A method for constructing a rainfall field spatial feature intelligent identification model

Abstract

The application discloses a kind of construction methods of rainfall field spatial feature intelligent identification model, comprising the following steps: step 1, collect rainfall data of research area;Step 2, build class real rainfall field;Step 3, determine and quantify influencing factor;Step 4, build rainfall field spatial feature intelligent identification model;Step 5, model training and verification.The method disclosed in the application fully considers a variety of factors that have significant influence on rainfall spatial distribution, which helps to improve the adaptability and migration ability of the model, and enhance its generalization performance;It is also helpful to reveal the change rule and response mechanism of the model in identifying rainfall spatial features under different conditions.Through the method disclosed in the application, the spatial information of multi-source precipitation data can be effectively fused, combined with spatial scale, rainfall order and terrain change influencing factor, to build a rainfall field spatial feature identification model with strong robustness and interpretability, providing reliable technical support for fine meteorological monitoring, disaster warning and other applications.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of meteorological monitoring, and particularly relates to a construction method of a rainfall field spatial feature intelligent identification model. BACKGROUND

[0002] With the development of remote sensing technology, multi-source rainfall products provide greater spatial coverage and higher spatio-temporal resolution, and each rainfall product contains a lot of spatial feature information of rainfall field. Reasonably utilizing the available spatial features of multi-source rainfall products is of great significance for accurately identifying and monitoring the spatial features of rainfall field.

[0003] In the field of meteorological monitoring, accurately identifying and monitoring the spatial features of rainfall field is of great significance for disaster warning, weather forecasting and other aspects. However, the traditional method often relies on artificially set assumptions, and has limitations in the face of complex and changeable natural environment. For example, the method based on statistical model usually needs a large amount of historical data to establish a reliable model, which is difficult to achieve in areas with scarce data or dramatic changes. In recent years, with the development of remote sensing technology and computer science, especially the progress of deep learning technology, it provides new possibilities for intelligent identification of rainfall field spatial features. For example, deep learning models, especially convolutional neural networks (CNN) and recurrent neural networks (RNN), have made remarkable achievements in image processing and time series analysis. These models can automatically learn complex patterns from a large amount of data without setting detailed assumptions by humans. However, although these models perform well on specific tasks, there are still some challenges in combining with rainfall products and processing data such as rainfall field spatial features, including how to effectively capture global spatio-temporal features and how to deal with the influence of different spatial scales, rainfall types and terrain changes.

[0004] Therefore, in order to more effectively identify the spatial features of rainfall field under different conditions, the intelligent identification method based on deep learning is gradually becoming a research hotspot, which can provide more comprehensive and accurate rainfall information, and how to construct a rainfall field spatial feature intelligent identification model is a technical problem urgently to be solved in the field. SUMMARY

[0005] The application aims to provide a construction method of a rainfall field spatial feature intelligent identification model to solve the above technical problems.

[0006] To achieve the above-mentioned purpose, the application provides the following technical scheme:

[0007] The application discloses a construction method of a rainfall field spatial feature intelligent identification model, which comprises the following steps:

[0008] Step 1, collecting rainfall data in the study area: collecting rainfall data in the study area, including multi-source satellite remote sensing quantitative estimation rainfall product data and rainfall data of rain gauges in the study area; the rainfall data of rain gauges in the study area includes the position coordinates of each rain gauge in the study area and its measured rainfall;

[0009] Step 2, constructing a real rainfall field: first, divide the study area, that is, in the range of the study area, divide the study area into training area, verification area and migration area according to rain gauge density and small watershed as unit; second, according to the boundary of each divided area, divide each area into regular rectangular grid, define the grid resolution, and process the multi-source rainfall product into grid data with the same resolution; then construct a real rainfall field according to the measured rainfall of rain gauges in the training area, that is, the measured field, and then determine the rainfall center position and rainfall space gradient of each rainfall product by using the constructed real rainfall field; finally, generate rainfall center position matrix A and rainfall space gradient matrix B;

[0010] Step 3, determining and quantifying the influence factor: considering the factors that significantly affect the spatial monitoring performance of multi-source rainfall products, taking spatial scale, rainfall order and terrain change as influence factors, quantifying the spatial scale, rainfall order and terrain change influence factors in each unit period respectively, forming spatial scale influence factor input matrix C, rainfall order influence factor input matrix D and terrain change influence factor input matrix F;

[0011] Step 4, constructing rainfall field spatial feature intelligent identification model: the rainfall field spatial feature intelligent identification model is a model framework based on generative adversarial network GAN, which is composed of generator and discriminator, and the ViT algorithm is integrated into the generator and discriminator of the generative adversarial network GAN; the input of the model is the rainfall center position matrix, the rainfall space gradient matrix and the influence factor matrix; the output of the model is the identified rainfall center position and rainfall space gradient, that is, the identified field; the generator and discriminator are alternately trained by using the loss function of the generative adversarial network GAN to make them optimal; the optimization goal of the generator is to make the distribution of generated data approach the real data respectively, and the optimization goal of the discriminator is to maximize its discrimination ability of generated data and real data;

