index calculated using satellite remote sensing technology based on the differences in reflectance of vegetation to specific spectral bands.

[0033] Specifically, the study acquires measured daily precipitation data from observation stations within the target area, the latitude and longitude of the stations, satellite remote sensing precipitation data, global and regional reanalysis data, station interpolated grid precipitation data, altitude, slope and aspect data of the study area, normalized vegetation index, and other data. Among these, the global reanalysis data variables include precipitation, 2m air temperature, 2m dew point temperature, wind speed, and air pressure.

[0034] The target area can be the Yarlung Tsangpo River basin and its surrounding areas. The measured daily precipitation data used in this application's embodiments comes from the China Meteorological Administration, collecting daily precipitation data from 162 meteorological stations over the period of 2001-2019. The station interpolated gridded precipitation data uses CN05.1 (China 0.25°×0.25° daily gridded precipitation dataset). Satellite remote sensing data is from Global Precipitation Measurement (GPM) data provided by NASA and the Japan Aerospace Exploration Agency (JAXA). This application's embodiments use GPMIMERG Final (the final version of the integrated multi-satellite precipitation retrieval data under the Global Precipitation Measurement Project), with a spatiotemporal resolution of 0.5h, 0.1°*0.1°, and a selected period of 2001-2019. The data download website is (https: / / pmm.nasa.gov / data-access / downloads / gpm). Global reanalysis data is from the European Centre for Medium-Range Weather (ECMWF). Forecasts (ECWMF) Fifth Generation Global Atmospheric Reanalysis - Land Component (ERA5-Land) is used. The spatiotemporal resolution is 1 hour, 0.1°*0.1°. Downloaded variables include precipitation, 2m dew point temperature, 2m air temperature, air pressure, and U / V (east-west / north-south) wind speed. The selected period is 2001-2019. The download URL is (https: / / cds.climate.copernicus.eu / ). Regional reanalysis data uses TPReanalysis (Third Pole Reanalysis), which employs the WRF model (Weather Research and Forecasting Model), centered on the Tibetan Plateau, and uses ERA5 (ECMWF Fifth Generation Global Atmospheric Reanalysis) as the data source. AtmosphericReanalysis was used to generate dynamic downscaling data for lateral boundary conditions. The spatiotemporal resolution of the data is 1 hour, and the data size is 9 km * 9 km. The download address is (http: / / biggeo.gvc.gu).(se / TPReanalysis / ); Elevation data for the target area were obtained using HydroSHEDS (Hydrological Data and Maps Based on Shuttle Elevation Derivatives at Multiple Scales), derived from the Shuttle Radar Topography Mission (SRTM), with a resolution of 3 arcseconds. Slope and aspect data were calculated using DEM (Digital Elevation Model). Normalized Difference Vegetation Index (NDVI) data were selected from SPOT NDVI (Système Probatoire d'Observation de la Terre), a dataset with a spatial resolution of 1 km, covering the period from 2001 to 2019.

[0035] In step S102, based on a preset geographic detector, the explanatory power of meteorological information, geographic information and normalized vegetation index on precipitation is quantified, and target environmental prediction factors are screened according to the quantification results.

[0036] Specifically, a method based on geographic detectors to quantify the explanatory power of environmental predictors on precipitation is introduced to identify the explanatory power of each environmental predictor on precipitation and to identify the interactions between environmental predictors. This helps to screen environmental predictors for subsequent machine learning models and avoid overfitting caused by excessively high machine learning dimensionality.

[0037] Furthermore, in some embodiments, based on a preset geographic detector, the explanatory power of meteorological information, geographic information, and normalized vegetation index on precipitation is quantified, and target environmental prediction factors are screened according to the quantification results. This includes: determining initial environmental prediction factors based on meteorological information, geographic information, and normalized vegetation index, and performing discretization and hierarchical processing on the initial environmental prediction factors to obtain discretized environmental prediction factors; calculating the explanatory power of the discretized environmental prediction factors on precipitation based on a preset precipitation explanatory power calculation formula, and screening qualified environmental prediction factors that meet preset conditions according to the calculation results; calculating the precipitation explanatory power when two factors in the qualified environmental prediction factors work together, and obtaining the screened target environmental prediction factors based on the precipitation explanatory power when two factors work together and the precipitation explanatory power of each factor in the qualified environmental prediction factors.

