Weather management system and weather management method

Abstract

To provide a weather management system that uses a weather phenomenon corresponding to observation data of a weather radar obtained from weather-related posted information, and outputs alert information on the basis of results of analyzing the observation data.SOLUTION: A weather management system of an embodiment classifies observation data of a weather radar on the basis of a predetermined determination criterion, outputs an analysis result based on the classification, and outputs alert information on the basis of the analysis result. The determination criterion can be changed on the basis of a weather phenomenon grasped from weather-related posted information which is posted from a submitter terminal and matches the observation data observed by the weather radar. Also, a message indicating that the weather phenomenon based on the analysis result may have already occurred, or a message indicating that the weather phenomenon based on the analysis result may occur in the future is output to a user of a user terminal that receives the alert information.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] Embodiments of the present invention relate to a technique using the true value (correct level) of a meteorological phenomenon corresponding to observation data of a weather radar, obtained by collating an observation result based on the observation data observed by the weather radar with the actual meteorological phenomenon.

Background Art

[0002] A weather radar emits radio waves (microwaves) and observes rain and snow existing within a predetermined range. The distance to the rain or snow can be measured from the time until the emitted radio waves return, and the intensity of the rain or snow can be observed from the intensity of the returned radio waves.

[0003] In recent years, dual-polarization weather Doppler radars have been introduced, making it possible to more accurately estimate the type discrimination of precipitation particles in clouds and the intensity of precipitation by using radio waves vibrating in the horizontal direction (horizontal polarization) and radio waves vibrating in the vertical direction (vertical polarization).

Prior Art Documents

Non-Patent Documents

[0004]

Non-Patent Document 1

Non-Patent Document 2

Summary of the Invention

Problems to be Solved by the Invention

[0005] Provided is a weather management system that uses weather phenomena corresponding to observation data of a weather radar obtained from weather-related posting information and outputs alert information based on the analysis result of the observation data.

Means for Solving the Problems

[0006] The weather management system according to the embodiment classifies observation data of a weather radar based on a predetermined determination criterion, and includes an analysis unit that outputs an analysis result based on the classification, a determination criterion control unit that changes the determination criterion based on a weather phenomenon grasped from the weather-related posting information that is posted from a contributor terminal and matches the observation data observed by the weather radar, and an alert control unit that outputs alert information based on the analysis result. The alert control unit outputs a message to the user of the user terminal that receives the alert information, indicating that there is a possibility that the weather phenomenon based on the analysis result has already occurred, or a message indicating that the weather phenomenon based on the analysis result may occur in the future.

Brief Description of the Drawings

[0007]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Figure 8

Figure 9

Figure 10

Figure 11

Figure 12

Figure 13

Figure 14

Figure 15

Figure 16

Figure 17

Best Mode for Carrying Out the Invention

[0008] Hereinafter, embodiments will be described with reference to the drawings.

[0009] As described above, a weather radar such as a dual-polarization weather Doppler radar enables more accurate estimation of the type discrimination of precipitation particles in clouds and the intensity of precipitation. On the other hand, the analysis results based on the observation data observed by the weather radar are only "estimates" and may differ from the actual weather phenomena occurring on site. Precipitation particles are classified into categories such as wet snow, dry snow, ice crystals, drizzle, rain, sleet, hail, and heavy rain. For example, in an observation area where it is estimated that "wet snow has fallen" based on the observation data, there may actually be a weather phenomenon where "sleet has fallen".

[0010] Therefore, in addition to the observation data of the weather radar, the Japan Meteorological Agency combines the data of observation devices such as rain gauges across the country held by the Japan Meteorological Agency, the Ministry of Land, Infrastructure, Transport and Tourism, and local governments, and uses the weather phenomena actually observed on site or in the vicinity thereof to strive to improve the accuracy of the estimation results based on the observation data.

[0011] However, with sensor devices such as rain gauges and snow gauges, although it is possible to grasp whether the rainfall or snow accumulation is large or small, it is impossible to grasp whether it was drizzle, wet snow, dry snow, or sleet or hail that fell. In this regard, although a disdrometer (precipitation particle size and velocity distribution measuring device) can observe drizzle, rain, hail, snow, etc., the disdrometer is mostly used for research purposes and is not deployed at the AMeDAS observation stations across the country of the Japan Meteorological Agency. In particular, hail and sleet occur infrequently and are rarely observable on the ground, and it is not realistic to promote the deployment of disdrometers.

[0012] Thus, it has been difficult to feedback to the analysis based on the observation data of the weather radar the weather phenomena that cannot be observed or are difficult to observe by the observation equipment at the AMeDAS observation stations, and it has been difficult to improve the analysis accuracy based on the observation data of the weather radar.

[0013] In addition to weather phenomena such as hail and sleet, the occurrence frequency of gusty weather phenomena such as tornadoes and squalls is also low, and it is rare to be observable on the ground. Although it is possible to observe the occurrence of cumulonimbus clouds based on the observation data of the weather radar and estimate the possibility of the occurrence of tornadoes and the like, in addition to the low occurrence frequency, the occurrence range is local and narrow, and it is difficult to observe with the observation equipment.

[0014] Therefore, a new technology for collating the observation results based on the observation data observed by the weather radar with the actual weather phenomena is constructed to provide the true value (correct level) of the weather phenomena corresponding to the observation data of the weather radar. Specifically, the weather-related posting information posted to the web server that receives posting information from the poster terminal is used.

[0015] Further, in a weather management device that classifies and analyzes the observation data of the weather radar based on a predetermined determination criterion, the determination criterion is changed based on the true value (correct level) of the weather phenomena based on the posting information. For example, by lowering the determination criterion, the overlooking of the occurrence of a predetermined weather phenomenon is suppressed.

[0016] (First Embodiment) FIGS. 1 to 4 are diagrams for explaining the first embodiment. FIG. 1 is a diagram showing the functional blocks and network configuration of the information management device 100 of the present embodiment.

[0017] The information management device 100 is provided with information regarding the observation data of the weather radar from the weather management device 200 and posting information posted by posters from the web server 300. The information management device 100 provides a function of extracting and generating, as the true value (correct label), the weather phenomenon grasped from the weather-related posting information posted by the poster as the weather phenomenon corresponding to the observation data observed by the weather radar. That is, it functions as a device that generates a dataset of the observation data and the correct label corresponding to the observation data.

[0018] The weather radar, for example, emits radio waves (microwaves) and observes rain and snow. It measures the distance to rain and snow from the time until the emitted radio waves return, and observes the intensity of rain and snow from the intensity of the returned radio waves. Also, as described above, a dual-polarization weather Doppler radar is introduced, and by using radio waves vibrating in the horizontal and vertical directions (horizontal polarization, vertical polarization), the type of precipitation particles in the cloud is discriminated and the intensity of precipitation is estimated.

[0019] In a dual-polarization weather Doppler radar, the shape of precipitation particles can be estimated from the ratio of amplitudes. The larger the precipitation particles, the more they are affected by air resistance and become flattened. This is observed using horizontal polarization and vertical polarization, and the shape of the precipitation particles is estimated from the amplitude ratio of the reflected waves. Also, the intensity of rain can be estimated from the phase difference. When radio waves travel through water such as raindrops, their speed is slightly slower than in the atmosphere where there is nothing. Utilizing the property that the speed of the horizontal polarization becomes slower with stronger rain, it is observed using horizontal polarization and vertical polarization, and the intensity of rain is estimated from the phase difference of the reflected waves.

[0020] In the estimation of precipitation particles by the ratio of amplitudes, categories such as wet snow, dry snow, ice crystals, drizzle, rain, hail, sleet, and heavy rain are prepared in advance. For example, when the ratio of amplitudes is equal to or greater than a predetermined value, it is determined (categorized) that the precipitation particles are sleet. The ratio of amplitudes is set as a determination criterion, and based on the observation data, the weather phenomenon in the observation area is estimated. Note that sleet is ice pellets with a diameter of 5 mm or more falling from cumulonimbus clouds, and ice pellets with a diameter less than 5 mm are defined as hail.

[0021] The web server 300 receives posting information posted from the poster terminal T and provides a website that publishes the received posting information on the web. The poster is a user registered with the web server 300 and includes an unspecified number of posters. As an example, posting information to SNSs (Social Networking Services) such as Twitter (registered trademark) and Facebook (registered trademark) can be used. In this case, posting information other than weather-related posting information is also accumulated in the web server 300.

[0022] As another example, posting information to a website that receives posting information specialized in weather-related matters can be used. For example, the poster becomes a weather reporter or an observer and posts daily weather information. In this case, only weather-related posting information is accumulated in the web server 300.

[0023] The information management device 100 is connected to one or more websites 300 to acquire weather-related posting information. Note that the poster terminal T is a mobile terminal such as a multifunctional mobile phone such as a smartphone or a tablet computer, and has a data communication function, an arithmetic function (such as a CPU), and a storage device (such as a memory and an auxiliary storage device) through an IP (Internet protocol) network or a mobile communication network.

[0024] In addition, the poster terminal T can include a display control application such as a browser, a photographing device that photographs a still image or / and a moving image, a touch panel type display input device, and a GPS device. As will be described later, the posting information posted from the poster terminal T includes the position information and time information acquired by the GPS device together with the posting content.

