JP7265915B2 - Tsunami height and tsunami arrival time prediction system - Google Patents

Description

The present invention relates to a tsunami height and tsunami arrival time prediction system.

In recent years, a tsunami observation system using ocean radar has been proposed (see Patent Document 1). Marine radar can measure ocean currents, waves, etc. in a wide area of about 100 km by irradiating radio waves to the sea surface from an antenna installed on land, receiving backscattered waves due to waves on the sea surface, and analyzing the frequency. The marine radar can only measure the sea surface velocity component in the line-of-sight direction of the radio waves emitted from the antenna, and cannot directly measure the wave height of an incoming tsunami.

In Patent Document 1, the radio wave irradiation area of the marine radar is divided into a plurality of areas, and a tsunami is detected based on a beat signal having a frequency difference between the transmission signal (radio wave irradiation signal) and the reception signal (reflected wave) of the marine radar. to detect the current velocity on the surface of the sea. In Patent Document 1, a tsunami wave height at each time is calculated by executing a tsunami behavior simulation for each measurement sample period based on the calculated sea surface flow velocity of the tsunami. In Patent Document 1, the tsunami arrival time and tsunami height are predicted based on the arrival wave height (tsunami height) of the tsunami at each time.

JP 2016-85206 A

By the way, in this tsunami observation system, a simulation of tsunami behavior is performed for each measurement sample period based on the calculated sea surface velocity of the tsunami, and processing such as calculating the wave height of the tsunami at each time is performed. There is a problem that the calculation takes a certain amount of time, that is, the prediction time until the tsunami arrival time and tsunami height are calculated becomes long. Also, since the calculation is complicated, errors are likely to occur.

An object of the present invention is to solve the above problems and provide a tsunami height and tsunami arrival time prediction system capable of shortening the calculation time of the tsunami height and tsunami arrival time at the tsunami arrival point and reducing the occurrence of errors. be.

In order to solve the above problems, the tsunami height and tsunami arrival time prediction system of the present invention learns in advance a tsunami flow velocity distribution image for learning, and prepares a measured tsunami flow velocity distribution image when the tsunami reaches a predetermined point. A tsunami height and tsunami arrival time prediction system with a convolutional neural network section that classifies combinations of tsunami arrival times and tsunami heights, wherein three measured tsunami velocity distributions are measured in succession over time based on ocean radar observations. A flow velocity distribution image generation unit for generating an image, wherein the convolutional neural network unit extracts each feature quantity of the three measured tsunami flow velocity distribution images, and based on the result of combining all the feature quantities , Among the three measured tsunami current velocity distribution images, the latest measured tsunami current velocity distribution image is classified into a combination of tsunami arrival time and tsunami height when reaching the predetermined point.

Also, the three measured tsunami current velocity distribution images may be taken at intervals of one minute.
Further, the predetermined point may be an installation point of a transmitting/receiving antenna of the marine radar.
Further, it is preferable that the three measured tsunami current velocity distribution images are input to three RGB channels of the convolutional neural network unit.

ADVANTAGE OF THE INVENTION According to this invention, it is effective in shortening the calculation time of the tsunami height of a tsunami arrival point, and tsunami arrival time, and reducing the generation|occurrence|production of an error.

1 is a block diagram of a tsunami height and tsunami arrival time prediction system according to an embodiment of the present invention; FIG. (a) is an electrical block diagram of the marine radar, and (b) is an electrical block diagram of the prediction device. Explanatory drawing of the example of arrangement|positioning of the observation point of a marine radar. Tsunami flow velocity distribution map of simulation of 20m class tsunami. Tsunami flow velocity distribution map of simulation of 20m class tsunami. Tsunami flow velocity distribution map of simulation of 20m class tsunami. Tsunami flow velocity distribution map of simulation of 20m class tsunami. Explanatory drawing of the data processing order in a prediction apparatus. Block diagram of a convolutional neural network unit. Line of sight no. and observation point No. matrix. Explanatory diagram of multi-level classification.

