CN111047099A

CN111047099A - Regional torrential flood risk prediction method and system

Info

Publication number: CN111047099A
Application number: CN201911294556.5A
Authority: CN
Inventors: 宋杰; 张亮; 董梅; 胡辉
Original assignee: Hangzhou Ruhr Technology Co Ltd
Current assignee: Hangzhou Ruhr Technology Co Ltd
Priority date: 2019-12-16
Filing date: 2019-12-16
Publication date: 2020-04-21
Anticipated expiration: 2039-12-16
Also published as: CN111047099B

Abstract

The invention discloses a regional torrential flood risk prediction method and a regional torrential flood risk prediction system, wherein the prediction method comprises the following steps: s1, forecasting weather information of all hills in the area based on the meteorological model; s2, screening out hills needing risk prediction based on the weather information; s3, dividing the hills needing risk prediction into different risk prediction grades according to the basic information of the hills; different risk prediction grades correspond to different risk prediction periods; s4, preliminarily determining a disaster-causing factor for mountain torrent evaluation; s5, screening the preliminarily determined disaster-causing factors to obtain main disaster-causing factors affecting the torrential flood; s6, training and generating a plurality of torrential flood risk prediction models based on the main disaster factors; s7, combining the several risk prediction models with the best selectivity to form a final risk prediction model; and S8, performing mountain flood risk prediction on the hill to be predicted. The method and the device realize the risk prediction of the regional hills, have low cost, wide coverage and high processing efficiency, and improve the safety of the hills.

Description

Regional torrential flood risk prediction method and system

Technical Field

The invention relates to the technical field of mountain torrent disaster analysis, in particular to a regional mountain torrent risk prediction method and system.

Background

The mountain torrent disasters in China occur frequently, and the safety of life, property and the like of people is seriously threatened. According to statistics, the loss of the country and the lives and properties of people caused by mountain torrent disasters in China accounts for about 40 percent of the total loss caused by natural disasters. Torrential flood generally refers to a flood of water occurring in small areas (typically within hundreds of km 2) of hilly areas, with sudden rises and falls (tens of minutes to hours) induced by heavy rainfall. The dangerousness of the torrential flood is mainly manifested by paroxysmal property, concentrated water quantity, large flow velocity and strong scouring force, and a large amount of silt and stone and the like are often wrapped and carried to form a debris flow and a landslide, so that the torrential flood has strong destructiveness. Therefore, research on risk assessment and distribution rules of the torrential flood is urgently needed to be comprehensively developed, the torrential flood is prevented, and the influence degree of loss is reduced.

In the existing mountain torrent forecasting and early warning method, a rainfall critical value of an area where mountain torrents occur is mostly counted by historical rainfall at home, a dynamic critical rainfall value is mostly taken as a threshold value at foreign countries, namely, the rainfall occurring on a small watershed is calculated and analyzed through a hydrological model to obtain real-time soil humidity of the watershed, the rainfall required by the situation that the peak flow of the outlet section of the watershed reaches a preset early warning flow value is reversely deduced, and when the rainfall reaches the value, mountain torrent early warning is issued. However, the problems of fast mountain torrent flow rate, short forecast period, data shortage, different prediction model and conventional hydrological prediction thinking and the like still exist in the simulation process. In addition, the existing application distributed hydrological model constructs a digital drainage basin by using a Digital Elevation Model (DEM) generated by a Geographic Information System (GIS), and the process of water level, flow, silt and the like of an early warning section is obtained by calculating the convergence of drainage basin production, so that the time and the level of mountain torrents outbreak can be conveniently mastered, and the model can be applied to mountainous areas and other areas without data or lack of monitoring data. However, since the model is established on the motion wave and the diffusion wave, although the internal water level and the flow velocity process in the mountain torrent area can be forecasted, the formation and the evolution process of the mountain torrent are often difficult to accurately describe due to neglect of flowing mechanical factors, and further the reliability of the hydrographic forecasting of the mountain torrent disaster is low. The duration of rainstorm inducing mountain torrents is often short, and the linear relation between rainfall and runoff can not be achieved in a short time, so that the hydrological method based on unit lines is not completely applicable. The existing torrential flood forecasting model based on hydrodynamics and hydrosand dynamics is still in a theoretical research stage, and is practically applied to torrential flood forecasting.

The monitoring processing efficiency of a pair of single-point hills one by one is low, the coverage is small, the regional mountain torrent risk prediction technology fuses the multivariate data, the mountain torrent risk development rule on the analysis area is excavated, the mountain torrent hidden danger points are investigated, the hills to be monitored are identified from numerous hills, and the low-cost, wide-coverage and high-efficiency risk monitoring is realized. At present, no effective prediction and evaluation method is available for regional risk of mountain torrent health monitoring.

Therefore, how to realize regional torrent risk prediction with high precision and high efficiency and convert passive monitoring and processing of torrent risk disasters into active prediction and response is a problem to be solved in the field.

Disclosure of Invention

The invention aims to provide a regional torrential flood risk prediction method and system aiming at the defects of the prior art. The method and the device realize the risk prediction of the regional hills, have low cost, wide coverage and high processing efficiency, and improve the safety of the hills.

