JP6813865B1 - Information processing method, program, information processing device and model generation method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an information processing method or the like capable of accurately predicting a water depth. SOLUTION: The information processing method obtains the water depth observed at an observation point in a unit section in which the ground surface corresponding to a calculation area is divided into a certain unit, and inputs the water depth of the observation point. The boundary condition of the unit section is estimated by inputting the acquired water depth of the observation point into the model that outputs the boundary condition regarding the water flow flowing in and out of the unit section, and the estimated boundary condition is applied to a predetermined calculation formula. Therefore, the computer executes a process of generating water depth distribution data indicating the water depth distribution in the unit section. [Selection diagram] Fig. 1

Description

The present invention relates to an information processing method, a program, an information processing device, and a model generation method.

There is a simulation technology that calculates the water depth of rivers, etc. for the prediction and verification of flood damage. For example, Patent Document 1 discloses a flooding simulation apparatus that divides topographical data of a target area into a grid pattern, applies a predetermined mathematical model to calculate the water depth of each grid, and simulates an inundation distribution.

Japanese Unexamined Patent Publication No. 2008-84243

When a simulation is performed as in the invention of Patent Document 1, it is necessary to set predetermined parameters necessary for the simulation, such as the boundary conditions of the water flow flowing into and out of the target area. However, these parameters are generally unknown, making accurate simulation difficult.

The information processing method according to one aspect is when the water depth observed at the observation points in the unit section in which the ground surface corresponding to the calculation area is divided into fixed units is acquired and the water depth of the observation points is input. The boundary condition of the unit section is estimated by inputting the acquired water depth of the observation point into the model that outputs the boundary condition regarding the water flow flowing in and out of the unit section, and the estimated boundary condition is applied to a predetermined calculation formula. Then, the computer executes a process of generating water depth distribution data indicating the water depth distribution in the unit section.

On one side, the water depth can be predicted accurately.

It is explanatory drawing which shows the configuration example of the water depth prediction system. It is a block diagram which shows the configuration example of a server. It is explanatory drawing which shows the generation process of an estimation model. It is explanatory drawing about the simulation calculation process of water depth. It is explanatory drawing about the simulation calculation process of water depth. It is explanatory drawing about a boundary condition. It is explanatory drawing about the altitude conversion processing. It is explanatory drawing about water depth correction processing. It is explanatory drawing about water depth prediction processing. It is explanatory drawing which shows the display screen example of a terminal. It is a flowchart which shows an example of the processing procedure which a server executes. It is a flowchart which shows the processing procedure of the subroutine of model generation. It is a flowchart which shows the processing procedure of the subroutine of water depth prediction. It is explanatory drawing about the altitude conversion processing and water depth conversion processing which concerns on Embodiment 2. It is a flowchart which shows the processing procedure of the subroutine of model generation which concerns on Embodiment 2. It is a flowchart which shows the processing procedure of the subroutine of water depth prediction which concerns on Embodiment 2. It is explanatory drawing which shows the outline of Embodiment 3. It is a flowchart which shows the processing procedure of the subroutine of model generation which concerns on Embodiment 3. It is a flowchart which shows the processing procedure of the subroutine of water depth prediction which concerns on Embodiment 3. It is a schematic diagram for demonstrating the concept of Bayesian optimization.

Hereinafter, the present invention will be described in detail with reference to the drawings showing the embodiments thereof.
(Embodiment 1)
FIG. 1 is an explanatory diagram showing a configuration example of a water depth prediction system. In this embodiment, a water depth prediction system that predicts the distribution of the water depth in the entire area from the observed values of the water depth observed in the area (calculation area) to be predicted will be described. The water depth prediction system includes an information processing device 1 and terminals 2, 2, 2, .... Each device is communicatively connected to a network N such as the Internet.

The information processing device 1 is an information processing device capable of transmitting and receiving various types of information processing and information, and is, for example, a server computer, a personal computer, or the like. In the present embodiment, it is assumed that the information processing device 1 is a server computer, and in the following, it is read as server 1 for the sake of brevity. The server 1 functions as a predictor for predicting the water depth, and distributes the water depth of the entire area from the topographical data showing the elevation of the ground surface of the target area and the water depth observed at a predetermined number of observation points in the area. Perform simulation calculations to predict.

In the present embodiment, the server 1 estimates the parameters necessary for performing the simulation calculation by using the estimation model 141 (see FIG. 3 and the like) constructed by machine learning. Specifically, the server 1 uses the estimation model 141 to estimate the boundary conditions (for example, the flow velocity) of the water flow flowing in and out of the target area from the water depths of a predetermined number of observation points. The server 1 performs a simulation calculation based on the boundary conditions estimated using the estimation model 141, and generates water depth distribution data showing the water depth distribution in the area.

The terminal 2 is a terminal device of a user who uses this system, and is, for example, a personal computer, a tablet terminal, a smartphone, or the like. The server 1 acquires the topographical data of the area to be predicted and the water depth observed in the area from the terminal 2 and generates the water depth distribution data.

FIG. 2 is a block diagram showing a configuration example of the server 1. The server 1 includes a control unit 11, a main storage unit 12, a communication unit 13, and an auxiliary storage unit 14.
The control unit 11 has one or a plurality of arithmetic processing units such as a CPU (Central Processing Unit), an MPU (Micro-Processing Unit), and a GPU (Graphics Processing Unit), and stores a program P stored in the auxiliary storage unit 14. By reading and executing, various information processing, control processing, etc. are performed. The main storage unit 12 includes SRAM (Static Random Access Memory) and DR.
It is a temporary storage area such as AM (Dynamic Random Access Memory) and flash memory, and temporarily stores data necessary for the control unit 11 to execute arithmetic processing. The communication unit 13 is a communication module for performing processing related to communication, and transmits / receives information to / from the outside.