[0012] Step 5, model training and verification: set the input mode of influence factors, including two modes:

[0013] Mode one: taking spatial scale, rainfall order, terrain change influence factor as sample state and rainfall center position and rainfall space gradient as input of rainfall field spatial feature intelligent identification model, that is, stacking rainfall center position matrix A, rainfall space gradient matrix B, spatial scale matrix C, rainfall order matrix D and terrain change matrix F into a tensor to input the model;

[0014] In the second way, a plurality of models are classified and trained respectively according to a spatial scale, a rainfall level, and a terrain change factor. Specifically, 1) a spatial scale is used as a condition to input a model, a spatial scale matrix C, a rainfall center position matrix A, and a rainfall spatial gradient matrix B are stacked into a tensor to input a model for separate training; 2) a rainfall level is used as a condition to input a model, a rainfall level matrix D, a rainfall center position matrix A, and a rainfall spatial gradient matrix B are stacked into a tensor to input a model for separate training; and 3) a terrain change is used as a condition to input a model, a terrain change matrix F, a rainfall center position matrix A, and a rainfall spatial gradient matrix B are stacked into a tensor to input a model for separate training.

[0015] The rainfall field spatial feature intelligent recognition model is input according to the above two input modes respectively, and the model output is the recognized rainfall center position and rainfall spatial gradient. The model is trained and verified and analyzed.

[0016] Further, the specific process of dividing the study area into a training area, a verification area, and a migration area in step 2 is as follows: an area with a rainfall station density less than or equal to 25 km 2 / station is regarded as a rainfall station dense area; an area with a rainfall station density greater than 25 km 2 / station is regarded as a rainfall station sparse area; a small watershed is a closed catchment unit divided by a watershed ridge and a river outlet; a part of the rainfall station dense small watershed is selected as the model training area, that is, the training area; another part of the rainfall station dense small watershed is selected as the model verification area, that is, the verification area; and the rainfall station sparse small watershed is selected as the model migration area, that is, the migration area.

[0017] Further, the specific process of constructing a real rainfall field according to the measured rainfall of the rainfall station in the training area in step 2 is as follows: the position coordinates of each rainfall station are (x a , y a )(a = 1, 2, 3…), the measured rainfall is Z a , and an inverse distance weighted interpolation method is used to construct a real rainfall field. For any grid point (x g , y g ), the estimated rainfall is calculated by using the inverse distance weighted interpolation method. The formula is as follows:

[0018]

[0019] In the formula, Z a is the measured rainfall of the a rainfall station; ω a is the a rainfall station for the grid point (x g , y gweight of the rain gauge; d a distance from the a-th rain gauge to the grid point (x g , y g ); r is the number of rain gauges participating in the calculation, determined by the search radius or the number of nearest neighbors; p is the power parameter, which is a constant.

[0020] Further, the specific process of determining the rainfall center position and the rainfall field spatial gradient of each rainfall product using the constructed pseudo-real rainfall field in step 2 is as follows:

[0021] The rainfall center position is determined using the constructed pseudo-real rainfall field, and the grid point with the maximum rainfall is the rainfall center position. The formula is as follows:

[0022]

[0023] In the formula, The corresponding position (x c , y c ) is the rainfall center position.

[0024] In n multi-source rainfall products S x (x = 1, 2, …, i, …, n), take rainfall product S i as an example, and the rainfall center position of rainfall product S i is used as the available information for calculation; for each unit time period, set the grid points with the rainfall center position (x i , y i ) of rainfall product S i as the center and the allowed error distance δ as the radius to 1, and the remaining grid points to 0, thereby generating the rainfall center position matrix A i . The formula is as follows:

[0025]

[0026] In the formula, k (x,y) is the setting value of the unit time period coordinate (x, y) grid point;

[0027] The rainfall center position matrix of n multi-source rainfall products is calculated respectively, and then stacked into an n-channel rainfall center position matrix A;

[0028] Then the rainfall field spatial gradient is determined. The rainfall field spatial gradient is the grid rainfall of the rainfall product, and after normalization, it is obtained.

[0029] In n multi-source rainfall products S x (x = 1, 2, …, j, …, n), take rainfall product S j as an example, and the rainfall center position of rainfall product S jThe rainfall field is subjected to Min-Max normalization processing, and a rainfall product S is obtained j The spatial gradient of the rainfall field is obtained, and a rainfall spatial gradient matrix B is generated j The Min-Max normalization formula is as follows:

[0030]

[0031] In the formula, h (x,y) is the setting value of the coordinate (x, y) grid point; Z j(x,y) is the rainfall product S j at the coordinate (x, y) grid point; Z min is the rainfall product S j at the minimum rainfall in the region; Z max is the rainfall product S j at the maximum rainfall in the region;

[0032] The rainfall spatial gradient matrix of each of the n multi-source rainfall products is obtained, and then stacked into an n-channel rainfall spatial gradient matrix B.