(First embodiment)
A tsunami height and tsunami arrival time prediction system (hereinafter simply referred to as a prediction system) according to one embodiment of the present invention will be described below with reference to FIGS. 1 to 11. FIG.

Prediction system 10 includes marine radar 20 , prediction device 30 , and display device 50 .
<Ocean Radar 20>
The prediction system 10 of the present embodiment irradiates the sea surface with radio waves from the transmitting and receiving antenna 22 (see FIG. 2A) installed on land with the marine radar 20, receives backscattered waves due to waves on the sea surface, and performs frequency analysis. By doing so, the velocity of ocean currents, waves, tsunamis, etc. can be measured in a wide measurement area of about 100 km. Then, the prediction device 30 generates a measured tsunami velocity distribution image with respect to the measurement area, and classifies the measured tsunami velocity distribution image and a combination of tsunami height and tsunami arrival time at a tsunami arrival point as a predetermined point. , and the result is displayed on the display device 50 .

In this embodiment, the tsunami arrival point is the point (installation point) where the transmitting/receiving antenna 22 of the marine radar 20 is installed, but it is not limited to the point where the transmitting/receiving antenna 22 of the marine radar 20 is installed. , or other locations.

As shown in FIG. 2( a ), the marine radar 20 includes a transmitting/receiving antenna 22 , a transmitting/receiving section 24 , a signal processing section 26 and a computing section 28 . The transmitting/receiving antenna 22 is, for example, an HF radar, and is generally installed on land. For example, it may be installed in the sea. Further, in FIG. 1, the shape of the transmitting/receiving antenna 22 may be cylindrical, linear, or arrayed. Further, the transmitting/receiving antenna 22 may be a transmitting antenna and a receiving antenna, which may be integrated or separated.

The transmitting/receiving section 24 includes a signal generating circuit and the like that generates a transmission signal having a predetermined frequency and outputs the transmission signal to the transmitting/receiving antenna 22 . The transmitting/receiving section 24 also includes receiving means for receiving a received signal from the transmitting/receiving antenna 22 and outputting the received signal to the signal processing section 26 . The transmitting/receiving unit 24 may be configured such that the transmitting unit and the receiving unit are integrated, or the transmitting unit and the receiving unit are separated.

The transmitting/receiving antenna 22 radiates a transmission signal for detecting a tsunami to the sea (sea surface) as a transmission radio wave, receives a reception radio wave strongly backscattered (Bragg scattering) on the sea surface as a reception signal, and converts the reception signal into a radio wave. Output to the transmission/reception unit 24 .

The signal processing unit 26 multiplies the transmission signal generated by the transmission/reception unit 24 by the reception signal received by the transmission/reception antenna 22, and outputs the signal resulting from the multiplication to the calculation unit 28 as a beat signal. That is, the signal processing unit 26 generates a beat signal having a frequency difference between the transmission signal and the reception signal. The received radio waves are modulated by the Doppler effect, and the amount of modulation depends on the current velocity on the sea surface and is calculated as a beat signal. Note that the signal processing unit 26 filters out harmonic components from the signal resulting from the multiplication of the transmission signal and the reception signal.

The calculation unit 28 receives the beat signal from the signal processing unit 26, and based on the beat signal, the transmission/reception antenna 22 observes, for example, an observation point P (n, m) in a sea area of several tens to several hundreds of kilometers square (Fig. 3) to calculate the velocity of the sea surface of the tsunami. The calculation of the sea surface velocity of this tsunami will be described in detail.

FIG. 3 shows the entire area (sea area) observed by the marine radar 20 . As shown in the figure, a plurality of observation points P (n, m) are set in a direction away from the transmitting/receiving antenna 22 on a plurality of lines of sight Ln (n=1, 2, . ing. Note that n of the observation point P(n,m) is the same as n of the line of sight Ln, and m(=1, 2, . It is a number attached to the side.