The auxiliary storage unit 14 is a non-volatile storage area such as a large-capacity memory or a hard disk, and stores a program P and other data necessary for the control unit 11 to execute processing. Further, the auxiliary storage unit 14 stores the estimation model 141. The estimation model 141 is a trained model generated by training training data, and as described later, flows out to the unit section by inputting the water depth of a predetermined number of observation points observed in a certain unit section. This is a model that outputs the boundary conditions related to the incoming water flow. The estimation model 141 is expected to be used as a program module that functions as a part of artificial intelligence software.

The auxiliary storage unit 14 may be an external storage device connected to the server 1. Further, the server 1 may be a multi-computer composed of a plurality of computers, or may be a virtual machine virtually constructed by software.

Further, in the present embodiment, the server 1 is not limited to the above configuration, and may include, for example, an input unit that accepts operation input, a display unit that displays an image, and the like. Further, the server 1 is provided with a reading unit that reads a portable storage medium 1a such as a CD (Compact Disk) -ROM, a DVD (Digital Versatile Disc) -ROM, and reads and executes the program P from the portable storage medium 1a. You can do it. Alternatively, the server 1 may read the program P from the semiconductor memory 1b.

Further, in the present embodiment, it is assumed that the server 1 on the cloud performs the estimation based on the estimation model 141. However, the present invention is not limited to this, and the estimation model 141 generated by the server 1 may be installed in the local terminal 2 so that the terminal 2 performs the estimation based on the estimation model 141.

Hereinafter, an outline of the present embodiment will be described with reference to FIGS. 3 to 10.
As will be described later, when the server 1 receives the input of the topographical data of the area to be predicted and the water depth observed at a predetermined number of observation points in the area from the terminal 2 (see FIG. 10), the following Execute the process. First, the server 1 sets an arbitrary boundary condition as a parameter for simulation and then performs a simulation calculation based on the terrain data. As a result, the server 1 generates water depth distribution data simulating the water depth distribution in the area as a simulation result (see FIGS. 4 to 8). Then, the server 1 performs machine learning using the generated water depth distribution data and the boundary conditions used at the time of simulation as training data. As a result, the server 1 generates an estimation model 141 that outputs the boundary conditions when the water depth of the observation point is input (see FIG. 3).

Next, the server 1 inputs the water depth of the observation point into the generated estimation model 141 and estimates the boundary conditions (see FIG. 9). Then, the server 1 performs a simulation calculation based on the estimated boundary conditions and the topographical data to generate water depth distribution data. The server 1 outputs the generated water depth distribution data to an arbitrary terminal 2.

In the present embodiment, learning is performed every time prediction is performed to generate an estimation model 141, but the method is not limited to such a method. For example, an estimation model 141 in which the water depth and boundary conditions of an arbitrary area have been learned is prepared in advance, and the water depth of the observation point acquired from the terminal 2 is input to the prepared estimation model 141 to estimate the boundary conditions. May be good.

FIG. 3 is an explanatory diagram showing a generation process of the estimation model 141. In Figure 3, includes a water depth of the ground surface which corresponds to the calculated area observation point in the unit section which is divided by a fixed unit hi (x, y), the boundary conditions u _BC about water flow to and from the flow to the unit block A conceptual diagram shows how machine learning is performed using training data to generate an estimation model 141.

Before explaining the estimation model 141, first, the unit partition and the boundary conditions will be described. The unit section is a ground surface area in which the ground surface is divided into fixed units, for example, a region in which the ground surface is divided into a rectangular shape with a width of 2 km and a height of 1 km. The unit such as the horizontal width and the vertical width of the unit section is not particularly limited, and the shape thereof is not limited to a rectangular shape.

The boundary condition is a parameter relating to the water flow on the boundary surface forming the unit compartment, for example, the flow velocity when the water flow flows in and out of the boundary surface. The boundary condition is not limited to the flow velocity, and may be, for example, a water flow rate. In FIG. 3, the water flow flowing in and out of the unit section is indicated by an arrow. In the example of FIG. 3, the water flow flows in from the left side surface and the lower side surface, and the water flow flows out from the right side surface and the upper side surface.

The server 1 learns the training data and generates an estimation model 141 that estimates the boundary conditions from the water depths of a predetermined number of observation points in the unit section. The estimation model 141 is, for example, a model constructed by a neural network. The estimation model 141 has an input layer that accepts input of water depth data of an observation point, an intermediate layer (hidden layer) that extracts a feature amount from the input data, and an output layer that outputs an estimation result based on the feature amount. Each layer consists of one or more neurons, and finally outputs the estimation result of the boundary condition from the output layer. However, the estimation model 141 is not limited to the neural network, and may be a model constructed by polynomial regression or the like.

In the present embodiment, the server 1 generates an estimation model 141 that outputs boundary conditions when the water depths of a plurality of observation points (for example, 5 points) are input. The observation point may be one point. The server 1 inputs the water depths at a plurality of points extracted from the training data into the estimation model 141, and acquires the boundary conditions as an output. The server 1 compares the boundary conditions output from the estimation model 141 with the boundary conditions of the correct answer in the training data, and optimizes parameters such as weights between neurons so that the two are approximated. As a result, the server 1 generates the estimation model 141.