In order to achieve the purpose, the invention adopts the following technical scheme:

a regional torrential flood risk prediction method comprises the following steps:

s1, forecasting weather information of all hills in the area based on the meteorological model;

s2, screening hills in the area based on the weather information, and selecting hills needing to be subjected to mountain flood risk prediction;

s3, dividing the hills needing to be subjected to the mountain torrent risk prediction into different risk prediction grades according to the basic information of the hills; predicting different hills according to the risk prediction periods corresponding to the risk prediction grades;

s4, preliminarily determining disaster-causing factors for mountain torrent risk evaluation from the data of the national departments, the national meteorological information center and literature research;

s5, screening the preliminarily determined disaster-causing factors to obtain main disaster-causing factors affecting the torrential flood;

s6, collecting historical torrential flood data, and training a logistic regression model, a random forest, a nearest neighbor classification, a K mean classification, a Bayesian method and a deep learning model based on a water body recognition technology respectively based on the determined main disaster causing factors to generate a plurality of torrential flood risk prediction models;

s7, evaluating the performance of a logistic regression model, a random forest, a nearest neighbor classification, a K mean classification, a Bayesian method and a torrential flood risk prediction model generated by deep learning model training based on a water body recognition technology, and selecting several risk prediction models with the best performance to jointly form a final risk prediction model;

s8, the values of main disaster-causing factors corresponding to the hills needing risk prediction are collected through mutual integration and application of a high-precision remote sensing monitoring technology and an internet of things sensing technology, the values of the main disaster-causing factors are input into a final risk prediction model formed in a combined mode, and the risk values of the hills and the torrential flood are predicted.

Further, the dividing into different risk prediction levels according to the basic information is specifically:

setting the weight of historical rainfall, gradient, vegetation characteristics, tributary (ditch) water system distribution condition and mountain torrents influence range as omega in sequence₁、ω₂、ω₃、ω₄、ω₅Wherein, ω is₁+ω₂+ω₃+ω₄+ω₅1 is ═ 1; the mountain torrent risk assessment value that the hillock corresponds is:

V＝V₁*ω₁+V₂*ω₂+V₃*ω₃+V₃*ω₄+V₅*ω₅

wherein, V₁、V₂、V₃、V₄、V₅The values of historical rainfall, gradient, vegetation characteristics, branch water system distribution condition and mountain torrents influence range are respectively; the value of the historical rainfall is the ratio of the historical rainfall of the hill to the average value of the historical rainfall nationwide; the values of the slopes are respectively endowed with different slopes, and the steeper the slope is, the higher the corresponding slope value is; the vegetation characteristics are values endowed with different vegetation coverage rates, the higher the vegetation coverage rate is, the lower the corresponding vegetation characteristic value is, the more tributary water systems are, and the lower the corresponding tributary water system distribution value is; the values of the torrential flood influence ranges are values respectively endowed with different influence ranges, and the larger the influence on human, the higher the corresponding torrential flood influence range value.

And dividing the hill into corresponding risk prediction levels based on the calculated risk assessment values, wherein the higher the risk assessment value is, the higher the risk prediction level is.

Further, the step S5 is specifically:

selecting important disaster-causing factors from the preliminarily determined disaster-causing factors according to the characteristics of the disaster-causing factors, wherein the characteristics comprise variance, inclination, peak state, frequency, vibration mode, modal curvature, regression residual, wavelet energy and fitting coefficients;

screening the important disaster-causing factors through correlation analysis, conditional entropy, posterior probability and logistic regression weight, selecting the most useful feature subset according to variable forecasting force, and respectively extracting static factors and dynamic factors from the disaster-causing factors; the static factors include elevation, gradient, slope, geology, landform, vegetation, land utilization, unobstructed river channels, population density and economic loss, and the dynamic factors include rainfall, rainfall intensity, rainfall duration, water level, soil humidity and human engineering.

Further, setting a threshold value of the variance, and when the variance of the disaster-causing factor is smaller than the threshold value of the variance, rejecting the disaster-causing factor; otherwise, the characteristic is taken as a relatively important disaster-causing factor;

calculating the correlation coefficient of the more important disaster causing factor and the torrential flood, setting the threshold value of the correlation coefficient, and rejecting the disaster causing factor when the correlation coefficient of the disaster causing factor and the torrential flood is smaller than the threshold value of the correlation coefficient; otherwise, the characteristic is taken as a main disaster-causing factor.

Further, the step S6 includes:

the logistic regression model has the parameter form:

wherein the content of the first and second substances,

n is the number of independent variables; and (3) forming a characteristic vector by using characteristic values corresponding to main disaster-causing factors in historical torrential.

Further, the step S6 includes:

constructing a convolutional neural network, wherein the convolutional neural network consists of an input layer, a convolutional layer, a PReLU layer, a pooling layer, a full-link layer and an output layer; the method specifically comprises 5 convolutional layers, each convolutional layer is attached with a nonlinear activation function PReLU layer with parameters, a pooling layer is connected behind a first convolutional layer, a second convolutional layer and a fourth convolutional layer, each pooling layer adopts a maximum pooling method, the first convolutional layer is connected with an input layer, the input layer inputs values corresponding to main disaster-causing factors in torrent data to be processed, a full-connection layer is located between the last pooling layer and an output layer, each neuron is connected with all neurons of the previous layer, and according to the requirement of risk detection, a feature vector is mapped to the output layer in a targeted manner, and the output layer outputs torrent risk values;

training a convolutional neural network model by using a large amount of historical torrent data, calculating a loss function of the convolutional neural network, iterating and updating the convolutional neural network by using the loss function, continuously training the convolutional neural network based on the risk prediction of deep learning, reducing the loss function to an expected value, and generating a final convolutional neural network model.

