Learning device, learning method, and prediction device

Abstract

To provide a learning device, a learning method, and a prediction device that can perform an accurate prediction by using information which can be easily prepared and easily acquired.SOLUTION: The learning device includes: a learning data acquisition unit 12 for acquiring, as learning data, a water level in a river at a first time point in a snow melting period of time and a set of representative point information at one or more second time points including the past before the first time point and weather item value information; and a learning processing unit 13 for causing a learning unit to mechanically learn by using the learning data. The representative point information relates to a representative point on which the intensity distribution of a weather item in a range including at least a part of a water collecting area is reflected. The weather item value information relates to a weather item value in the range.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a learning device, a learning method, and a prediction device. [Background technology]

[0002] In recent years, the effects of global warming have led to an increase in heavy rain disasters, and damage caused by external and internal flooding in rivers is on the rise. Even during river construction, rising water levels can cause damage (washout, sinking, etc.) to workers and construction equipment (heavy machinery and materials). Therefore, predicting water levels in advance and informing construction personnel is important for protecting workers and construction equipment (heavy machinery and materials). Flood warning systems have traditionally been applied to river construction sites. An example of the configuration of a conventional flood warning system is shown in FIG. 29. The flood warning system shown in FIG. 29 obtains information about rainfall from the Japan Meteorological Agency and river information from the Ministry of Land, Infrastructure, Transport and Tourism, and predicts river water levels by analyzing this information. The predicted river water levels and warnings based on the predictions are notified to construction workers at the construction site via the Internet, mobile email, etc. Conventional flood warning systems predict river water levels using physical models (water level prediction models) based on hydraulic formulas, and physical models such as (1) a "numerical model," (2) a "regression model," (3) a "cumulative rainfall model," and (4) a "conservation law model" are used (see, for example, Patent Documents 1 and 2). The "numerical model" is a physical model that uses distributed runoff analysis to determine the water level at the construction site using the water level, rainfall distribution, land use, and elevation upstream of the prediction point as input values. The "regression model" is a physical model that determines the regression equation between the water level at an observation station upstream of the prediction point and the water level at the prediction point, and predicts the water level at the prediction point from the regression equation. The "accumulated rainfall model" is a physical model that determines whether or not there will be flooding based on the relationship between the rainfall in the watershed upstream of the prediction point and the water level at the prediction point. The "conservation law model" is a physical model that predicts the water level at the prediction point based on the relationship between the rainfall in the watershed upstream of the prediction point and the water level at the prediction point. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Application Publication No. 2019-045290 [Patent Document 2] Japanese Patent Application Laid-Open No. 2008-015916 Summary of the Invention [Problem to be solved by the invention]

[0004] Generally, the construction period for river works is divided into the "flood season," which is the period when flooding is likely to occur due to heavy rain (rainy season) or typhoons, and the "non-flood season," which is the period outside of the flood season. As a guideline for carrying out river works, the flood season is defined as June 1st to October 31st, and the non-flood season is defined as November 1st to May 31st of the following year, and river works are often not carried out during the flood season. Flooding can occur even during the non-flood season due to heavy rain, for example, in rivers that flow through areas with heavy snowfall, flooding occurs from March to May due to snowmelt. Snowmelt is caused by multiple factors, including temperature, rainfall, wind speed, sunshine hours, and the associated heat balance and cooling suppression effects. Therefore, when focusing on the snowmelt period, the following challenges arise when predicting water levels. First, although there are water level prediction methods that take snowmelt into account, such as runoff analysis, there is the issue that it takes time and effort to construct equations that take into account meteorological factors related to snowmelt and to tune parameters. Second, while many methods use snowmelt volume as an input value to predict hourly water levels during the snowmelt season, there is a problem in that accurate prediction of snowmelt volume requires observational data from multiple locations. Specifically, these methods include (a) the degree-hour (degree-day) method, which simply calculates daily snowmelt volume using daily mean temperature; (b) the snow surface depression method, which calculates snowmelt volume per day from changes in snow depth; (c) the heat balance method, which calculates snowmelt volume by calculating the heat quantity related to radiation, sensible heat, and latent heat transfer from meteorological parameters such as snow depth, temperature, and solar radiation; and (d) the tank model method, which uses the snow layer as a tank to predict snowmelt water on an hourly basis. These methods require observational data such as precipitation, temperature, and wind speed from multiple locations to improve snowmelt volume prediction accuracy, and they may not be able to predict snowmelt volume with limited observational data. From this perspective, the present invention provides a learning device, a learning method, and a prediction device that require simple advance preparation and are capable of achieving highly accurate predictions using information that is easy to obtain. [Means for solving the problem]

[0005] A learning device according to the present invention includes a learning data acquisition unit and a learning processing unit. The learning data acquisition unit acquires, as learning data, a set of the water level of the river at a first time point during a snowmelt period when snow is present in at least a portion of the river's catchment area, and representative point information and meteorological item value information at at least one or more second time points that are earlier than the first time point. The learning processing unit has a learning device and uses the learning data to train the learning device through machine learning. The representative point information is information about a representative point that reflects the intensity distribution of meteorological items within a range that includes at least a portion of the watershed, and the meteorological item value information is information about meteorological item values ​​within the range. The learning data acquisition unit acquires learning data including the representative point information and meteorological item value information of meteorological items such as daily average temperature, daily accumulated precipitation, sunshine hours, global solar radiation, daily average wind speed, and snow depth. The learning processing unit is configured to Daily average temperature, daily accumulated precipitation, sunshine hours, total solar radiation, daily average wind speed and snow depth Representative point information and The learning device is trained to output the water level of the river at a fourth time point that is after the third time point by inputting meteorological item value information. 。 In the learning device according to the present invention, the use of machine learning reduces the effort required for parameter tuning (parameters are set during learning), thereby reducing the amount of work required for advance preparation compared to conventional methods. Furthermore, the representative point information used in the learning data reflects the intensity distribution of meteorological items related to snowmelt, so by using a learning device trained using this learning data, it is possible to predict water levels for various snowmelt patterns.

[0006] The weather forecasting system may further include an image processing unit that calculates the representative point information and the weather item value information from a weather distribution image that shows the intensity distribution of the weather items by changing colors. Weather distribution images published by the Japan Meteorological Agency and the National Agriculture and Food Research Organization (NARO) cover the entire country of Japan and are easy to obtain (for example, they can be purchased from the Japan Meteorological Agency). Therefore, using these weather distribution images makes it easy to predict water levels at any point throughout Japan. Furthermore, by using forecast weather distribution images (future weather distribution images), it is possible to predict water levels long into the future, allowing for ample time to evacuate and protect heavy machinery and materials in construction yards. The reason for not using the weather distribution images themselves for training is that if all the data (all pixel values) from the weather distribution images were used, the number of data points would amount to tens of thousands to hundreds of thousands per image, depending on the image resolution, which would lead to issues such as neural network coefficients (weights) not converging and requiring significant computational time. For example, the image processing unit converts the RGB values ​​of each pixel constituting the weather distribution image at the second time point into a luminance value, calculates the center of gravity of the range when the luminance value is regarded as weight as the representative point, and obtains the position of the center of gravity and the distance from a predetermined reference point to the center of gravity as the representative point information. Also, the image processing unit obtains a statistical value calculated from the luminance values ​​of each pixel in the range at the second time point as the weather item value information. The image processing unit processes the weather distribution image. NewA new weather distribution image may be created, and the representative point information and the weather item value information may be calculated from the new weather distribution image.

[0007] The learning data acquisition unit may acquire learning data that further includes a water level at a fifth time point that is an earlier time point than the first time point and is included in a time period affected by water level changes due to snowmelt. In this case, the learning processing unit trains the learning device to output the water level of the river at the fourth time point by inputting representative point information and meteorological item value information at at least one or more third time points and the water level at a sixth time point corresponding to the fifth time point. In this way, the reference water level information is reflected in the learning data, making it possible to achieve even more accurate predictions. Further, the learning processing unit Daily average temperature, daily accumulated precipitation, sunshine hours, total solar radiation, daily average wind speed and snow depth The representative point information and One or both of the weather item value information may be weighted. The learning processing unit The weight values ​​used for weighting are determined, for example, by multiple regression analysis using the water level as the dependent variable and the meteorological item values ​​as the explanatory variables. In this way, the contribution of meteorological items to the water level prediction results is reflected in the learning data, making it possible to achieve even more accurate predictions.

