Multi-source information fusion small watershed precipitation estimation method, device and storage device

Abstract

The application discloses a kind of small watershed precipitation estimation method, equipment and storage device of multi-source information fusion, method includes the following steps: obtaining satellite remote sensing precipitation data;Precipitation data is reduced scale processing, and precipitation sequence is obtained;Obtain multi-source information, multi-source information includes: air temperature sequence and historical precipitation sequence of rainfall site;Multi-source information is spliced with precipitation sequence, and sample set is obtained;Sample set is divided into training set and test set;LSTM neural network model is constructed;The LSTM neural network model is trained using training set, and test set is used for testing, and the trained model is obtained;The input data to be predicted is input to the trained model, and the final precipitation estimation value is obtained.The application has beneficial effects: it can improve the monitoring ability of precipitation and non-precipitation area and effectively solve the problem of missing rainfall sequence caused by equipment failure of rainfall station, and can effectively capture local rainstorm in small watershed of medium and high altitude mountainous area.

Description

Technical Field

[0001] This invention relates to the field of meteorological forecasting, and in particular to a method for estimating precipitation in small watersheds by fusing multi-source information. Background Technology

[0002] Precipitation is one of the most important pieces of information in meteorological and hydrological simulations. Precipitation amount is the main factor determining dryness and wetness. Spatialized rainfall information is of great significance for water resource management, establishing watershed hydrological models, ecological environment governance, and studying climate change.

[0003] In existing precipitation monitoring methods, we generally consider the measurements from rain gauge stations as the true precipitation values. However, due to the influence of topography and severe weather, ground-based rain gauge stations deployed within a watershed are prone to malfunction during extreme weather events, resulting in missing data from some stations. In recent years, various remote sensing precipitation products (such as TRMM and IMERG) have become important references in various research and planning projects because they can provide global-scale, high-temporal-resolution precipitation data.

[0004] However, the relatively coarse spatial resolution of satellite remote sensing data severely limits its application in estimating missing values ​​in rainfall sequences, especially in small watersheds in mountainous areas where rainfall distribution exhibits significant spatial variability. Fusion of satellite precipitation data with ground-based rain gauge data can effectively address this problem.

[0005] Current research on the fusion of satellite and ground-based rain gauge data mainly focuses on large-scale regions and watersheds, with time scales primarily at the annual, seasonal, and monthly levels. However, research on data fusion for small-scale regions and watersheds is scarce, and studies at the hourly level are even rarer.

[0006] The reasons are as follows:

[0007] 1) Satellite precipitation products inherently possess a certain degree of uncertainty and cannot reflect the true precipitation situation in the study area;

[0008] 2) Satellite precipitation products have insufficient resolution, with multiple ground rain gauge stations sharing the same raster precipitation data, which fails to reflect the spatial variability of precipitation in small watersheds in mountainous areas. Summary of the Invention

[0009] To address the technical problem that existing methods cannot accurately estimate hourly precipitation in small watersheds using satellite precipitation data, this invention provides a multi-source information fusion method for estimating precipitation in small watersheds, which includes the following steps:

[0010] S1. Acquire satellite remote sensing precipitation data;

[0011] S2. Downscale the satellite remote sensing precipitation data to obtain the downscaled precipitation sequence;

[0012] S3. Obtain multi-source information, including: temperature sequence and historical precipitation sequence from rainfall stations; S4. Concatenate the multi-source information with the downscaled precipitation sequence to obtain a sample set;

[0013] S5. Divide the sample set into a training set and a test set;

[0014] S6. Construct an LSTM neural network model;

[0015] S7. Train the LSTM neural network model using the training set and test it using the test set to obtain the final trained model.

[0016] S8. Input the input data to be predicted into the trained model to obtain the final precipitation estimate.

[0017] A storage device that stores instructions and data for implementing a small watershed precipitation estimation method based on multi-source information fusion.

[0018] A multi-source information fusion device for small watershed precipitation estimation includes: a processor and a storage device; the processor loads and executes instructions and data in the storage device to implement a multi-source information fusion method for small watershed precipitation estimation.

[0019] The beneficial effects provided by this invention are: it can improve the monitoring capabilities of precipitation and non-precipitation areas and effectively solve the problem of missing rainfall sequences caused by equipment failure of rain gauges, while also effectively capturing localized rainstorms in small watersheds in mid-to-high altitude mountainous areas. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0021] Figure 2 This is a schematic diagram of how an LSTM neural network model processes data.

[0022] Figure 3 This is a schematic diagram of the hardware device operation according to an embodiment of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0024] Please refer to Figure 1 , Figure 1 This is a schematic diagram of the method flow of the present invention. The present invention provides a method for estimating precipitation in a small watershed by fusing multi-source information, comprising the following steps:

[0025] S1. Acquire satellite remote sensing precipitation data;

[0026] In this embodiment of the invention, precipitation data from the IMERG-E-V06 satellite is acquired;

[0027] S2. Downscale the satellite remote sensing precipitation data to obtain the downscaled precipitation sequence;

[0028] It should be noted that this invention selects n rainfall stations g within the watershed. k (k = 1, 2, Λ, n) is used as the estimation target of this invention.