[0038] Optionally, in some embodiments, the initial environmental prediction factor is determined based on meteorological information, geographic information, and normalized vegetation index, including: determining the spatial environmental prediction factor of the observation station based on meteorological information, geographic information, and normalized vegetation index according to a preset interpolation strategy; unifying the spatial environmental prediction factor to a daily scale according to a preset time scale, dividing the season in which the station observes precipitation to obtain a temporal environmental prediction factor, and fusing the spatial environmental prediction factor and the temporal environmental prediction factor to obtain the initial environmental prediction factor.

[0039] Optionally, in some embodiments, the preset formula for calculating the degree of precipitation interpretation is: ; in, L For the number of categories of environmental predictors, N h and N These represent the number of units for the h-th type of environmental predictor and the number of units for the global environmental factor, respectively. σ 2 and σ These represent the precipitation variance of the h-th type environmental predictor and the precipitation variance of the global environmental predictor, respectively.

[0040] The initial environmental predictor factors are all candidate variables that may affect the spatial distribution or dynamic changes of precipitation, initially extracted from three types of basic data: meteorological information, geographic information, and normalized difference vegetation index. The preset precipitation explanatory power refers to the calculation method and judgment criteria pre-set before analysis to quantify the explanatory power of environmental predictor factors on the spatial differentiation or dynamic changes of precipitation. Preset conditions can be user-defined conditions, conditions obtained through a limited number of experiments, or conditions obtained through a limited number of computer simulations. The preset interpolation strategy can be user-defined, a strategy obtained through a limited number of experiments, or a strategy obtained through a limited number of computer simulations. The preset time scale can be user-defined, a time scale obtained through a limited number of experiments, or a time scale obtained through a limited number of computer simulations. The daily scale refers to the time dimension standard for observing, statistically analyzing, integrating, and analyzing data or phenomena using one day (24 hours) as the basic time unit.

[0041] Specifically, environmental prediction factors, including geographical factors (latitude and longitude, altitude, aspect, and slope), meteorological factors (2m air temperature, 2m dew point temperature, wind speed, and surface pressure), and NDVI, are interpolated to the location of the stations. Environmental prediction factors at different time scales are uniformly organized to a daily scale. The season in which the station observes precipitation is divided according to the criteria of winter (December-February), spring (March-May), summer (June-August), and autumn (September-November), which serve as the initial environmental prediction factors. Each environmental prediction factor is discretized, and continuous variables are divided into 3-9 levels using the natural breakpoint method, the equal interval method, and the quantile method to obtain discretized environmental prediction factors.

[0042] The explanatory power (q) of each environmental predictor factor on precipitation under different classification methods was calculated. The classification method and classification level corresponding to the largest q were selected as the optimal parameters for the geographic detector. Environmental factors were ranked according to their importance based on the q value, and the results of each environmental predictor factor were tested for significance. A threshold was set, and environmental predictors with q exceeding the threshold and passing the significance test were retained; these were considered qualified environmental predictors. The qualified environmental predictors were labeled as X1, X2, …X p These qualified environmental predictors are paired up, and the q-value of the combined effect of the two environmental predictors is calculated and denoted as qc. The qc is compared with the q-value of the two environmental predictors acting alone to determine the impact of the interaction of the environmental predictors on precipitation. If qc is greater than q, both environmental predictors are retained. If qc is less than either environmental predictor q, the environmental predictor with the larger q is retained for subsequent model input; target environmental predictors are then selected.

[0043] For example, such as Figure 2 As shown, Figure 2 This is a schematic diagram of a geographic detector filtering environmental prediction factors according to an embodiment of this application.

[0044] In step S103, the multi-source grid precipitation data is preprocessed. Grid precipitation data centered on the observation station or prediction grid and sample environmental prediction factors are extracted from the preprocessed multi-source grid precipitation data and target environmental prediction factors. The grid precipitation data and sample environmental prediction factors are then normalized to obtain a sample dataset.

[0045] Among them, the sample environmental predictor refers to the local environmental predictor data extracted from the determined target environmental predictor, around the specific spatial location of the observation station center, and spatially matched with the corresponding grid precipitation data.

[0046] Specifically, the target environmental prediction factors and multi-source grid precipitation data from step S102 are preprocessed. Hourly-scale precipitation data from each grid are accumulated to a daily scale to maintain consistency with daily precipitation data observed at the stations. The input prediction factors include two categories: grid precipitation data (ERA5-Land, TPReanalysis, GPM, CN05.1) and environmental information (geographic information, meteorological information, etc.). The study area is divided into grids, and the spatial resolution of the fused data is set to 3.3 km. The aforementioned prediction factors are interpolated onto the fused data grid. Centered on a station or prediction grid, prediction factors are extracted from the surrounding 5*5 grids. Min-Max normalization is performed on each prediction factor (sample environmental prediction factors) and daily precipitation (grid precipitation data).