[0008] The learning method according to the present invention includes a learning data acquisition step and a learning processing step. In the learning data acquisition step, a set of the water level of the river at a first time point during a snowmelt period when snow is present in at least a part of the river's catchment area, and representative point information and meteorological item value information at at least one or more second time points that are earlier than the first time point, is acquired as learning data. In the learning processing step, the learning data is used to train a learner. The representative point information is information about a representative point that reflects the intensity distribution of meteorological items within a range that includes at least a portion of the watershed, and the meteorological item value information is information about meteorological item values ​​within the range. In the learning data acquisition step, learning data is acquired that includes the representative point information and meteorological item value information for meteorological items such as daily average temperature, daily accumulated precipitation, sunshine hours, global solar radiation, daily average wind speed, and snow depth. In the learning process, at least one or more third points in time Daily average temperature, daily accumulated precipitation, sunshine hours, total solar radiation, daily average wind speed and snow depth Representative point information and The learning device is trained so that it outputs the water level of the river at a fourth time point that is after the third time point by inputting meteorological item value information. In the learning method of the present invention, the use of machine learning reduces the effort required for parameter tuning (parameters are set during learning), thereby reducing the amount of work required for advance preparation compared to conventional methods. Furthermore, since the representative point information used in the learning data reflects the intensity distribution of meteorological items related to snowmelt, a learning device trained using this learning data can predict water levels for various snowmelt patterns. The prediction device of the present invention is equipped with a trained learning device that has been trained using the learning method described above, and predicts the water level of the river at the fourth time point by inputting representative point information and meteorological item value information at the third time point into the trained learning device. By using the prediction device according to the present invention, it is possible to predict water levels for various snowmelt patterns. [Effects of the Invention]

[0009] According to the present invention, accurate prediction can be achieved using information that is easy to obtain and requires simple advance preparation. [Brief explanation of the drawings]

[0010] [Figure 1] 1 is a functional configuration diagram of a learning device and a prediction device according to a first embodiment of the present invention. FIG. [Figure 2] This is an image of a river and its catchment area. [Figure 3] 10A and 10B are diagrams for explaining the processing of the image processing unit, in which (a) is an example of a weather distribution image, (b) is an example of a watershed image from which a weather distribution image within the watershed has been extracted, and (c) is an example of a representative point. [Figure 4]These are images of the calculated center of gravity for each weather distribution image. (a) shows the calculated result for the daily mean temperature image, (b) shows the calculated result for the daily accumulated precipitation image, and (c) shows the calculated result for the sunshine duration image. [Figure 5] These are images of the calculated center of gravity for each weather distribution image, where (a) shows the calculated result for the global solar radiation image, (b) shows the calculated result for the daily average wind speed image, and (c) shows the calculated result for the snow depth image. [Figure 6] 3 is an example of a training data group in the first embodiment of the present invention. [Figure 7] 1 is a flowchart illustrating a learning method and a prediction method according to a first embodiment of the present invention. [Figure 8] This is a graph showing water level changes at Tsukigata Observatory from 2017 to 2019. [Figure 9] 1 is a flowchart showing the procedure for carrying out a prediction test. [Figure 10] FIG. 1 is a diagram illustrating the relationship between test data and training data in cross-validation. [Figure 11] This is a graph comparing the measured water level in 2018 with the predicted water level in the first prediction test in the first embodiment, where (a) is a graph for the range from 2018 / 3 / 1 to 2018 / 3 / 31, (b) is a graph for the range from 2018 / 4 / 1 to 2018 / 4 / 30, and (c) is a graph for the range from 2018 / 5 / 1 to 2018 / 5 / 31. [Figure 12] This is a graph comparing the measured water levels in 2019 with the predicted water levels in the first prediction test in the first embodiment, where (a) is a graph covering the range of 2019 / 3 / 1 to 2019 / 3 / 31, (b) is a graph covering the range of 2019 / 4 / 1 to 2019 / 4 / 30, and (c) is a graph covering the range of 2019 / 5 / 1 to 2019 / 5 / 31. [Figure 13] This is a graph comparing the measured water levels in 2020 with the predicted water levels in the first prediction test in the first embodiment, where (a) is a graph covering the range of 2020 / 3 / 1 to 2020 / 3 / 31, (b) is a graph covering the range of 2020 / 4 / 1 to 2020 / 4 / 30, and (c) is a graph covering the range of 2020 / 5 / 1 to 2020 / 5 / 31. [Figure 14] This is a graph comparing the measured water level with the predicted water level in the second prediction test in the first embodiment, where (a) is a graph for the range from 2015 / 4 / 1 to 2015 / 4 / 30, and (b) is a graph for the range from 2018 / 4 / 1 to 2018 / 4 / 30. [Figure 15] This figure explains the results of the verification of applicability to river construction work. (a) is a graph showing the total number of correct, missed, and missed events for each weather item, and their proportions. (b) is a graph showing the number of missed events for each weather item, and the proportion of the absolute value of the difference in water level between the measured water level and the predicted water level. (c) is a graph showing the number of missed events for each weather item, and the proportion of the absolute value of the difference in water level between the measured water level and the predicted water level. [Figure 16] 10 is an example of a training data group according to the second embodiment of the present invention. [Figure 17] This figure compares the measured water levels in 2018 with the predicted water levels in the prediction test in the second embodiment, where (a) is a graph covering the range of 2018 / 3 / 1 to 2018 / 3 / 31, (b) is a graph covering the range of 2018 / 4 / 1 to 2018 / 4 / 30, and (c) is a graph covering the range of 2018 / 5 / 1 to 2018 / 5 / 31. [Figure 18] This figure compares the measured water levels in 2019 with the predicted water levels in the prediction test in the second embodiment, where (a) is a graph covering the range of 2019 / 3 / 1 to 2019 / 3 / 31, (b) is a graph covering the range of 2019 / 4 / 1 to 2019 / 4 / 30, and (c) is a graph covering the range of 2019 / 5 / 1 to 2019 / 5 / 31. [Figure 19] This is a graph comparing the measured water levels in 2020 with the predicted water levels in the prediction test in the second embodiment, where (a) is a graph for the range of 2020 / 3 / 1 to 2020 / 3 / 31, (b) is a graph for the range of 2020 / 4 / 1 to 2020 / 4 / 30, and (c) is a graph for the range of 2020 / 5 / 1 to 2020 / 5 / 31. [Figure 20] This is a graph showing the average and standard deviation of the absolute value of the difference between the measured water level and the predicted water level, with the measured water level at the time of prediction being ``yes'' or ``no.'' [Figure 21]10A and 10B are diagrams for explaining an example of a method for calculating weights using multiple regression analysis in the third embodiment of the present invention, in which (a) is an example of a response variable and explanatory variables used in the multiple regression analysis, and (b) is an example of a weight value obtained by the multiple regression analysis. [Figure 22] This figure compares the measured water levels in 2018 with the predicted water levels in the prediction test in the third embodiment, where (a) is a graph covering the range of 2018 / 3 / 1 to 2018 / 3 / 31, (b) is a graph covering the range of 2018 / 4 / 1 to 2018 / 4 / 30, and (c) is a graph covering the range of 2018 / 5 / 1 to 2018 / 5 / 31. [Figure 23] This figure compares the measured water levels in 2019 with the predicted water levels in the prediction test in the third embodiment, where (a) is a graph covering the range of 2019 / 3 / 1 to 2019 / 3 / 31, (b) is a graph covering the range of 2019 / 4 / 1 to 2019 / 4 / 30, and (c) is a graph covering the range of 2019 / 5 / 1 to 2019 / 5 / 31. [Figure 24] This is a graph comparing the actual water levels in 2020 with the predicted water levels in the prediction test in the third embodiment, where (a) is a graph for the range of 2020 / 3 / 1 to 2020 / 3 / 31, (b) is a graph for the range of 2020 / 4 / 1 to 2020 / 4 / 30, and (c) is a graph for the range of 2020 / 5 / 1 to 2020 / 5 / 31. [Figure 25] This figure compares the measured water levels in 2018 with the predicted water levels in the prediction test in the third embodiment, where (a) is a graph covering the range of 2018 / 3 / 1 to 2018 / 3 / 31, (b) is a graph covering the range of 2018 / 4 / 1 to 2018 / 4 / 30, and (c) is a graph covering the range of 2018 / 5 / 1 to 2018 / 5 / 31. [Figure 26] This figure compares the measured water levels in 2019 with the predicted water levels in the prediction test in the third embodiment, where (a) is a graph covering the range of 2019 / 3 / 1 to 2019 / 3 / 31, (b) is a graph covering the range of 2019 / 4 / 1 to 2019 / 4 / 30, and (c) is a graph covering the range of 2019 / 5 / 1 to 2019 / 5 / 31. [Figure 27]This is a graph comparing the actual water levels in 2020 with the predicted water levels in the prediction test in the third embodiment, where (a) is a graph for the range of 2020 / 3 / 1 to 2020 / 3 / 31, (b) is a graph for the range of 2020 / 4 / 1 to 2020 / 4 / 30, and (c) is a graph for the range of 2020 / 5 / 1 to 2020 / 5 / 31. [Figure 28] This figure shows the average and standard deviation of the absolute value of the difference between the measured water level and the predicted water level, divided into cases of ``yes'' and ``no'' measured water level at the time of prediction and ``yes'' and ``no'' weighting. [Figure 29] 1 shows an example of the configuration of a conventional flood warning system. DETAILED DESCRIPTION OF THE INVENTION

[0011] Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings. Each drawing is merely a schematic illustration to allow a sufficient understanding of the present invention. Therefore, the present invention is not limited to the illustrated examples. In each drawing, common or similar components are designated by the same reference numerals, and redundant explanations thereof will be omitted.