4 to 7, divided regions are set for each observation point P(n, m) in the entire region irradiated with the transmitted radio wave. The computing unit 28 calculates the current velocity of the tsunami on the sea surface of each divided area based on the beat signal of the observation point P(n, m) of each divided area.

The calculation unit 28 also outputs the calculated flow velocity of each divided region to the prediction device 30 . That is, the calculation unit 28 observes the entire area observed by the marine radar 20 by the transmitting/receiving antenna 22 every second, and every second, the prediction device 30 reports the divided areas of the entire area observed by the marine radar 20. Output the velocity.

<Prediction device 30>
FIG. 2B shows the hardware configuration of the prediction device 30 according to this embodiment. As shown in the figure, the prediction device 30 is composed of a computer having a CPU 37, a ROM 38, a RAM 39, an external storage device (for example, flash memory) 40, and a communication interface 41 for communicating with external devices and the Internet. there is

The function of each part of the prediction device 30 described below is such that the CPU 37 stores control programs (eg, image processing programs) and various data (eg, CNN network structure and learning network parameters, etc.).

Some or all of the functions of each unit may be realized by processing by GPUs (Graphics Processing Units) in place of or together with processing by the CPU 37 . Also, similarly, part or all of each function may be performed by dedicated hardware circuit processing instead of or together with software processing.

Each part realized by the above CPU 37 will be described. As shown in FIGS. 1 and 8, the prediction device 30 includes a flow velocity distribution image generation unit 31, a normalization unit 32, a preprocessing unit 33, a CNN processing unit 34, a learning unit 35, and a learning data input unit 36. there is

The flow velocity distribution image generation unit 31 generates a flow velocity distribution image by using the flow velocity of each divided area of the entire area input from the ocean radar 20 or the learning data input unit 36 as image data. Here, for convenience of explanation, the identification of the divided area is substituted by the code P(n,m) attached to the observation point of the divided area.

The flow velocity distribution image is image data consisting of a group of pixels of vertical n×horizontal m, and the flow velocity of each divided area is treated as the brightness of each pixel.
The flow velocity distribution image generator 31 generates a flow velocity distribution image every minute based on the flow velocity of each divided area for the entire area input every second.

Here, the flow velocity distribution image generation unit 31 generates a tsunami flow velocity distribution image for learning when data is input from the learning data input unit 36, and a measured tsunami flow velocity distribution when data is input from the ocean radar 20. Generate an image.

The normalization unit 32 normalizes the brightness (flow velocity) of the image data output from the flow velocity distribution image generation unit 31, for example, by the min-max method. The min-max method is a process in which the maximum value of brightness (flow velocity) is set to "1" and the minimum value is set to "0", and other flow velocities are set to values between "1" and "0". be. Note that the normalization here is not limited to the min-max method, and may be performed by other appropriate methods.

The preprocessing unit 33 shapes the size (length x width) of the image data to a predetermined size. That is, the image data having the size of n (vertical)×m (horizontal) input from the normalization unit 32 is padded with 0 so as to have the size of N (vertical)×M (horizontal).

The preprocessing unit 33 stores the image data formed in the predetermined size as described above in a buffer (not shown). Then, the image data are accumulated in the buffer in the order in which they were stored.
Then, the preprocessing unit 33 forms the latest image data D, the image data D1 formed one minute before the latest image data, and the image data D1 formed two minutes before the latest image data among the image data formed as described above. The resulting image data D2 is read out from the buffer and output to the CNN processing unit 34 as RGB image data. Note that RGB in this specification is used simply for identifying channels, not as RGB for color.

The CNN processing unit 34 is, for example, AlexNet type CNN. The CNN processing unit 34 corresponds to a convolutional neural network unit.
The CNN processing unit 34 performs image processing on the image data D, D1, and D2 in each of the RGB channels using a convolutional neural network, and converts the combined image data of the image data D, D1, and D2 into It is classified into a combination of tsunami height and tsunami arrival time at the tsunami arrival point.

