Moisture damage prediction device, moisture damage prediction method, and moisture damage prediction program

Abstract

To enable easy identification of locations predicted to suffer moisture damage in a farm field at low cost.SOLUTION: A moisture damage prediction device is provided, comprising an identification unit (31) configured to identify dangerous locations predicted to suffer moisture damage in a particular farm field using a trained model generated by machine learning using a dataset comprising geospatial information of the particular farm field and moisture damage information showing locations that have suffered moisture damage in the past.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to a moisture damage prediction device, a moisture damage prediction method, and a moisture damage prediction program for identifying risky locations where moisture damage is predicted to occur. [Background technology]

[0002] In fields where vegetables, grains, and other crops are grown, moisture damage is a problem in which the amount of moisture in the soil is excessive, causing crops to grow poorly. Therefore, appropriate drainage measures are required for fields where moisture damage may occur. As a prerequisite for taking appropriate drainage measures, it is first necessary to investigate whether there are any fields where moisture damage may occur.

[0003] The possibility of moisture damage occurring is judged by monitoring the soil moisture and groundwater level in the field during the crop cultivation period. In this monitoring, various sensors are installed, such as a soil moisture sensor in the field and a water level sensor in a well dug for observation, and the values ​​obtained by these sensors are recorded over time and analyzed.

[0004] In recent years, methods have been proposed for evaluating whether or not moisture damage has occurred in a field by analyzing satellite images of the field or images of the field taken by an unmanned aerial vehicle (UAV) (Non-Patent Documents 1 and 2). [Prior art documents] [Non-patent literature]

[0005] [Non-Patent Document 1] Katsuhisa Tanba et al., Satellite remote sensing for understanding the actual condition of sugar beet moisture damage in the Tokachi region of Hokkaido in 2009, Abstracts of the Japanese Society of Soil Science and Plant Nutrition, Vol. 56, 16-1, 2010 [Non-Patent Document 2] Shuhei Yamamoto et al., Evaluation of soybean moisture damage in farm fields based on UAV remote sensing - Observation example in the Sendai coastal area in 2017 - (Flash report), Journal of the Crop Science Society of Japan, Vol. 88, No. 1, p. 48-49, 2019 Summary of the Invention [Problem to be solved by the invention]

[0006] However, the above-mentioned method of monitoring the state of a field using various sensors can only determine the location of the sensor in the field. Since the soil condition and topography of a field are usually uneven, there is a problem that the judgment result is likely to differ greatly depending on the location of the sensor. Therefore, it is preferable to install various sensors at multiple locations in the field, but in this case, a lot of effort is required to install and monitor the various sensors. In addition, a large number of sensors are required to comprehensively judge the entire field, but the equipment such as sensors used for monitoring is generally expensive and is not realistic from the viewpoint of cost.

[0007] On the other hand, the conventional techniques shown in Non-Patent Documents 1 and 2 evaluate whether or not moisture damage has occurred based mainly on the growth state of agricultural crops, such as leaf color, etc. In other words, these conventional techniques do not determine whether or not moisture damage is likely to occur in a farm field.

[0008] One aspect of the present invention aims to realize a moisture damage prediction device or the like that can identify locations in a farm field where moisture damage is predicted to occur simply and at low cost. [Means for solving the problem]

[0009] In order to solve the above problems, a moisture damage prediction device according to one embodiment of the present invention includes an identification unit that uses a trained model generated by machine learning using a dataset of geospatial information of a specific field and moisture damage information indicating locations where moisture damage has previously occurred in the specific field to identify danger areas in the specific field where moisture damage is predicted to occur from a target image showing the specific field.

[0010] According to the above configuration, the moisture damage prediction device can identify risk areas in the target image where moisture damage is predicted to occur, based on a trained model that has learned the relationship between geospatial information and moisture damage information. Therefore, the moisture damage prediction device can identify risk areas easily and at low cost.

[0011] In the moisture damage prediction device according to one embodiment of the present invention, the identification unit may divide the target image into two or more images and identify the risk areas from each of the divided images.

[0012] According to the above configuration, the identification unit divides the target image into two or more images and identifies dangerous areas where moisture damage is predicted to occur. Therefore, even if the target image is a large image that includes a wide area, the identification unit can divide the target image into images of appropriate size and identify dangerous areas for each divided image. Therefore, the moisture damage prediction device can reduce the processing load and efficiently identify dangerous areas.

[0013] In addition, the identification unit can adjust the size of the image for identifying the dangerous spots by changing the number of divisions. Therefore, the identification unit can identify the dangerous spots after adjusting the scale of the divided images to be optimal for processing according to the size of the target image or the size of the farm field included in the target image.

[0014] In one embodiment of the moisture damage prediction device of the present invention, the geospatial information includes a plurality of geospatial characteristics, and the identification unit may identify the danger areas using the trained model generated for each of the geospatial characteristics.

[0015] According to the above configuration, the identification unit identifies risk areas based on each of the trained models generated from a plurality of geospatial characteristics. Therefore, the moisture damage prediction device can identify risk areas according to the geospatial characteristics that cause the risk areas.

[0016] In one embodiment of the moisture damage prediction device of the present invention, the geospatial information may include, as the geospatial characteristics, at least information indicating a groundwater flow system and information indicating unevenness of the ground surface.