[0025] The weather prediction device 400 provides a weather prediction function using a weather prediction model, for example, the Meso Model of the Japan Meteorological Agency or a mesoscale weather model such as WRF, and provides prediction results of weather phenomena (snow, rain, hail, sleet, etc.) in the prediction area and at the prediction time. The weather prediction device 400 (weather prediction model) is a known technology, and detailed description thereof will be omitted.

[0026] As shown in FIG. 1, the information management device 100 is connected to the weather management information 200, the website 300, and the weather prediction device 400 through an IP network or a dedicated line. The communication device 110 controls data communication with each device. The information management device 100 includes a control device 120 and a storage device 130. The control device 120 includes an information acquisition unit 121, an extraction unit 122, and a data management unit 123.

[0027] FIG. 2 is a diagram for explaining the generation process of true value information that associates the weather phenomenon grasped from the contribution information with the weather phenomenon corresponding to the observation data of the present embodiment.

[0028] The contribution information posted to the web server 300 includes a contribution article or / and a contribution image, location information, time information, and contributor information. The contribution article is a contribution text, and the contribution image is an image taken by a photographing device provided in the contributor terminal T or an image taken by a photographing device different from the contributor terminal T. The location information is the current location of the contributor (contributor terminal T), and may include address information corresponding to the current location. The contribution time is the date and time information when the web server 300 receives the contribution information. The contributor information is a user identifier such as a user ID registered in the web server 300.

[0029] The contributor inputs a contribution article or a contribution image from a contribution screen provided by the web server 300 through the contributor terminal T, and transmits the contribution information to the web server 300. The web server 300 receives the contribution information transmitted from the contributor terminal T, and posts and publishes the contribution information on the contribution website provided by the web server 300.

[0030] Incidentally, as an example of the information management device 100 acquiring, i.e., collecting, the posted information, the mode of acquiring from the web server 300 that receives and publishes the posted information has been described, but it is not limited to this. For example, the posted information may be acquired via a search server (search site). On the search site, posted information related to search keys such as keywords can be collected from a plurality of different posting web sites (a plurality of different web sites 300). Therefore, the information management device 100 can acquire the posted information posted to the web server 300 from the web site 300 or from a search site that can collect the posted information of the web site 300.

[0031] As shown in FIG. 2, the information acquisition unit 121 acquires weather-related posted information from a plurality of posted information posted to the web server 300 based on keywords related to weather or the similarity of images related to weather. One or more pieces of weather-related posted information can be acquired. For example, posted information containing the keyword "hyou" can be acquired as weather-related information, or using a reference image in which "hyou" appears, posted information having a posted image similar to the reference image can be acquired as weather-related posted information. Although "hyou" has been described as an example, weather-related posted information related to "arare" can be collected in the same way. In addition, an image recognition AI model can also be used as a process for extracting posted information having a posted image related to "hyou". For example, a group of images in which "hyou" appears can be input in advance as learning data, and a learning process can be performed to generate an "hyou image recognition" AI model. Then, using the generated AI model, the similarity of the posted image to "hyou" can be calculated, and the posted information having a posted image with a similarity equal to or greater than a predetermined value can be configured to be acquired as weather-related information.

[0032] Next, the extraction unit 122 performs a matching process with the observation data of the weather radar. That is, it extracts weather-related posted information that matches the observation area and observation time of the observation data observed by the weather radar.

[0033] For example, in the analysis result of weather radar observation data, that is, in the precipitation particle estimation based on the amplitude ratio, the observed area and observation time determined as hail are matched with the position information and posting time of weather-related posting information related to "hail". The extraction unit 122 determines whether the position information and posting time of the weather-related posting information related to "hail" belong to a predetermined range that is sufficiently close to the observed area and observation time, and extracts the weather-related posting information related to "hail" within the sufficiently close predetermined range.

[0034] The data management unit 123 associates the weather-related posting information matched with the observation data as the weather phenomenon corresponding to the observation data. In other words, the weather phenomenon grasped from the weather-related posting information extracted by the extraction unit 122 is output as the weather phenomenon corresponding to the observation data.

[0035] The data management unit 123 can function as a data generation unit that outputs the weather phenomenon grasped from the weather-related posting information as the true value (correct label) of the weather phenomenon corresponding to the observation data, and generates a data set including the observation data and the true value (correct label) of the weather phenomenon. This data set can be fed back to the weather management device 200 as supervised learning data, and can be used for tuning the above-described analysis processing, analysis algorithm, and determination criteria (threshold values) used in the analysis processing of the weather radar observation data. Note that if the true value (correct label) of the weather phenomenon is associated with the corresponding observation data, it is not necessary to hold the observation data itself as a data set. For example, when feeding back to the weather management device 200, the information management device 100 (data management unit 123) can generate and provide a data set including the observed area and observation time and the true value (correct label) of the output weather phenomenon.

[0036] In the above example, in the precipitation particle estimation based on the observation data, the mode of matching the observation data determined as hail and the meteorological-related posting information related to "hail" has been described. However, for example, the observation area and observation time determined as sleet are matched with the location information and posting time of the meteorological-related posting information related to "hail", and for the observation data of the observation area and observation time determined as sleet, "hail" may be associated as the true value (correct label) of the meteorological phenomenon. Also, vice versa.

[0037] In this way, the information management device 100 of the present embodiment performs the verification of the meteorological phenomenon estimated from the analysis result of the observation data using the posting information that can be collected on the web, and generates the meteorological phenomenon grasped from the meteorological-related posting information as the true value (correct label) of the meteorological phenomenon corresponding to the observation data.

[0038] In particular, for meteorological phenomena with low occurrence frequencies such as hail and sleet, which are rarely observable on the ground, the verification of the analysis result of the observation data can be performed, and the analysis accuracy based on the observation data of the weather radar can be improved.

[0039] FIG. 3 is a diagram for explaining the process of extracting meteorological-related posting information from posting information using various evaluation information of the present embodiment.

[0040] The information collection unit 121 extracts meteorological-related posting information from a plurality of posting information, but can evaluate the reliability of the posting information and extract meteorological-related posting information whose evaluation is above a certain standard. The information collection unit 121 can include a first processing unit to a fourth processing unit.

[0041] As shown in FIG. 3, the first processing unit 121A of the information collection unit 121 extracts candidate posting information from a plurality of posting information posted on the web server 300 based on the similarity of keywords related to meteorology or images related to meteorology (first processing).

[0042] The second processing unit 121B of the information collection unit 121 generates a contributor evaluation value for the contributor of the candidate submission information based on a predetermined evaluation criterion (second processing). The contributor evaluation value can be calculated using evaluation parameters (evaluation criteria) such as the number of past submissions of the contributor, the number of times evaluated by other contributors, and the number of past submission information that has been adopted as the true value (correct label). The number of past submissions of the contributor and the number of times evaluated by other contributors can be obtained from the submission history managed by the web server 300. The number of past submission information that has been adopted as the true value (correct label) can be accumulated by the information management device 100 for the contributor of the weather-related submission information adopted as the true value (correct label). The number of past submission information that has been adopted as the true value (correct label) is stored in the storage device 130 as contributor evaluation information 132.

[0043] The third processing unit 121C of the information collection unit 121 calculates the number of submission information of other contributors similar to the candidate submission information extracted by the first processing unit 121A as the submission accuracy of the candidate submission information. Specifically, submission information of other contributors (similar submission information) that includes position information and time information within a predetermined range close to the position information and time information of the candidate submission information and has similarity in keywords related to weather or images related to weather is extracted from a plurality of submission information submitted to the web server 300. The third processing unit 121C generates a submission accuracy evaluation value for the candidate submission information based on the extracted similar submission information (third processing).

[0044] The fourth processing unit 121D of the information collection unit 121 evaluates the information accuracy of the candidate submission information extracted by the first processing unit 121A using the prediction information generated by the weather prediction device 400. Specifically, the fourth processing unit 121D generates a weather event evaluation value for the candidate submission information based on the prediction information generated by a predetermined weather prediction model corresponding to the position information and time information of the candidate submission information (fourth processing). If the weather phenomenon in the position information and time information of the candidate submission information is the same as the weather information in the prediction information, the weather event evaluation value is calculated to be high, and if the weather phenomenon in the position information and time information of the candidate submission information is different from the weather information in the prediction information, the weather event evaluation value is calculated to be low.

[0045] Then, based on the contributor evaluation value, the submission accuracy evaluation value, and the weather event evaluation value, the information collection unit 121 extracts weather-related submission information from the candidate submission information whose evaluation is equal to or higher than a predetermined value. For example, candidate submission information having each evaluation value equal to or higher than the predetermined value can be extracted as weather-related submission information, or candidate submission information whose total of each evaluation value is equal to or higher than the predetermined value can be extracted as weather-related submission information. Further, weight values may be applied to the contributor evaluation value, the submission accuracy evaluation value, and the weather event evaluation value. For example, the weight values of the contributor evaluation value and the submission accuracy evaluation value may be set high, and the weight value of the weather event evaluation value may be set low, and the total value of the evaluation values may be calculated. In this case, the weight values can be arbitrarily set.

[0046] Each of the contributor evaluation value, the submission accuracy evaluation value, and the weather event evaluation value can be configured independently or in any combination. The information collection unit 121 can be configured to evaluate candidate submission information based on at least one of the contributor evaluation value, the submission accuracy evaluation value, and the weather event evaluation value, or any combination thereof, and extract weather-related submission information.

