CN114742206B - Rainfall intensity estimation method for comprehensive multi-time space-scale Doppler radar data - Google Patents
Rainfall intensity estimation method for comprehensive multi-time space-scale Doppler radar data Download PDFInfo
- Publication number
- CN114742206B CN114742206B CN202210417830.9A CN202210417830A CN114742206B CN 114742206 B CN114742206 B CN 114742206B CN 202210417830 A CN202210417830 A CN 202210417830A CN 114742206 B CN114742206 B CN 114742206B
- Authority
- CN
- China
- Prior art keywords
- data
- factors
- scale
- model
- radar
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
- G06F16/21—Design, administration or maintenance of databases
- G06F16/215—Improving data quality; Data cleansing, e.g. de-duplication, removing invalid entries or correcting typographical errors
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
- G06F16/22—Indexing; Data structures therefor; Storage structures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
- G06F16/25—Integrating or interfacing systems involving database management systems
- G06F16/254—Extract, transform and load [ETL] procedures, e.g. ETL data flows in data warehouses
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Systems or methods specially adapted for specific business sectors, e.g. utilities or tourism
- G06Q50/10—Services
- G06Q50/26—Government or public services
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
- Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation
Abstract
The invention relates to the technical field of deep learning quantitative precipitation estimation, in particular to a rainfall intensity estimation method for comprehensive multi-time space-scale Doppler radar data, which comprises the following steps: doppler single polarization radar data and ground automatic weather station observation data are obtained; acquiring radar reflectivity factors, meteorological factors and geographic factors; the polar coordinates are converted into Cartesian coordinates; k neighbor interpolation and data slicing; making a label; designing a self-attention module and building a model; training a model and adjusting parameters; adjusting an optimal model to estimate the intensity of the precipitation of the test set; compared with the traditional rainfall estimation method, the method can well combine meteorological factors which are favorable for rainfall estimation in multiple scales, inhibit adverse factors to a certain extent, and the designed model can learn the microphysical characteristics in a rainfall field, effectively utilize the meteorological factors and the geographic factors, combine with radar reflectivity factors, reduce errors and perform more accurate estimation.
Description
Technical Field
The invention relates to the technical field of deep learning quantitative precipitation estimation, in particular to a rainfall intensity estimation method for comprehensive multi-time space-scale Doppler radar data.
Background
Rainfall intensity estimation is an important research direction in meteorology, and is closely related to our daily lives. In recent years, strong convection weather in summer is more frequent, natural disasters such as urban waterlogging, flood, debris flow and the like occur, serious threat is caused to life and property safety of people, and serious economic loss is caused to the country. The Doppler radar plays an increasing role in rainfall intensity estimation, and accurate and timely rainfall intensity estimation has great significance for disaster prevention and reduction.
However, the rainfall process is particularly complex, and high resolution, high accuracy rainfall intensity estimation is a challenging task. In the prior radar precipitation estimation research, small-scale radar reflectivity factors around the rain gauge are often concerned, and influence of weather factors and geographical environments around the rain gauge on rainfall is also rarely considered.
Disclosure of Invention
The invention aims to provide a rainfall intensity estimation method for synthesizing multi-time space-scale Doppler radar data, which aims to solve the problems in the background technology.
The technical scheme of the invention is as follows: the rainfall intensity estimation method for synthesizing the multi-time space-scale Doppler radar data comprises the following steps:
s1, doppler single polarization radar data and ground automatic weather station observation data are obtained;
s2, acquiring radar reflectivity factors, meteorological factors and geographic factors;
s3, adopting combined reflectivity, converting radar data in a polar coordinate system into grid data in a Cartesian coordinate system, then using inverse distance weighted interpolation regulation data, removing noise by using a Markov distance, performing quality control on the radar data, and acquiring accurate longitude and latitude data of a radar and the like;
s4, acquiring data sets of 1 multiplied by 400 of all stations at all moments under a Cartesian coordinate system, cutting to obtain three single-layer data sets, carrying out normalization processing on all data, and combining the three single-layer data sets into one data sample to be stored;
s5, dividing data into a test set and a training set according to the proportion of 2:8 by taking the actual precipitation of the station as a ground truth value tag, and finally storing all the data in a matrix form;
s6, establishing a rainfall intensity estimation model designed by using a deep learning technology;
s7, initializing the weight, training times, learning rate and learning rate attenuation coefficient of each neuron of the model, obtaining a precipitation estimated value through a feature extraction network and a fully-connected neural network, and calculating loss of a prediction result; obtaining an optimal network model and parameters;
s8, taking the data in the test set as input layer data, and inputting the input layer data into a network model to obtain corresponding predicted precipitation data;
s9, selecting an evaluation index for measuring the performance of the model, measuring the correlation between the true value and the estimated value, and respectively analyzing in time and space dimensions according to the result to obtain an optimal result.