[0033] Further, the specific process of quantifying the spatial scale, rainfall intensity, and terrain change influence factors in each unit period in step 3 to form the spatial scale influence factor input matrix C, the rainfall intensity influence factor input matrix D, and the terrain change influence factor input matrix F is as follows:

[0034] First, the spatial scale is divided into three categories: small, medium, and large. The range below 3km×3km is classified as small scale, and the grid point is set to W s =1; the range between 3km×3km and 10km×10km is classified as medium scale, and the grid point is set to W s =2; and the range above 10km×10km is classified as large scale, and the grid point is set to W s =3; the grid matrix formed is subjected to normalization processing to form the spatial scale influence factor input matrix C.

[0035] Secondly, the rainfall intensity is divided into small rain, moderate rain, heavy rain, heavy rain, and above, and the daily rainfall less than 10mm is classified as small rain, and the grid point is set to W p =1; the daily rainfall between 10mm and 25mm is classified as moderate rain, and the grid point is set to W p =2; the daily rainfall between 25mm and 50mm is classified as heavy rain, and the grid point is set to W p =3; the daily rainfall between 50mm and 100mm is classified as heavy rain, and the grid point is set to W p =4; and the daily rainfall above 100mm is classified as heavy rain and above, and the grid point is set to W p= 5; the formed grid matrix is normalized to form the rainfall intensity influence factor input matrix D;

[0036] Again, the terrain change is quantified by the slope ratio S, and the grid slope ratio calculation formula is as follows:

[0037]

[0038] In the formula, and is the gradient of the elevation in the x direction and the y direction;

[0039] The slope ratio of each grid point is calculated by the formula to quantify the terrain change, and the formed grid matrix is normalized to form the terrain change influence factor input matrix F.

[0040] Further, the specific process of alternating training the generator and the discriminator to reach the optimum by using the loss function of the generative adversarial network GAN in step 4 is as follows:

[0041] The loss function of the generative adversarial network is as follows:

[0042]

[0043] In the formula, V(D, G) is the cross-entropy loss; z is the tensor input to the generator; x is the real data tensor; G(z) is the tensor generated by the generator; D(x) is the probability that the discriminator judges the real rainfall field formed by the real data tensor to be real; D(G(z)) is the probability that the discriminator judges the tensor generated by the generator to be real;

[0044] Wherein, the real data tensor is stacked by two channels, including the rainfall center position hit rate HR and the spatial structure similarity index SSIM; based on the measured field, when the straight line distance between the measured field and the rainfall center of the identified field is less than a certain allowable error distance, it is counted as a hit, and its formula is as follows:

[0045]

[0046] In the formula, I i is the rainfall center hit indication function; d i is the straight line distance between the measured field and the rainfall center of the identified field in the i period; δ is the allowable error distance; N is the total number of samples;

[0047] The spatial structure similarity index SSIM is used to quantify the similarity of the spatial gradient of the identified field and the measured field, and its calculation formula is as follows:

[0048]

[0049] In the formula, x is the normalized measured field; y is the normalized identified field; μx and μ y These are the means of two normalized rainfall fields, respectively; σ x and σ y σ represents the variance of the two normalized rainfall fields. xy Let C1 and C2 be the covariance of the two normalized rainfall fields; C1 and C2 are constants to prevent the denominator from being zero.

[0050] Furthermore, the specific process of training and validating the model described in step 5 is as follows:

[0051] For the case where the model input is performed according to Method 1, in the training area, the measured rainfall from rain gauges is used to construct a near-real rainfall field to train the intelligent recognition model of rainfall field spatial features; in the validation area, a small number of rain gauges are extracted to construct a sparse rain gauge network, and the intelligent recognition model of rainfall field spatial features is applied; the remaining rain gauges are used to validate the model output results, and the Nash coefficient (NSE) is used to evaluate the model performance.

[0052] For the case where the model is input according to method two, three models under different influencing factor conditions are trained in the training area and then the models are validated. A small number of rain gauge stations are extracted in the validation area to construct a sparse rain gauge network. The intelligent identification model of the spatial characteristics of the rainfall field under different influencing factor conditions is applied. The remaining rain gauge stations are used to validate the model results. The output results of the model under different influencing factor conditions are analyzed, and the Nash coefficient (NSE) is used to evaluate the model effect.

[0053] The formula for calculating the Nash coefficient (NSE) is as follows:

[0054]

[0055] In the formula, Q b To verify the observation value of one of the s remaining rain gauges after sampling within the area, rain gauge station b; Q v This is the model output value for the grid point where rain gauge b is located; This represents the average of the observations from the remaining s rain gauge stations;

[0056] When NSE ≥ 0.5, it indicates that the model has passed validation, and the model can be applied to the migration area to achieve accurate identification of the rainfall field in sparse rain gauge areas; when NSE < 0.5, it indicates that the model has not passed validation and needs to be retrained for optimization and improvement.

