Flood prediction method and flood prediction system

Abstract

With this flood prediction method, a computer system 1 creates an observed inundation map 64 on the basis of observation data before and after past flood events in a given region, creates a simulated inundation map 65 by performing physical flood simulation using estimated precipitation data, estimated runoff data, and river topography data related to the flood events as input data, generates a neural network model for predicting an inundation area on the basis of elevation data 62 and precipitation data 63 by machine learning using the photographed image inundation map and the simulated inundation map as training data, and creates a prediction flood map 60 for enabling identification of an inundation area in a flood prediction target area through input of elevation data and precipitation data related to the flood prediction target area into the trained neural network model.

Description

Flood forecasting method and flood forecasting system

[0001] The present invention relates to a flood forecasting method and a flood forecasting system.

[0002] Floods are a natural disaster that has a serious impact on the socio-economy, and in recent years, the damage caused by them has become even more severe due to the effects of climate change. In order to minimize damage caused by floods, advance preparation and rapid response in the event of an emergency are necessary. In order to properly carry out preparations and immediate response, it is necessary to accurately grasp rainfall conditions, flood occurrence, and flooded areas due to floods, as well as to predict flooded areas in advance.

[0003] Hazard maps are a well-known method for predicting flood risk over the long term. Hazard maps are created based on flood physics simulation models that are based on land surface physical processes. There is also a method for identifying areas where floods and inundation will occur using satellite images. The following non-patent documents 1 and 2 describe simulation techniques that can predict inundation areas using topographical information, precipitation, runoff, and other input information, while the following non-patent document 3 describes methods for obtaining satellite images.

[0004] Furthermore, the following Non-Patent Documents 4, 5, and 6 describe near-real-time precipitation, elevation data, and river topography data from all over the country that can be used to predict various disasters.

[0005] Global Hydrology Group, University of Tokyo, "CaMa-Flood: Global River Hydrodynamics Model," [online], [Retrieved July 12, 2023], Internet <URL: https: / / hydro.iis.u-tokyo.ac.jp / ~yamadai / cama-flood / > Japan Agency for Marine-Earth Science and Technology, "Development of an Earth System Model for Predicting / Reproducing Climate Change," [online], [Retrieved July 12, 2023], Internet <URL: https: / / www.jamstec.go.jp / tougou / event / sympo / 2019 / images / 3_yoshimura_m.pdf> THE EUROPEAN SPACE AGENCY, "Sentinel-1 SAR User Guide”, [online], [searched July 12, 2023], Internet <URL: https: / / sentinel.esa.int / web / sentinel / user-guides / sentinel-1-sar / > Japan Meteorological Agency, “Analyzed Rainfall”, [online], [searched July 12, 2023], Internet <URL: https: / / www.jma.go.jp / jma / kishou / know / kurashi / kaiseki.html> Institute of Industrial Science, The University of Tokyo, “MERIT DEM: Improved topographic DEM with multiple errors removed”, [online], [searched August 4, 2023], Internet <URL: http: / / hydro.iis.u-tokyo.ac.jp / ~yamadai / MERIT_DEM / > Institute of Industrial Science, The University of Tokyo, “MERIT Hydro: Global Hydrographic Dataset,” [online], [Retrieved August 4, 2023], Internet <URL: http: / / hydro.iis.u-tokyo.ac.jp / ~yamadai / MERIT_Hydro / >

[0006] Hazard maps are an effective means of estimating future risks, but near-real-time flood estimation and short-term forecasting pose challenges in terms of computational efficiency and input data preparation. In other words, current methods for creating hazard maps are difficult to apply to applications that require repeated flood forecasts in response to constantly changing precipitation conditions. Therefore, research and proposals have been made to speed up flood physics simulation calculations so that near-real-time flood forecasts can be performed. However, at present, no such speed-up methods have been put into practical use. Furthermore, even if flood physics simulation calculations are accelerated in the future, the current simulation methods themselves result in a discrepancy between predictions and reality. Therefore, new methods are needed to correct the error between simulation predictions and reality.

