Multi-source rainfall data fusion method based on meta-heuristic and machine learning

Abstract

The invention relates to the technical field of rainfall fusion, and discloses a multi-source rainfall data fusion method based on meta-heuristic and machine learning, and the method comprises the following steps: collecting ground rainfall observation data, multi-source satellite rainfall product data and auxiliary variable data, and carrying out the data preprocessing; dividing the preprocessed ground rainfall observation data, multi-source satellite rainfall product data and auxiliary variable data into a training set and a verification set in proportion; setting an initial value range of hyper-parameters of the multi-layer perceptron model, inputting the initial value range and the training set into the multi-layer perceptron model for training, and optimizing the hyper-parameters of the multi-layer perceptron model by adopting a sparrow search method to obtain a trained multi-layer perceptron model; inputting the verification set into the trained multilayer perceptron model for identifying the rainfall observation value of the verification set, and obtaining predicted ground rainfall observation data, namely fused rainfall data; according to the method, the precision of the fused rainfall data is improved, and meanwhile, an unestablished hypothesis depending on reality is avoided.

Description

Technical Field

[0001] The present invention relates to the field of precipitation fusion technology, and in particular to a multi-source precipitation data fusion method based on meta-heuristic and machine learning. Background Art

[0002] Precipitation is the most active meteorological element in the global hydrological cycle, climate and ecological balance. Obtaining high-precision precipitation data is crucial to improving the accuracy of hydrological modeling and flood prediction, as well as accurately revealing the laws of water cycle changes. At present, precipitation information mainly comes from three aspects: ground rain gauge observations, satellite remote sensing observations and inversion, and radar estimation. Rain gauge observations are the most common, direct and accurate method for obtaining precipitation information. However, due to the limitations of natural conditions and the number and spatial distribution of precipitation stations, precipitation observation data in some sparse rain gauge areas are not representative. Radar and satellite remote sensing observations can obtain effective rainfall information by detecting physical quantities related to rainfall in the atmosphere, which to a certain extent makes up for the limitations of ground observation data in spatial distribution and temporal continuity. However, affected by retrieval algorithms and sensor errors, radar and satellite remote sensing observation data are difficult to accurately estimate the corresponding precipitation data that changes over time. In the existing research on reducing the uncertainty of precipitation data products, most of them focus on the fusion of a single satellite precipitation product with precipitation ground observation data. This method is single and cannot accurately capture the spatial variation of precipitation. At the same time, the common methods used to fuse satellite precipitation products, such as inverse error variance weighting (IEVW), geographically weighted regression (GWR), triple collocation (TC), etc., rely on assumptions that do not hold in reality. Limited by the density of ground observation stations, the quality of data sets and the strategies adopted by researchers, most of these methods fuse a single satellite precipitation product with precipitation observation data, or only fuse multiple satellite precipitation products without considering observed precipitation, resulting in a small number of fused precipitation products. In addition, machine learning algorithms are widely used in precipitation product fusion because they can well capture the complex nonlinear relationship between variables and target outputs, do not rely on strict statistical assumptions, have efficient self-learning ability and data computing and processing capabilities, and can handle high-dimensional and large-scale data sets. However, due to the control of hyperparameters (the setting of model hyperparameters will affect the reliability of the learning process and results) and the influence of data sets, machine learning algorithms will overfit or fail to achieve the best fitting effect. Hyperparameters are parameters that are set before the learning process begins, rather than learned from training data. More accurate model hyperparameters are more conducive to accurate prediction and reduced model error. Determining hyperparameters is a challenging task. In machine learning, there are many types of model hyperparameters, and each hyperparameter has multiple potential optional values, which increases the search space for parameter values. However, in most studies, hyperparameters are usually manually adjusted or use default values, especially on a daily or smaller time scale, which greatly affects model accuracy and the model effect is not optimal. Summary of the invention

[0003] In view of the above-mentioned deficiencies in the prior art, the present invention provides a multi-source precipitation data fusion method based on metaheuristics and machine learning, which is used to solve the problems that the commonly used methods in the prior art rely on assumptions that are not true in reality, are limited by the density of ground data observation stations, the quality of data sets and the strategies adopted by researchers, the amount of fused precipitation product data is small, and the model prediction accuracy is low due to inaccurate hyperparameter selection in machine learning.

