JP3674707B1 - Disaster prevention business plan support system and method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a disaster prevention business plan support system and method capable of calculating CL accurately and easily, and capable of being executed in a short time and at low cost while ensuring accuracy of updating with additional data.
An information input device, an information storage device, an information operation device, and an information output device are provided. The information input device includes analysis data in each region, analysis parameters and an analysis model for calculating CL. The information calculation device includes an analysis parameter setting unit that inputs an analysis parameter to an analysis model, and a discriminating boundary surface analysis unit that calculates CL. , Means for outputting at least one piece of information among analysis data, analysis parameter, analysis model, and CL.
[Selection] Figure 1

Description

  The present invention relates to a disaster prevention business plan support system and method for calculating the occurrence limit line, evacuation reference line, and warning reference line of earth and sand disasters, and in particular, it is possible to verify the accuracy of these CLs using required levels. Furthermore, the present invention relates to a disaster prevention business plan support system and method capable of efficiently executing CL update by adopting a support vector machine as an analysis model for calculating CL.

  Sediment disasters (debris flow, landslides, landslides) occur every year across the country, precious lives are lost and valuable assets are destroyed. About 70% of Japan's land is mountainous, with many geologically vulnerable regions, and many geographical features such as steep topography. Due to the social conditions such as population increase to mountain streams, steep slope collapse risk points, landslide risk points), and also meteorological conditions such as typhoons and rainy seasons that are prone to sediment disasters, Sediment disasters are one of the fatal natural disasters in Japan.

  There are about 520,000 dangerous places for landslide disasters nationwide, and the maintenance rate by hardware measures is as low as about 20%. In addition, it is a budget to implement hard measures for all of these dangerous places. Due to the limitations of time and time, the necessity of covering the delay of hardware measures with software measures has been recognized. The purpose of soft measures is to protect human lives from earth and sand disasters and to minimize the destruction of property. Soft measures include warnings, evacuation instructions, emergency responses according to the damage situation, and secondary actions. A function to accurately and promptly support disaster prevention is necessary, and it is required how quickly various disaster prevention information can be collected, organized and transmitted. In particular, accurate warnings and evacuation instructions are important, and these are usually based on warning and evacuation baselines set using short-term and long-term rainfall indicators, In many cases, one reference line is used in the entire region including a plurality of mountain streams, and most of the reference lines are approximated by straight lines. In other words, many conventional baselines do not consider the risk of landslide disasters (latent risk) due to different terrain factors for each slope or mountain stream, and represent complex natural phenomena by linear approximation. For this reason, there remains a problem regarding its accuracy.

Therefore, in recent years, studies have been made to display the reference line in a non-linear manner in order to express complex natural phenomena.
For example, in Non-Patent Document 1, as a “study on the setting of a non-linear collapse occurrence limit rainfall line using an RBF network”, while using a Gaussian function that is a representative function of RBF (radial basis function), an input layer, A technique relating to the setting of an objective and highly accurate non-linear collapse occurrence limit rainfall line by an RBF network having a hierarchical structure including an intermediate layer and an output layer is disclosed.
In this way, the setting of the critical rainfall occurrence line using the RBF network is performed by setting the teacher value of the generated data to −1 and the non-occurrence teacher value to 1 and superimposing the Gaussian function on the discrimination boundary surface. If a discriminant boundary surface related to the collapse phenomenon is constructed using the generated data and the non-occurrence data, the landslide occurrence limit rainfall line can be easily obtained from the contour line.
In addition, since a Gaussian function is used as the output function of the intermediate layer, the output value is 0 in a range where no data exists. Accordingly, the output value is close to 1 in an area where there is a large amount of non-occurrence data, but the output value decreases in an area where there is little non-occurrence data. Therefore, since it is possible to consider a region with little non-occurrence data even if there is no occurrence data as a dangerous region, it is possible to draw a collapse-occurrence limit rainfall line with only non-occurrence data.

However, the discriminant boundary surface for obtaining the landslide occurrence limit rainfall line using this RBF network is formed by overlapping the radial basis functions such as the Gaussian function for the generated data and the non-generated data as described above. As the number increases, the amount of data used for analysis increases, resulting in an increase in analysis time and cost.
In addition to the analysis using such an RBF network, for example, as disclosed in Patent Document 1, using a support vector machine as an analysis model, type discrimination in the vicinity of a boundary where data that cannot be correctly discriminated may exist. A “class identification device and class identification method” with improved accuracy is disclosed.

This Patent Document 1 discloses not only a support vector machine but also an analysis using a neural network at the same time. When two types of teacher data are used to classify them, a support vector machine is used. Analyzing.
Further, Patent Document 2 discloses a “Tens / Aspect / Modality Translation Processing Method, Tens / Aspect / Modality Translation System”, which is a support vector for accurately translating a tense / aspect / modality at the time of machine translation. A technique using a machine is disclosed.
Although the use of the support vector machine in Patent Document 1 and Patent Document 2 is not directly related to the technical field of the present invention, as shown in FIG. 2 of Patent Document 1 and FIG. The machine is used when discriminating data having two types of features or attributes.
Therefore, it can also be used to draw a boundary line for discriminating data relating to two types of events of occurrence and non-occurrence of landslide disasters in the present invention. From such a viewpoint, Patent Document 1 and Patent Document 2 are cited and disclosed.

Kazumasa Kuramoto et al. 5: Study on the setting of non-linear landslide generation limit rain line using RBF network, Proceedings of Japan Society of Civil Engineers, No.672 / VI-50, pp.117-132, 2001.3 JP 2003-208594 A JP 2003-16067 A

However, since the technique disclosed in Non-Patent Document 1 is formed by superimposing a radial basis function such as a Gaussian function on generated data and non-generated data as described above, it is used for analysis as the number of data increases. There is a problem that the amount of data to be increased also increases the analysis time and cost.
Furthermore, when the RBF network is used when analyzing the CL, two types of each of the lattice spacing, the radius of the basis function, the non-occurrence rainfall suppression parameter, and the occurrence rain suppression parameter are set in trial and error. Therefore, there is also a problem that it takes time to study the optimum parameter.

  Further, in the techniques disclosed in Patent Document 1 and Patent Document 2, a line serving as a boundary is drawn in order to discriminate between two types of data having some characteristics, but a means for verifying their accuracy is provided. However, it was not possible to judge how much accuracy the boundary line can guarantee.

  The present invention has been made in response to such a conventional situation, in view of the above-described situation relating to a disaster prevention business plan support system and method for calculating the occurrence limit line of landslide disaster, evacuation reference line and warning reference line, It is possible to calculate at least one of the sediment disaster occurrence limit line, the evacuation reference line, and the warning reference line with high accuracy and easily, and it is possible to execute update with additional data in a short time and at low cost while ensuring accuracy. The purpose is to provide a disaster prevention business plan support system and method.

In order to achieve the above object, a disaster prevention business plan support system according to claim 1 has an information input device, an information storage device, an information calculation device, and an information output device, and a short-term rainfall index in each region. And the sedimentation disaster occurrence limit line, evacuation reference line, or warning reference line (hereinafter, these may be abbreviated as CL in some cases) using the rainfall data including the long-term rainfall index and the actual information on the occurrence / non-occurrence of the sediment disaster. The information input device analyzes the rainfall data and the actual information on the occurrence / non-occurrence of landslide disasters (hereinafter referred to as the rainfall data and the actual information on the occurrence / non-occurrence of landslide disasters). storage and may be referred to as data.), and an analysis parameter and an analysis model for calculating the CL, a possible input means to the information storage device, the information calculating device, in the information storage device And model selection unit is selected and read analytical model, to the selected analysis model, analyze inputs the analysis parameters inputted from the read and input or information input device analysis parameters from the information storage device parameter setting unit And reading the analysis data from the information storage device or using the analysis data input from the information input device , forming the short-term rainfall index and the long-term rainfall index of the rain data included in this analysis data as the vertical axis or the horizontal axis, respectively. On the grid points of the grid on the two-dimensional plane, plot the actual information on the occurrence or non-occurrence of sediment disasters to form a discrimination boundary surface, and calculate CL as a contour line corresponding to the desired threshold from this discrimination boundary surface to provided a discrimination boundary analysis unit, the information output device, of the analytical model and CL and analysis data and the analysis parameters of at least 1 It is an output means possible information.

Also, disaster prevention business plan support system as defined in claim 2 is the invention of claim 1, wherein, the information input device, for the CL that will be computed, generated hit against rainfall data when landslides occurred The input level from the information input device can be set to the required level consisting of the non-occurrence medium rate for the rainfall data when there is no rate and / or sediment disaster , or stored in advance in the information storage device The information calculation device is provided with a request level setting unit that can be set by reading out the required level data and reading out the rainfall data and the record information of occurrence / non-occurrence of sediment disaster from the information storage device or input from the information input device. Using the rainfall data and actual information on occurrence / non-occurrence of landslide disaster, for the CL calculated by the discriminant interface analysis unit, Or calculating a non-generation predictive value for rainfall data when landslides did not occur, as compared to the requirement level set by demanding setting unit as it has a CL accuracy verification unit for verifying the accuracy of the CL is there.

The disaster prevention business plan support system according to claim 3 is the invention according to claim 1 or claim 2, wherein the analysis data includes factors other than rainfall data that cause sediment disasters at points included in each area. The analysis parameter setting unit reads the attribute value or evaluation value from the information storage device as the risk of occurrence of a landslide disaster , and the discriminating boundary surface analysis unit includes the risk value classifying said point in response to the magnitude, and thereby calculates the CL per this classification.

The disaster prevention business plan support system according to claim 4 is the invention according to claim 1 or claim 2, wherein the analysis data includes factors other than rainfall data that cause sediment disasters at points included in each area. Including an attribute value or an evaluation value related to at least one factor, and the analysis parameter setting unit reads analysis data including the attribute value or the evaluation value from the information storage device, or uses the analysis data input from the information input device, Calculate the sediment disaster occurrence rate from the actual information on the occurrence or non-occurrence of sediment disasters in each category section of the factor, the cumulative value of all the factors of the sediment disaster occurrence rate as the risk of the point, the discriminant boundary surface analysis unit, Points are classified according to the degree of risk, and CL is calculated for each classification.

The disaster prevention business plan support system described in claim 5 is the invention described in any one of claims 1 to 4, wherein the analysis model is a support vector machine (hereinafter abbreviated as SVM). The analysis parameter setting unit reads and inputs the analysis parameter from the information storage device to the support vector machine, or inputs the analysis parameter input from the information input device, and the discriminant boundary surface analysis unit stores the information Two-dimensional formed by reading the analysis data from the device or using the analysis data input from the information input device, with the short-term rainfall index and the long-term rainfall index of the rain data included in this analysis data as the vertical axis or horizontal axis, respectively Plot information on the occurrence or non-occurrence of landslide disasters on the grid points of the grid on the plane, and use the support vector machine to And calculating the line vector and bias values to form a judgment boundary surface, it calculates the CL as contour corresponding to the desired threshold from the determination interface, among the analytical data used for calculation of the CL, the support vector machine If the support vector to be classified is extracted and stored in the information storage device, and the analysis data accumulates additional data thereafter, the CL can be updated using the support vector and the additional data. is there.