The learning data input unit 36 inputs learning data for performing learning processing of the convolutional neural network referred to by the CNN processing unit 34 .
<CNN processing unit 34>
A process of classifying the combined image data of the image data D, D1, and D2 into combinations of tsunami height and tsunami arrival time at the tsunami arrival point by the prediction device 30 using a convolutional neural network will be described.

As shown in FIG. 9 , the CNN processing section 34 has a convolution section 70 and a full coupling section 80 . The convolution unit 70 performs processing for extracting image features from the input image, and the total combining unit 80 converts the combined image data of the image data D, D1, and D2 from the extracted image features to the tsunami at the tsunami arrival point. Classification into combination of height and tsunami arrival time is performed.

The convolution unit 70 has three RGB channels, and includes image processing units 170, 270 and 370 for R image (image data D), G image (image data D1) and B image (image data D2). I have.

The image processing unit 170 for the R image (image data D) has a structure in which a plurality of layers of feature amount extraction layers 170-1, . . . , 170-p (p=1, 2, . ing. The feature quantity extraction layers 170-1, .

The feature quantity extraction layer 170-1 of the first layer scans an input image by raster scanning for each predetermined size. Then, the convolution layer 172 and the pooling layer 173 perform feature amount extraction processing on the scanned data, thereby extracting feature amounts included in the R image (image data D).

The second feature amount extraction layer (not shown) extracts an image (hereinafter also referred to as a "feature map") input from the previous feature amount extraction layer 170-1, for example, by raster scanning every predetermined size. Scan to Then, the scanned data is similarly subjected to feature quantity extraction processing by the convolution layer 172 and the pooling layer 173, thereby extracting the feature quantity contained in the input image. Note that the feature amount extraction layers below the second layer are integrated while considering the positional relationship of the plurality of feature amounts from which the feature amount extraction layer 170-1 of the first layer is extracted. are extracted.

More specifically, each feature quantity extraction layer 170-1, . . . , 170-p executes the following processing. The convolution layer 172, for example, performs a convolution operation on each pixel value of a scanned image of a predetermined size using a feature extraction filter in which a weighting factor (and bias) is set. Then, the convolution layer 172 sequentially executes such arithmetic processing for each scanning and performs mapping. Then, the convolution layer 172 performs a convolution operation on each of the images input from the previous layer using the feature extraction filter, and adds them to the corresponding mapping positions to form one feature map. to generate

The pooling layer applies a maximum pooling function or an average pooling function to the feature map output by the previous layer for each predetermined size (for example, every 2×2 pixels) to compress the feature amount. do.

Each feature extraction layer 170-1, . Then, each feature amount extraction layer 170-1, . A map (hereinafter referred to as "feature map group") is generated.

The image processing unit 270 for the G image (image data D1) has a structure in which a plurality of layers of feature amount extraction layers 270-1, . . . , 270-p (p=1, 2, . ing. The feature quantity extraction layers 270-1, . . . , 270-p include a convolution layer 272 and a pooling layer 273, respectively.

The feature amount extraction layer 270-1 of the first layer scans the input image for each predetermined size by raster scanning. Then, the convolution layer 272 and the pooling layer 273 perform feature amount extraction processing on the scanned data, thereby extracting feature amounts included in the G image (image data D1).

The second feature quantity extraction layer (not shown) scans the feature map input from the previous feature quantity extraction layer 270-1 by raster scanning, for example, for each predetermined size. Then, the scanned data is similarly subjected to the feature amount extraction processing by the convolution layer 272 and the pooling layer 273, thereby extracting the feature amount included in the input image. Note that the feature amount extraction layers below the second layer are integrated while considering the positional relationship of the plurality of feature amounts extracted by the feature amount extraction layer 270-1 of the first layer. are extracted.

, 270-p are the same as those of the feature amount extraction layers 170-1, .
The image processing unit 370 for the B image (image data D2) has a structure in which a plurality of layers of feature amount extraction layers 370-1, . . . , 370-p (p=1, 2, . ing. The feature quantity extraction layers 370-1, . . . , 370-p include a convolution layer 372 and a pooling layer 373 respectively.

The feature quantity extraction layer 370-1 of the first layer scans an input image for each predetermined size by raster scanning. Then, the convolution layer 372 and the pooling layer 373 perform feature amount extraction processing on the scanned data, thereby extracting feature amounts included in the B image (image data D2).

The second feature amount extraction layer (not shown) scans the feature map input from the previous layer feature amount extraction layer 370-1, for example, by raster scanning every predetermined size. Then, the scanned data is similarly subjected to the feature amount extraction processing by the convolution layer 372 and the pooling layer 373, thereby extracting the feature amount included in the input image. Note that the feature amount extraction layers below the second layer are integrated while considering the positional relationship of the plurality of feature amounts from which the feature amount extraction layer 370-1 of the first layer is extracted. are extracted.

, 370-p are the same as those of the feature extraction layers 170-1, .
Note that an activation layer may be provided between the convolution layer and the pooling layer in each of the RGB channels. In this case, the activation layer applies, for example, a well-known logistic sigmoid function, ReLU function (Rectified Linear Units), etc. to the pixel value of each pixel of the feature map after the convolution operation by the convolution layer. Execute.

, 170-p, 270-1 . -1 . . . , 370-p are repeated to extract various feature amounts of the subject in the image at high dimensions. Then, the convolution unit 70 uses the feature map group generated by the last layer of the plurality of feature quantity extraction layers 170-p, 270-p, and 370-p as the final calculation result, and all Output to the combining unit 80 .

The fully connected section 80 is composed of, for example, a multilayer perceptron in which a plurality of fully connected layers 80-1, . . . , 80-q (fully connected) are hierarchically connected. Note that q=1, 2, . . .

The fully-connected layer on the input side of the fully-connecting unit 80 fully connects to each value of the feature map group acquired from the connecting unit 75, performs a sum-of-products operation on the result of the connection while varying the weight coefficient, and outputs the result. . The fully connected layer on the input side of the fully connected unit 80 is provided with the same number of elements as the values of the feature map group in order to convert each value of the feature map group into a feature quantity vector.

The fully-connected layer next to the fully-connected layer 80-1 is fully connected to the values output by the elements of the previous fully-connected layer, and performs the sum-of-products operation on the result of the connection while varying the weight coefficients. . Through such processing, the full combining unit 80 outputs the identification result of the classification item to be identified in the image based on the features obtained from the image. Similarly, the fully connected layer 80-q in the lower hierarchy outputs the identification result of the classification item to be identified in the image.

Note that the total coupling unit 80, for example, executes a process of applying a softmax function or the like in the softmax unit 81 to the output value from each element of the output layer of the multi-layer perceptron, and selects , the classification result is output to the output unit 82 so that the value of the calculation result by the sum-of-products calculation becomes large for the applicable classification item. That is, the measured tsunami velocity distribution image based on the data from the ocean radar 20 is classified into a combination of arrival time at the tsunami arrival point and tsunami height at the arrival point.

A convolutional neural network is subjected to learning processing using an image attached with learning data that is a correct answer, so that network parameters (for example, weighting coefficients and biases of convolutional layers, weighting coefficients and biases of fully connected layers) is adjusted and works as above.

(Action of Embodiment)
The operation of the prediction system 10 configured as described above will be described.
<Learning of prediction device 30>
The learning data is created in advance by the simulator, and is data of the flow velocity in the divided area (observation point) P(n,m) in the line of sight Ln of the marine radar 20 .

That is, the learning data is the first wave, Leading Wave, created by the simulator in order to detect the tsunami as far offshore as possible. is predicted. This Leading Wave is a solitary wave, and the horizontal flow velocity distribution is characterized by a drastic change in flow velocity at the front of the solitary wave, especially in the direction of propagation.