[0012] [First embodiment] <Configuration of the learning device and prediction device according to the first embodiment> A learning device 10 and a prediction device 20 according to the first embodiment will be described with reference to Fig. 1. Fig. 1 is a functional configuration diagram of the learning device 10 and the prediction device 20 according to the first embodiment. Note that the learning device 10 and the prediction device 20 can also be configured as a single device. The learning device 10 is a device that learns the relationship between the state of snowmelt in a river's catchment area (also called a "watershed") and the water level of the river. The learning device 10 performs machine learning using information obtained by processing a weather distribution image F. The weather distribution image F is an image showing the intensity distribution of weather-related items (sometimes abbreviated to "weather items"). In this embodiment, a weather distribution image F showing the intensity distribution of weather items related to snowmelt is particularly used. Weather items related to snowmelt include, for example, "daily mean temperature (°C)," "daily accumulated precipitation (mm / day)," "sunshine hours (h / day)," and "global solar radiation (MJ / m 2 ) / day), "daily average wind speed (m / s)", and "snow depth (cm)". The daily average temperature is the average temperature on a daily basis (the average of 24 observations taken every hour from 1:00 to 24:00). The daily accumulated precipitation is the cumulative value of precipitation in a day (24 observations taken every hour from 1:00 to 24:00). The sunshine hours are the number of hours during a day when direct sunlight hits the earth's surface (direct solar radiation is 0.12kW / m 2 Global solar radiation is the amount of global solar radiation received in a day (the sum of solar radiation incident from all directions in the sky due to scattering, solar radiation reflected from clouds, and direct solar radiation). Daily average wind speed is the average wind speed on a daily basis (the average of values ​​observed every 10 minutes from 1:00 to 24:00). Snow depth is the depth of snow on a daily basis (the average of snow depth from 1:00 to 24:00). The prediction device 20 is a device that predicts the water level of a river at a predetermined time based on the state of snowmelt in the river's catchment area. The prediction device 20 predicts the water level using information obtained by processing the weather distribution image F.

[0013] The weather distribution image F is an RGB image that shows the intensity distribution of meteorological items (weather items) using color variations. The weather distribution image F may be, for example, one published by the Japan Meteorological Agency or the National Agriculture and Food Research Organization (NARO). The intensity distribution in the weather distribution image F indicates that the intensity of meteorological item values ​​increases in the order of white, light blue, light blue, dark blue, yellow, orange, red, and brown. The weather distribution image F shows the intensity distribution of meteorological items in the catchment area of ​​the river for which learning (or prediction) is performed. The weather distribution image F preferably shows the intensity distribution of the entire river catchment area, but does not necessarily have to show the entire area (i.e., it may show only a portion of the river catchment area). The weather distribution image F preferably includes information indicating the point in time at which the intensity distribution is shown (for example, the date and time). Hereinafter, the weather distribution image F used in the learning stage may be referred to as the "weather distribution image Fa," and the weather distribution image F used in the forecasting stage may be referred to as the "weather distribution image Fb." The weather distribution image F whose weather item is "daily mean temperature" may be referred to as the "daily mean temperature image F1 (or F1a, F1b)." The weather distribution image F whose weather item is "daily accumulated precipitation" may be referred to as the "daily accumulated precipitation image F2 (or F2a, F2b)." The weather distribution image F whose weather item is "sunshine hours" may be referred to as the "sunshine hours image F3 (or F3a, F3b)." The weather distribution image F whose weather item is "global solar radiation" may be referred to as the "global solar radiation image F4 (or F4a, F4b)." Furthermore, a weather distribution image F whose weather item is "daily average wind speed" may be particularly referred to as "daily average wind speed image F5 (or F5a, F5b)." Furthermore, a weather distribution image F whose weather item is "snow depth" may be particularly referred to as "snow depth image F6 (or F6a, F6b)."

[0014] (Learning device configuration) 1, the learning device 10 includes an image processing unit 11, a learning data acquisition unit 12, and a learning processing unit 13. The image processing unit 11, the learning data acquisition unit 12, and the learning processing unit 13 are realized by program execution processing by a CPU (Central Processing Unit), dedicated circuits, etc. When these functions are realized by program execution processing, programs for realizing the functions are stored in a storage unit (not shown). A weather distribution image Fa is input to the image processing unit 11. There is no particular limit to the number of weather distribution images Fa input to the image processing unit 11, and for example, weather distribution images Fa for each day or each hour may be input. The image processing unit 11 obtains representative point information and weather item value information from the weather distribution image Fa. The representative point here reflects the intensity distribution of meteorological items in the river's catchment area, and is obtained by image processing the weather distribution image Fa. The representative point is the center of gravity of the catchment area when the intensity of meteorological item values ​​(such as "daily average temperature," "daily accumulated precipitation," "sunshine hours," "global solar radiation," "daily average wind speed," and "snow depth") is considered to be weight. The representative point information is, for example, the position of the representative point and the distance from a predetermined reference point to the representative point. The meteorological item value information is information about the values ​​(weather item values) of meteorological items in the river catchment area (e.g., "average daily temperature," "daily accumulated precipitation," "sunshine hours," "global solar radiation," "average daily wind speed," "snow depth"), and is, for example, a statistical value (e.g., average value) for the entire catchment area. The meteorological item value information is obtained by image processing the weather distribution image Fa. Note that the meteorological item value information may be an actual measured value.

[0015] The processing of the image processing unit 11 will be described with reference to Figures 2 and 3. Figure 2 is an image diagram of a river and a watershed. Figure 3 is a diagram for explaining the processing of the image processing unit, in which (a) is an example of a weather distribution image Fa, (b) is an example of a watershed image E obtained by extracting the weather distribution image Fa within the watershed, and (c) is an example of representative points. In FIG. 2, the river is denoted by the symbol K1, and the watershed is denoted by the symbol K2. The watershed K2 indicates the area where rain flows into the river K1, and the rain that falls in the watershed K2 flows out into the sea K3 via the river K1. The watershed K2 may be based on, for example, a river basin map published by the Ministry of Land, Infrastructure, Transport and Tourism. An observation station K4 is installed downstream of the river K1, and the water level of the river K1 can be measured there. In this embodiment, the water level at the observation station K4 is predicted, and attention is focused on the watershed K2A upstream of the observation station K4 (i.e., the melting of snow accumulated in the upstream watershed K2A). It is preferable to select a portion of the watershed K2 to be focused on depending on the location of the water level to be predicted. For example, when predicting the water level near the mouth of the river K1, it is preferable to focus on the entire watershed K2, including the upstream watershed K2A and the downstream watershed K2B. The time it takes for the water from melting snow accumulated in the upstream watershed K2A to affect the water level at observation station K4 is determined by the characteristics of river K1, such as its length and elevation difference. Note that river K1 shown in Figure 2 is the Ishikari River.

[0016] The image processing unit 11 extracts a watershed image E (see FIG. 3(b)), which is an image of the watershed K2A portion, from the weather distribution image Fa (see FIG. 3(a)), and converts the RGB values ​​of each pixel constituting the watershed image E into a brightness value. Brightness values ​​are also called gray values. The conversion from RGB values ​​to brightness values ​​is performed using, for example, equation (1). Brightness value = 255 - (0.587 * r value + 0.114 * g value + 0.299 * b value) Equation (1) For example, if the intensity of meteorological item values ​​increases in the order of white, light blue, light blue, dark blue, yellow, orange, red, and brown, a brightness conversion formula is constructed so that the brightness value of white approaches "0 (zero)" and the brightness value of brown approaches "255" through brightness conversion. In other words, the red, blue, and green RGB values ​​are weighted according to the intensity of the meteorological items in the intensity distribution. In this way, the image processing unit 11 creates a gray image D (see FIG. 3(c)) by brightness converting the watershed image E (see FIG. 3(b)).