[0007] On the other hand, satellite images can detect flooded areas with high accuracy, but it takes time to obtain the image data. Furthermore, identifying flooded areas from satellite images requires expert knowledge. Therefore, it is difficult to identify flooded areas in near real time based on satellite images and then quickly issue warnings, etc. One concept is to use a "satellite constellation" to link multiple satellites and operate them in an integrated manner to obtain satellite images in near real time, but of course, it takes a great deal of time and money to realize this concept. Furthermore, at present, the technology to predict floods from image data obtained by satellite constellations is still in the research stage.

[0008] As mentioned above, of the meteorological information that can be obtained in near real time, precipitation is the only meteorological information currently available that can be used to predict flood inundation areas. Furthermore, predicting floods caused by events such as typhoons and heavy rain, and the areas that will be inundated by such floods, using only precipitation is still in the research stage and has not yet been put to practical use. However, given the increasing severity of flood damage caused by heavy rain due to climate change in recent years, it is necessary to quickly prepare for flood disasters. Therefore, there is an urgent need to develop methods for grasping flood conditions in near real time and for making highly accurate short-term flood forecasts.

[0009] Therefore, the present invention aims to provide a flood prediction method and flood prediction system that can calculate flood inundation areas with high accuracy in a short period of time based on precipitation, grasp flood conditions in quasi-real time, and make short-term flood predictions.

[0010] One aspect of the present invention for achieving the above object is a method for predicting flood inundation areas using a computer system, the computer system comprising: an observed inundation map creation step for creating an observed inundation map, which is image data that enables identification of flooded areas in an arbitrary region, based on observation data obtained by observing the conditions before and after the occurrence of a past flood event in the region from above; a simulated inundation map creation step for creating a simulated inundation map, which is image data that enables identification of flooded areas in the arbitrary region, by performing a physical flood simulation using as input data elevation data for the arbitrary region, estimated precipitation data for the flood event, estimated runoff data for the flood event, and river topography data that defines the waterway; and a model learning step for generating a neural network model for predicting flooded areas based on the elevation data and precipitation data by machine learning using the observed inundation map and the simulated inundation map as training data. This is a flood prediction method that includes a predicted inundation map creation step in which elevation data and precipitation in a flood-predicted area are input into the trained neural network model generated by the model learning step, and image data that makes it possible to identify flooded areas in the predicted area is created as a predicted inundation map.

[0011] The flood forecasting method may also be such that, in the model learning step, the number of epochs of learning using the observed inundation map as training data and the number of epochs of learning using the simulated inundation map as training data are set to a predetermined ratio. The flood forecasting method may also be such that, in the model learning step, the number of epochs of learning using the observed inundation map as training data is set to A, and the number of epochs of learning using the simulated inundation map as training data is set to B, where A:B = 20:80, or further, that the model learning step is set to A + B = 100.

[0012] The observation data may be data observed by an artificial satellite. The precipitation data input in the predicted inundation map creation step may be rainfall observed over a predetermined period of time up to the present. The flood forecasting method may also include a flood information output step of publishing the predicted inundation map on a web page on the Internet.

[0013] Another aspect of the present invention is a flood prediction system that is made up of one or more computers and that predicts areas that will be inundated by floods, and that comprises: an observed inundation map creation means that creates image data as an observed inundation map that makes it possible to identify inundated areas in an arbitrary area based on observation data obtained by observing the conditions before and after the occurrence of a past flood event in that area from the air; a simulated inundation map creation means that creates image data as a simulated inundation map that makes it possible to identify inundated areas in that area by performing a physical flood simulation using as input data elevation data for the arbitrary area, estimated precipitation data for the flood event, estimated runoff data for the flood event, and river topography data that defines the waterway; a model learning means that generates a neural network model for predicting inundated areas based on the elevation data and precipitation data by machine learning using the observed inundation map and the simulated inundation map as training data; and a predicted inundation map creation means for inputting elevation data and precipitation in a flood-predicted area into the trained neural network model generated by the model learning step, and creating a predicted inundation map as image data that makes it possible to identify flooded areas in the predicted area.

[0014] According to the present invention, a flood forecasting method and a flood forecasting system are provided that can quickly predict areas that will be inundated by floods based on the amount of precipitation. Other advantages will be apparent from the following description.