Further, in S1, doppler radar base data and ground station precipitation data are obtained in the chinese weather data network, respectively.
Further, in S2, the radar reflectivity factor is a main input, the weather factor and the geographic factor are input in an auxiliary mode, the weather factor mainly uses temperature and humidity, and the geographic factor adopts elevation to perform preliminary pretreatment on data.
Further, in S3, the grid point closest to the national weather station is selected as the center of the reflectivity factor, and the radar-based data in polar coordinates is then converted into grid data in cartesian coordinates by adopting multi-scale input.
Further, in S3, the step of removing noise includes: denoising and filtering pixel points smaller than 70 by using a conventional echo picture processing method.
Further, in S4, the data sets of 1×400×400 of all sites in the cartesian coordinate system at all times are obtained, the single-layer data sets of 1×100×100, 1×50×50 and 1×25×25 are obtained by cutting with the national weather site as the center, all the data are normalized, and the three single-layer data sets are combined into one data sample to be stored.
Further, in S6, the radar data image precipitation feature extraction is performed by using mixed cavity convolution, the downsampling is performed by using maximum pooling, redundant information is removed, the feature is compressed, the network complexity is simplified, the receptive field of the high-level network is increased by using the non-local module, and the acquired information distribution is wider. The designed multi-scale attention module is used for balancing a large-scale image and a small-scale image which are centered on a site.
Further, in S7, a weighted combination of Mean Square Error (MSE) and average absolute error (MAE) is used as a loss function to calculate a loss for the prediction result; and (3) back propagation is carried out by utilizing a neural network, each weight gradient is calculated, the weights are updated according to a gradient descent algorithm, the weights of the neurons are continuously adjusted until the error of the training set is within a reasonable range, and the network training is stopped, so that the optimal network model and parameters are obtained.
Further, in S9, root Mean Square Error (RMSE), mean Absolute Error (MAE) and Correlation Coefficient (CC) are used as evaluation indexes for measuring the performance of the model.
The rainfall intensity estimation method for the comprehensive multi-time space-scale Doppler radar data provided by the invention is improved, and compared with the prior art, the rainfall intensity estimation method has the following improvement and advantages:
the method comprises the following steps: the radar reflectivity factor is used as main input, the meteorological factors and the geographic factors are used as auxiliary input variables, the meteorological factors adopt temperature and humidity, the geographic factors adopt elevation, the data are processed, the training set training model is utilized, the parameters are adjusted, and finally the model which can be applied to actual rainfall intensity estimation is obtained. The model reasonably utilizes historical rainfall observation data, improves the accuracy of rainfall intensity estimation, reasonably refers to the influence of cloud cluster motion paths and cloud cluster sizes on rainfall values measured by the rainfall gauge and the influence of geographical environments around the rainfall gauge on rainfall, can estimate regional rainfall intensity, and has good application prospect;
and two,: the model is integrally based on a convolutional neural network, and a brand-new multi-scale self-attention module is designed to better integrate factors which are beneficial to rainfall estimation in different scales. Under the condition of keeping the network result unchanged, the input data has a plurality of good dimensions, the difference value is carried out on the meteorological factors in the range according to a spherical model by using a Kriging interpolation method according to the meteorological factor data of the national meteorological station, and then the difference value and the reflectivity factor are matched in time and space and are used as input together, so that the influence of the multi-scale space-time data on the actual precipitation of the station is reasonably considered;
and thirdly,: the invention improves the accuracy of estimating the precipitation amount according to the radar data, and proves the effectiveness of taking the radar reflectivity factors with multiple scales and the weather and geographic factors as auxiliary input variables in the rainfall intensity estimation task by using a deep learning technology. The designed multi-scale attention module can better combine meteorological factors which are favorable for rainfall estimation in multiple scales, inhibit adverse factors to a certain extent, the designed model can learn micro-physical characteristics in a rainfall field, effectively utilize meteorological factors and geographic factors, combine with radar reflectivity factors, reduce errors and perform more accurate estimation.