[0017] According to the above configuration, the geospatial information includes information on the groundwater flow system and information on the unevenness of the ground surface. Therefore, the moisture damage prediction device can identify dangerous areas with higher accuracy by using the information on the groundwater flow system and the information on the unevenness of the ground surface, which are considered to be closely related to the occurrence of moisture damage.

[0018] In one embodiment of the moisture damage prediction device of the present invention, the trained model is generated by machine learning a generative adversarial network consisting of a generator and a discriminator, and the trained model may be a generator that inputs the target image as the geospatial information to the generator, causes the generator to generate a specific image indicating the dangerous area, inputs a generated pair consisting of the target image and the specific image and a correct pair consisting of the target image and a correct image indicating the moisture damage information to the discriminator, causes the discriminator to determine whether the generated pair is a pair including the specific image generated by the generator, and inputs the correctness of the determination to the generator and the discriminator to train the model.

[0019] According to the above configuration, the moisture damage prediction device can identify dangerous areas using a trained model that can identify dangerous areas with higher accuracy.

[0020] In order to solve the above problems, a moisture damage prediction method according to one embodiment of the present invention includes an identification step of identifying danger points in a specific field where moisture damage is predicted to occur from a target image showing the specific field, using a trained model generated by machine learning using a dataset of geospatial information of the specific field and moisture damage information indicating locations where moisture damage has previously occurred in the specific field.

[0021] The moisture damage prediction device of each aspect of the present invention may be realized by a computer. In this case, the moisture damage prediction program of the moisture damage prediction device, which realizes the moisture damage prediction device by causing the computer to operate as each part (software element) of the moisture damage prediction device, and the computer-readable recording medium on which it is recorded, are also included in the scope of the present invention. Effect of the Invention

[0022] According to one aspect of the present invention, a moisture damage prediction device or the like can be realized that is capable of identifying locations in a farm field where moisture damage is predicted to occur, simply and at low cost. [Brief description of the drawings]

[0023] [Figure 1] FIG. 2 is a diagram showing functional blocks of a moisture damage prediction device according to one embodiment. [Diagram 2] FIG. 2 is a diagram showing an overview of the processing of a control device provided in the moisture damage prediction device. [Diagram 3] FIG. 11 is another diagram showing an overview of the processing of the control device. [Figure 4] 4 is a flowchart showing a process flow of the control device. [Diagram 5] FIG. 2 is a block diagram showing an overview of the processing performed by the model generation unit using a generative adversarial network. [Figure 6] 13 is a flowchart showing the flow of a trained model generation process performed by the model generation unit. [Figure 7] A diagram showing functional blocks of a moisture damage prediction system according to a modified example of an embodiment. [Figure 8] A diagram showing functional blocks of a moisture damage prediction system relating to another modified example of an embodiment. [Figure 9] FIG. 13 is a diagram showing an example of a moisture damage prediction map according to a modified example of an embodiment. [Figure 10] FIG. 4 is a diagram showing a process of generating a moisture damage prediction map in the first embodiment. [Figure 11] FIG. 11 is a diagram showing a process of generating a moisture damage prediction map in Example 2. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0024] An embodiment of the present invention will now be described with reference to FIGS.

[0025] [Configuration of the moisture damage prediction device] The configuration of the moisture damage prediction device 100 will be described with reference to Fig. 1. The moisture damage prediction device 100 identifies dangerous areas in a field where moisture damage is predicted to occur, using a trained model generated by machine learning from various information about the field. For "dangerous areas", see, for example, "dangerous area R" in Fig. 2, "dangerous areas R1 and R2" in Fig. 3, and "dangerous area R3" in Fig. 9.

[0026] Then, the moisture damage prediction device 100 generates a moisture damage prediction map showing the identified dangerous spots. In this embodiment, the moisture damage prediction map is an image that allows a user of the moisture damage prediction device 100 to visually check the presence or absence and distribution of dangerous spots in the field. As shown in FIG. 1, the moisture damage prediction device 100 includes a control device 10, an input unit 40, an output unit 50, and a memory unit 60.

[0027] The control device 10 comprehensively controls each part of the moisture damage prediction device 100. The control device 10 includes a model generation device (trained model generation device) 20 and a map generation device 30.

[0028] The model generating device 20 generates a trained model used by the moisture damage prediction device 100 to identify risky locations. The model generating device 20 includes a data acquiring unit 21 and a model generating unit 22. The data acquiring unit 21 acquires geospatial information and moisture damage information input to the input unit 40. Details of the geospatial information and moisture damage information will be described later. The data acquiring unit 21 also acquires various types of information input to the input unit 40 other than the geospatial information and moisture damage information.

[0029] The data acquisition unit 21 may be included in the map generating device 30. In this case, the data acquisition unit 21 also has the function of the data acquisition unit of the model generating device 20.

[0030] The model generation unit 22 inputs the data set of the geospatial information and the moisture damage information acquired by the data acquisition unit 21 into the learning model as teacher data, and generates a trained model by subjecting the learning model to machine learning. Details of the learning model and the machine learning thereof will be described later.

[0031] The map generating device 30 includes an identifying unit 31 and a map generating unit 32. The identifying unit 31 identifies dangerous areas from a target image by using a trained model generated by the model generating unit 22. The target image is an image showing a specific farm field, and indicates an image used for identifying dangerous areas by the map generating device 30. Examples of the target image include a satellite photograph showing the specific farm field and an image taken by an unmanned aerial vehicle. Moreover, the target image is not limited to such photographs, and may be a map image showing the specific farm field.