[0015] 1 is a network configuration diagram including a computer system (flood forecasting system) that implements a flood forecasting method according to an embodiment. FIG. 2 is a diagram illustrating the hardware configuration of a computer included in the flood forecasting system. FIG. 3 is a diagram illustrating the flow of information processing in the flood forecasting method according to an embodiment. FIG. 4 is a diagram illustrating a predicted inundation map created by the flood forecasting method according to an embodiment, showing an example of a predicted inundation map displayed on a user terminal. FIG. 5 is a diagram illustrating a predicted inundation map corresponding to a past flood event. FIG. 6 is a graph illustrating the relationship between accuracy and the ratio (A:B) of the number of epochs A when a simulated inundation map created by the flood forecasting method according to an embodiment is used as training data and the number of epochs B when an observed inundation map is used as training data. FIG. 7 is a graph illustrating the relationship between the ratio (A:B) and precision. FIG. 8 is a graph illustrating the relationship between the ratio (A:B) and recall. FIG. 9 is a graph illustrating the relationship between the ratio (A:B) and F1 score.

[0016] An embodiment of the present invention will be described with reference to the accompanying drawings. == ... Since elevation is a constant, the flood forecasting method according to the embodiment can estimate flood inundation areas in near real time based solely on precipitation, which is variable information. The input precipitation is not limited to the observed value of analyzed precipitation, but can also be a forecast value based on analyzed precipitation. This allows for advance prediction of flood inundation areas for areas where precipitation forecast values ​​exist. By setting the precipitation amount arbitrarily, it is also possible to quickly create an inundation map showing the relationship between precipitation and flooded areas. <Computer System> A computer system for forecasting floods using the method according to the embodiment (hereinafter sometimes referred to as a "flood forecasting system") can be configured, for example, with a single computer or multiple computers distributed over a communication network such as the Internet or a LAN.

[0017] An outline of a network configuration including a flood prediction system is shown in Figure 1. The flood prediction system 1 shown in Figure 1 is composed of three computers (10, 20, 30) that are connected to each other via the Internet 100 so that they can communicate with each other.

[0018] The Internet 100 is also connected to an external database (hereinafter sometimes referred to as the "satellite image database 40") that stores image data created based on observation data, and a computer (hereinafter sometimes referred to as the "user terminal 50") that enables users to enjoy information services related to flood predictions provided by the flood prediction system 1. The user terminal 50 is, for example, a computer equipped with a browser, such as a personal computer (PC), smartphone, or tablet terminal. Of course, the user terminal 50 may also be, for example, a dedicated information processing terminal installed in a public facility or the like.

[0019] The three computers (10, 20, 30) that make up the flood prediction system 1 basically have the same hardware configuration. Figure 2 shows an example of the hardware configuration of the three computers (10, 20, 30). As shown in Figure 2, each computer (10, 20, 30) includes a control unit 101, a main memory unit 102, a communication unit 103, an input unit 104, an output unit 105, and an external memory unit 106.

[0020] The control unit 101 is configured using, for example, a CPU (Central Processing Unit), an MPU (Micro Processing Unit), a GPU (Graphics Processing Unit), an AI (Artificial Intelligence) chip, an FPGA (Field Programmable Gate Array), a SoC (System on Chip), an ASIC (Application Specific Integrated Circuit), etc.

[0021] The main memory unit 102 is, for example, a read-only memory (ROM), a random access memory (RAM), or a non-volatile memory (NVRAM (Non Volatile RAM)), and stores programs and data.

[0022] The communication unit 103 is a wireless or wired communication module (wireless communication module, communication network adapter, USB module, etc.) and is configured to communicate with other computers (10, 20, 30, 40, 50) via a communication network, communication cable, etc.

[0023] The input unit 104 and output unit 105 are user interfaces of the flood prediction system 1. The input unit 104 is a user interface that accepts data input, and is, for example, a keyboard, mouse, touch panel, card reader, or audio input device (e.g., microphone). The output unit 105 is a user interface that outputs various information to the outside, and is, for example, a display device (liquid crystal display, organic EL panel, etc.) that displays various information, an audio output device (e.g., speaker) that outputs various information by audio, or a printer that prints on paper media.