Further, the performance evaluation method of the torrential flood risk prediction model comprises the steps of mean square error, coefficient determination, accuracy rate, recall rate, accuracy rate, a receiving sensitivity curve and area under the curve.

And further, carrying out mountain torrent risk mapping and display by adopting ArcGIS, and carrying out image layer superposition on the collected mountain spatial data and the mountain torrent risk value after cleaning treatment.

Further, the risk prediction method further includes:

and (3) performing regional hill evaluation by adopting a sampling evaluation method: calculating the proportion of the number of hills corresponding to each mountain torrent early warning level occupying the total number of hills in the area, setting the total number of sampled samples, and calculating the number of the sampled samples corresponding to each mountain torrent early warning level according to the total number of the sampled samples and the proportion of each mountain torrent early warning level; randomly extracting a corresponding number of hills from the hills at each mountain flood early warning level; extracting hills which do not need to be subjected to the torrential flood risk prediction until the number of the extracted hills reaches the total number of the sampled samples; and calculating the overall mountain flood risk value of the region according to the sampled hill samples, and taking the average value of the mountain flood risk values of the hill samples as the mountain flood risk value of the regional hill.

The invention also provides a regional torrential flood risk prediction system, which is used for realizing the regional torrential flood risk prediction method, and comprises the following steps:

the weather information module is used for predicting weather information of all hills in the area based on the meteorological model;

the first screening module is used for screening hills in the area based on the weather information and selecting hills needing to be subjected to mountain flood risk prediction;

the grade division module is used for dividing the hills needing to be subjected to the mountain torrent risk prediction into different risk prediction grades according to the basic information of the hills; predicting different hills according to the risk prediction periods corresponding to the risk prediction grades;

the disaster-causing factor determination module is used for primarily determining disaster-causing factors for mountain torrent risk evaluation from the data of the national and local departments, the national meteorological information center and the literature investigation;

the second screening module is used for screening the preliminarily determined disaster-causing factors to obtain main disaster-causing factors influencing the mountain torrents;

the training module is used for acquiring historical torrential flood data, and training a logistic regression model, a random forest, a nearest neighbor classification, a K mean classification, a Bayesian method and a deep learning model based on a water body recognition technology respectively based on the determined main disaster causing factors to generate a plurality of torrential flood risk prediction models;

the evaluation module is used for evaluating the performance of a logistic regression model, a random forest, a nearest neighbor classification, a K mean value classification, a Bayesian method and a torrential flood risk prediction model generated by deep learning model training based on a water body recognition technology, and a final risk prediction model is formed by combining a plurality of risk prediction models with the best selectivity;

and the prediction module is used for acquiring the values of main disaster-causing factors corresponding to the hills needing risk prediction through the mutual integration application of the high-precision remote sensing monitoring technology and the Internet of things sensing technology, inputting the values of the main disaster-causing factors into a final risk prediction model formed by combination, and predicting the risk values of the hills and the torrential flood.

Compared with the prior art, the invention has the following advantages:

(1) the regional mountain torrent risk prediction method based on the multi-point risk prediction solves the problem that the existing single-point risk prediction is low in processing efficiency, and makes up the vacancy of the existing regional mountain torrent risk prediction and evaluation;

(2) due to the fact that single-point prediction processing of hills one by one is complex, existing mountain flood risk prediction is low in coverage, effective supervision on a large number of hills is not achieved, all hills in the area are supervised simultaneously, and the hill coverage rate of the mountain flood risk prediction is improved;

(3) according to the method, the hills needing risk prediction are screened, and the hills needing risk prediction are predicted according to different risk prediction grades and different risk prediction periods, so that the problems that the existing mountain torrent risk prediction method is single and solidification is processed are solved, and the complexity of mountain torrent risk prediction is reduced while safety supervision is carried out on a large number of hills;

(4) according to the method, a plurality of torrential flood risk prediction models are generated through training, combined risk prediction is performed by selecting the prediction models with the best performance, and the accuracy of risk prediction is improved by combining the advantages of the plurality of risk prediction models;

(5) according to the method, the ArcGIS is used for map making and displaying the mountain torrent risk, the map making of the mountain torrent easiness is carried out according to the mountain torrent early warning level corresponding to the mountain torrent risk, red, orange, yellow and blue are respectively used as early warning colors corresponding to the high, medium and low four-level early warning levels, the condition of a hill can be visually displayed, and the development trend of the mountain torrent risk can be detected; carrying out early warning of different levels according to different risks, and timely carrying out risk early warning processing;

(6) the sampling evaluation method is adopted to evaluate the regional hills, so that the overall level of the regional hills can be known, the hills with different risks can be treated, the risk of the hills is reduced, and the passive monitoring and treatment of the torrential flood risk disasters are converted into active prediction and response.