[0004] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is:

[0005] A multi-source precipitation data fusion method based on metaheuristics and machine learning includes the following steps:

[0006] S1. Collect ground precipitation observation data, multi-source satellite precipitation product data and auxiliary variable data;

[0007] Among them, auxiliary variable data include DEM data, longitude and latitude data, slope and aspect data, ground temperature data and spatial autocorrelation variable data;

[0008] S2, performing data preprocessing on the ground precipitation observation data, multi-source satellite precipitation product data and auxiliary variable data to obtain preprocessed ground precipitation observation data, multi-source satellite precipitation product data and auxiliary variable data;

[0009] S3, dividing the pre-processed ground precipitation observation data, multi-source satellite precipitation product data and auxiliary variable data into a training set and a validation set in proportion;

[0010] S4, setting the initial value range of the hyperparameters of the multilayer perceptron model, and inputting the hyperparameters of the multilayer perceptron model together with the training set for training, and optimizing the hyperparameters of the multilayer perceptron model by using the sparrow search method to obtain a trained multilayer perceptron model;

[0011] Among them, the hyperparameters include the number of hidden layers, the number of neurons in each hidden layer, and the learning rate;

[0012] S5. Input the validation set into the trained multi-layer perceptron model to identify the precipitation observation values ​​of the validation set, and obtain the predicted ground precipitation observation data, that is, the fused precipitation data.

[0013] Furthermore, the spatial autocorrelation variable data in step S1 is the precipitation calculated by the rainfall station defined by the inverse distance weighted method based on the neighboring observations, that is:

[0014]

[0015] Among them, SAC represents the precipitation calculated by the rain gauge station based on the nearby observations. represents the square of the distance from the rainfall station to the i1th adjacent station, w i1 represents the weight of the i1th neighboring site, obs i1 represents the precipitation at the i1th neighboring station.

[0016] Furthermore, step S2 specifically includes:

[0017] S21, using a resampling method to unify the spatial resolution of the auxiliary variable data and the multi-source satellite precipitation product data to obtain preprocessed auxiliary variable data;

[0018] S22. Match the multi-source satellite precipitation product data with the ground precipitation observation data in terms of time and position to obtain pre-processed ground precipitation observation data and multi-source satellite precipitation product data.

[0019] Furthermore, step S3 specifically includes:

[0020] The preprocessed multi-source satellite precipitation product data and auxiliary variable data are used as raw data, the preprocessed ground precipitation observation data are used as labeled data, and the labeled raw data are divided into training set and validation set in proportion.

[0021] Furthermore, the specific process of optimizing the hyperparameters of the multilayer perceptron model using the sparrow search method in step S4 is as follows:

[0022] Step 1: Set the initial value range of the hyperparameters of the multilayer perceptron model, and use the initial value range of the hyperparameters as the upper and lower bounds of the search space of the sparrow search method, and set the maximum number of iterations of the sparrow search method;

[0023] Step 2: Based on the upper and lower bounds of the search space, randomly generate an initial sparrow population in the search space, and calculate the fitness value of each individual sparrow in the initial sparrow population;

[0024] Among them, the fitness value of each individual sparrow in the initial sparrow population is the performance index of the multi-layer perceptron model trained by the hyperparameter combination on the validation set;

[0025] Step 3: Based on the fitness value of each individual sparrow in the initial sparrow population, find the best and worst sparrow individuals at present, retain the sparrow individual with the best current fitness value, and take the position of the sparrow individual with the best current fitness value as the global best position of the sparrow population;

[0026] Step 4: Determine whether the current number of iterations has reached the maximum number of iterations. If so, output the global best position of the sparrow population as the optimal solution. Otherwise, execute Step 5.