[0047] Note that the extraction process of weather-related submission information using the above-described evaluation values may not be applied according to the characteristics of the submission information. For example, in the case of weather-related submission information collected from a weather information submission web server 300 (website) that specifically accepts submissions of weather information, it can be handled as submission information for which the reliability of the contributor and the information accuracy as submission information are guaranteed in advance, and the information management device 100 side may not perform selection based on each evaluation value. In other words, when the web server 300 side that accepts submission information selects and publicly discloses submission information for which the reliability of the contributor and the information accuracy are guaranteed on the web, the candidate submission information extracted by the first processing unit 121A can be directly extracted as weather-related submission information.

[0048] FIG. 4 is a diagram showing a processing flow executed by the information management device 100 of the present embodiment.

[0049] Posting information is accumulated in the web server 300 from the poster through the poster terminal T. At this time, posting information other than weather-related information is also included.

[0050] The information management device 100 extracts candidate posting information from a plurality of pieces of posting information posted to the web server 300 based on keywords related to weather or the similarity of images related to weather (S101). For example, using keywords such as "hyou" and "arare", or sample images of "hyou" and "arare" as extraction keys, the posting information stored in the web server 300 is searched. Note that the extraction keys are arbitrarily set by the operation side of the information management device 100.

[0051] Note that the candidate posting information is posting information corresponding to either one of the keywords related to weather and the sample images, or posting information corresponding to both the keywords related to weather and the sample images.

[0052] Next, the information management device 100 performs the evaluation value calculation process described with reference to FIG. 3 on the extracted candidate posting information (S102). The information management device 100 extracts weather-related posting information from one or more pieces of candidate posting information whose evaluation is equal to or higher than a predetermined value based on the calculated poster evaluation value, posting accuracy evaluation value, and weather event evaluation value (S103). For example, the candidate posting information with the highest evaluation result based on the evaluation value can be extracted as the weather-related posting information.

[0053] The information management device 100 acquires observation data from the weather management device 200. The information management device 100 extracts weather-related posting information that matches the observation area and observation time of the observation data observed by the weather radar (S104). Then, the information management device 100 outputs the weather phenomenon grasped from the extracted weather-related posting information as the true value (correct label) of the weather phenomenon corresponding to the observation data (S105). The true value (correct label) of the weather phenomenon based on the output posting information is stored in the storage device 130 (S106).

[0054] As described above, the information management device 100 can also function as a data generation unit that generates a data set including observation data and the true value (correct label) of the weather phenomenon. The information management device 100 can provide the data set to the weather management device 200 or the weather prediction device 400 at an appropriate timing, which can be used to improve the accuracy of the analysis process of the weather management device 200 and the prediction process of the weather prediction device 400.

[0055] In addition, in the processing steps of the above-described information management device 100, the matching process centered on step S104 includes two modes.

[0056] The first one is that based on the location information and time information of the weather-related posting information extracted in step S103, the information management device 100 (control device 120) acquires the observation data including the corresponding observation area and observation time from the weather management device 200, and in step S104, they can be configured to be matched.

[0057] The second one is that in the weather management device 200, the observation data observing specific weather phenomena such as "hyou" and "arare" is acquired in advance, and in the process of extracting the weather-related posting information from step S101 to step S103, the weather-related posting information corresponding to the observation area and observation time of the observation data can be configured to be extracted.

[0058] Both processes match the weather-related posting information collected from the web server 300 and the observation data with the location information and time information, and extract the weather-related posting information that matches the observation area and observation time of the observation data observed by the weather radar, and output it as the true value (correct label) of the weather phenomenon corresponding to the observation data.

[0059] (Second Embodiment) Figures 5 to 7 are diagrams for explaining the second embodiment. This embodiment provides a weather management function in which the weather management device 200 analyzes observation data and outputs an alert based on the analysis result. At this time, the true value (correct label) of the weather phenomenon for the observation data generated by the information management device 100 of the first embodiment is fed back in real time. The weather management device 200 of this embodiment constitutes a weather management system in cooperation with the information management device 100 of the first embodiment.

[0060] Figure 5 is a diagram showing the functional blocks and network configuration of the weather management device 200 of this embodiment. The weather management device 200 includes a communication device 210, a control device 220, and a storage device 230. The control device 220 includes an analysis unit 221, a determination criterion control unit 222, an alert control unit 223, and a true value information acquisition unit 224.

[0061] The analysis unit 221 classifies the observation data of the weather radar based on a predetermined determination criterion and outputs an analysis result based on the classification. As described above, the analysis unit 221 can perform precipitation particle estimation processing (analysis processing) based on the ratio of amplitudes, and categories (classifications) such as wet snow, dry snow, ice crystals, drizzle, rain, hail, sleet, and heavy rain are prepared in advance. The analysis unit 221 sets the ratio of amplitudes as a determination criterion and estimates the weather phenomenon in the observation area based on the observation data.

[0062] The alert control unit 223 outputs alert information based on the analysis result of the analysis unit 221. For example, the alert information can be output to an information distribution system registered by a plurality of users or a web server that publishes information on a website, or the alert information can be provided to a user (user terminal) registered in the weather management device 200.

[0063] And in this embodiment, a determination criterion control unit 222 is provided, and the determination criterion is changed based on the weather phenomenon (true value) grasped from the weather-related posting information corresponding to the observation data output from the information management device 100 of the first embodiment.

[0064] FIG. 6 is a diagram for explaining the determination criterion change process of the present embodiment.

[0065] The weather management device 200 performs precipitation particle estimation processing using the observation data A of the weather radar. The observation data A has an observation time of 15:31 on January 5, 2023, and the observation area is City A. The analysis unit 221 outputs a particle diameter of "3 mm" as the analysis result of the observation data A based on the determination criterion. The analysis unit 221 outputs the weather phenomenon of "hail" corresponding to the particle diameter of "3 mm" and the analysis result. The alert control unit 223 generates and outputs alert information such as "There may be hail falling in City A. Please be careful."

[0066] This observation data A is also provided to the information management device 100 in real time in parallel with the analysis process of the weather management device 200, and a generation process of the true value (correct label) of the weather phenomenon based on the weather-related posting information for the observation area and the observation time is performed. Then, the true value (correct label) is associated with the observation data A and provided from the information management device 100 to the weather management device 200. The true value information acquisition unit 224 performs the process of providing the observation data A to the information management device 100 and the process of acquiring the true value (correct label) from the information management device 100.

[0067] When the determination criterion control unit 222 receives the true value (correct label) of the observation data A, it performs a collation process between the analysis result and the true value (correct label). If it is determined as a result of the collation process that the weather phenomenon of the analysis result is different from the true value (correct label) of the observation data A, the determination criterion control unit 222 temporarily changes the determination criterion. In the example of FIG. 6, the analysis result is "hail", the true value (correct label) is "snow", and the two are different. The determination criterion control unit 222 changes the determination criterion with the weather phenomenon based on the observation data A as the true value (correct label).

[0068] For example, the determination criteria for precipitation particle estimation processing are as follows: when the ratio of amplitudes is equal to or greater than a predetermined value X and less than a predetermined value Y, it is determined as "hail pellets", and when the ratio of amplitudes is equal to or greater than the predetermined value Y, it is determined as "hail". As the analysis result of the meteorological phenomenon, "hail pellets", "hail", etc. are output. In the example of FIG. 6, since the ratio of the amplitude of the observation data A was equal to or greater than the predetermined value X and less than the predetermined value Y, the determination criteria control unit 222 output "hail pellets" as the analysis result. However, in reality, "hail" was falling, so the predetermined value Y, which is the boundary threshold between "hail pellets" and "hail", is changed to a predetermined value Y1. The predetermined value Y1 is a value smaller than the predetermined value Y, and a threshold change is performed to expand the region determined as "hail" in the observation data.

[0069] In this way, a threshold change is made to the determination criteria to temporarily narrow the region determined as "hail pellets" and expand the region determined as "hail". Then, the changed determination criteria are applied to the analysis processing of the observation data B observed after the observation data A.

[0070] The observation data B has an observation time of 15:42 on January 5, 2023, and the observation area is City B (adjacent to City A). At this time, the observation data B is the same as the observation data A. According to the analysis result using the determination criteria before the change, it is determined as "hail pellets". However, according to the true value (correct label) of the observation data A, with the determination criteria after the change, the analysis unit 221 outputs the meteorological phenomenon "hail" as the analysis result of the observation data B. The alert control unit 223 generates and outputs alert information such as, for example, "There may be hail falling in City B. Please be careful" or "There may be hail falling in City B soon. Please be careful". Since the weather radar can observe precipitation particles in the upper air, there is a possibility of outputting alert information several minutes to several tens of minutes before it falls to the ground.

[0071] Here, the change in the determination criteria can be applied for each observation area. For example, for the observation data C in an observation area that is far from the observation area of the observation data A, the analysis process can be performed without changing the determination criteria. On the other hand, for the analysis process of the observation data B that is close to the observation area of the observation data A and is observed within a predetermined time from the observation time of the observation data A, the changed determination criteria can be configured to be applied.