  The disaster prevention business plan support system according to claim 6 is the invention described in claim 5, wherein the analysis vector that has not been classified into support vectors by the support vector machine is separated hyperplane f (x ) = ± 1 includes the analysis data included in the predetermined distance ε in the support vector and the additional data, and calculates CL.

  The disaster prevention business plan support system according to claim 7 is the invention according to claim 5 or claim 6, wherein the support vector machine of the analysis model is a support vector having a soft margin expressed by equation (1). What is a machine.

  In the disaster prevention business plan support system described in claim 8, in the invention described in claim 5 or claim 6, the support vector machine of the analysis model is a ν support vector machine expressed by equation (2). There is something.

The disaster prevention business plan support method according to claim 9 is a computer-implemented method for performing disaster prevention data , including rainfall data including short-term rainfall indicators and long-term rainfall indicators , actual information on occurrence / non-occurrence of landslide disasters, and landslide disasters. a disaster prevention business planning support method for calculating the CL using the evaluation value for each factor causing the model selection unit of the computer, the steps of be read out by selecting an analysis model stored in advance in the information storage device, analyzing the parameter setting unit of the computer comprises the steps of inputting an analysis parameter inputted from the information storage device after reading the analysis parameters inputted in the analysis model or an information input apparatus, the discrimination boundary analysis unit of the computer, the information storage device analysis data using analysis data inputted from the reading out or information input device from rainfall data contained in the analysis data Plotting the actual information on the occurrence or non-occurrence of landslide disasters on the grid points of the grid on the two-dimensional plane formed with the short-term rainfall index and the long-term rainfall index as vertical axis or horizontal axis respectively. a step of calculating the CL as contour lines corresponding to the discrimination boundary from the desired surface threshold, the computer of the information output device, comprising the step of outputting at least information of the analytical model and CL and analysis data and the analysis parameters Is.

The disaster prevention business plan support method according to claim 10 is a computer-implemented method for performing disaster prevention data , including rainfall data including short-term rainfall indicators and long-term rainfall indicators , actual information on occurrence / non-occurrence of landslide disasters, and landslide disasters. a disaster prevention business plan support method calculates the CL using the evaluation value of each factor verifying the CL in demanding that causes, required level setting unit of the computer, with respect to calculated the CL, the landslides Reads the required level data stored in the information storage device in advance to the required level consisting of the occurrence probability for the rain data when it occurs and / or the non-occurrence probability for the rain data when no sediment disaster has occurred. or a step of adopting to set an input value from the information input device, the model selection unit of the computer, select the analysis model stored in advance in the information storage device And to step out seen, the analysis parameter setting part of the computer, a step of inputting an analysis parameters inputted from the analysis parameters or information input apparatus read out from the information storage device in the analysis model, determine the boundary surface analysis unit of the computer However , by reading the analysis data from the information storage device or using the analysis data input from the information input device , the short-term rainfall index and the long-term rainfall index of the rain data included in this analysis data are formed as the vertical axis or the horizontal axis, respectively. On the grid points of the grid on the two-dimensional plane, plot the actual information on the occurrence or non-occurrence of sediment disasters to form a discrimination boundary surface, and calculate CL as a contour line corresponding to the desired threshold from this discrimination boundary surface reading a step of, CL accuracy verification unit of the computer, the record information of rainfall data and landslides generation and non-generation from the information storage device Or, using the rainfall data input from the information input device and the record information of occurrence / non-occurrence of landslide disaster, the occurrence of landslide data when landslide disaster occurs for CL calculated by the discriminant boundary surface analysis unit Calculating the non-occurrence probability for the rainfall data when there is no rate and / or sediment disaster, and verifying the accuracy of the CL by comparing with the required level set by the required level setting unit, and outputting information The apparatus includes a step of outputting at least one piece of information among analysis data, analysis parameter, analysis model, and CL.

In the present invention, CL is set by using actually acquired rainfall data and actual data of sediment disaster, so that CL can be set objectively without requiring user's experience and intuition. A high CL can be created. Therefore, it is possible to update the database for sediment disasters in a short period of time, and to cope with disaster countermeasures that require quickness, and it is possible to exert great benefits for national land conservation and lifesaving.
In particular, the disaster prevention business plan support system described in claim 2 calculates the probability of occurrence / non-occurrence of rainfall data using the rainfall data and the actual information on occurrence / non-occurrence of landslide disasters. Therefore, it is possible to provide a disaster prevention business plan support system that can guarantee high accuracy at all times.
Further, particularly in claims 5 to 8, by introducing a support vector machine into the analysis model, it is possible to easily calculate in addition to high accuracy, and also when updating with additional data, in a short time and at low cost. It is feasible and enables highly efficient analysis in terms of time and cost.

Before describing the best mode and embodiments of the present invention, in order to facilitate understanding of the invention, embodiments and examples described in the claims and specification of the present application, the present specification and patents are described. The definition of the word used in a claim is shown. Some of them have been described in the prior art, but are described collectively here.
First, “disaster prevention project” as used in this application is not limited to the construction of facilities for disaster prevention directly, but prioritization of construction, such as survey projects conducted to investigate the dangers of slopes or mountain streams. This includes all disaster prevention projects that require
In addition, the “factor” related to the risk of occurrence of sediment disaster in this application refers to factors related to topography and geology such as topsoil thickness, ground condition, relationship between slopes and discontinuities, presence of faults, rock classification, etc. Including environmental factors such as vegetation, age of trees, presence or absence of logging, presence or absence of spring water, presence or absence of collapse history, rainfall related factors such as hourly rainfall, effective rainfall, continuous rainfall, etc. It can be used as an attribute for executing the analysis of the degree of risk provided to. For example, when analyzing the occurrence / non-occurrence of landslide disasters under the prescribed rainfall factor conditions and adding other factors other than the rain factor as attributes, CL is analyzed after evaluating the potential risk. Can be set.

  “Category” as used in the present application refers to a classification based on a physical quantity or a non-physical quantity for evaluating each factor, and “category section” refers to the divided range. For example, if it is a topographic factor called river basin average slope, 0 ° to 10 ° is a category 1, 10 ° to 20 ° is a category 2, etc., and a 0 ° to 10 ° section with respect to a physical quantity of “°” This is called a category section. In addition, in the topographic factor of mountain stream direction, the east is category 1 and the west is category 2 with respect to the non-physical quantity “east, west, south, and north”. In such a case, the category section is one direction.

The “attribute value” in the present application refers to an ordinal number indicating a plurality of category section names or category sections in each factor, that is, 1 or 2 such as Category 1 and Category 2. Further, the “evaluation value” means a measured value of the average gradient in an actual region or point if it is a topographic factor called the river-basin average gradient of the river described above, and specifically, a numerical value such as 15 °.
In addition, the “probability of occurrence” of sediment-related disasters means the ratio of the number of precipitations that exceeded CL before the disaster occurrence time to the total occurrence of rainfall, and the “non-occurrence target” is the total non-occurrence It means the ratio of the number of non-occurring rainfalls in which all rains did not exceed CL during a series of rains.

Hereinafter, a disaster prevention business plan support system and a disaster prevention business plan support method according to an embodiment of the present invention will be described with reference to FIGS.
FIG. 1 is a configuration diagram of a disaster prevention business plan support system according to the present embodiment. 2 to 5 show flowcharts of the disaster prevention project plan support method. In this embodiment, a support vector machine with a hard margin that completely separates a set of occurrence data of sediment-related disasters and a non-occurrence set for analyzing the discriminant boundary surface and setting CL based on the analysis result, a certain amount of error An explanation will be given using a support vector machine with a soft margin that allows the above, and a one-class support vector machine (hereinafter abbreviated as νSVM) that uses only non-occurrence data of earth and sand disasters.

  The disaster prevention business plan support system includes an information input device 1, an information calculation device 2, an information output device 3, and an information storage device 4. The information input device 1 is used to input data stored in the information storage device 4 in advance, or to directly input data when the information processing device 2 is operated. The information processing device 2 includes a request level setting unit 5, a model selection unit 6, an analysis parameter setting unit 7, a discriminant boundary surface analysis unit 8, and a CL accuracy verification unit 9, and is read from the information storage device 4. Using the data and data input from the information input device 1 and the analysis model, CL is set based on the analysis of the discriminant boundary surface and the analysis result of the discriminant boundary surface, and further analysis for updating the CL is performed. Is what you do.

Before describing the elements constituting the information processing device 2, data stored in the information storage device 4 will be described first.
In the information storage device 4, the actual data 11, the request level data 15, the analysis parameter 16, the model data 17, the CL data 18, and the support vector data 19 are stored. These data are input from the information input device 1 described above. A part of the analysis parameter 16 is obtained by calculation in the information calculation device 2.
The actual data 11 includes rainfall data 12 at a certain area or each point included in the area, and disaster record data indicating whether or not a landslide disaster has occurred or not occurred in those areas or points. 13. It includes factor evaluation value data 14 relating to an evaluation value relating to a factor causing a sediment disaster or an attribute value. The rain data 12 may be actual measurement values, data obtained by standardizing the actual measurement values, or a rain index such as effective rainfall data with a half-life of 21 days. These rainfall data 12, disaster record data 13, and factor evaluation value data 14 are stored as a data table with ID codes relating to areas and points to be analyzed, and these ID codes are used as keys. Search is possible.
The required level data 15 is a required level for the accuracy of the CL obtained from the result of the discriminant boundary surface analyzed using the model, and is expressed by a hit probability of occurrence and a hit probability of non-occurrence.
Although the analysis parameter 16 varies depending on the analysis model used, the radius r of the kernel function common to all the support vector machines, the variable w representing the normal vector of the separation hyperplane, and the variable b called the bias term. , Ρ, a weight parameter C for slack variable ξ _i , and ν that is a parameter representing a ratio of data misclassified in one class of support vector machine. It refers to a set of numerical values including these variables, parameters or coefficients.

First, the model data 17 has a distinction between a linear model and a non-linear model. For these distinctions, a support vector machine model with a hard margin, a support vector machine model with a soft margin, and a ν support vector machine, respectively. Model is included. In this embodiment, although the case where a support vector machine is selected as a model is used, other analysis models may be used, and in this case, data on the analysis model to be employed is included.
The CL data 18 is data relating to CL obtained based on the discrimination boundary surface obtained by analysis. The CL data 18 may be rewritten by updating the analysis, or the old data obtained in the past may be stored while being distinguished from the updated data. In addition, data regarding the discrimination boundary surface for setting CL may be included.
The support vector data 19 indicates a set of model data 17 and analysis parameters 16, and specifically, for example, in a support vector machine with a soft margin, a radius r of a specific kernel function, a specific parameter C, etc. from the analysis parameter 16 Means the data related to each support vector to which is assigned. Further, as will be described later in the description of the embodiment, it includes data relating to information such as information relating to the importance of the support vector and whether it can be said to be an abnormal value.

Next, the information arithmetic device 2 will be described.
The information processing device 2 is also connected to the information input device 1 and the information output device 3, and further to the information storage device 4, and setting, analysis, verification, etc. using information input via the information input device 1 It is also possible to read out the data stored in advance via the information input device 1 from the information storage device 4 and use it.
Data used for setting, analysis or verification executed in the information processing device 2 and the results of those operations are output or displayed via the information output device 3.
The required level setting unit 5 of the information processing device 2 sets the required level prior to calculations such as setting and updating of CL. The requirement level data 15 for setting the requirement level is expressed by the hit probability of occurrence of landslide disaster and the hit probability of non-occurrence as described above. Is to do. The request level setting unit 5 can be set by adopting an input value from the information input device 1, but may read any of the request level data 15 stored in the information storage device 4 in advance. . At the time of reading, it is preferable that a plurality of pre-stored requirement level data 15 can be selected from among the requirement level data 15 that are candidates and displayed on the information output device 3.