[0047] In step S104, a convolutional neural network is constructed, and the sample dataset is divided into a first dataset and a second dataset. The convolutional neural network is then trained using the first dataset and the second dataset to obtain a precipitation event classification model and a precipitation prediction model.

[0048] The first dataset and the second dataset can be datasets pre-defined by the user, datasets obtained through a limited number of experiments, or datasets obtained through a limited number of computer simulations.

[0049] Specifically, a machine learning multi-source precipitation data fusion model is constructed, namely a convolutional neural network (CNN). The CNN is trained based on the first and second datasets to obtain a precipitation event classification model and a precipitation prediction model.

[0050] Optionally, in some embodiments, training a convolutional neural network using a first dataset to obtain a precipitation event classification model includes: dividing the first dataset into a first training set and a first test set based on a preset partitioning ratio, and determining a first validation set from the first training set using a K-fold cross-validation method; inputting the first training set into the convolutional neural network for training to obtain first initial model parameters; inputting the first validation set into the convolutional neural network for performance evaluation based on the first initial model parameters, and determining the first model parameters based on the first performance evaluation results; inputting the first test set into the convolutional neural network for model testing based on the first model parameters, and obtaining a precipitation event classification model when the first test results meet preset requirements.

[0051] Here, the first training set and the first test set refer to the subsets separated from the first dataset according to a preset division ratio, which are core sets used for learning or testing convolutional neural network parameters. The first validation set refers to the subset obtained from the first training set through K-fold cross-validation, used to evaluate the generalization ability of the first initial model parameters and adjust hyperparameters. The first initial model parameters refer to the set of model parameters obtained through preliminary training after inputting the first training set into the convolutional neural network, without being fined by the first validation set. The first performance evaluation result refers to the set of quantitative results used to measure the generalization ability of the first initial model parameters, calculated by using a preset classification task evaluation index after inputting the first validation set into the convolutional neural network based on the first initial model parameters. The first model parameters refer to the set of learnable parameters of the convolutional neural network with reliable generalization ability, obtained by retraining with the first training set after optimizing hyperparameters through the performance evaluation of the first validation set based on the first initial model parameters.

[0052] Specifically, in this embodiment, 0.1 mm / d is used as the precipitation event threshold. Daily precipitation greater than 0.1 mm / d is designated as a wet day, and daily precipitation less than 0.1 mm / d is designated as a dry day. The first dataset is randomly shuffled and divided into a first training set (Trainset) and a first test set (Testset) according to a preset ratio, which can be 7:3. Each sample in the first training set contains multi-dimensional information, and the data format of the input prediction factor in each sample is as follows: ,…, ; in, This represents the numerical value of a certain environmental predictor at a single spatial location (row i, column j). i The value of j is 1-5; the value of j is 1-5. t This represents the number of time steps. x The input predictive factors include precipitation data from different grid sources, geographic information, meteorological information, etc.

[0053] That is, each sample has a dimension of 5*5*t, and each predictor in the sample has a dimension of 5*5. The hyperparameters of the convolutional neural network are divided into three categories: structural hyperparameters (number of convolutional layers, number of fully connected layers), sensitive algorithm hyperparameters (kernel size, learning rate), and other hyperparameters (number of convolutional kernels, dropout rate, number of neurons). Common hyperparameters are selected, and the performance of all hyperparameter combinations in each category is tested sequentially. In the process of optimizing hyperparameters, K-fold cross-validation (K can be 5) is used. The first training set is further divided into a training set and a first validation set, that is, the first training set is divided into 5 groups (trainset1, trainset2, ..., trainset5). Each time, one group is used as the validation set, and the remaining 4 groups are used as the training set. The experiment is repeated 5 times for each group of parameters. For example, such as... Figure 3 As shown, Figure 3 This is a schematic diagram of a multichannel convolutional neural network model provided according to an embodiment of this application.