Drawings

Fig. 1 is a flowchart of a regional torrential flood risk prediction method according to an embodiment;

fig. 2 is a structural diagram of a regional torrential flood risk prediction system according to the second embodiment.

Detailed Description

The embodiments of the present invention are described below with reference to specific embodiments, and other advantages and effects of the present invention will be easily understood by those skilled in the art from the disclosure of the present specification. The invention is capable of other and different embodiments and of being practiced or of being carried out in various ways, and its several details are capable of modification in various respects, all without departing from the spirit and scope of the present invention. It is to be noted that the features in the following embodiments and examples may be combined with each other without conflict.

It should be noted that the drawings provided in the following embodiments are only for illustrating the basic idea of the present invention, and the components related to the present invention are only shown in the drawings rather than drawn according to the number, shape and size of the components in actual implementation, and the type, quantity and proportion of the components in actual implementation may be changed freely, and the layout of the components may be more complicated.

The invention is further described with reference to the following drawings and specific examples, which are not intended to be limiting.

Example one

As shown in fig. 1, the present embodiment provides a regional torrential flood risk prediction method, including:

and collecting weather information of all hills in the region, predicting the weather information by adopting a meteorological model, and predicting the weather condition of a future period of time. The invention does not define a specific meteorological model, and usually predicts weather for 7 days in the future.

china is a mountainous country, and the area of a hilly area is 2/3 which is about the area of the national soil and is far higher than the average level of the world. Since ancient times, China has a habit of living by mountains and water, so that hilly areas are densely populated, and disaster-bearing bodies are provided for the occurrence of mountain flood disasters. If all hills in the area are monitored, the cost of hardware and software is a great burden in mountain torrent risk prediction. Therefore, the method and the device firstly screen the hills needing to be subjected to the torrential flood risk prediction based on the weather information of the areas where the hills are located. For example, if the hill does not rain for a future period of time, the probability of the hill's risk of developing a mountain flood is extremely low, and thus, the mountain flood risk prediction is not performed. The screening of the monitored hills is not invariable and can be updated regularly according to weather information. The frequency of weather updates varies from season to season, e.g., the predicted frequency of weather in spring is lower than in summer because the frequency of rainfall in spring is lower than in summer.

the method and the device set different risk prediction levels, and the different risk prediction levels correspond to different risk prediction periods. Specifically, the hills are graded according to their basic information, and different weights are set for different basic information categories. The basic information of the hill includes the historical rainfall, the slope, the vegetation characteristics, the branch (trench) water system distribution, the mountain torrent influence range, and the like of the hill.

For example, the weight of setting the historical rainfall, gradient, vegetation characteristics, distribution of branch (ditch) water system, and mountain torrents influence range is ω in order₁、ω₂、ω₃、ω₄、ω₅Wherein, ω is₁+ω₂+ω₃+ω₄+ω₅1. Therefore, the mountain torrent risk assessment value corresponding to the hill is:

V＝V₁*ω₁+V₂*ω₂+V₃*ω₃+V₃*ω₄+V₅*ω₅

wherein, V₁、V₂、V₃、V₄、V₅The values of the historical rainfall, the gradient, the vegetation characteristics, the distribution condition of the branch (ditch) water system and the mountain torrents influence range are respectively. The value of the historical rainfall is the ratio of the historical rainfall of the hill to the average value of the historical rainfall nationwide; the values of the slopes are respectively endowed with different slopes, and the steeper the slope is, the higher the corresponding slope value is; the vegetation characteristics are values endowed with different vegetation coverage rates, the higher the vegetation coverage rate is, the lower the corresponding vegetation characteristic value is, correspondingly, the more tributary water systems are, the lower the corresponding tributary (ditch) water system distribution value is; the value of the mountain torrent influence range is respectively endowed with different influencesThe greater the impact on humans, the higher the corresponding torrent impact range value.

And dividing the hills into corresponding risk prediction levels based on the calculated mountain torrent risk assessment values. For example, three grades I, II and III of mountain torrent risk prediction are set, the mountain torrent risk assessment value corresponding to the grade I is [20, ∞ ], and the mountain torrent is a mountain torrent with large historical rainfall and less vegetation coverage, and such a mountain torrent needs to be frequently subjected to mountain torrent risk prediction and maintenance; the risk assessment value corresponding to the II-grade hill is [10, 20 ], and the II-grade hill is a hill with moderate historical rainfall and moderate vegetation coverage, and the prediction period of the hill is longer than that of the I-grade hill; the risk assessment value corresponding to the level III is [0, 10 ], the method is a hill with low historical rainfall and much vegetation coverage, the prediction period of the hill is shortest, frequent prediction is not needed, and data processing cost is saved. For example, the prediction period for a grade I hill is 1 hour, the prediction period for a grade II hill is 12 months, and the prediction period for a grade III hill is 24 months.