[0057] The method has the beneficial effects that: the method uses the spatial structure characteristics of the multi-source rainfall product as input to drive the deep learning model, so that the model output approximates the spatial distribution characteristics of the measured rainfall field as much as possible, thereby realizing intelligent identification and restoration of the rainfall field structure. The method fully considers various factors that have a significant influence on the spatial distribution of rainfall, which helps to improve the adaptability and migration ability of the model under different regional and climate conditions, enhance the generalization performance, and help to reveal the change rule and response mechanism of the model in identifying rainfall spatial characteristics under different conditions. Through the method, the spatial information of multi-source rainfall data can be effectively fused, combined with the spatial scale, rainfall level and terrain change influencing factors, to construct a rainfall field spatial feature identification model with strong robustness and interpretability, thereby providing reliable technical support for fine meteorological monitoring, disaster warning and other applications.

[0058] The application will be further described in detail below with reference to the accompanying drawings and specific embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0059] Figure 1 The method flowchart of the application is shown in the following figure:

[0060] Figure 2 The real rainfall field constructed for the rain gauge station is shown in the following figure:

[0061] Figure 3 The rainfall center position matrix data of the rainfall product S i is shown in the following figure:

[0062] Figure 4 The rainfall field spatial gradient matrix data of the rainfall product S j is shown in the following figure:

[0063] Figure 5 The intelligent identification model structure diagram of the rainfall field spatial feature is shown in the following figure: DETAILED DESCRIPTION

[0064] The application discloses a method for constructing an intelligent identification model of rainfall field spatial features, as shown in the following figure, which comprises the following steps: Figure 1

[0065] Step 1, collecting rainfall data in the research area: collecting rainfall data in the research area, including multi-source high-temporal and spatial resolution, high-time efficiency satellite remote sensing quantitative estimation rainfall product data and rainfall data of rain gauges in the research area; wherein the rainfall data of rain gauges in the research area includes the position coordinates and the measured rainfall of each rain gauge in the research area.

[0066] ​Step 2, Constructing the pseudo-real rainfall field: First, divide the study area, that is, within the scope of the study area, based on the density of rain stations and small watersheds as units, the study area is divided into training area, verification area, and migration area. Specifically, the area with a rain station density less than or equal to 25 km 2 / station is considered a rain station dense area; the area with a rain station density greater than 25 km 2 / station is considered a rain station sparse area; the small watershed is a closed catchment unit divided by the watershed ridge and river outlet; select a part of the rain station dense small watershed as the model training area, i.e. the training area; another part of the rain station dense small watershed as the model verification area, i.e. the verification area; the rain station distribution sparse small watershed is the model migration area, i.e. the migration area.

[0067] Second, according to the boundaries of each divided area, each area is divided into a regular rectangular grid, with a grid resolution of X km x X km, and the multi-source rainfall product is processed into grid data with the same resolution. The coordinates of each grid point are (x g , y g ), (g = 1, 2, …, m).

[0068] According to the measured rainfall of the rain stations in the training area, a pseudo-real rainfall field is constructed. Let the position coordinates of each rain station be (x a , y a ) (a = 1, 2, 3, …), and the measured rainfall be Z a . The inverse distance weighted interpolation method is used to construct the pseudo-real rainfall field (i.e. the measured field), as shown in Figure 2 . For any grid point (x g , y g ), the estimated rainfall is calculated using the inverse distance weighted interpolation method, as follows:

[0069]

[0070] In the formula, Z a is the measured rainfall of the a-th rain station; ω a is the weight of the a-th rain station for the grid point (x g , y g ); d a is the distance from the a-th rain station to the grid point (x g , y g ); r is the number of rain stations participating in the calculation, determined by the search radius or the number of nearest neighbors; p is the power parameter, which is a constant.

[0071] The constructed pseudo-real rainfall field is used to determine the rainfall center position, and the grid point with the maximum rainfall is the rainfall center position, as follows:

[0072] ​

[0073] In the formula, The corresponding position (x c , y c ) is the rainfall center position.

[0074] In n multi-source rainfall products S x (x = 1, 2, …, i, …, n), taking rainfall product S i as an example, the rainfall center position of rainfall product S i is calculated as available information. For each unit time period, the grid point with the rainfall center position (x i , y i ) of rainfall product S i as the center and the allowable error distance δ as the radius is set to 1, and the remaining grid points are set to 0, thereby generating a rainfall center position matrix A i As shown in Figure 3 , the formula is as follows:

[0075]

[0076] In the formula, k (x,y) is the setting value of the unit time period coordinate (x, y) grid point.

[0077] The rainfall center position matrix of n multi-source rainfall products is calculated respectively, and then stacked into an n-channel rainfall center position matrix A.

[0078] Then determine the rainfall field spatial gradient, which is the grid rainfall of the rainfall product, and do normalization processing to obtain

[0079] In n multi-source rainfall products S x (x = 1, 2, …, j, …, n), taking rainfall product S j as an example, the rainfall field rainfall of rainfall product S j is subjected to Min-Max normalization processing, and the rainfall field spatial gradient of rainfall product S j is obtained, thereby generating a rainfall spatial gradient matrix B j As shown in Figure 4 , the Min-Max normalization formula is as follows:

[0080]

[0081] In the formula, h (x,y) is the setting value of the coordinate (x, y) grid point; Z j(x,y) is the rainfall of rainfall product S j at the coordinate (x, y) grid point; and Z min is the rainfall of rainfall product S jMinimum rainfall in the region; Z 　ax Rainfall product S j Maximum rainfall in the region.