[0029] Step S2 is as follows:

[0030] S21. Expand outwards by m grids from the area to be predicted, collecting satellite grid s. i Satellite precipitation data for the location (i = 1, 2, Λ, m);

[0031] S22. Use the Kriging method to downscale the spatial input vector to obtain the spatial output vector; where the spatial input vector R sat =[R sat (s1),R sat (s2),ΛR sat (s m )] T For the satellite grid center s i Satellite precipitation data for the location; the calculation process for the spatial output vector is as follows:

[0032]

[0033] Among them, R dow (g k ) is the rain gauge g k Satellite precipitation estimates; λ = [λ1, λ2, Λ, λ m [ ] represents the site weight coefficient, indicating the R value within the study area. sat For R dow (g k The degree of contribution of ); the spatial output vector is the downscaled precipitation sequence.

[0034] The calculation process for the site weight coefficient is as follows:

[0035] Construct the empirical variogram γ'(h) and the variogram γ(h):

[0036] Convert equation (4) into matrix form

[0037]

[0038] Kλ'=D (6)

[0039] λ'=K -1 D (7)

[0040] In equations (2)-(7), γ'(h) is the empirical variability function, and h = ||s i -s j || is the center position of the satellite grid. i and s j The Euclidean distance, in meters; N(h) is the set of observation pairs, representing the points within the given range ||s|. i -s j ||=h'inners i and s j The observed logarithm; γ(h) is the Gaussian model variogram, which is fitted by minimizing the error |γ'(h)-γ(h)|, l is the effective range, c is the maximum variogram value; μ is the Lagrange multiplier, γ(s) i -s j ) and γ(s j -g k ) represent the Euclidean distance as ||s i -s j ||and||s j -g k The variogram value of ||;

[0041] After solving λ' according to equation (7), the site weight coefficient λ is obtained according to equation (5).

[0042] S3. Obtain multi-source information, including: temperature sequence and historical precipitation sequence from rainfall stations;

[0043] S4. The multi-source information is spliced ​​with the downscaled precipitation sequence to obtain the sample set;

[0044] The process of obtaining the sample set in step S4 is as follows:

[0045] The downscaled precipitation series is concatenated with the temperature series and historical precipitation series from rainfall stations, and used as the input matrix X of the LSTM neural network model. t , where X t As shown in the following formula:

[0046] X t ={I t-M+1 ,LI t-1 ,I t} F×M (8)

[0047]

[0048] In equations (8)-(9), The measured precipitation data is for time t-1. For satellite precipitation data from the rain gauge at time t, Tem1 t Let Tem2 be the ambient temperature of the watershed at time t. t Let I be the dew point temperature at time t; t Let M be the input vector, M be the time step of the input matrix, and F be the number of variables in the input vector.

[0049] It should be noted that before inputting the training set into the LSTM neural network model, the Min-Max method is used to normalize both the training and test sets, resulting in the normalized training set X. * train Test set X * test The normalization process is as follows:

[0050]

[0051] In equation (10), a * 'a' is the normalized data sequence; 'a' is the original data sequence; 'a' is the normalized data sequence. min ,a max These are the minimum and maximum values ​​in the original data sequence.

[0052] S5. Divide the sample set into a training set and a test set;

[0053] In this embodiment of the invention, 80% of the data is used as the training sample set, and the remaining data is used as the test sample set.

[0054] S6. Construct an LSTM neural network model;

[0055] It should be noted that this invention employs a sliding window method, where the window length is n, and the sliding step is 1 each time, starting from X. * train Extract continuous data of length n and use it as the training set X for the LSTM network. train Extract the (n+1)th data point from the historical precipitation sequence of rainfall stations as X. train The label Y train At the same time from X * test Extract continuous data of length n and use it as the test set X for the LSTM network. test Extract the (n+1)th data point from the historical precipitation sequence of rainfall stations as X. test The label Y test The training and test sets are used as normalized historical data.

[0056] S7. Train the LSTM neural network model using the training set and test it using the test set to obtain the final trained model.

[0057] Please refer to Figure 2 , Figure 2 This is a schematic diagram of how an LSTM neural network model processes data.

[0058] It should be noted that the network training process is as follows:

[0059] S71. Set the number of training iterations of the LSTM network to epochs, and use the mean squared error (MSE) as the loss function. When the loss gradually decreases and stabilizes during the iteration process, the training can be considered to have ended.

