Method for carrying out extreme precipitation simulation by using CNN-LSTM model based on multi-source data fusion

Abstract

The invention discloses a method for performing extreme precipitation simulation by using a CNN-LSTM model based on multi-source data fusion. The method comprises the following steps: firstly, acquiring multi-source data of a research area, calculating the extreme precipitation of the area, and screening the most important key atmospheric circulation index influencing the extreme precipitation of a target area by using an extreme gradient lifting model; performing unified preprocessing on the multi-source data; then constructing a space-time prediction model fusing a convolutional neural network and a long-short-term memory network, and generating a deep learning prediction model through multi-feature joint training; and carrying out space-time prediction on the extreme precipitation by utilizing the preprocessed multi-source data and a deep learning model. According to the method, the extreme rainfall data, the atmospheric circulation index and the terrain DEM data are fused in a unified input framework, so that the model can perform joint prediction on rainfall by simultaneously utilizing the space-time rainfall distribution characteristics, the atmospheric circulation mode and the terrain information, and the prediction and simulation precision is high.

Description

Technical Field

[0001] The present invention belongs to the technical field of hydrology and meteorology, and in particular relates to a method for simulating extreme precipitation by using a CNN-LSTM model based on multi-source data fusion. Background Art

[0002] Extreme precipitation, as an extreme climate event, often leads to disasters such as floods, debris flows, landslides, and urban flooding, causing great harm to the natural environment and human society. Atmospheric circulation is one of the important factors affecting the occurrence of extreme precipitation events. Atmospheric circulation directly or indirectly affects the occurrence of extreme precipitation by affecting water vapor transport and the evolution of local climate dynamic conditions. For example, during the South Asian monsoon and East Asian monsoon, a large amount of water vapor is transported from the ocean to the land, which easily triggers heavy rains and floods. In addition to the dynamic characteristics of atmospheric circulation, regional terrain characteristics (such as altitude, elevation gradient, slope aspect, etc.) also have a significant impact on the distribution, intensity and frequency of extreme precipitation events. Changes in altitude, differences in slope and slope aspect, and the interaction between complex terrain and the atmosphere determine the uneven distribution of precipitation in space.

[0003] Traditional extreme precipitation prediction methods mainly include statistical analysis methods (such as regression analysis, time series models, extreme value theory) and model methods based on physical processes (such as global climate model GCM, regional climate model RCM). These methods have obvious limitations in comprehensively considering the joint effects of atmospheric circulation and terrain characteristics. In the formation of extreme precipitation, the global-scale atmospheric circulation driving is closely related to the local-scale terrain forcing. Traditional methods are difficult to fully capture this multi-scale, multi-factor nonlinear mechanism, and it is difficult to achieve high accuracy in both time and space. Statistical methods tend to model time series and are difficult to capture spatial distribution characteristics; although physical models can describe spatial patterns, their temporal resolution is insufficient, and local characteristics are not handled in a detailed manner. Summary of the invention

[0004] The purpose of the present invention is to provide a method for simulating extreme precipitation based on multi-source data fusion using a CNN-LSTM model, so as to solve the technical problems in the prior art that only a single feature type is focused on, the synergistic effect of terrain characteristics and key meteorological factors on extreme precipitation is ignored, and effective information in spatial and temporal dimensions is not fully mined.

[0005] In order to solve the above technical problems, the present invention adopts the following solutions: The method for simulating extreme precipitation using the CNN-LSTM model based on multi-source data fusion includes the following steps: Step S1: Acquire multi-source data of the target area, including historical precipitation raster data, atmospheric circulation index time series data, and terrain DEM data.

[0006] Step S2: Calculate the extreme precipitation index Rx5day in the target area and save it as a raster file with the same coordinate system as the source file. The calculation formula is as follows: (1) (2) In the above formula, S i is the sum of precipitation for 5 consecutive days starting from the 𝑖th day, P i For the i The daily precipitation of the day, Rx5day It is the maximum precipitation for 5 consecutive days each month.

[0007] Step S3: Using the extreme gradient boosting algorithm XGBoost, the key atmospheric circulation indexes that have the greatest impact on extreme precipitation in the target area are screened.