the method collects characteristic data, rainfall data, annual average 1-12-month rainfall distribution, large-scale latest topographic map of the typical mountain torrent disaster basin, research hydrological zoning conditions, regional station network conditions and hydrological data. The typical mountain torrent disaster basin characteristic data mainly comprises: river basin area, river channel length, river channel gradient, critical river reach section characteristic data and the like; the large-scale latest topographic map of the typical mountain torrent disaster drainage basin is mainly used for measuring and calculating characteristic values of the drainage basin and a river channel above a control section; the regional station network condition mainly comprises the distribution conditions of the existing weather stations (stations), rainfall stations and hydrological stations (including hydrological experiment stations and water level stations) in the region; the rainfall data comprises actual measurement data of rainfall process of each rainfall station in the period of multiple rainfall disasters of the mountain torrents, sorted daily rainfall data, characteristic value (including 10 min, 30 min, 1 h, 3 h, 6 h, 12 h and 24 h maximum rainfall series) data of maximum rainfall in the time period of the calendar year, rainfall data, rainfall duration in and adjacent to the area of the mountain torrents, rainfall intensity, total rainfall in the occurrence process of the mountain torrents and the like, the hourly rainfall data of the rainfall process in the area corresponding to the calendar mountain torrents are collected, the total rainfall in the process is counted, the maximum rainfall in the hourly period (10 min, 30 min, 1 h, 3 h, 6 h, 12 h and 24 h), the latest rainstorm contour map, and a rainstorm statistical parameter contour map (including maximum 10 min, 30 min, 1 h, 3 h, 6 h, 12 h and the like), 24-hour rainstorm contour map and corresponding statistical parameter (mean, skewness coefficient Cv, dispersion coefficient Cs) contour map); the hydrological data mainly comprise actually-measured flood peak water level, flood peak flow and occurrence time corresponding to historical mountain flood disasters, mountain flood disaster occurrence processes, time intervals from rainstorm beginning to disaster occurrence and the like, wherein the water level data are a water level element extract table in the mountain flood disaster occurrence period, the flow data are a flood element extract table in the mountain flood disaster occurrence period, actually-measured flood ratio reduction and the like; hydrological data include short-duration rainstorm time-surface-depth relationship maps or tables, hilly area rainfall-runoff correlation maps, runoff coefficient contour maps, and the like. According to the related information of typical mountain torrent disasters, disaster-causing factors such as rainfall time, rainfall intensity, vegetation and the like of mountain torrent risk evaluation are determined.

the hills have complex environment, so that the disaster factors of the mountain floods are many. And a large amount of disaster-causing factors are detected and processed in a complex way. And some unimportant disaster-causing factors exist, and the influence on the torrential flood is small. Therefore, the invention needs to screen the preliminarily determined disaster-causing factors to obtain the main disaster-causing factors affecting the mountain torrents.

Each disaster-causing factor is a feature of the hill, so that the invention analyzes and investigates the monitoring data acquired from different channels, and extracts the index sensitive to the mountain torrent of the hill through feature engineering. Firstly, the invention selects the important disaster-causing factor from the preliminarily determined disaster-causing factors through the characteristics of the disaster-causing factors. The statistical features include variance, slope and kurtosis, the frequency domain features include frequency, mode shape and modal curvature, and the other features include regression residuals, wavelet energies and fitting coefficients. For example, features with larger variance may be considered useful. If the variance is small, such as less than 1, then this feature may not work as much for the method of torrential flood risk prediction. Most extremely, if the variance of a certain feature is 0, that is, the value of the feature is the same in all samples, it has no effect on the training of the risk prediction model and can be directly discarded. Specifically, the invention sets a threshold value of variance, and rejects the feature when the variance of the feature is smaller than the set threshold value.

And for the screened important disaster-causing factors, further screening by methods such as correlation analysis, conditional entropy, posterior probability, logistic regression weight and the like, selecting the most useful feature subset according to variable forecasting force, and respectively extracting static factors and dynamic factors from the disaster-causing factors. Wherein, static factors include elevation, slope, geology, landform, vegetation, land utilization, unobstructed river channels, population density, economic loss and the like, and dynamic factors include rainfall, rainfall intensity, rainfall duration, water level, soil humidity, human engineering and the like. For example, in the correlation analysis, the invention calculates the correlation coefficient between each disaster-causing factor and the mountain torrents in the more important disaster-causing factors, and the larger the correlation coefficient is, the greater the correlation between the factor and the mountain torrents occurring on the hills is, that is, the greater the influence of the factor on the mountain torrents occurring on the hills is. The method collects the events of the mountain torrents and the corresponding disaster-causing factor statistical data, and calculates the specific mountain torrents and the disaster-causing factor correlation coefficient in the hill damage events.

the method and the system for obtaining the mountain torrent data have different mountain torrent risk values corresponding to different mountain torrent occurrence times. The higher the time interval from the beginning of the lowering to the occurrence of a mountain torrent, the higher the risk of a mountain torrent. Therefore, a corresponding torrential flood risk value is set for each acquired torrential flood data.

And taking the acquired historical torrent data as sample data, and training a logistic regression model, a random forest, a nearest neighbor classification, a K mean classification, a Bayesian method and a deep learning model based on a water body recognition technology. After the sample data is acquired, objects (noise, missing values, outliers and drifts) to be preprocessed by the sample data need to be cleaned, such as a band-pass filtering method, a sigma-yellow transform method, a blind source separation method, a wavelet transform method, a moving average method and the like. Specifically, for each historical torrent data, a value corresponding to the main disaster causing factor is obtained as an input of the model, a corresponding torrent risk value is used as an output of the model, and the prediction model is continuously trained.