[0032] The identification unit 31 may acquire the target image via a data acquisition unit 21 included in the model generation device 20. Furthermore, the map generation device 30 may have a data acquisition unit dedicated to the map generation device 30, separate from the data acquisition unit 21.

[0033] The map generating unit 32 generates a moisture damage prediction map by adding the moisture damage prediction information indicating the dangerous points identified by the identifying unit 31 to the target image. Here, the map generating unit 32 adds the moisture damage prediction information to the target image such that the target image and the moisture damage prediction information are displayed in a superimposed manner on the moisture damage prediction map.

[0034] The moisture damage prediction information may be added to the target image in such a manner that, for example, the image portion in the target image corresponding to the dangerous area is colored or framed on the target image. A moisture damage prediction map including such moisture damage prediction information makes it easy to visually recognize dangerous areas. The moisture damage prediction information added to the target image may be information in which an address, a place name, or coordinates such as latitude and longitude are expressed by characters and numbers. According to such a moisture damage prediction map including moisture damage prediction information, when the target image is an image showing a wide area including many farm fields, the possibility that a user of the moisture damage prediction map will overlook a small dangerous area can be reduced. Therefore, a user of such a moisture damage prediction map can reliably recognize, for example, a dangerous area that is narrow but highly in need of moisture damage countermeasures.

[0035] The input unit 40 receives data such as geospatial information and moisture damage information, and various other information input to the moisture damage prediction device 100. The input unit 40 is not particularly limited as long as it is a device that receives input of such data to the moisture damage prediction device 100. The input unit 40 may be, for example, an input port that inputs information from an external device, a camera, an image scanner, or a keyboard and mouse. The input unit 40 may also be a communication device that can receive various information transmitted from an external communication device (either wired communication or wireless communication is acceptable).

[0036] The output unit 50 outputs the moisture damage prediction map generated by the control device 10 to the user of the moisture damage prediction device 100. In addition, the output unit 50 outputs the moisture damage prediction map, as well as the processing results by the control device 10 and various information stored in the storage unit 60. The output unit 50 may be, for example, a display device for displaying the moisture damage prediction map, such as a liquid crystal display, or a printing device such as a printer that outputs the moisture damage prediction map on paper media. The output unit 50 may also be a communication device capable of transmitting the moisture damage prediction map to a terminal such as a smartphone. In other words, the output unit 50 can support at least one of the three output modes of "display", "printing" and "transmission". The output unit 50 may be configured as a touch panel that also functions as the input unit 40.

[0037] The storage unit 60 is a storage medium for storing information input to the moisture damage prediction device 100 and information generated by the moisture damage prediction device 100. The storage unit 60 may store programs for executing various functions of the control device 10. Specifically, the storage unit 60 may store a program for executing a machine learning algorithm for the model generation unit 22 to generate a trained model, and a trained model generated by such machine learning.

[0038] [Control device processing] (Overview of control unit processing) The processing of the control device 10 provided in the moisture damage prediction device 100 will be outlined with reference to Fig. 2 and Fig. 3. Note that in this specification, the processing of the control device 10 is mainly described as an example using a generative adversarial network (GAN) as a machine learning algorithm. However, as will be described later, the machine learning algorithm used in the processing of the control device 10 is not limited to a generative adversarial network.

[0039] As shown in Fig. 2, the information acquired by the moisture damage prediction device 100 includes geospatial information indicating a specific field and moisture damage information in the specific field. Geospatial information is information indicating the position of a specific point or area in space and information on various phenomena associated therewith. Geospatial information is utilized, for example, in Geographic Information Systems (GIS). Fig. 2 shows an overview of processing when the geospatial information includes information on a satellite photo, which is a target image.

[0040] The moisture damage information indicates locations of past moisture damage occurrences in the specific field. Here, the moisture damage information may indicate locations of past moisture damage occurrences in a part of the specific field. Furthermore, in a case where the specific field indicated in the geospatial information includes multiple fields, the moisture damage information may indicate locations of past moisture damage occurrences in only a part of the multiple fields.

[0041] The moisture damage prediction device 100 divides the acquired satellite photo or the like into two or more images. Then, the moisture damage prediction device 100 selects one of the divided images as a selected image. Note that if the geospatial information does not include image information such as a satellite photo, the geospatial information may be divided into two or more information groups for each range based on the position information or the like included in the geospatial information.

[0042] The moisture damage prediction device 100 generates a trained model and identifies a risk location R for each selected image. At this time, the moisture damage prediction device 100 generates a correct pair to be used for machine learning. The correct pair is a data set consisting of a selected image and a correct image. The correct image is an image showing a moisture damage occurrence location (moisture damage information) T in a specific area in a farm field corresponding to the selected image.

[0043] The moisture damage prediction device 100 generates a trained model by machine learning, and uses the trained model to identify risk areas R for all selected images. The moisture damage prediction device 100 adds moisture damage prediction information indicating the risk areas R to each selected image in a superimposed manner, and recombines these selected images to generate a moisture damage prediction map.