[0024] The external storage unit 106 is, for example, a solid state drive (SSD), a hard disk drive, an optical storage medium (such as a compact disc (CD) or a digital versatile disc (DVD)), an IC card, or an SD card, and stores programs and data. Each of the three computers (10, 20, 30) reads the programs and data stored in its external storage unit 106 into the main storage unit 102 as needed, and executes various information processes described below.

[0025] Of the three computers (10, 20, 30) that make up the flood prediction system 1, computer 10 (hereinafter referred to as the "simulation device 10") performs flood physical simulations based on the techniques described in Non-Patent Documents 1 and 2, for example, to create simulated inundation maps. Computer 20 (hereinafter referred to as the "model generation device 20") trains a model using simulated inundation maps and observed inundation maps as training data. Computer 30 (hereinafter referred to as the "flood area prediction device 30") creates observed inundation maps based on observation data acquired from the satellite image database 40 and predicts flood-prone areas based on the trained model. The inundation area prediction device 30 also functions as a web server, providing predicted flood information to users of user terminals 50 via a web page.

[0026] In this embodiment, the external storage unit 106 of the flood area prediction device 30 stores various data necessary for model generation. The flood area prediction device 30 also performs preprocessing to convert the data stored in the external storage unit 106 and satellite image data obtained from the satellite image database 40 into data suitable for model training and generation. Of course, the data stored in the external storage unit 106 of the flood area prediction device 30 may also be obtained from an external database. The model generation device 20 is a so-called "GPU server" equipped with a GPU. The following describes the various data used in the flood prediction method of this embodiment and the information processing flow in the computers (10, 20, 30) that make up the flood prediction system 1. <Input Information> The main information processing steps involved in the flood prediction method of this embodiment are flood physics simulation, observation data analysis, model training and generation, and flood (flood area) prediction using the model. Table 1 below lists examples of data used in the various information processing steps described above.

[0027]

[0028] In Table 1, the satellite image data is observation data obtained by a synthetic aperture radar (SAR) mounted on an artificial satellite such as "Sentinel-1" described in Non-Patent Document 3. Specifically, it is microwave image data that associates a position on the Earth expressed by latitude and longitude with the reception intensity of C-band (6 GHz band) microwaves reflected at that position. Note that while observation data obtained by an artificial satellite is used in this embodiment, data obtained by observation from an aircraft such as a drone may also be used.

[0029] The estimated precipitation data is, for example, an estimated value of precipitation calculated by "ERA5 (ECMWF Reanalysis v5)," a numerical forecast model provided by the European Centre for Medium-Range Weather Forecasts (ECMWF). Similarly, the estimated runoff data is an estimated value of runoff calculated by ERA5, also provided by ECMWF. Using ERA5, it is possible to estimate precipitation and runoff for any region and any period in the past.

[0030] The simulated inundation map is an inundation map created by flood physical simulation using a global river model (Catchment-based Macro-scale Floodplain: CaMa-Flood) as described in Non-Patent Documents 1 and 2. The observed inundation map is an inundation map created based on satellite image data. The elevation data is the elevation values ​​of various locations as published in the "Multi-Error-Removed Improved-Terrain Digital Elevation Models (MERIT-DEM)" described in Non-Patent Document 5. The river topography data is a hydrographic map based on the latest topographical dataset, and is data showing river channels and catchments on land on Earth. Note that the "Multi-Error-Removed Improved-Terrain global Hydrography map (MERIT-Hydro)" described in Non-Patent Document 6 can be used as the river topography data.

[0031] Water mask data is data used to identify river basins on Earth (global water area data). Here, we used original water mask data created based on OpenStreetMap data. As is well known, OpenStreetMap is a collaborative project to create a world map that is freely available and editable.

[0032] The analyzed rainfall data is near-real-time precipitation data (analyzed rainfall) across the country, measured within a specified time period in the target area for flood prediction. For example, the analyzed rainfall data collected by the Automated Meteorological Data Acquisition System (AMeDAS) published by the Japan Meteorological Agency can be used. The analyzed rainfall data from AMeDAS combines weather radar observation data with rain gauge data from across the country to analyze precipitation distribution every hour (30 minutes for the flash version) at high resolution over a 1-km square area. Non-Patent Document 4, cited above, provides an explanation of the analyzed rainfall data and how to use the published analyzed rainfall data. <Information Processing> The information processing in the flood prediction method according to the embodiment mainly includes pre-processing, which involves preparing the learning data necessary for training and generating a model; model generation processing, which involves training and generating a model; and flood prediction processing, which predicts the flood inundation level based on the generated model. The model generation processing includes the first and second learning stages.