[0017] Furthermore, the image processing unit 11 calculates the center of gravity G of the watershed K2A as a representative point when the brightness value of each pixel constituting the gray image D (see FIG. 3(c)) is considered to be a weight. The calculation of the center of gravity of the watershed K2A can be found from the distance (e.g., coordinate value) from a reference point (here, origin O) set on the weather distribution image Fa (see FIG. 3(a)) to each pixel and the magnitude of the brightness value, and is performed using, for example, equation (2). Here, origin O is the lower left vertex of the weather distribution image Fa, and the X-axis and Y-axis are set on the sides constituting the weather distribution image Fa, and the X-coordinate value and Y-coordinate value correspond, for example, to the position (arrangement) of the pixel. X coordinate of center of gravity G G = (X1M1+ X2M2+ … + X n M n ) / (M1+ M2+ … + M n ) ···Formula (2) Here, "X1,X2, ... ,X n " is the distance in the X-axis direction from the origin O to each pixel (X coordinate value), and "M1,M2, ... ,M n " is the brightness value of each pixel. The Y coordinate value of the center of gravity G is calculated in the same way, and the center of gravity coordinate (X G , Y G ) is obtained. Then, the image processing unit 11 calculates the position of the center of gravity G (center of gravity coordinates (X G , Y G ) and the distance L from the reference point (here, the origin O) to the center of gravity G are used as representative value information. The image processing unit 11 also calculates a statistical value R of the brightness values ​​of each pixel that makes up the gray image D (see FIG. 3(c)), and uses the calculated statistical value R as meteorological item value information. The statistical value R of the brightness values ​​may be, for example, the average value of the brightness values ​​of each pixel that makes up the gray image D.

[0018] Referring to Figures 4 and 5, we will explain the process when using six types of weather distribution images F: a daily average temperature image F1a, a daily accumulated precipitation image F2a, a sunshine duration image F3a, a global solar radiation image F4a, a daily average wind speed image F5a, and a snow depth image F6a. Figures 4 and 5 show calculation images of the center of gravity G for each weather distribution image Fa. Figure 4(a) shows the calculation results for the daily average temperature image F1a, Figure 4(b) shows the calculation results for the daily accumulated precipitation image F2a, and Figure 4(c) shows the calculation results for the sunshine duration image F3a. Figure 5(a) shows the calculation results for the global solar radiation image F4a, Figure 5(b) shows the calculation results for the daily average wind speed image F5a, and Figure 5(c) shows the calculation results for the snow depth image F6a. As shown in FIG. 4(a), the coordinates of the center of gravity (X1, Y1), the distance L1 to the center of gravity G, and the statistical value R1 of the brightness value are calculated from the daily mean temperature image F1a. As shown in FIG. 4(b), the coordinates of the center of gravity (X2, Y2), the distance L2 to the center of gravity G, and the statistical value R2 of the brightness value are calculated from the daily accumulated precipitation amount image F2a. Furthermore, as shown in FIG. 4(c), the coordinates of the center of gravity (X3, Y3), the distance L3 to the center of gravity G, and the statistical value R3 of the brightness value are calculated from the sunshine duration image F3a. Also, as shown in FIG. 5(a), the centroid coordinates (X4, Y4), the distance L4 to the centroid G, and the luminance statistical value R4 are calculated from the global solar radiation image F4a. As shown in FIG. 5(b), the center of gravity coordinates (X5, Y5), the distance L5 to the center of gravity G, and the luminance value statistical value R5 are calculated from the daily average wind speed image F5a. Furthermore, as shown in FIG. 5(c), the center of gravity coordinates (X6, Y6), the distance L6 to the center of gravity G, and the luminance value statistics R6 are calculated from the snow depth image F6a.

[0019] The learning data acquisition unit 12 shown in FIG. 1 acquires the water level of the river measured at the observation station and the position of the center of gravity G (center of gravity coordinates (X G , Y G)), the distance L from the origin O to the center of gravity G, and the average brightness value in the gray image D are combined and organized as training data. The water level of the river measured at an observation station may be, for example, that published by the Ministry of Land, Infrastructure, Transport and Tourism. The training data acquisition unit 12 organizes the training data, for example, according to instructions from a person who predicts the water level (hereinafter referred to as the "user"). Note that the water level of the river, the position of the center of gravity G, the distance L to the center of gravity G, and the average brightness value may be registered in advance in a storage unit as a combination, and the training data acquisition unit 12 may acquire the registered information as training data. The storage unit may be included in a device connected via a communication line (for example, a cloud system). Hereinafter, a collection of organized training data may be referred to as a "training data group."

[0020] An example of the learning data group organized by the learning data acquisition unit 12 is shown in Fig. 6. The learning data group shown in Fig. 6 is composed of multiple pieces of learning data with different water level measurement times. Here, it is assumed that the maximum water level of a river is predicted on a daily basis. Note that prediction does not have to be performed on a daily basis, and prediction may be performed, for example, on an hourly basis. The training data set shown in Figure 6 includes information from 2008 to 2020, and the information for each year corresponds to the water level (for example, the maximum daily water level) at a river observation station from March to May. Each training data is composed of the "forecast date," the "forecasted maximum water level (daily)," and "information on meteorological items" for each day over several days (for example, "average daily temperature," "daily accumulated precipitation," "sunshine hours," "global solar radiation," "average daily wind speed," and "snow depth"). In Figure 6, the information on meteorological items is assumed to be information for three days: "d+1 (the next day)," "d (the day)," and "d-1 (the day before)," assuming that the forecast date (forecast date) is "d+1 (sometimes written as "the next day")." The information on meteorological items is calculated by dividing the "centroid coordinates (X G , Y G), "distance to the center of gravity L," and "average value R within the catchment area (related to meteorological items)." As the average value R within the catchment area shown in Figure 6, the average value of the brightness values ​​of each pixel that makes up the gray image D is registered, but it may also be an actually measured value. In this way, the learning data acquisition unit 12 acquires, as learning data, a set of the river water level (here, the maximum water level on a daily basis) at a certain first point in time (the prediction date (next day) in FIG. 6) and representative point information and meteorological item value information at at least one or more second points in time (the three days from the previous day to the next day in FIG. 6) that include a time period earlier than the first point in time. Note that in the prediction stage, the river water level (here, the maximum water level on a daily basis) is the target of prediction, and the representative point information and meteorological item value information are the data used for prediction.

[0021] The learning processing unit 13 shown in FIG. 1 has a learning device, and performs machine learning on the learning device using the learning data acquired by the learning data acquisition unit 12. The learning device is a learning system in machine learning, and can derive good results by comparing the results of classification, prediction, judgment, etc. based on given data with the actual results that are correct answers, and adjusting various parameters. The learning device is, for example, a neural network, and the present embodiment will be described assuming a neural network. The learning device is also referred to as a "learning model." The learning device may also be, for example, a device that uses deep learning or other machine learning methods (SVM, RF, etc.) other than a neural network. As is well known, neural networks are an information processing method developed with the aim of having computers imitate the workings of the human brain. Neural networks excel at complex nonlinear processing, and by using learning functions, they are characterized by the fact that there is no need to formulate the relationship between explanatory variables and target variables. Neural networks generally consist of an input layer, a hidden layer, and an output layer. Explanatory variables are input to the input layer, and target variables are output from the output layer. The hidden layer plays the role of relating the two. There are no restrictions on the number of hidden layers or the number of neurons in each hidden layer, and they can be set arbitrarily. For example, there could be three hidden layers, with the number of nodes in the hidden layer being twice the number of nodes in the input layer.

[0022] The method for training the learning device is not particularly limited, and the learning device can be trained, for example, by backpropagation. Specifically, the "centroid coordinates," "distance to the centroid," and "average values ​​within the catchment area for each weather item" for each day from the previous day to the next day that constitute the training data are input to the input layer, and the intermediate layer is adjusted based on the error between the result output from the output layer and the "water level" for the next day, which is the predicted date. In other words, the learning processing unit 13 inputs representative point information and weather item value information for at least one or more third time points and trains the learning device to output the river water level at a fourth time point, which is after the third time point. The number of training data to be trained is not particularly limited, and learning is terminated, for example, when the desired accuracy is reached.

[0023] (Configuration of prediction device) 1, the prediction device 20 includes an image processing unit 21, a prediction data acquisition unit 22, and a prediction processing unit 23. The image processing unit 21, the prediction data acquisition unit 22, and the prediction processing unit 23 are realized by program execution processing by a CPU (Central Processing Unit), a dedicated circuit, etc. When these functions are realized by program execution processing, a program for realizing the function is stored in a storage unit (not shown). A weather distribution image Fb is input to the image processing unit 21. There is no particular limit to the number of weather distribution images Fb input to the image processing unit 21; for example, weather distribution images Fb for each day or each hour may be input. The image processing unit 21 obtains representative point information and weather item value information from the weather distribution image Fb. The processing of the image processing unit 21 is similar to the processing of the image processing unit 11 in the learning stage, so a detailed description will be omitted. The weather distribution image Fb may be of the past (for example, an actual measurement) or of the future (for example, a forecast). In other words, it is possible to predict water levels one day or several hours from now (for example, the present or near future) from an actual past weather distribution image Fb, and it is also possible to predict water levels one day or several hours from a future weather distribution image Fb published as a forecast.