[0054] The first training set data is input into the convolutional layer of the convolutional neural network. The convolutional layer includes f convolutional kernels, each with a size of m*m*t. The t predictors correspond to t channels. During convolution, each predictor is convolved with its corresponding kernel channel, and the values ​​of t channels are superimposed to obtain the result of one convolutional kernel. This process extracts multi-dimensional spatial feature information from a 5*5 grid of t predictors for each sample. The formula for calculating the j-th convolutional kernel is: ; in, This represents the convolution operation. Z j , W i,j , b j Let represent the feature map corresponding to the j-th convolutional kernel, the weight matrix of the i-th channel of the j-th convolutional kernel, and the bias of the j-th convolutional kernel, respectively. This represents the activation function of the convolutional neural network. The sigmoid function (S-shaped function) is used in precipitation event recognition (classification) models, while the ReLU function (rectified linear unit) is used in precipitation prediction models. A dropout layer (regularization technique) is added after the convolutional layers to randomly discard some neurons during model training, preventing the network from over-relying on specific neurons, enhancing the network's generalization ability, and preventing overfitting. The dropout layer results are input into a pooling layer, which reduces computational complexity by decreasing the size of the feature map; in this embodiment, a max pooling layer is used. A flattening layer follows the pooling layer, which transforms the output into a one-dimensional array. The one-dimensional data is then input into a fully connected layer, where the extracted feature maps are transformed into the network's final output. For precipitation event classification models, the model outputs predicted probabilities. Considering that the proportions of wet and dry days are not entirely the same, an F1 score is introduced. The probability value corresponding to the maximum F1 score is selected as the optimal classification threshold, and this index is used to evaluate the performance of the hyperparameter combination. The F1 score is related to the precision and recall of the classification model. The precision calculation formula is: ; The formula for calculating recall is: ; The formula for calculating F1 score is: ; in, T The number of samples correctly predicted as wet days by the model. P This represents the number of samples that the model incorrectly predicted as wet days. N This represents the number of samples that the model incorrectly predicted as dry days.

[0055] For precipitation prediction models, previous operations were all based on normalized values. Therefore, it is necessary to perform inverse normalization before outputting the data to obtain the predicted precipitation y. predict The root mean square error (RMSE) is used to evaluate the performance of different hyperparameters. The calculation formula is as follows: ; in, s The number of data points.

[0056] Repeat the above steps, compare the model validation set performance under different hyperparameter combinations, and select the best hyperparameter combination for precipitation event identification and precipitation prediction models for subsequent model training and simulation.

[0057] A precipitation event classification model is constructed using optimized hyperparameters to identify precipitation events; a convolutional neural network classification model is trained to predict precipitation events in the prediction grid. The specific method is as follows: A precipitation event classification model was constructed using optimized hyperparameters. A 5-fold cross-validation method was employed to further divide the first training set (Trainset) into a training set and a first validation set. Specifically, the training set was divided into five groups, with one group used as the validation set each time, and the remaining groups used as the training set to train the model parameters. The validation set predictions were then input into the trained model. This experiment was repeated five times to obtain the predictions for the complete training set. A global classification threshold was determined based on the Precision, Recall, and F1 score of the complete training set. Specifically, if the model's predicted probability was greater than or equal to the global classification threshold, the predicted day was considered a wet day; otherwise, it was considered a dry day. The classification model was trained using the first training set. The first test set (Testset) was input into the trained model, and the predicted probability values ​​were output. Based on the previously determined global classification threshold, the output predicted probabilities were classified. The evaluation metric, F1 score, was calculated to verify whether the trained classification model could effectively identify precipitation events.

[0058] Optionally, in some embodiments, training a convolutional neural network using a second dataset to obtain a precipitation prediction model includes: dividing the second dataset into a second training set and a second test set based on a preset division ratio, and determining a second validation set from the second training set using a K-fold cross-validation method; inputting the second training set into the convolutional neural network for training to obtain second initial model parameters; based on the second initial model parameters, inputting the second validation set into the convolutional neural network for performance evaluation, and determining second target model parameters based on the second performance evaluation results; based on the second model parameters, inputting the second test set into the convolutional neural network for model testing, and obtaining a precipitation prediction model when the second test results meet preset requirements.

[0059] Specifically, a precipitation prediction model is constructed using the selected hyperparameters described above, and trained using the second training set. The second test set is then input into the trained model to output predicted values, which are inversely normalized to obtain the predicted precipitation. The RMSE (Real-Time Sequence of Precipitation) metric for the test set is calculated to verify whether the prediction model can predict precipitation effectively. It should be noted that the method for constructing the precipitation prediction model is similar to the method for constructing the precipitation event classification model described above; to avoid redundancy, it will not be elaborated upon here.

[0060] In step S105, grid precipitation events are predicted using a precipitation event classification model, and wet-day grid precipitation is predicted using a precipitation amount prediction model. A multi-source precipitation data fusion dataset is obtained based on the predicted grid precipitation events and wet-day grid precipitation.