[0082] The rainfall spatial gradient matrix of n multi-source rainfall products is obtained respectively, and then stacked into an n-channel rainfall spatial gradient matrix B.

[0083] Step 3, determine and quantify the influence factor: considering the factors such as spatial scale, rainfall order, and terrain change which have significant influence on the spatial monitoring performance of multi-source rainfall products, taking spatial scale, rainfall order, and terrain change as influence factors, quantifying the spatial scale, rainfall order, and terrain change in each unit period, the specific process is as follows:

[0084] Firstly, the spatial scale can be divided into three categories: small, medium and large scale, and the range below 3km×3km is divided into small scale, and the grid point is set as W s =1; the range between 3km×3km to 10km×10km is medium scale, and the grid point is set as W s =2; the range above 10km×10km is large scale, and the grid point is set as W s =3. The formed grid matrix is normalized to form the spatial scale influence factor input matrix C.

[0085] Secondly, the rainfall order can be divided into small rain, moderate rain, heavy rain, rainstorm and above, and the daily rainfall less than 10mm is small rain, and the grid point is set as W p =1; the daily rainfall between 10mm and 25mm is moderate rain, and the grid point is set as W p =2; the daily rainfall between 25mm and 50mm is heavy rain, and the grid point is set as W p =3; the daily rainfall between 50mm and 100mm is rainstorm, and the grid point is set as W p =4; the daily rainfall above 100mm is rainstorm and above, and the grid point is set as W p =5. The formed grid matrix is normalized to form the rainfall order influence factor input matrix D.

[0086] Thirdly, the terrain change is quantified by using the slope ratio S, and the grid point slope ratio calculation formula is as follows:

[0087]

[0088] In the formula, and are the gradients of the elevation in the x direction and the y direction.

[0089] The terrain change is quantified by calculating the slope ratio of each grid point, and the formed grid matrix is normalized to form a terrain change influence factor input matrix F.

[0090] Step 4, constructing a rainfall field spatial feature intelligent identification model: a rainfall field spatial feature intelligent identification model is constructed, which is a model framework based on a generative adversarial network (GAN). The generative adversarial network (GAN) is composed of a generator and a discriminator. The Vision Transformer (ViT) algorithm is integrated into the generator and the discriminator of the generative adversarial network (GAN), solving the limitations of traditional architecture and training instability. The mode input is a rainfall center position matrix, a rainfall spatial gradient matrix and an influence factor matrix; the model output is the identified real rainfall center position and rainfall spatial gradient (i.e. the identification field). Among them, the display form of the rainfall center position is consistent with the form of the rainfall center position of a certain rainfall product (such as S i ), and the display form of the rainfall spatial gradient is consistent with the form of the rainfall spatial gradient of a certain rainfall product (such as S j ). The structure of the rainfall field spatial feature intelligent identification model is shown in Figure 5 . The generator and the discriminator are alternately trained to reach the optimal state by using the loss function of the generative adversarial network. The optimization goal of the generator is to make the distribution of the generated data approach the real data respectively, while the optimization goal of the discriminator is to maximize its discrimination ability between the generated data and the real data.

[0091] The loss function of the generative adversarial network is as follows:

[0092]

[0093] In the formula, V(D, G) is the cross-entropy loss; z is the tensor input to the generator; x is the real data tensor; G(z) is the tensor generated by the generator; D(x) is the probability that the discriminator judges the tensor formed by the real rainfall field to be real; D(G(z)) is the probability that the discriminator judges the tensor generated by the generator to be real.

[0094] Among them, the real data tensor is stacked by two channels, including the identification field rainfall center position hit rate HR and the spatial structure similarity index SSIM. Taking the measured field as the benchmark, when the straight line distance between the measured field and the identification field rainfall center is less than a certain allowable error distance, it is counted as a hit, and its formula is as follows:

[0095]

[0096] In the formula, I i is the rainfall center hit indication function; d i is the straight line distance between the measured field and the identification field rainfall center in the i period; δ is the allowable error distance; and N is the total number of samples.

[0097] The spatial structure similarity index SSIM is used to quantify the similarity of the identified field and the measured field rainfall spatial gradient, SSIM comprehensively considers the spatial distribution mean, variance and correlation, and is suitable for similarity evaluation of multi-scale rainfall field, and the calculation formula is:

[0098]

[0099] In the formula, x is the normalized measured field; y is the normalized identified field; μ x and μ y are the mean values of the two normalized rainfall fields respectively; σ x and σ y are the variances of the two normalized rainfall fields respectively; σ xy is the covariance of the two normalized rainfall fields; C1 and C2 are constants to prevent the denominator from being zero.