[0027] Among them, the optimal solution is the optimal hyperparameter combination;

[0028] Step 5: Based on the alarm value and safety threshold, update the location of the discoverer in the next iteration;

[0029] Step 6: Based on the position of the discoverer, update the position of the joiner in the next iteration;

[0030] Step 7: Update the position of the sentinel in the next iteration;

[0031] Step8: After executing steps Ste5-Step7, recalculate the fitness value of each sparrow individual, find out the current best and worst sparrow individuals, retain the sparrow individual with the best current fitness value, and use the position of the sparrow individual with the best current fitness value as the global optimal position of the sparrow population, add 1 to the number of iterations, and return to step Step4.

[0032] Furthermore, the formula for updating the position of the finder at the next iteration is:

[0033]

[0034] Among them, d represents the total number of dimensions, t represents the current number of iterations, P represents the finder, represents the position of the finder P in the jth dimension at the t+1th iteration, represents the position of the finder P in the jth dimension at the tth iteration, e represents the exponential function, α represents a random number between 0 and 1, iter max represents the maximum number of iterations, R2 represents the alarm value, ST represents the safety threshold, Q represents a random number that obeys the normal distribution, and L represents a 1×d matrix whose elements are all 1.

[0035] Furthermore, the formula for updating the position of the joiner in the next iteration is:

[0036]

[0037] Among them, n represents the total number of sparrows, J represents the joiner, represents the position of participant J in the jth dimension at the t+1th iteration, represents the position of participant J in the jth dimension at the tth iteration, represents the global worst position at the tth iteration, A represents a 1×d matrix whose elements are randomly assigned 1 or -1, T represents transpose, otherwise represents the judgment condition, and A + Represents an intermediate temporary variable and has no practical meaning.

[0038] Furthermore, the formula for updating the position of the sentinel at the next iteration is:

[0039]

[0040] Among them, V represents the vigilant, represents the position of the vigilant V in the jth dimension at the t+1th iteration, represents the position of the vigilant V in the jth dimension at the tth iteration, represents the global optimal position at the tth iteration, β represents the step size control parameter, K represents a random number between -1 and 1, and f V represents the fitness value of the vigilant V at the tth iteration, f g represents the global best fitness value at the tth iteration, f ω represents the global worst fitness value at the tth iteration, and ε represents the minimum constant to avoid the denominator being zero.

[0041] Furthermore, in step S5, the validation set is input into the trained multi-layer perceptron model to identify the precipitation observation values ​​of the validation set, and the predicted ground precipitation observation data, that is, the calculation formula of the fused precipitation data is obtained:

[0042] P o =MLP(P1,P2,…,P m ,X)

[0043] Among them, P o represents the predicted ground precipitation observation data, MLP represents the trained multi-layer perceptron model, P1, P2, P m represent the preprocessed precipitation product data of the 1st, 2nd and mth satellites respectively, and X represents the preprocessed auxiliary variable data.

[0044] The present invention has the following beneficial effects:

[0045] 1. The multi-source precipitation data fusion method based on metaheuristics and machine learning proposed in the present invention uses the sparrow search method to optimize the hyperparameters of the multilayer perceptron model, and can automatically optimize the hyperparameters on a daily or smaller time scale instead of default or manual adjustment, thereby reducing the occurrence of overfitting of the multilayer perceptron model during the training process, making the performance of the multilayer perceptron model optimal, and improving the prediction accuracy of the multilayer perceptron model;

[0046] 2. At the ground precipitation observation site, multi-source satellite precipitation product data and auxiliary variable data are used as the input of the multi-layer perceptron model. At the same time, spatial autocorrelation variable data is introduced into the auxiliary variable data, so that the precipitation observation data finally predicted by the multi-layer perceptron model is better and distributed more naturally, thereby improving the prediction accuracy;

[0047] 3. Compared with other machine learning models, the introduction of the multi-layer perceptron model can capture more data features through its network structure, better fit the nonlinear relationship between multi-source satellite precipitation product data and auxiliary variables, and avoid relying on assumptions that are not valid in reality, the density of ground observation stations, the quality of the data set, and the influence of the strategies adopted by researchers, thus solving the problem of small amount of precipitation data. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 A schematic diagram of a process of a multi-source precipitation data fusion method based on metaheuristics and machine learning proposed in the present invention;