[0061] Specifically, t predictor data points for each prediction grid (5x5 grid) are extracted, normalized, and input into a trained precipitation event classification model to obtain daily, grid-by-grid prediction probability values. Based on a global classification threshold, the output prediction probabilities are classified, and the daily wet / dry condition of each grid is output. Similarly, t predictor data points for each prediction grid (5x5 grid) are extracted, normalized, and input into a trained precipitation prediction model to output predicted values. After inverse normalization, daily, grid-by-grid precipitation prediction values ​​are obtained. This yields the predicted precipitation events and wet day precipitation amounts for each grid, and based on this, a multi-source precipitation data fusion dataset is generated.

[0062] Optionally, in some embodiments, a multi-source precipitation data fusion dataset is obtained based on the predicted grid precipitation event and the wet day grid precipitation, including: determining whether the predicted grid precipitation event is a rainless day; if the predicted grid precipitation event is a rainless day, then recording the daily precipitation as 0, otherwise, recording the daily precipitation as the wet day grid precipitation; and obtaining a multi-source precipitation data fusion dataset based on multiple daily precipitation values.

[0063] Among them, a rainless day refers to a day in the grid precipitation event prediction where the predicted or observed precipitation is lower than the preset meteorological threshold.

[0064] Specifically, based on the prediction results of the precipitation event classification model and the precipitation prediction model, for the same date and the same grid, if the precipitation classification model predicts a rainless day, the precipitation at that grid point on that day is recorded as 0; if the precipitation classification model predicts a rainy day, the precipitation on that day is recorded as the precipitation predicted by the precipitation prediction model. Following this principle, a fusion dataset of daily 3.3km*3.3km resolution multi-source precipitation data from 2001 to 2019 is generated.

[0065] It should be noted that the above embodiments are merely illustrative examples of the technical solutions of this application and are not intended to limit the application in any way. Without departing from the scope of the technical solutions of this application, some modifications or alterations can be made to the above-disclosed technical content to create equivalent embodiments. All such modifications and alterations that do not depart from the scope of the technical solutions of this application shall still fall within the scope of the technical solutions of this application.

[0066] To facilitate a deeper understanding of the machine learning-based multi-source precipitation data fusion method for mining multi-dimensional spatial information presented in this application, the following section combines... Figure 4 Please provide a detailed explanation.

[0067] like Figure 4 As shown, Figure 4This is a flowchart of a machine learning method for fusing multi-source precipitation data to mine multi-dimensional spatial information, according to an embodiment of this application. The method includes: first, introducing a geographic detector to identify the explanatory power of environmental predictive factors on precipitation and the interactions between various environmental predictive factors, thereby achieving effective screening of environmental predictive factors; then, extracting multiple predictive factors, such as regional reanalysis data, and multi-dimensional grid information centered on stations and prediction grid points for each predictive factor, improving the insufficient description of spatial information in previous data fusion processes; using a convolutional neural network to achieve multi-channel convolution, obtaining multi-dimensional spatial features in the predictive factors, establishing precipitation event classification and precipitation prediction models respectively, simulating precipitation events and precipitation amounts sequentially, and finally integrating the simulation results of precipitation events and precipitation amounts to achieve daily, grid-point-by-grid fusion of multi-source precipitation data within the prediction range. The fused precipitation data can absorb the advantages of precipitation data from different sources, significantly improving accuracy compared to the original data.

[0068] The machine learning-based multi-source precipitation data fusion method proposed in this application, which mines multi-dimensional spatial information, acquires measured daily precipitation, multi-source grid precipitation, meteorological, geographical, and normalized vegetation indices. It uses a geographic detector to determine and extract sample environmental prediction factors, preprocesses multi-source grid precipitation data, extracts grid precipitation data and sample environmental prediction factors centered on observation stations, normalizes the data to obtain a sample dataset, and uses it to train a convolutional neural network to obtain a precipitation event classification model and a precipitation prediction model. The model then predicts grid precipitation events and wet day precipitation, forming a multi-source precipitation data fusion dataset. This method solves the problem of insufficient consideration of multi-dimensional spatial information in related technologies, fully mining the effective information from various environmental variables and multi-dimensional data during the multi-source precipitation data fusion process.

[0069] Next, referring to the accompanying drawings, a machine learning multi-source precipitation data fusion device for mining multi-dimensional spatial information according to an embodiment of this application is described.

[0070] Figure 5 This is a block diagram of a machine learning multi-source precipitation data fusion device for mining multi-dimensional spatial information according to an embodiment of this application.

