A real-time rainfall field assimilation method integrating rainfall stations, monitoring images and numerical weather forecasts
Through deep learning model and Bayesian fusion technology, the monitoring images and data of fusion rainfall stations, monitoring images and numerical weather forecasts are decomposed, and the real-time rainfall field estimation is insufficient, achieving more accurate rainfall field data. In urban flood control, disaster reduction and emergency management.
Patent Information
- Application Number
- CN202410558495.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-05-08
- Publication Date
- 2024-11-01
- Estimated Expiration
- 2044-05-08
AI Technical Summary
In the prior art, the estimation accuracy of real-time rainfall fields is not high, and it is difficult to effectively integrate rainfall stations, monitoring images and numerical weather forecast information, resulting in insufficient accuracy of rainfall spatial distribution, affecting the accuracy of meteorological and hydrological forecasts.
The deep learning model is used to decompose the monitoring images, extract rainfall layer information, combine matrix low-rank decomposition and Bayesian fusion method to fuse the data of rainfall stations, monitoring images and numerical weather forecasts, and obtain more accurate through spatial interpolation and probability density function matching. real-time rainfall data.
It improves the estimation accuracy of rainfall sites, is suitable for real-time forecasting and early warning of urban heavy rainstorms and floods, and provides technical support for flood prevention and disaster reduction and emergency management.
Smart Images

Figure CN118503906B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of atmospheric science and water conservancy engineering technology, and particularly relates to a real-time rainfall field assimilation method integrating a rainfall station, a monitoring image and a numerical weather forecast. Background Art
[0002] In meteorological and hydrological business applications, high-precision rainfall fields are particularly important, which determines the accuracy of meteorological and hydrological forecasts based on rainfall as input. However, rainfall is a highly variable atmospheric variable, so it is difficult to obtain accurate estimates of rainfall amount and spatial field characteristics.
[0003] At present, the traditional rainfall measurement at rain gauges can obtain relatively accurate point estimates of rainfall, but due to the insufficient spatial density of rain gauges, it cannot accurately describe the spatial variation of rainfall. In recent years, short-term numerical weather forecasting technology has gradually matured. Although there is still a large error in the accuracy of rainfall forecasts at specific points, it can provide a relatively reliable description of the spatial distribution characteristics of rainfall. In addition, government business departments such as water conservancy and meteorology have deployed a large number of dedicated video surveillance, and the public has also deployed a large number of private video surveillance, which can provide continuous video surveillance images and obtain massive rainfall image information in real time during rainfall. How to integrate the rainfall information from the above three different sources, make full use of the advantages of the "point estimation" accuracy of rain gauges, the "spatial variation" reliability of numerical weather forecasts, and the "spatial distribution and real-time" of monitoring images, and obtain a more accurate rainfall field than previous rainfall observation data processing technologies has important application value. Summary of the invention
[0004] In order to solve the technical problem of low accuracy of real-time rainfall field estimation in the prior art, the present invention proposes a real-time rainfall field assimilation method that integrates rain gauges, monitoring images and numerical weather forecasts, providing a technical basis for urban flood prevention, disaster reduction and emergency management.
[0005] The object of the present invention is achieved in that:
[0006] The present invention provides a real-time rainfall field assimilation method integrating rain gauges, monitoring images and numerical weather forecasts, determining the scope of a region to be studied, in which a number of monitoring video points and ground rain gauges are provided, and collecting grid rainfall forecast data in numerical weather forecasts for a rainfall event in the region, the method comprises the following steps:
[0007] Step 1: Decompose the monitoring image to obtain the rainfall layer
[0008] Decompose the monitoring image obtained from the monitoring video point into a background layer B, a rainfall layer R and a noise layer N;
[0009] The background layer represents the main body of the image; the rain layer represents raindrops; and the noise layer represents random disturbances in the image due to factors including wind or brightness changes.