[0100] Step 5, model training and verification: first set the influence factor input mode, including two modes, which are:

[0101] Mode one, the spatial scale, rainfall intensity, and terrain change influence factors are input as sample states and rainfall center position and rainfall spatial gradient to the rainfall field spatial feature intelligent identification model, that is, the rainfall center position matrix A, the rainfall spatial gradient matrix B, the spatial scale matrix C, the rainfall intensity matrix D and the terrain change matrix F are stacked into a tensor to input the model.

[0102] Mode two, the spatial scale, rainfall intensity, and terrain change influence factors are used as classification conditions to train multiple models. Specifically: 1) input the model with spatial scale as the condition, stack the spatial scale matrix C and the rainfall center position matrix A, the rainfall spatial gradient matrix B into a tensor to input the model for separate training; 2) input the model with rainfall intensity as the condition, stack the rainfall intensity matrix D and the rainfall center position matrix A, the rainfall spatial gradient matrix B into a tensor to input the model for separate training; 3) input the model with terrain change as the condition, stack the terrain change matrix F and the rainfall center position matrix A, the rainfall spatial gradient matrix B into a tensor to input the model for separate training.

[0103] According to the above two input modes, the rainfall field spatial feature intelligent identification model is input, and the model output is the identified rainfall center position and rainfall spatial gradient. The model is trained and verified.

[0104] For the case of model input according to mode one, in the training area, the measured rainfall of the rainfall stations is used to construct a real rainfall field for training the intelligent recognition model of rainfall field spatial characteristics; in the verification area, a small number of rainfall stations are extracted to construct a sparse rainfall station network (more than 25km 2 / station), which is applied to the intelligent recognition model of rainfall field spatial characteristics, and the remaining rainfall stations are used to verify the model output results.

[0105] The Nash coefficient NSE is used to evaluate the model effect, and the calculation formula is as follows:

[0106]

[0107] In the formula, Q b is the observation value of a rainfall station b in the remaining s rainfall stations in the verification area; Q v is the model output value of the grid point where the rainfall station b is located; and Q is the average value of the observation values of the remaining s rainfall stations.

[0108] When NSE≥0.5, it indicates that the model has good simulation effect, that is, the model passes the verification, and then the model is applied to the migration area to realize accurate recognition of the rainfall field in the sparse rainfall station area; when NSE<0.5, it indicates that the model performance is poor and does not pass the verification, and needs to be retrained for optimization and improvement. This input mode is conducive to training the intelligent recognition model of rainfall field spatial characteristics in the dense rainfall station area and migrating the model to the sparse rainfall station area.

[0109] For the case of model input according to mode two, three models under different influencing factor conditions (spatial scale, rainfall order, and terrain change) are trained in the training area, and then the model verification is performed. In the verification area, a small number of rainfall stations are extracted to construct a sparse rainfall station network (more than 25km 2 / station), different influencing factor conditions are applied to the intelligent recognition model of rainfall field spatial characteristics, and the remaining rainfall stations are used to verify the model results. The output results of the models under different influencing factor conditions are analyzed, and the Nash coefficient NSE is used to evaluate the model effect. Similarly, when NSE≥0.5, it indicates that the model has good simulation effect, that is, the model passes the verification, and then the model can be further applied to the migration area; when NSE<0.5, it indicates that the model performance is poor and does not pass the verification, and needs to be retrained for optimization and improvement. Thus, the influence mechanism of spatial scale, rainfall order, and terrain change on the recognition of rainfall center position and spatial gradient by the model can be systematically revealed. This comparison is not only helpful for identifying the performance difference of the model under different conditions, but also helpful for analyzing the rules of the model recognizing spatial characteristics under different conditions.

[0110] It should be pointed out finally that the above description is only used to explain the technical solutions of the present application but not to limit the present application. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or equivalently replaced without departing from the spirit and scope of the present application.

Claims

1. A method for constructing a rainfall field spatial feature intelligent identification model, characterized in that, The method comprises the following steps: Step 1, collecting rainfall data of the study area: collecting rainfall data of the study area, including multi-source satellite remote sensing quantitative estimation rainfall product data and rainfall data of rain gauges in the study area; the rainfall data of the rain gauges in the study area includes the position coordinates of each rain gauge in the study area and the measured rainfall thereof; Step 2, constructing a pseudo-real rainfall field: first, dividing the study area, that is, in the range of the study area, dividing the study area into training area, verification area and migration area according to the density of rain gauges and small watersheds as units; secondly, according to the boundaries of each divided area, dividing each area into a regular rectangular grid, defining the grid resolution, and processing the multi-source rainfall products into grid data of the same resolution; then constructing a pseudo-real rainfall field according to the measured rainfall of the rain gauges in the training area, that is, a measured field, and then determining the rainfall center position and rainfall space gradient of each rainfall product by using the constructed pseudo-real rainfall field, and finally generating a rainfall center position matrix A and a rainfall space gradient matrix B; Step 3, determining and quantifying the influence factors: considering the factors that significantly affect the spatial monitoring performance of multi-source rainfall products, taking spatial scale, rainfall order and terrain change as influence factors, respectively quantifying the spatial scale, rainfall order and terrain change influence factors in each unit period to form a spatial scale influence factor input matrix C, a rainfall order influence factor input matrix D and a terrain change influence factor input matrix F; Step 4, constructing a rainfall field spatial feature intelligent identification model: the rainfall field spatial feature intelligent identification model is a model framework based on a generative adversarial network GAN, the generative adversarial network GAN is composed of a generator and a discriminator, the ViT algorithm is integrated into the generator and the discriminator of the generative adversarial network GAN, the model input is the rainfall center position matrix, the rainfall space gradient matrix and the influence factor matrix; the model output is the identified rainfall center position and rainfall space gradient, that is, the identified field; the generator and the discriminator are alternately trained to reach the optimal by using the loss function of the generative adversarial network GAN, the optimization goal of the generator is to make the distribution of the generated data approach the real data respectively, and the optimization goal of the discriminator is to maximize the discrimination ability of the generated data and the real data; Step 5, model training and verification: setting the influence factor input mode, including two modes: Mode one: taking the spatial scale, rainfall order and terrain change influence factors as sample states and rainfall center position and rainfall space gradient to input the rainfall field spatial feature intelligent identification model together, that is, stacking the rainfall center position matrix A, the rainfall space gradient matrix B, the spatial scale matrix C, the rainfall order matrix D and the terrain change matrix F into a tensor to input the model; In the second mode, a plurality of models are classified and trained by taking spatial scale, rainfall level, and terrain change influencing factor as classification conditions. Specifically, 1) a model is inputted by taking spatial scale as a condition, and a spatial scale matrix C, a rainfall center position matrix A, and a rainfall spatial gradient matrix B are stacked into a tensor to input the model for separate training; 2) a model is inputted by taking rainfall level as a condition, and a rainfall level matrix D, a rainfall center position matrix A, and a rainfall spatial gradient matrix B are stacked into a tensor to input the model for separate training; and 3) a model is inputted by taking terrain change as a condition, and a terrain change matrix F, a rainfall center position matrix A, and a rainfall spatial gradient matrix B are stacked into a tensor to input the model for separate training. The spatial feature intelligent recognition model is inputted according to the above two input modes, and the model outputs are the recognized rainfall center position and rainfall spatial gradient. The model is trained and verified.

2. The method of claim 1, wherein, The specific process of dividing the study area into training area, verification area and migration area in the study area range based on the rain gauge station density and small watershed unit in step 2 is as follows: the area with rain gauge station density less than or equal to 25km 2 / station is regarded as a rain gauge dense area; and the area with rain gauge station density greater than 25km 2 / station is regarded as a rain gauge sparse area. A small watershed is a closed catchment unit divided by a watershed ridge and a river outlet. A part of small watersheds with dense rain stations is selected as a training area of the model, and another part of small watersheds with dense rain stations is selected as a verification area of the model. Small watersheds with sparse rain stations are used as a migration area of the model.

3. The method of claim 1, wherein the method further comprises: The specific process of constructing the real rainfall field according to the measured rainfall of the rainfall station in the training area in step 2 is as follows: assuming that the position coordinates of each rainfall station are (x a , y a )(a = 1, 2, 3,...), and the measured rainfall is Z a , a real rainfall field is constructed by using the inverse distance weighted interpolation method, and for any grid point (x g , y g ), the estimated rainfall is calculated by using the inverse distance weighted interpolation method, and the formula is as follows: In the formula, Z a is the measured rainfall of the a-th rainfall station; ω a is the weight of the a-th rainfall station for the grid point (x g , y g ); d a is the distance from the a-th rainfall station to the grid point (x g , y g ); r is the number of rainfall stations participating in the calculation, determined by the search radius or the number of nearest neighbors; and p is the power parameter, which is a constant.

4. The method of claim 3, wherein, The specific process of determining the rainfall center position and the rainfall spatial gradient by using the constructed quasi-real rainfall field in step 2 is as follows: The rainfall center position is determined by using the constructed quasi-real rainfall field, and the grid point with the maximum rainfall is the rainfall center position. The formula is as follows: wherein the corresponding position (x c , y c ) is the position of the rainfall center; In n multi-source rainfall products S x (x = 1, 2, …, i, …, n), take the rainfall product S i as an example, the rainfall center position of the rainfall product S i is calculated as available information; for each unit time period, the grid point with the rainfall center position (x i , y i ) of the rainfall product S i as the center and the allowable error distance δ as the radius is set to 1, and the rest of the grid points are set to 0, thereby generating a rainfall center position matrix A i , the formula is as follows: In the formula, k (x,y) is the setting value of the unit period coordinate (x, y) grid point; The rainfall center position matrix of n multi-source rainfall products is calculated respectively, and then stacked into an n-channel rainfall center position matrix A. Then, the rainfall spatial gradient is determined. The rainfall spatial gradient is the grid rainfall of the rainfall product, and the normalized processing is obtained. In n multi-source rainfall products S x (x = 1, 2, …, j, …, n), taking the rainfall product S j as an example, the rainfall field rainfall of the rainfall product S j is Min-Max normalized, and the rainfall field spatial gradient of the rainfall product S j is obtained, thereby generating a rainfall spatial gradient matrix B j , and the Min-Max normalization formula is as follows: wherein h (x,y) is the set value of the coordinate (x, y) grid; Z j(x,y) is the rainfall product S j is the rainfall at the coordinate (x, y) grid; Z min is the rainfall product S j is the minimum rainfall in the region; Z max is the rainfall product S j is the maximum rainfall in the region; The rainfall spatial gradient matrix of n multi-source rainfall products is calculated respectively, and then stacked into an n-channel rainfall spatial gradient matrix B.