[0008] Step S4: resample the terrain DEM data to align it with the grid of the precipitation raster data, while ensuring that the terrain DEM data and the precipitation raster data have the same coordinate system.

[0009] Step S5: Min-Max normalization is performed on the above multi-source data, and a spatiotemporal prediction model integrating convolutional neural network CNN and long short-term memory network LSTM is constructed to perform spatiotemporal prediction of extreme precipitation using multi-source data. The process can be expressed as follows: (3) In the above formula, P extreme is the simulated value of extreme precipitation, X P , X Cn , X T They are the input historical extreme precipitation data, key atmospheric circulation index data and terrain DEM data. f It is a spatiotemporal prediction model driven by multi-source data that integrates convolutional neural network (CNN) and long short-term memory network (LSTM).

[0010] Step S6: Select the top 80% time series of extreme precipitation data and atmospheric circulation index data as the training set, build and train the model, use the mean square error (MSE) as the optimization target, calculate the square error between the predicted value and the true value, and through multiple rounds of training, make full use of the deep learning model driven by multi-source data to complete the simulation and prediction of the complex spatiotemporal patterns of extreme precipitation.

[0011] Step S7: Calculate the correlation coefficient R of each pixel prediction value to evaluate the accuracy of the simulation results.

[0012] As a core technology in the field of artificial intelligence, deep learning has demonstrated strong data-driven capabilities and nonlinear modeling advantages in the field of hydrology and meteorology, and can efficiently capture complex spatiotemporal characteristics and multi-factor interactions. Based on convolutional neural networks (CNN) and long short-term memory networks (LSTM), this invention overcomes the shortcomings of traditional methods in describing multi-scale and multi-dimensional interactions in spatial feature extraction and time series modeling; at the same time, deep learning can process large-scale, multi-source heterogeneous data, thereby improving prediction accuracy and generalization capabilities.

[0013] Further optimization, the specific process of step S3 includes: Step S3.1: Use ArcGIS pixel statistics method to calculate the mean time series of extreme precipitation in the target area. The calculation formula is as follows: (4) In the above formula, For the target area in the time period T The average extreme precipitation value in N is the total number of pixels, d T,j For raster files at time Segment T Medium pixel j The value of .

[0014] Step S3.2: Align the time series of atmospheric circulation index and extreme precipitation in time, merge them into a set of data as input data, input them into the extreme gradient boosting algorithm XGBoost, and use the first 80% of the time series as the training set and the rest as the test set. Use the extreme gradient boosting algorithm XGBoost to learn the impact of each index on extreme precipitation. Decision tree is the basic building block of XGBoost. During the construction of each tree, XGBoost will select the best split point according to the different values ​​of the features to maximize the model's predictive ability for the target variable. In this step, the algorithm calculates the importance of each index based on the improvement of the extreme precipitation time series prediction results at the split node, and finally selects the four indexes with the largest contribution as the key atmospheric circulation indexes according to the order of importance. The importance can be calculated using the following formula: (5) In the above formula, I k Index k The importance of Tr is the set of all decision trees in the model, St For Trees t The set of all split nodes in , f tr,s Number tr At the split node sThe features selected above; α ( ) is the indicator function, when f tr,s = k hour, α ( f tr,s = k )=1, otherwise 0.

[0015] Further optimization, the specific process of step S5 includes: Step S5.1: Copy the atmospheric circulation index data in the spatial dimension to align it with the pixel, merge the extreme precipitation data, atmospheric circulation index data, and terrain DEM data in the channel dimension to form a multi-channel input feature tensor, use a one-dimensional convolutional neural network CNN to extract the local spatial feature map of each pixel in the spatial dimension, and fuse the multi-channel information of extreme precipitation, atmospheric circulation index, and terrain DEM data to generate a discriminative spatial feature map, where the convolution process can be expressed by the following formula: (6) In the above formula, F is the output feature map of the convolutional layer, X is the input feature map, W is the convolution kernel weight, b is the bias and G is the activation function.