[0010] Step 2: Extract raindrop information from the rainfall layer of the monitoring image and estimate the rainfall intensity
[0011] S201, establish a functional relationship between the rainfall intensity I and the number of raindrops d and the size of raindrops s in the rainfall image Z(·), specifically:
[0012] I=f(Z(d,s))
[0013] Where f(·) represents the nonlinear relationship between rainfall intensity and rainfall image;
[0014] S202, using a deep learning model to establish a solution model for the function I=f(Z(d,s)) in step S201; using a publicly available massive image dataset as a pre-training input to obtain parameter values of the deep learning model;
[0015] S203, using the rainfall layer R obtained in step 1 as input, verifying the deep learning model determined in step S202, and fine-tuning its parameter values, and finally obtaining the parameter values of the applicable deep learning model;
[0016] S204, using the deep learning model determined in step S203, taking the image captured in real time at the monitoring video point as input, to obtain the real-time rainfall intensity of the monitoring video point in the area to be studied.
[0017] Step 3: Establish a rainfall spatial field based on monitoring images
[0018] The spatial resolution of the rainfall spatial field is selected, and the rainfall spatial interpolation method is used to spatially interpolate the real-time rainfall intensity data of the point-distributed monitoring video points obtained in step 2 to obtain the rainfall spatial field Ω1 based on the monitoring image.
[0019] Step 4: Establish a rainfall spatial field based on ground rainfall stations
[0020] The same spatial resolution as step 3 is selected, and the rainfall spatial interpolation method is also used to spatially interpolate the rainfall intensity data of the point-distributed rain gauge stations to obtain the rainfall spatial field Ω2 based on the rain gauge stations.
[0021] Step 5: Fusion monitoring image and rainfall field assimilation of rain gauge station
[0022] The Bayesian fusion method is used to assimilate the two independent source rainfall spatial field data Ω1 and Ω2 obtained in steps 3 and 4 to obtain the assimilated new rainfall spatial field data Ω 1+2 .
[0023] Step 6: Assimilation of rainfall fields by integrating rainfall stations, monitoring images and numerical weather forecast information
[0024] S601, selecting grid rainfall forecast data in the numerical weather forecast for the rainfall event to be studied, and processing its spatial resolution (i.e., grid size) to be consistent with the spatial resolution of rainfall data Ω1 and Ω2, recorded as Ω3;
[0025] S602, using the Bayesian fusion method, two rainfall spatial field data from different sources Ω 1+2 Assimilate the data with Ω3 to obtain the new rainfall spatial field data Ω after assimilation 1+2+3 , thus obtaining real-time rainfall field assimilation data that integrates three independent sources: rainfall stations, monitoring images, and numerical weather forecasts.
[0026] Furthermore, in step 1, the specific method of decomposing the monitoring image includes the following steps:
[0027] S101: Consider the surveillance image source data as a two-dimensional matrix X∈R h×w , specifically:
[0028] X=B+R+N
[0029] Where B∈R h×w , R∈R h×w and N∈R h×w are the background layer, rainfall layer and noise layer respectively, h and w are the height and width of the image respectively;
[0030] S102: Using the matrix low-rank decomposition theory, decompose and obtain the rainfall layer R∈R in step S101 h×w , which is the processed raindrop attribute matrix.
[0031] Furthermore, in step 5, the specific method of the Bayesian fusion method is:
[0032] S501: Calculate the probability density function of the monitoring image rainfall field Ω1 and the rainfall station rainfall field Ω2 respectively, take the probability density function of the ground rainfall station network rainfall field Ω2 as the benchmark, use the probability density function matching technology to eliminate the systematic error of the monitoring image rainfall field Ω1, and obtain the rainfall field after data accuracy correction.
[0033] S502: Using the dynamic Bayesian model averaging technology, the two rainfall fields Ω2 and Fusion is performed to obtain the new rainfall spatial field data Ω after assimilation 1+2 .