5. The method of claim 1, wherein the method further comprises: The specific process of quantifying the spatial scale, rainfall level, and terrain change influencing factor in each unit time period respectively to form the spatial scale influencing factor input matrix C, the rainfall level influencing factor input matrix D, and the terrain change influencing factor input matrix F in step 3 is as follows: First, the spatial scale is divided into three categories of small, medium and large scale, and the range below 3km x 3km is divided into small scale, the grid is set to W s =1; the range between 3km x 3km to 10km x 10km is medium scale, the grid is set to W s =2; the range above 10km x 10km is large scale, the grid is set to W s =3; the formed grid matrix is normalized to form the spatial scale influence factor input matrix C; Secondly, the rainfall is divided into light rain, moderate rain, heavy rain, rainstorm and above, the daily rainfall less than 10mm is light rain, the grid point is set as W p =1; the daily rainfall between 10mm to 25mm is moderate rain, the grid point is set as W p =2; the daily rainfall between 25mm to 50mm is heavy rain, the grid point is set as W p =3; the daily rainfall between 50mm to 100mm is rainstorm, the grid point is set as W p =4; the daily rainfall above 100mm is above rainstorm, the grid point is set as W p =5; the formed grid matrix is normalized to form the rainfall level influence factor input matrix D; Again, the terrain change is quantified by using the slope ratio S. The grid point slope ratio calculation formula is as follows: wherein and are the gradients of the elevation in the x-direction and the y-direction; The terrain change is quantified by calculating the slope ratio of each grid point. The formed grid matrix is normalized to form the terrain change influencing factor input matrix F.

6. The method of claim 1, wherein, The specific process of alternately training the generator and the discriminator to reach the optimum by using the loss function of the generative adversarial network (GAN) in step 4 is as follows: The loss function of the generative adversarial network is as follows: In the formula, V(D, G) is the cross-entropy loss; z is the tensor input to the generator; x is the real data tensor; G(z) is the tensor generated by the generator; D(x) is the probability that the discriminator judges that the tensor formed by the quasi-real rainfall field is real; and D(G(z)) is the probability that the discriminator judges that the tensor generated by the generator is real. Wherein, the real data tensor is stacked by two channels, including the hit rate HR of identifying the center position of the field rainfall and the spatial structure similarity index SSIM; taking the measured field as the benchmark, when the straight-line distance between the measured field and the identified field rainfall center is less than a certain allowable error distance, it is counted as a hit, and its formula is: where I i is the hit indicator function for rainfall centers; d i is the straight-line distance between the i period observed and identified rainfall centers; δ is the allowable error distance; N is the total number of samples; The spatial structure similarity index SSIM is used to quantify the similarity of the spatial gradient of the identified field and the measured field rainfall, and its calculation formula is: where x is the normalized observed field; y is the normalized identified field; μ x and μ y are the mean of the two normalized rainfall fields, respectively; σ x and σ y are the variance of the two normalized rainfall fields, respectively; v xy is the covariance of the two normalized rainfall fields; and C1, C2 are constants to prevent the denominator from being zero.

7. The method of claim 2, wherein the method further comprises: The specific process of training and verifying the model in step 5 is as follows: For the case of model input according to mode one, in the training area, the measured rainfall of the rainfall station is used to construct a real rainfall field for training the rainfall field spatial feature intelligent identification model; in the verification area, a small amount of rainfall stations are extracted to construct a sparse rainfall station network, and the rainfall field spatial feature intelligent identification model is applied, and the remaining rainfall stations are used to verify the model output results, and the Nash coefficient NSE is used to evaluate the model effect; For the case of model input according to mode two, in the training area, three models under different influence factor conditions are trained respectively, and then model verification is performed; in the verification area, a small amount of rainfall stations are extracted to construct a sparse rainfall station network, and the rainfall field spatial feature intelligent identification model under different influence factor conditions is applied respectively, and the remaining rainfall stations are used to verify the model results, the output results of the model under different influence factor conditions are analyzed, and the Nash coefficient NSE is used to evaluate the model effect; The calculation formula of the Nash coefficient NSE is as follows: where Q b is the observation value of one of the remaining s rain gauges b in the validation area after being extracted; Q v is the model output value of the grid point where the rain gauge b is located; is the average of the observation values of the remaining s rain gauges; When NSE≥0.5, it indicates that the model passes the verification, and then the model is applied to the migration area to realize accurate identification of the rainfall field in the sparse area of the rainfall station; When NSE<0.5, it indicates that the model does not pass the verification, and needs to be retrained for optimization and improvement.

Images