[0016] Step S5.2: The spatial feature sequence output by the convolutional neural network CNN is input into the long short-term memory neural network LSTM for time series modeling. LSTM selectively memorizes important information and forgets irrelevant information through a gating mechanism, thereby capturing the dynamic change trend in the past time and outputting the hidden state vector at the last moment. This vector condenses the comprehensive information of the entire time series as the basis for subsequent predictions; the process is expressed by the following formula: (7) In the above formula, h t is the hidden state at the current time step, x t is the input spatial feature map of the current time step, h t-1 is the hidden state at the previous time step.

[0017] Step S5.3: The hidden state vector output by the long short-term memory neural network LSTM is mapped back to the spatial feature through the fully connected layer and associated with each pixel position, where the fully connected layer mapping can be expressed as follows: (8) In the formula, zis the feature vector output by the fully connected layer, h T is the last hidden state of LSTM, W fc and b fc are the weights and biases of the fully connected layer.

[0018] Step S5.4: The mapped features are further converted into extreme precipitation Rx5day prediction values ​​consistent with the pixel distribution through the final convolution layer, and then the spatial distribution prediction of future time points is generated. The process is expressed by the following formula: (9) In the formula, Y is the predicted value of extreme precipitation for each pixel; z is the multidimensional feature vector obtained in step 5.3; Conv It is a one-dimensional convolution process; b Conv For bias.

[0019] Further optimization, in step S7, the correlation coefficient R is calculated by the following formula: (10) In the formula, R is the correlation coefficient for each pixel; M is the sample size; X m and Y m For each pixel m The true value and predicted value of samples; and is the mean of the true and predicted values ​​for each pixel.

[0020] Compared with the prior art, the present invention has the following beneficial effects: 1. The present invention constructs a spatiotemporal prediction method for extreme precipitation driven by deep learning based on multi-source data, which makes up for the limitation of traditional methods in insufficient utilization of terrain features and key atmospheric circulation factors, further mines the effective information in multi-source data, and makes full use of the synergistic effect of terrain features and key meteorological factors on extreme precipitation, providing an innovative solution for the prediction of extreme precipitation.

[0021] 2. Compared with traditional methods, the spatiotemporal prediction of extreme precipitation provided by the present invention integrates extreme precipitation data with atmospheric circulation index and terrain DEM data in a unified input framework, so that the model can simultaneously use spatiotemporal precipitation distribution characteristics, atmospheric circulation patterns and terrain information to jointly predict precipitation, with high prediction simulation accuracy. This solves the problem that the existing technology only focuses on a single feature type. 3. The method of simulating extreme precipitation using the CNN-LSTM model based on multi-source data fusion described in the present invention provides an innovative solution for the spatiotemporal prediction of extreme precipitation, can be widely used in the fields of water resources management and disaster warning, and provides strong support for scientific decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 This is a flow chart of the method for simulating extreme precipitation using a CNN-LSTM model based on multi-source data fusion according to the present invention; Figure 2 A schematic diagram of a spatiotemporal prediction model integrating a convolutional neural network (CNN) and a long short-term memory (LSTM) network according to the present invention; Figure 3 The spatial distribution map of the correlation of the prediction results from 2011 to 2022 was obtained for the Yarlung Zangbo River Basin simulation; Figure 4 This is a comparison chart between the predicted value and the true value of a certain pixel in the simulation results of the Yarlung Zangbo River Basin; Figure 5 The spatial distribution map of the correlation of the prediction results from 2011 to 2022 is obtained by using only a single feature (atmospheric circulation index). DETAILED DESCRIPTION

[0023] The technical solution of the present invention is described in detail below in conjunction with embodiments, but the protection scope of the present invention is not limited to the embodiments.