[0034] The advantages and beneficial effects of the present invention are:
[0035] The real-time rainfall field assimilation method integrating rain gauges, monitoring images and numerical weather forecasts described in the present invention overcomes the problems that real-time rainfall field estimation in the current meteorological and hydrological fields is mainly based on rain gauges, monitoring image information is insufficiently utilized, and spatial distribution accuracy is low during rainfall estimation. It is suitable for real-time forecasting and early warning of urban rainstorms and floods, and provides a technical basis for urban flood prevention, disaster reduction and emergency management. BRIEF DESCRIPTION OF THE DRAWINGS
[0036] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.
[0037] Figure 1 It is a technical framework diagram of the real-time rainfall field assimilation method integrating rainfall stations, monitoring images and numerical weather forecasts according to the present invention. DETAILED DESCRIPTION
[0038] Example:
[0039] In this example, a rainfall event occurs in a certain urban study area. There are n1 monitoring video points, n2 ground rain gauges, and grid rainfall forecast data in the numerical weather forecast in the area. According to the method described in the present invention, the process of real-time rainfall field assimilation for this rainfall event is as follows:
[0040] Step 1: Decompose the monitoring image to obtain the rainfall layer
[0041] S101, at the t hour, n1 monitoring images of each monitoring video point are collected every minute. There are a total of n1×60 monitoring images in the t hour. It is assumed that each monitoring image represents the average rainfall intensity in the 1-minute period.
[0042] S102, at the first monitoring point, the first minute monitoring image is regarded as a two-dimensional matrix X∈R h×w , decomposed into background layer B, rainfall layer R and noise layer N. Among them, the background layer represents the main body of the image, which is assumed to remain unchanged in the consecutive frames of the video; the rainfall layer represents raindrops, which are characterized by changes in frame-to-frame contrast and sparse distribution in space; the noise layer represents random disturbances of the image caused by wind or brightness changes, and accounts for a relatively small proportion of the signal compared with the other two layers.
[0043] X=B+R+N
[0044] Where B∈R h×w , R∈R h×w and N∈R h×w They are background layer, rainfall layer and noise layer respectively, and h and w are the height and width of the image respectively.
[0045] S103, using the mature matrix low-rank decomposition theory, decompose and obtain the rainfall layer R∈R h×w, which is the processed raindrop attribute matrix.
[0046] S104, repeating steps S102 and S103, obtaining all N2 monitoring points and all monitoring image rainfall layers from the 1st to the 60th minute (ie, raindrop attribute matrix data).
[0047] Step 2: Extract raindrop information from the rainfall layer of the monitoring image and estimate the rainfall intensity
[0048] S201, the rainfall image Z(·) in each monitoring image in step 1 includes information such as the number of raindrops d and the size of raindrops s, which can establish a functional relationship with the rainfall intensity I, specifically:
[0049] I=f(Z(d,s))
[0050] Where f(·) represents the nonlinear relationship between rainfall intensity and rainfall image.
[0051] S202, using an existing, technically mature deep learning model (such as a convolutional neural network model CNN) to establish a solution model for the function I = f(Z(d,s)) in step S201; using a large amount of image data sets that are publicly available on the Internet (these images are not necessarily related to rainfall) as pre-training input to obtain parameter values of the deep learning model.
[0052] S203, taking the rainfall layer R (i.e., raindrop attribute matrix data) obtained in step 1 as input, verifies the deep learning model determined in step S202, and fine-tunes its parameter values, and finally obtains the parameter values of the applicable deep learning model.
[0053] S204, using the deep learning model determined in step S203, taking the real-time image captured by the monitoring video point as input, the rainfall intensity of all 1 to n1 monitoring points in step 1 and the 1st to 60th minutes can be solved, and the rainfall data of the tth hour of the 1st to n1 monitoring video points can be accumulated. in refers to the rainfall at the tth hour at the first monitoring video point, It refers to the rainfall in the tth hour at the second monitoring video point, and so on.