[0024] The predicted data acquisition unit 22 shown in FIG. 1 acquires the position of the center of gravity G (center of gravity coordinates (X G , Y G )), the distance L from the origin O to the center of gravity G, and the average brightness value in the gray image D are combined and organized as predicted data. Note that the position of the center of gravity G, the distance L to the center of gravity G, and the average brightness value may be registered in advance in a storage unit, and the predicted data acquisition unit 22 may acquire the registered information as predicted data. The storage unit may be included in a device connected via a communication line (for example, a cloud system). It is desirable that the predicted data is calculated using a weather distribution image Fb in the same format as the weather distribution image Fa. In this way, the predicted data acquisition unit 22 acquires, as learning data, pairs of representative point information and weather item value information for at least one or more third time points (in this embodiment, for three days from the previous day to the next day (prediction day)). The prediction processing unit 23 shown in Fig. 1 has a trained learning device. The trained learning device is a learning device that has been machine-learned by the learning device 10. In other words, the trained learning device has been trained to output the water level of the river at a fourth time point that is after the third time point by inputting representative point information and meteorological item value information for at least one or more third time points. The prediction processing unit 23 inputs the prediction data acquired by the prediction data acquisition unit 22 into the trained learning device, thereby predicting the water level of the river at the fourth time point that is after the third time point and that is included in the prediction data, and outputs the prediction result.

[0025] <Regarding the learning method and prediction method according to the first embodiment> The learning method and prediction method according to the first embodiment will be described with reference to Fig. 7 (and also with reference to Figs. 1 to 6 as appropriate). Fig. 7 is a flowchart illustrating the learning method and prediction method according to the first embodiment. 7, the process of constructing a prediction model (step S10) corresponding to the learning method according to the first embodiment mainly includes a process of preparing input data (past) (step S11), a parameter tuning process (step S12), and a process of determining prediction accuracy (step S13). Also, the process of predicting water level (step S20) corresponding to the prediction method according to the first embodiment mainly includes a process of preparing input data (real-time) (step S21), a calculation process using a prediction method (step S22), a confirmation process of prediction results (step S23), and a process of providing the results to construction personnel (step S24).

[0026] (Preparation process of input data (past) (step S11)) The user prepares a weather distribution image Fa and measured river water level information to be used in training the prediction model (learner). The user prepares, for example, a daily or hourly weather distribution image Fa and river water level information. The river water level is the water level at the location corresponding to the point to be predicted, and may be published by, for example, the Ministry of Land, Infrastructure, Transport and Tourism. The weather distribution image Fa may be published by, for example, the Japan Meteorological Agency or the National Agriculture and Food Research Organization, and since the information published by the Japan Meteorological Agency and the National Agriculture and Food Research Organization covers the entire country of Japan, the weather distribution image Fa can be prepared without much effort or time. The image processing unit 11 of the learning device 10 calculates the representative point information and meteorological item value information from the weather distribution image Fa (see FIG. 3(a)) using the above-mentioned method. The representative point information is, for example, the position of the center of gravity G of the watershed K2 (center of gravity coordinates (X G , Y G )) and the distance L from the reference point (here, the origin O) to the center of gravity G. The meteorological item value information is, for example, the average brightness value of each pixel that constitutes the watershed K2. The learning data acquisition unit 12 acquires the water level of the river measured at the observation station and the position of the center of gravity G (center of gravity coordinates (X G , Y G )), the distance L from the origin O to the center of gravity G, and the average brightness value in the gray image D are combined and organized as training data. The organized training data is stored in a storage unit (not shown).

[0027] (Parameter tuning process (step S12)) The learning data acquisition unit 12 of the learning device 10 sequentially acquires learning data, and the learning processing unit 13 uses the acquired learning data to train a learning device. If the learning device is a neural network, the intermediate layer is adjusted using the learning data. The learning processing unit 13 trains the learning device by machine learning so that the learning device outputs the river water level at a fourth time point, which is after the third time point, by inputting representative point information and meteorological item value information at at least one third time point. (Prediction accuracy determination process (step S13)) In this step, the predictive accuracy of the learned learning device is determined, and if the predictive accuracy meets the standard set by the user, the learning step is terminated (if the predictive accuracy is "good"). On the other hand, if the predictive accuracy of the learning device does not meet the set standard (if the predictive accuracy is "poor"), the parameter tuning step (step S12) is continued, and the learning step continues. Note that the parameter tuning step (step S12) and the predictive accuracy determination step (step S13) may be realized by cross-validation. When learning is performed using cross-validation, for example, data from the organized data for the year to be predicted is used as test data, and the rest is classified as learning data to predict the water level for each year. Details will be described later.

[0028] (Input data (real-time) preparation process (step S21)) The user prepares a weather distribution image Fb to be used for predicting the water level of a river. The user prepares, for example, a daily or hourly weather distribution image Fb. The weather distribution image Fb may be one published by, for example, the Japan Meteorological Agency or the National Agriculture and Food Research Organization. The image processing unit 21 of the prediction device 20 calculates representative point information and meteorological item value information from the weather distribution image Fb in the same manner as in the preparation process of step S11. Then, the prediction data acquisition unit 22 calculates the position of the center of gravity G (center of gravity coordinates (X G , Y G )), the distance L from the origin O to the center of gravity G, and the average brightness value in the gray image D are combined and organized as predicted data. (Calculation process using prediction method (step S22)) The prediction data acquisition unit 22 of the prediction device 20 acquires the prediction data, and the prediction processing unit 23 predicts the water level of the river by inputting the prediction data into a trained learning device. The trained learning device has been trained to output the water level of the river at a fourth time point, which is after the third time point, by inputting representative point information and meteorological item value information at at least one or more third time points. (Step S23 for confirming the prediction results and step S24 for providing the results to the construction personnel) The user checks the water level prediction result calculated in step S22 and provides the information on the predicted water level to people involved in the construction work, if necessary. Based on the provided predicted water level, the people involved in the construction work can evacuate heavy machinery and materials in the construction yard or take measures such as protecting them.

[0029] <Effects of the learning device and prediction device according to the first embodiment> As described above, the learning device 10 and the prediction device 20 according to the first embodiment can predict water levels for various snowmelt patterns by using the weather distribution image F for learning and prediction. In addition, the weather distribution image F published by the Japan Meteorological Agency, National Agriculture and Food Research Organization, and other organizations covers the entire country, making it possible to predict water levels at any point. In addition, weather distribution image F is easy to obtain, and the use of a neural network as a learning device reduces the effort required for parameter tuning, so the effort and time required to develop a water level prediction method and collect the necessary data is reduced compared to conventional methods. In addition, by using weather distribution images F published by the Japan Meteorological Agency and the National Agriculture and Food Research Organization, it is possible to predict water levels long into the future, allowing for ample time to evacuate and protect heavy machinery and materials in the construction yard.

[0030] To verify the effectiveness of the learning device 10 and prediction device 20 according to the first embodiment, prediction tests using actual data were conducted, which will now be described. The prediction tests were conducted in two cases: one using all six types of weather distribution images F (first prediction test) and the other using one of the six types of weather distribution images F (second prediction test). The weather distribution image F cited information on six types of meteorological items related to snowmelt from the Mesh Agricultural Weather Data System provided by the National Agriculture and Food Research Organization. In the prediction test, the water level at Tsukigata Observatory on the Ishikari River was predicted using weather distribution image F with deep learning (neural network (NN)), a machine learning method. The Ishikari River is the river indicated by symbol K1 in Figure 2, and Tsukigata Observatory is located at symbol K4 in Figure 2. The catchment area of ​​the Ishikari River is approximately 8,900 km 2 As shown in Figure 2, Tsukigata Observatory is located downstream of the Ishikari River.

[0031] (First predictive test) A prediction test was conducted using 13 years of data from 2008 to 2020. The water level changes at Tsukigata Observatory are shown in Figure 8, for example. Figure 8 is a graph showing the water level changes at Tsukigata Observatory from 2017 to 2019. The flood danger level at Tsukigata Observatory is 15.6 m, and the water level without flooding is approximately EL + 5 to 6 m (EL: observation station water level). Although not shown, the maximum water level from 2010 to 2020 was EL + 12.73 m (2:00 p.m. on September 3, 2011). As shown in Figure 8, during the snowmelt season (March to May), the water level of the river rises as the snow melts. In this invention, the water level at Tsukigata Observatory during the snowmelt season (March to May) is predicted using weather distribution image F for that season.