[0024] The Yarlung Zangbo River Basin is located in the southern Qinghai-Tibet Plateau and the northern foothills of the Himalayas. The average altitude in my country is over 4,000 m. It is one of the highest international rivers in the world, with high altitude, heavy precipitation, high heterogeneity, and high gradient. The Yarlung Zangbo River Basin in my country covers an area of ​​about 240,000 km 2 The main stream of the river is more than 2,000 km long and has an annual runoff of more than 160 billion m 3 , the water level drop in the lower reaches of the river is 2725m. Under the influence of climate warming, the precipitation in the Yarlung Zangbo River Basin has shown an overall upward trend at all scales in the past 40 years. Therefore, the spatiotemporal prediction of extreme precipitation in the basin has a far-reaching impact on disaster prevention and mitigation, comprehensive basin management and sustainable development.

[0025] In this embodiment, if Figure 1 As shown in FIG. 1 , the method for simulating extreme precipitation using the CNN-LSTM model based on multi-source data fusion specifically includes the following steps: Step 1: Obtain historical precipitation raster data, terrain DEM data and atmospheric circulation index data in the Yalong River Basin.

[0026] In this embodiment, precipitation grid data with a time span of January 1, 1961 to December 31, 2022 and a spatial resolution of 0.1° is used, with a total of 2402 pixels. The terrain DEM data DEM is SRTM elevation data with a spatial resolution of 90m. The atmospheric circulation index data is a total of 88 types of monthly time series with a time span of January 1961 to December 2022.

[0027] Step 2: Calculate the extreme precipitation index Rx5day in the Yajiang River Basin. Based on the daily historical precipitation grid data obtained in step 1, taking the extreme precipitation index Rx5day (maximum continuous 5-day precipitation, subsequent extreme precipitation represents Rx5day) as an example, the monthly extreme precipitation grid data in the Yajiang River Basin is obtained by calculation. The data maintains the same coordinate system as the source file. The calculation formula of Rx5day is as follows: (1) (2) In the above formula, S i is the sum of precipitation for 5 consecutive days starting from the 𝑖th day, P i is the daily precipitation, Rx5day It is the maximum precipitation for 5 consecutive days each month. The extreme precipitation data mentioned in all subsequent steps refer to the data calculated in this step.

[0028] Step 3: Use the extreme gradient boosting algorithm XGBoost to screen out the four key atmospheric circulation indices that have the greatest impact on extreme precipitation in the Yajiang River Basin. The specific steps are as follows: Step 3.1: Using ArcGIS pixel statistics method, calculate the mean time series of extreme precipitation in the basin. Some of the calculated data are shown in Table 1. The calculation formula is as follows: (3) In the above formula, For the Yajiang River Basin in the time period T The average extreme precipitation value in N is the total number of pixels, d T,j For raster files in time period T Medium pixel j The value of .

[0029] In this embodiment, the time period T is each month from 1961 to 2022, and the average extreme precipitation in the basin during this period is shown in Table 1.

[0030] Table 1 Average extreme precipitation in the basin

[0031] Step 3.2: Align the atmospheric circulation index time series and the average extreme precipitation time series calculated in step 3.1 in time, merge them into a set of data as input data, and bring them into the extreme gradient boosting algorithm XGBoost. The first 80% of the time series is used as the training set, and the rest is used as the test set. The extreme gradient boosting algorithm XGBoost is used to learn the impact of each index on extreme precipitation. Decision tree is the basic building block of XGBoost. During the construction of each tree, XGBoost will select the best split point according to the different values ​​of the features to maximize the model's prediction ability for the target variable. In this step, the extreme gradient boosting algorithm XGBoost calculates the importance of each index based on the improvement of the atmospheric circulation index in the split node on the prediction results of the extreme precipitation time series, and finally sorts them according to importance, and selects the four indices with the largest contribution as the key atmospheric circulation indices. The importance can be calculated using the following formula: (4) In the above formula, I k Index k The importance of Tr is the set of all decision trees in the model, St For Trees t The set of all split nodes in , f tr,s Number tr At the split node s The features selected above; α ( ) is the indicator function, when f tr,s = k hour, α ( f tr,s = k )=1, otherwise 0.

[0032] In this embodiment, the key atmospheric circulation indices finally selected are: East Asian trough intensity index, Northern Hemisphere polar vortex center intensity index, North Atlantic subtropical high area index and Northern Hemisphere polar vortex intensity index. Some data of each index are shown in Table 2.