[0072] Therefore, the determination criteria can be changed according to the observation area and the observation time of the observation data corresponding to the true value provided by the information management device 100. And the change value of the determination criteria is set with an effective area and an expiration date. That is, as the observation area to which the change value is applied, an area within a predetermined range from the observation area of the observation data A corresponding to the true value (correct label) is set, and as the expiration date, a predetermined time (for example, 30 minutes) from the observation time of the observation data A is set. When the expiration date passes, the change value is initialized, and thereafter, the analysis process is performed using the determination criteria value (default value) before the change.

[0073] FIG. 7 is a diagram showing a processing flow executed by the weather management device 200 of the present embodiment.

[0074] The weather radar makes continuous observations in time series at predetermined intervals and outputs the observation data. When the observation data A of the weather radar is input, the weather management device 200 performs an analysis process on the observation data A (S201). At this time, the weather management device 200 provides the information management device 100 with the observation data A to be analyzed.

[0075] The weather management device 200 performs the above-described precipitation particle estimation process. For example, it classifies based on a predetermined determination criteria and outputs an analysis result based on the classification. The analysis result is stored in the storage device 230 (S202). The weather management device 200 outputs alert information based on the analysis result (S203).

[0076] The weather management device 200 receives the true value information of the weather phenomenon for the observation data A from the information management device 100. When receiving the true value information from the information management device 100, the weather management device 200 starts the determination criterion change process (S204). The weather management device 200 compares the analysis result of the observation data A with the received true value (correct label) information. If it is determined that they are the same weather phenomenon, the determination criterion change process is not performed (NO in S205). On the other hand, if it is determined that they are different weather phenomena (YES in S205), the determination criterion is changed based on the true value (correct label) grasped from the weather-related posting information corresponding to the observation data A (S206).

[0077] The weather management device 200 repeatedly performs the analysis process of the subsequent weather radar observation data using the determination criterion value changed in step S206 (from NO in S208 to S201).

[0078] At this time, as described above, when the determination criterion control unit 222 changes the determination criterion based on the weather phenomenon grasped from the weather-related posting information corresponding to the observation data A, the effective area and effective time to which the changed determination criterion is applied are set based on the observation area and observation time of the observation data A (S207). Then, in step S201, the analysis unit 221 determines whether the observation area and observation time of the subsequent observation data B are within the effective range corresponding to the set effective area and effective time. If it is determined that they are within the effective range, the changed determination criterion is applied to the analysis process of the subsequent observation data B.

[0079] In this embodiment, in the analysis process of the observation data in the weather management device 200, the true value (correct label) provided by the information management device 100 in real time is applied and reflected in the subsequent analysis process of the observation data. That is, when the analysis result of the observation data and the true value (correct label) based on the weather-related posting information are different from each other, by changing the determination criterion based on the true value (correct label), the analysis result of the subsequent observation data can be made more accurate.

[0080] In particular, hail causes significant damage as a natural disaster. Even if the analysis result of the observation data determines that it is not hail but sleet is falling, there may actually be hail falling (in the posted information). Therefore, in subsequent analysis processing of the observation data, even for the same observation data determined to be sleet, it is strictly determined that hail is falling. That is, the frequency of being determined as hail is increased, and the oversight of not being determined as hail is reduced. By configuring in this way, it is possible to easily issue an alert for a meteorological phenomenon that causes significant damage as a natural disaster, and appropriately raise awareness of disaster prevention.

[0081] Also, based on the observation area and observation time of the observation data that triggered the change in the determination criteria, set the area and time range to which the changed value of the determination criteria is applied. Hail is a special meteorological phenomenon with a low frequency of being observed as described above, and is not a meteorological phenomenon that extends over a long time and a wide area. Therefore, from the observation area of the observation data that triggered the change in the determination criteria, for example, within a range of 50 km radius and for observation data within 2 hours, apply the changed value of the determination criteria to easily issue an alert for a meteorological phenomenon that causes significant damage as a natural disaster. By configuring in this way, while appropriately limiting the area where an alert indicating that hail is falling is output, it is possible to suppress confusion and the like caused by the alert without overly expanding the area where an alert indicating that hail is falling is output.

[0082] (Third Embodiment) Figures 8 to 12 are diagrams for explaining the third embodiment. This embodiment is provided with a classification function for meteorological-related posted information based on the insurance payment history, compared to the above first embodiment. Regarding the same configuration as the above first embodiment, the same reference numerals are given in each figure and the description is omitted.

[0083] Figure 8 is a diagram showing the functional blocks and network configuration of the information management device 100 of this embodiment. The control device 120 of the information management device 100 further includes a classification control unit 124, and classification-related information 134 is stored in the storage device 130.

[0084] FIG. 9 is a diagram for explaining a first function of classifying weather-related posting information based on the insurance payment history of the present embodiment and a second function of classifying weather-related posting information using a classification model. The first function is a function of generating learning data from past weather-related posting information and generating a classification model using the generated learning data. The second function is a function of performing a classification determination process on the weather-related posting information newly extracted by the extraction unit 122 using the generated classification model, and associating the result of the classification determination process with the weather phenomenon corresponding to the observation data output from the data management unit 123, that is, the true value (correct label).

[0085] The insurance payment history is a history of actually paying insurance money for damages caused by weather phenomena such as hail, snow, tornadoes, and gusts in insurance products provided by insurance companies. Insurance products include personal insurance products covering natural disasters such as fire insurance, flood insurance, hail insurance, and vehicle insurance, and insurance products covering natural disasters for businesses.

[0086] The history of paying insurance money is the actual record of the insurance company grasping the damage situation and paying insurance money based on the claim from the policyholder, and includes information such as payment items, damage location, damage date and time, and payment target.

[0087] The payment item is a weather phenomenon related to the occurrence of damage. For example, "hail damage" is included as a payment item. The damage location is the location where hail fell and the payment target suffered damage (location information where the damage occurred), for example, address information such as XX town, XX city, XX prefecture. The payment target is the asset that suffered damage and is the insurance target. For example, in the case of vehicle insurance, it is the vehicle, and in the case of fire insurance, it is the house. This insurance payment history information is provided by the insurance company. The insurance payment history information can be obtained from the insurance company (servers and terminals on the insurance company side) through the network, or obtained as physical list information and input into the information management device 100 by the operation administrator. The insurance payment history information is stored in the storage device 130 as classification-related information 134.

[0088] In this embodiment, insurance payment history information is used to classify weather-related posting information into a group of weather phenomena with a high damage risk (a group of weather phenomena with an insurance payment record) and other groups (a group of weather phenomena without an insurance payment record). That is, among a plurality of weather-related posting information, the weather-related posting information corresponding to the damage location and date and time with an insurance payment history is extracted, and information indicating that it is a weather phenomenon with a high damage risk is associated. As a result, even for the same weather phenomenon, weather phenomena with a high damage risk based on the insurance payment history (record) can be classified. For example, the weather-related posting information belonging to the weather phenomenon "hail" can be classified into the weather-related posting information belonging to the weather phenomenon "hail (high damage risk)" and other weather-related posting information.

[0089] Therefore, the classification control unit 124 associates (labels) the classification information "high damage risk" with the weather-related posting information corresponding to the damage location and date and time with an insurance payment history. Then, the data management unit 123 can control so that the classification information associated with the weather-related posting information used to generate the true value (correct label) is included when generating the true value (correct label). For example, the true value information 133 of weather events for each observation data stored in the storage device 130 can be configured to include the classification information "high damage risk" in addition to the observation time, observation area, and weather phenomenon based on the weather-related posting information.

[0090] Hail is an ice pellet with a diameter of 5 mm or more as described above, and even if the diameter is 10 mm, it is still the same hail. However, the greater the diameter, the greater the damage, and the smaller the diameter, the smaller the damage may be. Therefore, even though the weather phenomenon is the same hail, the damage caused by the weather phenomenon is different. In this embodiment, insurance payment history information (damage record) is used as evidence of actual damage caused by the weather phenomenon "hail".

[0091] Then, the weather-related posting information corresponding to the past insurance payment history (the weather-related posting information used to generate the true value (correct label)) is accumulated as learning data. The learning data is stored in the storage unit 130 as classification-related information 134.

[0092] For example, the images and posting texts of the weather-related posting information corresponding to the past insurance payment history can be accumulated as learning data. The images are used as learning data for the images of actual damages, and the posting texts are used as learning data for the texts (keywords, etc.) related to the images of actual damages. For example, posting texts containing keywords related to the image itself such as "it was 30 mm in diameter", "the size of a golf ball", "the hood of the car was dented", "the roof of the garage was broken through", and keywords related to the damages (losses) caused by the image are used as learning data.

[0093] The accumulated learning data is used in the generation process (learning process) of a classification model that determines whether the newly extracted weather-related posting information corresponds to a weather phenomenon with a high damage risk. The classification model calculates, from the image or / and posting text of the newly extracted weather-related posting information, the similarity of the image and the similarity of the posting text based on keywords, etc., and classifies the weather-related posting information having a similarity equal to or higher than a predetermined value into the "group of weather phenomena with a high damage risk".

[0094] FIG. 10 is a diagram showing the processing flow of the first function executed by the classification control unit 124. The classification control unit 124 performs the following processing.

[0095] 1) Based on the insurance payment history information including the location where the damage related to the insurance payment occurred, the date and time when the damage occurred, and the weather phenomenon related to the occurrence of the damage, the past weather-related posting information extracted by the extraction unit 122 is classified into the group of weather phenomena with a high damage risk in the first process (S1001, S1002). 2) The second process (S1003) of storing the weather-related posting information classified into the group of weather phenomena with a high damage risk as learning data in a predetermined storage area. 3) Using the learning data, perform a third process (S1004) of generating a classification model for classifying the weather-related post information to be extracted by the extraction unit 122 in the future into weather groups with a high risk of damage.