[0032] The procedure for conducting the prediction test is shown in Figure 9. Figure 9 is a flowchart showing the procedure for conducting the prediction test. In the prediction test, daily weather distribution images F from March to May 2008 to 2020 are used. First, data within the watershed (catchment area) is extracted from six types of weather distribution images F related to snowmelt (daily mean temperature image F1, daily accumulated precipitation image F2, sunshine duration image F3, global solar radiation image F4, daily mean wind speed image F5, and snow depth image F6) (step S121), and the RGB values ​​of each pixel in the extracted area are converted to brightness values ​​between 0 and 255 (step S122). This creates six types of grayscale images D on a daily basis. Next, the distance L to the center of gravity G and the brightness statistic R are calculated for the six types of grayscale images D (step S123). Next, the data is organized by combining the daily maximum water levels at Tsukigata Observatory from March to May 2008 to 2020 with representative point information and meteorological item value information for three days (the day before, the day of, and the day after the observation of the daily maximum water level, assuming that the day on which the daily maximum water level was observed is the following day) (step S124). Using representative point information and meteorological item value information for three days allows for consideration of the amount and direction of change in meteorological distribution. Note that the reason for not using the weather distribution image F itself for learning is that if all the data (all pixel values) of the weather distribution image F were used, the number of data points would amount to tens of thousands to hundreds of thousands per image, depending on the image resolution, which would result in issues such as neural network coefficients (weights) not converging and requiring a significant amount of calculation time.

[0033] Next, among the organized data, data for the year to be predicted was used as test data, and the rest was classified as learning data, and cross-validation was performed to predict water levels for each year from 2008 to 2020 (steps S125 to S126). The relationship between test data and learning data in cross-validation is shown in Fig. 10. As shown in Fig. 10, for example, if the year to predict water levels is 2008, the data for 2008 is used as test data, and the remaining years from 2009 to 2020 are classified as learning data. As an example of the results of the first prediction test, a comparison of the measured water levels from 2018 to 2020 and the predicted water levels from the prediction test is shown in Figures 11 to 13. In Figures 11 to 13, the measured water levels are shown in thick lines, and the predicted water levels are shown in thin lines. As shown in Figures 11 to 13, when the actual water levels rose and fell, the predicted water levels also rose and fell, which means that it was possible to predict water level fluctuations (for example, determining whether there were large-scale water level fluctuations). Note that, as shown by the two-dot chain line in the figures, there were cases where water level differences of more than 1 meter occurred.

[0034] (Second predictive test) In the second forecast test, a single type of weather distribution image F was used to make predictions. As an example of the results of the second forecast test, Figure 14 shows a comparison of the measured water levels in April 2015 and April 2018 with the predicted water levels in the forecast test. In Figure 14, the measured water levels are shown with thick solid lines, and the predicted water levels for each item are shown with thin lines using different line types and symbols. Figure 14 also shows the predictions using all six types of weather distribution image F (i.e., the results of the first forecast test) with intermediate solid lines. As shown in Figure 14, even when one type of weather distribution image F was used, as in the first prediction test, the predicted water level rose and fell when the actual water level rose and fell, and it can be said that the water level fluctuation situation (for example, determining whether there was a large-scale water level fluctuation) was able to be predicted.

[0035] Furthermore, in river construction work, temporary cofferdams are often set up within the river as work proceeds, and it is sometimes necessary to predict whether the river's water level will exceed a danger level set based on the temporary cofferdam (for example, set at the top or below the top of the cofferdam). For this reason, from the perspective of applicability to river construction work, the number of events involving flooding exceeding the danger level and the absolute value of the difference in water level between the measured water level and the predicted water level are organized and verified. Here, the danger level is assumed to be equivalent to the height of the top of the temporary cofferdam, and is set at "EL+8m." Figure 15 is a diagram used to explain the results of the verification of applicability to river engineering works. (a) is a graph showing the total number of correct, missed, and missed events for each weather item, and their proportions. (b) is a graph showing the number of missed events for each weather item and the proportion of the absolute value of the difference in water level between the measured and predicted water levels. (c) is a graph showing the number of missed events for each weather item and the proportion of the absolute value of the difference in water level between the measured and predicted water levels. "Correct" events are cases where both the measured and predicted water levels exceeded the danger level. "Missed" events are cases where the measured water level exceeded the danger level but the predicted water level did not. "Missed" events are cases where the measured water level did not exceed the danger level, but the predicted water level did.

[0036] When predicting river construction, keeping the number of "oversights" low is important from the perspective of safety management. Looking at Figure 15(a), we can see that predictions using one meteorological item had 12 to 17 "oversights," while predictions using six meteorological items had the best result of only 9 "oversights." The reason that predictions using six meteorological items had more "correct answers" (fewer "oversights") is thought to be because the predicted water level was higher due to the increased number of items that responded compared to predictions using one meteorological item. As shown in Figure 15(b), the most common "misses" were for water level differences of 0.5-1.0 m for each forecast result, and the percentage of "misses" that were close to the "correct answer" was small. Also, as shown in Figure 15(c), the number of "misses" was the highest for forecasts using six meteorological items. This is thought to be due to the same reason as the high number of "correct answers" for forecasts using six meteorological items.

[0037] [Second embodiment] In the first embodiment, as shown in the results of the prediction test, the rising and falling movements of the predicted water level were similar to those of the measured water level, demonstrating the feasibility of water level prediction using the weather distribution image F. It was also possible to predict whether or not large-scale water level fluctuations would occur. However, there were cases where the water level difference between the measured and predicted values ​​was 1 meter or more (see the area indicated by the two-dot chain lines in Figures 11 to 13). The cause of this difference was thought to be that the input information did not include a reference value for the water level at the prediction location. Therefore, in the second embodiment, past water levels (e.g., the maximum water level on the day) at the Tsukigata Observatory, which is the prediction location, are added to the learning data and prediction data. Note that the configuration, other than the content of the learning data and prediction data, is the same as in the first embodiment. Therefore, only the differences will be described below.

[0038] FIG. 16 shows an example of a learning data group organized by the learning data acquiring unit 12 in the second embodiment. The learning data group shown in FIG. 16 is composed of multiple learning data with different water level measurement times. The learning data group shown in FIG. 16 includes information from 2008 to 2020, and the information for each year corresponds to the water level (e.g., daily maximum water level) at a river observation station from March to May. Reference past water level information (in FIG. 16, the daily maximum water level on the day before the prediction date (i.e., the day)) is newly added to each learning data. As a result, each learning data in the second embodiment is composed of the "prediction date," the "predicted maximum water level (daily)," the "daily maximum water level on the day before the prediction date," and "information on weather items (e.g., "daily average temperature," "daily accumulated precipitation," "sunshine hours," "global solar radiation," "daily average wind speed," and "snow depth")" for each day over several days. In FIG. 16, the information about the weather items is assumed to be information for three days, "d+1 (next day)", "d (current day)", and "d-1 (previous day)", assuming that the predicted date (prediction date) is "d+1 (sometimes written as "next day")". The information about the weather items is calculated by dividing the "center of gravity coordinates (X G , Y G ) "distance to the center of gravity L" and "average value R within the catchment area (for meteorological items)".

[0039] In this way, the learning data acquisition unit 12 in the second embodiment acquires as learning data a set of the water level of the river at a certain first point in time (the predicted date (next day) in Figure 16), the water level at a fifth point in time that is earlier than the first point in time and that is included in the time period affected by water level changes due to snowmelt (the daily maximum water level on the day before the predicted date (i.e., the day) in Figure 16), and representative point information and weather item value information at at least one or more second point in time (the three days from the previous day to the next day in Figure 16) that is earlier than the first point in time. The trained learning device, which is trained using training data including the water level at the fifth time point, is trained to output the water level of the river at the fourth time point by inputting representative point information and meteorological item value information for at least one or more third time points and the water level at a sixth time point corresponding to the fifth time point. The "sixth time point corresponding to the fifth time point" is the time point at which the time from the sixth time point to the fourth time point is equal to the time from the fifth time point to the first time point.