[0044] In this way, with a configuration in which dangerous areas R in a target image are identified from a trained model generated by machine learning, the moisture damage prediction device 100 can identify dangerous areas R easily and at low cost. That is, a user of the moisture damage prediction device 100 can identify dangerous areas R in a field from the geospatial information and moisture damage information of the field, without the effort and expense of installing sensors or the like in the field for monitoring. In addition, the moisture damage prediction device 100 can generate an image in which the identified dangerous areas R are easily visually recognized by displaying the target image and the moisture damage prediction information in a superimposed manner.

[0045] Furthermore, with a configuration in which the target image is divided and processed, the moisture damage prediction device 100 can divide the target image into images of appropriate size even if the target image is a large image including a wide range. Then, the moisture damage prediction device 100 can generate a trained model and identify dangerous areas R for each divided image. Therefore, the moisture damage prediction device 100 can reduce the processing load and efficiently generate a trained model and identify dangerous areas R.

[0046] Furthermore, the moisture damage prediction device 100 can adjust the size of each image for generating a trained model or identifying a hazardous area R by changing the number of divisions. Therefore, the moisture damage prediction device 100 can generate a trained model and identify a hazardous area R after adjusting the scale of the divided images to be optimal for processing according to the size of the target image or the size of the farm field included in the target image.

[0047] In addition, the geospatial information input to the moisture damage prediction device 100 preferably includes a plurality of geospatial characteristics. Examples of such geospatial characteristics include, but are not limited to, information indicating a groundwater flow system and information indicating the unevenness of the ground surface. In this specification, the information indicating the unevenness of the ground surface may be referred to as a "topographical moisture index." Information on the groundwater flow system and the topographical moisture index are related to the drainage of the field and are geospatial characteristics closely related to moisture damage. In addition, the geospatial information input to the moisture damage prediction device 100 may include any geospatial characteristics related to moisture damage.

[0048] If the geospatial information includes multiple geospatial characteristics, the model generation device 20 can generate multiple trained models for each geospatial characteristic. As shown in Fig. 3, if the geospatial information includes information on the groundwater flow system and the topographic wetness index, the model generation unit 22 performs machine learning for each of these geospatial characteristics separately and independently to generate trained models for each of them. Then, the identification unit 31 uses the trained models generated for each of these geospatial characteristics to identify a dangerous area R1 caused by groundwater and a dangerous area R2 caused by unevenness on the ground surface.

[0049] Therefore, by using information on the groundwater flow system and information on the topographical wetness index, which are considered to be closely related to the occurrence of moisture damage, the identifying unit 31 can identify the dangerous places R1 and R2 corresponding to these pieces of information separately. In the following explanation, all the dangerous places identified by the identifying unit 31, such as the dangerous places R1 and R2, may be collectively referred to as the dangerous places R.

[0050] Furthermore, the map generating unit 32 can add, to the target image, a plurality of pieces of moisture damage prediction information indicating the dangerous points R1 and R2 identified for each trained model by the identifying unit 31 in a distinguishable display mode. Therefore, the map generating unit 32 can generate a moisture damage prediction map in which the causes of moisture damage can be easily and visually identified.

[0051] (Control device operation flow) An example of the processing of the control device 10 will be specifically described with reference to Fig. 4 to Fig. 6. The control device 10 executes a trained model generation method and a map generation method according to an embodiment of the present invention.

[0052] 4, first, the data acquisition unit 21 included in the model generating device 20 acquires a data set of geospatial information and moisture damage information indicating a specific farm field via the input unit 40 (S1, data acquisition step). Here, an example is shown in which the geospatial information includes a satellite photo, which is a target image, and information on the groundwater flow system and topographical moisture index.

[0053] Next, the model generation unit 22 divides a satellite photo included in the geospatial information showing a specific field into two or more images (S2). The number of divisions may be appropriately selected as an optimal number depending on the size of the satellite photo or the size of the field included in the satellite photo. The model generation unit 22 then selects one of the divided images as a selected image (S3). Thereafter, the processes from S4 to S7 are performed on the selected image selected by the model generation unit 22 in S3.

[0054] Next, the model generation unit 22 executes a trained model generation process for the information on the groundwater flow system and the information on the topographical wetness index separately to generate trained models for each (S4, model generation step). Details of the trained model generation process will be described later. The trained models generated by the model generation unit 22 may be directly input to the map generation device 30 or may be stored in the storage unit 60.

[0055] Next, the identification unit 31 included in the map generation device 30 acquires the selected image from the model generation device 20. Then, the identification unit 31 inputs the selected image into the trained model generated by the model generation unit 22, and identifies a dangerous spot R in the selected image (S5, identification step). At this time, the identification unit 31 uses the multiple trained models generated by the model generation unit 22 to identify a dangerous spot R1 caused by groundwater and a dangerous spot R2 caused by unevenness on the ground surface.

[0056] Next, the map generating unit 32 plots the moisture damage prediction information indicating the ranges of the danger points R1 and R2 identified by the identifying unit 31 on the selected image (S6, S7). At this time, the map generating unit 32 plots the moisture damage prediction information indicating the danger points R1 and R2 on the selected image so that they are visually identifiable.

[0057] Then, the map generating unit 32 determines whether or not the processes from S4 to S7 have been performed for all images divided by the model generating unit 22 (S8). When the map generating unit 32 determines that the processes have not been performed for all images divided (no in S8), it inputs the result to the model generating unit 22. In this case, the model generating unit 22 selects an image that has not been processed as a new selected image, and the control device 10 performs the processes from S4 to S7 on the selected image.