[0096] FIG. 11 is a diagram showing a processing flow executed by the information management device 100, which corresponds to the processing flow of FIG. 4 in the first embodiment. In FIG. 11, steps S104A and S104B are added, and these steps correspond to the second function by the classification control unit 124.

[0097] For the weather-related post information extracted in step S104, the classification control unit 124 performs the following processing. 4) Perform a fourth process (S104A) of performing a classification determination process for determining whether the weather-related post information extracted by the extraction unit 122 belongs to a weather phenomenon group with a high risk of damage using the generated classification model. 5) Perform a fifth process (S104B) of associating the result of the classification determination process with the weather phenomenon (true value (correct label)) corresponding to the observation data output from the data management unit 123.

[0098] Then, the true value (correct label) output from the data management unit 123 is added with the classification determination result (identification information indicating that it is a weather phenomenon with a high risk of damage) by the classification model. That is, the true value (correct label) output from the data management unit 123 is configured to have information on whether it is a weather phenomenon group with a high risk of damage (S105).

[0099] Next, a case where the true value (correct label) including the identification information indicating "weather phenomenon with a high risk of damage" in the present embodiment is applied to the second embodiment will be described.

[0100] First, as shown in FIG. 6, in the second embodiment, the determination criteria can be changed using the true value (correct label) provided by the information management device 100. Here, even for the true value (correct label) including the identification information indicating that it is a "weather phenomenon with a high damage risk" provided by the information management device 100 of the present embodiment, regardless of whether it is a "weather phenomenon with a high damage risk", the weather management device 200 can change the determination criteria.

[0101] On the other hand, the weather management device 200 can also control so that the change width of the determination criteria is different between the true value (correct label) belonging to the "weather phenomenon with a high damage risk" and the true value (correct label) not belonging to the "weather phenomenon with a high damage risk". For example, when changing the determination criteria based on the true value (correct label) belonging to the "weather phenomenon with a high damage risk" (S206), it is changed to the change value P1, and when changing the determination criteria based on the true value (correct label) not belonging to the "weather phenomenon with a high damage risk", it can be changed to a change value P2 smaller than the change value P1 (P1 > P2).

[0102] Also, as shown in FIG. 12, the alert control unit 223 of the weather management device 200 determines whether the true value (correct label) provided from the information management device 100 belongs to the "weather phenomenon with a high damage risk" (S2004). When it is determined that it belongs to the "weather phenomenon with a high damage risk" (YES in S2004), the alert control unit 223 can control to output alert information (for example, a warning alert) indicating that it is a weather phenomenon for which actual damage has been reported in the past in the alert output (S2005). On the other hand, when it is determined that it does not belong to the "weather phenomenon with a high damage risk" (NO in S2004), the alert control unit 223 can control to output normal alert information in the alert output (S2006).

[0103] Therefore, even in the same alert output process, an alert with a high damage risk having actual damage in the past can be output, and for example, for the current typhoon or a typhoon that may come in the future, stronger recommendations for evacuation indoors, etc. can be made.

[0104] Also, it is possible to control so that the output destination of the alert information is different depending on whether it belongs to the "weather phenomenon with high damage risk" or not. As described above, in the second embodiment, alert information can be output to an information distribution system registered by a plurality of users or a web server that publishes information on a website. In this case, for example, when it belongs to the "weather phenomenon with high damage risk", the alert control unit 223 determines that it is urgent and directly provides the alert information to the terminals of the users registered in the information distribution system. When it does not belong to the "weather phenomenon with high damage risk", the alert information can be output to the information distribution system or the web server, and the terminals of each user can be configured to receive the alert information via the information distribution system or the web server. Note that, even when it belongs to the "weather phenomenon with high damage risk", the alert information belonging to the "weather phenomenon with high damage risk" may be provided to the information distribution system or the web server without changing the output destination, and may be provided from the information distribution system or the web server to the terminals of each user.

[0105] As described above, the information management device 100 of the present embodiment classifies weather-related posting information used to generate a true value (correct label) based on the past insurance payment history into a group of weather phenomena with high damage risk and other groups. That is, a plurality of weather-related posting information (true values (correct labels)) are classified according to the risk level based on the past insurance payment history. For example, even for the true value (correct label) of the same weather phenomenon "hail", the difference in risk level can be distinguished. Thereby, feedback data according to the risk level can be provided for the analysis based on the observation data of the weather radar, and the analysis accuracy can be further improved.

[0106] Furthermore, the information management device 100 of the present embodiment accumulates, as learning data, weather-related post information classified into the "weather phenomena with high damage risk" group based on past insurance payment histories, and generates a classification model for determining "weather phenomena with high damage risk" using the learning data. As a result, when applied to the second embodiment described above, the weather management device 200 can provide different alert outputs according to the classification determination results of "weather phenomena with high damage risk" with true values (correct labels) added, and can control the change values of the determination criteria to be different according to the degree of risk. Thereby, the weather management device 200 can appropriately raise awareness of disaster prevention in real time.

[0107] For example, when large hail with a high degree of risk falls, there is a possibility that the bonnet of a car will be damaged or that damage will occur to agricultural crops. However, if this can be known even immediately before or in advance, it may be possible to reduce damage by taking actions such as moving the car under a roof or covering the agricultural crops with a sheet. The same applies to weather phenomena other than hail. If the occurrence of gusts can be known in advance, it may be possible to prevent damage by suspending railway operations, closing highways to traffic, or stopping work at heights. Although normal alert output is also an effective means of prompting actions to prevent damage, it can be used as a means of strongly recommending actions to prevent damage as an alert output with a high degree of risk for which actual damage has occurred in the past.

[0108] In addition, prompting actions to prevent damage also has the following merits from the perspective of insurance contracts. That is, an insurance company that provides an insurance product that covers damage by insurance notifies the policyholder of the alert information and has the policyholder take actions to reduce damage. As a result, the policyholder can prevent damage, and the insurance company can reduce the payment of insurance benefits, which is beneficial for both parties.

[0109] Although the first to third embodiments have been described above, the information management device 100 can provide true values (correct labels) of meteorological phenomena other than hail and sleet, such as tornadoes and gusts. In this case, the weather-related posting information is posting information related to hail, sleet, tornadoes or gusts, and the information collection unit 121 can obtain weather-related posting information including tornadoes and gusts as keywords or / and similar images from the web server 300. Then, the extraction unit 122 extracts weather-related posting information that matches the observation area and observation time of the observation data of tornadoes or gusts observed by the weather radar. The data management unit 123 can output the meteorological phenomena of tornadoes or gusts grasped from the extracted weather-related posting information as meteorological phenomena corresponding to the observation data.

[0110] One method of observing tornadoes and gusts is weather Doppler radar observation. In addition to the position and intensity of precipitation, the Doppler effect of radio waves reflected from precipitation particles carried by the wind can be used to measure the component of the wind approaching the weather radar and the component of the wind moving away. This is called the Doppler velocity. When the approaching velocity is equal to or greater than a predetermined value and the receding velocity is equal to or greater than a predetermined value, it can be determined that a wind vortex (rotation in a developed cumulonimbus cloud that brings about a tornado; mesocyclone) is occurring. For example, when the Doppler velocity is equal to or greater than a predetermined value, it can be determined that a tornado is occurring.

[0111] In this way, the weather management device 200 can obtain observation data of the Doppler velocity from the weather Doppler radar and perform analysis processing on the possibility of the occurrence of tornadoes and gusts. Then, the weather management device 200 can provide the observation data of the Doppler velocity to the information management device 100.

[0112] In addition, the weather prediction device 400 in the first embodiment can use a weather prediction model (prediction algorithm) of EHI (Energy-Helicity Index) for tornadoes and gusts. In this case, as the fourth processing unit 121D, the information collection unit 121 can generate a weather event evaluation value for the candidate post information based on the prediction information generated by the weather prediction model of tornadoes or gusts corresponding to the position information and time information of the candidate post information. Then, the information collection unit 121 can extract weather-related post information from the candidate post information in consideration of the weather event evaluation value (extract weather-related post information from candidate post information whose evaluation value of the prediction of the occurrence of a tornado or gust is equal to or higher than a predetermined value). Known methods other than EHI can also be applied to the weather prediction model for tornadoes and gusts.

[0113] Then, the weather management device 200 of the second embodiment described above can perform each process of outputting alert information and changing the determination criteria for tornadoes and gusts. In the process of changing the determination criteria, the threshold value of the Doppler velocity for determining the occurrence of a tornado can be changed based on the true value (correct label) provided by the information management device 100. In addition, the function of the classification control unit 124 of the third embodiment described above can also be configured for tornadoes, gusts, and squalls. In this case, the insurance payment history information includes tornadoes, gusts, and squalls as weather phenomena related to the occurrence of damage.

[0114] Note that each of the devices 100 and 200 in the first, second, and third embodiments described above may have a device configuration targeting any one of hail, snow pellets, tornadoes, and gusts, or may have a device configuration targeting all of hail, snow pellets, tornadoes, and gusts, or any combination thereof.

[0115] (Fourth Embodiment) Figures 13 to 17 are diagrams for explaining the fourth embodiment. This embodiment has a mechanism for improving the analysis function (weather detection function) of the weather management device 200 of the second embodiment above by using weather-related posting information (true value information) provided from the information management device 100 of the first embodiment above. For the same configurations as those in the second embodiment and the like, the same reference numerals are given in each figure and the description thereof is omitted.

[0116] Figure 13 is a diagram showing the functional blocks and network configuration of the weather management device 200A of this embodiment. Basically, it has the same configuration as FIG. 5, but the analysis unit 221 is configured as a weather detection unit 221A having a weather detection model 221a. The weather detection model 221a performs a type discrimination process for precipitation particles in the upper-air clouds in the observation data based on a determination criterion, and performs a weather detection process for outputting the type of precipitation particles pouring onto the ground surface based on the result of the type discrimination process.

[0117] That is, the weather detection model 221a (analysis unit 221) performs a type discrimination process (type estimation process) for precipitation particles based on the ratio of amplitudes observed by a weather radar, and classifies them into various types of precipitation particles such as wet snow, dry snow, ice crystals, drizzle, rain, hail, sleet, and heavy rain according to the magnitude of the ratio of amplitudes. At this time, a determination criterion is set for each type of precipitation particle, and the weather detection model 221a determines which type of determination criterion the ratio of amplitudes observed by the weather radar belongs to, and discriminates the type of one or more precipitation particles existing in the upper-air clouds. Then, the weather detection model 221a outputs the type of precipitation particles pouring onto the ground surface based on the result of the type discrimination process for the precipitation particles in the upper-air clouds.

[0118] The information output from the weather detection unit 221A can be configured to include the type of precipitation particles pouring onto the ground surface (for example, sleet) discriminated by the weather detection model 221a, and the location and time based on the observation area and observation time of the observation data. The location can also be configured to output the corresponding area (for example, XX town, XX city, XX prefecture) from the latitude and longitude of the observation area based on the map information.

[0119] Further, it can also be configured to output information representing the reliability of the detected weather phenomenon together with the location and time. The reliability will be described later.

[0120] The determination criterion control unit 222 is configured as a learning unit 222A of the weather detection model 221a. The learning unit 222A generates the weather detection model 221a, uses the weather phenomena grasped from the weather-related posting information as learning data (correct labels), learns the types of precipitation particles that actually poured onto the ground surface with respect to the types of upper-air precipitation particles based on the observation data, and performs a learning process (tuning process) to change the determination criterion.

[0121] Moreover, the weather management device 200A of the present embodiment can be configured to include a setting unit 221B for setting priorities for a plurality of types of precipitation particles. In the clouds in the upper air, there are a plurality of different types of precipitation particles. For example, from the observation data, it is possible to grasp the distribution situation of the hail area, sleet area, snow area, and rain area in the clouds in the upper air. That is, in the clouds in the upper air, there are a plurality of different types of precipitation particles in three dimensions (latitude, longitude, altitude).

[0122] On the other hand, the type of precipitation particles that actually pour onto the ground surface needs to be captured in two dimensions (latitude, longitude). In the present embodiment, the setting unit 221B controls so that priorities can be set for each type of precipitation particle. Then, when the weather detection model 221a discriminates the type of each upper-air precipitation particle existing within the observation area of the observation data and outputs the type of precipitation particle that pours onto the ground surface, it outputs the type of precipitation particle that pours onto the ground surface in descending order of the set priorities. For example, assume that the priorities are set in descending order of hail > sleet > snow > rain. When the weather detection model 221a detects two precipitation particles of "hail" and "sleet", it can output the "hail" with the higher priority as the type of precipitation particle that pours onto the ground surface.

[0123] Although the mode of output control using the priority by the setting unit 221B has been described, it is not limited to this. For example, control may be performed to output only "hail" with large damage as a natural disaster. That is, the weather detection model 221a may be configured to output the type of precipitation particles pouring on the ground surface when a specific weather phenomenon is detected. Further, the setting change of the priority can be performed at an arbitrary timing. For example, during the weather detection process, the priority of snow may be made higher than that of hail, and the output may be switched so that "snow" is preferentially output as precipitation particles from the middle.

[0124] As described above, the weather management device 200A of this embodiment includes the communication device 210, the control device 220, and the storage device 230. The control device 220 includes a weather detection unit 221A (corresponding to the analysis unit 221) including the weather detection model 221a, a setting unit 221B, a learning unit 222A (corresponding to the determination criterion control unit 222), an alert control unit 223, and a true value information acquisition unit 224.

[0125] Next, the weather detection model 221a and the learning unit 222A will be described. FIG. 14 is a diagram showing a processing flow executed by the weather management device 200A. FIG. 14 corresponds to FIG. 7 of the second embodiment and is a flowchart adapted to the processing of this embodiment. Steps S201a to S207a correspond to steps S201 to 207 of FIG. 7.

[0126] The weather detection model 221a is a functional unit that sequentially receives observation data observed by a weather radar, discriminates and outputs the current weather phenomenon (the type of precipitation particles pouring on the ground surface) using the observation data as an input (S201a). The weather detection unit 221A stores the output weather detection result in the storage device 230 (S202a), and the alert control unit 223 performs alert output processing (203a). In the example of FIG. 14, the priority setting process by the setting unit 221B is performed in advance (S202A).

[0127] The learning unit 222A performs learning processing using learning data to generate a weather detection model 221a (S201A). The learning data is observation data with weather-related posting information as the correct level, and is information obtained by setting past observation data by a weather radar and the weather-related posting information corresponding to the observation data. In the example of FIG. 13, it corresponds to the observation data 231 and the true value information 233, and is stored in the storage device 230.

[0128] The learning unit 222A performs a first learning process (step S201A) for generating the weather detection model 221a using the accumulated past learning data, and a second learning process for tuning (adjusting) the weather detection model 221a generated by the first learning process. The second learning process tunes (adjusts) the criteria for classifying the observation data into multiple types of precipitation particles, and updates the weather detection model 221a or generates individual weather detection models 221a with different criteria by tuning as described later.

[0129] The second learning process is divided into a real-time process (step S206a) that is executed in real time at the timing when weather-related posting information (true value information) is input, and a batch process (S210a) that holds the second learning process even when weather-related posting information is input and executes the second learning process at a predetermined timing. And the tuning methods include a first tuning method that executes only the real-time process, a second tuning method that combines the real-time process and the batch process, and a third tuning method that executes only the batch process.

[0130] The second tuning method performs the second learning process in real time to tune the judgment criteria at the timing when weather-related posting information (true value information) is input during the weather detection process by the weather detection model 221a (S206a). Then, at an arbitrary timing when the weather detection process is not being performed, for example, during a maintenance period, the judgment criteria are retuned (relearning process) using a plurality of observation data and weather-related posting information accumulated from the previous maintenance to the current maintenance, and the weather detection model 221a is updated (S210a). FIG. 14 shows the processing flow of this second tuning method.

[0131] Although the first tuning method does not perform the second learning process by batch processing, since it tunes by performing the second learning process during the weather detection process in the same manner as the second tuning method, when the weather phenomenon grasped from the observation data is different from the weather phenomenon grasped from the weather-related posting information, the judgment criteria can be changed in real time using the weather-related posting information as the correct label. Therefore, the weather detection process can be performed on subsequent observation data using the tuned judgment criteria.

[0132] On the other hand, in the case of the third tuning method, the weather detection model 221a is not updated in real time. Therefore, when the weather phenomenon grasped from the observation data is different from the weather phenomenon grasped from the weather-related posting information, the weather-related posting information corresponding to the previous observation data is not reflected as the correct label in the output result (the type of precipitation particles pouring down on the ground surface) of the weather detection unit 221A for subsequent observation data. Therefore, when applying the third tuning method, as a temporary measure until the second learning process is performed, a replacement process is performed to replace the type of precipitation particles output from the weather detection model 221a with the weather phenomenon (correct label) grasped from the weather-related posting information corresponding to the previous observation data.

[0133] FIG. 15 is a diagram showing a processing flow according to the third tuning method. The weather detection unit 221A determines whether the type of precipitation particles output from the weather detection model 221a based on the observation data is different from the weather phenomenon grasped from the corresponding weather-related posting information (S205a). When the weather detection unit 221A determines that they are different weather phenomena (YES in S205a), it temporarily holds the feature amount of the observation data and the weather phenomenon grasped from the weather-related posting information (S2010), and turns on the replacement processing flag (S2011). Note that the feature amount is information representing the weather state grasped from the observation data, for example, the weather state in the upper air clouds such as the distribution range of each precipitation particle and the altitude at which the precipitation particles exist. Note that the feature amount extracted from the observation data can be calculated by a known method and can be represented by, for example, a vector of parameters indicating the weather state.

[0134] Then, when weather detection processing is performed on subsequent observation data and the type is determined, in step S2012, it is confirmed whether the replacement processing flag is ON. If the replacement processing flag is ON, the process proceeds to step S2013, and the similarity between the feature amount of the subsequent observation data and the feature amount of the previously held observation data is determined. The determination of similarity is, for example, to determine whether the feature amount of the subsequent observation data is in a certain weather state at a certain distance from the feature amount (weather state vector) of the previously held observation data. And if the feature amount (weather state vector) of the subsequent observation data is at a certain distance from the feature amount of the previous observation data, it can be determined that they are similar. For this vector distance, known methods such as Euclidean distance, Manhattan distance, and Mahalanobis distance can be applied.

[0135] Then, in step S2014, when it is determined that the subsequent observation data is similar to the previous observation data, the correct label based on the previous weather-related posting information is compared with the result of the weather detection process for the subsequent observation data. When the result of the weather detection process for the subsequent observation data is different from the correct label based on the previous weather-related posting information, the weather detection unit 221A outputs the type of precipitation particles with the correct label based on the previous weather-related posting information being positive. On the other hand, when it is determined that the subsequent observation data has no similarity with the previous observation data, the result of the weather detection process for the subsequent observation data is output.

[0136] By configuring in this way, even if the determination criteria (weather detection model 221a) are not changed in real time, it is possible to output the result of the weather detection process that reflects the correct label based on the previous weather-related posting information in real time.

[0137] Next, as described in the above second embodiment, in accordance with the mode of holding different determination criteria respectively, also in this embodiment, the weather detection model 221a after tuning can be generated as an individual weather detection model different from the weather detection model 221a before tuning. FIG. 16 is a diagram showing a processing flow in which a plurality of weather detection models with different determination criteria are generated in the real-time processing of the second learning process, and the corresponding weather detection model is selected for the input observation data to perform the weather detection process.

[0138] As shown in FIG. 16, when the weather phenomenon grasped from the observation data is different from the weather phenomenon grasped from the weather-related posting information in the tuning process (YES in S205a), the learning unit 222A performs a learning process with the weather-related posting information as the correct label based on the currently used weather detection model 221a, and individually generates a weather detection model 221a (corresponding to the first weather detection model) with the determination criteria changed (S2061a). Then, the learning unit 222A sets the effective area and effective time for the individually generated weather detection model 221a based on the observation area and observation time of the observation data (S2071a).

[0139] When subsequent observation data is input, the weather detection unit 221A selects a weather detection model 221a that matches the observation area and observation time of the subsequent observation data from a plurality of weather detection models 221a with different effective areas and effective times (S2001a), and controls to perform weather detection processing on the subsequent observation data using the selected weather detection model 221a (S201a).

[0140] In step S2001a, when there is no weather detection model 221a that matches the observation area and observation time of the subsequent observation data, the weather detection model 221A can be configured to select the weather detection model 221a (reference model) generated in step 201A or re-learned in step S210a and perform weather detection processing.

[0141] With this configuration, a plurality of each weather detection model can be switched and used. That is, the weather detection model 221a generated in step 201A or re-learned in step S210a can be configured as a reference model for long-term use. Each weather detection model 221a individually generated in the real-time processing in the second learning process can be applied as a short-term usable model with at least one of the judgment criteria, effective area, and expiration date being different. Then, the weather detection unit 221A can switch from the reference model for long-term use to the model for short-term use according to the observation data, or switch between a plurality of models for short-term use, so as to improve the processing accuracy of the weather detection processing. When the effective time (expiration date) of the applied weather detection model 221a has passed, the weather detection unit 221A can be controlled to switch to the reference model and perform weather detection processing.

[0142] Here, the reliability output from the weather detection unit 221A will be described. The weather detection model 221a uses weather-related posting information used as learning data. Therefore, based on each evaluation value regarding the weather-related posting information of the first embodiment, the reliability of the weather detection model 221a can be calculated. And the reliability of the weather detection model 221a can be configured to be output together with the result of the weather detection process. The weather detection unit 221A can output information including, for example, the weather phenomenon "hail", the area "〇〇 City, 〇〇 Prefecture", and the reliability "80%".

[0143] Specifically, as described in the first embodiment, the weather-related posting information can be evaluated by the contributor evaluation value, the posting accuracy evaluation value, and the weather event evaluation value. Therefore, the information management device 100 provides the weather management device 200A with the ground truth information and each evaluation value of the weather-related posting information on which the ground truth information is based. Then, the reliability of the weather detection model 221a can be calculated using the sum of each evaluation value of the weather-related posting information used as learning data, or any one of the evaluation values. When performing the learning process using a plurality of different weather-related posting information, the average value, median value, etc. of the evaluation values of each weather-related posting information can be calculated and used as the reliability of the weather detection model 221a. The method for calculating the reliability is arbitrary.

[0144] On the other hand, a highly reliable weather detection model 221a can also be generated. That is, the evaluation value of weather-related posting information can be used as an index for selecting the weather-related posting information to be used as learning data. For example, the weather-related posting information accumulated over a certain period of time is sorted in descending order of reliability, and learning is performed in order from the most reliable. Then, the learning process is terminated when the desired result is obtained. For example, as sample test data, weather detection processing is performed with another observation data that outputs a different type from the correct label as input, and it is verified whether the same type of weather phenomenon as the correct label is output. When the verification result is equal to or higher than a predetermined value, the learning may be terminated. Alternatively, the learning process may be configured to use learning data having an evaluation value equal to or higher than a certain value. By configuring in this way, it is possible to suppress the learning process using learning data with low reliability as weather-related posting information, and avoid overfitting of the weather detection model 221a.

[0145] The learning method of the weather detection model 221a, that is, the method of tuning the determination criteria by the learning unit 222A, can be performed by any of the following means, or by combining two or more of the following means. The type of precipitation particles obtained from the weather detection model 221a through the type discrimination process of the observation data is "A", and the type of precipitation particles (weather phenomenon) grasped from the weather-related posting information is described as "C".

[0146] (A) Extract observation data within a certain similarity range from the weather phenomena (weather conditions) extracted from the input observation data from among the past learning data learned so far. Then, replace the correct label associated with the extracted observation data from "A" to "C". That is, as learning data, the correct label of the observation data with a certain similarity is replaced with the type "C" of the precipitation particles grasped from the weather-related posting information. Then, using the observation data group whose correct label has been replaced with "C" as learning data, re-learning processing is performed to update the weather detection model 221a.

[0147] In addition, for the determination of similarity, a similarity determination process for comparing the feature quantities between the above-described observation data can be applied. At this time, the threshold value for determining whether or not it is within a certain similarity range can be dynamically changed. For example, when the output result of the updated weather detection model 221a is still the precipitation particle type "A", the threshold value for determining whether or not it is within a certain similarity range can be updated (increased), the distribution range applied as learning data can be increased, and the learning process can be performed again to update the weather detection model.

[0148] (B) In the case of weather detection processing using a probability density function or the like, the probability of each precipitation particle type (classification) can be calculated based on the input observation data, and the type with the highest probability can be output. In this case, the learning unit 222A can increase the probability that the input observation data is classified into the type "C" or decrease the probability that it is classified into the type "A". The increase amount and decrease amount of the probability may use fixed values set in advance, or the values of the increase amount and decrease amount may be dynamically changed.

[0149] In addition, when increasing and decreasing the probability, an offset may be uniformly added or subtracted, or the probability density function itself may be changed. When changing the probability density function, the mean and variance of the probability density function may be changed respectively, and this change amount may also be a value set in advance or a dynamically changed value. For example, to increase the probability, the mean value of the probability density function of the type "C" is changed so as to be close to the input value of the probability density function extracted from the input observation data. Conversely, to decrease the probability, the mean value of the probability density function of the type "A" is changed so as to be away from the input value of the probability density function extracted from the input observation data. The result based on this updated probability can be used as the final output. If the result based on the updated probability is still the type "A", the same operation may be repeated until the desired result is obtained.

[0150] (C) As described in (B) above, a probability density function or the like can be used, and a probability density function or the like can be provided for each feature quantity of the meteorological phenomenon (feature quantity of the high-dimensional vector) extracted from the input observed data. In this case, the probabilities of the respective feature quantities can be multiplied to calculate the final probabilities for each category. At this time, the probabilities of each feature quantity may be uniformly increased, or the amount of probability to be dynamically increased may be allocated according to the result based on the probability. For example, it is configured to have a probability density function for each of feature quantity 1, feature quantity 2, and feature quantity 3, and the probabilities of each feature quantity are calculated for category "B". When the probability calculated from feature quantity 1 is 0.99, the probability calculated from feature quantity 2 is 0.91, and the probability calculated from feature quantity 3 is 0.95, since the probability of feature quantity 2 is low, the probability of feature quantity 2 is increased significantly. As the method for increasing or decreasing the probability, the same method as in (B) above can be applied.

[0151] (D) When using the probability density function or the like in (B) above, if a plurality of meteorological-related posting information with high evaluation values (reliability) is obtained, the probability density function may be configured to be updated using these plurality of correct labels. For example, using the probability density function as a prior distribution, a Bayesian update method for obtaining a posterior distribution from reliable results can be utilized. At this time, this posterior distribution can be used as the probability density function to update the probability density function of category "C", and the result based on this updated probability can be used as the final output. If the result based on the updated probability still outputs category "A", the same operation can be repeated until the desired result is obtained.

[0152] In the above description, an example has been described in which the weather management device 200 is in cooperation with the individually configured information management device 100, but it is not limited thereto. For example, as shown in FIG. 17, it can also be configured as a weather management device or a weather management system having each functional part and each information of the information management device 100. In this case, it can be configured as a weather management system having the following configuration.