Drawings
The invention is further explained below with reference to the drawings and examples:
FIG. 1 is a schematic flow chart of the method of the present invention;
FIG. 2 is a schematic diagram of a hybrid cavity convolution in the present embodiment;
FIG. 3 is a schematic diagram of a non-local module structure in the present embodiment;
FIG. 4 is a schematic diagram of a multi-scale attention module structure in the present embodiment;
FIG. 5 is a schematic diagram of the model structure in this embodiment;
fig. 6 is a schematic diagram of the full connection layer structure of the present embodiment.
Detailed Description
The following detailed description of the present invention clearly and fully describes the technical solutions in the embodiments of the present invention, and it is obvious that the described embodiments are only some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
Aiming at the defect of insufficient precision of the traditional precipitation amount estimation method, the invention aims to solve the problem of improving the precision of estimating precipitation data according to radar data and designs a deep learning model of a multi-scale self-attention module. The effectiveness of multi-scale radar reflectivity factors, as well as weather and geographic factors, as covariates in quantitative precipitation estimation tasks was demonstrated. Because the change of the cloud cluster affects the rain gauge, a large-scale characteristic diagram in multiple scales is adopted to learn the complex change and movement condition of the cloud cluster in a wider area, and a small-scale characteristic diagram is adopted to learn the spatial information of stronger correlation between the vicinity of the rain gauge and precipitation. Taking into account the spatial correlation of weather and geographic factors, two-dimensional weather and geographic factors are employed as auxiliary variables to capture their spatial characteristics. The designed self-attention mechanism module better combines factors which are favorable for rainfall estimation in multiple scales, and suppresses adverse factors to a certain extent. Compared with a traditional rainfall estimation method, which takes a single scale as an input deep learning model, the model learns the micro-physical process characteristics in a rainfall field, effectively utilizes meteorological factors and geographic factors, combines the meteorological factors and radar reflectivity factors, objectively characterizes the rainfall phenomenon, further reduces errors and obtains more accurate rainfall intensity estimation.
The specific technical scheme of the invention is as follows:
as shown in fig. 1, the rainfall intensity estimation method for synthesizing the multi-time space-scale doppler radar data comprises the following steps:
s1, respectively acquiring Doppler radar reflectivity data and ground station precipitation data in a Chinese meteorological data network (http:// data. Cma. Cn) according to historical precipitation data;
s2, taking radar reflectivity factors as main input, and weather factors and geographic factors as auxiliary input, wherein the weather factors mainly adopt temperature and humidity, and the geographic factors adopt elevation, so that preliminary pretreatment is carried out on data;
s3, adopting combined reflectivity, converting radar data in a polar coordinate system into grid data in a Cartesian coordinate system, and then using inverse distance weighted interpolation regulation data;
p is any positive real number, typically, p=2;h i is the distance from the discrete point to the differential point(x, y) is the difference point coordinates, (x) i ,y i ) Is a discrete point coordinate.