The model selection unit 6 first selects whether to update CL or to newly set CL. This selection may be input from the information input device 1 when analysis is started. In response to the input of information from the information input device 1, the model selection unit 6 determines whether it is a CL update or a new one.
In addition, a linear / non-linear selection of whether to use a linear model or a non-linear model for the CL analysis, and in this embodiment, two types of selection of which type of support vector machine to select are selected. Is included. However, linear or non-linear selection is not essential, and either one type of analysis model may always be selected.
These analysis models may be selected by reading the model data 17 stored in the information storage device 4 in advance, or may be input from the information input device 1 when analyzing.
The analysis parameter setting unit 7 sets an analysis parameter used for the analysis model selected by the model selection unit 6. The analysis parameter is read from the analysis parameter 16 stored in the information storage device 4. The information is displayed on the information output device 3 and selected from them. Of course, the information may be input from the information input device 1. Since the analysis parameters used differ depending on the analysis model, the analysis parameter setting unit 7 displays the analysis parameters on the information output device 3 according to the determination result while determining the analysis model selected by the model selection unit 6. It is desirable to be selected.
Specifically, an ID code that can recognize the analysis model is attached to the analysis model, and the analysis parameter is stored in the information storage device 4 with an ID code that is common to the analysis model. It is good to keep. The model selection unit 6 determines the ID code of the analysis model selected by the model selection unit 6 and transmits information on the ID code to the analysis parameter setting unit 7. The analysis parameter setting unit 7 recognizes the ID code. However, the ID code is searched from the analysis parameters 16 stored in the information storage device 4, and the matching analysis parameters 16 are read out and displayed on the information output device 3.

Further, the analysis parameter setting unit 7 reads out the above-described factor evaluation value data 14 and disaster record data 13 for each factor in the region to be analyzed or each point in the region, and the category section for each factor. By calculating the sediment disaster occurrence rate from the disaster record data 13 and using the sediment disaster occurrence rate as the sediment disaster occurrence score for that section, the risk level for each region or multiple points is calculated. It can be reflected in analysis and settings. That is, the landslide disaster occurrence score functions as a risk against the landslide disaster. In calculating the sediment disaster rate, the attribute values in the category section may be arranged, or the evaluation values for each factor in the category section may be arranged.
Furthermore, if the risk level considering a plurality of factors is obtained by combining and accumulating the risk levels calculated for each factor, the accuracy with respect to the risk level can be improved.
The analysis of the risk level is selectively performed. Whether the CL is set after the risk level is calculated in advance or whether the CL is set only by the rainfall data 12 without calculating the risk level. Should be selectable. This selection is controlled so that a message is displayed on the information output device 3 in the analysis parameter setting unit 7 or in advance in the request level setting unit 5 and the model selection unit 6, and information related to the selection is input from the information input device 1. You should encourage them to do so. Also, once the risk level analysis is performed, the risk level analysis is not required again when the CL is updated. Therefore, when the update analysis is performed, the analysis algorithm must be updated. It is recommended that a flag for recognizing be set to avoid selection of calculation regarding the degree of risk. However, when the factor evaluation value data 14 is added or changed, it is desirable to reset this flag so that calculation regarding the degree of risk can be selected. Further, the calculated degree of risk may be displayed by the information output device 3.
The analysis of the risk level in the analysis parameter setting unit 7 will be described later in detail with reference to FIG.

  Next, in the discriminant boundary surface analysis unit 8, first, data to be analyzed is selected from the actual data 11 stored in the information storage device 4. The selected actual data 11 is input to the previously selected analysis model, the discrimination boundary surface is analyzed, and CL is set from the discrimination boundary surface. The set CL is stored in the information storage device 4 as CL data 18. In the discriminating boundary surface analysis unit 8, the calculation differs depending on whether or not the risk analyzed in the analysis parameter setting unit 7 is reflected in the CL. Details thereof will be described later with reference to FIG.

Finally, the CL accuracy verification unit 9 will be described. The CL accuracy verification unit 9 reads the CL analyzed by the discriminant boundary surface analysis unit 8 from the information storage device 4 and verifies the prediction accuracy. Prediction accuracy is calculated by calculating the probability of occurrence for rainfall data when a sediment disaster occurs and / or the probability of non-occurrence for rainfall data when a sediment disaster does not occur. This is done by reading 15 and comparing it. The verification result of the CL prediction accuracy is performed by displaying the comparison result with the required level data 15 on the information output device 3. If the required level is satisfied, the verification is completed. If not satisfied, the model selection unit 6 selects the analysis model again, and the analysis parameter setting unit 7 also selects the analysis parameter. become.
If the prediction accuracy of the CL still does not satisfy the required level by selecting these analysis models and analysis parameters, the required level may be set again by the required level setting unit 5. In such a case, it is preferable to display the required level data 15 on the information output device 3 in advance so that the required level data 15 lower than the initial required level data 15 can be selected.

Hereinafter, the present invention will be described specifically by way of examples. First, as Example 1, a process of accurately and easily setting at least one of an occurrence limit line of an earth and sand disaster, an evacuation reference line, and a warning reference line for each region or slope, and each mountain stream, and these reference lines A disaster prevention project plan support method for efficiently updating
2 to 4 show that it is possible to accurately and easily set at least one of the non-linear occurrence limit line of landslide disaster, evacuation reference line, and warning reference line for each region or slope, and each mountain stream. It is the general | schematic flowchart which showed the preferable process and procedure of the disaster prevention business plan support method for updating the reference line efficiently. In addition, these processes and procedures introduced in the present embodiment are common to the disaster prevention business plan support system described above. In these drawings, in order to clearly show the relationship with FIG. 1, that is, specific cooperation with hardware, the configuration of hardware is also shown. In this connection, the embodiment referring to FIGS. 2 to 4 will be described with reference to the reference numerals of the components shown in FIG.

  First, as shown in FIG. 2, in step S1, a required level for analyzing CL is set. As described above, there are two required levels: the target ratio for occurrence of sediment-related disasters and the target ratio for non-occurrence rainfall. For this setting, it may be stored in the information storage device 4 in advance as the request level data 15 or may be input from the information input device 1.

Next, at the same time before or after step S1, evaluation values and attribute values for various factors related to the risk of occurrence of landslide disasters at various terrain and locations, and the time series of each station obtained from rainfall stations In the information storage device, the attribute values of the rainfall data and the historical information on occurrence / non-occurrence of landslide disasters such as the date and time of landslide disaster occurrence are readable. In addition, an ID code or the like is attached to a point where each evaluation value or attribute value is obtained so that the point can be identified.
For multiple locations to be analyzed, collect the data represented by the combination of the evaluation value for each factor, the actual information on occurrence / non-occurrence of landslide disasters, and the name of the rainfall observatory from which rainfall data is extracted. For each factor related to the risk of occurrence of landslide disasters in Japan, each category segment is divided into a plurality of category sections, and the category section names are assigned corresponding to the evaluation values for each of the plurality of factors in the data. Category category name of attribute, attribute value, actual information on occurrence / non-occurrence of landslide disaster, list of rainfall station names for extracting rainfall data, time-series rainfall data attribute value for each rainfall station, sediment It is better to organize a list that shows a combination of actual and non-occurrence of disasters. In addition, the list may be arranged as the evaluation value without using the attribute value.
In addition, after selecting the factor in advance, the evaluation value, the attribute value, the record information of occurrence / non-occurrence of landslide disaster, the ID code, and the like may be stored in the information storage device.
The rain data in FIG. 2 corresponds to the rain data 12 in FIG. 1, actual information on occurrence / non-occurrence of landslide disasters corresponds to the disaster actual data 13, and attribute values and evaluation values correspond to the factor evaluation value data 14. It can be said that the list shows the data structure of the actual data 11 that collectively shows the rainfall data 12, the disaster record data 13, and the factor evaluation value data 14.

  The rainfall data 12 may be converted into a rainfall index such as an effective rainfall if necessary, and arranged in time series for each rainfall station. Moreover, after converting into effective rainfall etc. beforehand, you may store in the information storage apparatus 4 in time series. In order to shorten the analysis time, the rain data 12 without landslide disaster may be converted into a representative value by clustering processing, or converted into a representative value and stored in the information storage device 4 in advance. Also good.

  Next, in step S2, whether or not to update the current CL or set a new CL is selected. This selection may be performed by inputting from the information input device 1 or, if possible, storing information on the selection in the information storage device 4 in advance and reading it out. Predicting the current CL when factors related to the risk of occurrence or non-occurrence of earth and sand disasters such as ground strength have changed significantly, when a large-scale disaster has occurred compared to past disaster results, or against current rainfall When accuracy is low, CL needs to be updated.

Next, CL is set and updated in steps S3 to S6. First, in step S3, an analysis model for setting CL is selected. As an analysis model of CL, firstly, there is a distinction between linear CL and nonlinear CL, and there are selection of several analysis models for setting each. For example, there is a support vector machine with a hard margin for complete separation, a support vector machine with a soft margin that allows a certain amount of error, and a one-class support vector machine that performs CL setting only with precipitation that does not cause sediment disasters. Make these choices. Data relating to these analysis models is stored in advance in the information storage device 4 as model data 17.
When the model to be used for the analysis is determined, the analysis parameter 16 necessary for each model is input in step S4. As described above, it is necessary to determine the radius r of the kernel function for the analysis parameter 16 in common with all models. Further, in a support vector machine with a soft margin, it is necessary to determine the parameter C related to the allowance of misclassification, and in a one-class support vector machine, it is necessary to designate ν as an upper limit value of a ratio regarded as an outlier. These analysis parameters may be input from the information input device 1. It may be stored in advance in the information storage device 4 and read from there.

  In step S5, it is selected whether to reflect the risk of landslide disasters due to differences in multiple topography, geology, vegetation factors, etc., in the CL set as a rain factor, or whether to set a CL only as a rain factor. For selection, information related to selection stored before analysis may be read from the information storage device 4 or displayed on the information output device 3 so as to be input from the information input device 1 and selected by the operator. You may make it prompt.

  Step S6 will be described in detail with reference to FIG. First, in step S61, factor evaluation value data 14 related to attribute values and evaluation values such as terrain, geology, and vegetation factors arranged for each of a plurality of points in the analysis target area, and disaster result data 13 related to sediment disaster information. Is read from the information storage device 4. Next, the landslide disaster occurrence rate is calculated from the disaster record data 13 in each category section for each read factor, and the landslide disaster occurrence rate is set as the landslide disaster occurrence score of the section (step S62). Topographical and geological factors do not change significantly in several years, and it may not be necessary to recalculate the risk level for risk points that have already been calculated when the CL is updated. Therefore, in step 63, when the risk level for each of a plurality of points is calculated in advance, the past risk level is read from the information storage device and used as it is. In step S64, the attribute value or evaluation value of each factor is associated with the landslide disaster occurrence score calculated above for each of a plurality of points, and the cumulative value of all the factors of the landslide disaster occurrence score is calculated as the risk for each point. The calculated risk level is output and displayed by the information output device 3.