[0153] (1) The weather management system (weather management device 200A) acquires weather-related posting information from a plurality of posting information posted to a web server that receives posting information from a poster terminal, and extracts the weather-related posting information that matches the observation area and observation time of the observation data observed by a weather radar. An information management unit, A weather detection unit having a weather detection model that performs a type discrimination process on precipitation particles in clouds in the sky based on observation data and outputs the type of precipitation particles pouring onto the ground based on the result of the type discrimination process. Using the weather phenomenon grasped from the weather-related posting information as learning data, learning the type of precipitation particles that actually poured onto the ground with respect to the type of precipitation particles in the sky based on the observation data, and generating a weather detection model. It has a learning unit.

[0154] And the type discrimination process of the weather management system of (1) above can be configured as a process of classifying the weather phenomenon in the sky based on the observation data into the corresponding type based on the determination criteria set for each type of precipitation particle. The learning unit performs a learning process using the weather phenomenon grasped from the weather-related posting information corresponding to the observation data as the correct label, and tunes the determination criteria. Furthermore, the weather management system of (1) above can include the functions of a weather detection unit 221A and a learning unit 222A including the weather detection model 221a of the present embodiment.

[0155] Also, the replacement process (steps S201a to S208) described in FIG. 15 can be configured as a weather management device that reflects true value information by the replacement process without changing the determination criteria, and can have the following configuration.

[0156] (3) The weather management device An analysis unit that classifies the observation data of the weather radar based on a predetermined determination criterion and outputs an analysis result based on the classification. An information acquisition unit that acquires weather-related posting information corresponding to the observation data output from the information management device according to claim 1. The analysis unit performs classification processing of precipitation particles in the upper clouds in the observation data based on the determination criteria set for each type of precipitation particle, and outputs the type of precipitation particles pouring onto the ground surface based on the result of the classification processing. The analysis unit compares the meteorological phenomenon grasped from the meteorological-related contribution information corresponding to the observation data with the type of precipitation particles pouring onto the ground surface that is output, and if it is determined that they are different meteorological phenomena, the observation data and the meteorological-related contribution information determined to be different meteorological phenomena are retained. The analysis unit determines the similarity between the subsequent observation data and the retained observation data, and if it is determined that there is similarity between the subsequent observation data and the retained observation data, the meteorological phenomenon grasped from the retained meteorological-related contribution information is compared with the type of precipitation particles pouring onto the ground surface output for the subsequent observation data. If the meteorological phenomenon obtained from the subsequent observation data is different from the meteorological phenomenon grasped from the retained meteorological-related contribution information, the meteorological phenomenon grasped from the retained meteorological-related contribution information is output as the meteorological detection result for the subsequent observation data.

[0157] Furthermore, similar to the example of FIG. 17, as a weather management device or a weather management system having each functional unit and each piece of information of the information management device 100, it can be configured to have the following configuration.

[0158] (4) The weather management system includes an information management unit that acquires meteorological-related contribution information from a plurality of contribution information posted to a web server that receives contribution information from a contributor terminal, and extracts the meteorological-related contribution information that matches the observation area and observation time of the observation data observed by a weather radar, a weather detection unit that performs weather detection processing for performing classification processing of precipitation particles in the upper clouds based on the observation data and outputting the type of precipitation particles pouring onto the ground surface based on the result of the classification processing. Then, the weather detection unit compares the weather phenomenon grasped from the weather-related posting information corresponding to the observation data with the type of precipitation particles pouring onto the ground surface output by the weather detection process. When it is determined that they are different weather phenomena, the observation data and the weather-related posting information determined to be different weather phenomena are retained. Furthermore, the weather detection unit determines the similarity between the subsequent observation data and the retained observation data. When it is determined that there is similarity between the subsequent observation data and the retained observation data, the weather phenomenon grasped from the retained weather-related posting information is compared with the type of precipitation particles pouring onto the ground surface output by the weather detection process using the subsequent observation data as input. When the weather phenomenon obtained from the subsequent observation data is different from the weather phenomenon grasped from the retained weather-related posting information, the weather phenomenon grasped from the retained weather-related posting information is output as the weather detection result for the subsequent observation data.

[0159] In this way, the weather management device or weather management system described in (3) and (4) above can have an aspect of a system configuration that does not assume learning processing or the like that reflects the true value information based on the weather-related posting information in the weather detection process itself. For example, a system that performs existing weather detection processing is configured to have a replacement processing function that replaces the type of precipitation particles output in the weather detection process with the weather phenomenon (correct label) grasped from the weather-related posting information corresponding to the previous observation data while cooperating with the information management device 100 that manages the weather-related posting information. By configuring in this way, it is possible to improve the accuracy of weather detection without changing the weather detection process itself.

[0160] In the alert output process of the alert control unit 223, it may be configured to obtain a large amount of highly reliable weather-related posting information as a weather phenomenon. As described in the second embodiment, the alert control unit 223 can generate and output alert information such as "There may be hail in City A. Please be careful." At this time, a message such as "What kind of weather phenomenon do people near City A see? Please post it." is inserted into the alert information. By configuring in this way, the information management device 100 can recruit (collect) a large amount of weather-related posting information, and the weather management device 200A can obtain a large amount of learning data. Also, when providing alert information to the users (user terminals) registered in the weather management device 200, it is also possible to push notifications to the user terminals so as to collect a large amount of weather-related posting information. In this way, by introducing a mechanism that promotes posting in conjunction with the alert output process, it is possible to obtain postings for highly reliable (certain) weather phenomena and improve the accuracy of the weather detection model 221a.

[0161] Also, in the same manner as in the first to third embodiments, this embodiment can also perform a weather detection process capable of detecting weather phenomena such as tornadoes and gusts.

[0162] As described above, each function constituting the above-described information management device 100 and weather management device 200 can be realized by a program. A computer program prepared in advance to realize each function is stored in an auxiliary storage device, and a control unit such as a CPU reads the program stored in the auxiliary storage device into the main storage device and executes the program read into the main storage device, whereby the functions of each part can be operated.

[0163] Also, the above program can be provided to a computer in a state recorded on a computer-readable recording medium. Examples of computer-readable recording media include optical discs such as CD-ROMs, phase change optical discs such as DVD-ROMs, magneto-optical discs such as MO (Magneto Optical) and MD (Mini Disk), magnetic discs such as floppy (registered trademark) discs and removable hard discs, and memory cards such as compact flash (registered trademark), smart media, SD memory cards, and memory sticks. In addition, hardware devices such as integrated circuits (IC chips, etc.) specially designed and configured for the purpose of the present invention are also included as recording media.

[0164] Note that although the embodiments of the present invention have been described, these embodiments are presented as examples and are not intended to limit the scope of the invention. This novel embodiment can be implemented in various other forms, and various omissions, replacements, and changes can be made without departing from the gist of the invention. These embodiments and their modifications are included in the scope and gist of the invention, and are also included in the invention described in the claims and its equivalent scope.

Explanation of Reference Numerals

[0165] 100 Information management device 110 Communication device 120 Control device 121 Information collection unit 121A First processing unit 121B Second processing unit 121C Third processing unit 121D Fourth processing unit 122 Extraction unit 123 Data management unit 124 Classification control unit 130 Storage device 131 Submitted information 132 Submitter evaluation information 133 True value information 134 Classification-related information 200, 200A Weather management device 210 Communication device 220 Control device 221 Analysis unit 221A Weather detection unit 221a Weather detection model 221B Setting unit 222 Judgment criterion control unit 222B Learning unit 223 Alert control unit 224 True value information acquisition unit 230 Memory device 231 Observation data 232 Judgment criterion information 233 True value information 234 Setting information 300 Web server 400 Weather prediction device T Terminal (contributor terminal)

Claims

1. An analysis unit that classifies the observation data of a weather radar based on a predetermined criterion and outputs an analysis result based on the classification; A criterion control unit that changes the criterion based on a weather phenomenon grasped from the weather-related posting information posted from a poster terminal and matching the observation data observed by the weather radar; An alert control unit that outputs alert information based on the analysis result, and has, The alert control unit outputs a message to the user of the user terminal that receives the alert information, indicating that a weather phenomenon based on the analysis result may have already occurred, or a message indicating that a weather phenomenon based on the analysis result may occur in the future. A weather management system characterized by that.

2. The weather management system according to claim 1, wherein the alert control unit outputs a message related to weather-related posting information.

3. A program executed by a computer, which causes the computer to, Have a first function of classifying the observation data of a weather radar based on a predetermined criterion and outputting an analysis result based on the classification; Have a second function of changing the criterion based on a weather phenomenon grasped from the weather-related posting information posted from a poster terminal and matching the observation data observed by the weather radar; Have a third function of outputting alert information based on the analysis result, and The third function outputs a message to the user of the user terminal that receives the alert information, indicating that a weather phenomenon based on the analysis result may have already occurred, or a message indicating that a weather phenomenon based on the analysis result may occur in the future. A program characterized by that.

4. A weather management method, wherein a computer, Performs a first step of classifying the observation data of a weather radar based on a predetermined criterion and outputting an analysis result based on the classification; Performs a second step of changing the criterion based on a weather phenomenon grasped from the weather-related posting information posted from a poster terminal and matching the observation data observed by the weather radar; Performs a third step of outputting alert information based on the analysis result. The third step is a weather management method characterized by outputting a message to the user of the user terminal that receives the alert information, indicating that there may already be a weather phenomenon based on the analysis result, or a message indicating that a weather phenomenon based on the analysis result may occur in the future.

Images