[0049] Figure 2 Schematic diagram of the training principle of the multi-layer perceptron model in the embodiment. DETAILED DESCRIPTION

[0050] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0051] like Figure 1 As shown, a multi-source precipitation data fusion method based on metaheuristic and machine learning includes the following steps S1-S5:

[0052] S1. Collect ground precipitation observation data, multi-source satellite precipitation product data and auxiliary variable data; the auxiliary variable data include DEM data, latitude and longitude data, slope and aspect data, ground temperature data and spatial autocorrelation variable data.

[0053] In this embodiment, the collected ground precipitation observation data, multi-source satellite precipitation product data and auxiliary variable data are daily scale data; wherein, the multi-source satellite precipitation product data include precipitation product data collected by satellites such as IMERG_Early, GSMap_NRT, PDIR-Now and SM2RAIN-ASCAT, wherein the purpose of collecting multi-source satellite precipitation product data is to solve the problem of small amount of fused precipitation product data in the prior art. Compared with the fusion with a single satellite precipitation product data, the fusion of multiple satellite precipitation products with ground observation data can more accurately capture the spatial variation of precipitation, and avoid the lack of single satellite precipitation product data. Auxiliary variable data include DEM data, latitude and longitude data, slope and aspect data, ground temperature data and spatial autocorrelation variable data, among which DEM data is digital elevation model data. The purpose of introducing auxiliary variable data is to supplement and correct the deficiencies of multi-source satellite precipitation product data, enhance the model's understanding and description of precipitation phenomena from multiple dimensions, and ultimately improve the accuracy and spatial resolution of precipitation prediction; at the same time, the purpose of collecting ground precipitation observation data is to use the ground precipitation observation data collected by the rain gauge station as label data or dependent variable, and then input it into the model for training, so as to provide the model with a high-precision and reliable reference benchmark to improve the model training accuracy.

[0054] Specifically, the spatial autocorrelation variable data in step S1 is the precipitation calculated by the rainfall station based on the neighboring observations defined by the inverse distance weighted method, that is:

[0055]

[0056] Among them, SAC represents the precipitation calculated by the rain gauge station based on the nearby observations. represents the square of the distance from the rainfall station to the i1th adjacent station, w i1 represents the weight of the i1th neighboring site, obs i1 represents the precipitation at the i1th neighboring station.

[0057] In this embodiment, the purpose of collecting spatial autocorrelation variable data is to improve the model's understanding and prediction of the spatial distribution of precipitation by introducing spatial dependency information, so that the fused precipitation data predicted by the model is more accurate and distributed more evenly.

[0058] S2. Preprocessing the ground precipitation observation data, multi-source satellite precipitation product data and auxiliary variable data to obtain preprocessed ground precipitation observation data, multi-source satellite precipitation product data and auxiliary variable data.

[0059] In this embodiment, the purpose of data preprocessing for ground precipitation observation data, multi-source satellite precipitation product data and auxiliary variable data is to facilitate the extraction of multi-source satellite precipitation product data values ​​corresponding to the positions of the auxiliary variable data and the ground precipitation observation data, as well as the covariate data values ​​in the auxiliary variable data (i.e., DEM data, latitude and longitude data, slope and aspect data, ground temperature data and data values ​​in spatial autocorrelation variable data), so as to facilitate the subsequent input of these data into the model for training.

[0060] Specifically, step S2 specifically includes S21-S22:

[0061] S21. Use the resampling method to unify the spatial resolution of the auxiliary variable data and the multi-source satellite precipitation product data to obtain the preprocessed auxiliary variable data.

[0062] S22. Match the multi-source satellite precipitation product data with the ground precipitation observation data in terms of time and position to obtain pre-processed ground precipitation observation data and multi-source satellite precipitation product data.