[0033] Table 2 Key atmospheric circulation indices

[0034] Step 4: Using the basin precipitation raster data as the target template, resample the terrain DEM data to align it with the grid of the precipitation raster data, while ensuring that the terrain DEM data and the precipitation raster data have the same coordinate system.

[0035] Step 5: Perform Min-Max normalization on the extreme precipitation grid data obtained in step 2, the key atmospheric circulation index obtained in step 3, and the terrain DEM data obtained in step 4, and input them into the constructed spatiotemporal prediction model integrating the convolutional neural network CNN and the long short-term memory network LSTM, as shown in the figure. Figure 2 The process is shown as follows: (5) In the above formula, P extreme is the simulated value of extreme precipitation, X P , X Cn , X T They are respectively the normalized historical extreme precipitation data, n key atmospheric circulation index data, and terrain DEM data. f It is a spatiotemporal prediction model driven by multi-source data that integrates convolutional neural network (CNN) and long short-term memory network (LSTM).

[0036] The detailed steps are as follows: Step 5.1: In deep learning, data is usually input into the model in the form of multidimensional arrays (tensors). The feature tensor in the model can have multiple channels (dimensions), each of which represents different features or information. In this model, the atmospheric circulation index data is first copied in the spatial dimension to align with the pixels; the terrain DEM data is copied in the temporal dimension to align with each time step (monthly scale). Secondly, the extreme precipitation data, atmospheric circulation index data, and terrain DEM data are superimposed as different channels and combined into a multi-channel input feature tensor. Subsequently, a one-dimensional convolutional neural network CNN is used to extract the local spatial feature map of each pixel in the spatial dimension, and the multi-channel information of precipitation, atmospheric circulation index and terrain is fused to generate a discriminative spatial feature map, where the convolution process is expressed by the following formula: (6) In the above formula, F is the output feature map of the convolutional layer; X Is the input feature map. In this model, X It is multi-channel information consisting of precipitation data, atmospheric circulation index data and terrain DEM data; W The convolution kernel weights are randomly initialized before training and automatically adjusted according to the results after each training; b is the bias, which is initially set to 0 by default and will be automatically adjusted according to the results after each training; G is the activation function, and the ReLU function is selected in this embodiment.

[0037] Step 5.2: Input the spatial feature sequence output by the convolutional neural network CNN into the long short-term memory neural network LSTM for time series modeling. LSTM selectively memorizes important information and forgets irrelevant information through the gating mechanism, thereby capturing the dynamic change trend in the past time and outputting the hidden state vector at the last moment. This vector condenses the comprehensive information of the entire time series as the basis for subsequent predictions. The process is expressed by the following formula: (7) In the formula, h t is the hidden state at the current time step, x t is the input feature of the current time step, when t =0 x 0 Calculated by formula (6) F , h t-1 is the hidden state at the previous time step.

[0038] Step 5.3: The hidden state vector output by the long short-term memory neural network LSTM is mapped back to the spatial features through the fully connected layer and associated with each pixel position, where the fully connected layer mapping can be expressed as follows: (8) In the formula, z is the feature vector output by the fully connected layer, h T is the last hidden state of LSTM, W fc is the weight matrix of the fully connected layer. In this case, the Kaiming Uniform method is used for default initialization. b fc is the sum bias vector of the fully connected layer, which is initialized to 0 in this case and will be automatically adjusted according to the results after each training; Step 5.4: Through the final convolutional layer, the information of multiple feature channels of the multidimensional feature vector is integrated into one output channel to generate the extreme precipitation Rx5day forecast value for each pixel, and then generate the spatial distribution forecast of future time points. This process can be expressed by the following formula: (9) In the above formula, Y is the predicted value of extreme precipitation for each pixel; z is the multidimensional feature vector obtained in step 5.3; Conv It is a one-dimensional convolution process; b Conv It is the bias, which is initially set to 0 by default and will be automatically adjusted based on the results after each training.

[0039] Step 6: Select the first 80% time series of extreme precipitation data and atmospheric circulation index data in step 5.1 as the training set, build and train the model, and use the mean square error (MSE) as the optimization target to reflect the square error between the predicted value and the true value.