Since different training models have different predictive effects on different data. Therefore, the method trains and generates various prediction models, selects the optimal model according to the performance of each model, and further optimizes the prediction effect of the mountain torrent risk.

Taking the logistic regression model as an example, the logistic regression model is also called a generalized linear regression model, and it has substantially the same form as the linear regression model, and all have a₀+a₁X, wherein, a₀And a₁Are parameters to be found which differ in their dependent variables. Multiple linear regression directly combines a₀+a₁X is a dependent variable, i.e. y ═ a₀+a₁X. The logistic regression model is a through a function S₀+a₁X corresponds to a hidden state p, p ═ S (a)₀+a₁X) and then the value of the dependent variable is determined according to p and the size of 1-p. The function S is a Sigmoid function

Changing t to a₀+a₁X, the parametric form of the logistic regression model can be obtained:

as can be seen from the parameter form of the logistic regression, there are two undetermined parameters a in the logistic regression model₀And a₁And in practice the features are usually a plurality,i.e., the parameters and arguments of the logistic regression model are multiple, another representation of logistic regression can be obtained:

wherein the content of the first and second substances,

n is the number of arguments. Therefore, the logistic regression model is trained to determine an appropriate parameter vector a so that a p-value is given as accurately as possible for a new argument vector X.

And (3) forming a characteristic vector by using characteristic values corresponding to main disaster-causing factors in historical torrential.

For the deep learning method, the invention firstly constructs a convolutional neural network, and the convolutional neural network is composed of an input layer, a convolutional layer, a PReLU layer, a pooling layer, a full-link layer and an output layer. The method specifically comprises 5 convolutional layers, each convolutional layer is attached with a nonlinear activation function PReLU layer with parameters, the first convolutional layer, the second convolutional layer and the fourth convolutional layer are connected with pooling layers, each pooling layer adopts a maximum pooling method, the first convolutional layer is connected with an input layer, the input layer inputs values corresponding to main disaster-causing factors in torrent data to be processed, a full-connection layer is located between the last pooling layer and an output layer, each neuron is connected with all neurons of the previous layer, and the feature vectors are mapped to the output layer in a targeted manner according to the requirement of risk detection. And the output layer outputs the mountain torrent risk value. The method comprises the steps of training a convolutional neural network model by using a large amount of historical torrent data, calculating a loss function of the convolutional neural network, iterating and updating the convolutional neural network by using the loss function, continuously training the convolutional neural network based on deep learning risk prediction to reduce the loss function to an expected value, and generating a final convolutional neural network model.

The invention can collect mountain torrent data, water level and other mountain torrent data through remote sensing technology to detect on the premise of water body identification, and the water body identification technology is a technology for extracting water body information by excluding other non-water body information through analyzing based on spectral characteristics and spatial position relation of water body. One idea is to perform pixel-by-pixel classification by comprehensively extracting the characteristics of the shape, spectrum, texture, neighborhood and the like of the ground features in the remote sensing image to realize intelligent segmentation. Another idea is to perform multi-scale segmentation of the remote sensing image and then perform water body identification using a classifier, such as a classification neural network. The invention is not limited to the water body identification technology.

the risk prediction method trains a plurality of risk prediction models, and selects the risk prediction model with the optimal performance so as to improve the efficiency of risk prediction. Specifically, for the logistic regression model, Mean Squared Error (MSE) is generally used to estimate the Error between the predicted data and the acquired data, the more accurate the value is as close to 0, the more accurate the model is, and a determination coefficient (R-Square) is used to estimate the interpretability of the variable to the model, and the closer to 1, the better the fitting ability of the model to the data is. For classification models such as time series clustering and nearest neighbor classification, K-folding cross validation is adopted, the accuracy rate, the recall rate and the accuracy rate are calculated through the obtained confusion matrix, the evaluation estimated value of each model is obtained, and the closer the value is to 1, the higher the accuracy rate is. In addition, the receiving sensitivity curve (ROC) can be used to evaluate the quality of the model, and the optimal threshold is selected by that the larger the Area Under the curve (AUC), the higher the prediction accuracy probability, and generally, the AUC value higher than 0.75 indicates that the model prediction is effective. Based on different characteristics of the models, the method is combined with various evaluation methods, and finally, a plurality of risk prediction models with the best performance are selected to realize combined prediction of the mountain torrent risks so as to ensure the interpretability, effectiveness and accuracy of the models.

Specifically, different weights are respectively given to the selected several risk prediction models with better performance, and the weights represent the influence of the models on the final risk prediction. The higher the weight, the greater its impact on torrential flood risk prediction outcomes. Thus, the risk prediction models are ranked by performance, with higher ranked models being weighted more heavily.

The method carries out data acquisition on the hills needing to carry out the mountain torrent risk prediction, and particularly carries out the acquisition by mutually integrating and applying a high-precision remote sensing monitoring technology and an internet of things sensing technology. A large number of various sensors are distributed on a hill needing prediction, for example, a humidity sensor is arranged to detect soil humidity. The invention also acquires through a high-precision remote sensing monitoring technology, acquires image data of a hill and the like through a scanning imaging type sensor, identifies the water body through water body identification and other technologies, and further monitors the water quantity.