As described above, in the present embodiment, simulation calculation based on arbitrary boundary conditions is performed to generate water depth distribution data for training, and the generated water depth and the boundary conditions used at the time of simulation are used as training data. However, the present invention is not limited to this, and the water depth and boundary conditions used as training data may be actually observed values.

4 and 5 are explanatory views relating to the simulation calculation process of the water depth. In FIG. 4, the unit section is conceptually illustrated. FIG. 5 conceptually illustrates a cross-sectional view at a certain point in the unit compartment. The details of the simulation calculation will be described below.

The server 1 performs a simulation of calculating the water depth at each point in the unit section using the topographical data indicating the elevation of the target unit section. Topographical data is data in which elevation values are associated with position information (for example, latitude and longitude) of each point, for example, DSM (Digital Surface Model), DEM (Digital Elevation Model; numerical elevation model). ), 3D map and other map data. The topographical data may be data indicating the altitude of each point in the unit section, and may be data in another format different from DSM or the like in the surveying method or the like.

The topographical data includes the first topographical data that includes features such as buildings and trees existing on the ground surface and is regarded as the ground surface, and the second topographical data that excludes features such as buildings and trees from the first topographical data. There is data. In the above example, the DSM and 3D map correspond to the first terrain data and the DEM corresponds to the second terrain data. In the simulation for calculating the water depth, either the first topographical data or the second topographical data can be used, but in the present embodiment, the DSM data which is the first topographical data is used. In the cross-sectional view of FIG. 5, the solid line represents the ground surface, the dotted line represents the water surface, and the alternate long and short dash line represents the reference plane that serves as the reference for altitude. The convex part corresponds to a feature such as a building.

The specific simulation method is not particularly limited, but in the present embodiment, the calculation is performed using the shallow water equation, which is a kind of fluid calculation method. The shallow water equation is an equation that performs fluid calculation by approximating three-dimensional fluid equations (Navier-Stokes equations) to two dimensions. Specifically, in the shallow water equation, the flow velocity in the direction perpendicular to the ground surface is ignored, and the fluid calculation is performed considering only the direction parallel to the ground surface.

The server 1 discretizes the unit partition for each fixed area and performs the calculation. For example, as shown in FIG. 4, the server 1 divides a unit partition into a grid using the finite volume method, and puts a water depth h (scalar amount) at the center of each grid and a flow velocity u (vector amount) on the boundary between the grids. ) Is placed and the calculation is performed. The finite element method or the like may be used for discretizing the unit compartment. In the following description, for convenience, each divided grid is referred to as a "cell".

As described above, when the water depth is h, the flow velocity is u, the gravitational acceleration is g, and the altitude is h ₀ , the shallow water equation is expressed by the following equation.

Equation (1) represents the equation of motion, and equation (2) represents the continuity equation (conservation law of flow rate). Note that τ ^ω is a term representing the influence of the roughness of the ground surface, and τ ^bld is a term representing the influence of a feature (for example, building).

The server 1 applies an arbitrary boundary condition to the above equation to perform a simulation calculation of the water depth. The boundary condition used as the training data may be, for example, a random number value calculated by the server 1 using a random number generator, or a value manually input by the user. The simulation calculation calculates the boundary conditions that include the range of water depth that can actually occur.

FIG. 6 is an explanatory diagram relating to the boundary conditions. For example, as shown in FIG. 6, the server 1 has two boundary surfaces (left side surface and lower side surface) as inflow surfaces of water flow among the four boundary surfaces, and the other two boundary surfaces (right side surface and upper side surface). Is the outflow surface. For the outflow surface, the spatial gradient of physical quantities (flow velocity, water depth, etc.) is set to zero, and the outflow is free. Also, if the inflow is the same, the distribution of the overall water depth is not considered to change much, so of the two inflow surfaces, one inflow surface (left side) has a fixed flow velocity and the other inflow surface ( The flow velocity is variable only on the lower side). The surface that fixes or changes the flow velocity may be the reverse of the above. The server 1 sets the variable flow velocity of the boundary surface as a boundary condition.

In this embodiment, for the sake of simplicity, the boundary condition is set so that only one inflow surface is variable. However, the above-mentioned method of setting the boundary condition is an example, and the present embodiment is not limited to this. For example, both of the two inflow surfaces may be variable, and other boundary conditions may be set for the outflow surface. Further, the inflow surface may be one surface or three surfaces.

In addition to the above boundary conditions, the server 1 simulates the initial values of the water depth and the flow velocity in each cell as fixed values. The server 1 applies the set values such as boundary conditions and the altitude of each cell (point) indicated by the topographical data to the shallow water equation, and calculates the water depth h of each cell. As a result, the server 1 generates water depth distribution data showing the water depth distribution in the unit section.

The server 1 executes the above simulation calculation a predetermined number of times while changing the boundary conditions. As a result, the server 1 generates a predetermined number of training data which is a combination of water depth and boundary conditions. The server 1 performs learning using each of the generated training data and generates an estimation model 141.

FIG. 7 is an explanatory diagram relating to the altitude conversion process. As described above, the server 1 performs the simulation calculation of the water depth with reference to the altitude of the ground surface indicated by the topographical data. However, for example, as illustrated in FIG. 5, when a feature having a height exceeding the water depth exists, the water depth cannot be defined in the region corresponding to the feature and cannot be calculated directly. Therefore, in the present embodiment, processing is performed so that the water depth can be calculated.