[0040] In this embodiment, a time series from 1961 to 2010 is selected, the model is set to perform 100 rounds of training, and an early stopping mechanism is embedded. When the MSE does not decrease within 10 consecutive training rounds, the training is automatically terminated. Through the above steps, the deep learning model driven by multi-source data is fully utilized to complete the simulation and prediction of complex spatiotemporal patterns of extreme precipitation.

[0041] Step 7: Calculate the correlation coefficient R of the predicted value of each pixel during the validation period (i.e., 2011-2022) to evaluate the accuracy of the simulation results. The correlation coefficient can be calculated using formula (10): (10) In the formula, R is the correlation coefficient for each pixel; M is the sample size, which is 120 in this case; X m and Y m For each pixel m The true value and predicted value of samples; and is the mean of the true value and the predicted value for each pixel.

[0042] The spatiotemporal prediction accuracy of the model in the Yalong River Basin is as follows: Figure 3 As shown in the figure, the average value of the correlation coefficient of all pixels is 0.82, and the overall accuracy of the model is good. Select the pixel in row 20 and column 115, read the true value and predicted value of the pixel at each time step in the validation period, compare the true value and predicted value in the prediction period, and perform Pearson correlation calculation, as shown in the figure. Figure 4 As shown in Figure 2, the time series correlation coefficient reaches 0.84. In addition, the spatial distribution of the correlation coefficient of the model using only the atmospheric circulation index for prediction under the same model structure and parameters is shown in Figure 2. Figure 5 The average correlation coefficient of all pixels is 0.75, and its accuracy is worse than that of the scheme that uses both atmospheric circulation index and terrain characteristics.

[0043] It can be seen that the prediction model based on the fusion of atmospheric circulation index and terrain DEM data can accurately capture extreme precipitation events and has a good prediction ability for extreme precipitation.

[0044] The above is only an example of the implementation of the present invention in a specific river basin, but the protection scope of the present invention is not limited thereto. If a person skilled in the art is inspired by it and designs a method and embodiment similar to the technical solution without creativity without departing from the purpose of the present invention, they should all fall within the protection scope of the present invention.

Claims

1. A method for simulating extreme precipitation using a CNN-LSTM model based on multi-source data fusion, characterized in that: The steps include: Step S1: Obtain multi-source data of the target area, including historical precipitation raster data, atmospheric circulation index time series data, and terrain DEM data; Step S2: Calculate the extreme precipitation index Rx5day in the target area and save it as a raster file with the same coordinate system as the source file. The calculation formula is as follows: （1）； （2）； In the above formula, S i is the sum of precipitation for 5 consecutive days starting from the 𝑖th day, P i For the i The daily precipitation of the day, Rx5day The maximum precipitation for 5 consecutive days each month; Step S3: using the extreme gradient boosting algorithm XGBoost to screen the key atmospheric circulation indexes that have the greatest impact on extreme precipitation in the target area; Step S4: resampling the terrain DEM data to align it with the grid of the precipitation raster data, while ensuring that the terrain DEM data and the precipitation raster data have the same coordinate system; Step S5: Min-Max normalization is performed on the above multi-source data, and a spatiotemporal prediction model integrating convolutional neural network CNN and long short-term memory network LSTM is constructed to perform spatiotemporal prediction of extreme precipitation using multi-source data. The process can be expressed as follows: （3）； In the above formula, P extreme is the simulated value of extreme precipitation, X P , X Cn , X T They are the input historical extreme precipitation data, key atmospheric circulation index data and terrain DEM data. f It is a spatiotemporal prediction model that integrates convolutional neural network (CNN) and long short-term memory (LSTM) network driven by multi-source data. Step S6: Select the top 80% time series of extreme precipitation data and atmospheric circulation index data as training sets, train the model, use the mean square error (MSE) as the optimization target, calculate the square error between the predicted value and the true value, and through multiple rounds of training, fully utilize the deep learning model driven by multi-source data to complete the simulation and prediction of complex spatiotemporal patterns of extreme precipitation; Step S7: Calculate the correlation coefficient R of each pixel prediction value to evaluate the accuracy of the simulation results.