The learning data includes flow velocities 1, 2, 3, . Here, when the tsunami flow velocity in each divided area is known, the tsunami height in each divided area can be calculated by a known method based on the flow velocity. Although the tsunami height is affected by the topography of the seafloor in the observation area, coastal currents, etc., it can be calculated based on empirical rules.

The learning data, as shown in FIG. 10, is data every minute from the occurrence of the tsunami until it reaches the tsunami arrival point (the coast closest to the transmitting/receiving antenna 22 in this embodiment), and is the line of sight Ln ( ) and observation point P(n,m) (observation point No. in FIG. 10), and the flow velocity is embedded in each column corresponding to the divided area.

The learning data of the present embodiment is data every minute from the occurrence of the tsunami to the arrival point of the tsunami, with respect to tsunami heights of various heights in units of m, including tsunami heights of less than 1 m to more than 22 m. It's becoming

4 to 7 show examples similar to those observed by the transmitting/receiving antenna 22 of the above data, showing an example from the occurrence of a tsunami with a tsunami height exceeding 20 m until reaching the tsunami arrival point. there is

FIG. 4 shows a state in which a tsunami with a tsunami height exceeding 20 m reaches the divided area P(1,20) on the line of sight L1. FIG. 5 shows a state in which a tsunami with a tsunami height exceeding 20 m reaches the divided area P (1, 16) on the line of sight L1 several minutes after the state in FIG. FIG. 6 shows a state in which a tsunami with a tsunami height exceeding 20 m reaches the divided area P(1,8) on the line of sight L1 several minutes after the state in FIG. FIG. 7 shows a state in which a tsunami with a tsunami height exceeding 20 m reaches the divided area P(1,1) on the line of sight L1 several minutes after the state in FIG.

4 to 7, tsunamis in other divided regions are omitted for convenience of explanation.
As described above, the learning data created for a tsunami with a tsunami height (H) of less than 1m to over 22m is the elapsed time (T) The data are sequentially input from the learning data input unit 36 to the learning unit 35 . The learning unit 35 inputs the input learning data to the flow velocity distribution image generation unit 31, and converts the data generated through the normalization unit 32 and the preprocessing unit 33 to the tsunami arrival point at the tsunami arrival point as shown in FIG. The CNN processing unit 34 deep-learns labels based on combinations of the time T and the tsunami height H at the arrival point.

In FIG. 11, the tsunami height H is divided into items of H<1m, H<2m, . Each column of combinations shows a classification from 1 to 374.

When one of the classification results is input to the output unit 82, the output unit 82 displays a combination of the tsunami arrival time T at the tsunami arrival point and the tsunami height H at the arrival point based on this classification result. Display at 50.

Therefore, the CNN processing unit 34 learns whether the learning data is correct based on the tsunami flow velocity distribution image based on the learning data (for learning). The CNN processing unit 34 starts the first learning from the first learning data and ends the learning when all the learning data are used.

Here, if the correct answer differs from the classification result of the combination of the tsunami arrival time T at the actual tsunami arrival point and the tsunami height H at the arrival point, the learning unit 35 uses network parameters (for example, convolution Update layer weighting factors and biases, fully connected layer weighting factors and biases). This update in deep learning is repeated by the learning unit 35 until the correct answer matches the classification result of the combination of the tsunami arrival time T at the actual tsunami arrival point and the tsunami height H at the arrival point.

Here, it is characteristic that the image data output from the preprocessing unit 33 to the CNN processing unit 34 is the latest image data D, the image data D1 formed one minute before the latest image data, and the latest data The image data D2 formed two minutes before is read from the buffer and output to the CNN processing unit 34 as RGB image data.

As a result, the CNN processing unit 34 can recognize the time change by the three image data, and the classification accuracy of the CNN processing unit 34 is improved. Note that the CNN processing unit 34 before the supply of learning data is started prepares a new data set and controls and tunes the propagation function.

<Actual Observation of Prediction System 10>
The operation of the prediction device 30 after completing the deep learning using the learning data as described above will be described.

As shown in FIGS. 1 and 2(a), the prediction device 30 inputs the flow velocity of each divided area for the entire area observed by the ocean radar 20, and the flow velocity distribution image generator 31 generates the flow velocity of each divided area as an image. A flow velocity distribution image is generated as data and output to the normalization unit 32 . The current velocity distribution image created based on the actual observation results corresponds to the measured tsunami current velocity distribution image.

The normalization unit 32 shown in FIG. 8 normalizes the brightness (flow velocity) of the image data output from the flow velocity distribution image generation unit 31 by, for example, the min-max method, and supplies the normalized data to the preprocessing unit 33. Output to the preprocessing unit 33 .

The preprocessing unit 33 shapes the size (length x width) of the image data to a predetermined size. That is, the image data having a size of n (vertical)×m (horizontal) input from the normalization unit 32 is padded with 0 so as to have a size of N (vertical)×M (horizontal), and is stored in a buffer (not shown). Then, the preprocessing unit 33 forms the latest image data D, the image data D1 formed one minute before the latest image data, and the image data D1 formed two minutes before the latest image data among the image data formed as described above. The resulting image data D2 is read out from the buffer and output to the CNN processing unit 34 as RGB image data.

The convolution unit 70 of the CNN processing unit 34 shown in FIG. 9 includes a plurality of hierarchically connected feature amount extraction layers 170-1, . -p, 370-1, . Then, the convolution unit 70 uses the feature map group generated by the last layer of the plurality of feature quantity extraction layers 170-p, 270-p, and 370-p as the final calculation result, and all Output to the combining unit 80 .

The fully-connected unit 80 fully-connects the fully-connected layer of each hierarchy to the values output by the elements of the fully-connected layer of the previous hierarchy, and performs a sum-of-products operation on the result of the connection while varying the weighting factor. Through such processing, the full combining unit 80 outputs to the softmax unit 81 the identification result of the classification item to be identified in the image based on the features obtained from the image.

Then, the softmax unit 81 executes a process of applying a softmax function or the like to the output value, and for the corresponding classification item among the plurality of classification items, the value of the calculation result by the sum-of-products calculation is large. The classification result is output to the output unit 82 so that Based on the classification result, the output unit 82 causes the display device 50 to display a combination of the tsunami arrival time T at the tsunami arrival point and the tsunami height H at the arrival point.

As described above, in this embodiment, by generating three chronologically continuous measured tsunami current velocity distribution images based on the observation results of the ocean radar 20, the latest measured tsunami current velocity distribution in the measured tsunami current velocity distribution images Images can be classified into combinations of tsunami arrival time and tsunami height when reaching a tsunami arrival point (predetermined point). As a result, the tsunami arrival time and tsunami height at the arrival point of the tsunami can be known at an early stage.

It should be noted that the term “continuous over time” in this specification means that the observation data of the ocean radar 20 is obtained every second as in the present embodiment, but the image data D, D1, D2 (for example, one minute apart).

For example, if the observation data of the ocean radar 20 is not every second but every minute or several tens of seconds, the preprocessing unit 33 changes the time intervals of the image data D, D1, and D2 to the minutes or This includes the case where there is an interval of several tens of seconds.

This embodiment has the following features.
(1) The prediction system 10 of this embodiment includes a flow velocity distribution image generator 31 that generates three chronologically continuous measured tsunami flow velocity distribution images based on the observation results of the ocean radar 20 . In addition, in the prediction system 10, the CNN processing unit 34 (convolutional neural network unit) selects the latest measured tsunami current velocity distribution image among the three measured tsunami current velocity distribution images based on the three measured tsunami current velocity distribution images, The tsunami arrival time at the tsunami arrival point (predetermined point) and the combination of the tsunami height are classified.

As a result, the calculation up to the classification is simplified, and the calculation time for the tsunami height at the wave arrival point and the tsunami arrival time is shortened, and the occurrence of errors can be reduced. In addition, in this embodiment, since pre-learning is performed using convolutional neural network technology, observation data input to this system can be instantly classified and discriminated. In addition, by performing many verifications in advance, it is possible to verify the accuracy of the tsunami arrival time and tsunami height at the time of arrival, and to understand the error in tsunami judgment.

(2) In the prediction system 10 of this embodiment, the three measured tsunami current velocity distribution images are taken at intervals of one minute. As a result, the CNN processing unit 34 (convolutional neural network unit) can recognize temporal changes from the three image data, and the classification accuracy of the CNN processing unit 34 can be improved.

(3) In the prediction system 10 of the present embodiment, the predetermined point is the installation point of the transmitting/receiving antenna of the marine radar 20 . As a result, the above effect (1) can be easily obtained with respect to the tsunami arrival time and tsunami height toward the transmitting/receiving antenna.

(4) In the prediction system 10 of the present embodiment, the three measured tsunami current velocity distribution images are input to three RGB channels of the CNN processing section 34 (convolutional neural network section). A convolutional neural network for inputting a color image can be used as the convolutional neural network unit.

In addition, the embodiment of the present invention is not limited to the above embodiment, and may be modified as follows.
-Although AlexNet type CNN is used as the convolutional neural network unit, it is not limited to this. For example, the following convolutional neural network may be used as the convolutional neural network unit.

BN-AlexNet, BN-NIN, ENet, GoogleNET, ResNet-18, VGG-16, ResNet-50, ResNet-101, inception-v3, inception-v4.

- In the above embodiment, three measured tsunami current velocity distribution images are taken at intervals of one minute, but the interval is not limited to one minute. For example, an interval of several tens of seconds, such as an interval of 30 seconds, may be used, and is not limited.

10 ... prediction system,
20... marine radar, 22... transmitting/receiving antenna, 24... transmitting/receiving section,
26... Signal processing unit, 28... Calculation unit, 30... Prediction device,
31... flow velocity distribution image generation unit, 32... normalization unit, 33... preprocessing unit,
34... CNN processing unit, 35... learning unit, 36... learning data input unit,
37...CPU, 38...ROM, 39...RAM, 40...external storage device,
41... communication interface, 60... integration processing unit, 70... convolution unit,
80... fully connected section, 80-1, ..., 80-q... fully connected layer, 82... output section,
170... image processing unit, 170-1,..., 170-p... feature quantity extraction layer,
172... convolution layer, 173... pooling layer,
270... image processing unit, 270-1,..., 270-p... feature quantity extraction layer,
272... convolution layer, 273... pooling layer,
370... image processing unit, 370-1,..., 370-p... feature quantity extraction layer,
372... convolution layer, 373... pooling layer.

Claims (4)

Tsunami height and height with a convolutional neural network unit that pre-learns tsunami velocity distribution images for learning and classifies measured tsunami velocity distribution images into combinations of tsunami arrival time and tsunami height when the tsunami reaches a predetermined point. A tsunami arrival time prediction system,
Equipped with a current velocity distribution image generation unit that generates three chronologically continuous measured tsunami current velocity distribution images based on the observation results of the ocean radar,
The convolutional neural network unit extracts each feature quantity of the three measured tsunami flow velocity distribution images, and based on the result of combining all the feature quantities , the latest A tsunami height and tsunami arrival time prediction system that classifies the measured tsunami current velocity distribution image into a combination of tsunami arrival time and tsunami height when reaching the predetermined point. The tsunami height and tsunami arrival time prediction system according to claim 1, wherein the three measured tsunami velocity distribution images are spaced one minute apart. The tsunami height and tsunami arrival time prediction system according to claim 1 or 2, wherein said predetermined point is an installation point of a transmitting/receiving antenna of said marine radar. The tsunami height and tsunami arrival according to any one of claims 1 to 3, wherein the three measured tsunami current velocity distribution images are respectively input to three channels of RGB of the convolutional neural network unit. time prediction system.

Images