Returning to FIG. In step S7 in FIG. 2, the current CL is updated or a new CL is set. Step S7 will be described in detail with reference to FIG.
First, in step S71, analysis data, that is, actual data 11 is selected. When setting up a new CL, the sediment-related disaster occurrence rainfall (hereinafter referred to as “occurrence rainfall”) and the sediment-free disaster occurrence rainfall (hereinafter referred to as “non-occurrence rainfall”) based on the past sediment disaster information. select. When updating the CL, select the update data extracted when setting the current CL and the newly generated or non-occurring rainfall. Accordingly, information indicating whether the CL is updated or newly established is stored in the information storage device 4 in advance, and the selection is read out or the input from the information input device 1 is prompted. It is necessary to determine which period of actual data 11 is to be read. It is preferable to add an ID code to information on whether it is updated or newly established, read the ID code, and select the actual data 11 in a period corresponding to the ID code.
In addition, since the setting method differs depending on whether or not the risk level is reflected in CL, when the risk level is reflected in CL, the processing from step S72 to step S77 is performed and the risk level is not reflected. In this case, the processing from step S78 to step S80 is performed. Each CL setting method is shown below
When the risk level is reflected in the CL, points such as a risk location to be set in step S72 are divided into several groups according to the level of the risk level. Next, in step S73, occurrence / non-occurrence rainfall is extracted for each group. In step S74, w representing the normal vector of the separation hyperplane and the value of the bias term b are calculated for each group by the support vector machine according to the CL model selected in step S3, and the discriminant boundary surface is constructed using them. To do. The optimum threshold value is examined from each constructed discrimination boundary surface, and these are set as the basic CL. In step S75, the normal vector of the separation hyperplane used for setting the basic CL and the value of the bias term are stored in the information storage device 4 in a readable manner as the analysis parameter 16. Further, support vectors (important data) with high importance for the basic CL, unnecessary data with low importance, abnormal values, and update data are extracted and stored as support vector data 19 in the information storage device 4 so as to be readable.
Data related to CL settings, that is, data around the CL is called highly important data. Sporadic occurrence data in an area where non-occurrence data is concentrated is called an abnormal value. Other data is regarded as less important data. Furthermore, update data necessary for CL update is extracted from these data.
These levels of importance, update data, and abnormal values will be described later.
In step S76, a grid with constant intervals is set on the vertical axis of the short-term rainfall index and the horizontal axis of the long-term rainfall index within the range where the rainfall data exists, and the representative risk ( Set an approximate expression of the average value) and the weight of each grid. Finally, in step S77, the degree of danger at each point that is the target of CL setting is input to the above-described approximate expression, and CL is constructed for each point. The constructed CL is output and displayed by the information output device 4 and stored in the information storage device 4 so as to be readable.

When the degree of risk is not reflected in CL, occurrence / non-occurrence rainfall in the target area is extracted in step S78. In step S79, the support vector machine calculates w and the value of the bias term representing the normal vector of the separation hyperplane according to the CL model selected in step S3, and constructs a discrimination boundary surface using them. The optimum threshold value is examined from the constructed discriminant boundary surface, and is defined as CL. The normal vector of the separation hyperplane and the value of the bias term used for the construction of CL are stored in the information storage device 4 as the analysis parameter 16 so as to be readable.
The constructed CL is output / displayed by the information output device or the information display device and is stored in the information storage device so as to be readable. In step S80, support vectors (important data) having high importance for CL settings, unnecessary data having low importance, abnormal values, and update data are extracted and stored in an information storage device so as to be readable.

Returning to FIG. In step S8 of FIG. 2, the prediction accuracy for occurrence / non-occurrence rainfall is calculated for the set CL. Prediction accuracy is evaluated by the accuracy of occurrence / non-occurrence. Step S8 will be described in detail with reference to FIG.
First, in step S81, CL for accuracy verification is read from the information storage device 4. Next, in step S82, the generation / non-occurrence rainfall (learning data) used for the CL setting from the actual data 11 of the information storage device 4 or the occurrence / non-occurrence precipitation (unlearned data) not used for the CL setting is set. Extract a series of rainfall. A series of these occurrence / non-occurrence rainfalls may be input from the information input device 1 or may be stored in the information storage device 4 in advance and read from there. In step S83, the probability of occurrence / non-occurrence is calculated from the relationship between the extracted series of occurrence / non-occurrence rainfall and CL. The probability of occurrence means the ratio of the number of occurrences of rainfall that exceeded CL before the disaster occurrence time to the total occurrence of rainfall, and the probability of occurrence of occurrence is CL for all rainfalls in a series of precipitations for all non-occurrence rainfall. Refers to the percentage of the number of non-occurring rainfall that did not exceed. The prediction accuracy is output and displayed by the information output device.

  Returning to FIG. In step S9, it is verified whether or not the prediction accuracy calculated in step S8 satisfies the required level set in step S1. If the required level is satisfied, the process moves to step S10. If the required level is not satisfied, it is necessary to reconstruct the CL by changing the analysis model or analysis parameters of the CL setting, or to reconstruct the CL using all the occurrence / non-occurrence rainfall up to the present time. Move to S2.

  In step S10, using the set CL, the timing of the warning recommendation and the evacuation instruction is determined in real time against the rainfall that changes every moment. The results are also displayed on the GIS (Geographic Information System) in a real-time display with a color-coded rectangle according to the degree of danger, along with the topographic map, so that the operator can visually check the time and place where there is a high risk of landslide disaster. Can be grasped.

  Next, the second embodiment will be described with reference to the above-described first embodiment and showing more specific examples and calculation examples.

  Countermeasure work maintenance (hardware countermeasures) for landslide disaster risk points (hereinafter referred to as “dangerous points”) is a direct landslide disaster prevention measure, but there are approximately 520,000 landslide disaster risk points nationwide. However, the maintenance rate is as low as about 20%, and it is difficult to respond quickly because it is budgetary, time-consuming, and limited to implement hard countermeasures for all of these dangerous locations. It is in. Therefore, it has become an important issue to implement software measures such as vigilance and evacuation in order to supplement hardware measures, and to protect human life with the highest priority.

  The MLIT has proposed linear CL based on past disaster records as a standard for measuring the timing of evacuation and evacuation for sediment-related disasters. However, the above CL has the following problems in terms of setting method and prediction accuracy. The current state of linear CL often relies on the subjectivity of engineers to select the precipitation that will cause landslide disasters and to set the slope of the CL. In recent years, public administration has been required to disclose information and an objective method is desired. Therefore, the current setting method is not practical. In addition, since the shape of the CL is linear, it cannot respond to sediment-related disasters that are complex natural phenomena, and problems such as low prediction accuracy have been pointed out. Therefore, it is necessary to establish a nonlinear CL setting method based on an objective method and to predict sediment disasters with high accuracy.

In addition to the above problems, there are the following problems in CL update. The CL is set in consideration of the distribution status of landslide disaster occurrence and non-occurrence rainfall based on past disaster history. CL is therefore dependent on the density and accuracy of data on rainfall data and disaster conditions. In particular, since there are fewer occurrences of rainfall than non-occurrence rainfall, in the event of a disaster that exceeds past disaster results, it is a more dangerous rainfall area (hereinafter referred to as “dangerous area”) or safer rainfall area than before. (Hereinafter referred to as “safe area”) may change. In addition, when the disaster occurrence conditions such as the ground strength or the way of raining change, the above-mentioned region may also change when the countermeasure work is advanced. For these reasons, it is important to review the CL regularly or in response to changing circumstances.
However, since data is accumulated year by year along with non-occurrence rainfall, it is not efficient in terms of labor and time when all these data are used for CL update. In addition, past data may not always be useful due to changes in ground conditions and the effects of countermeasures. For example, in Yamaguchi Prefecture, where there are many landslide disasters in the Chugoku region, several tens of landslide disasters occur each year. In addition, about 10,000 non-occurrence rainfalls, which are indispensable for setting CL, are accumulated every year. For the occurrence and non-occurrence of rainfall that accumulates year by year, there is no way to select important or unnecessary data for CL setting with conventional technology, data accumulated from the past and newly obtained data CL must be set for a large amount of occurrence / non-occurrence rainfall including In recent years, information on rainfall for 10 minutes has begun to be provided in order to grasp the situation of sediment-related disaster occurrence or non-occurrence rainfall in more detail than the conventional one-hour rainfall. If 10-minute rainfall is used, the number of data will be 6 times that of 1-hour rainfall, and it will be difficult to set and update CL with the prior art using the RBF network.
A technology using an RBF network (hereinafter referred to as “RBFN”) has been published as a prior art for setting non-linear CL. In particular, this technology requires a lot of non-occurring rainfall for setting non-linear CL. It is expected that the above problems will occur in the review. Therefore, it is necessary to select the past occurrence / non-occurrence rainfall according to the importance to the sediment disaster and establish a method to update the CL except for the data with low importance.
In the present embodiment, a support vector machine that is currently attracting attention as a pattern classification method is used to try to solve the above-described CL setting method, accuracy problems, and CL update problems.

In the prior art, generated and non-occurring rainfall are indispensable for setting CL. In areas where there are few disasters due to earth and sand disasters, depending on the distance between the available observation stations and the locations where they occur, there may be no generation rain that can be used for CL settings. In addition, there are many data in the past disaster history where the occurrence location is unclear, and it is not possible to associate it with the rainfall data.
In the conventional method, when the generated rainfall is not available as described above, the CL is set at the upper limit of the non-generated rain. In this method, the safety area of the CL may become extremely large, and it is often a dangerous CL. Therefore, there is a risk of landslide disasters being missed during heavy rains. At present, there are few methods for objectively accurate CL using only non-occurring rainfall.
Therefore, there is a need for a method for objectively setting accurate CL even with only non-occurrence rainfall, until CL is set only with non-occurrence rainfall accuracy, accurate occurrence rain is obtained, and CL is reviewed. In the meantime, it can be used for temporary evacuation as temporary CL. In this embodiment, νSVM is applied, and the setting of CL is attempted only by non-occurrence rainfall even in an area having a small disaster record by objectively determining the outlier.

Support vector machine is a method proposed by Vapnik et al. In 1992 and is currently attracting attention as a method for pattern classification and function approximation. In general pattern classification, there are overwhelmingly more problems that are not linearly separable than problems that are linearly separable. In the support vector machine, when a pattern classification problem cannot be linearly separated, it can be brought into a state where linear separation is possible by a non-linear mapping, and an optimal separation hyperplane can be obtained.
Support vector machines include a hard margin method that aims to completely separate two sets A and B (occurrence and non-occurrence in the present invention) and a soft margin method that allows a certain degree of misclassification (noise). .
First, the concept of the hard margin method for the purpose of complete separation will be described. As shown in FIG. 6, when there are two sets A (occurrence) and B (non-occurrence), the optimal separation hyperplane H ₀ that separates the two sets is expressed by Equation (3),

This H ₀ can be obtained by maximizing the distance (margin) between the two separated hyperplanes H ₁ and H ₂ passing through the closest data expressed by the equations (4) and (5). .