[0040] <Effects of the learning device and prediction device according to the second embodiment> To verify the effects of the learning device 10 and prediction device 20 according to the second embodiment, a prediction test using actual data was conducted, which will now be described. The prediction test in the second embodiment was similar to that in the first embodiment, and using the Tsukigata Observatory on the Ishikari River as an example, the water level at Tsukigata Observatory was predicted by deep learning (neural network (NN)), a machine learning technique, using six types of weather distribution image F. The Ishikari River is the river indicated by symbol K1 in Figure 2, and the Tsukigata Observatory is located at the position indicated by symbol K4 in Figure 2. The catchment area of ​​the Ishikari River is approximately 8,900 km 2 As shown in Figure 2, Tsukigata Observatory is located downstream of the Ishikari River. As an example of the results of the prediction test in the second embodiment, a comparison of the measured water levels from 2018 to 2020 and the predicted water levels in the prediction test is shown in Figures 17 to 19. Figures 17 to 19 show both the results when the measured water levels were not used during prediction (as in the first embodiment) and the results when the measured water levels were used. In Figures 17 to 19, the measured water levels are shown in bold lines, the predicted water levels when the measured water levels were not used during prediction are shown in thin lines, and the predicted water levels when the measured water levels were used during prediction are shown in dotted lines. Looking at Figures 17 to 19, it can be seen that the water level difference between the measured values ​​and the predicted values ​​is smaller than in the results of the prediction test in the first embodiment. FIG. 20 shows a comparison of the difference between the measured water level and the predicted water level in the first and second embodiments. In FIG. 20, the average and standard deviation of the absolute value of the difference between the measured water level and the predicted water level are shown as "yes" or "no" for the measured water level at the time of prediction. As shown in FIG. 20, by using the measured water level at the time of prediction, both the average and standard deviation of the absolute value of the difference between the measured water level and the predicted water level decreased. In other words, it was confirmed that using the measured water level at the time of prediction improves prediction accuracy.

[0041] The learning device 10 and the prediction device 20 according to the second embodiment described above can also achieve substantially the same effects as those of the first embodiment. Furthermore, in the second embodiment, the water level can be predicted with higher accuracy by including past water levels that serve as a reference at the prediction point in the learning data.

[0042] [Third embodiment] The contribution of meteorological factors related to snowmelt may vary depending on the region and time of year. For example, there are known physical formulas (such as the heat balance method) for calculating the amount of snowmelt. These formulas take into account the heat balance of snowmelt-related factors and calculate the amount of snowmelt by multiplying the related meteorological factors by a coefficient. Therefore, in the third embodiment, the contribution of a weather item to a water level prediction is set as a weight, and the learning data and prediction data are weighted (for example, multiplied by the weight) to predict the water level of a river. The weight value used for weighting may be calculated independently of the weather distribution image F, or may be calculated based on the weather distribution image F. In this embodiment, a case where the weight value is calculated independently of the weather distribution image F will be described as an example.

[0043] <How weights are calculated and how they are used for learning and prediction> In this embodiment, weighting is performed using multiple regression analysis. In multiple regression analysis, the relationship between the objective variable and the explanatory variables is modeled as in equation (3), and the explanatory variables (x in equation (3)) are i ) to construct an equation to find the objective variable (y in equation (3)). i corresponds to the weight. ·y=w0+w1x1+w2x2+w3x3+···+w n x n ...Equation (3) Multiple regression analysis is a method for analyzing multiple x i Using the combination of and y, find the w that is closest to y i The value of the explanatory variable x is calculated using a mathematical method (such as Lagrange's method of undetermined multipliers). i is the weather item value, the objective variable y is the river water level, and the weight of the weather item is w i An example of applying six meteorological parameters related to snowmelt to equation (3) is shown in equation (4). Equation (4) is used to analyze the relationship between river water levels and the six meteorological parameters using multiple regression analysis. y = w0 + (w1 × daily mean temperature) + (w2 × daily accumulated precipitation) + (w3 × sunshine hours) + (w4 × total solar radiation) + (w5 × daily mean wind speed) + (w6 × snow depth) Equation (4)

[0044] The river water level and meteorological item values ​​used in the multiple regression analysis (i.e., information used to calculate weight values) are preferably those within the catchment area of ​​the river whose water level is to be predicted, and are preferably observation data from the prediction point where the river water level is predicted or a point upstream of the prediction point. Furthermore, the river water level and meteorological item values ​​used in the multiple regression analysis are preferably those for a period related to the learning data used to learn the river water level and the prediction data used for the prediction, and are preferably those for a time point corresponding to the learning data and the prediction data. Furthermore, the types and number of meteorological items used in the multiple regression analysis are not particularly limited, but are preferably meteorological items corresponding to the learning data and the prediction data. In this embodiment, weight values ​​are calculated using multiple regression analysis, but weight values ​​may also be set using methods other than multiple regression analysis. Weight values ​​may be set not only based on mathematical calculations but also, for example, based on experience. During learning and prediction, weighting is performed on one or both of the representative point information and the weather item value information using a weight value, and learning or prediction is performed using the weighted information. For example, the learning processing unit 13 multiplies one or both of the representative point information and the weather item value information included in the learning data by a weight value, and performs learning using the information after multiplication by the weight value. Furthermore, the prediction processing unit 23 multiplies one or both of the representative point information and the weather item value information included in the prediction data by a weight value, and performs prediction using the information after multiplication by the weight value. Note that the representative point information and the weather item value information may be weighted directly, or the representative point information and the weather item value information may be weighted indirectly by weighting the weather distribution image F.

[0045] Next, we will explain an example of how to calculate weights using multiple regression analysis. Data from the Ishikari River period "2007 to March 1st to May 31st, 2021" is targeted, and the weight values ​​in equation (4) are calculated using data from this period. The water level at Tsukigata Observatory is used as the dependent variable y. The water level at Tsukigata Observatory is obtained from the water gate water quality database of the Ministry of Land, Infrastructure, Transport and Tourism every hour. The water level change at Tsukigata Observatory is shown in Figure 8, for example. The explanatory variable x iFor the weather parameter values, we use the values ​​of meteorological parameters from the Asahikawa Observatory, which is 75 km upstream from the Tsukigata Observatory. For the weather parameter values ​​from the Asahikawa Observatory, we use the corresponding observation items from the AMeDAS (Automated Meteorological Data Acquisition System) observations. There are multiple AMeDAS observation items, and hourly observation values ​​can be obtained. In this example, we focus on six types of weather parameters (temperature, precipitation, sunshine hours, global solar radiation, wind speed, and snow depth) and obtain their hourly observation values.

[0046] Next, daily values ​​are calculated from the hourly observations. The maximum daily water level from 0:00 to 23:00 is calculated from the hourly observations. The average daily temperature and wind speed values ​​from 0:00 to 23:00 are calculated from the hourly observations. The daily integrated values ​​from 0:00 to 23:00 for precipitation, sunshine hours, and global solar radiation are calculated from the hourly observations. The maximum daily snow depth value from 0:00 to 23:00 is calculated from the hourly observations. Figure 21(a) shows the dependent variables and explanatory variables calculated using the method described above. Figure 21(a) shows an example of dependent variables and explanatory variables used in multiple regression analysis. Finally, the relationship between the daily maximum water level for the following day (called "d+1") and six meteorological parameters for that day (called "d") is analyzed using program code that performs multiple regression analysis to determine weights. Specifically, the explanatory variables are the daily mean temperature, daily accumulated precipitation, sunshine hours, global solar radiation, daily mean wind speed, and snow depth for "d (that day)," and the dependent variable is the water level for "d+1 (the following day)." Using a combination of explanatory and dependent variables for a total of approximately 1,380 days, multiple regression analysis is performed to determine weights w1 through w6 (daily mean temperature: w1, daily accumulated precipitation: w2, sunshine hours: w3, global solar radiation: w4, daily mean wind speed: w5, and snow depth: w6). The weights determined using the method described above are shown in Figure 21(b). FIG. 21(b) is an example of weight values ​​obtained by multiple regression analysis, where the weight for snow depth is the maximum value "0.417" and the weight for daily average wind speed is the minimum value "0.059".

[0047] <Effects of the learning device and prediction device according to the third embodiment> In order to verify the effect of the third embodiment, a prediction test was conducted using the weight values ​​shown in FIG. 21(b) and actual data, which will now be described. The prediction test in the third embodiment is similar to that in the first and second embodiments, and using the Tsukigata Observatory on the Ishikari River as an example, the water level at Tsukigata Observatory was predicted using six types of weather distribution images F by deep learning (neural network (NN)), which is a machine learning technique. The weighting was performed based on representative point information (the position of the center of gravity G of the watershed (center of gravity coordinates (X G , Y G ) and the distance L from the reference point (here, the origin O) to the center of gravity G, as well as meteorological item value information (average watershed R). As an example of the results of the prediction test in the third embodiment, a comparison of the measured water levels (daily maximum water levels) from 2018 to 2020 and the predicted water levels in the prediction test is shown in Figures 22 to 27. 22 to 24 show both the results when the measured water level is used but the weight is not used during prediction (the same as in the second embodiment) and the results when the measured water level is not used during prediction but the weight is used during prediction. In Figures 22 to 24, the daily maximum measured water level is shown in bold, the daily maximum predicted water level when the measured water level is used but the weight is not used during prediction is shown in dotted line, and the daily maximum predicted water level when the measured water level is not used during prediction but the weight is used during prediction is shown in dashed dot line. 25 to 27 show both the results when the measured water level is used but no weight is used during prediction (as in the second embodiment) and the results when both the measured water level and weight are used during prediction. In Figures 25 to 27, the daily maximum measured water level is shown in bold, the daily maximum predicted water level when the measured water level is used but no weight is used during prediction is shown in dotted line, and the daily maximum predicted water level when both the measured water level and weight are used during prediction is shown in chain double-dashed line. 22 to 27, it can be seen that the prediction results using the weights in the third embodiment (graphs indicated by dashed and dotted lines) tend to be closer to the actual measured values ​​than the results of the prediction test in the second embodiment where the weights were not used (graph indicated by the dotted line).