Specifically, the server 1 lowers the elevation of the entire area of the calculation area to be lower than the actual altitude so that the altitude of the area corresponding to the feature (hereinafter referred to as "feature area") does not exceed the water depth. Simulation calculation of water depth is performed in the state converted to height. The server 1 then transforms the water depth according to the original elevation.

In FIG. 7, a cross-sectional view corresponding to the original altitude is shown on the left side, and a cross-sectional view after elevation conversion is shown on the right side. When the original altitude is ch _{0, DSM} (c> 1), the server 1 converts the altitude of the feature area into a fixed value h _{0, DSM} . This prevents the water surface from becoming discontinuous in the feature area. The server 1 performs a simulation calculation based on the converted elevation terrain data, and calculates the water depth h of each cell.

After calculating the water depth h, the server 1 converts the water depth h according to the original altitude. When the water depth to be finally obtained is h _{r, DSM ,} the relationship of h _{r, DSM} + ch _{0, DSM} = h + h _{0, DSM} holds, so that the server 1 has h _{r, DSM} = h + h _{0, DSM −} ch _{0. Converts} water depth as _DSM . From the above, the water depth of the feature area can be calculated appropriately.

FIG. 8 is an explanatory diagram regarding the water depth correction process. As described above, the server 1 converts the altitude of the actual ground surface and then performs a simulation to calculate the water depth of each cell (point). Here, the server 1 further corrects the water depth of each cell converted according to the original altitude by taking a weighted average with the surrounding cells.

Specifically, when the water depth converted according to the original altitude is _{hi, org} (i is the cell number :), the target cell and another cell are as shown in the following formula (3). weighting factor _{w i} in accordance with the distance r _i
Is calculated, and the weighted average of each cell is taken to calculate the water depth h. Note that h _{i, org} corresponds to h _{r, DSM} in each cell.

On the other hand, the water depth correction may not be necessary at the points where the water depth correction does not correspond to the feature area. Therefore the server 1, the point of performing depth correction, rather than the feature region, and, if simulation computed water depth h _{i, org} is 0 or more, according to the following equation (4), different weighting factors w _i Is used to increase the weight. The feature area is specified by using the multi-scale feature extraction method. The server 1 performs multi-scale calculation using the position and altitude of each point indicated by the topographical data as a three-dimensional point cloud, and detects the undulations (edges, etc.) of the ground surface due to the existence of the feature.

The server 1 calculates the water depth h of each cell according to the formulas (3) and (4). When the calculated water depth h is smaller than the water depth h _{i, org} calculated by the simulation, the server 1 uses the water depth h _{i, org} calculated by the simulation as the final water depth h instead of the calculated water depth h. adopt.

As described above, the server 1, the point of performing depth correction, whether true feature region, and the water depth h _i, to change the weight coefficient w _i, depending on whether _org is equal to or greater than zero , Calculate the water depth h of each cell.

FIG. 9 is an explanatory diagram relating to the water depth prediction process. The server 1 estimates the boundary conditions of the unit section to be predicted using the estimation model 141 generated above, performs simulation calculation, and generates (predicts) water depth distribution data.

That is, the server 1 inputs the water depths of the plurality of observation points acquired from the terminal 2 into the estimation model 141, and estimates the boundary condition (flow velocity) of the water flow flowing into the unit section. The server 1 performs a simulation calculation based on the estimated boundary conditions and the topographical data acquired from the terminal 2, and generates water depth distribution data showing the water depth distribution in the unit section. The simulation calculation method is the same as during learning. The server 1 outputs the generated water depth distribution data to the terminal 2 and displays it. The terminal 2 that inputs data to the server 1 and the terminal 2 that outputs water depth distribution data may be different terminals.

FIG. 10 is an explanatory diagram showing an example of a display screen of the terminal 2. FIG. 10 illustrates an example of a screen displayed by the terminal 2 when the server 1 executes the above processing. The screen includes a terrain data input field 91, a water depth input field 92, an address input field 93, a water depth list display field 94, and a water depth distribution display field 95.

The terminal 2 accepts input of terrain data and a data file related to the water depth of the observation point in response to the operation input to the terrain data input field 91 and the water depth input field 92. Further, the terminal 2 accepts the input of the address of one or a plurality of points for displaying the water depth in the water depth list display field 94 in response to the operation input in the address input field 93. The terminal 2 outputs the data input in each input field to the server 1 and requests the processing of the water depth prediction.

The server 1 executes a series of processes according to the request from the terminal 2 and outputs the prediction result to the terminal 2. Specifically, the server 1 outputs a predicted value of the water depth of each point corresponding to the address entered in the address input field 93, and at the same time, with respect to the map (topographic data) of the unit section targeted for prediction, each point. Outputs a water depth distribution map that superimposes the prediction results of the water depth. The terminal 2 displays in the water depth list display field 94 a water depth list that lists each address entered in the address input field 93 and the water depth of the point corresponding to the address. Further, the terminal 2 displays the water depth distribution map in the water depth distribution display column 95. The water depth distribution map is a map in which the display mode of each point is changed according to the water depth, for example, a map in which the display color of each cell is changed according to the water depth. Note that, for convenience of illustration, FIG. 10 only shows how the display colors of the cells are different by hatching, and does not show the map. The user can grasp the water depth of the desired address by referring to the water depth list display field 94, and can grasp the water depth of the entire target area (unit section) in the water depth distribution display field 95.