[0063] In this embodiment, the purpose of matching the multi-source satellite precipitation product data with the ground precipitation observation data in time and position is to ensure that the ground precipitation observation data and the multi-source satellite precipitation product data correspond to each other in time and position, avoid errors caused by time or space misalignment, and ensure that the multi-source satellite precipitation product data and the ground observation data are compared and fused under the same conditions, thereby maximizing the scientificity and accuracy of the fused data.

[0064] S3. Divide the preprocessed ground precipitation observation data, multi-source satellite precipitation product data and auxiliary variable data into training set and validation set in proportion.

[0065] Specifically, step S3 includes:

[0066] The preprocessed multi-source satellite precipitation product data and auxiliary variable data are used as raw data, the preprocessed ground precipitation observation data are used as labeled data, and the labeled raw data are divided into training set and validation set in proportion.

[0067] S4. Set the initial value range of the hyperparameters of the multilayer perceptron model, and input the hyperparameters together with the training set into the multilayer perceptron model for training. Use the sparrow search method to optimize the hyperparameters of the multilayer perceptron model to obtain a trained multilayer perceptron model; wherein the hyperparameters include the number of hidden layers, the number of neurons in each hidden layer, and the learning rate.

[0068] In this embodiment, the hyperparameters of the multilayer perceptron model include the number of hidden layers, the number of neurons in each hidden layer, and the learning rate; since the setting of the hyperparameters will affect the reliability of the learning process and results of the multilayer perceptron model, setting the optimal hyperparameter combination can prevent the multilayer perceptron model from overfitting during training and achieve the best fitting effect; and in this embodiment, since the sparrow search method has a simple structure and fewer parameters, the algorithm is easy to understand and implement, and at the same time, it has strong global search capabilities, excellent ability to jump out of local optimality, fast convergence speed, and shows good adaptability and high robustness when dealing with various optimization problems. Therefore, the present invention uses the sparrow search method to optimize the hyperparameters, thereby obtaining the optimal hyperparameter combination; the multilayer perceptron model is trained together with the training set using the optimal hyperparameter combination, and the trained multilayer perceptron model can be obtained for predicting fused precipitation data.

[0069] Specifically, the specific process of optimizing the hyperparameters of the multilayer perceptron model using the sparrow search method in step S4 is:

[0070] Step 1: Set the initial value range of the hyperparameters of the multilayer perceptron model, and use the initial value range of the hyperparameters as the upper and lower bounds of the search space of the sparrow search method, and set the maximum number of iterations of the sparrow search method.

[0071] In this embodiment, the parameters of the sparrow search method include the population size, the maximum number of iterations, and the upper and lower bounds of the search space (i.e., the value range of the hyperparameters) so as to be used for the hyperparameter optimization of the multilayer perceptron model.

[0072] Step 2: Based on the upper and lower bounds of the search space, randomly generate an initial sparrow population in the search space, and calculate the fitness value of each individual sparrow in the initial sparrow population; wherein, the fitness value of each individual sparrow in the initial sparrow population is the performance indicator of the multilayer perceptron model trained by the hyperparameter combination on the validation set.

[0073] Step 3: Based on the fitness value of each sparrow individual in the initial sparrow population, find out the current best and worst sparrow individuals, retain the sparrow individual with the best current fitness value, and use the position of the sparrow individual with the best current fitness value as the global optimal position of the sparrow population.

[0074] Step 4: Determine whether the current number of iterations has reached the maximum number of iterations. If so, output the global best position of the sparrow population as the optimal solution. Otherwise, execute Step 5; the optimal solution is the optimal hyperparameter combination.

[0075] Step 5: Based on the alarm value and the safety threshold, update the location of the discoverer in the next iteration, that is:

[0076]

[0077] Among them, d represents the total number of dimensions, t represents the current number of iterations, P represents the finder, represents the position of the finder P in the jth dimension at the t+1th iteration, represents the position of the finder P in the jth dimension at the tth iteration, e represents the exponential function, α represents a random number between 0 and 1, iter max represents the maximum number of iterations, R2 represents the alarm value, ST represents the safety threshold, Q represents a random number that obeys the normal distribution, and L represents a 1×d matrix whose elements are all 1.