[0054] Step 3: Establish a rainfall spatial field based on monitoring images
[0055] S301, select the spatial resolution of the rainfall spatial field, such as 1km×1km, and divide the entire study area into x 1km horizontal areas. 2 Square grid, y 1km in length 2 square grid;
[0056] S302, based on the rainfall data of the 1st to n1th monitoring video points in step 2 Using a mature rainfall spatial interpolation method (such as block kriging estimation), the real-time rainfall intensity data of the point-distributed monitoring video points are spatially interpolated to obtain the rainfall spatial field Ω1 based on the monitoring image. Specifically:
[0057]
[0058] Among them, A x,y Represents the rainfall at the tth hour based on the monitoring image on the xth column and yth row grid in the study area.
[0059] Step 4: Establish a rainfall spatial field based on ground rainfall stations
[0060] The same grid spatial resolution of 1km×1km is used as in step 3. A mature rainfall spatial interpolation method (such as block kriging estimation) is used to interpolate the rainfall data of the n2 rainfall stations at the t hour. Perform spatial interpolation to obtain the rainfall spatial field Ω2 based on the rainfall station, specifically:
[0061]
[0062] Among them, B x,y It represents the t-th hour rainfall based on the ground rain gauge station on the x-th column and y-th row grid in the study area.
[0063] Step 5: Fusion monitoring image and rainfall field assimilation of rain gauge station
[0064] S501: Calculate the probability density functions F(Ω1) and F(Ω2) of the monitoring image rainfall field Ω1 and the rainfall station rainfall field Ω2 respectively. Take the probability density function F(Ω2) of the ground rainfall station network rainfall field as the benchmark, use the mature probability density function matching technology to eliminate the systematic error of the monitoring image rainfall field Ω1, and obtain the rainfall field after data accuracy correction.
[0065] S502: Using the mature dynamic Bayesian model averaging technology, the two rainfall fields Ω2 and Fusion is performed to obtain the new rainfall spatial field data Ω after assimilation 1+2 .
[0066] Step 6: Assimilation of rainfall fields by integrating rainfall stations, monitoring images and numerical weather forecast information
[0067] S601, select the grid rainfall forecast data in the numerical weather forecast of the study area at the tth hour, the data spatial resolution is 1km×1km, denoted as Ω3, specifically:
[0068]
[0069] Among them, C x,y It represents the t-th hour rainfall of the grid rainfall forecast on the x-th column and y-th row grid in the study area.
[0070] S602: Calculate the probability density function F(Ω3) of the rainfall field Ω3 in the numerical weather forecast, take the probability density function F(Ω2) of the rainfall field of the ground rainfall station network as the benchmark, use the mature probability density function matching technology to eliminate the systematic error of the monitoring image rainfall field Ω3, and obtain the rainfall field after data accuracy correction
[0071] S603: Using the mature dynamic Bayesian model averaging technology, the two rainfall fields Ω 1+2 and Fusion is performed to obtain the new rainfall spatial field data Ω after assimilation 1+2+3 .
[0072] In summary, the real-time rainfall field assimilation method for integrating rain gauges, monitoring images and numerical weather forecasts disclosed in the present invention decomposes the rainfall layer in the monitoring image, establishes a deep learning model to solve the rainfall intensity, and spatially interpolates the point-distributed ground rain gauge data and the monitoring image rainfall data to obtain the spatial distribution of rainfall from two independent sources; and combines the rainfall spatial distribution data in the numerical weather forecast, and uses the Bayesian fusion technology to obtain a more accurate real-time rainfall field result after assimilation.
[0073] Finally, it should be noted that the above is only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred arrangement scheme, a person skilled in the art should understand that the technical solution of the present invention (such as the use of various formulas, the sequence of steps, etc.) can be modified or replaced by equivalents without departing from the spirit and scope of the technical solution of the present invention.