For example, the terminal 2 may superimpose a point corresponding to the address specified in the address input field 93 on the water depth distribution map (for example, display an icon). As a result, the user can get a bird's-eye view of the depth of water around the desired address.

FIG. 11 is a flowchart showing an example of a processing procedure executed by the server 1. The processing contents executed by the server 1 will be described with reference to FIG.
The server 1 acquires from the terminal 2 the water depths of a predetermined number of observation points in the unit section in which the ground surface is divided into fixed units and the topographical data indicating the elevation of the ground surface in the unit section (step S1). ..

Next, the server 1 executes a model generation subroutine (step S2). Specifically, the server 1 performs a simulation calculation based on the acquired topographical data, generates water depth distribution data simulating the distribution of the water depth in the unit section, and calculates the generated water depth and the boundary conditions used at the time of the simulation. Machine learning as training data is performed to generate an estimation model 141.

Next, the server 1 executes a subroutine for predicting the water depth (step S3). Specifically, the server 1 inputs the water depth of the observation point into the generated estimation model 141, estimates the boundary conditions, performs simulation calculations based on the estimated boundary conditions and the topographical data, and generates water depth distribution data. ..

Then, the server 1 outputs the generated water depth distribution data to the terminal 2 and ends a series of processes (step S4).

FIG. 12 is a flowchart showing a processing procedure of the “model generation subroutine” in step S2.
The control unit 11 of the server 1 identifies the feature area where the feature exists from the terrain data acquired from the terminal 2 (step S11). The control unit 11 converts the elevation of the entire region of the calculation region into a fixed value so that the elevation of the feature region does not exceed the water depth (step S12).

The control unit 11 sets the boundary conditions of the unit section corresponding to the terrain data (step S13). The boundary condition may be a random value or a value set and input by the user. The control unit 11 performs a simulation calculation based on the set boundary conditions and the topographical data whose altitude is converted in step S12, and generates water depth distribution data simulating the distribution of water depth in the unit section (step S14). At this time, the control unit 11 converts the water depth according to the altitude before conversion.

The control unit 11 corrects the generated water depth distribution data (step S15). Specifically, the control unit 11 uses different weighting coefficients depending on whether or not the point where the water depth correction is performed corresponds to the feature area, and sets the water depth at each point (cell) to the water depth at the surrounding point. Take a weighted average and correct.

The control unit 11 determines whether or not the number of simulations has reached a predetermined number (step S16). When it is determined that the predetermined number of times has not been reached (S16: NO), the control unit 11 returns the process to step S13. In this case, the control unit 11 sets the boundary condition again (step S13) and generates the water depth distribution data (step S14). As a result, a predetermined number of training data consisting of water depth and boundary conditions are generated.

When it is determined that the predetermined number of times has been reached (S16: YES), the control unit 11 extracts the water depths of a predetermined number of points corresponding to the observation points from one training data (step S17). The control unit 11 inputs the water depth at each extracted point into the estimation model 141 (step S18). The control unit 11 compares the boundary conditions output from the estimation model 141 with the boundary conditions of the correct answer in the training data, and updates the parameters (weights between neurons, etc.) of the estimation model 141 (step S19).

The control unit 11 determines whether or not the training of the training data is sufficient (step S20). When it is determined that the learning is not sufficient (S20: NO), the control unit 11 returns the process to step S17. In this case, the control unit 11 extracts the water depths at a predetermined number of points from other training data, inputs them to the estimation model 141, and updates the parameters of the estimation model 141. When it is determined that the learning is sufficient (S20: YES), the control unit 11 ends the subroutine. Whether or not the learning is sufficient is determined by using the loss function.

FIG. 13 is a flowchart showing a processing procedure of the “subroutine for predicting water depth” in step S3.
The control unit 11 of the server 1 inputs the water depths of a predetermined number of observation points acquired from the terminal 2 into the estimation model 141 (step S31). Then, the control unit 11 estimates the boundary conditions of the water flow flowing in and out of the unit section based on the estimation model 141 (step S32).

The control unit 11 performs a simulation calculation based on the estimated boundary conditions and the topographical data whose altitude is converted in step S12, and generates water depth distribution data indicating the distribution of water depth in the unit section (step S33). The control unit 11 corrects the water depth at each point indicated by the water depth distribution data (step S34). The control unit 11 ends the subroutine.

In the above, only the water depth is input as the input to the estimation model 141, but other parameters such as the amount of rainfall at the time of observation may also be used as the input.

From the above, according to the first embodiment, the simulation accuracy of the water depth prediction can be improved by estimating the boundary conditions of the water flow flowing in and out of the unit section using the estimation model 141.

Further, according to the first embodiment, when handling the first topographical data including the feature, the simulation can be preferably performed by converting the altitude of the calculation area to a sufficiently low value.

Further, according to the first embodiment, the water depth of the feature region can be suitably calculated by converting the water depth of the calculation region calculated by the simulation into the original altitude.

Further, according to the first embodiment, the water depth of each point (for example, the weighted average with the water depth of the surrounding points) is used by using different weighting factors depending on whether or not the point where the water depth correction is performed corresponds to the feature area. ), It is possible to more preferably predict the water depth in consideration of the features.

(Embodiment 2)
In the present embodiment, a mode for predicting the water depth by combining the first topographical data including the feature and the second topographical data from which the feature is removed will be described. In each of the following embodiments, the same reference numerals are given to the contents overlapping with the first embodiment, and the description thereof will be omitted.