[0071] like Figure 5 As shown, the machine learning multi-source precipitation data fusion device 10 for mining multi-dimensional spatial information includes: an acquisition module 100, a determination module 200, a processing module 300, a construction module 400, and a prediction module 500.

[0072] Among them, the acquisition module 100 acquires measured daily precipitation data, multi-source grid precipitation data, meteorological information, geographic information and normalized vegetation index of observation stations in the target area; The determination module 200, based on a preset geographic detector, quantifies the explanatory power of meteorological information, geographic information, and normalized vegetation index on precipitation, and selects target environmental prediction factors based on the quantification results. The processing module 300 preprocesses the multi-source grid precipitation data, extracts grid precipitation data and sample environmental prediction factors centered on the observation station or prediction grid from the preprocessed multi-source grid precipitation data and target environmental prediction factors, and normalizes the grid precipitation data and sample environmental prediction factors to obtain the sample dataset. Module 400 is constructed to build a convolutional neural network. The sample dataset is divided into a first dataset and a second dataset. The convolutional neural network is trained using the first dataset and the second dataset respectively to obtain a precipitation event classification model and a precipitation prediction model. The prediction module 500 predicts grid precipitation events through a precipitation event classification model and predicts wet day grid precipitation through a precipitation amount prediction model. Based on the predicted grid precipitation events and wet day grid precipitation, a multi-source precipitation data fusion dataset is obtained.

[0073] According to one embodiment of this application, the determining module 200 is specifically used for: Based on meteorological information, geographic information, and normalized vegetation index, initial environmental prediction factors are determined, and the initial environmental prediction factors are discretized and graded to obtain discretized environmental prediction factors. Based on the preset formula for calculating the degree of precipitation interpretation, the degree of interpretation of discrete environmental predictors on precipitation is calculated, and qualified environmental predictors that meet the preset conditions are selected based on the calculation results. The precipitation explanation degree of each pair of qualified environmental predictors is calculated, and the target environmental predictors are obtained based on the precipitation explanation degree of each pair of factors and the precipitation explanation degree of each factor in the qualified environmental predictors.

[0074] According to one embodiment of this application, the determining module 200 is specifically used for: Based on meteorological information, geographic information, and normalized vegetation index, spatial environment prediction factors for observation stations are determined according to a preset interpolation strategy. The spatial environment prediction factors are unified into a daily scale according to a preset time scale. The seasons in which the precipitation is observed at the stations are divided to obtain the temporal environment prediction factors. The spatial environment prediction factors and the temporal environment prediction factors are then integrated to obtain the initial environment prediction factors.

[0075] According to one embodiment of this application, the preset formula for calculating the degree of precipitation interpretation is: ; Among them, among them, L For the number of categories of environmental predictors,N h and N These represent the number of units for the h-th type of environmental predictor and the number of units for the global environmental factor, respectively. σ 2 and σ These represent the precipitation variance of the h-th type environmental predictor and the precipitation variance of the global environmental predictor, respectively.

[0076] According to one embodiment of this application, the construction module 400 is specifically used for: Based on the preset partitioning ratio, the first dataset is divided into a first training set and a first test set, and the first validation set is determined from the first training set using the K-fold cross-validation method. The first training set is input into the convolutional neural network to obtain the first initial model parameters; Based on the first initial model parameters, the first validation set is input into the convolutional neural network for performance evaluation, and the first model parameters are determined based on the first performance evaluation results. Based on the first model parameters, the first test set is input into the convolutional neural network for model testing, and when the first test result meets the preset requirements, a precipitation event classification model is obtained.

[0077] According to one embodiment of this application, the construction module 400 is specifically used for: Based on the preset partitioning ratio, the second dataset is divided into a second training set and a second test set, and the second validation set is determined from the second training set using the K-fold cross-validation method. The second training set is input into the convolutional neural network to obtain the second initial model parameters; Based on the second initial model parameters, the second validation set is input into the convolutional neural network for performance evaluation, and the second target model parameters are determined based on the second performance evaluation results. Based on the second model parameters, the second test set is input into the convolutional neural network for model testing, and the precipitation prediction model is obtained when the second test results meet the preset requirements.

[0078] According to one embodiment of this application, the prediction module 500 is specifically used for: Determine whether the predicted precipitation event in the grid is a rainless day; If the predicted precipitation event is a rainless day, the daily precipitation is recorded as 0; otherwise, the daily precipitation is recorded as the wet day precipitation. A multi-source precipitation data fusion dataset was obtained based on multiple daily precipitation data.