2. The method for simulating extreme precipitation using a CNN-LSTM model based on multi-source data fusion according to claim 1, characterized in that: In step S3, the key atmospheric circulation indexes that have the greatest impact on the extreme precipitation in the target area are screened, specifically including: Step S3.1: Use ArcGIS pixel statistics method to calculate the mean time series of extreme precipitation in the target area. The calculation formula is as follows: （4）； In the above formula, For the target area in the time period T The average extreme precipitation value in N is the total number of pixels, d T，j For raster files at time Segment T Medium pixel j The value of Step S3.2: Combine the time series of atmospheric circulation index and extreme precipitation, and use the first 80% of the time series as the training set and the rest as the test set. Use the extreme gradient boosting algorithm XGBoost to learn the impact of each index on extreme precipitation. The algorithm calculates the importance of each index based on the improvement of the prediction results of each index in the split node, and finally sorts them according to the importance, and selects the four indices with the largest contribution as the key atmospheric circulation indices; the importance is calculated by the following formula: （5）； In the above formula, I k Index k The importance of Tr is the set of all decision trees in the extreme gradient boosting algorithm XGBoost. St For Trees t The set of all split nodes in , f tr,s Number tr At the split node s The features selected above; α ( ) is the indicator function, when f tr,s = k hour, α ( f tr,s = k )=1, otherwise 0.

3. The method for simulating extreme precipitation using a CNN-LSTM model based on multi-source data fusion according to claim 2, characterized in that: In step S5, a CNN-LSTM model is constructed, which specifically includes: Step S5.1: Copy the atmospheric circulation index data in the spatial dimension to align it with the pixel, merge the extreme precipitation data, atmospheric circulation index data, and terrain DEM data in the channel dimension to form a multi-channel input feature tensor, use a one-dimensional convolutional neural network CNN to extract the local spatial feature map of each pixel in the spatial dimension, and fuse the multi-channel information of extreme precipitation, atmospheric circulation index, and terrain DEM data to generate a discriminative spatial feature map, where the convolution process is expressed by the following formula: （6）； In the formula, F is the output feature map of the convolutional layer, X is the input feature map, W is the convolution kernel weight, b is the bias, G is the activation function; Step S5.2: The spatial feature sequence output by the convolutional neural network CNN is input into the long short-term memory neural network LSTM for time series modeling. The long short-term memory neural network LSTM selectively memorizes important information and forgets irrelevant information through a gating mechanism, thereby capturing the dynamic change trend in the past time and outputting the hidden state vector at the last moment. This vector condenses the comprehensive information of the entire time series as the basis for subsequent predictions; the process is expressed by the following formula: （7）； In the above formula, h t is the hidden state at the current time step, x t is the input spatial feature map of the current time step, h t-1 is the hidden state of the previous time step; Step S5.3: The hidden state vector output by the long short-term memory neural network LSTM is mapped back to the spatial feature through the fully connected layer and associated with each pixel position, where the fully connected layer mapping is expressed by the following formula: （8）； In the above formula, z is the feature vector output by the fully connected layer, h T is the last hidden state of LSTM, W fc and b fc are the weights and biases of the fully connected layer; Step S5.4: The mapped features are further converted into extreme precipitation Rx5day prediction values ​​consistent with the pixel distribution through the final convolution layer, and then the spatial distribution prediction of future time points is generated. The process is expressed by the following formula: （9）； In the formula, Y is the predicted extreme precipitation value of the pixel; z is the multidimensional feature vector obtained in step 5.3; Conv It is a one-dimensional convolution process; b Conv For bias.

4. The method for simulating extreme precipitation using a CNN-LSTM model based on multi-source data fusion according to claim 3, characterized in that: In step S7, the correlation coefficient R is calculated by the following formula: （10）； In the formula, R is the correlation coefficient for each pixel; M is the sample size; X m and Y m For each pixel m The true value and predicted value of samples; and is the mean of the true and predicted values ​​for each pixel.

Images