FIG. 14 is an explanatory diagram relating to the altitude conversion process and the water depth conversion process according to the second embodiment. In FIG. 14, for the same point in the unit section, the cross-sectional view (left side) corresponding to the first topographical data, the cross-sectional view (center) obtained by converting the elevation of the feature area of the first topographical data, and the second topographical data. The corresponding cross-sectional views (right side) are shown respectively. An outline of the present embodiment will be described with reference to FIG.

In the present embodiment, the server 1 performs a simulation calculation using the first topographical data showing the elevation of the ground surface including the features and the second topographical data showing the elevation of the ground surface excluding the features. For example, the server 1 uses DSM data as the first terrain data and DEM data as the second terrain data. For example, the server 1 performs a simulation calculation based on the first terrain data and predicts the water depth based on the second terrain data.

Similar to the first embodiment, the server 1 converts the water depth of the feature area according to the altitude of the original feature area. Here, the server 1 converts the water depth of the feature area based on the difference between the altitude of the feature area in the first terrain data and the altitude of the feature area in the second terrain data.

Ideally, the first topographical data and the second topographical data have the same elevation of the ground surface excluding the features, but they may actually be different. In the example of FIG. 14, the second terrain data is slightly lower than the first terrain data. In this state, if the water depth calculated from the second topographical data is simply converted by the altitude of the feature region indicated by the first topographical data, the prediction accuracy may decrease due to the relative deviation.

Therefore, the server 1 converts the water depth of the feature region based on the difference in elevation between the first terrain data and the second terrain data. For example, as shown in FIG. 14, the calculated water depth in the first terrain data obtained by converting the altitude is h _, the elevation of the first terrain data after the elevation conversion is h _{0, DSM} , and the water depth before the elevation conversion based on the first terrain data. _Is hr _{, DSM} , the original elevation of the first terrain data is ch _{0, DSM} , the water depth in the second terrain data is hr _{, DEM} , the elevation is ch _{0, DEM,} and the first terrain data and the second terrain data. Let δh _{DEM, DSM} be the difference in elevation. In this case, _since the relationship of h _{r, DSM} + ch _{0, DSM} = h _{r, DEM} + δh _{DEM, DSM} + ch _{0, DEM} holds, h _{r, DEM} = h _{r, DSM} + ch _{0, DSM} − ch _{0, DEM −} δ h It becomes _{DEM and DSM} . Thereby, the water depth distribution data can be suitably generated by combining the first topographic data and the second topographic data while considering the relative elevation deviations.

FIG. 15 is a flowchart showing a processing procedure of the “model generation subroutine” according to the second embodiment. In the following description, for convenience, the server 1 will be described as having acquired the first terrain data and the second terrain data from the terminal 2.
The control unit 11 of the server 1 converts the elevation of the entire area of the calculation area in the first terrain data (step S201), and shifts the process to step S13.

After setting the boundary conditions (step S13), the control unit 11 performs a simulation calculation using the first terrain data and the second terrain data converted in step S201 based on the set boundary conditions, and performs simulation calculation, and the water depth distribution data. Is generated (step S202). Specifically, as described above, the control unit 11 performs a simulation calculation using the first topographical data and generates water depth distribution data. The control unit 11 corrects the water depth using the second terrain data with respect to the predicted value of the water depth simulated and calculated using the first terrain data, and generates the water depth distribution data.

The control unit 11 converts the water depth of the feature area in the generated water depth distribution data based on the difference between the altitude of the feature area before conversion in the first terrain data and the altitude of the feature area in the second terrain data. (Step S203). Specifically, as described above, the control unit 11 adjusts the difference in elevation between the first terrain data and the second terrain data to obtain the water depth of the final feature region. The control unit 11 shifts the process to step S16.

FIG. 16 is a flowchart showing a processing procedure of the “subroutine for predicting water depth” according to the second embodiment. After inputting the water depth of the observation point into the estimation model 141 and estimating the boundary conditions (step S32), the server 1 executes the following processing.
Based on the estimated boundary conditions, the control unit 11 of the server 1 generates water depth distribution data using the first terrain data obtained by converting the elevations of the entire region in step S201 and the second terrain data (step S221). Specifically, as in the case of learning, the control unit 11 converts the water depth using the second terrain data with respect to the water depth calculated by the simulation using the first terrain data. That is, the control unit 11 converts the water depth of the feature region based on the difference between the elevation of the original feature region in the first terrain data and the elevation of the feature region in the second terrain data (step S222). .. The control unit 11 ends the subroutine.

Based on the above, according to the second embodiment, the water depth is predicted by combining the first topographical data showing the elevation of the ground surface including the features and the second topographical data showing the elevation of the ground surface excluding the features. You can also do it.

Further, according to the second embodiment, the water depth can be predicted more accurately by converting the water depth based on the difference in elevation between the first terrain data and the second terrain data.

(Embodiment 3)
In the present embodiment, a mode in which the boundary conditions are estimated a plurality of times and the simulation calculation is performed from each of the estimated boundary conditions will be described.

FIG. 17 is an explanatory diagram showing an outline of the third embodiment. In FIG. 17, among the observation points of the first number of points (for example, 5 points), the water depths of the observation points of the second points (for example, 3 points) extracted in different combinations are input to the estimation model 141, and the boundary conditions are set. The state of estimating multiple times is shown. An outline of the present embodiment will be described with reference to FIG.