After the spatial data of the hill are acquired, the hill data need to be preprocessed. Cleaning objects (noise, missing values, outliers and drifts) of the acquired spatial data needing to be preprocessed, for example, cleaning methods such as band-pass filtering, Hi-Huang transformation, blind source separation, wavelet transformation, moving average and the like, inputting the preprocessed hill data into a final risk prediction model, and obtaining a predicted torrential flood risk value.

Because the final risk prediction model is formed by combining a plurality of initial risk prediction models, the final risk prediction value is the weighted sum of the risk values predicted by the initial risk prediction models.

In order to more intuitively display the influence of each disaster-causing factor on the mountain torrents risk and the development trend of the mountain torrents risk, the method adopts ArcGIS to map and display. And the collected hill space data are subjected to cleaning treatment and then are superposed with the mountain torrent risk value in a map layer. Specifically, a display scale is selected, transformed by ArcGIS based on geographic coordinates or projection coordinates, clipped and superimposed. Further, according to the mountain torrent risk value, the mountain torrent early warning level is set and is divided into four levels, namely a high level, a medium level and a low level. The higher the mountain torrent risk value is, the higher the probability of indicating that the hill takes place mountain torrents, the larger the damage of mountain torrents, the wider the range of involvement, the higher the early warning level of mountain torrents, therefore, the higher the corresponding early warning level thereof. And according to the mountain torrent early warning level corresponding to the mountain torrent risk, carrying out mountain torrent easiness-to-occur mapping. The red, orange, yellow and blue are used as the early warning colors corresponding to the high, medium and low levels of early warning levels, and a natural breakpoint method, a standard deviation method, an equidistant segmentation method and the like can be adopted for color grading, which is not limited herein.

According to the method, the mountain torrent risk prediction period corresponding to the mountain torrent risk prediction level of the mountain torrent is predicted, and the mountain torrent risk value and the mountain torrent early warning level of the mountain torrent are continuously updated. The risk early warning means adopted by different risk early warning levels are different. For example, when the risk early warning level is extremely high or high, an alarm is given out, and corresponding risk early warning information is timely transmitted to a supervisor in a telephone manner; when the risk early warning level is medium, transmitting corresponding risk early warning information to a supervisor in a short message mode; and when the risk early warning level is low, displaying the risk prediction result on the monitoring equipment only through the ArcGIS image.

Meanwhile, in order to convert passive monitoring and processing of the mountain torrent risk disasters into active prediction and response, when the mountain torrent early warning level of a mountain is predicted to be extremely high and high, the value of a main disaster-causing factor of the mountain is obtained, and the mountain torrent is repaired in a targeted manner, so that the mountain torrent risk of the corresponding mountain is reduced, the service life of the mountain is prolonged, and vegetation coverage is increased, water flow circulation is enhanced and the like. Meanwhile, when the mountain torrent early warning level is extremely high, vehicles, pedestrians and the like are suspended to pass through a hill, and personal safety of people is guaranteed. After the hills are subjected to targeted restoration, the mountain flood risk value and the mountain flood risk early warning level are predicted again, and when the mountain flood risk early warning level of the hills is lowered to high, medium and low, the use of the hills is recovered.

The method can be used for integrally evaluating regional mountain torrent risks. Specifically, the method adopts a sampling evaluation method to evaluate the regional hills. The method comprises the steps of firstly, calculating the proportion of the number of hills corresponding to each hill flood early warning level to the total number of hills in the area, setting the total number of sampled samples, and calculating the number of the sampled samples corresponding to each hill flood early warning level according to the total number of the sampled samples and the proportion of each hill flood early warning level. And randomly extracting a corresponding number of hills from the hills at each flood early warning level. Then, the hills which do not need to be subjected to the flood risk prediction are extracted until the number of the extracted hills reaches the total number of the sampled samples. And calculating the overall torrent risk value of the region according to the sampled hill sample. Specifically, the average value of the flood risk values of the respective hill samples may be used as the flood risk value of the regional hill. For hills that do not need to be subjected to mountain torrent risk prediction, the mountain torrent risk value is 0.

Example two

As shown in fig. 2, the present embodiment provides a regional torrential flood risk prediction system, including:

V＝V₁*ω₁+V₂*ω₂+V₃*ω₃+V₃*ω₄+V₅*ω₅

wherein, V₁、V₂、V₃、V₄、V₅Are respectively asHistorical rainfall, slope, vegetation characteristics, tributary (ditch) water system distribution, and mountain torrents influence range. The value of the historical rainfall is the ratio of the historical rainfall of the hill to the average value of the historical rainfall nationwide; the values of the slopes are respectively endowed with different slopes, and the steeper the slope is, the higher the corresponding slope value is; the vegetation characteristics are values endowed with different vegetation coverage rates, the higher the vegetation coverage rate is, the lower the corresponding vegetation characteristic value is, correspondingly, the more tributary water systems are, the lower the corresponding tributary (ditch) water system distribution value is; the values of the torrential flood influence ranges are values respectively endowed with different influence ranges, and the larger the influence on human, the higher the corresponding torrential flood influence range value.