[0078] Step 6: Based on the position of the discoverer, update the position of the joiner in the next iteration, that is:

[0079]

[0080] Among them, n represents the total number of sparrows, J represents the joiner, represents the position of participant J in the jth dimension at the t+1th iteration, represents the position of participant J in the jth dimension at the tth iteration, represents the global worst position at the tth iteration, A represents a 1×d matrix whose elements are randomly assigned 1 or -1, T represents transpose, otherwise represents the judgment condition, and A + It represents an intermediate temporary variable and has no practical meaning. Step 7: Update the position of the sentinel in the next iteration, that is:

[0081]

[0082] Among them, V represents the vigilant, represents the position of the vigilant V in the jth dimension at the t+1th iteration, represents the position of the vigilant V in the jth dimension at the tth iteration, represents the global optimal position at the tth iteration, β represents the step size control parameter, K represents a random number between -1 and 1, and f V represents the fitness value of the vigilant V at the tth iteration, f g represents the global best fitness value at the tth iteration, f ω represents the global worst fitness value at the tth iteration, and ε represents the minimum constant to avoid the denominator being zero.

[0083] Step8: After executing steps Ste5-Step7, recalculate the fitness value of each sparrow individual, find out the current best and worst sparrow individuals, retain the sparrow individual with the best current fitness value, and use the position of the sparrow individual with the best current fitness value as the global optimal position of the sparrow population, add 1 to the number of iterations, and return to step Step4.

[0084] S5. Input the validation set into the trained multi-layer perceptron model to identify the precipitation observation values ​​of the validation set, and obtain the predicted ground precipitation observation data, that is, the fused precipitation data.

[0085] Specifically, in step S5, the validation set is input into the trained multi-layer perceptron model to identify the precipitation observation value of the validation set, and the predicted ground precipitation observation data, that is, the calculation formula of the fused precipitation data is obtained:

[0086] P o =MLP(P1,P2,…,P m ,X)

[0087] Among them, P o represents the predicted ground precipitation observation data, MLP represents the trained multi-layer perceptron model, P1, P2, P m represent the preprocessed precipitation product data of the 1st, 2nd and mth satellites respectively, and X represents the preprocessed auxiliary variable data.

[0088] In this embodiment, the training set and the optimal hyperparameters are input into the multilayer perceptron model for training. The training process is as follows: Figure 2 As shown, Figure 2 The input data of the multi-layer perceptron model includes pre-processed multi-source satellite precipitation product data and auxiliary variable data, ground precipitation observation data and optimal hyperparameters. After training, the multi-layer perceptron model can output fused precipitation products, that is, predicted ground precipitation observation data. Among them, the multi-layer perceptron model consists of an input layer, a hidden layer and an output layer, and each layer contains multiple neurons, and the neurons are connected with weights; the input layer is used to receive input data, and each input neuron corresponds to a feature of the input data, that is, the number of neurons in the input layer is equal to the number of features of the input data; the hidden layer is used to capture the complex interaction relationship between the input data (input features); the output layer is used to output the predicted data, that is, each neuron in the output layer corresponds to a possible category or predicted value. Among them, Figure 2 The LST-Night and LST-Day in the table are both ground temperature data.

[0089] S6. Input the validation set into the trained multi-layer perceptron model to identify the precipitation observation values ​​of the validation set, and obtain the predicted ground precipitation observation data, that is, the fused precipitation data.

[0090] Specifically, in step S6, the validation set is input into the trained multi-layer perceptron model to identify the precipitation observation value of the validation set, and the predicted ground precipitation observation data, that is, the calculation formula of the fused precipitation data is obtained:

[0091] P o =MLP(P1,P2,…,P m ,X)

[0092] Among them, P o represents the predicted ground precipitation observation data, MLP represents the trained multi-layer perceptron model, P1, P2, P m Both represent preprocessed multi-source satellite precipitation product data, and X represents preprocessed auxiliary variable data.