In the present embodiment, when the server 1 acquires the water depth of the number of observation points of the first point from the terminal 2, it extracts the water depth of the number of second points less than the number of first points and inputs it to the estimation model 141. Specifically, the server 1 extracts the water depths of the number of second points of a plurality of patterns so that the combination of observation points is different. For example, when the number of the first points is 5 and the number of the second points is 3, the water depth of the number of the second points is extracted by 10 combinations.

For example, during training, the server 1 similarly extracts the water depths of the number of second points from the training water depth distribution data generated by the simulation calculation in different combinations, and trains the water depths of each pattern to train a plurality of estimation models 141. Is generated. The server 1 inputs the water depths of the number of second points of each pattern extracted above into different estimation models 141, and estimates the boundary conditions.

The server 1 generates water depth distribution data based on the boundary conditions output from each estimation model 141. Specifically, the server 1 calculates the average value of the boundary conditions output from each estimation model 141, performs simulation calculation based on the average value, and generates water depth distribution data.

In this embodiment, the average value of the boundary conditions output from each estimation model 141 is used. For example, the server 1 uses the boundary conditions output from each estimation model 141 as they are to perform a plurality of simulation calculations. It may be executed several times to generate a plurality of water depth distribution data. In this case, for example, the average value of each water depth distribution data is taken to obtain the final water depth distribution data. As described above, the server 1 may estimate a plurality of boundary conditions and generate water depth distribution data using the estimated plurality of boundary conditions, and the configuration in which the average value of the boundary conditions is used for the simulation is not essential.

FIG. 18 is a flowchart showing a processing procedure of the “model generation subroutine” according to the third embodiment. If YES in step S16, the server 1 executes the following processing.
The control unit 11 of the server 1 extracts the water depth of the number of first points corresponding to one boundary condition from the training data (step S301). Further, the control unit 11 extracts the water depth of the second point from the extracted water depth of the first point (step S302). The control unit 11 inputs the water depth of the number of extracted second points into one estimation model 141 (step S303). The control unit 11 compares the boundary conditions output from the estimation model 141 with the boundary conditions of the correct answer in the training data, and updates the parameters of the estimation model 141 (step S304).

The control unit 11 determines whether or not all the patterns of the number of second points have been learned for the water depth of the number of first points extracted in step S301 (step S305). When it is determined that there is a pattern that has not been learned (S305: NO), the control unit 11 returns the process to step S302. In this case, the control unit 11 extracts the water depth of the second number of points by another combination (step S302), inputs the extracted water depth of the second point number into the other estimation model 141 (step S303), and performs the estimation. The parameters of the model 141 are updated (step S304).

When it is determined that the patterns of all the second points have been learned (S305: YES), the server 1 determines whether or not the training of the training data is sufficient (step S306). When it is determined that the learning is not sufficient (S306: NO), the control unit 11 returns the process to step S301. In this case, the control unit 11 changes the training data, executes the processes of steps S301 to S305, and generates a plurality of estimation models 141. When it is determined that the training of the training data is sufficient (S306: YES), the control unit 11 ends a series of processes.

FIG. 19 is a flowchart showing a processing procedure of the “subroutine for predicting water depth” according to the third embodiment. For convenience of explanation, it is assumed that the server 1 has already acquired the water depth of the observation points of the first number of points and the topographical data of the unit section to be predicted.
The control unit 11 of the server 1 extracts the water depth of the second point from the water depth of the first point and inputs it to the estimation model 141 (step S321). Then, the control unit 11 estimates the boundary condition of the unit section to be predicted (step S322).

The control unit 11 determines whether or not the boundary conditions have been estimated for all the patterns of the number of second points (step S323). When it is determined that there is a pattern for which the boundary condition has not been estimated (S323: NO), the control unit 11 returns the process to step S321. In this case, the control unit 11 extracts the water depth of the second point number by another combination. Then, the control unit 11 inputs the extracted water depth into a different estimation model 141 to estimate the boundary conditions.

When it is determined that the boundary conditions have been estimated for all the patterns (S323: YES), the control unit 11 calculates the average value of the estimated plurality of boundary conditions (step S324). The control unit 11 performs simulation calculation based on the calculated average value of the boundary conditions and the acquired topographical data, and generates water depth distribution data (step S325). Then, the control unit 11 shifts the process to step S36.

From the above, according to the third embodiment, the prediction accuracy can be improved by estimating the boundary conditions a plurality of times and using them in the simulation.

(Embodiment 4)
In this embodiment, water depth distribution data is generated using a Bayesian optimization method.
FIG. 20 is a schematic diagram for explaining the concept of Bayesian optimization. The horizontal axis of FIG. 20 shows the flow velocity _UBC which is a boundary condition, and the vertical axis shows the loss function (function to be minimized) L. The loss function L here is calculated from the difference between the water depth value simulated and calculated based on the boundary conditions on the horizontal axis and the observed water depth value. The black circles B1 to B3 in FIG. 20 mean the points (arbitrary) to be calculated at the initial stage, the solid line J1 is the average value in the regression by the Gaussian process, and the dotted line J2 shows the standard deviation.

The server 1 according to the present embodiment acquires the water depth observed at the observation points in the unit section in which the ground surface corresponding to the calculation area is divided into fixed units. Then, the sum of the square errors of the water depth value and the water depth value obtained by simulation calculation under a predetermined boundary condition is calculated and calculated as the value of the loss function L.