[0033] Figure 3 shows the flow of information processing in the flood prediction method according to the embodiment. Figure 3 also illustrates the information processing performed by the computers (10, 20, 30) and user terminal 50 that make up the flood prediction system 1, as well as the communication procedures between these computers (10, 20, 30, 50). ==Pre-Processing== <Pre-Processing by the Flooded Area Prediction Device 30> As shown in Figure 3, the flooded area prediction device 30 accesses the satellite image database 40, searches for areas where actual floods have occurred in the past (hereinafter referred to as "flood events"), and downloads two sets of satellite image data taken at the closest times before and after the flood event (s1, s2). Next, the flooded area prediction device 30 processes the satellite image data before and after the flood event obtained from the satellite image database 40 (s3). Specifically, the flooded area prediction device 30 converts the two sets of satellite image data before and after the flood event into binary image data that become water mask data using a threshold method, and then applies majority filtering to the water mask data for smoothing. This unifies areas of water that were separated by noise in the binary image data.The flooded area prediction device 30 then calculates the pixels flooded by the flood event from the difference between the two satellite image data that have been smoothed, and creates image data of an inundation map in which flooded areas are clearly indicated in the satellite image as an observed inundation map as shown in Table 1 (s4).The flooded area prediction device 30 then transmits this observed inundation map, along with precipitation data, elevation data, and water mask data, to the model generation device 20 (s5).

[0034] The observed inundation map, precipitation data, elevation data, and water mask data all have different resolutions. Therefore, the flooded area prediction device 30 performs the following preprocessing on these data before transmitting them to the model generation device 20. First, the flooded area prediction device 30 maps the corresponding elevation data to the area corresponding to the satellite image data, and aligns the spatial coordinates of the observed inundation map (which serves as training data) with those of the precipitation data, elevation data, and water mask data. The flooded area prediction device 30 then standardizes the resolution of each piece of data to the resolution of the satellite image data (10 m) using linear interpolation. Furthermore, to allow the GPU of the model generation device 20 to efficiently process each piece of data converted to the standardized resolution, the flooded area prediction device 30 cuts out each piece of image data into patches of 128 x 128 pixels. To verify the learning results, the flooded area prediction device 30 then divides the data into training data and test data, with 80% and 20% of the data being training data. <Pre-processing by simulation device 10> Meanwhile, the simulation device 10 creates simulated inundation maps for various locations by performing flood physical simulations such as those described in Non-Patent Documents 1 and 2 (s11 to s13). Specifically, the simulation device 10 creates simulated inundation maps by inputting precipitation data, runoff data, and river topography data required to cause floods into the global river model. The precipitation data and runoff data input into the global river model are, for example, estimated precipitation and runoff for a predetermined period of time (e.g., three hours) in the past up to the date and time of the satellite image data corresponding to the flood event.

[0035] To create a simulated inundation map, the simulation device 10 first creates inundation depth data by flood physics simulation based on the above input data (precipitation data, runoff data, and river topography data). Next, the simulation device 10 downscales the inundation depth data using the above elevation data and river topography data to create high-resolution simulated inundation map data for each time point.

[0036] Regarding downscaling, in the flood physics simulation using the global river model, the inundation depth is calculated for each catchment area, and since each catchment area has an area of ​​several square kilometers or more, it is necessary to determine whether or not there will be inundation within that catchment area. Therefore, in this embodiment, the simulation device 10 uses high-resolution elevation data to assume that the elevation difference within each catchment area is the water depth difference, and downscales the coarse-resolution water depth data obtained in the flood physics simulation to the resolution of the elevation data.

[0037] The simulation device 10 then performs the same preprocessing on the created simulation inundation map as the inundation area prediction device 30, and then transmits this simulation inundation map to the model generation device 20 (s14). ==Model generation process== The model generation device 20 acquires the simulation inundation map transmitted from the simulation device 10, and the observed inundation map, precipitation data, elevation data, and water mask data transmitted from the inundation area prediction device 30 (s21).

[0038] The model generation device 20 then trains the model by performing a first learning stage in which machine learning is performed using the simulated inundation map as training data, and a second learning stage in which machine learning is performed using the observed inundation map as training data (s22, s23). In the first learning stage, the physical mechanisms related to flooding are learned, and in the second learning stage, local phenomena that cannot be explained by physical laws are learned.

[0039] In each stage of machine learning, the model generation device 20 uses cross-entropy error as a loss function and optimizes model parameters by backpropagating the error between predicted data and training data. The model generation device 20 also uses stochastic gradient descent (learning rate: 0.01) as an optimization algorithm. Furthermore, the model generation device 20 performs training so that the number of epochs for the first stage of training, in which simulated inundation maps are used as training data, and the number of epochs for the second stage of training, in which observed inundation maps are used as training data, are each a predetermined ratio relative to the total number of epochs. In this embodiment, the model generation device 20 uses simulated inundation maps as training data for the first 20 epochs of the 100 epochs of training, and satellite image data as training data for the remaining 80 epochs. The ratio of the number of epochs when the simulated inundation maps and observed inundation maps are used as training data can be determined based on a preliminary experiment, as described below.

[0040] In this embodiment, the model generation device 20 creates a model using a deep learning convolutional neural network (CNN). The model architecture is based on U-Net, and uses CNN to extract features from two 128 x 128 pixel input images of precipitation data and elevation data, ultimately outputting a 128 x 128 pixel flood prediction image. As is well known, U-Net is a neural network architecture widely used in the field of deep learning, and is used for image segmentation, which identifies labels assigned to each pixel in an image. Therefore, in this embodiment, the model generation device 20 generates a predicted flood map by identifying the presence or absence of flooding for each pixel.

[0041] The model generation device 20 transmits the model generated in this manner to the flooded area prediction device 30 (s24). ==Flood prediction process== The flooded area prediction device 30 acquires this trained model from the model generation device 20 (s6).

[0042] When any precipitation data (e.g., the cumulative analyzed rainfall for the three hours prior to the prediction time) is input, the flood area prediction device 30 predicts the areas to be flooded by flooding by inputting this precipitation data and elevation data stored in the flood area prediction device 30 into a model, and outputs map image data showing the predicted flood areas (s7, s8). Furthermore, because the output map image data contains location information, the flood area prediction device 30 maps the output map image data onto map data that allows identification of place names, etc., to generate an image in which the flood area is mapped onto the map data. The image output by the model is a grayscale image of 128 x 128 pixels, with each pixel assigned a value between 0 and 1.0. The flood area prediction device 30 converts the grayscale image into a binary image representing the presence or absence of flooding using a predetermined threshold. The predicted flood map is the binary image mapped onto the map.

[0043] In this embodiment, the flood area prediction device 30 functions as a web server. The flood area prediction device 30 posts an image of a flood map (hereinafter, sometimes referred to as a "predicted flood map"), which maps flooded areas on a geographical map, on a web page identified by a specific URL, updating the image as needed (s9). In this embodiment, the flood area prediction device 30 also posts, in addition to the predicted flood map, simulation results based on a flood physics model and estimated images of flooded areas obtained through satellite observations on the web page. The flood area prediction device 30 then returns image data of the predicted flood map to a user terminal 50 that accesses the web page via the Internet 100 (s51, s10), and the user terminal 50 displays the image of the predicted flood map (s52). This allows users who view the predicted flood map 60 to prepare for disaster prevention and mitigation.

[0044] Figure 4 shows an example of a predicted inundation map 60 displayed on a user terminal 50. As shown in Figure 4, the predicted inundation map 60 is a map 70 that includes the area targeted for flood prediction (the area within the dotted rectangular frame in the figure; hereinafter, sometimes referred to as the "target area for flood prediction"). Within the target area for flood prediction, flooded areas are displayed as a binary image with darker colors than other areas. <Evaluation> Next, to evaluate the accuracy of the flood prediction method according to the embodiment, a past flood event was reproduced using a model. Figures 5A and 5B show a predicted inundation map 60 corresponding to an actual flood event that occurred in the past, as well as the precipitation data 61, elevation data 62, water mask 63, observed inundation map 64, and simulated inundation map 65 that were used to generate the predicted inundation map 60. While the actual precipitation data 61 and elevation data 62 are images in which the rectangular area targeted for flood prediction is color-coded according to the precipitation and elevation values, in Figures 5A and 5B, the colors are displayed as a single shade of color. 5A and 5B also show the Geospatial Information Authority of Japan's estimated inundation map 66 for the flood event as a true value. As is well known, the estimated inundation map 66 is a map that uses images and elevation data collected by the Geospatial Information Authority of Japan to calculate the water depth in the flooded area and expresses the depth as a shade of gray. Note that the flood event shown in Fig. 5A is the flooding of the Abukuma River caused by Typhoon Hagibis in 2019, and the flood event shown in Fig. 5B is the flooding of the Yoshida River caused by the same typhoon.

[0045] As described above, conventional flood physical simulations require input of multiple physical variables (such as runoff and water surface gradient) and are computationally expensive. However, the flood forecasting method according to the embodiment applies machine learning to input only precipitation data as a variable, enabling flood area estimation in a shorter time. Furthermore, because forecast data for precipitation data exists, future flood predictions and simulations can also be performed. Therefore, the flood forecasting method according to the embodiment is also useful for pre-emptive disaster response. <Regarding the Combination of Epochs> As described above, in the flood forecasting method according to the embodiment, the number of epochs for learning the model using simulated inundation maps as training data and the number of epochs for learning using observed inundation maps as training data were set based on the results of a prior experiment. In the experiment, the ratio between the number of epochs for learning using simulated inundation maps as training data and the number of epochs for learning using observed inundation maps as training data was varied, and the models created according to this ratio were evaluated using F1 scores.

[0046] The F1 score is used to classify imbalanced data, such as flooded areas. Imbalanced data is data in which class labels are unevenly distributed. For example, in a two-class classification problem, a dataset with very few positive examples and an overwhelming number of negative examples is imbalanced data. Relying on accuracy alone can lead to inappropriate evaluation of model performance in imbalanced data. This is because a model can achieve high accuracy by simply predicting the majority class. Therefore, the F1 score is used when accuracy alone is not sufficient and precision and recall are required. Because the F1 score is the harmonic mean of precision and recall, it is well-suited to evaluating model performance in imbalanced data. Precision indicates the percentage of samples predicted to belong to a particular class that actually belong to that class. In imbalanced data, prediction of the minority class is often particularly important, and precision is useful for evaluating this accuracy. Recall indicates the percentage of samples correctly predicted to belong to that class. With imbalanced data, it is important whether samples from minority classes are properly detected, and recall evaluates this detection ability. Regarding the balance of the evaluation, the F1 score is the harmonic mean of precision and recall, so it is possible to achieve a balance between the two. A balance between precision and recall is particularly important with imbalanced data, as it is an excellent evaluation method for improving the prediction accuracy of minority classes while maintaining overall performance. Accuracy, precision, recall, and F1 are defined by the following Equations 1 to 4.

[0047] Accuracy = (TP + TN) / (TP + TN + FP + FN) ... Equation 1 Precision = TP / (TP + FP) ... Equation 2 Recall = TP / (TP + FN) ... Equation 3 F1 = (2 × Precision × Recall) / (Precision + Recall) ... Equation 4 Note that TP, FP, FN, and TN in Equations 1 to 4 are combinations of predicted and actual positive and negative examples shown in Table 2 below.

[0048]

[0049] Figures 6A to 6D show the evaluation results of the preliminary experiment. Figures 6A to 6D are graphs showing the relationship between the ratio (A:B) of the number of epochs A when simulated inundation maps were used as training data and the number of epochs B when observed inundation maps were used as training data, out of a total of 100 epochs, and the accuracy, precision, recall, and F1 score. In this example, the ratio of the number of epochs for learning based on flood physics simulation to the number of epochs for learning based on satellite images was set to 20:80 based on the F1 score. ===Other Examples==== The above describes examples and application examples of the present invention. However, the present invention is not limited to the above and includes various modifications and applications. Furthermore, the above examples are intended to clearly explain the configuration and information processing of the flood forecasting system used in the flood forecasting method of the present invention. It is not necessarily limited to systems that include all of the described configurations and information processing. Other configurations and other information processing may be added, or parts of the configurations and information processing may be deleted or replaced with other configurations and information processing. The order of one information processing and another information processing can also be reversed.

[0050] For example, in the flood prediction method according to the above-described embodiment, a model is generated using a simulated inundation map based on a physical flood simulation and an observed inundation map based on satellite image data as training data. Of course, the observed inundation map can be replaced with an inundation map based on images of the ground surface taken by an aircraft or drone.

[0051] Furthermore, the various data used in the physical flood simulation are not limited to those shown in Table 1. For example, the estimated precipitation data and estimated runoff data do not have to be based on ERA5. Data sets owned by organizations other than ECMWF may be used as long as they can estimate the precipitation and runoff amounts of past flood events.

[0052] 1 Flood prediction system, 10 Simulation device, 20 Model generation device, 30 Flood area prediction device, 40 Satellite image database, 50 User terminal, 60 Forecasted inundation map, 61 Precipitation data, 62 Elevation data, 63 Water mask, 64 Observed inundation map, 65 Simulated inundation map, 66 True value (estimated inundation map), 100 Internet, 101 Control unit, 102 Main memory unit, 103 Communication unit, 104 Input unit, 105 Output unit, 106 External memory unit

Claims

1. A method for predicting flood inundation areas using a computer system, the method comprising: an observed flood map creation step of creating, as an observed flood map, image data that enables identification of flood inundation areas in an arbitrary region based on observation data obtained by observing the state before and after a past flood event in the arbitrary region from above; a simulated flood map creation step of creating, as a simulated flood map, image data that enables identification of flood inundation areas in the region by physical flood simulation using elevation data for the arbitrary region, estimated precipitation data in the flood event, estimated runoff data in the flood event, and river terrain data defining a watercourse as input data; a model learning step of generating, by machine learning with the observed flood map and the simulated flood map as teacher data, a neural network model for predicting flood inundation areas based on elevation data and precipitation data; and a predicted flood map creation step of inputting elevation data and precipitation in the flood prediction target region into the learned neural network model generated in the model learning step and creating, as a predicted flood map, image data that enables identification of flood inundation areas in the prediction target region.

2. The flood prediction method according to claim 1, wherein in the model learning step, the number of epochs of learning using the observed flood map as teacher data and the number of epochs of learning using the simulated flood map as teacher data are set to a predetermined ratio.

3. The flood prediction method according to claim 2, wherein in the model learning step, the number of epochs of learning using the observed flood map as teacher data is set as A, and the number of epochs of learning using the simulated flood map as teacher data is set as B, and A:B = 20:

80.

4. The flood prediction method according to claim 3, wherein in the model learning step, A + B = 100 is set.

5. The flood prediction method according to claim 1, wherein the observation data is data observed by an artificial satellite.

6. The flood prediction method according to claim 1, wherein the precipitation data input in the predicted flood map creation step is rainfall observed in a predetermined past time up to the present time.

7. The flood prediction method according to any one of claims 1 to 6, which executes a flood information output step of publishing the predicted inundation map on a web page on the Internet.

8. A flood prediction system configured by one or more computers for predicting an inundated area due to a flood, comprising: an observed inundation map creation means for creating, as an observed inundation map, image data that enables identification of an inundated area in a region based on observed data obtained by observing the state before and after a past flood event in an arbitrary region from above; a simulation inundation map creation means for creating, as a simulation inundation map, image data that enables identification of an inundated area in the region by physical flood simulation using, as input data, elevation data for the arbitrary region, estimated precipitation data in the flood event, estimated outflow data in the flood event, and river terrain data defining a watercourse; a model learning means for generating, by machine learning using the observed inundation map and the simulation inundation map as teacher data, a neural network model for predicting an inundated area based on elevation data and precipitation data; and a predicted inundation map creation means for inputting elevation data and precipitation in a flood prediction target region into the learned neural network model generated in the model learning step and creating, as a predicted inundation map, image data that enables identification of the inundated area in the prediction target region.

Images