R is the distance from the interpolation point to the furthest discrete point, and n is the total number of discrete points. Removing noise by utilizing the mahalanobis distance, performing quality control on radar data, finally obtaining accurate longitude and latitude data of the radar and the like, and executing the step 4;
s4, acquiring data sets of 1 multiplied by 400 of all stations at all moments under a Cartesian coordinate system, cutting to obtain three single-layer data sets, carrying out normalization processing on all data, and combining the three single-layer data sets into one data sample to be stored;
s5, dividing data into a test set and a training set according to the proportion of 2:8 by taking the actual precipitation of the station as a ground truth value tag, and finally storing all the data in a matrix form;
s6, establishing a rainfall intensity estimation model designed by using a deep learning technology, and executing a step 7;
in this embodiment, the extraction of the radar data image features is performed using hybrid hole convolution. The mixed cavity convolution adds zero pixels into the characteristic mapping of the standard convolution kernel to fill, so that the calculated amount is reduced, and the purpose of expanding the receptive field is achieved. Compared with the common convolution, the cavity convolution can improve the resolution of the sampled image under the condition of not increasing the parameter quantity, and realizes dense feature extraction in depth CNN. For a common convolution kernel with the size of K, the corresponding hole convolution kernel is K+ (K-1) (R-1), wherein R is the hole rate when the feature map is sampled. The invention builds the network by adopting a mixed cavity convolution (HDC) mode, avoids the grid effect from breaking the continuity between local information, and can sample a complete area of the original characteristic diagram. I.e. for a number N of convolutional layers, eachThe convolution kernel of one layer has a size K and an expansion rate r 1 ,r 2 ,...,r i ]Its maximum expansion rate needs to satisfy the following formula:
M i ＝max[M i+1 -2r i ,M i+1 -2(M i+1 -r i ),r i ]
wherein r is i For the expansion ratio of the ith layer, M i Is the maximum expansion rate of the i-th layer. With HDC, the receptive field can be enlarged without losing local information, capturing more global information, as shown in fig. 2.
The non-local module considers all positions, but convolution and sequence cannot consider so much information, the non-local module can directly calculate interaction of two positions without considering the distance problem, the efficiency is high, the effect is good, only a few layers are used, the input scale is various, the input scale is easy to combine with other models, the global space-time characteristics can be captured, different weights are distributed, finally, the convolution and sequence cannot be realized at each position, the specific structure is shown in fig. 3, and the equation is as follows:
where x is the input, y is the output, and the f function computes the feature of the ith location of x and the feature similarity of the jth location of x. The g function computes a representation of the features of the jth position of x, C (x) being used for normalization. It can be seen that the feature of the ith position of y is a weighted average of the features of all positions of x. When the f-function selects an ededdgaussian,is equal to->Solving for softMax in the j dimension, therefore
Also in self-attentive form. In the present invention, we use a 1 x 1 convolution operation to implement three different feature mappings of the output. Multiplying the different modules to obtain similarity scores between every two global pixel points, and converting the similarity scores into weight scores of global information for each pixel point through a softMax function. Z for output of each position i Is a weighted sum of global information.
The addition of an input as a residual term in the equation makes the non-local module more stable.
The multiscale attention module balances a large-scale image and a small-scale image which are centered on a site, and receives two inputs, namely a small-scale characteristic image x M And the large-scale feature map xL, the specific model structure is shown in fig. 4, the small-scale feature map is used as a Key module through feature mapping, and the large-scale feature map is used as a Query module through feature mapping. Multiplying the Key module and the Query module to obtain x M And x L A pixel-by-pixel similarity scoring matrix, as follows:
to make the module more stable, the input is often connected at the end of the model as shortcut;
s7, initializing the weight, training times, learning rate and learning rate attenuation coefficient of each neuron of the model, obtaining a precipitation estimated value through a feature extraction network and a fully connected neural network, and calculating loss of a prediction result by adopting a weighted combination of Mean Square Error (MSE) and average absolute error (MAE) as a loss function; counter-propagating by utilizing a neural network, calculating each weight gradient, updating the weights according to a gradient descent algorithm, continuously adjusting the weights of the neurons until the error of a training set is within a reasonable range, stopping network training, and obtaining an optimal network model and parameters, as shown in fig. 5;
s8, data in the test set are used as input layer data and are input into the network model, and corresponding predicted precipitation data are obtained, as shown in FIG. 6;
s9, adopting Root Mean Square Error (RMSE), mean Absolute Error (MAE) and Correlation Coefficient (CC) as evaluation indexes for measuring the performance of the model, wherein the specific formula is as follows
And measuring the correlation between the true value and the estimated value by using a Correlation Coefficient (CC), and analyzing in time and space dimensions according to the result to obtain an optimal result.