[0093] In summary, the present invention proposes a multi-source precipitation data fusion method based on metaheuristic and machine learning. First, the sparrow search method is used to optimize the hyperparameters of the multilayer perceptron model. The hyperparameters can be automatically optimized on a daily or smaller time scale, rather than by default or manual adjustment, thereby reducing the occurrence of overfitting of the multilayer perceptron model during the training process, so that the performance of the multilayer perceptron model is optimized, and the prediction accuracy of the multilayer perceptron model is improved; secondly, multi-source satellite precipitation product data and auxiliary variable data are used as inputs of the multilayer perceptron model at the ground precipitation observation station, and spatial autocorrelation variable data is introduced into the auxiliary variable data so that the precipitation observation data finally predicted by the multilayer perceptron model is better and more naturally distributed, thereby improving the prediction accuracy; in addition, the introduction of the multilayer perceptron model can capture more data features through its network structure compared to other machine learning models, can better fit the nonlinear relationship between precipitation and auxiliary variables, and avoid relying on assumptions that are not established in reality, the density of ground observation stations, the quality of the data set, and the influence of the strategies adopted by researchers, thereby solving the problem of small amount of precipitation data.

[0094] The present invention uses specific embodiments to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.

[0095] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. A multi-source precipitation data fusion method based on metaheuristics and machine learning, characterized in that: The following steps are involved: S1. Collect ground precipitation observation data, multi-source satellite precipitation product data and auxiliary variable data; Among them, auxiliary variable data include DEM data, longitude and latitude data, slope and aspect data, ground temperature data and spatial autocorrelation variable data; S2, performing data preprocessing on the ground precipitation observation data, multi-source satellite precipitation product data and auxiliary variable data to obtain preprocessed ground precipitation observation data, multi-source satellite precipitation product data and auxiliary variable data; S3, dividing the pre-processed ground precipitation observation data, multi-source satellite precipitation product data and auxiliary variable data into a training set and a validation set in proportion; S4, setting the initial value range of the hyperparameters of the multilayer perceptron model, and inputting the hyperparameters of the multilayer perceptron model together with the training set for training, and optimizing the hyperparameters of the multilayer perceptron model by using the sparrow search method to obtain a trained multilayer perceptron model; Among them, the hyperparameters include the number of hidden layers, the number of neurons in each hidden layer, and the learning rate; S5. Input the validation set into the trained multi-layer perceptron model to identify the precipitation observation values ​​of the validation set, and obtain the predicted ground precipitation observation data, that is, the fused precipitation data.

2. The multi-source precipitation data fusion method based on metaheuristics and machine learning according to claim 1 is characterized in that: The spatial autocorrelation variable data in step S1 is the precipitation calculated by the rainfall station based on the neighboring observations defined by the inverse distance weighted method, that is: Among them, SAC represents the precipitation calculated by the rain gauge station based on the nearby observations. represents the square of the distance from the rainfall station to the i1th adjacent station, w i1 represents the weight of the i1th neighboring site, obs i1 represents the precipitation at the i1th neighboring station.

3. The multi-source precipitation data fusion method based on metaheuristics and machine learning according to claim 2 is characterized in that: Step S2 specifically includes: S21, using a resampling method to unify the spatial resolution of the auxiliary variable data and the multi-source satellite precipitation product data to obtain preprocessed auxiliary variable data; S22. Match the multi-source satellite precipitation product data with the ground precipitation observation data in terms of time and position to obtain pre-processed ground precipitation observation data and multi-source satellite precipitation product data.

4. The multi-source precipitation data fusion method based on metaheuristic and machine learning according to claim 3 is characterized in that: Step S3 specifically includes: The preprocessed multi-source satellite precipitation product data and auxiliary variable data are used as raw data, the preprocessed ground precipitation observation data are used as labeled data, and the labeled raw data are divided into training set and validation set in proportion.