[0079] It should be noted that the foregoing explanation of the machine learning multi-source precipitation data fusion method for mining multi-dimensional spatial information also applies to the machine learning multi-source precipitation data fusion device for mining multi-dimensional spatial information in this embodiment, and will not be repeated here.

[0080] The machine learning-based multi-source precipitation data fusion device for mining multi-dimensional spatial information proposed in this application acquires measured daily precipitation, multi-source grid precipitation, meteorological, geographical, and normalized vegetation indices. It determines and extracts sample environmental prediction factors using a geographic detector, preprocesses the multi-source grid precipitation data, extracts grid precipitation data and sample environmental prediction factors centered on observation stations, normalizes the data to obtain a sample dataset, and uses it to train a convolutional neural network to obtain a precipitation event classification model and a precipitation prediction model. The model then predicts grid precipitation events and wet day precipitation, forming a multi-source precipitation data fusion dataset. This addresses the problem of insufficient consideration of multi-dimensional spatial information in related technologies, fully mining the effective information from various environmental variables and multi-dimensional data during the multi-source precipitation data fusion process.

[0081] Figure 6 This is a schematic diagram of an electronic device provided in an embodiment of the present invention. The electronic device may include: The memory 601, the processor 602, and the computer program stored on the memory 601 and capable of running on the processor 602.

[0082] When the processor 602 executes the program, it implements the machine learning multi-source precipitation data fusion method for mining multi-dimensional spatial information provided in the above embodiments.

[0083] Furthermore, electronic devices also include: Communication interface 603 is used for communication between memory 601 and processor 602.

[0084] The memory 601 is used to store computer programs that can run on the processor 602.

[0085] The memory 601 may include high-speed RAM (Random Access Memory) memory, and may also include non-volatile memory, such as at least one disk storage.

[0086] If the memory 601, processor 602, and communication interface 603 are implemented independently, then the communication interface 603, memory 601, and processor 602 can be interconnected via a bus to complete communication between them. The bus can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 6 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0087] Optionally, in a specific implementation, if the memory 601, processor 602, and communication interface 603 are integrated on a single chip, then the memory 601, processor 602, and communication interface 603 can communicate with each other through an internal interface.

[0088] The processor 602 may be a CPU (Central Processing Unit), an ASIC (Application Specific Integrated Circuit), or one or more integrated circuits configured to implement embodiments of the present invention.

[0089] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described machine learning multi-source precipitation data fusion method for mining multi-dimensional spatial information.

[0090] This application also provides a computer program product, which includes a computer program. When the computer program is executed by a processor, it implements the steps in any of the above embodiments of the machine learning multi-source precipitation data fusion method for mining multi-dimensional spatial information.

[0091] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0092] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0093] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A machine learning method for fusing multi-source precipitation data to mine multi-dimensional spatial information, characterized in that, Includes the following steps: Acquire measured daily precipitation data, multi-source grid precipitation data, meteorological information, geographic information, and normalized vegetation index from observation stations within the target area; Based on a preset geographic detector, the explanatory power of the meteorological information, the geographic information, and the normalized vegetation index on precipitation is quantified, and target environmental prediction factors are screened based on the quantification results. The multi-source grid precipitation data is preprocessed, and grid precipitation data centered on the observation station or prediction grid and sample environmental prediction factors are extracted from the preprocessed multi-source grid precipitation data and the target environmental prediction factors. The grid precipitation data and the sample environmental prediction factors are then normalized to obtain a sample dataset. Convolutional neural networks are constructed, and the sample dataset is divided into a first dataset and a second dataset. The convolutional neural network is trained using the first dataset and the second dataset respectively to obtain a precipitation event classification model and a precipitation prediction model. The precipitation event classification model predicts grid precipitation events, and the precipitation amount prediction model predicts wet day grid precipitation. Based on the predicted grid precipitation events and the wet day grid precipitation, a multi-source precipitation data fusion dataset is obtained.

2. The method according to claim 1, characterized in that, The method, based on a preset geographic detector, quantifies the explanatory power of the meteorological information, the geographic information, and the normalized vegetation index on precipitation, and selects target environmental prediction factors based on the quantification results, including: Based on the meteorological information, the geographical information, and the normalized vegetation index, initial environmental prediction factors are determined, and the initial environmental prediction factors are discretized and graded to obtain discretized environmental prediction factors. Based on the preset formula for calculating the degree of precipitation interpretation, the degree of interpretation of the discrete environmental prediction factors on precipitation is calculated, and qualified environmental prediction factors that meet the preset conditions are selected according to the calculation results. The precipitation explanation degree of each pair of qualified environmental predictors is calculated, and the target environmental predictors are obtained based on the precipitation explanation degree of each pair of factors and the precipitation explanation degree of each factor in the qualified environmental predictors.