In S9, root Mean Square Error (RMSE), mean Absolute Error (MAE) and Correlation Coefficient (CC) are used as evaluation indexes for measuring performance of the model, correlation between the real value and the estimated value is measured by using the Correlation Coefficient (CC), and analysis is performed on time and space dimensions according to the result to obtain an optimal result. MAE is less affected by outliers and can reflect the overall error of the true and estimated values. The correlation of the actual and estimated values for CC was measured and the study was such that the larger the CC values, the smaller the RMSE and MAE values represent the superiority of the final model.
Specifically, specific parameters of the model are set: the total training times are 500 times, the initial value of the learning rate is set to be 0.0001, the learning rate is dynamically changed, when the learning rate exceeds 20 rounds and does not change, the training can be automatically stopped, the model is ensured to quickly converge, the diverging effect of the model is not caused, all data are normalized, and the convergence rate of the model is further improved.
By comparing with the traditional Z-R relationship, the Z-R relationship is found to be not very good for fitting the relationship between the reflectivity factor and the rainfall, and has larger difference with the ground true value. But the performance of the used BPN network is superior to that of the traditional Z-R relation, and the advantages of the deep learning method in fitting the rainfall and the reflectivity factor are effectively proved. However, the CNN captures the spatial structure ignored by the BP network through CNN model comparison results. For different CNN methods, weather factors and geographic factors are added as auxiliary variables, and more accurate rainfall values can be obtained compared with the method that the reflectivity factors are taken as the input, so that the correlation of rainfall and weather and geographic space environments is further verified.
The previous description is provided to enable any person skilled in the art to make or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other embodiments without departing from the spirit or scope of the invention. Thus, the present invention is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims (5)
1. The rainfall intensity estimation method for the comprehensive multi-time space-scale Doppler radar data is characterized by comprising the following steps of: the method comprises the following steps:
s1, doppler single polarization radar data and ground automatic weather station observation data are obtained;
s2, acquiring radar reflectivity factors, meteorological factors and geographic factors, wherein the radar reflectivity factors are used as main inputs, the meteorological factors and the geographic factors are used as auxiliary inputs, the meteorological factors use temperature and humidity, and the geographic factors adopt elevation to perform preliminary pretreatment on data;
s3, adopting combined reflectivity, selecting a grid point closest to a national weather station as the center of a reflectivity factor, converting radar data in a polar coordinate system into grid data in a Cartesian coordinate system, using inverse distance weighted interpolation regulation data, removing noise by utilizing a Markov distance, performing quality control on the radar data, and acquiring accurate radar longitude and latitude data;
s4, acquiring single-layer 1×400×400 data sets of all stations in a Cartesian coordinate system at all times, cutting by taking national weather stations as the center, respectively acquiring single-layer data sets with the sizes of 1×100×100 and 1×25×25, carrying out normalization processing on all the data, and combining the three single-layer data sets into one data sample for storage;
s5, dividing data into a test set and a training set according to the proportion of 2:8 by taking the actual precipitation of the station as a ground truth value tag, and finally storing all the data in a matrix form;
s6, a rainfall intensity estimation model designed by using a deep learning technology is established, mixed cavity convolution is used for radar data image precipitation feature extraction, maximum pooling is used for downsampling, redundant information is removed, features are compressed, network complexity is simplified, specifically, for a common convolution kernel with the size of K, the corresponding cavity convolution kernel is K+ (K-1) ×R-1, wherein R is the cavity rate when a feature map is sampled, for convolution layers with the number of N, the convolution kernel size of each layer is K, the expansion rate is [ R1, R2, …, ri ], and the maximum expansion rate needs to meet the following formula:
M i ＝max[M i+1 -2r i ，M i+1 -2(M i+1 -r i )，r i ]
where ri is the expansion rate of the i-th layer and Mi is the maximum expansion rate of the i-th layer;