This is the hard margin method. However, in real problems, as shown in FIG. 7, there is a high possibility that the data contains noise. Therefore, the hard margin method that performs complete separation is prone to over-learning (generalization ability (additional data) Is not necessarily good in terms of discriminating power). In particular, among the landslide disasters, there are sporadic landslides that occur in light rain, and these disasters need to be removed as a pretreatment when setting CL in the past. The process tends to be subjective as well as CL setting.
On the other hand, as shown in FIG. 8, the soft margin method maximizes the margin and simultaneously introduces the slack variable ξ to minimize it. By introducing the slack variable ξ, the constraint condition can be relaxed and misclassification can be allowed. The degree to which misclassification is allowed is determined by C. If C is increased, the degree of misclassification is further reduced, so that a discrimination result close to the hard margin method is obtained.
In this embodiment, paying attention to the above-mentioned soft margin method among support vector machines in particular, and interpreting sporadic cracks as noise, using this soft margin method objectively selecting sporadic cracks, Try setting CL.
All of the above ideas are effective for problems that can be linearly separated. When nonlinearity is strong and linear separation is impossible, the following examination is performed. As shown in FIG. 9, the original space X is mapped to a high-dimensional feature space Z by a certain non-linear mapping (z = φ (x)), and is in a linearly separable state. If the above-described soft margin method is applied to this state, an optimal nonlinear separation hyperplane can be obtained. Although the present invention can be applied to the case where linear separation is possible with respect to the setting of the sediment disaster occurrence limit line (linear CL), in this embodiment, an example of the case where linear separation is not particularly possible (nonlinear CL) will be taken up. .

  The soft margin method is formulated by introducing a slack variable ξ to the hard margin method formula that completely separates the set A and the set B, thereby relaxing the constraint as shown in the following formula (6). Misclassification can be considered.

  In addition, since it is necessary to minimize misjudgment, the formulation of the soft margin method is expressed by the following equation (7).

  Here, C is a weight parameter for the slack variable ξ, and if C is increased, the degree of erroneous discrimination is further minimized, so that the discrimination result is close to complete separation. Further, x represents occurrence rain and non-occurrence rainfall, and y represents a teacher value (-1 for occurrence rain, 1 for non-occurrence rainfall). w is a variable called a normal vector of a separated hyperplane, and b is a variable called a bias term.

  Here, considering the Lagrangian dual problem for equation (7), it becomes as follows.

Here, α is called a Lagrangian variable. Furthermore, the following equation holds from the dual relationship between Equation (7) and Equation (8).

  At this time, the inner product of the vector expressed by the equation (10) exists in the second term of the objective function of the equation (8). However, since the feature space is generally high-dimensional, the calculation of the inner product takes a lot of time. Cost.

  Therefore, by using a kernel function that preserves the inner product, the calculation can be simplified and the discriminant function can be extended nonlinearly. In this embodiment, a Gaussian kernel is applied among the kernel functions. Therefore, Equation (8) and the discriminant function f (x) can be rewritten as follows using Equation (9) and Equation (10).

  Among the data used in the analysis, data in which α in Expression (12) satisfies α> 0 is called a support vector, and these exist mainly around the separation hyperplane. Therefore, the support vector is indispensable data for setting the separation hyperplane. In this embodiment, among the support vectors, data included between two separate hyperplanes f (x) = ± 1 is defined as a support vector having high importance (hereinafter referred to as “important data”), Support vectors not included in this period were recognized as outliers. In other words, the sporadic landslides that are conventionally said to be unpredictable can be considered as a support vector for the occurrence in the safe area. Data other than the support vector was defined as data of low importance for setting the separation hyperplane (hereinafter referred to as “unnecessary data”).

  Note that important data (support vector existing in the range of -1 ≤ f (x) ≤ +1) is data near the boundary between occurrence and non-occurrence of landslides, and it is difficult to judge whether or not it actually occurs. This is zone data. In this embodiment, since the non-linear CL is set only by the rainfall factor, it is inferred that the gray zone data affected the occurrence / non-occurrence of landslide disasters due to differences in topography and geological conditions. In other words, if the gray zone data can be extracted objectively as in the present invention, it is possible to identify a point or area where the cause of the occurrence of landslide disasters cannot be clarified only by rainfall factors. Therefore, if this information obtained from the present invention is effectively used, it is possible to easily narrow down the locations that require particularly detailed field surveys from a large number of survey locations, and to carry out disaster prevention projects efficiently. It is done.

  νSVM can set a separation hyperplane in one class, and the basic idea is that when mapping to feature space Z by nonlinear mapping, the off-point that is isolated from others in the input space is near the origin of the feature space It takes into account the separation hyperplane that separates the origin and the sample group by utilizing the property of being mapped to. The support vector machines up to the above correspond to two classes (occurrence: -1, non-occurrence: +1), and both occurrence and non-occurrence rainfall are indispensable. On the other hand, νSVM has an advantage that a non-linear CL can be set only by non-occurring rainfall.

  Then, the formulation of νSVM is as follows.

Here, x is a generated rainfall and a non-occurring rainfall, and ν is a parameter representing a ratio of misclassified data. w is a variable representing the normal vector of the separation hyperplane, ρ is a variable called a bias term, and ξ _i is a slack variable.

  Here, the dual problem for equation (15) is considered as follows. The non-linear extension is performed using a Gaussian kernel as in the support vector machine of Expression (12).

The optimum separation hyperplane is set so that a predetermined group ν of sample groups (outliers) remain on the origin side as shown in FIG. Therefore, ν in the equation (15) is a parameter that determines the range of outliers. The larger the ν, the wider the range of outliers. That is, the safety area of the nonlinear CL is reduced. This ν must be studied experimentally as well as the support vector machine parameters. The support vector is non-occurrence rainfall with α> 0.
By solving the equations (15) and (16) and obtaining α and ρ, the discriminant function is as follows.

Next, as a specific calculation example, in this example, the support vector machine described above was applied to the setting of the rainfall limit line for the occurrence of landslide disasters in Shimonoseki City, Yamaguchi Prefecture, especially landslide disasters, and a non-linear CL setting example as a reference example 1 will be described with reference to FIGS. 11 to 22 and Tables 1 to 6. FIG.
In the Chugoku region, Yamaguchi Prefecture has a lot of sediment-related disasters, and Shimonoseki City has one of the most dangerous places in the prefecture, and is one of the areas where early enhancement of the warning and evacuation system is desired. In this example, among the landslide disasters that occurred in Shimonoseki City, the landslide disasters occurred within a radius of 5 km centered on the Shimonoseki Meteorological Observatory (1975-1999: 80, 2000-2003: 33) Subject to analysis. In this embodiment, the minimum unit of CL setting is set within a radius of 5 km from Shimonoseki Meteorological Observatory. Therefore, differences (risk levels) such as topography, geology, and vegetation factors that differ from point to point are not considered.

In this embodiment, in order to show the usefulness of the present invention with respect to the prior art, a linear CL proposed by the former Ministry of Construction (current Ministry of Land, Infrastructure, Transport and Tourism) and a nonlinear CL (hereinafter referred to as a non-linear CL) using a method based on RBFN published as a prior art. , Referred to as “RBFN-CL”) and comparison with the nonlinear CL according to the present invention is performed. First, the linear CL is set. As the linear CL, “Precautionary evacuation reference rainfall setting method for concentrated landslides (draft)” examined in the “Comprehensive Sediment Disaster Countermeasure Study Committee” sponsored by the former Ministry of Construction (current Ministry of Land, Infrastructure, Transport and Tourism) (Hereinafter referred to as “proposal proposal”). This proposal is effective only for landslides that occur intensively. When setting, the past landslide-generated rainfall is classified as intensively generated rainfall (hereinafter referred to as “intensive occurrence”) and sporadic generated rainfall. (Hereinafter referred to as “occurrence of sporadic”).
Concentrated landslides mean that the area around the peak of a series of rainfalls (a series of rains separated by a period of no rainfall for 24 consecutive hours) when the effective rainfall exceeds a certain level qualitatively. It is defined as a collapse that occurs in a limited range. In this example, intensive occurrence and sporadic occurrence were classified under the following conditions with reference to this definition.
Concentrated occurrence: Among disasters that occurred within 3 hours after the peak of hourly rainfall and effective rainfall (half-life 72 hours) of a series of rainfalls, effective rainfall (half-life 72 hours) is 140 or more. According to the above conditions, in this embodiment, 80 cases of landslide occurrence in 1975 to 1999 were classified into 39 cases of concentrated occurrence and 41 cases of sporadic occurrence. The classification results of concentrated occurrence and sporadic occurrence are as shown in FIG. The occurrence of rainfall used to set the linear CL uses only concentrated occurrence in the combination of effective rainfall (half-life 1.5 hours) and effective rainfall (half-life 72 hours) at the time of occurrence, and non-occurrence rainfall does not occur. The combination of the effective rainfall (half-life 1.5 hours) and the effective rainfall (half-life 72 hours), which is the farthest from the origin, was used. As shown in FIG. 11 (b), the linear CL is set as the lower limit for the rainfall from the origin most in the concentration occurrence, and the slope of the straight line is set so that the hit rate for the non-occurrence rainfall is the highest. From FIG. 11 (b), it can be seen that there are a plurality of non-occurring rainfalls exceeding the CL because the shape of the linear CL is a straight line.

Next, RBFN-CL setting using RBFN is performed. The RBFN-CL was set by applying the prior art to the occurrence / non-occurrence rainfall from 1975 to 1999 as well as the proposal. The generated rainfall used for the setting did not classify concentrated occurrence and sporadic occurrence, and used the combination of effective rainfall (half-life 1.5 hours) and effective rainfall (half-life 72 hours) at the time of occurrence in all occurrences. Non-occurrence rainfall is a combination of the effective rainfall (half-life 1.5 hours) and effective rainfall (half-life 72 hours) of all rains except the occurrence of rain (hereinafter referred to as “all non-occurrence rainfall”). (See FIG. 12). However, in this case, the total non-occurrence rainfall is within the range where rainfall data exists as shown in Fig. 13 (a) according to the method of the prior art, the vertical axis (effective rainfall (half life 1.5 hours)), the horizontal axis (effective rainfall). (Half-life 72 hours)) Set only a grid at regular intervals (ΔRx, ΔRy), and only when data exists in the grid as shown by the hatched area (cluster 1 to cluster 4) in FIG. For the data in the grid, clustering processing using the centroid method was performed. The representative points of each region by this clustering were regarded as non-occurrence rainfall used for analysis.
In this example, ΔRx and ΔRy were set to ΔRx = 5.0 mm and ΔRy = 1.5 mm, respectively, according to the method of the prior art. To set the nonlinear CL is also utilized RBFN other, setting the radius Rx of the basis functions, Ry, suppression parameter lambda _max for a non-occurrence rainfall, lambda _min, suppression parameter lambda _max for generating rainfall, the lambda _min as a parameter There is a need to. In this embodiment, these parameters are set to Rx = 35.0 mm, Ry = 10.5 mm, non-occurrence rainfall suppression parameters λ _max = 100, λ _min = 1, and generated rain suppression parameters λ _max = λ _min according to the prior art method. = 10. FIG. 14 shows the set RBFN-CL.

Next, non-linear CL (hereinafter referred to as “SVM-CL”) using a support vector machine is set. SVM-CL was set by applying the soft margin of the support vector machine using the 1975-1999 landslide occurrence / non-occurrence rainfall as well as the proposal proposal, RBFN-CL. The generated rainfall used for the setting does not classify concentrated occurrence and sporadic occurrence as in the RBFN-CL setting, and the effective rainfall (half-life 1.5 hours) and effective rainfall (half-life 72 Time) combination was used. In addition, in order to compare with RBFN-CL, non-occurrence rainfall was clustered in the same way as RBFN-CL, and representative points of each region by clustering were used as non-occurrence rainfall for analysis.
In the prior art by RBFN, when the basis functions are arranged in all the occurrence / non-occurrence rainfall, there is a risk that a non-linear CL with low generalization ability is constructed due to extreme density of data. Therefore, in the prior art based on RBFN, it is necessary to perform the above-described clustering as preprocessing and equalize the data density to some extent. Note that the support vector machine is easier to analyze than RBFN-CL because it is not necessary to perform the above clustering process for non-occurrence rainfall when setting nonlinear CL.

In the soft margin of the support vector machine, it is necessary to determine the radius r and the parameter C of the kernel function shown in Expression (12). First, regarding C, if C is made too large, generally the generalization ability is reduced. Known to tend to be lower. Therefore, each separation hyperplane is constructed by changing C, and the generalization ability and discrimination ability are evaluated from the shape and prediction accuracy of CL to determine the optimum value. Regarding the radius r of the Gaussian kernel, there is no particular optimal determination method, and there are many cases that can be determined empirically in the previous research. Also in this patent, as a result of examining various r, the standard deviation σ of data is used. Although there is no theoretical basis, it is known that the discrimination ability is good when the standard deviation σ of the data is used through many numerical examples so far.
On the other hand, the parameters determined on the basis of normalized occurrence / non-occurrence rainfall can be used as almost optimal parameters even in regions with different rainfall conditions. The parameters of the support vector machine obtained by this embodiment are Know-how can be used, and SVM-CL with high generalization ability can be set without requiring trial and error. In this embodiment, the generated and non-occurring rainfall using the SVM-CL is normalized by the average value and standard deviation according to the following equation (18).

  Here, x ′ means occurrence / non-occurrence rainfall after normalization. When the above equation (18) is applied to the equation (12), the radius r is fixed to 1, and the equation (12) is modified, the following equation (19) is obtained.

  This is equivalent to the case where the data before normalization and the radius r are substituted into the equation (12) as the standard deviation σ. Further, as described above, it is known that the radius r has a high discrimination ability when the standard deviation σ is empirically set. Therefore, normalizing the data with the equation (18) and analyzing the radius as r = 1 is the same as analyzing the radius r as the standard deviation σ for the data that is not normalized. For these reasons, in this embodiment, after normalizing the data, the radius r = 1 is analyzed.

  Next, we try to set nonlinear CL using the soft margin of the support vector machine. Here, specific input data and output data obtained by analysis will be described using Tables 1 to 4 so that the analysis procedure can be understood. Table 1 shows a part of the landslide disaster report (hereinafter referred to as “disaster data list”), in which the date and time of occurrence, the name of the station from which rainfall data is extracted, etc. are organized. In addition, it is desirable to select a nearby station as a disaster occurrence site for the rain station from which rainfall data is extracted. In this embodiment, a disaster occurrence location within a radius of 5 km centered on the Shimonoseki weather station is targeted.

  Table 2 shows an example of rainfall data extracted from the observation stations (hereinafter referred to as “rainfall data list”), and such a list is organized for each observation station. Most of the rainfall data at the observatory is the hourly rainfall every hour. In this example, the effective rainfall (half-life 1.5 hours) is used as the short-term rainfall index for nonlinear CL, and the effective rainfall (half-life 72 hours) is used as the long-term rainfall index. Is also shown. In addition, in order to set a highly accurate nonlinear CL, it is necessary to select an optimum rainfall index for each region. Short-term and long-term rainfall indices include effective rainfall, hourly rainfall, cumulative rainfall, and other effective rainfall using other half-lives.Therefore, the rain data list includes all possible rainfall indices based on hourly rainfall. It is good to calculate and organize.

  Next, extraction of generated rainfall and non-generated rainfall will be described. For occurrence rainfall, the rainfall index at the date and time of occurrence in the disaster data list is extracted from the rain data list, and for non-occurrence rainfall, all rainfall indexes except for a series of rainfall including occurrence rainfall are used. Table 3 (a) is a list of generated and non-occurring rainfalls extracted from the rainfall data list. The teacher value shown in Table 3 (a) corresponds to y in equation (12), where the rainfall index with a value of 1 is non-occurring rainfall, and the rainfall index with a value of -1 is generating rainfall. Table 3 (c) calculates the average value and standard deviation for each of the rainfall indicators in Table 3 (a) (see Table 3 (b)), and uses multiple normalized rainfall indicators and teachers according to Equation (18). The value (hereinafter referred to as “data set”), the radius r of the Gaussian kernel shown in the equation (12), and C for determining the ratio for allowing misidentification are organized (hereinafter referred to as “analysis data list”). .

The analysis of the support vector machine is optimized from the data in the analysis data list using equation (12), and α _i (i = 1, 2,..., N) and the value of b which is a bias term are calculated. It is in. Table 4 is a list of analysis results of support vector machines (hereinafter referred to as “analysis result list”). In Table 4, r and C values, α _i (i = 1, 2,..., N) and b values, f (x) values, identification results, and importance levels for each data set are organized. Yes. Hereinafter, the value of f (x), the identification result, and the importance will be described.

Since the discriminant function f (x) is given by Expression (11), each data set obtained by analysis, α _i (i = 1, 2,..., N) and the values of b and expressions for each data set, Input to (11) and calculate the value of f (x).

If the value of f (x) is positive, the identification is 1 (non-occurrence), and if it is negative, it is -1 (occurrence). If they do not match, it is misjudged.
The importance level indicates a classification result of each data set based on definitions of important data, unnecessary data, and abnormal values. According to Table 4, the data set of ID1 is positively determined as non-occurrence, and α _i > 0, so it becomes a non-occurrence support vector. The value of f (x) is 0.989, and is in the range of −1 ≦ f (x) ≦ + 1.
According to Table 4, the data set of ID747 is a generation support vector, and the value of f (x) is -0.561, so it is classified as important data like ID1. The data set of ID 487 is a support vector for occurrence, but the identification result is an erroneous determination (f (x) = 1.003), and thus an abnormal value (walk occurrence). All other data sets with α _i = 0 are unnecessary data.

Next, setting examples of the separation hyperplane set according to the analysis procedure are shown in FIGS. The separation hyperplane is a value obtained by displaying the value of f (x) in each data set with contour lines. 15 to 20 show separation hyperplanes in which the radius r of the Gaussian kernel is fixed to 1 and C is changed. As shown in FIGS. 15 (a) to 15 (d), when C is increased, the degree of misjudgment (generated rain in the safety area or non-occurrence rainfall in the dangerous area) is further reduced. A close discrimination result is obtained, and the shape is strongly nonlinear.
As shown in FIG. 21, the separation hyperplane for the occurrence of landslide disaster and non-occurrence rainfall must always have a negative slope so as not to contradict natural phenomena. The separation hyperplane shown in FIG. 15 is natural in that there is a region where the non-linearity is strong and the slope is partly positive. It can be said that the shape contradicts the phenomenon (see FIG. 21). Therefore, it can be determined that the range C shown in FIG. 15 is not suitable for setting the separation hyperplane.
16 to 20 show examples in which C is gradually reduced and a separation hyperplane is set. It can be seen that by reducing C, a large number of misclassifications are allowed, and the separation hyperplane has a simple shape. As shown in FIG. 16 to FIG. 20, as a result of setting the separation hyperplane by changing some C, the separation hyperplane showed a smooth downward-sloping shape with C = 1 to 10 and consistent with the natural phenomenon. Among them, C = 1 captures more non-occurrence rainfall. Further, as shown in FIGS. 19 and 20, there is no difference in the shape and the hit rate of non-occurrence rainfall even if C is made smaller than 1. In the same condition, learning is more accurate as C is larger. In this embodiment, C = 1 is determined as the optimum value. The radius r was set to r = 1 as described above.
Further, it can be seen from FIG. 18 that by setting the separation hyperplane by the support vector machine, it is possible to objectively classify support vectors (important data), unnecessary data, and abnormal values that are highly important with respect to the setting of the separation hyperplane.

As shown in FIG. 18, three types of separation hyperplanes f (x) = 0, f (x) = + 1, and f (x) = − 1 can be obtained from the analysis result of the support vector machine. In this embodiment, f (x) = + 1 on the origin side is adopted for the non-linear CL in consideration of safety in order not to overlook landslide disasters. Hereinafter, the separation hyperplane f (x) = + 1 is denoted as SVM-CL. FIG. 22 (a) shows SVM-CL, RBFN-CL, and a proposal. In RBFN-CL, the contour line 0.8 is adopted so that the hitting ratio for the generated rainfall is the highest among several contour lines (0.2 to 0.8) shown in FIG. From FIG. 22 (a), it can be seen that SVM-CL and RBFN-CL have a non-linear shape that reproduces the rainfall distribution, and capture more non-occurrence rainfall than the proposal. In Fig. 22 (a), there is rainfall that occurs in the safe area of SVM-CL and RBFN-CL, but this disaster occurred in SVM-CL, RBFN- three hours before the disaster time as shown in Fig. 22 (b). CL exceeded. Therefore, it was judged that this outbreak was occurring because it could be dealt with sufficiently if a warning evacuation was issued when CL was exceeded.
Table 5 shows the verification results of the prediction accuracy of SVM-CL, RBFN-CL, and the proposed proposal. Table 5 shows the results of verification of the prediction accuracy for the occurrence and non-occurrence rainfall from 1975 to 1999 used in the analysis. In order to make a comparison with the proposal, the occurrence of rainfall was performed only for concentrated occurrence based on the definition of the proposal. Table 5 shows that SVM-CL, RBFN-CL, and the proposed proposal have a hit rate of 100% for concentrated occurrence, but SVM-CL and RBFN-CL are about 7% of the proposed proposal for non-occurring rainfall. Improvement in accuracy was observed. The prediction accuracy of SVM-CL was about 0.6% higher than that of RBFN-CL.

  Next, Table 6 shows the accuracy verification results for occurrence / non-occurrence rainfall (hereinafter referred to as “unlearned data”) from 2000 to July 2003, which is not used for analysis of support vector machines or RBFN. From Table 6, it can be said that SVM-CL and RBFN-CL have high prediction accuracy for unlearned data, so that the generalization ability is superior to the proposal. The prediction accuracy of SVM-CL was 100% for both occurrence and non-occurrence rainfall, 0.2% higher than RBFN-CL. From the above, nonlinear CL using a support vector machine has higher prediction accuracy than conventional technologies, and the setting method itself is less dependent on subjectivity, so it is considered to be effective as a landslide disaster (slope failure) occurrence limit rainfall line. Compared with the prior art by RBFN, the support vector machine does not have much difference in the prediction accuracy of the nonlinear CL, but the setting method itself has the following advantages.