3. The method according to claim 2, characterized in that, The determination of initial environmental prediction factors based on the meteorological information, the geographical information, and the normalized vegetation index includes: Based on the meteorological information, the geographical information, and the normalized vegetation index, the spatial environment prediction factors of the observation station are determined according to a preset interpolation strategy. The spatial environment prediction factors are unified into a daily scale according to a preset time scale, and the seasons in which the station observes precipitation are located are divided to obtain the temporal environment prediction factors. The spatial environment prediction factors and the temporal environment prediction factors are then fused to obtain the initial environment prediction factors.

4. The method according to claim 2, characterized in that, The preset formula for calculating the degree of precipitation interpretation is as follows: ； in, L For the number of categories of environmental predictors, N h and N These represent the number of units for the h-th type of environmental predictor and the number of units for the global environmental factor, respectively. σ 2 and σ These represent the precipitation variance of the h-th type environmental predictor and the precipitation variance of the global environmental predictor, respectively.

5. The method according to claim 1, characterized in that, The convolutional neural network is trained using the first dataset to obtain a precipitation event classification model, including: Based on a preset partitioning ratio, the first dataset is divided into a first training set and a first test set, and the first validation set is determined from the first training set using the K-fold cross-validation method. The first training set is input into the convolutional neural network for training to obtain the first initial model parameters; Based on the first initial model parameters, the first validation set is input into the convolutional neural network for performance evaluation, and the first model parameters are determined based on the first performance evaluation result. Based on the first model parameters, the first test set is input into the convolutional neural network for model testing, and the precipitation event classification model is obtained when the first test result meets the preset requirements.

6. The method according to claim 1, characterized in that, The convolutional neural network is trained using the second dataset to obtain a precipitation prediction model, including: Based on a preset partitioning ratio, the second dataset is divided into a second training set and a second test set, and the second validation set is determined from the second training set using the K-fold cross-validation method. The second training set is input into the convolutional neural network for training to obtain the second initial model parameters; Based on the second initial model parameters, the second validation set is input into the convolutional neural network for performance evaluation, and the second target model parameters are determined based on the second performance evaluation results; Based on the second model parameters, the second test set is input into the convolutional neural network for model testing, and the precipitation prediction model is obtained when the second test result meets the preset requirements.

7. The method according to claim 1, characterized in that, The process of obtaining a multi-source precipitation data fusion dataset based on the predicted grid precipitation events and the wet day grid precipitation includes: Determine whether the predicted precipitation event in the grid is a rainless day; If the predicted precipitation event is a rainless day, the daily precipitation is recorded as 0; otherwise, the daily precipitation is recorded as the wet day precipitation. A multi-source precipitation data fusion dataset was obtained based on multiple daily precipitation data.

8. A machine learning-based multi-source precipitation data fusion device for mining multi-dimensional spatial information, characterized in that, include: The acquisition module acquires measured daily precipitation data, multi-source grid precipitation data, meteorological information, geographic information, and normalized vegetation index from observation stations within the target area; The determination module, based on a preset geographic detector, quantifies the explanatory power of the meteorological information, the geographic information, and the normalized vegetation index on precipitation, and selects target environmental prediction factors based on the quantification results. The processing module preprocesses the multi-source grid precipitation data, extracts grid precipitation data centered on the observation station or prediction grid and sample environment prediction factors from the preprocessed multi-source grid precipitation data and the target environment prediction factors, and normalizes the grid precipitation data and the sample environment prediction factors to obtain a sample dataset. The module constructs a convolutional neural network and divides the sample dataset into a first dataset and a second dataset. The convolutional neural network is then trained using the first dataset and the second dataset to obtain a precipitation event classification model and a precipitation prediction model. The prediction module predicts grid precipitation events using the precipitation event classification model and predicts wet day grid precipitation using the precipitation amount prediction model. Based on the predicted grid precipitation events and the wet day grid precipitation, a multi-source precipitation data fusion dataset is obtained.

9. An electronic device, characterized in that, include: The memory, the processor, and the computer program stored in the memory and executable on the processor, the processor executing the program to implement the machine learning multi-source precipitation data fusion method for mining multi-dimensional spatial information as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, The program is executed by the processor to implement the machine learning multi-source precipitation data fusion method for mining multi-dimensional spatial information as described in any one of claims 1-7.

Images