[0058] On the other hand, when the map generating unit 32 determines that all the divided images have been processed (yes in S8), it recombines all the images that have been processed from S4 to S7 to generate a moisture damage prediction map (map generating step).Then, the map generating unit 32 causes the output unit 50 to output the generated moisture damage prediction map (S9).

[0059] In the above, the control device 10 successively performs the process of generating a trained model and the process of identifying a dangerous spot R, but the process of the control device 10 is not limited to this. For example, if a trained model has already been generated, the control device 10 may omit the trained model generation process of S4. In this case, the process of dividing a satellite photo and selecting one of the divided parts (the processes of S2 and S3) may be performed by the identification unit 31 instead of the model generation unit 22.

[0060] (Trained model generation process) The trained model generation process shown in S4 of Fig. 4 will be specifically described with reference to Fig. 5 and Fig. 6. In this embodiment, the control device 10 performs machine learning for generating the trained model and identifies the dangerous spot R by using a generative adversarial network.

[0061] 5, the generative adversarial network X1 includes a generator X2 and a discriminator X3. Here, in one embodiment of the present invention, the "learning model" refers to the generative adversarial network X1 before machine learning is performed and during machine learning, and the "trained model" refers to the generator X2 that has completed machine learning.

[0062] After the trained model is generated by the model generation unit 22, machine learning of the trained model may be further performed so that the dangerous spot R can be identified with higher accuracy. Here, the "trained model" on which the further machine learning is performed is a generative adversarial network X1 including a machine-learned generator X2 (i.e., a trained model) and a discriminator X3.

[0063] 5 and 6, in the trained model generation process, the model generation unit 22 inputs a selected image to the generator X2 (S41). The model generation unit 22 causes the generator X2 to generate a specific image by adding moisture damage prediction information indicating the risky area R to the selected image (S42, image generation step).

[0064] Next, the model generation unit 22 judges whether the data acquisition unit 21 has acquired the moisture damage information corresponding to the selected image input to the generator X2 (S43). When the model generation unit 22 judges that the data acquisition unit 21 has acquired the moisture damage information (yes in S43), it generates a correct pair, which is a data set consisting of a correct image showing the moisture damage information and the selected image, and inputs the correct pair to the discriminator X3 (S44). As a result, the discriminator X3 learns the information of the correct pair.

[0065] On the other hand, if the model generation unit 22 determines that the data acquisition unit 21 has not acquired the moisture damage information (no in S43), it skips S44. Then, the model generation unit 22 generates a generation pair, which is a data set consisting of the selected image and the specific image, and inputs the generation pair to the discriminator X3 (S45).

[0066] Next, the model generation unit 22 causes the discriminator X3 to discriminate whether the input generated pair is a pair including a specific image generated by the generator X2 or a pair including a correct image (S46, discrimination step). The model generation unit 22 inputs information on the accuracy of the discrimination result by the discriminator X3 to the generator X2 and the discriminator X3 (S47, learning step).

[0067] As a result, the generator X2 receives feedback on whether the specific image it generated was close enough to the correct image to fool the discriminator X3. The generator X2 then performs machine learning by adjusting parameters so that it can generate a specific image closer to the correct image. The discriminator X3 also receives feedback on whether the input image pair includes the specific image generated by the generator X2. The discriminator X3 then performs machine learning by adjusting parameters so that it can more reliably distinguish between the correct image and the specific image.

[0068] The model generation unit 22 performs machine learning of the generative adversarial network X1 by repeatedly performing machine learning of the generator X2 and the discriminator X3 on all selected images. In this manner, the model generation unit 22 generates a trained model, which is the trained generator X2, by the machine learning of the generative adversarial network X1.

[0069] In the above example, the discriminator X3 performs a process of determining whether or not there is corresponding moisture damage information each time a selected image is input. However, the process of the model generation device 20 is not limited to this. For example, the model generation device 20 may perform a trained model generation process on a selected image including corresponding moisture damage information before other selected images after dividing the target image acquired by the data acquisition unit 21. According to this configuration, the model generation unit 22 can perform the processes of S42, S43, and S45 to S47 on the other selected images in a state in which the discriminator X3 has accumulated sufficient information on correct pairs. Therefore, the model generation unit 22 can perform machine learning of the adversarial generative network X1 more efficiently.

[0070] The model generating device 20 may also perform the trained model generation process only on the selected image containing the corresponding moisture damage information, or on the selected image and some of the other selected images. With this configuration, the model generating device 20 can simultaneously achieve high-speed processing, reduced load, and generation of a trained model with the minimum necessary accuracy.

[0071] Furthermore, the generator X2 may generate a specific image by applying noise to the correct image. The noise may be set randomly by the generator X2, or an appropriate noise parameter may be set in advance by the user of the moisture damage prediction device 100. With this configuration, the model generation unit 22 can generate an infinite number of specific images similar to the correct image. Therefore, even if there are few selected images with corresponding moisture damage information, the model generation unit 22 can generate a trained model without reducing the accuracy of identifying the dangerous area R.

[0072] Moreover, what is input to the discriminator X3 does not have to be a data set of two images, such as a generating pair and a correct answer pair. The model generation unit 22 may input a specific image or a correct answer image directly to the discriminator X3, and the discriminator X3 may determine whether the input image is a correct answer image or not.