(1) In RBFN, when setting non-linear CL, it is necessary to perform clustering of non-occurrence rainfall as preprocessing, but support vector machine does not need preprocessing.
(2) In order to set nonlinear CL with RBFN, lattice spacing ΔRx, ΔRy, basis function radii Rx, Ry, non-occurrence rainfall suppression parameters λ _max , λ _min , generated rain suppression parameters λ _max , λ _min , It is necessary to examine the optimum value by trial and error in total of 8 parameters. On the other hand, in the support vector machine, it is only necessary to consider the optimum values of the two parameters in total, that is, the radius r of the basis function, C that determines the rate at which misclassification is allowed. Therefore, support vector machines have fewer parameters to consider than RBFN.
(3) The support vector machine can not only set nonlinear CL as a result of analysis, but can also objectively classify support vectors (important data), unnecessary data, and abnormal values that are highly important for setting nonlinear CL. As described above, it can be said that the nonlinear CL setting method using the support vector machine is superior to the prior art using RBFN in the above three points. An embodiment using the above feature (3) will be described below as a reference example 2.

In this example, in order to improve the problem of CL update, CL with the minimum necessary data added newly generated / non-occurring rainfall to the support vector (important data) with high importance extracted when setting the current CL Try to update. By narrowing down all the past occurrence / non-occurrence rainfall to important data without updating, it is expected to improve the efficiency of CL review work. However, there is a possibility that unnecessary data (other than support vectors) or abnormal values determined when setting the current CL will be important when updating the CL. Therefore, in this embodiment, unnecessary data and abnormal values included in the distance ε from the separation hyperplane f (x) = ± 1 are added to the important data as generated / non-occurring rainfall used for CL update as shown in FIG. I decided to consider it. Hereinafter, this data is referred to as update data together with important data. The definition of update data is as shown below.
Update data:-(1 + ε) ≦ f (x) ≦ (1 + ε) important data, unnecessary data, abnormal value

Hereinafter, in this example, as an example of updating the non-linear CL, reference example 2 in which the non-linear CL from 1975 to 1999 set in the above-described examination was updated every year until 2002 is shown in FIGS. Reference Example 3 in which a non-linear CL was set in 1975 to 1983 and updated every 10 years in 1993 and 2003 is shown in FIGS. 30 to 38 and Table 7. FIG. At the time of examination, some ε were changed.
The results of Reference Example 2 are shown in FIGS. 24 to 29, three cases of ε = 0 (use only important data), ε = 0.001, and ε = 0.005 were examined. 24 (a), (b), FIG. 25 (a), (b) when ε = 0, and FIGS. 26 (a), 26 (b), 27 (a), 27 when ε = 0.001. FIGS. 28 (a), 28 (b), 29 (a), and 29 (b) show cases where (b) and ε = 0.005.
From FIG. 24 to FIG. 29, it can be seen that the nonlinear CL updated every year in all three cases has no significant difference in shape. This is because the renewal period is only 3 years, so the conditions such as ground strength have hardly changed, and unless there is a large-scale occurrence or non-occurrence rainfall that exceeds nonlinear CL compared to past disaster records. , CL is presumed not to change greatly, and it can be judged as a reasonable result. In fact, there were only 3 disasters in the target area that fell between 2000 and 2002, and it was not particularly heavy compared to past disasters.

Subsequently, as Reference Example 3, an example in which the nonlinear CL is updated every 10 years is shown in FIGS. Compared to Reference Example 2, in the renewal period of Reference Example 3, since large-scale landslide disasters occurred in 1993, 1999, and 2003, it is estimated that nonlinear CL is greatly improved. In Reference Example 3, as in Reference Example 2, three cases of ε = 0 (only important data used), ε = 0.001, and ε = 0.005 were studied. 30 (a), (b), FIG. 31, when ε = 0, FIGS. 32 (a), (b), FIG. 33 when ε = 0.001, and FIG. a), (b) and FIG.
From Fig. 30 to Fig. 35, in all three cases, the non-linear CL in 1993 is safer than the non-linear CL in 1983 because there is important data that does not occur in the vicinity of 60mm to 70mm effective rainfall (half life 1.5hr). The area is greatly enlarged in the vertical axis direction. In addition, the frequent occurrence of landslides in 2003 and the availability of rainfall data from sabo rain gauges densely arranged in the prefecture enabled the reliable occurrence of rainfall to be reflected in SVM-CL. As a result, at the time of renewal in 2003, the non-occurrence rainfall was low and the occurrence of rain increased in a range that was not reliable as a safe area, so the non-linear CL after renewal has a smaller vertical safety area than before. I understand that. For this reason, CL needs to be updated sequentially when accurate and non-occurrence rainfall is available, and for that purpose it is renewed that it is important to update CL as easily and efficiently as possible. it can.
In the data used in this example, there was no significant difference in the update results even when ε was changed. However, when applying the present invention in other regions, it is necessary to experimentally examine the optimum ε. There is.

  Next, regarding the update of CL, in order to compare the prior art by RBFN and the support vector machine, RBFN-CL was updated as shown in FIGS. 36 to 38 using the same data in RBFN. 36 to 38, (a) shows CL by a support vector machine, and (b) shows CL by RBFN. Table 7 shows the number of data and analysis time used at the time of 1993 update and 2003 update. From Table 7, it can be seen that the number of data used for the update is as small as about 1/4 of RBFN-CL in SVM-CL. This is because all data is not used for updating the non-linear CL, and updating is performed only with highly important data. In RBFN-CL, the number of data increased in 2003 compared to 1993, but SVM-CL shows no significant difference. This suggests that when the nonlinear CL is updated using the prior art by RBFN, the amount of data to be handled will become enormous and some time and effort will be required for setting the nonlinear CL. This can also be easily estimated from the analysis time in Table 7. From the above, as a nonlinear CL setting method, it is considered that the utility of the present invention is higher than the prior art in that the nonlinear CL can be updated with the minimum necessary data.

In areas where there have been few disasters in the past, there are cases where it is not possible to obtain the generated rainfall that can be used for CL setting due to problems such as the reliability of the disaster time and the distance between the disaster occurrence location and the rainfall observation station. In such an area, there is no choice but to respond to warning evacuation by setting CL only with non-occurrence rainfall as provisional CL until accurate rainfall is available.
In recent years, prefectures have begun to closely maintain rainfall stations for the purpose of warning and evacuation of sediment-related disasters and flood control activities. In Yamaguchi Prefecture, 110 Sabo Rainfall Bureaus have been set up in the prefecture, and it is possible to obtain accurate rainfall data in units of 10 minutes. That is, in the future, it is considered that non-occurrence rainfall with high accuracy including occurrence rain can be obtained in finer units. Therefore, it is important to set a non-linear CL with high precision non-occurrence rainfall until the occurrence of rainfall is available even when warning evacuation is carried out in units of small areas.
In the present embodiment, description will be given with reference to Table 8, Table 9, and FIG. 39 as Reference Example 4 in which nonlinear CL is set only by non-occurring rainfall. Using the non-occurrence rainfall (2000-2003) of the Shinmachi Rainfall Bureau, which has not had a disaster record in recent years, among the above Sabo Rainfall Bureau, nonlinear CL was set by νSVM as follows.

As in Reference Example 1, the occurrence and non-occurrence rainfall used for the analysis of νSVM is extracted from the rain data list for the occurrence rain and the occurrence date and time of the disaster data list. Use all rainfall indicators except for. Table 8 (a) is a list in which generated and non-occurring rainfalls extracted from the rainfall data list are arranged. Table 8 (c) calculates the average value and standard deviation for each of the rainfall indices in Table 8 (a) (see Table 8 (b)). The values (hereinafter referred to as “data set”), the radius r of the Gaussian kernel shown in the equation (16), and the parameter ν that determines the range of outliers are arranged.
The analysis of νSVM is optimized from the data in Table 8 (c) using Equation (16), and α _i (i = 1, 2,..., n) and the value of ρ which is a bias term are calculated. It is in.
Table 9 is a list of the results of analysis by νSVM as Reference Example 4. Table 9 summarizes the values of r and for each data set, values of α _i (i = 1, 2,..., N) and ρ, values of f (x), identification results, and importance. . The values of f (x), identification results, and significance are the same as in Reference Example 1.

  39A to 39D show the results of setting nonlinear CL in four cases of ν = 0.025, ν = 0.050, ν = 0.075, and ν = 0.100, respectively. In ν = 0.025 in FIG. 39 (a), the shape of the nonlinear CL did not smoothly fall to the right, and became a shape inconsistent with the natural phenomenon. In the other three cases where ν was increased, the shape became simple as ν increased, and the safety area of the nonlinear CL tended to decrease. Among these 4 cases, it can be judged that the non-linear CL of ν = 0.050, which has a simple shape and captures more non-occurrence rainfall, is appropriate. Non-linear CL set only by non-occurrence rainfall, when highly accurate generated rain is obtained in the future, reset CL using a 2-class support vector machine, so that more reliable non-linear CL can be used for warning evacuation It is presumed that it can respond.

  As mentioned above, although the Example of this invention was described, it is clear that various changes may be made to the form and detail, without deviating from the essence and scope of this invention prescribed | regulated by the claim. .

  It has a wide range of uses such as drafting disaster prevention plans and creating hazard maps in public institutions such as local governments and disaster prevention centers. It is also expected to be used as a teaching material for disaster prevention and evacuation drills at educational institutions, etc., and for private companies operating construction and civil engineering projects, It can be used as a shared tool for cooperation with public institutions, and can also be applied to research and development and design projects related to corporate disaster prevention technology.