Taking the logistic regression model as an example, the logistic regression model is also called a generalized linear regression model, and it has substantially the same form as the linear regression model, and all have a₀+a₁X, wherein, a₀And a₁Are parameters to be found which differ in their dependent variables. Multiple linear regression directly combines a₀+a₁X is a dependent variable, i.e. y ═ a₀+a₁X. The logistic regression model is a through a function S₀+a₁X corresponds to a hidden state p, p ═ S (a)₀+a₁X) then rootThe value of the dependent variable is determined according to the size of p and 1-p. The function S is a Sigmoid function

as can be seen from the parameter form of the logistic regression, there are two undetermined parameters a in the logistic regression model₀And a₁In practical applications, there are usually a plurality of features, that is, there are a plurality of parameters and independent variables of the logistic regression model, so that another expression form of logistic regression can be obtained:

wherein the content of the first and second substances,

Therefore, the regional torrential flood risk prediction method and the regional torrential flood risk prediction system, provided by the invention, realize the prediction of regional torrential flood risks, overcome the problem of low processing efficiency of the existing single-point risk prediction, and make up the vacancy of the existing regional torrential flood risk prediction and evaluation; due to the fact that single-point prediction processing of hills one by one is complex, existing mountain flood risk prediction is low in coverage, effective supervision on a large number of hills is not achieved, all hills in the area are supervised simultaneously, and the hill coverage rate of the mountain flood risk prediction is improved; meanwhile, the hills needing risk prediction are screened and predicted according to different risk prediction grades and different risk prediction periods, so that the problems of single existing mountain flood risk prediction method and solidification treatment are solved, and the complexity of mountain flood risk prediction is reduced while safety supervision is carried out on a large number of hills; training to generate a plurality of torrential flood risk prediction models, selecting the prediction models with the best performance to carry out combined risk prediction, and combining the advantages of the plurality of risk prediction models to improve the accuracy of risk prediction; in addition, ArcGIS is adopted to map and display the mountain torrent risk, the mountain torrent easiness map is carried out according to the mountain torrent early warning level corresponding to the mountain torrent risk, red, orange, yellow and blue are respectively used as early warning colors corresponding to the high, medium and low four-level early warning levels, the condition of a hill can be visually displayed, and the development trend of the mountain torrent risk can be detected; carrying out early warning of different levels according to different risks, and timely carrying out risk early warning processing; the sampling evaluation method is adopted to evaluate the regional hills, so that the overall level of the regional hills can be known, the hills with different risks can be treated, the risk of the hills is reduced, and the passive monitoring and treatment of the torrential flood risk disasters are converted into active prediction and response.

It is to be noted that the foregoing is only illustrative of the preferred embodiments of the present invention and the technical principles employed. It will be understood by those skilled in the art that the present invention is not limited to the particular embodiments described herein, but is capable of various obvious changes, rearrangements and substitutions as will now become apparent to those skilled in the art without departing from the scope of the invention. Therefore, although the present invention has been described in greater detail by the above embodiments, the present invention is not limited to the above embodiments, and may include other equivalent embodiments without departing from the spirit of the present invention, and the scope of the present invention is determined by the scope of the appended claims.

Claims

1. A regional torrential flood risk prediction method is characterized by comprising the following steps:

2. The method for regional torrential flood risk prediction according to claim 1, wherein the risk prediction levels divided into different levels according to the basic information are specifically:

setting calendarThe weights of the historical rainfall, the slope, the vegetation characteristics, the distribution condition of the branch (ditch) water system and the mountain torrents influence range are omega in sequence₁、ω₂、ω₃、ω₄、ω₅Wherein, ω is₁+ω₂+ω₃+ω₄+ω₅1 is ═ 1; the mountain torrent risk assessment value that the hillock corresponds is:

V＝V₁*ω₁+V₂*ω₂+V₃*ω₃+V₃*ω₄+V₅*ω₅

3. The method for predicting regional torrential flood risk according to claim 1, wherein the step S5 specifically comprises:

4. The regional torrential flood risk prediction method according to claim 3, wherein a variance threshold is set, and when the variance of the disaster-causing factor is smaller than the variance threshold, the disaster-causing factor is removed; otherwise, the characteristic is taken as a relatively important disaster-causing factor;

5. The method for predicting regional torrential flood risk according to claim 1, wherein the step S6 comprises:

the logistic regression model has the parameter form:

wherein the content of the first and second substances,

6. The method for predicting regional torrential flood risk according to claim 1, wherein the step S6 comprises:

7. The method for predicting regional torrential flood risk according to claim 1,

the performance evaluation method of the torrential flood risk prediction model comprises the steps of mean square error, determination coefficient, accuracy rate, recall rate, accuracy rate, receiving sensitivity curve and area under the curve.

8. The regional mountain torrent risk prediction method according to claim 1, wherein ArcGIS is adopted to map and display mountain torrent risk, and collected mountain hill space data is cleaned and then overlaid with mountain torrent risk values.

9. The method for regional torrential flood risk prediction according to claim 1, wherein the risk prediction method further comprises:

10. A regional torrential flood risk prediction system for implementing the regional torrential flood risk prediction method according to any one of claims 1 to 9, comprising: