Small watershed flood early warning method based on X-band dual-polarization phased array radar

Abstract

The invention discloses a small watershed flood early warning method based on an X-band dual-polarization phased array radar, and relates to the technical field of flood early warning, and the method comprises the following steps: obtaining an X-band dual-polarization phased array radar data set and a rainfall station actual measurement rainfall data set, and building a radar rainfall calculation model through linear fitting; acquiring an X-band dual-polarization phased array radar image set, and performing prediction by adopting a short-term rainfall prediction model to obtain a predicted radar image and corresponding predicted radar data; inputting the predicted radar data into a radar rainfall calculation model to obtain predicted rainfall; and carrying out small watershed flood early warning by adopting the flood forecasting model. According to the method, the X-band dual-polarization phased array radar data and image are combined with the radar rainfall capacity calculation model, the short-term rainfall forecasting model and the flood forecasting model, so that the accuracy of rainfall capacity forecasting is improved, and the reliability of flood early warning is further improved.

Description

Technical Field

[0001] The present application relates to the field of flood warning technology, and in particular to a small watershed flood warning method based on X-band dual-polarization phased array radar. Background Art

[0002] Due to factors such as large population concentrations, lower surface water storage and drainage capacity in cities than in natural areas, shrinking forest and lake areas, illegal river occupation in some areas, and recurring extreme climate events, deaths from floods in small and medium-sized watersheds consistently account for over 60% of all flood-related deaths. Small watersheds, due to their widespread distribution, sudden onset, short runoff generation and convergence times, limited monitoring networks, and lack of data, make it difficult to meet the required spatial and temporal scales for early warning and forecasting. Therefore, research on early warning methods for small watershed floods is needed.

[0003] In the prior art, Chinese patent CN112837508A discloses a flood early warning system, comprising: a first measuring station for monitoring upstream rainfall and riverbed water level; a second measuring station, located in the living camp of the hydropower station, for monitoring rainfall in the area where the living camp is located, thereby estimating the water level and flow of regional tributaries; a third measuring station, located on the dam of the hydropower station, for monitoring the riverbed water level, rainfall, wind direction and temperature in the area where the dam is located; a fourth measuring station, located downstream of the hydropower station, for measuring rainfall, temperature and wind speed in the downstream area; and a main server, connected to the first, second, third and fourth measuring stations to receive information.

[0004] The accuracy of rainfall monitoring is a key factor in small-basin flood warnings, and requires high resolution. However, the above-mentioned existing technologies do not take into account the high-resolution requirements of rainfall monitoring for small-basin flood warnings. The accuracy of rainfall prediction is poor, and thus the reliability of flood warnings is poor. Summary of the Invention

[0005] This application provides a small watershed flood warning method based on X-band dual-polarization phased array radar to solve the problem that the existing technology does not consider the high-resolution requirement of rainfall monitoring for small watershed flood warning, the accuracy of rainfall prediction is poor, and the reliability of flood warning is poor.

[0006] On the one hand, the present application provides a small watershed flood early warning method based on an X-band dual-polarization phased array radar, comprising the following steps:

[0007] Step 1: Obtain the X-band dual-polarization phased array radar dataset and the measured rainfall dataset at the rain gauge, and establish a radar rainfall estimation model through linear fitting.

[0008] Step 2: Obtain an X-band dual-polarization phased array radar image set, and use a short-term rainfall forecast model to perform prediction based on the X-band dual-polarization phased array radar image set to obtain a predicted radar image and corresponding predicted radar data.

[0009] Step three: input the predicted radar data into the radar rainfall estimation model to obtain the predicted rainfall.

[0010] Step 4: Use a flood forecasting model to issue a small watershed flood warning based on the predicted rainfall.

[0011] In a possible implementation, in step 1, after obtaining the X-band dual-polarization phased array radar dataset, data quality control processing and mixed elevation angle particle value processing are performed on the data in the X-band dual-polarization phased array radar dataset.

[0012] The data quality control process includes: outlier removal, secondary echo filtering, ground object echo removal and attenuation correction.

[0013] In a possible implementation, in step 1, after obtaining the rainfall data set measured at the rain gauge station, data in the rainfall data set measured at the rain gauge station is subjected to data anomaly elimination and data cleaning.

[0014] In a possible implementation, in step 1, the accuracy of the radar rainfall estimation model is evaluated using the correlation coefficient, root mean square error, relative error, and relative deviation.

[0015] In a possible implementation, in step 2, the short-term rainfall forecast model is obtained by training a convolutional neural network model using an X-band dual-polarization phased array radar historical image training set.

[0016] In one possible implementation, the convolutional neural network model adopts a UNet network framework, which includes downsampling and upsampling structures.

[0017] In one possible implementation, the UNet network framework is provided with a convolutional block attention module.

[0018] In one possible implementation, the UNet network framework is provided with depthwise separable convolution.

[0019] In a possible implementation, a generative adversarial network is introduced based on the UNet network framework.

[0020] In a possible implementation, in step 4, the flood forecasting model adopts the Xin'anjiang model or the empirical unit line model.

[0021] The small watershed flood early warning method based on X-band dual-polarization phased array radar in this application has the following advantages:

[0022] By combining X-band dual-polarization phased array radar data and images with radar rainfall estimation models, short-term rainfall forecast models, and flood forecast models, the accuracy of rainfall forecasts has been improved, thereby enhancing the reliability of flood warnings. X-band dual-polarization phased array radar data can meet the high-resolution rainfall monitoring requirements of small watershed flood warnings.

[0023] The proposed data quality control processing and mixed elevation angle particle value processing for the data in the X-band dual-polarization phased array radar dataset improves the reliability of radar data and thus improves the reliability of rainfall prediction.

[0024] The proposed method uses correlation coefficient, root mean square error, relative error, and relative deviation to evaluate the accuracy of the radar rainfall estimation model. When the accuracy does not meet the requirements, the radar rainfall estimation model can be adjusted until the accuracy meets the requirements.

[0025] The proposed short-term rainfall forecast model is obtained by training a convolutional neural network model using a historical image training set of X-band dual-polarization phased array radar. Compared with recurrent neural networks, convolutional neural network models are computationally efficient and can forecast extrapolated results in one go, without the need for iterative forecasting and taking less time.

[0026] The proposed convolutional neural network model adopts the UNet network framework, which includes downsampling and upsampling structures and has the advantages of lightweight and high efficiency.

[0027] The proposed UNet network framework is equipped with a convolutional block attention module, which guides the network to pay more attention to features that are more important to the task.

[0028] The proposed UNet network framework is set with depth-wise separable convolution, which can reduce the parameters of the model without significantly sacrificing performance.

[0029] The proposed method introduces a generative adversarial network based on the UNet network framework, and adopts an adversarial learning strategy to force the predicted image effect to be closer to the real image. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0031] Figure 1A schematic flow chart of a small watershed flood warning method based on an X-band dual-polarization phased array radar provided in an embodiment of the present application. DETAILED DESCRIPTION

[0032] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0033] like Figure 1 As shown, the embodiment of the present application provides a small watershed flood warning method based on an X-band dual-polarization phased array radar, comprising the following steps:

[0034] Step 1: Obtain the X-band dual-polarization phased array radar dataset and the measured rainfall dataset at the rain gauge, and establish a radar rainfall estimation model through linear fitting.

[0035] Step 2: Obtain an X-band dual-polarization phased array radar image set, and use a short-term rainfall forecast model to perform prediction based on the X-band dual-polarization phased array radar image set to obtain a predicted radar image and corresponding predicted radar data.

[0036] Step three: input the predicted radar data into the radar rainfall estimation model to obtain the predicted rainfall.

[0037] Step 4: Use a flood forecasting model to issue a small watershed flood warning based on the predicted rainfall.

[0038] Exemplarily, in step 1, after obtaining the X-band dual-polarization phased array radar dataset, data quality control processing and mixed elevation angle particle value processing are performed on the data in the X-band dual-polarization phased array radar dataset.

[0039] The data quality control process includes: outlier removal, secondary echo filtering, ground object echo removal and attenuation correction.

[0040] Specifically, in this embodiment, the data in the X-band dual-polarization phased array radar data set includes: horizontal reflectivity factor Z H (dBZ), differential reflectivity factor Z DR (dB), differential propagation phase shift rate K DP (° / km), correlation coefficient ρ HV .

[0041] Horizontal reflectivity factor Z HIt is related to the number and size of particles and is defined as follows when the Rayleigh scattering condition is met: the sum of the sixth power of the diameter of the rainfall particles per unit volume. Therefore, the larger the particle size, the stronger the reflectivity. H As shown in the following formula:

[0042]

[0043] Among them, D i represents the diameter of the i-th rainfall particle, D represents the diameter of the rainfall particle, and N(D) represents the particle droplet spectrum structure.

[0044] Differential reflectivity factor Z DR As shown in the following formula:

[0045]

[0046] Among them, Z H Represents the horizontal reflectivity factor, Z V Represents the vertical reflectivity factor. The closer the differential reflectivity factor is to 0, the closer the raindrop shape is to a sphere. When the differential reflectivity factor is less than 0, the raindrop shape is a vertical ellipsoid. When the differential reflectivity factor is greater than 0, the raindrop shape is a horizontal ellipsoid, regardless of the number of particles in the rainfall area.

[0047] Differential propagation phase shift rate K DP As shown in the following formula:

[0048]

[0049] in, represents the forward phase difference propagation phase shift, r n and r n+1 Respectively represent the nth and n+1th detection distances, and They represent the two-way propagation phase shift obtained by measuring the nth and n+1th detection distances respectively. When the precipitation particles in the air receive the electromagnetic waves emitted by the detection radar, they also receive the electromagnetic waves forward scattered by other precipitation particles between the radar and the air. There is a phase difference between these two electromagnetic waves. Because the precipitation particles are not all standard spherical particles, the forward scattering will vary with the incident angle in different directions. The forward phase difference propagation phase shift It is the difference between the vertical and horizontal phase differences. Usually, as the detection distance increases, The value of will also be accumulated. The accumulated amount is very small in weak precipitation and large in heavy precipitation. DP That is, the impact of particles on the phase difference of radar wave propagation is not affected by the calibration error of the radar system, and is very little affected by attenuation, partial occlusion, and droplet spectrum changes. Therefore, it is widely used in attenuation correction, system calibration and particle phase identification.

[0050] Correlation coefficient ρ HV It refers to the amplitude of the correlation coefficient between horizontal and vertical polarization echoes, which indicates the correlation between the backscattering characteristics of electromagnetic waves in two directions. Its size is related to the phase state, axial ratio, tilt angle and irregularity of the particles. Generally, the correlation coefficient of liquid precipitation is greater than 0.95, which can be used as an important parameter for identifying precipitation particles and ground objects.

[0051] Specifically, in this embodiment, outlier removal is a fundamental method for data quality control. Because various errors may exist during the acquisition of radar reflectivity factor data, such as radar system failures and weather interference, outliers need to be removed. Outlier removal typically utilizes statistical methods, such as those based on local variance, median, and standard deviation.

[0052] Secondary echo filtering effectively removes multiple reflections from radar data. These reflections are often caused by the interaction between the radar beam and ground objects, resulting in false high-reflectivity areas in the data. To eliminate these false high-reflectivity areas, a secondary echo filtering method based on radar range and time can be used.

[0053] In addition, radar data may contain electromagnetic interference signals, such as lightning and television signals, which can interfere with the collection and analysis of radar reflectivity factor data. To eliminate these electromagnetic interference signals, time- and space-based filtering methods can be used.

[0054] Ground object echo removal usually adopts methods based on terrain and vegetation, such as terrain height profile analysis and vegetation index calculation.

[0055] Attenuation correction is to modify the radar reflectivity factor data to eliminate the attenuation effect of the atmosphere on the radar beam.

[0056] Specifically, mixed-elevation particle extraction is a key step in radar-based quantitative precipitation estimation. Factors such as radar elevation angle, particle size, and particle morphology must be considered when performing this extraction. Data from different elevation angles and particle sizes require different processing methods to obtain accurate precipitation estimates. Furthermore, validating and correcting the mixed-elevation particle extraction results can improve their reliability and accuracy.

[0057] However, complex terrain can introduce errors into precipitation estimates. When a single radar scans an area with complex terrain, the radar's electromagnetic beam can be obstructed, resulting in inaccurate data in some areas. This is especially true in areas with complex terrain, where obstruction is particularly severe. In such cases, using multiple radars to cover the same area within their scanning range allows for complementary verification and improved accuracy.

[0058] Illustratively, in step one, after obtaining the actual rainfall data set of the rain gauge station, data anomaly removal and data cleaning are performed on the data in the actual rainfall data set of the rain gauge station.

[0059] Specifically, data anomaly removal involves eliminating cases of missing values ​​and statistical anomalies. Missing values ​​refer to situations where a rain gauge does not record rainfall data due to instrument or transmission problems, resulting in a null value in the table. This data can be directly removed. Statistical anomalies refer to situations where the rainfall data recorded at a certain station is significantly higher or lower than that of other nearby stations. There is no fixed criterion for judging this type of anomaly, and the presence of statistical anomalies can usually be determined by examining outliers.

[0060] The purpose of data cleaning is to obtain data pairs that can be used to fit and evaluate radar rainfall estimation relationships. First, it is necessary to extract the station data for the time period and region of interest and merge these data together. Then, perform a check for station location anomalies and remove stations with abnormal locations. Note that the rainfall measurement accuracy of most rain gauges is 0.1mm or 0.5mm, so it is necessary to retain stations with rainfall of at least 0.1mm or 0.5mm. In addition, hourly cumulative rainfall exceeding 150mm is generally considered an outlier (although extremely heavy rainfall can even reach 200mm / h) and needs to be removed.

[0061] Furthermore, during the rain gauge data cleaning process, it is necessary to compare the rain gauge data with X-band dual-polarization phased array radar data from the corresponding locations for quality control. Specifically, data pairs with high radar reflectivity and abnormally low rainfall, or vice versa, need to be eliminated to reduce the impact of abnormal distributions on the generated fitted rainfall relationship.

[0062] Specifically, in this embodiment, in step 1, the X-band dual-polarization phased array radar dataset and the rain gauge station measured rainfall dataset are respectively 5-minute X-band dual-polarization phased array radar data and rain gauge station measured rainfall data from April to July 2022 in a certain location to obtain parameter values ​​for a localized quantitative precipitation estimation relationship. By linearly fitting the X-band dual-polarization phased array radar dataset and the rain gauge station measured rainfall dataset, the quantitative precipitation estimation relationship parameters (i.e., the relationship between the X-band dual-polarization phased array radar data and the rain gauge station measured rainfall data) are determined, thereby achieving the process of predicting and monitoring rainfall in the region.

[0063] Exemplarily, in step one, the accuracy of the radar rainfall estimation model is evaluated using the correlation coefficient, root mean square error, relative error, and relative deviation.

[0064] Specifically, the formulas for the correlation coefficient (CC), root mean square error (RMSE), relative error (RMAE), and relative bias (RMB) are as follows:

[0065]

[0066] Where n represents the number of matching data points between the X-band dual-polarization phased array radar dataset and the rainfall dataset measured at the rain gauge, and R i represents the hourly precipitation estimated by the i-th group of radars (i.e., the i-th group of outputs of the radar rainfall estimation model) G i represents the measured rainfall data at the corresponding rain gauge, R represents the mean hourly rainfall estimated by n radars, and G represents the mean rainfall data measured at n rain gauges. The correlation coefficient (CC) reflects the correlation between the radar estimate and the ground-based measurement. The closer it is to 1, the higher the agreement. The root mean square error (RMSE) reflects the degree to which the radar estimate deviates from the measured rainfall. Smaller values ​​indicate lower dispersion, indicating more concentrated data and better algorithm stability. The relative error (RMAE) reflects the reliability of the data; smaller values ​​indicate smaller actual errors. The relative deviation (RMB) reflects the average deviation between the radar observations and the ground-based measurements. A positive (negative) RMB indicates that the radar-estimated precipitation overestimates (underestimates) the observed precipitation.

[0067] Illustratively, in step 2, the short-term rainfall forecast model is obtained by training a convolutional neural network model using an X-band dual-polarization phased array radar historical image training set.

[0068] Specifically, at a macro level, we typically consider a radar image of the study area at a specific moment in time as a two-dimensional image frame. Therefore, the entire rainfall process is composed of a sequence of two-dimensional radar rainfall images. These radar images have strong spatiotemporal correlations between adjacent frames, making them amenable to spatiotemporal sequence prediction methods. For short-term rainfall forecasting, the observed data at each moment is a two-dimensional M×N radar rainfall map. If the radar rainfall image is partitioned along a P-dimensional scale, it can be considered a many-to-many spatiotemporal sequence prediction problem. This modeling approach, known as end-to-end processing, allows for training and prediction using deep learning methods. This modeling approach does not require feature selection, but instead performs semantic segmentation on the entire feature set to automatically acquire image features. However, this approach is computationally demanding and is therefore well-suited for simulation and prediction using deep learning methods.

[0069] Compared to recurrent neural networks, convolutional neural networks are computationally efficient and can predict extrapolated results in a single pass, eliminating the need for iterative predictions and resulting in a shorter prediction time. A convolutional neural network typically consists of four components: an input layer, alternating convolutional and pooling layers, and a final output layer. Optionally, a normalization layer and a dropout layer may also be included in between to prevent overfitting. By repeatedly stacking different layers, a deep and complete convolutional neural network is constructed. The process is as follows: the original image enters the network through the input layer, where the size of the input vector is determined by the original image size. The convolutional neural network uses different convolution kernels to extract different features of the image. The data is then reduced in dimensionality through a pooling layer. Finally, the output layer is preceded by a fully connected layer to produce the network's final output.

[0070] Exemplarily, the convolutional neural network model adopts the UNet network framework, which includes downsampling and upsampling structures.

[0071] Specifically, in this embodiment, the convolutional neural network model adopts the UNet network framework, downsampling is used to gradually display environmental information, and the upsampling process is to combine the downsampled layer information and the upsampled input information to restore the detail information and gradually restore the image accuracy. This model is a lightweight but very efficient model.

[0072] Exemplarily, the UNet network framework is provided with a convolutional block attention module.

[0073] Specifically, the convolutional block attention module sequentially applies the attention mechanism to the channel and spatial dimensions. Attention is a mechanism that amplifies useful signals and suppresses less important ones, guiding the network to focus more on features that are more important to the task. In this embodiment, the convolutional block attention module is inserted into the skip connection between the encoder and decoder of the UNet network framework. During training, a loss function with the attention mechanism is used to optimize the network parameters.

[0074] Exemplarily, the UNet network framework is provided with depthwise separable convolution.

[0075] Specifically, the depthwise separable convolution splits the traditional convolution operation into depthwise convolution and one pointwise convolution, which can reduce the parameters of the model without significantly sacrificing performance. In this embodiment, the traditional standard convolution is replaced with depthwise separable convolution in the encoder and decoder of the UNet network framework. This includes convolutional layers, pooling layers, and convolution operations after upsampling layers. Since the number of output channels of the depthwise separable convolution is the same as the number of input channels (in the depthwise convolution stage), the number of output channels is adjusted in the pointwise convolution stage to match the subsequent layers of UNet, ensuring that the replaced network structure still maintains the U-shaped structure and jump connections of UNet.

[0076] Exemplarily, a generative adversarial network is introduced based on the UNet network framework.

[0077] Specifically, in this embodiment, the UNet network framework is used as the main structure of the generator of the generative adversarial network, maintaining its U-shaped structure and skip connections. The discriminator of the generative adversarial network adopts a convolutional neural network structure to distinguish the generated predicted images from the real images. The adversarial learning strategy is used to force the predicted images to be closer to the real images.

[0078] Illustratively, in step 4, the flood forecasting model adopts the Xin'anjiang model or the empirical unit line model.

[0079] Specifically, in this embodiment, the Xin'an River model includes a three-source full-flow model, a three-source hysteresis model, and a Muskingum River segmented continuous model. Considering the small catchment area of ​​the small watershed, the calculation time step size is set to 1 hour.

[0080] The three-source full-storage runoff model includes three parts: evapotranspiration calculation, runoff calculation and water source calculation.

[0081] In this embodiment, evaporation is calculated using a three-layer evaporation model, and the calculation formula is as follows:

[0082] E p =K×E0.

[0083] Among them, E p represents evaporation capacity, K represents evaporation conversion coefficient, and E0 represents the measured evaporation amount.

[0084] The flow yield calculation uses the watershed storage capacity curve to consider that the soil water deficit on the watershed surface is equal to the water storage capacity, as shown in the following formula:

[0085]

[0086] Among them, f / F represents a parabola, Wm represents the point storage capacity, Wmm represents the maximum value of the point storage capacity, and b represents the parabola power of the basin storage capacity curve.

[0087] Then the average basin storage capacity WM is obtained as shown in the following formula:

[0088]

[0089] When the ratio of the basin impervious area IMP is not equal to 0, the above formula is rewritten as:

[0090]

[0091] The ordinate value a corresponding to a certain soil moisture content W is as shown in the following formula:

[0092]

[0093] When the rainfall PE after deducting evaporation is less than 0, there is no runoff generation; when it is greater than 0, runoff generation occurs.

[0094] Runoff generation is divided into two cases: partial runoff generation and full-basin runoff generation:

[0095] When PE + a < Wmm, at this time it is partial runoff generation, and the runoff volume is as shown in the following formula:

[0096]

[0097] When PE + a ≥ Wmm, at this time it is full-basin runoff generation, and the runoff volume is as shown in the following formula:

[0098] R = PE - (WM - W).

[0099] Among them, R represents the runoff volume, and PE represents the rainfall after deducting evaporation.

[0100] For the analysis of the flow hydrograph during the flood season in humid and semi-humid regions by calculating water sources separately, the runoff components generally include surface, subsurface, and groundwater components. Since there are obvious differences in the confluence velocities of the runoff of various components, the division of water sources is a very important link. In this embodiment, the division of water sources is carried out through a free water storage reservoir.

[0101] The runoff volume R obtained from runoff generation enters the free water storage reservoir, together with the water that has not flowed out completely in the reservoir originally, to form the real-time storage volume S. The bottom width of the free water storage reservoir is the current ratio of the runoff generation area FR, which is time-varying. KI and KG are the outflow coefficients of subsurface flow and groundwater respectively. The calculation formulas for the runoff volumes of various water sources are as follows:

[0102] When S + R ≤ SM, RS = 0, RI = (S + R) × KI × FR, RG = (S + R) × KG × FR.

[0103] When S + R > SM, RS = (S + R - SM) × FR, RI = SM × KI × FR, RG = SM × KG × FR.

[0104] Among them, SM represents the free - water average storage capacity, RS represents the surface runoff, RI represents the subsurface runoff, and RG represents the groundwater runoff.

[0105] Since the free - water point storage capacity Sm on the runoff - generating area ratio FR is not evenly distributed, it is not appropriate to take the free - water average storage capacity SM as a constant. It is also necessary to consider its area distribution in a way similar to the basin storage capacity curve. For this purpose, a parabola is also adopted, and EX is introduced as its power, then there is:

[0106]

[0107] Smm = (1 + EX)SM.

[0108]

[0109] Among them, Sm represents the free - water point storage capacity, Smm represents the maximum value of the free - water point storage capacity, EX represents the parabola power of the free - water storage capacity curve, SM represents the free - water average storage capacity, and AU represents the ordinate value corresponding to a certain storage capacity S.

[0110] When PE + AU < Smm, the surface runoff is shown as follows:

[0111]

[0112] When PE + AU ≥ Smm, the surface runoff is shown as follows:

[0113] RS = (PE + S - SM)FR.

[0114] The three - source lagging routing model is used for routing calculation, including unit - basin routing and channel routing. Unit - basin routing includes hillslope routing and river - network routing.

[0115] Hillslope routing refers to the process of water body gathering on the hillslope. In this routing stage, the regulation of the surface runoff obtained through water - source division in the three - source runoff - generating model has little effect, and it directly enters the river network, becoming the total inflow of the surface runoff to the river network; the subsurface flow enters the subsurface - water reservoir, and after the recession of the subsurface - water reservoir, it becomes the total inflow of the subsurface flow to the river network; the groundwater runoff enters the groundwater storage reservoir, and after the recession of the groundwater storage reservoir, it becomes the total inflow of the groundwater to the river network.

[0116] River network confluence refers to the process by which water flows from the slope into the river channel and converges along the river network. During this confluence phase, the flow characteristics are governed by the hydraulic conditions of the river channel, and the various water sources are consistent. The sum of the three is the total inflow to the river network, which is then aggregated to the unit outlet after a hysteresis calculation.

[0117] River confluence refers to the use of segmented Muskingum continuous routing to calculate the flow of each unit outlet to the basin outlet based on the hydraulic characteristics of each unit outlet to the basin outlet and river channel, and then perform linear superposition.

[0118] The basic principle of Muskingum method is based on the water balance equation and the tank storage equation.

[0119] The water balance equation is as follows:

[0120]

[0121] The tank storage equation is as follows:

[0122] W=KQ′.

[0123] Q′=xI+(1-x)O.

[0124] Where I1 and I2 represent the upstream inflow of the reach at the beginning and end of the time period, respectively; O1 and O2 represent the downstream outflow of the reach at the beginning and end of the time period, respectively; Δt represents the calculation time step; W1 and W2 represent the reach channel storage at the beginning and end of the time period, respectively. W represents the reach channel storage, K represents the storage constant (with time dimension), Q′ represents the storage flow, I represents the inflow, O represents the outflow, and x represents the flow weighting factor (dimensionless).

[0125] The Muskingum confluence calculation formula obtained by combining the water balance equation and the tank storage equation is as follows:

[0126] O2=C0I2+C1I1+C2O1.

[0127] Among them, C0, C1, and C2 are functions of the storage constant K and the flow rate weight factor x, as follows:

[0128]

[0129] C0+C1+C2=1.

[0130] When Δt<2Kx, C0<0, I2 has a negative effect on O2, and negative flow is likely to occur in the rising section of the outflow process line; when Δt>2K-2Kx, C2<0, O1 has a negative effect on O2, and negative flow is likely to occur in the receding section of the outflow process line. In order to avoid unreasonable phenomena such as negative flow, to ensure that the flow of the upper and lower sections changes linearly during the calculation period and that the flow changes linearly along the river section at any time, it is generally required that Δt≈K. The Muskingum River Segmented Continuous Routing Model is a segmented continuous routing model based on the Muskingum method, which divides the routing river section into N sub-river sections, and each sub-river section has a parameter K L 、x L The relationship with the parameters before river section division is as follows:

[0131]

[0132] The formula for segmented continuous routing still uses the Muskingum confluence routing formula and the function formulas of C0, C1, and C2, but K and x in the function formulas are replaced by the segmented sub-section parameter K. L 、x L .

[0133] In the empirical unit line model, the unit line refers to the surface runoff hydrograph formed at the outlet section of a given watershed by a unit of net surface rainfall uniformly distributed over a unit time period. This indicates that the shape of the surface (direct) runoff hydrograph for a given watershed is influenced by all of its physical characteristics. The unit line method makes three assumptions: the total duration of the resulting surface runoff hydrograph remains constant despite varying net rainfall over a unit time period; the outflow process generated by N times the net rainfall over a unit time period has a flow rate that is N times that of the unit line; and the outflow processes generated by net rainfall over each unit time period do not interfere with each other, with the flow rate at the outlet section equal to the sum of the flows generated by the net rainfall over each unit time period.

[0134] The unit line is applicable to the principles of multiple ratio and superposition, because its input (net rainfall) and output (outflow) take discrete values, so the convolution integral of the system can be expressed as follows:

[0135]

[0136] Among them, Q i represents the flow rate at the end of the i-th outlet section period, I j represents the average net rainfall in the jth period, m represents the number of net rainfall periods, q i-i+1 It represents the flow rate at the end of the unit line in the i-j+1th period.

[0137] In this embodiment, the warning level is set to four levels, represented by four colors: blue, yellow, orange, and red.

[0138] When the maximum grid rainfall reaches a set threshold of 30 mm (a customizable threshold), or when the percentage of areas within a region / basin with rainfall greater than 20 mm (a customizable threshold) exceeds 20% (a customizable threshold), a page and SMS alert will be issued for that region / basin. The resolution of the grid rainfall is tentatively set at 1 km. When the threshold is reached, an alert is issued every 30 minutes. If the alert level is raised by one level, the alert message is resent, and the alert interval is refreshed.

[0139] In this embodiment, the default values ​​of the precipitation warning level thresholds are shown in Table 1:

[0140] Table 1 Default values ​​of precipitation warning level thresholds

[0141] Precipitation warning level 1 hour (milliseconds) 3 hours (mm) River rise (meters) blue 30 50 1～2 yellow 50 70 2～3 orange color 70 100 3～4 red 100 200 More than 4 meters

[0142] In a possible embodiment, the small watershed flood warning method based on X-band dual-polarization phased array radar of the present application is applied to small watersheds such as the Guitang River, Huangni River, Baisha River, and Jinjing River, and 10 flood and disaster risk warning text messages are issued. The average forecast period for small watershed flood and disaster risk warnings is 1.29 hours, of which the average forecast period for Guitang River flood and disaster risk warnings is 1.8 hours.

[0143] This embodiment of the application improves the accuracy of rainfall forecasts and, in turn, the reliability of flood warnings by combining X-band dual-polarization phased array radar data and images with radar rainfall estimation models, short-term rainfall forecast models, and flood forecast models. X-band dual-polarization phased array radar data can meet the high-resolution rainfall monitoring requirements for small watershed flood warnings.