It is a block diagram of the disaster prevention business plan support system which concerns on this Embodiment. It is a flowchart of a disaster prevention business plan support method. It is a flowchart for demonstrating the detail of step S5 in FIG. It is a flowchart for demonstrating the detail of step S6 in FIG. It is a flowchart for demonstrating the detail of step S7 in FIG. It is a conceptual diagram for demonstrating the support vector machine of a hard margin. It is a conceptual diagram for demonstrating the support vector machine of a soft margin. It is a conceptual diagram for demonstrating the support vector machine of a soft soft margin. It is a conceptual diagram for demonstrating a support vector machine. It is a conceptual diagram for demonstrating 1 class of support vector machines. (A), (b) is both a conceptual diagram at the time of setting linear CL using the disaster prevention business plan support system which concerns on this Embodiment, or its method. It is a conceptual diagram of the generation | occurrence | production / non-occurrence | production rain used for the analysis for the setting of CL. (A), (b) is the conceptual diagram of the non-occurrence | production rain extraction used for the analysis using RBFN shown compared with this Embodiment. It is a conceptual diagram of the nonlinear CL set by the analysis using RBFN shown in comparison with the present embodiment. (A)-(d) are all the conceptual diagrams of the nonlinear CL analyzed using the disaster prevention business plan support system which concerns on this Embodiment, or its method. It is a conceptual diagram of nonlinear CL analyzed using the disaster prevention business plan support system concerning this embodiment, or its method. It is a conceptual diagram of nonlinear CL analyzed using the disaster prevention business plan support system concerning this embodiment, or its method. It is a conceptual diagram of nonlinear CL analyzed using the disaster prevention business plan support system concerning this embodiment, or its method. It is a conceptual diagram of nonlinear CL analyzed using the disaster prevention business plan support system concerning this embodiment, or its method. It is a conceptual diagram of nonlinear CL analyzed using the disaster prevention business plan support system concerning this embodiment, or its method. It is a conceptual diagram of CL provided with the shape inconsistent with a natural phenomenon. (A), (b) is a figure which compares CL set using the support vector machine used for this Embodiment, and RBFN made into the comparison object. It is a conceptual diagram of the update method of CL using (epsilon). (A) is a conceptual diagram of non-linear CL analyzed using data of ε = 0 from 1975 to 1999, and (b) is a CL updated in 2000 using updated data of ε = 0. It is the conceptual diagram shown. (A), (b) is the conceptual diagram which showed CL updated in 2001 and 2002, respectively using the update data of (epsilon) = 0. (A) is a conceptual diagram of nonlinear CL analyzed using data of ε = 0.001 from 1975 to 1999, and (b) is a CL updated in 2000 using updated data of ε = 0.001. It is the conceptual diagram shown. (A), (b) is the conceptual diagram which showed CL updated in 2001 and 2002, respectively using the update data of (epsilon) = 0.001. (A) is a conceptual diagram of a nonlinear CL analyzed using data of ε = 0.005 from 1975 to 1999, and (b) is a CL updated in 2000 using updated data of ε = 0.005. It is the conceptual diagram shown. (A), (b) is the conceptual diagram which showed CL updated in 2001 and 2002, respectively using the update data of (epsilon) = 0.005. (A) is a conceptual diagram of nonlinear CL analyzed using data of ε = 0 from 1975 to 1983, and (b) is a CL updated in 1993 using updated data of ε = 0. It is the conceptual diagram shown. It is the conceptual diagram which showed CL updated in 1993 using CL updated in 1993 using the update data of (epsilon) = 0. (A) is a conceptual diagram of nonlinear CL analyzed using data of ε = 0.001 from 1975 to 1983, and (b) is a CL updated in 1993 using updated data of ε = 0.001. It is the conceptual diagram shown. It is the conceptual diagram which showed CL updated in 1993 using the update data of (epsilon) = 0.001 for CL updated in 1993. FIG. (A) is a conceptual diagram of nonlinear CL analyzed using data of ε = 0.005 from 1975 to 1983, and (b) is a CL updated in 1993 using updated data of ε = 0.005. It is the conceptual diagram shown. It is the conceptual diagram which showed CL updated in 1993 using the update data of (epsilon) = 0.005. (A), (b) is a conceptual diagram showing SVM-CL and RBFN-CL constructed from 1975 to 1983, respectively. (A), (b) is the conceptual diagram which showed SVM-CL and RBFN-CL which were updated in 1993, respectively. (A), (b) is the conceptual diagram which showed SVM-CL and RBFN-CL which were updated in 2003, respectively. (A) to (d) show the results of setting nonlinear CL in four cases of ν = 0.025, ν = 0.050, ν = 0.075, and ν = 0.100, respectively.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 ... Information input device 2 ... Information operation device 3 ... Information output device 4 ... Information storage device 5 ... Request level setting part 6 ... Model selection part 7 ... Analysis parameter setting part 8 ... Discrimination boundary surface analysis part 9 ... CL accuracy verification part DESCRIPTION OF SYMBOLS 11 ... Actual data 12 ... Rainfall data 13 ... Disaster performance data 14 ... Factor evaluation value data 15 ... Request level data 16 ... Analysis parameter 17 ... Model data 18 ... CL data 19 ... Support vector data

Claims (10)

It has an information input device, an information storage device, an information calculation device, and an information output device. It uses precipitation data including short-term rainfall indicators and long-term rainfall indicators in each region, and information on the occurrence and non-occurrence of landslide disasters. A disaster prevention business plan support system that calculates a disaster occurrence limit line, an evacuation reference line, or a warning reference line (hereinafter sometimes collectively referred to as CL),
The information input device includes rainfall data and actual / non-occurrence of sediment disaster data in each region (hereinafter, the rain data and actual / non-occurrence information of sediment-related disaster may be referred to as analysis data), and the above. An analysis parameter and an analysis model for calculating CL may be input to the information storage device,
The information calculation device includes a model selection unit that selects and reads out the analysis model stored in the information storage device;
An analysis parameter setting unit that reads and inputs an analysis parameter from the information storage device or inputs an analysis parameter input from the information input device to the selected analysis model;
Reading analysis data from the information storage device or using analysis data input from the information input device , forming a short-term rainfall index and a long-term rainfall index of the rain data included in the analysis data as a vertical axis or a horizontal axis, respectively. On the grid points of the grid on the two-dimensional plane, the actual information on occurrence / non-occurrence of sediment disasters is plotted to form a discrimination boundary surface, and the CL as a contour line corresponding to a desired threshold is formed from the discrimination boundary surface and a discrimination boundary analysis unit for calculating a,
The information output device is a means capable of outputting at least one of the analysis data, analysis parameter, analysis model, and CL, and is a disaster prevention business plan support system. The information input apparatus, for the CL where the Ru is calculated, from the non-generation predictive value for rainfall data when generating predictive value for rainfall data when landslides occur and / or landslides did not occur A request level setting unit that can be set by adopting an input value from the information input device, or can be set by reading the request level data stored in the information storage device in advance,
The information calculation device reads out the rainfall data and the information on the occurrence / non-occurrence of sediment disaster from the information storage device or uses the rainfall data and the information on the occurrence / non-occurrence of sediment disaster that are input from the information input device. Thus, with respect to the CL calculated by the discriminant boundary surface analysis unit, the occurrence probability for the rain data when a sediment disaster occurs and / or the non-occurrence for the rain data when no sediment disaster occurs. The disaster prevention project planning support system according to claim 1, further comprising a CL accuracy verification unit that calculates a rate and verifies the accuracy of the CL in comparison with the required level set by the required level setting unit . In addition to rainfall data, the analysis data includes an attribute value or an evaluation value related to at least one factor among factors causing a sediment disaster at a point included in each area,
The analysis parameter setting unit reads the attribute value or the evaluation value from the information storage device as a risk of occurrence of a landslide disaster ,
The determination boundary surface analysis unit, the classifying said point in response to the risk of large and small, Disaster Prevention Project planning support according to claim 1 or claim 2, characterized in that computing the CL per this classification system. In addition to rainfall data, the analysis data includes an attribute value or an evaluation value related to at least one factor among factors causing a sediment disaster at a point included in each area,
The analysis parameter setting unit reads the analysis data including the attribute value or the evaluation value from the information storage device, or uses the analysis data input from the information input device, and performs a sediment disaster in each of the category sections of the factor Calculate the sediment disaster occurrence rate from the actual occurrence / non-occurrence information , the cumulative value of all the factors of the sediment disaster occurrence rate as the risk of the point,
3. The disaster prevention business plan according to claim 1, wherein the discriminant boundary surface analysis unit classifies the points according to the degree of risk and calculates the CL for each classification. Support system. The analysis model is a support vector machine (hereinafter sometimes abbreviated as SVM),
The analysis parameter setting unit reads and inputs analysis parameters from the information storage device to the support vector machine, or inputs analysis parameters input from the information input device,
The discriminant boundary surface analysis unit reads analysis data from the information storage device or uses analysis data input from the information input device, and calculates a short-term rainfall index and a long-term rainfall index of the rain data included in the analysis data. Plot the actual information on the occurrence or non-occurrence of landslide disasters on the grid points of the grid on the two-dimensional plane formed as the vertical axis or the horizontal axis, respectively. A value is calculated to form a discrimination boundary surface, CL is calculated from the discrimination boundary surface as a contour line corresponding to a desired threshold value, and the analysis data used for the calculation of CL is classified by a support vector machine. When a support vector is extracted and stored in the information storage device, and then the analysis data accumulates additional data , Disaster Prevention Project planning support system according to any one of claims 1 to 4, characterized in that it is updatable CL by using the additional data and the support vectors.   Among the analysis data not classified into the support vector by the support vector machine, the analysis data included in a predetermined distance ε from the separation hyperplane f (x) = ± 1 of the support vector machine is the support vector. 6. The disaster prevention business plan support system according to claim 5, wherein CL is calculated by being included in the additional data. 7. The disaster prevention business plan support system according to claim 5, wherein the support vector machine of the analysis model is a support vector machine having a soft margin expressed by the following equation:
However, x is generated rain and non-occurring rain, and y is a teacher value (-1 for rain generated, 1 for rain not generated). w is a variable called a normal vector of the separated hyperplane, b is a variable called a bias term, and ξ _i is a slack variable. C is a weighting parameter representing a trade-off between the margin (w ^T w) and the slack variable (ξ _i ). 7. The disaster prevention business plan support system according to claim 5 or 6, wherein the support vector machine of the analysis model is a ν support vector machine expressed by the following equation:
Here, x is a generated rain and a non-occurring rain, and ν is a parameter representing a ratio of misclassified data. w is a variable called a normal vector of the separated hyperplane, ρ is a variable called a bias term, and ξ _i is a slack variable. While the computer executes each process, it calculates CL using rainfall data including short-term and long-term rainfall indices in each region , actual information on occurrence / non-occurrence of landslide disasters, and evaluation values for each factor causing landslide disasters. A disaster prevention business plan support method,
Model selection unit of the computer, the steps of be read out by selecting an analysis model stored in advance in the information storage device,
An analysis parameter setting unit of the computer reads an analysis parameter from the information storage device and inputs the analysis parameter to the analysis model or inputs an analysis parameter input from the information input device;
The determination boundary analysis unit of the computer, using an analysis data input from reading the analysis data from the information storage device or the information input device, short-term rainfall index and long-term rain rainfall data included in the analysis data On the grid points of the grid on the two-dimensional plane formed with the index as the vertical axis or the horizontal axis, respectively, the discriminant boundary surface is formed by plotting the record information of occurrence / non-occurrence of the earth and sand disaster. Calculating CL as a contour line corresponding to a desired threshold ;
Disaster Prevention Project planning support method characterized by comprising the step information output device of the computer, for outputting at least one of the information of the analytical model and the CL and the analysis data and analysis parameters. While the computer executes each process, it calculates CL using rainfall data including short-term and long-term rainfall indexes in each region , actual information on occurrence / non-occurrence of landslide disasters, and evaluation values for each factor causing landslide disasters. A disaster prevention business plan support method for verifying this CL at the required level,
Required level setting section of the computer, relative to the CL is the arithmetic, non-occurring against rainfall data when generating predictive value for rainfall data when landslides occur and / or landslides did not occur A step of reading the request level data stored in advance in the information storage device or adopting the input value from the information input device to the request level consisting of a medium rate ; and
Model selection unit of the computer, the steps of be read out by selecting an analysis model stored in advance in the information storage device,
A step analysis parameter setting unit of the computer, for inputting an analysis parameter or the analysis parameters inputted from the information input apparatus read out from the information storage device on the analysis model,
The determination boundary analysis unit of the computer, using an analysis data input from reading the analysis data from the information storage device or the information input device, short-term rainfall index and long-term rain rainfall data included in the analysis data On the grid points of the grid on the two-dimensional plane formed with the index as the vertical axis or the horizontal axis, respectively, the discriminant boundary surface is formed by plotting the record information of occurrence / non-occurrence of the earth and sand disaster. Calculating CL as a contour line corresponding to a desired threshold ;
CL accuracy verification unit of the computer, the rainfall data and landslides occurrence or non-occurrence of rainfall data inputted actual information from the read out or the information input device from the information storage device and actual sediment disaster-nonoccurrence Using the information, the incidence rate for the rainfall data when a sediment disaster occurs for the CL calculated by the discriminant boundary surface analysis unit and / or the non-occurrence for the rain data when no sediment disaster occurs Calculating a hit rate and verifying the accuracy of the CL in comparison with the required level set by the required level setting unit ;
A disaster prevention project planning support method comprising a step of outputting at least one of the analysis data, analysis parameter, analysis model, and CL by an information output device.