[0060] S72, Transfer the training set X train The initial hidden layer state h0 and the initial hidden layer state are treated as a whole and input into the LSTM network. First, the output value f of the forget gate is calculated according to equation (11). t The generated data f t It is a number between 0 and 1, representing the cell state c. t-1 The percentage that should be retained. 0 represents c. t-1 All data is forgotten; 1 represents c. t-1 All data is retained. Next, the output value i of the input gate is calculated according to equations (12) and (13). t and candidate vectors of cellular state Next, the cell state needs to be updated according to equation (14). Finally, the probability vector of the "output gate" is combined with the cell state c activated by the tanh function. t Perform a dot product to obtain the final output result h. t (Equation 15).

[0061] f t =σ(W f [h t-1 ,x t ]+b f (11)

[0062] i t =σ(W i [h t-1 x t ]+b i (12)

[0063]

[0064] In equations (11)-(15), W c The weight matrix representing the cell state, b c W represents the bias of the cellular state.i Let b represent the weight matrix of the input gate. i W represents the bias of the input gate. f The weight matrix of the forget gate, b f W represents the bias of the forget gate. o Let b represent the weight matrix of the output gate. o h represents the bias of the output gate. t-1 x represents the output value of the LSTM network at time t-1. t Let σ represent the input value of the LSTM network at time t, and σ be the sigmoid function transform.

[0065] S73. Using a time-based backpropagation algorithm, the output value Y after model fitting is obtained. * train and Y train Fine-tuning the LSTM network updates its parameters, and through continuous iteration, the trained LSTM network is obtained.

[0066] S8. Input the input data to be predicted into the trained model to obtain the final precipitation estimate.

[0067] The input data to be predicted is used as the input matrix X′=[I t-2 ,I t-1 ,I t ] F×n The normalized output value h is obtained by inputting it into the trained model. t and for h t Inverse normalization is performed to obtain the combined precipitation estimate from the rain gauges at time t.

[0068] As an example, this case uses data from 7 typhoons and 19 heavy rainfall events recorded by 9 controlled rain gauge stations within the study basin from 2012 to 2022. These data are divided into training and test sets according to the number of typhoons and the amount of data, with a data ratio of 7:3. All 9 rain gauge stations within the basin are used as estimation targets. To illustrate the superiority of the Kriging-LSTM coupled model of this invention, the kriging model and the IMERG-LSTM model are used as comparative models. The estimation results of the test set using the three trained models are analyzed to demonstrate the advantages of this invention compared to other models.

[0069] Table 1 shows the evaluation metrics of the model. The results indicate that for satellite precipitation data with significant uncertainties, the Kriging-LSTM coupled model greatly improves the satellite precipitation error. The Kriging-LSTM coupled model shows the most significant improvement in the estimation accuracy of the nine targets: the mean CC value increased from 0.54 to 0.73; the mean MAE value decreased from 2.94 to 2.51; the mean RMSE value decreased from 6.33 to 5.29; and the mean RB value increased from -0.042 to -0.037.

[0070] Table 1 Comparison of accuracy of different estimation methods

[0071]

[0072] By applying different spatial downscaling methods to satellite precipitation data, a cumulative precipitation distribution map of Typhoon Maria was generated for this watershed.

[0073] In terms of rainfall distribution, the Kriging and LSTM models struggled to capture the spatial distribution of precipitation during Typhoon Maria, which exhibited significant spatial variations, particularly failing to pinpoint the concentrated areas of heavy rainfall. However, the Kriging-LSTM model accurately captured the center of the heavy rainfall during Maria.

[0074] Please see Figure 3 , Figure 3 This is a schematic diagram of the hardware device in operation according to an embodiment of the present invention. The hardware device specifically includes: a multi-source information fusion small watershed precipitation estimation device 401, a processor 402, and a storage device 403.

[0075] A multi-source information fusion small watershed precipitation estimation device 401: The multi-source information fusion small watershed precipitation estimation device 401 implements the multi-source information fusion small watershed precipitation estimation method.

[0076] Processor 402: The processor 402 loads and executes the instructions and data in the storage device 403 to implement the multi-source information fusion method for small watershed precipitation estimation.

[0077] Storage device 403: The storage device 403 stores instructions and data; the storage device 403 is used to implement the multi-source information fusion method for small watershed precipitation estimation.

[0078] The key technical point of this invention is:

[0079] (1) By integrating environmental variables, historical precipitation data from rain gauge stations, and downscaled remote sensing precipitation data, we can not only improve the monitoring capabilities for precipitation and non-precipitation areas, but also estimate the missing values ​​of precipitation sequences from rain gauge stations, effectively solving the problem of missing data caused by equipment failure at rain gauge stations.

[0080] (2) This method is well applicable to capturing local rainstorms in small watersheds in mountainous areas at medium and high altitudes, and the relevant results can provide technical support for the study of the spatiotemporal distribution of precipitation in small watersheds in mountainous areas.