5. The multi-source precipitation data fusion method based on metaheuristics and machine learning according to claim 4 is characterized in that: The specific process of optimizing the hyperparameters of the multilayer perceptron model using the sparrow search method in step S4 is as follows: Step 1: Set the initial value range of the hyperparameters of the multilayer perceptron model, and use the initial value range of the hyperparameters as the upper and lower bounds of the search space of the sparrow search method, and set the maximum number of iterations of the sparrow search method; Step 2: Based on the upper and lower bounds of the search space, randomly generate an initial sparrow population in the search space, and calculate the fitness value of each individual sparrow in the initial sparrow population; Among them, the fitness value of each individual sparrow in the initial sparrow population is the performance index of the multi-layer perceptron model trained by the hyperparameter combination on the validation set; Step 3: Based on the fitness value of each individual sparrow in the initial sparrow population, find the best and worst sparrow individuals at present, retain the sparrow individual with the best current fitness value, and take the position of the sparrow individual with the best current fitness value as the global best position of the sparrow population; Step 4: Determine whether the current number of iterations has reached the maximum number of iterations. If so, output the global best position of the sparrow population as the optimal solution. Otherwise, execute Step 5. Among them, the optimal solution is the optimal hyperparameter combination; Step 5: Based on the alarm value and safety threshold, update the location of the discoverer in the next iteration; Step 6: Based on the position of the discoverer, update the position of the joiner in the next iteration; Step 7: Update the position of the sentinel in the next iteration; Step8: After executing steps Ste5-Step7, recalculate the fitness value of each sparrow individual, find out the current best and worst sparrow individuals, retain the sparrow individual with the best current fitness value, and use the position of the sparrow individual with the best current fitness value as the global optimal position of the sparrow population, add 1 to the number of iterations, and return to step Step4.

6. The multi-source precipitation data fusion method based on metaheuristics and machine learning according to claim 5 is characterized in that: The formula for updating the finder's position at the next iteration is: Among them, d represents the total number of dimensions, t represents the current number of iterations, P represents the finder, represents the position of the finder P in the jth dimension at the t+1th iteration, represents the position of the finder P in the jth dimension at the tth iteration, e represents the exponential function, α represents a random number between 0 and 1, iter max represents the maximum number of iterations, R2 represents the alarm value, ST represents the safety threshold, Q represents a random number that obeys the normal distribution, and L represents a 1×d matrix whose elements are all 1.

7. The multi-source precipitation data fusion method based on metaheuristics and machine learning according to claim 6 is characterized in that: The formula for updating the position of the joiner at the next iteration is: Among them, n represents the total number of sparrows, J represents the joiner, represents the position of participant J in the jth dimension at the t+1th iteration, represents the position of participant J in the jth dimension at the tth iteration, represents the global worst position at the tth iteration, A represents a 1×d matrix whose elements are randomly assigned 1 or -1, T represents transpose, otherwise represents the judgment condition, and A + Represents an intermediate temporary variable and has no practical meaning.

8. The multi-source precipitation data fusion method based on metaheuristics and machine learning according to claim 7 is characterized in that: The formula for updating the position of the sentinel at the next iteration is: Among them, V represents the vigilant, represents the position of the vigilant V in the jth dimension at the t+1th iteration, represents the position of the vigilant V in the jth dimension at the tth iteration, represents the global optimal position at the tth iteration, β represents the step size control parameter, K represents a random number between -1 and 1, and f V represents the fitness value of the vigilant V at the tth iteration, f g represents the global best fitness value at the tth iteration, f ω represents the global worst fitness value at the tth iteration, and ε represents the minimum constant to avoid the denominator being zero.

9. The multi-source precipitation data fusion method based on metaheuristics and machine learning according to claim 8, characterized in that: In step S5, the validation set is input into the trained multi-layer perceptron model to identify the precipitation observation values ​​of the validation set, and the predicted ground precipitation observation data is obtained, that is, the calculation formula of the fused precipitation data is: P o =MLP(P1,P2,…,P m ,X) Among them, P o represents the predicted ground precipitation observation data, MLP represents the trained multi-layer perceptron model, P1, P2, P m represent the preprocessed precipitation product data of the 1st, 2nd and mth satellites respectively, and X represents the preprocessed auxiliary variable data.

Images