[0048] Furthermore, Fig. 28 shows a comparison of the differences between the measured water level and the predicted water level in the first to third embodiments. Fig. 28 shows the average and standard deviation of the absolute values ​​of the differences between the measured water level and the predicted water level, divided into cases of "yes" or "no" measured water level at the time of prediction and "yes" or "no" weighting. In other words, Fig. 28 is a quantitative summary of the differences between the measured water level and the predicted water level in Figs. 22 to 27. As shown in Figure 28, when the average and standard deviation of the differences were calculated, the difference when the weights of meteorological items were taken into account was smaller than when the weights were not taken into account. This result confirmed that taking the weights of meteorological items into account improves prediction accuracy. Furthermore, when comparing the prediction results using measured water levels without taking weights into account (prediction results according to the second embodiment) with the prediction results using weights but without using measured water levels (prediction results according to the third embodiment), the latter prediction results were equal to or better than the former prediction results. As can be seen from this, even if a weather distribution image F for one week ahead is obtained from a weather forecast, even if a prediction is made using only the weather distribution image F, by taking the weights of meteorological items into account, prediction results equivalent to those using measured water levels can be obtained. Since measured water levels are not required, water levels can be predicted earlier. Furthermore, advance preparation is simple and more accurate predictions can be achieved using information that is easily obtained. This is effective for safety management, such as early evacuation.

[0049] The third embodiment described above can also achieve substantially the same effects as the first embodiment. Furthermore, in the third embodiment, the contribution of meteorological items to the water level prediction result is used as a weight, thereby making it possible to predict the water level with higher accuracy. Although the embodiment of the present invention has been described above, the present invention is not limited to this and can be practiced within the scope of the claims.

[0050] In the first to third embodiments, a representative point in the upstream part of the catchment basin is obtained, but a representative point in the entire catchment basin or in another range of the catchment basin may also be obtained. In other words, the representative point may reflect the intensity distribution of meteorological items within an area including at least a part of the catchment basin of the river, and the representative point information is information reflecting the intensity distribution of meteorological items within an area including at least a part of the catchment basin of the river. The same applies to the meteorological item value information. Alternatively, the watershed may be divided and a representative point for each divided area may be determined. The method for dividing the watershed is not particularly limited, and for example, the watershed may be divided equally, or if a river branches, the area may be divided for each branching river. In this case, the training data may include representative point information for each divided area, or information determined from each representative point (for example, information obtained by adding or averaging each representative point) may be included in the training data.

[0051] Furthermore, in the first to third embodiments, six types of weather distribution images F (daily mean temperature image F1, daily accumulated precipitation image F2, sunshine duration image F3, global solar radiation image F4, daily mean wind speed image F5, and snow depth image F6) are shown, but other images may also be used. For example, an image showing the temperature distribution of the earth's surface or an image showing the distribution of snow density may also be used. Furthermore, a new weather distribution image F may be created by processing the weather distribution image F, and representative point information and weather item value information may be calculated from the new weather distribution image F. Furthermore, a new weather distribution image F may be created by mathematically calculating multiple weather distribution images F, and representative point information and weather item value information may be calculated from the new weather distribution image F. For example, a new weather distribution image F may be created by adding or multiplying the daily mean temperature image F1 and the daily accumulated precipitation image F2, or by subtracting or dividing the daily mean temperature image F1 from the daily accumulated precipitation image F2. In the first to third embodiments, a total of six representative points are calculated from each of the six weather distribution images F, and the positions of the six representative points and the distances from the reference point to the representative points are used as representative point information for learning. However, values ​​obtained by processing the six representative points may also be used as the representative point information. This processing is not limited to obtaining six new pieces of representative point information from the six representative points, but also includes obtaining a different number of pieces of representative point information (for example, one). [Explanation of symbols]

[0052] 10 Learning Device 11 Image processing section 12 Learning data acquisition unit 13 Learning processing unit 20 Prediction Device 21 Image processing section 22 Prediction data acquisition section 23 Prediction processing unit F, Fa, Fb weather distribution image

Claims

1. a learning data acquisition unit that acquires, as learning data, a set of a water level of the river at a first time point during a snowmelt period when snow is present in at least a part of the catchment area of ​​the river, and representative point information and meteorological item value information at at least one or more second time points that are earlier than the first time point; a learning processing unit having a learning device and causing the learning device to perform machine learning using the learning data; The representative point information is information about a representative point that reflects an intensity distribution of meteorological items within an area that includes at least a part of the watershed, the weather item value information is information about weather item values ​​in the range, the learning data acquisition unit acquires learning data including the representative point information and meteorological item value information of meteorological items such as daily average temperature, daily accumulated precipitation, sunshine hours, global solar radiation, daily average wind speed, and snow depth; the learning processing unit trains the learning device to output the water level of the river at a fourth time point that is after the third time point by inputting representative point information of daily average temperature, daily accumulated precipitation, sunshine hours, global solar radiation, daily average wind speed, and snow depth at at least one or more third time points and meteorological item value information; A learning device characterized by:

2. an image processing unit that calculates the representative point information and the meteorological item value information from a meteorological distribution image that shows the intensity distribution of the meteorological items by changing colors; 2. The learning device according to claim 1 .

3. The image processing unit converting the RGB values ​​of each pixel constituting the weather distribution image at the second time point into a luminance value, calculating the center of gravity of the range when the luminance value is regarded as a weight as the representative point, and obtaining the position of the center of gravity and the distance from a predetermined reference point to the center of gravity as the representative point information; a statistical value calculated from the brightness value of each pixel in the range at the second time point as the weather item value information; 3. The learning device according to claim 2.

4. the image processing unit processes the weather distribution image to create a new weather distribution image, and calculates the representative point information and the weather item value information from the created new weather distribution image.

4. The learning device according to claim 2 or 3.

5. the learning data acquisition unit acquires learning data further including a water level at a fifth time point that is earlier than the first time point and is included in a time period affected by a water level change due to snowmelt, the learning processing unit inputs representative point information and meteorological item value information at at least one third time point and a water level at a sixth time point corresponding to the fifth time point, thereby causing the learning device to learn to output the water level of the river at the fourth time point; 5. The learning device according to claim 1, wherein the learning device is a learning device for learning a plurality of learning operations.

6. the learning processing unit weights one or both of the representative point information and the meteorological item value information of the daily average temperature, daily accumulated precipitation, sunshine hours, global solar radiation, daily average wind speed, and snow depth; 6. The learning device according to claim 1, wherein the learning device is a learning device for learning a plurality of learning operations.

7. The weight values ​​used by the learning processing unit for weighting are obtained by multiple regression analysis using water level as the objective variable and meteorological item values ​​as the explanatory variables.

7. The learning device according to claim 6.

8. a learning data acquisition step of acquiring, as learning data, a set of the water level of the river at a first time point during a snowmelt period when snow is present in at least a part of the catchment area of ​​the river, and representative point information and meteorological item value information at at least one or more second time points including information that is past the first time point; a learning process step of performing machine learning on a learning device using the learning data, The representative point information is information about a representative point that reflects an intensity distribution of meteorological items within an area that includes at least a part of the watershed, the weather item value information is information about weather item values ​​in the range, In the learning data acquisition step, learning data including the representative point information and the meteorological item value information of meteorological items such as daily average temperature, daily accumulated precipitation, sunshine hours, global solar radiation, daily average wind speed, and snow depth is acquired; In the learning process, the learning device is trained to output the water level of the river at a fourth time point that is after the third time point by inputting representative point information of daily mean temperature, daily accumulated precipitation, sunshine hours, global solar radiation, daily mean wind speed, and snow depth at at least one or more third time points and meteorological item value information. A learning method characterized by:

9. a trained learning device that has been trained by the learning method according to claim 8, and that predicts the water level of the river at the fourth time point by inputting representative point information and meteorological item value information at the third time point into the trained learning device; A prediction device characterized by:

Images