Claims
1. A real-time rainfall field assimilation method integrating rainfall stations, monitoring images and numerical weather forecasts, characterized in that: Determine the scope of the area to be studied, where a number of monitoring video points and ground rain gauges are set, and collect grid rainfall forecast data in numerical weather forecast for a rainfall event in the area. The method comprises the following steps: Step 1: Decompose the monitoring image to obtain the rainfall layer Decompose the monitoring image obtained from the monitoring video point into a background layer B, a rainfall layer R and a noise layer N; The background layer represents the main body of the image; the rain layer represents raindrops; the noise layer represents random disturbances in the image due to factors including wind or brightness changes; Step 2: Extract raindrop information from the rainfall layer of the monitoring image and estimate the rainfall intensity S201, establish a functional relationship between the rainfall intensity I and the number of raindrops d and the size of raindrops s in the rainfall image Z(·), specifically: I=f(Z(d,s)) Where f(·) represents the nonlinear relationship between rainfall intensity and rainfall image; S202, using a deep learning model to establish a solution model for the function I=f(Z(d,s)) in step S201; using a publicly available massive image dataset as a pre-training input to obtain parameter values of the deep learning model; S203, using the rainfall layer R obtained in step 1 as input, verifying the deep learning model determined in step S202, and fine-tuning its parameter values, and finally obtaining the parameter values of the applicable deep learning model; S204, using the deep learning model determined in step S203, taking the monitoring image taken in real time at the monitoring video point as input, to obtain the real-time rainfall intensity of the monitoring video point in the area to be studied; Step 3: Establish a rainfall spatial field based on monitoring images Select the spatial resolution of the rainfall spatial field, use the rainfall spatial interpolation method, and spatially interpolate the real-time rainfall intensity data of the point-distributed monitoring video points obtained in step 2 to obtain the rainfall spatial field data Ω1 based on the monitoring image; Step 4: Establish a rainfall spatial field based on ground rainfall stations Select the same spatial resolution as step 3, and also use the rainfall spatial interpolation method to spatially interpolate the rainfall intensity data of the point-distributed rain gauge stations to obtain the rainfall spatial field data Ω2 based on the rain gauge stations; Step 5: Fusion monitoring image and rainfall field assimilation of rain gauge station The Bayesian fusion method is used to assimilate the two independent source rainfall spatial field data Ω1 and Ω2 obtained in steps 3 and 4 to obtain the assimilated new rainfall spatial field data Ω 1+2 ; The specific method of Bayesian fusion method is: S501: Calculate the probability density functions of the rainfall spatial field data Ω1 based on the monitoring image and the rainfall spatial field data Ω2 based on the rain gauge station respectively, take the probability density function of the rainfall spatial field data Ω2 based on the rain gauge station as the benchmark, use the probability density function matching technology to eliminate the systematic error of the rainfall spatial field data Ω1 based on the monitoring image, and obtain the rainfall spatial field data after data accuracy correction. S502: Using the dynamic Bayesian model averaging technology, the two rainfall spatial field data Ω2 and Fusion is performed to obtain the new rainfall spatial field data Ω after assimilation 1+2 ; Step 6: Assimilation of rainfall fields by integrating rainfall stations, monitoring images and numerical weather forecast information S601, selecting grid precipitation forecast data in the numerical weather forecast for the rainfall event to be studied, and processing the spatial resolution of the data to be consistent with the spatial resolution of the rainfall spatial field data Ω1 and Ω2, recorded as Ω3; S602, using the Bayesian fusion method, two rainfall spatial field data from different sources Ω 1+2 Assimilate the data with Ω3 to obtain the new rainfall spatial field data Ω after assimilation 1+2+3 , thus obtaining real-time rainfall field assimilation data that integrates three independent sources: rainfall stations, monitoring images, and numerical weather forecasts.
2. The real-time rainfall field assimilation method integrating rainfall stations, monitoring images and numerical weather forecasts according to claim 1 is characterized in that: In step 1, the specific method of decomposing the monitoring image includes the following steps: S101: Consider the surveillance image source data as a two-dimensional matrix X∈R h×w , specifically: X=B+R+N Among them, B∈R h×w , R∈R h×w and N∈R h×w are the background layer, rainfall layer and noise layer respectively, h and w are the height and width of the image respectively; S102: Using the matrix low-rank decomposition theory, decompose and obtain the rainfall layer R∈R in step S101 h×w , which is the processed raindrop attribute matrix.
Citation Information
Cited By
Mountainous area low-order small-watershed rainfall spatial distribution mode analysis method and system
CN120028886A