[0073] [Modifications] A number of modified examples are conceivable for the moisture damage prediction device 100 according to this embodiment. Such modified examples will be exemplified and described below with reference to Fig. 7 to Fig. 9. For ease of explanation, the same reference numerals are used for members having the same functions as those described in the embodiment of the present invention, and the description thereof will not be repeated.

[0074] (Moisture Damage Prediction System) 7, a moisture damage prediction system 200 is an example of a modified example of the moisture damage prediction device 100 according to the present embodiment. The moisture damage prediction system 200 differs from the moisture damage prediction device 100 in that it is configured with a server 201 and a terminal 202.

[0075] The server 201 performs substantial processing for generating a moisture damage prediction map. The server 201 includes a server control device 210, a server storage unit 260, and a server communication unit 270. The server control device 210 controls each unit of the server 201. The server control device 210 includes a model generating device 20 and a map generating device 30, similar to the control device 10.

[0076] The server storage unit 260 is a storage medium for storing information input to the server 201 and information generated by the server 201. The server storage unit 260 may also store programs for executing various functions of the server control device 210.

[0077] The server communication unit 270 is a communication device for transmitting and receiving data to and from the terminal communication unit 271 included in the terminal 202. The server communication unit 270 may transmit and receive data to and from the terminal communication unit 271 via a network such as the Internet, or may transmit and receive data directly to and from the terminal communication unit 271.

[0078] The terminal 202 transmits input data to the server 201 and outputs data received from the server 201. The terminal 202 includes an input unit 40, an output unit 50, a terminal control device 211, a terminal storage unit 261, and a terminal communication unit 271.

[0079] The terminal control device 211 comprehensively controls each unit of the terminal 202. The terminal storage unit 261 is a storage medium for storing information input to the terminal 202 and information generated by the terminal 202. The terminal communication unit 271 is a communication device for transmitting and receiving data to and from the server communication unit 270 included in the server 201.

[0080] The moisture damage prediction system 200 may include one or more servers 201. The moisture damage prediction system 200 may also include one or more terminals 202.

[0081] According to such a moisture damage prediction system 200, the high-load processes executed by the model generating device 20 and the map generating device 30 can be collectively performed by a high-performance and expensive server 201. Users of the moisture damage prediction system 200 can access the server 201 from their own terminals 202, transmit geospatial information and the like to the server 201, and receive and output the moisture damage prediction map generated by the server 201 based on the transmitted information. Therefore, users of the moisture damage prediction system 200 can use the moisture damage prediction system 200 inexpensively and easily.

[0082] 8, a moisture damage prediction system 300 is another modified example of the moisture damage prediction device 100 according to this embodiment. The moisture damage prediction system 300 is composed of a server 301 and a terminal 302. The moisture damage prediction system 300 differs from the moisture damage prediction system 200 in that the server 301 includes a model generation device 320 and the terminal 302 includes a map generation device 330.

[0083] The server 301 includes a server storage unit 260, a server communication unit 270, and a model generation device 320. In addition to the same functions as the model generation device 20 described above, the model generation device 320 also functions as a control device that collectively controls each unit of the server 301. The model generation device 320 includes a server data acquisition unit 321 and a model generation unit 322.

[0084] The server data acquisition unit 321 functions in the same manner as the data acquisition unit 21. However, the moisture damage prediction system 300 has a terminal data acquisition unit 333 in the map generating device 330 as a data acquisition unit in addition to the server data acquisition unit 321.

[0085] The model generation unit 322 generates a trained model in the same manner as the model generation unit 22. However, unlike the model generation unit 22, the model generation unit 322 transmits the generated trained model to the terminal 302 via the server communication unit 270. This is because the terminal 302 needs to receive the trained model in order for the map generation device 330 included in the terminal 302 to perform processing. The terminal 302 may directly input the received trained model to the map generation device 330, or may store it in the terminal storage unit 261.

[0086] The terminal 302 includes an input unit 40, an output unit 50, a terminal storage unit 261, a terminal communication unit 271, and a map generating device 330. The map generating device 330 has the same functions as the map generating device 30 described above, and also functions as a control device that controls each unit of the terminal 302.

[0087] The map generating device 330 also includes a terminal data acquiring unit 333. The terminal data acquiring unit 333 acquires at least a target image from the geospatial information input from the input unit 40. The identifying unit 331 performs the same processing as the identifying unit 31 described above, except for acquiring the target image from the terminal data acquiring unit 333.

[0088] The moisture damage prediction system 300 may include one or more servers 301. The moisture damage prediction system 300 may also include one or more terminals 302.

[0089] In one embodiment of the present invention, the model generating device 320 and the map generating device 330 may be configured separately, as in the moisture damage prediction system 300. With such a configuration, for example, the process of creating a trained model, which is a high load, may be performed by a high-performance and expensive server 301, and the process of identifying dangerous areas R, which is a relatively low load, may be performed by a terminal 302 used by a user of the moisture damage prediction system 300. This makes it possible to effectively reduce an increase in the load on the server 301, even in cases where the use of the moisture damage prediction system 300 is concentrated at a specific time, for example.

[0090] (Countermeasures map) The moisture damage prediction device 100 also adds moisture damage prediction information indicating dangerous points R1, R2 for each of a plurality of geospatial characteristics to the moisture damage prediction map that is generated, in a display form that allows each of the dangerous points to be identified. Here, the moisture damage prediction device 100 may include information on appropriate countermeasures in accordance with the geospatial characteristics that cause moisture damage in the moisture damage prediction information.