the non-local module is used for increasing the receptive field of a high-level network, the acquired information is distributed more widely, specifically, the non-local module directly calculates interaction of two positions, only a few layers are used, input scales are various, the interaction is combined with other models, global space-time characteristics are captured, different weights are distributed, and finally the interaction is aggregated at each position, wherein the equation is as follows:
where x is the input and y is the output, the f function computes the feature of the ith location of x and the feature similarity of the jth location of x, the g function computes a representation of the feature of the jth location of x, and C (x) is used for normalization; it can be seen that the feature of the ith position of y, which is a weighted average of the features of all positions of x, when the f function selects an ededdgaussian,equal to->Solving for softMax in the j dimension, therefore
Also in self-attentive form; the three different feature mappings of the output are realized by using a 1 multiplied by 1 convolution operation, different modules are multiplied to obtain similarity scores between every two global pixel points, the similarity scores are converted into weight scores of global information for each pixel point through a softMax function, and the output of each position is realized by Z i Is expressed as a weighted sum of global information;
the non-local module is more stable by adding input as a residual term in the formula;
balancing site-centric large scale graphs using a multi-scale attention moduleImage and small-scale image, in particular, multi-scale attention module, accepts two inputs, namely small-scale feature map x M And large scale feature map x L The small-scale feature map is used as a Key module through feature mapping, and the large-scale feature map is used as a Query module through feature mapping; multiplying the Key module and the Query module to obtain x M And x L A pixel-by-pixel similarity scoring matrix, as follows:
G i，j ＝(W k *(x M ) i ) T *(W Q *(x L ) j )
connecting the input as shortcut at the end of the model;
s7, initializing the weight, training times, learning rate and learning rate attenuation coefficient of each neuron of the model, obtaining a precipitation estimated value through a feature extraction network and a fully-connected neural network, and calculating loss of a prediction result; obtaining an optimal network model and parameters;
s8, taking the data in the test set as input layer data, and inputting the input layer data into a network model to obtain corresponding predicted precipitation data;
s9, selecting an evaluation index for measuring the performance of the model, measuring the correlation between the true value and the estimated value, and respectively analyzing in time and space dimensions according to the result to obtain an optimal result.
2. The method for estimating rainfall intensity of integrated multi-time-space-scale doppler radar data according to claim 1, characterized by: in S1, doppler radar base data and ground station precipitation data are respectively acquired in a China weather data network.
3. The method for estimating rainfall intensity of integrated multi-time-space-scale doppler radar data according to claim 1, characterized by: in S3, the step of removing noise includes: denoising and filtering pixel points smaller than 70 by using a conventional echo picture processing method.
4. The method for estimating rainfall intensity of integrated multi-time-space-scale doppler radar data according to claim 1, characterized by: s7, adopting a weighted combination of mean square error and average absolute error as a loss function, and calculating loss for a prediction result; and (3) back propagation is carried out by utilizing a neural network, each weight gradient is calculated, the weights are updated according to a gradient descent algorithm, the weights of the neurons are continuously adjusted until the error of the training set is within a reasonable range, and the network training is stopped, so that the optimal network model and parameters are obtained.
5. The method for estimating rainfall intensity of integrated multi-time-space-scale doppler radar data according to claim 1, characterized by: in S9, the root mean square error, the average absolute error and the correlation coefficient are used as evaluation indexes for measuring the performance of the model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210417830.9A CN114742206B (en) | 2022-04-20 | 2022-04-20 | Rainfall intensity estimation method for comprehensive multi-time space-scale Doppler radar data |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210417830.9A CN114742206B (en) | 2022-04-20 | 2022-04-20 | Rainfall intensity estimation method for comprehensive multi-time space-scale Doppler radar data |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114742206A CN114742206A (en) | 2022-07-12 |
| CN114742206B true CN114742206B (en) | 2023-07-25 |
Family
ID=82282727
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210417830.9A Active CN114742206B (en) | 2022-04-20 | 2022-04-20 | Rainfall intensity estimation method for comprehensive multi-time space-scale Doppler radar data |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114742206B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116824372B (en) * | 2023-06-21 | 2023-12-08 | 中国水利水电科学研究院 | Urban rainfall prediction method based on Transformer |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107807907A (en) * | 2017-09-08 | 2018-03-16 | 中国电力科学研究院 | A kind of precipitation classification method and system |
| CN109782375A (en) * | 2019-01-07 | 2019-05-21 | 华东交通大学 | Precipitation intensity estimating and measuring method and system based on big data |
| CN110346844A (en) * | 2019-07-15 | 2019-10-18 | 南京恩瑞特实业有限公司 | Quantitative Precipitation estimating and measuring method of the NRIET based on cloud classification and machine learning |
| CN112965146A (en) * | 2021-04-14 | 2021-06-15 | 中国水利水电科学研究院 | Quantitative precipitation estimation method combining meteorological radar and rainfall barrel observation data |
| CN113791415A (en) * | 2021-09-15 | 2021-12-14 | 南京信息工程大学 | Radar quantitative precipitation estimation method based on deep learning |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111625993B (en) * | 2020-05-25 | 2020-12-18 | 中国水利水电科学研究院 | Small watershed surface rainfall interpolation method based on mountainous terrain and rainfall characteristic prediction |
| CN113936142A (en) * | 2021-10-13 | 2022-01-14 | 成都信息工程大学 | Rainfall approach forecasting method and device based on deep learning |
-
2022
- 2022-04-20 CN CN202210417830.9A patent/CN114742206B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107807907A (en) * | 2017-09-08 | 2018-03-16 | 中国电力科学研究院 | A kind of precipitation classification method and system |
| CN109782375A (en) * | 2019-01-07 | 2019-05-21 | 华东交通大学 | Precipitation intensity estimating and measuring method and system based on big data |
| CN110346844A (en) * | 2019-07-15 | 2019-10-18 | 南京恩瑞特实业有限公司 | Quantitative Precipitation estimating and measuring method of the NRIET based on cloud classification and machine learning |
| CN112965146A (en) * | 2021-04-14 | 2021-06-15 | 中国水利水电科学研究院 | Quantitative precipitation estimation method combining meteorological radar and rainfall barrel observation data |
| CN113791415A (en) * | 2021-09-15 | 2021-12-14 | 南京信息工程大学 | Radar quantitative precipitation estimation method based on deep learning |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114742206A (en) | 2022-07-12 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US20220043182A1 (en) | Spatial autocorrelation machine learning-based downscaling method and system of satellite precipitation data | |
| Si et al. | Hybrid solar forecasting method using satellite visible images and modified convolutional neural networks | |
| CN110363327B (en) | ConvLSTM and 3D-CNN-based short rainfall prediction method | |
| CN109001736B (en) | Radar echo extrapolation method based on deep space-time prediction neural network | |
| CN111368769B (en) | Ship multi-target detection method based on improved anchor point frame generation model | |
| CN114139690A (en) | Short-term rainfall prediction method and device | |
| CN113936142A (en) | Rainfall approach forecasting method and device based on deep learning | |
| CN111145245A (en) | Short-time approaching rainfall forecasting method and system and computer readable storage medium | |
| CN113255972B (en) | Short-term rainfall prediction method based on Attention mechanism | |
| CN112415521A (en) | CGRU (China-swarm optimization and RU-based radar echo nowcasting) method with strong space-time characteristics | |
| CN114742206B (en) | Rainfall intensity estimation method for comprehensive multi-time space-scale Doppler radar data | |
| CN115393712B (en) | SAR image road extraction method and system based on dynamic hybrid pooling strategy | |
| CN110933633A (en) | Onboard environment indoor positioning method based on CSI fingerprint feature migration | |
| CN115691049A (en) | Convection birth early warning method based on deep learning | |
| CN117131991A (en) | Urban rainfall prediction method and platform based on hybrid neural network | |
| CN116699731A (en) | Tropical cyclone path short-term forecasting method, system and storage medium | |
| CN116720156A (en) | Weather element forecasting method based on graph neural network multi-mode weather data fusion | |
| CN114663307B (en) | Integrated image denoising system based on uncertainty network | |
| CN113341419B (en) | Weather extrapolation method and system based on VAN-ConvLSTM | |
| CN116012618A (en) | Weather identification method, system, equipment and medium based on radar echo diagram | |
| CN115049013A (en) | Short-term rainfall early warning model fusion method combining linearity and SVM | |
| CN116307267B (en) | Rainfall prediction method based on convolution | |
| CN110727719A (en) | Lightning positioning data assimilation method based on dynamic relaxation approximation | |
| Zhou et al. | Forecasting the western Pacific subtropical high index during typhoon activity using a hybrid deep learning model | |
| CN116580328B (en) | Intelligent recognition method for leakage danger of thermal infrared image dykes and dams based on multitasking assistance |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |