JP7529059B2 - Flood prediction program, flood prediction device, and machine learning method - Google Patents

Description

An embodiment of the present invention relates to flood prediction technology.

Conventionally, there is known technology that predicts future flooding based on observational information such as precipitation and typhoon conditions. Conventional technology for this flood prediction involves constructing a mathematical model that derives optimal parameters by machine learning past precipitation and water level data. The constructed mathematical model is then used to accurately predict future water levels based on current precipitation and water level data and weather (rainfall forecast) data for the next few hours distributed to each local government by meteorological agencies. There is also conventional technology that predicts areal flooding conditions in urban water areas in real time and presents evaluation results when the analysis grid is set to standard 5m and 10m meshes.

Developing AI technology that can predict river water levels using limited past rainfall and water level data, Fujitsu Limited, Fujitsu Laboratories Ltd., PRESS RELEASE, Internet <URL: https://pr.fujitsu.com/jp/news/2019/08/16-2.html> Kimura, Makoto; Kido, Yuno; Nakakita, Eiichi. Fundamental study on real-time flood prediction method for localized heavy rainfall in urban water areas. Journal of the Japan Society of Civil Engineers, B1 (Hydraulic Engineering), 2011, 67.4: I_931-I_936. Internet: URL: https://www.jstage.jst.go.jp/article/jscejhe/67/4/67_4_I_931/_pdf/-char/ja

However, with the above-mentioned conventional technology, when attempting to perform accurate flood predictions using machine learning models with a fine mesh (analysis grid) and high-resolution planar scale, the amount of data required for processing becomes large, resulting in enormous amounts of memory and calculation required.

One aspect aims to reduce the computational costs of flood prediction.

In one proposal, the flood prediction program causes a computer to execute an acquisition process, a generation process, and an output process. The acquisition process acquires observation information related to the target area. The generation process inputs the observation information into a first machine learning model to generate a first image showing a predicted result of the flooding situation in the target area, and inputs the first image into a second machine learning model to generate a second image having a higher resolution than the first image. The output process outputs the second image as a flood prediction result for the target area.

It reduces the computational costs of flood prediction.

FIG. 1 is an explanatory diagram illustrating an overview of flood prediction (during machine learning) according to an embodiment. FIG. 2 is an explanatory diagram illustrating downsampling and super-resolution. FIG. 3 is an explanatory diagram illustrating an overview of flood prediction (at prediction time) according to the embodiment. FIG. 4 is a block diagram illustrating an example of a functional configuration of the information processing device according to the embodiment. FIG. 5 is a flowchart illustrating an example of an operation during machine learning of the information processing device according to the embodiment. FIG. 6 is an explanatory diagram for explaining an overview of downsampling. FIG. 7 is a flowchart illustrating an example of an operation during machine learning of the information processing device according to the embodiment. FIG. 8 is a flowchart illustrating an example of an operation at the time of prediction of the information processing device according to the embodiment. FIG. 9 is a block diagram showing an example of a computer configuration.

Below, a flood prediction program, a flood prediction device, and a machine learning method according to the embodiments will be described with reference to the drawings. Configurations having the same functions in the embodiments will be given the same reference numerals, and duplicate descriptions will be omitted. Note that the flood prediction program, the flood prediction device, and the machine learning method described in the following embodiments are merely examples, and do not limit the embodiments. Furthermore, the following embodiments may be combined as appropriate within a range that does not cause contradictions.

In an embodiment, a machine learning model is used to generate a forecast map showing predicted flood conditions at each point in a target area for which flooding is predicted, based on observational information related to flooding in the target area, such as precipitation and water levels. The target area for flooding prediction is a specific area (e.g., a city or town area) separated by an administrative body, etc. The observational information includes water levels, flow rates, inflow rates, and discharge rates of rivers and sea areas related to the target area, and precipitation in the entire area including the target area.

Specifically, in the embodiment, based on observation information about the target area and a predictive map corresponding to the observation information, machine learning is performed to create a machine learning model so that a predictive map is output as a correct answer in response to input of observation information. Then, at the time of prediction, actual observation information about the target area is input to the machine learning model, and a predictive map is obtained from the output of the machine learning model.

Figure 1 is an explanatory diagram illustrating an overview of flood prediction according to an embodiment (during machine learning). As shown in Figure 1, in S1, the information processing device according to the embodiment prepares a set of training data to be used in machine learning.

Specifically, in S1, the information processing device performs a simulation or the like (S1a) based on observation information D1 relating to the target area, and obtains high-resolution flood prediction data D2, which is a forecast map with a high surface resolution (e.g., 5m mesh) obtained by finely dividing the mesh.

The high-resolution flood forecast data D2 is a correct prediction map corresponding to the observation information D1, showing predicted values of the flooding conditions (water level above ground level, flooding below floor level, flooding above floor level, etc.) at each point corresponding to each mesh. The pixel value of each pixel in the high-resolution flood forecast data D2 corresponds to the predicted value of the flooding conditions at each point in the target area.

The method of obtaining high-resolution inundation prediction data D2 by the information processing device is not limited to obtaining it by simulation from observation information D1. For example, in S1a, the information processing device may obtain high-resolution inundation prediction data D2 based on observation values (such as water levels at each point) obtained by actual observation.

Next, in S1, the information processing device downsamples the predicted map in the high-resolution inundation prediction data D2 (S1b) to obtain low-resolution inundation prediction data D3, which has a lower resolution than the high-resolution inundation prediction data D2.

Figure 2 is an explanatory diagram explaining downsampling and super-resolution. As shown in Figure 2, for example, an information processing device downsamples high-resolution inundation prediction data D2 of 5 m mesh to obtain low-resolution inundation prediction data D3 of 50 m mesh.

In S1, the information processing device prepares a dataset of high-resolution inundation prediction data D2 and low-resolution inundation prediction data D3 corresponding to observation information D1 from observation information D1 for various cases as training data.

Next, the information processing device generates a first machine learning model M1 by performing known machine learning using a regression prediction technique such as a neural network based on the observation information D1 in the prepared training data and the low-resolution inundation prediction data D3 (S2). Specifically, the information processing device sets parameters of the first machine learning model M1 using known techniques such as a gradient method or an error backpropagation method so that the output from the first machine learning model M1 when the observation information D1 is input to the first machine learning model M1 becomes the low-resolution inundation prediction data D3 that is the correct answer.

The information processing device also generates a second machine learning model M2 by performing machine learning of a model in a known super-resolution technology using a CNN (Convolutional Neural Network) or the like based on the low-resolution flood prediction data D3 in the prepared training data and the high-resolution flood prediction data D2 (S3). Specifically, the second machine learning model M2 is a CNN that provides a single-image super-resolution method for obtaining an image with a higher resolution by increasing the resolution (super-resolution) from a single image. The information processing device sets parameters of the second machine learning model M2 using a known method such as a gradient method or an error backpropagation method so that the output from the second machine learning model M2 when the low-resolution flood prediction data D3 is input to the second machine learning model M2 becomes the low-resolution flood prediction data D3 that is the correct answer.

In general super-resolution methods, various images (with different structures at the sub-grid scale) are subjected to high resolution, and therefore machine learning is performed using a large number of training data (examples). In contrast, in the super-resolution method of the embodiment, the predicted values (pixel values) of each point in the target area (with the same topography and building shapes at the sub-grid scale) are subjected to high resolution. Therefore, in generating the second machine learning model M2, less training data (examples) than in general super-resolution is sufficient.

Figure 3 is an explanatory diagram illustrating an overview of flood prediction (at prediction time) according to an embodiment. As shown in Figure 3, at prediction time (S4), the information processing device of the embodiment collects information distributed by meteorological organizations and measurement data from measuring devices to obtain observation information D10 related to the target area.

Next, the information processing device inputs the acquired observation information D10 into a first machine learning model M1 to generate low-resolution flood prediction data D11 that shows the predicted results (predictive map) of the flood situation in the target area.

Next, the information processing device generates high-resolution inundation prediction data D12 having a higher resolution than the low-resolution inundation prediction data D11 by inputting the generated low-resolution inundation prediction data D11 into a second machine learning model M2. Specifically, as shown in Fig. 2, the information processing device obtains high-resolution inundation prediction data D12 by increasing the resolution (super-resolution) of the low-resolution inundation prediction data D11 using the second machine learning model M2. This allows the information processing device to obtain high-resolution inundation prediction data D12, which is a predicted map of inundation conditions with a high surface resolution (e.g., 5m mesh).

If a machine learning model is used to directly obtain high-resolution flood prediction data D12, which is a high-resolution prediction map, from observation information D10 (task A), the amount of data required for processing will be large, and the amount of memory and calculation required will be enormous. The reason for this is that, for example, if a predicted value is y and an observed value included in observation information D10 is x, in y = f(x), the number of data elements for y is an order of magnitude larger than that for x, so the amount of calculation will be large and machine learning requires a huge amount of data.

Therefore, in the information processing device of the embodiment, a first machine learning model M1 is used to obtain low-resolution inundation prediction data D11, which is a predictive map with a lower resolution than the high-resolution inundation prediction data D12, from the observation information D10 (task B). Next, the information processing device of the embodiment increases the resolution of the low-resolution inundation prediction data D11 using a second machine learning model M2 to obtain high-resolution inundation prediction data D12 (task C).

As an example, the amount of data (number of cases) during machine learning in task (A) which attempts to directly obtain high-resolution inundation forecast data D12, which is a high-resolution predictive map, from observation information D10, is set to N_A, and the data size for each case is set to S_A. Furthermore, the amount of data (number of cases) during machine learning in task (B) which attempts to obtain low-resolution inundation forecast data D11, which is a high-resolution predictive map, from observation information D10, is set to N_B, and the data size for each case is set to S_B. Furthermore, the amount of data (number of cases) during machine learning in task (C) which attempts to obtain high-resolution inundation forecast data D12 from low-resolution inundation forecast data D11 is set to N_C, and the data size for each case is set to S_C.

Task (B) uses machine learning to learn how flooding occurs from a variety of observational information D10, so the number of cases is about the same as task (A) (N_A ≒ N_B). However, the amount of data for each case used in machine learning is low-resolution flood forecast data D3 downsampled from high-resolution flood forecast data D2, so S_A >> S_B.

Task (C) seeks high-resolution flood forecast data D12 obtained by increasing the resolution of low-resolution flood forecast data D11, and differences in observation information D1 are not an issue during machine learning. Furthermore, in machine learning for task (C), the predicted values (pixel values) of each point in the same target area (same topography and building shapes at the sub-grid scale) are the targets for high resolution, so the number of examples can be small (N_A>>N_C).

Therefore, when comparing the data size (number of cases x amount of data per case) during machine learning in task (B) + task (C) with the data size during machine learning in task (A), the result is as follows.
N_A×S_A >> N_B×S_B+N_C×S_C

In other words, the data size of task (B) + task (C) is smaller than that of task (A). Specifically, the difference may be about 100 times. Therefore, in the information processing device of the embodiment, the amount of data during machine learning can be reduced, and the amount of calculations can be prevented from becoming large.

Figure 4 is a block diagram showing an example of the functional configuration of an information processing device according to an embodiment. As shown in Figure 4, the information processing device 1 has a communication unit 10, a display unit 11, an operation unit 12, an input/output unit 13, a memory unit 14 and a control unit 15. This information processing device 1 is an example of a flood prediction device, and may be, for example, a PC (Personal Computer).

The communication unit 10 is realized, for example, by a NIC (Network Interface Card) or the like. The communication unit 10 is connected to other information processing devices via a network (not shown) in a wired or wireless manner, and is a communication interface that manages the communication of information between the communication unit 10 and the other information processing devices.

The display unit 11 is a display device for displaying various information. The display unit 11 is realized, for example, by a liquid crystal display or the like as a display device. The display unit 11 displays various screens, such as a display screen input from the control unit 15.

The operation unit 12 is an input device that accepts various operations from a user of the information processing device 1. The operation unit 12 is realized, for example, as an input device, by a keyboard, a mouse, etc. The operation unit 12 outputs the operation input by the user to the control unit 15 as operation information. Note that the operation unit 12 may be realized as an input device by a touch panel, etc., and the display device of the display unit 11 and the input device of the operation unit 12 may be integrated.

The input/output unit 13 is, for example, a memory card R/W (Reader/Writer). Instead of the observation information 141 etc. received by the communication unit 10, the input/output unit 13 may read the observation information 141 etc. stored in the memory card and store it in the storage unit 14. The input/output unit 13 also stores, for example, the prediction result output from the control unit 15 in the memory card. Note that, for example, an SD memory card or the like can be used as the memory card.

The memory unit 14 is realized by, for example, a semiconductor memory element such as a random access memory (RAM) or a flash memory, or a storage device such as a hard disk or an optical disk. The memory unit 14 stores observation information 141, training data 142, prediction data 143, first machine learning model information 144, and second machine learning model information 145.

Observation information 141 is observation information regarding the target area obtained from a server or the like, and corresponds to the observation information D1 and D10 described above.

The training data 142 is data used for machine learning of the first machine learning model M1 and the second machine learning model M2. Specifically, the training data 142 is data that is a set of the above-mentioned observation information D1, and high-resolution inundation prediction data D2 and low-resolution inundation prediction data D3 corresponding to the observation information D1 for each case used in machine learning.

The prediction data 143 is data showing prediction results from the observation information D10. Specifically, the prediction data 143 corresponds to low-resolution inundation prediction data D11 obtained by inputting the observation information D10 into a first machine learning model M1, and high-resolution inundation prediction data D12 obtained by inputting the low-resolution inundation prediction data D11 into a second machine learning model M2.

The first machine learning model information 144 is information about the first machine learning model M1, and is parameters for constructing the first machine learning model M1, such as a neural network. The second machine learning model information 145 is information about the second machine learning model M2, and is parameters for constructing the second machine learning model M2, such as a CNN.

The control unit 15 is a processing unit that controls the operation of the information processing device 1. The control unit 15 has an acquisition unit 151, a training data creation unit 152, a first machine learning unit 153, a second machine learning unit 154, an estimation unit 155, and an output unit 156. The control unit 15 can be realized by a CPU (Central Processing Unit) or an MPU (Micro Processing Unit), etc. The control unit 15 can also be realized by hardwired logic such as an ASIC (Application Specific Integrated Circuit) or an FPGA (Field Programmable Gate Array).

The acquisition unit 151 is a processing unit that acquires the observation information 141. For example, the acquisition unit 151 acquires the observation information 141 by communicating with a specific server that provides the observation information 141 related to the target area via the communication unit 10. The acquisition unit 151 stores the acquired observation information 141 in the memory unit 14.

The training data creation unit 152 is a processing unit that creates training data 142 used for machine learning of the first machine learning model M1 and the second machine learning model M2. Specifically, the training data creation unit 152 performs a simulation or the like using the observation information D1 (141) to obtain high-resolution inundation prediction data D2 corresponding to the observation information D1. The training data creation unit 152 may directly apply the observation values included in the observation information D1 to obtain the high-resolution inundation prediction data D2.

Next, the training data creation unit 152 downsamples the forecast map in the high-resolution inundation forecast data D2 to obtain low-resolution inundation forecast data D3 with a lower resolution than the high-resolution inundation forecast data D2. Next, the training data creation unit 152 stores training data 142, which is a set of observation information D1, and the high-resolution inundation forecast data D2 and low-resolution inundation forecast data D3 corresponding to this observation information D1, in the memory unit 14. The training data creation unit 152 performs the above process for the number of cases used in machine learning to create training data 142 for the number of cases.

The first machine learning unit 153 is a processing unit that generates a first machine learning model M1 based on the training data 142. Specifically, the first machine learning unit 153 reads out observation information D1 and low-resolution inundation prediction data D3 for each case included in the training data 142. Next, the first machine learning unit 153 sets (adjusts) parameters of the first machine learning model M1 so that when the read-out observation information D1 is input to the first machine learning model M1, the output from the first machine learning model M1 becomes low-resolution inundation prediction data D3 that is the correct answer.

Next, the first machine learning unit 153 performs the above processing for each of a predetermined number of cases prepared as training data 142, and then stores the parameters of the first machine learning model M1 in the memory unit 14 as first machine learning model information 144.

The second machine learning unit 154 is a processing unit that generates a second machine learning model M2 based on the training data 142. Specifically, the second machine learning unit 154 reads out low-resolution inundation prediction data D3 and high-resolution inundation prediction data D2 for each case included in the training data 142. Next, the second machine learning unit 154 sets (adjusts) parameters of the second machine learning model M2 so that when the read low-resolution inundation prediction data D3 is input to the second machine learning model M2, the output from the second machine learning model M2 becomes the high-resolution inundation prediction data D2 that is the correct answer.

Next, the second machine learning unit 154 performs the above processing for each of a predetermined number of cases prepared as training data 142, and then stores the parameters of the second machine learning model M2 in the memory unit 14 as second machine learning model information 145.

The estimation unit 155 is a processing unit that estimates high-resolution inundation prediction data D12, which is a predicted map of the inundation situation in the target area, based on observation information 141 (D10) of the target area. Specifically, the estimation unit 155 constructs a first machine learning model M1 based on the first machine learning model information 144 read from the memory unit 14. Next, the estimation unit 155 generates low-resolution inundation prediction data D11 by inputting the observation information D10 to the first machine learning model M1. The estimation unit 155 stores the generated low-resolution inundation prediction data D11 in the memory unit 14 as prediction data 143.

Next, the estimation unit 155 constructs a second machine learning model M2 based on the second machine learning model information 145 read from the memory unit 14. Next, the estimation unit 155 generates high-resolution inundation prediction data D12 with increased resolution by inputting the low-resolution inundation prediction data D11 into the second machine learning model M2. The estimation unit 155 stores the generated high-resolution inundation prediction data D12 in the memory unit 14 as prediction data 143.

The output unit 156 is a processing unit that outputs the estimation result of the estimation unit 155. Specifically, the output unit 156 reads out the high-resolution inundation forecast data D12 included in the forecast data 143 from the memory unit 14, and generates a display screen of a forecast map showing the inundation status of each point in the target area based on the high-resolution inundation forecast data D12. Next, the output unit 156 outputs the data of the generated display screen to the display unit 11, and causes the forecast map to be displayed on the display device.

The output destination of the output unit 156 is not limited to the display unit 11. For example, the output unit 156 may output the high-resolution flood prediction data D12 read from the memory unit 14 to a memory card via the input/output unit 13.

Figure 5 is a flowchart showing an example of the operation of the information processing device 1 during machine learning according to the embodiment. In the flowchart shown in Figure 5, training data 142 is created by a simulation based on observation information D1.

As shown in FIG. 5, when processing begins, the training data creation unit 152 performs N simulations of various types (e.g., different weather conditions, etc.) based on the training data 142 (D1) (S10) and obtains N sets of high-resolution flood prediction data D2.

Next, the training data creation unit 152 obtains low-resolution inundation prediction data D3 by downsampling from the high-resolution inundation prediction data D2 for N sets of high-resolution inundation prediction data D2 (S11).

Figure 6 is an explanatory diagram outlining downsampling. As shown in Figure 6, the training data creation unit 152 performs downsampling at a predetermined scale (e.g., 5m mesh → 50m mesh) to obtain low-resolution inundation prediction data D3 from high-resolution inundation prediction data D2.

Here, when downsampling from pixels contained in a specific region of the high-resolution flood prediction data D2, the training data creation unit 152 obtains low-resolution flood prediction data D3 by sampling the pixel with the worst flood prediction evaluation among the pixels contained in that region. The good or bad flood evaluation here refers to the height of the water level above the ground, and the higher the water level, the worse the evaluation (flooding above the floor is evaluated as worse than flooding below the floor). In this way, the training data creation unit 152 prevents the pixel with the worst flood prediction evaluation from being removed by downsampling.

Specifically, in case C1, high-resolution inundation prediction data D2a for a specific area includes a right-diagonal shaded portion (predicted value a) and a left-diagonal shaded portion (predicted value b), with predicted value b being evaluated worse than a. In this case, the training data creation unit 152 samples the worse-evaluated predicted value b. In other words, the training data creation unit 152 downsamples the high-resolution inundation prediction data D2a to low-resolution inundation prediction data D3a in which predicted value b corresponds to one pixel.

Similarly, in case C2, the high-resolution flood prediction data D2b for a specific area includes a right-diagonal shaded portion (predicted value a) and a left-diagonal shaded portion (predicted value b). Therefore, the training data creation unit 152 downsamples the high-resolution flood prediction data D2b to low-resolution flood prediction data D3b in which the predicted value b corresponds to one pixel.

In case C3, the high-resolution flood prediction data D2c for a specific area includes the right-diagonal shaded portion (predicted value a) but does not include the left-diagonal shaded portion (predicted value b). Therefore, the training data creation unit 152 downsamples the high-resolution flood prediction data D2c to low-resolution flood prediction data D3c in which the predicted value a corresponds to one pixel.

Next, the training data creation unit 152 prepares N sets of observation information D1 and low-resolution inundation prediction data D3 as training data 142 (S12). Next, the first machine learning unit 153 divides the N sets of training data 142 into N = N_train + N_test (S13).

Next, the first machine learning unit 153 performs machine learning of the first machine learning model M1 using the set of N_train observation information D1 and low-resolution inundation prediction data D3. The first machine learning unit 153 also performs validation of the first machine learning model M1 after machine learning using the set of N_test observation information D1 and low-resolution inundation prediction data D3. The first machine learning unit 153 adopts the first machine learning model M1 for which the error contained in the prediction has been confirmed to be within a predetermined range by the validation, and constructs the first machine learning model M1 (S14).

Furthermore, following S11, the training data creation unit 152 prepares M sets of high-resolution inundation forecast data D2 and low-resolution inundation forecast data D3 as training data 142 (S15). Next, the second machine learning unit 154 divides the M sets of training data 142 into M = M_train + M_test (S16). Note that the number of sets is M << N.

Next, the second machine learning unit 154 performs machine learning of the second machine learning model M2 using the set of high-resolution inundation prediction data D2 of M_train and low-resolution inundation prediction data D3. The second machine learning unit 154 also performs validation of the second machine learning model M2 after machine learning using the set of high-resolution inundation prediction data D2 of M_test and low-resolution inundation prediction data D3. The second machine learning unit 154 adopts the second machine learning model M2 for which the error contained in the prediction has been confirmed to be within a predetermined range by the validation, and constructs the second machine learning model M2 (S17).

Following S14 and S17, the control unit 15 stores the first machine learning model information 144 and the second machine learning model information 145 relating to the constructed first machine learning model M1 and second machine learning model M2 in the memory unit 14, and ends the process. Note that the processes of S13 and S14 by the first machine learning unit 153 and the processes of S16 and S17 by the second machine learning unit 154 may be performed in parallel.

Figure 7 is a flowchart showing an example of the operation of the information processing device 1 during machine learning according to the embodiment. Note that in the flowchart shown in Figure 7, no simulation is performed, and the training data 142 is created by using the observation data included in the observation information D1.

As shown in Figure 7, when the process starts, the training data creation unit 152 prepares N pieces of various types of observation information D1 (S10a). Here, the observation information D1 is assumed to include high-resolution inundation observation data (actually measured inundation situation map) corresponding to the high-resolution inundation prediction data D2.

The training data creation unit 152 downsamples the high-resolution inundation observation data to the low-resolution inundation observation data for N sets of observation information D1 (S11a). Next, the training data creation unit 152 prepares the N sets of observation information D1 and the low-resolution inundation observation data as training data 142 (S12a). Thereafter, the first machine learning unit 153 performs processing similar to S13 and S14 described above based on the prepared training data 142 to construct a first machine learning model M1.

Furthermore, following S11a, the training data creation unit 152 prepares M sets of high-resolution inundation observation data and low-resolution inundation observation data as training data 142 (S15a). Thereafter, the second machine learning unit 154 performs processing similar to S16 and S17 described above based on the prepared training data 142 to construct a second machine learning model M2.

Figure 8 is a flowchart showing an example of the operation of the information processing device 1 during prediction according to the embodiment. As shown in Figure 8, when processing is started, the acquisition unit 151 acquires observation information D10 at the time of the occurrence of a disaster from a specific server that provides observation information related to a target area (S20).

Next, the estimation unit 155 constructs a first machine learning model M1 from the first machine learning model information 144. Next, the estimation unit 155 acquires low-resolution inundation prediction data D11 by inputting the acquired observation information D10 into the constructed first machine learning model M1 (S21).

Next, the estimation unit 155 constructs a second machine learning model M2 from the second machine learning model information 145. Next, the estimation unit 155 inputs the low-resolution inundation prediction data D11 acquired in S21 into the constructed second machine learning model M2 to acquire high-resolution inundation prediction data D12 (S22). The output unit 156 outputs the high-resolution inundation prediction data D12 acquired in S22 to the display unit 11 (S23), and displays a forecast map based on the observation information D10 at the time of the disaster occurrence.

As described above, the information processing device 1 acquires observation information D10 relating to the target area. The information processing device 1 generates a first image (low-resolution inundation prediction data D11) showing the predicted results of the flooding situation in the target area by inputting the acquired observation information D10 into the first machine learning model M1. The information processing device 1 generates a second image (high-resolution inundation prediction data D12) having a higher resolution than the first image by inputting the first image into the second machine learning model M2. The information processing device 1 outputs the second image generated as the flooding prediction result for the target area.

This allows the information processing device 1 to perform accurate flood predictions using machine learning models at high resolution while minimizing increases in memory and calculation volume required for processing. In this way, the information processing device 1 can support the improvement of the accuracy of flood predictions.

The first image generated by the information processing device 1 is a map image showing the flooding conditions at each point overlooking the target area. This allows the information processing device 1 to obtain a more detailed map image showing the flooding conditions at each point overlooking the target area.

The information processing device 1 also downsamples the observation information D1 related to the target area and the corresponding flood forecast image (high-resolution flood forecast data D2). The information processing device 1 generates a first machine learning model M1 based on the downsampled flood forecast image (low-resolution flood forecast data D3) and the observation information D1. The information processing device 1 generates a second machine learning model M2 based on the downsampled flood forecast image (low-resolution flood forecast data D3) and the original flood forecast image (high-resolution flood forecast data D2), which takes the output of the first machine learning model M1 as input and outputs an image in which the input image has been made high-resolution.

This allows the information processing device 1 to create a machine learning model that performs accurate flood prediction at high resolution while minimizing increases in memory and calculation volume required for processing. In this way, the information processing device 1 can support the improvement of the accuracy of flood prediction.

Furthermore, when downsampling pixels contained in a specific region of the flood prediction image (high-resolution flood prediction data D2), the information processing device 1 samples the pixel with the worst flood prediction evaluation among the pixels contained in that region. This allows the information processing device 1 to prevent the pixel with the worst flood prediction evaluation from being removed by downsampling. Furthermore, by creating a first machine learning model M1 using the thus downsampled flood prediction image (low-resolution flood prediction data D3), the information processing device 1 becomes able to accurately predict the worst case in which the evaluation is the worst in flood prediction using the first machine learning model M1.

Note that each component of each device shown in the figure does not necessarily have to be physically configured as shown in the figure. In other words, the specific form of distribution and integration of each device is not limited to that shown in the figure, and all or part of them can be functionally or physically distributed and integrated in any unit depending on various loads and usage conditions. For example, the functional configuration related to machine learning in the information processing device 1 (e.g., the training data creation unit 152, the first machine learning unit 153, and the second machine learning unit 154) and the functional configuration related to prediction (e.g., the estimation unit 155 and the output unit 156) may be separate configurations, or each may be realized by an independent device configuration.

In addition, the various processing functions performed by the information processing device 1 may be executed in whole or in part on a CPU (or a microcomputer such as an MPU or MCU (Micro Controller Unit)) or a GPU (Graphics Processing Unit). It goes without saying that the various processing functions may be executed in whole or in part on a program analyzed and executed by the CPU (or a microcomputer such as an MPU or MCU) or a GPU, or on hardware based on wired logic. In addition, the various processing functions performed by the information processing device 1 may be executed by multiple computers working together through cloud computing.

The various processes described in the above embodiments can be realized by executing a program prepared in advance on a computer. Therefore, an example of a computer (hardware) that executes a program having the same functions as those in the above embodiments will be described below. Figure 9 is a block diagram showing an example of a computer configuration.

9, computer 200 has a CPU 201 that executes various types of arithmetic processing, a GPU 201a specialized for predetermined arithmetic processing such as image processing and machine learning processing, an input device 202 that accepts data input, a monitor 203, and a speaker 204. Computer 200 also has a medium reading device 205 that reads programs and the like from a storage medium, an interface device 206 for connecting to various devices, and a communication device 207 for connecting to external devices by wire or wirelessly. Computer 200 also has a RAM 208 that temporarily stores various types of information, and a hard disk device 209. Each unit (201 to 209) in computer 200 is connected to bus 210.

The hard disk drive 209 stores a program 211 for executing various processes in the acquisition unit 151, training data creation unit 152, first machine learning unit 153, second machine learning unit 154, estimation unit 155, output unit 156, etc. in the control unit 15 described in the above embodiment. The hard disk drive 209 also stores various data 212 such as observation information referenced by the program 211. The input device 202, for example, accepts input of operation information from an operator. The monitor 203, for example, displays various screens operated by the operator. The interface device 206 is connected to, for example, a printing device. The communication device 207 is connected to a communication network such as a LAN (Local Area Network) and exchanges various information with external devices via the communication network.

The CPU 201 or the GPU 201a reads out the program 211 stored in the hard disk drive 209, expands it in the RAM 208, and executes it to perform various processes related to the acquisition unit 151, the training data creation unit 152, the first machine learning unit 153, the second machine learning unit 154, the estimation unit 155, and the output unit 156. The program 211 does not have to be stored in the hard disk drive 209. For example, the program 211 stored in a storage medium readable by the computer 200 may be read out and executed. Examples of the storage medium readable by the computer 200 include portable storage media such as CD-ROMs, DVD disks, USB (Universal Serial Bus) memories, semiconductor memories such as flash memories, and hard disk drives. The program 211 may also be stored in a device connected to a public line, the Internet, a LAN, or the like, and the computer 200 may read out and execute the program 211 from the device.

1... Information processing device 10... Communication unit 11... Display unit 12... Operation unit 13... Input/output unit 14... Memory unit 15... Control unit 141... Observation information 142... Training data 143... Prediction data 144... First machine learning model information 145... Second machine learning model information 151... Acquisition unit 152... Training data creation unit 153... First machine learning unit 154... Second machine learning unit 155... Estimation unit 156... Output unit 200... Computer 201... CPU
201a...GPU
202: Input device 203: Monitor 204: Speaker 205: Media reader 206: Interface device 207: Communication device 208: RAM
209...Hard disk drive 210...Bus 211...Program 212...Various data C1 to C3...Cases D1, D10...Observation information D2, D2a to D2c, D12...High-resolution flood prediction data D3, D3a to D3c, D11...Low-resolution flood prediction data M1...First machine learning model M2...Second machine learning model

Claims (5)

Obtaining observational information on the target area;
inputting the observation information into a first machine learning model to generate a first image showing a predicted result of a flood condition in the target area;
generating a second image having a higher resolution than the first image by inputting the first image into a second machine learning model;
outputting the second image as a flood prediction result for the target area;
A flood prediction program that causes a computer to execute processing. The first image is a map image showing a flooding situation at each point overlooking the target area,
2. The flood prediction program according to claim 1, Obtaining observational information on the target area;
inputting the observation information into a first machine learning model to generate a first image showing a predicted result of a flood condition in the target area;
generating a second image having a higher resolution than the first image by inputting the first image into a second machine learning model;
outputting the second image as a flood prediction result for the target area;
A flood prediction device comprising a control unit that executes processing. The observation information for the target area and the corresponding flood prediction images are downsampled.
Generate a first machine learning model based on the downsampled flood prediction image and the observation information;
generating a second machine learning model based on the downsampled flood prediction image and the flood prediction image, the second machine learning model receiving an output of the first machine learning model as an input and outputting an image in which the input image has been made high resolution;
A machine learning method characterized in that processing is executed by a computer. The machine learning method according to claim 4, characterized in that the downsampling process samples the pixel having the worst flood prediction evaluation among the pixels included in a specific area of the flood prediction image when downsampling from pixels included in the specific area.

Images