[0091] It is preferable to take appropriate measures against moisture damage, each of which differs depending on the geospatial characteristics that cause the damage. For example, measures such as the formation of drainage channels or the creation of ridges are preferable for moisture damage caused by groundwater, while leveling the slope is preferable for moisture damage caused by uneven ground surfaces. Conversely, leveling the slope for moisture damage caused by groundwater is almost ineffective. Thus, taking appropriate measures depending on the cause of moisture damage is very important in preventing moisture damage.

[0092] Here, Fig. 9 shows an example in which three types of danger points R1, R2, and R3 exist due to different geospatial characteristics for each field or each section of a field included in the moisture damage prediction map. In this example, it is preferable to apply different types of measures to each danger point, such as measure 1 for danger point R1, measures 1 and 2 for danger point R2, and measure 3 for danger point R3. In such a case, as shown in Fig. 9, the moisture damage prediction device 100 may add information on the types of measures corresponding to each of the danger points R1, R2, and R3 to the moisture damage prediction information.

[0093] With this configuration, a user of the moisture damage prediction device 100 who views the moisture damage prediction map can visually and easily recognize appropriate measures for each risk location.

[0094] (Machine learning algorithms) In one embodiment of the present invention, an example has been described in which the model generation unit 22 uses the generative adversarial network X1 as a machine learning algorithm for generating a trained model. However, the model generation unit 22 may generate a trained model using other machine learning algorithms. In this case, the trained model used by the specification unit 31 is a trained model corresponding to the machine learning algorithm used by the model generation unit 22.

[0095] For example, any one of the following machine learning methods or a combination of them can be used to generate a trained model by the model generation unit 22. However, these specific methods do not limit the configuration of the present invention.

[0096] Clustering ·Principal Component Analysis (PCA) Neural Network (NN) ·Bayesian Network (BN) Random Forests Support Vector Machine (SVM) · Genetic Algorithm (GP; Genetic Programming) Among these, it is preferable to use clustering or principal component analysis, which are so-called unsupervised learning algorithms that enable accurate machine learning from a small amount of training data, similar to the generative adversarial network X1. In addition, when a wealth of moisture damage information is available, the model generation unit 22 may use a machine learning algorithm other than the unsupervised learning algorithm. For example, a data set consisting of geospatial information and moisture damage information may be used as training data, and a trained model may be generated using a known machine learning algorithm such as multinomial logistic regression analysis.

[0097] In addition, when the generative adversarial network X1 is used for machine learning, a neural network may be used for the individual machine learning of the generator X2 and the discriminator X3, or other machine learning algorithms may be used. When a neural network is used, for example, a convolutional neural network (CNN) or a recurrent neural network (RNN) may be used.

[0098] [Software implementation example] The control block of the map generating device 30 (particularly the identifying unit 31 and the map generating unit 32) may be realized by a logic circuit (hardware) formed in an integrated circuit (IC chip) or the like, or may be realized by software.

[0099] In the latter case, the map generating device 30 includes a computer that executes instructions of a program (map generating program), which is software that realizes each function. The computer includes, for example, one or more processors, and a computer-readable recording medium that stores the program. The object of the present invention is achieved by the processor reading the program from the recording medium and executing it in the computer. The processor may be, for example, a CPU (Central Processing Unit). The recording medium may be a "non-transient tangible medium," such as a ROM (Read Only Memory), as well as a tape, a disk, a card, a semiconductor memory, a programmable logic circuit, or the like. The device may further include a RAM (Random Access Memory) in which the program is expanded. The program may be supplied to the computer via any transmission medium (such as a communication network or a broadcast wave) that can transmit the program. Note that one aspect of the present invention may also be realized in the form of a data signal embedded in a carrier wave, in which the program is embodied by electronic transmission. EXAMPLES

[0100] An embodiment of the present invention will be described below with reference to Figures 10 and 11. However, the present invention is not limited to this embodiment.

[0101] In Example 1, a group of rice terraces formed on sloping land was treated as a specific field, and an attempt was made to generate a moisture damage prediction map using the moisture damage prediction device 100. In Example 2, a paddy field formed on relatively flat land was treated as a specific field, and an attempt was made to generate a moisture damage prediction map using the moisture damage prediction device 100. In both examples, geospatial information including satellite photographs, information on the groundwater flow system, and information on the topographical wetness index was input to the moisture damage prediction device 100. In addition, moisture damage information indicating past moisture damage occurrence locations T in some of these fields was input to the moisture damage prediction device 100 together with the geospatial information.

[0102] As shown in Figs. 10 and 11, the moisture damage prediction device 100 generated trained models for the information on the groundwater flow system and the information on the topographical moisture index, and identified the dangerous areas R1 and R2, respectively. Note that Figs. 10 and 11 show schematic diagrams of satellite photographs, etc., but the actual processing of the moisture damage prediction device 100 was performed using the satellite photograph data itself. In addition, in the images showing the topographical moisture index in Figs. 10 and 11, the unevenness of the ground surface is represented by shades of gray. The closer the shade is to white, the more convex the ground surface is, and the closer the shade is to gray, the more concave the ground surface is. However, the display mode of the moisture damage prediction location T superimposed on the image does not show the unevenness of the ground surface.

[0103] [Results of Example 1] 10 shows the process in which the moisture damage prediction device 100 identifies dangerous locations and the moisture damage prediction map generated by the moisture damage prediction device 100 in Example 1. The accuracy of identifying the dangerous location R by the moisture damage prediction device 100 was examined from the moisture damage prediction map.

[0104] The area of ​​the "true positive" where the moisture damage prediction device 100 correctly identified the moisture damage occurrence location T as the dangerous location R was 728 m 2 The area of ​​the "false positives" where areas that were not damp damage areas T were mistakenly identified as dangerous areas R was 730 m 2 In addition, the area of ​​the "false negatives" where the moisture damage prediction device 100 erroneously identified the moisture damage occurrence location T as not being the dangerous location R was 434 m 2 The area of ​​the "true negatives" that correctly identified areas that were not damp damage areas T as dangerous areas R was 4807 m 2 It was.

[0105] From these values, the accuracy of the moisture damage prediction device 100 in identifying dangerous locations R was evaluated by calculating the indices of accuracy, precision, and recall. The accuracy rate indicates the proportion of locations that the moisture damage prediction device 100 identified as dangerous locations R or as not dangerous locations R that were actually correct, and can be calculated using the following formula (1). The precision rate indicates the proportion of locations that the moisture damage prediction device 100 identified as dangerous locations R that were actually moisture damage occurrence locations T, and can be calculated using the following formula (2). The recall rate indicates the proportion of locations that the moisture damage prediction device 100 identified as dangerous locations R, among locations that were actually moisture damage occurrence locations T, and can be calculated using the following formula (3).

[0106] Correct answer rate (%) = (true positive + true negative) / (true positive + false positive + false negative + true negative)…(1) Precision rate (%) = true positive / (true positive + false positive)…(2) Recall rate (%) = true positive / (true positive + false negative)…(3) In Example 1, the accuracy rate was 83%, the precision rate was 50%, and the recall rate was 63%.

[0107] [Results of Example 2] 11 shows the process in which the moisture damage prediction device 100 identifies dangerous locations and the moisture damage prediction map generated by the moisture damage prediction device 100 in Example 2. In Example 2, the area of ​​the “true positive” for the dangerous location R identified by the moisture damage prediction device 100 is 501 m 2 The area of ​​the "false positive" was 927 m 2 The area of ​​the "false negative" was 199 m 2 The area of ​​the "true negative" is 1904 m 2 It was.

[0108] From these values, in Example 2, the accuracy rate was 68%, the precision rate was 35%, and the recall rate was 72%.

[0109] [Consideration] The most important thing for the moisture damage prediction device 100 is to accurately identify locations where moisture damage may actually occur as dangerous locations R. From this perspective, it can be said that the reproducibility rate is a particularly important index among the above-mentioned indices. In this regard, both Example 1 and Example 2 show good reproducibility rates equal to or higher than those of conventional methods using sensors, etc., which shows that the moisture damage prediction device 100 can identify dangerous locations R with an accuracy that is practical.

[0110] The present invention is not limited to the above-described embodiments and examples, and various modifications are possible within the scope of the claims. Embodiments obtained by appropriately combining the technical means disclosed in the different embodiments and examples are also included in the technical scope of the present invention. [Explanation of symbols]

[0111] 20 Model generation device (trained model generation device) 21 Data Acquisition Section 22 Model Generation Unit 30 Map Generator 31 Specific section 32 Map Generation Unit 100 Moisture damage prediction device T Location of moisture damage (location of moisture damage in the past) R, R1, R2, R3 Danger points X1 Generative Adversarial Network X2 generator X3 discriminator

Claims

1. A moisture damage prediction device comprising an identification unit that identifies danger areas in a specific field where moisture damage is predicted to occur from a target image showing the specific field, using a trained model generated by machine learning using a dataset of geospatial information of the specific field and moisture damage information indicating locations where moisture damage has occurred in the past in the specific field.

2. The moisture damage prediction device according to claim 1 , wherein the identification unit divides the target image into two or more images and identifies the risk areas from each of the divided images.

3. the geospatial information includes a plurality of geospatial characteristics; The moisture damage prediction device according to claim 1 or 2, characterized in that the identification unit identifies the dangerous areas using the trained models generated for each of the geospatial characteristics.

4. 4. The moisture damage prediction device according to claim 3, wherein the geospatial information includes, as the geospatial characteristics, at least information indicating a groundwater flow system and information indicating unevenness of the ground surface.

5. The trained model is generated by subjecting a generative adversarial network including a generator and a discriminator to the machine learning. The trained model is The moisture damage prediction device of any one of claims 1 to 4, characterized in that the target image as the geospatial information is input to the generator, the generator is caused to generate a specific image indicating the danger area, a generated pair consisting of the target image and the specific image and a correct pair consisting of the target image and a correct image indicating the moisture damage information are input to the discriminator, and the discriminator is caused to determine whether the generated pair is a pair including the specific image generated by the generator, and the correctness of the determination is input to the generator and the discriminator to allow them to learn.

6. A moisture damage prediction method comprising the steps of: identifying danger points in a specific field where moisture damage is predicted to occur, from a target image showing the specific field, using a trained model generated by machine learning using a dataset of geospatial information of the specific field and moisture damage information showing locations where moisture damage has occurred in the past in the specific field.

7. A moisture damage prediction program for causing a computer to function as the moisture damage prediction device according to claim 1, the moisture damage prediction program causing a computer to function as the specifying unit.

Images