[0144] The proposed data quality control processing and mixed elevation angle particle value processing for the data in the X-band dual-polarization phased array radar dataset improves the reliability of radar data and thus improves the reliability of rainfall prediction.

[0145] The proposed method uses correlation coefficient, root mean square error, relative error, and relative deviation to evaluate the accuracy of the radar rainfall estimation model. When the accuracy does not meet the requirements, the radar rainfall estimation model can be adjusted until the accuracy meets the requirements.

[0146] The proposed short-term rainfall forecast model is obtained by training a convolutional neural network model using a historical image training set of X-band dual-polarization phased array radar. Compared with recurrent neural networks, convolutional neural network models are computationally efficient and can forecast extrapolated results in one go, without the need for iterative forecasting and taking less time.

[0147] The proposed convolutional neural network model adopts the UNet network framework, which includes downsampling and upsampling structures and has the advantages of lightweight and high efficiency.

[0148] The proposed UNet network framework is equipped with a convolutional block attention module, which guides the network to pay more attention to features that are more important to the task.

[0149] The proposed UNet network framework is set with depth-wise separable convolution, which can reduce the parameters of the model without significantly sacrificing performance.

[0150] The proposed method introduces a generative adversarial network based on the UNet network framework, and adopts an adversarial learning strategy to force the predicted image effect to be closer to the real image.

[0151] Although the preferred embodiments of the present application have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present application.

[0152] Obviously, those skilled in the art may make various changes and modifications to this application without departing from the spirit and scope of this application. Thus, if these modifications and variations of this application fall within the scope of the claims of this application and their equivalents, this application is intended to include these modifications and variations.

Claims

1. A small watershed flood warning method based on X-band dual-polarization phased array radar, characterized in that: The following steps are involved: Step 1: Obtain the X-band dual-polarization phased array radar data set and the measured rainfall data set of the rain gauge station, and establish a radar rainfall estimation model through linear fitting; Step 2: obtaining an X-band dual-polarization phased array radar image set, and using a short-term rainfall forecast model to perform prediction based on the X-band dual-polarization phased array radar image set to obtain a predicted radar image and corresponding predicted radar data; Step 3, inputting the predicted radar data into the radar rainfall estimation model to obtain predicted rainfall; Step 4: Use a flood forecasting model to issue a small watershed flood warning based on the predicted rainfall.

2. The small watershed flood early warning method based on X-band dual-polarization phased array radar according to claim 1 is characterized in that: In step 1, after acquiring the X-band dual-polarization phased array radar data set, data quality control processing and mixed elevation angle particle value processing are performed on the data in the X-band dual-polarization phased array radar data set; The data quality control process includes: outlier removal, secondary echo filtering, ground object echo removal and attenuation correction.

3. The small watershed flood early warning method based on X-band dual-polarization phased array radar according to claim 1 is characterized in that: In step one, after obtaining the measured rainfall data set of the rain gauge station, data anomalies are eliminated and data cleaning is performed on the data in the measured rainfall data set of the rain gauge station.

4. The small watershed flood early warning method based on X-band dual-polarization phased array radar according to claim 1 is characterized in that: In step 1, the accuracy of the radar rainfall estimation model is evaluated using correlation coefficient, root mean square error, relative error, and relative deviation.

5. The small watershed flood early warning method based on X-band dual-polarization phased array radar according to claim 1 is characterized in that: In step 2, the short-term rainfall forecast model is obtained by training a convolutional neural network model using an X-band dual-polarization phased array radar historical image training set.

6. The small watershed flood early warning method based on X-band dual-polarization phased array radar according to claim 5 is characterized in that: The convolutional neural network model adopts the UNet network framework, which includes downsampling and upsampling structures.

7. The small watershed flood early warning method based on X-band dual-polarization phased array radar according to claim 6 is characterized in that: The UNet network framework is provided with a convolutional block attention module.

8. The small watershed flood early warning method based on X-band dual-polarization phased array radar according to claim 6 is characterized in that: The UNet network framework is provided with depthwise separable convolutions.

9. The small watershed flood early warning method based on X-band dual-polarization phased array radar according to claim 6 is characterized in that: A generative adversarial network is introduced based on the UNet network framework.

10. The small watershed flood early warning method based on X-band dual-polarization phased array radar according to claim 1, characterized in that: In step 4, the flood forecasting model adopts the Xin'anjiang model or the empirical unit line model.

Images