Next, the server 1 uses a model capable of executing Bayesian optimization, calculates the mean / variance (standard deviation) by a Gaussian process, and selects the points C1 to Cn that maximize the acquisition function, and the loss function L Search for the minimum value D. As a result, the server 1 estimates the boundary condition corresponding to the searched minimum value D. In other words, when the water depth of the observation point is input, the server 1 inputs the water depth of the acquired observation point to the model that outputs the boundary condition regarding the water flow flowing in and out of the unit section, and estimates the boundary condition of the unit section. To do.

Then, the server 1 applies the estimated boundary conditions to the shallow water equation described above to generate water depth distribution data showing the water depth distribution in the unit section.

From the above, according to the fourth embodiment, the simulation accuracy of the water depth prediction can be improved by estimating the boundary conditions of the water flow flowing in and out of the unit section using a model capable of executing Bayesian optimization. ..

In the above description, one variable with only the boundary conditions is estimated, but it is also possible to estimate a plurality of variables.

Also in the fourth embodiment, the shallow water equation reflects the topographical data indicating the altitude in the unit section. Therefore, the discussion of water depth conversion and water depth correction described in the first to third embodiments is established.

The embodiments disclosed this time should be considered to be exemplary in all respects and not restrictive. The scope of the present invention is indicated by the scope of claims, not the above-mentioned meaning, and is intended to include all modifications within the meaning and scope equivalent to the scope of claims.

1 Server (information processing device)
11 Control unit 12 Main storage unit 13 Communication unit 14 Auxiliary storage unit 141 Estimated model 2 Terminal P program

Claims (16)

Obtain the water depth observed at the observation points in the unit section where the ground surface corresponding to the calculation area is divided into fixed units.
When the water depth of the observation point is input, the boundary condition of the unit section is estimated by inputting the acquired water depth of the observation point into the model that outputs the boundary condition regarding the water flow flowing in and out of the unit section.
An information processing method in which a computer executes a process of applying the estimated boundary conditions to a predetermined arithmetic expression to generate water depth distribution data indicating the water depth distribution in the unit section. The calculation formula reflects topographical data indicating the altitude within the unit section.
The information processing method according to claim 1. The topographical data includes first topographical data indicating the elevation of the ground surface including the feature.
The first terrain data is converted so that the feature in the calculation area does not exceed the water depth, and the water depth distribution data is generated by using the calculation formula that reflects the converted first terrain data. Item 2. The information processing method according to item 2. The information processing method according to claim 3, wherein the water depth distribution data is converted according to the altitude of the ground surface before conversion. The method according to any one of claims 3 or 4, wherein the water depth at each point in the water depth distribution data is corrected by using different weighting factors depending on whether or not the point corresponds to the area corresponding to the feature. Information processing method. The terrain data further includes a second terrain data indicating the elevation of the ground surface excluding features.
The information processing method according to any one of claims 3 to 5, wherein the water depth distribution data is corrected based on the second topographical data. Obtain the water depth of the observation point of the first point,
From the water depth of the first point, a plurality of patterns of the water depth of the second point are extracted in different combinations.
The water depth of the second point number of each pattern is input to the model, and a plurality of the boundary conditions are estimated.
The information processing method according to any one of claims 1 to 6, which generates the water depth distribution data based on the plurality of boundary conditions. Calculate the average value of the plurality of boundary conditions
The information processing method according to claim 7, wherein the water depth distribution data is generated based on the calculated average value. Of the water depths of the first number of points in the unit section, the water depths of the second number of points of each pattern are input to the plurality of models to estimate the plurality of boundary conditions. The information processing method according to 8. The model is realized by neural network, polynomial regression, or Bayesian optimization.
The information processing method according to any one of claims 1 to 9. The arithmetic expression is a shallow water equation.
The information processing method according to any one of claims 1 to 10. Obtain the water depth observed at the observation points in the unit section where the ground surface corresponding to the calculation area is divided into fixed units.
When the water depth of the observation point is input, the boundary condition of the unit section is estimated by inputting the acquired water depth of the observation point into the model that outputs the boundary condition regarding the water flow flowing in and out of the unit section.
A program that causes a computer to execute a process of applying the estimated boundary conditions to a predetermined arithmetic expression to generate water depth distribution data indicating the water depth distribution in the unit section. An acquisition unit that acquires the water depth observed at the observation points in the unit section that divides the ground surface corresponding to the calculation area into fixed units, and
An estimation unit that estimates the boundary conditions of the unit section by inputting the acquired water depth of the observation point into a model that outputs the boundary conditions related to the water flow flowing in and out of the unit section when the water depth of the observation point is input. An information processing device including a generation unit that applies the boundary conditions estimated to be, to a predetermined calculation formula, and generates water depth distribution data indicating the water depth distribution in the unit section. The training data including the water depth at the point in the unit section in which the ground surface corresponding to the calculation area is divided into fixed units and the boundary condition regarding the water flow flowing in and out of the unit section is acquired.
A model generation method in which a computer executes a process of generating a model that outputs a boundary condition of the unit section when the water depth observed at an observation point in the unit section is input based on the training data. Obtain topographical data indicating the altitude within the unit section,
Set the boundary conditions of the unit section corresponding to the terrain data,
The training data is generated by performing a simulation for predicting the water depth in the unit section based on the topographical data and the boundary conditions.
The model generation method according to claim 14, wherein the model is generated based on the generated training data. By applying the topographical data and boundary conditions to a predetermined calculation formula, water depth distribution data showing the distribution of water depth within the unit section is generated.
From the generated water depth distribution data, the water depths at a predetermined number of points are extracted.
The model generation method according to claim 15, wherein the model is generated using the extracted water depth at each point and the boundary conditions as the training data.

Images