[0081] In summary, the beneficial effects of this invention are: it can improve the monitoring capabilities of precipitation and non-precipitation areas and effectively solve the problem of missing rainfall sequences caused by equipment failure at rain gauge stations, while also effectively capturing localized rainstorms in small watersheds in mid-to-high altitude mountainous areas.

[0082] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for estimating precipitation in a small watershed by fusing multi-source information, characterized in that: include: S1. Acquire satellite remote sensing precipitation data; S2. Downscale the satellite remote sensing precipitation data to obtain the downscaled precipitation sequence; S3. Obtain multi-source information, including: temperature sequence and historical precipitation sequence from rainfall stations; S4. The multi-source information is spliced ​​with the downscaled precipitation sequence to obtain the sample set; S5. Divide the sample set into a training set and a test set; S6. Construct an LSTM neural network model; S7. Train the LSTM neural network model using the training set and test it using the test set to obtain the final trained model. S8. Input the data to be predicted into the trained model to obtain the final precipitation estimate; Step S2 is as follows: S21. Expand outwards by m grids from the area to be predicted, collecting data from the satellite grid centers s. i Satellite precipitation data for the location i = 1, 2, Λ, m; S22. Use the Kriging method to downscale the spatial input vector to obtain the spatial output vector; where the spatial input vector R sat =[R sat (s1),R sat (s2),ΛR sat (s m )] T For the satellite grid center s i Satellite precipitation data for the location; the calculation process for the spatial output vector is as follows: Among them, R dow (g k ) is the rain gauge g k Satellite precipitation estimates; λ = [λ1, λ2, Λ, λ m [ ] represents the site weight coefficient, indicating the R value within the study area. sat For R dow (g k The degree of contribution of ); the spatial output vector is the downscaled precipitation sequence; The calculation process for the site weight coefficient is as follows: Construct the empirical variogram γ'(h) and the variogram γ(h): Convert equation (4) into matrix form Kλ'=D (6) λ'=K -1 D (7) In formulas (2)-(7), γ'(h) is the empirical variogram, h = ||s i - s j || is the Euclidean distance between the central positions s i and s j of the satellite grids, with the unit of m; N(h) is the set of observation pairs, representing the number of observations of s i [[ID= (10]]- s j || = h; γ(h) is the Gaussian model variogram, which is fitted by minimizing the error of |γ'(h) - γ(h)|. l is the effective range, c is the maximum variogram value; μ is the Lagrange multiplier, γ(s i - s j [[ID= ( ]]16)) and γ(s i - s j ) and γ(s j - g k ) represent the variogram values with the Euclidean distances of ||s i - s j || and ||s j 0)- g k || respectively; After solving λ' according to equation (7), the site weight coefficient λ is obtained according to equation (5).

2. The method for estimating precipitation in a small watershed by multi-source information fusion as described in claim 1, characterized in that: The process of obtaining the sample set in step S4 is as follows: The downscaled precipitation series is concatenated with the temperature series and historical precipitation series from rainfall stations, and used as the input matrix X of the LSTM neural network model. t , where X t As shown in the following formula: X t ={I t-M+1 ,Λ I t-1 ,I t } F×M (8) In equations (8)-(9), The measured precipitation data is for time t-1. For satellite precipitation data from the rain gauge at time t, Tem1 t Let Tem2 be the ambient temperature of the watershed at time t. t Let I be the dew point temperature at time t; t Let M be the input vector, M be the time step of the input matrix, and F be the number of variables in the input vector.

3. The method for estimating precipitation in a small watershed by multi-source information fusion as described in claim 2, characterized in that: Before inputting the training set into the LSTM neural network model, the Min-Max method is used to normalize both the training and test sets, resulting in the normalized training set X. * train Test set X * test The normalization process is as follows: In equation (10), a * 'a' is the normalized data sequence; 'a' is the original data sequence; 'a' is the normalized data sequence. min ,a max These are the minimum and maximum values ​​in the original data sequence.

4. The method for estimating precipitation in a small watershed by multi-source information fusion as described in claim 1, characterized in that: Step S8 is as follows: The input data to be predicted is used as the input matrix X′=[I t-2 ,I t-1 ,I t ] F×n The normalized output value h is obtained by inputting it into the trained model. t and for h t Inverse normalization is performed to obtain the combined precipitation estimate from the rain gauges at time t.

5. A storage device, characterized in that: The storage device stores instructions and data to implement the small watershed precipitation estimation method based on multi-source information fusion as described in any one of claims 1 to 4.

6. A multi-source information fusion device for estimating precipitation in a small watershed, characterized in that: include: A processor and a storage device; the processor loads and executes instructions and data in the storage device to implement the small watershed precipitation estimation method based on multi-source information fusion as described in any one of claims 1 to 4.

Images