Prediction device, learning device, prediction method, learning method, prediction program, and learning program

Abstract

To accurately predict the future state of ears of grain in relation to disease. [Solution] The prediction device (1) includes an acquisition unit (111) that acquires an image including a panicle as the subject during the panicle-turning stage, a determination unit (112) that determines the location of the panicle affected by panicle blast disease by inputting the image to a first machine learning model (TM1) that outputs information indicating the location of the panicle affected by panicle blast disease included as the subject in the input image, and a prediction unit (113) that makes a prediction about the state of the panicle after a predetermined period of time by referring to the determination result.

Description

Technical Field

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

Background Art

[0002] There is a need for a technique for grasping the state of plants against diseases. For example, Patent Document 1 discloses a diagnostic device that inputs a three-dimensional image of a damaged part of a test plant into an estimation model and diagnoses the state of the test plant.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] One of the diseases of plants is rice blast. In particular, for the panicles directly related to the yield and quality of rice, it is required to grasp the degree of damage at an early stage. In other words, there is a need for a technique to accurately predict the future state of the panicles against rice blast. However, the above-described diagnostic device does not assume predicting the future state of the panicles against rice blast.

[0005] One aspect of the present invention aims to realize a technique for accurately predicting the future state of panicles against rice blast.

Means for Solving the Problems

[0006] To solve the above problems, a prediction device according to one aspect of the present invention includes: an acquisition unit that acquires an image including a panicle in the panicle stage as its subject; a determination unit that determines which parts of the panicle included as its subject in the image are affected by rice blast disease by inputting the image to a first machine learning model that has been trained to output information indicating which parts of the panicle included as its subject in the input image are affected by rice blast disease; and a prediction unit that makes a prediction about the state of the panicle included as its subject in the image after a predetermined period of time by referring to the determination result by the determination unit.

[0007] To solve the above problems, a prediction method according to one aspect of the present invention includes: an acquisition process to acquire an image in which a panicle in the panicle stage is included as the subject; a determination process to determine which parts of the panicle included as the subject in the image are affected by rice blast disease by inputting the image into a first machine learning model that has been trained to output information indicating which parts of the panicle included as the subject in the input image are affected by rice blast disease; and a prediction process to make a prediction about the state of the panicle included as the subject in the image after a predetermined period of time by referring to the determination result in the determination process.

[0008] Each aspect of the present invention may be implemented by a computer, in which case the prediction program for the prediction device, which enables the computer to implement the prediction device by operating the computer as each part (software element) of the prediction device, and a computer-readable recording medium on which the program is recorded also fall within the scope of the present invention.

[0009] To solve the above problems, a learning device according to one aspect of the present invention includes an acquisition unit that acquires training data consisting of an image including a panicle in the panicle stage as the subject and information indicating the location of the panicle affected by rice blast disease, and a learning unit that trains a machine learning model using the training data.

[0010] To solve the above problems, a learning method according to one aspect of the present invention includes an acquisition process for acquiring training data consisting of an image including a panicle in the panicle stage as the subject and information indicating the location of the panicle affected by rice blast disease, and a learning process for training a machine learning model using the training data.

[0011] Each aspect of the present invention may be implemented by a computer, in which case the learning program for the learning device, which enables the computer to implement the learning device by operating the computer as each part (software element) of the learning device, and a computer-readable recording medium on which the program is recorded also fall within the scope of the present invention. [Effects of the Invention]

[0012] According to one aspect of the present invention, it is possible to make highly accurate predictions about the future state of rice panicles in relation to blast disease. [Brief explanation of the drawing]

[0013] [Figure 1] This is a block diagram showing the configuration of a prediction device according to Embodiment 1 of the present invention. [Figure 2] This figure shows an example of a score in Embodiment 1 of the present invention. [Figure 3] This figure shows the score of a rice panicle infected with blast disease in multiple locations in Embodiment 1 of the present invention. [Figure 4] A flowchart illustrating an example of the learning process flow executed by the prediction device according to Embodiment 1 of the present invention. [Figure 5] A flowchart illustrating another example of the learning process flow performed by the prediction device according to Embodiment 1 of the present invention. [Figure 6] This is an example of a graph that serves as training data used when training a second machine learning model according to Embodiment 1 of the present invention. [Figure 7] A flowchart illustrating an example of the flow of prediction processing performed by the prediction device according to Embodiment 1 of the present invention. [Figure 8]This is a diagram showing an example of a score pattern in Embodiment 1 of the present invention. [Figure 9] This is a graph in Example 1.

Embodiments for Carrying out the Invention

[0014] 〔Embodiment 1〕 Hereinafter, an embodiment of the present invention will be described in detail.

[0015] (Overview of Prediction Device 1) The prediction device 1 according to this embodiment is a device that predicts the future state of ears. Specifically, the prediction device 1 determines the locations where the ears included in an image with an early ear after heading as the subject are infected with rice blast. Then, the prediction device 1 refers to the determination result and predicts the state of the ears included in the image as the subject after a predetermined period.

[0016] In the present disclosure, the "ear" may be an ear of paddy rice or upland rice. Also, rice blast is a disease caused by filamentous fungi and is a disease that infects rice plants. The state where a rice ear is infected with rice blast is also referred to as "ear blast", and in ear blast, symptoms such as withered ear tissues or no formation of grains occur.

[0017] The early ear after heading refers to the ear in the tilting ear stage. Also, the ear in the tilting ear stage can be rephrased as an ear 10 to 14 days after heading, an ear one week after the end of flowering, or an ear around the cell division stop period of the embryo. Also, in the case of an ear naturally infected with rice blast, the early ear after heading is an ear around two weeks after heading. Also, when artificially inoculating potted rice grown in a greenhouse, the early ear after heading is an ear around 10 days after heading.

[0018] As an example of the state after a predetermined period of the ear, it includes where the ear is infected with blast disease after the predetermined period. As another example of the state after a predetermined period of the ear, for a plurality of ears of a certain variety, after a predetermined period, it includes how many ears among the plurality of ears are infected with blast disease and where they are infected with blast disease. Further, for a plurality of ears of a certain variety, after a predetermined period, how many ears among the plurality of ears are infected with blast disease and where they are infected with blast disease is also referred to as the "resistance strength" of the variety.

[0019] The predetermined period is not particularly limited, but in this embodiment, the maturity survey period will be taken as an example for explanation. Also, the maturity survey period can be described as the day when about two weeks have passed since the ear-tilting stage, or it can be rephrased as the yellow ripening stage of the ear.

[0020] (Configuration of Prediction Device 1) The configuration of the prediction device 1 will be described with reference to FIG. 1. FIG. 1 is a block diagram showing the configuration of the prediction device 1 according to this embodiment.

[0021] As shown in FIG. 1, the prediction device 1 includes a control unit 11, a storage unit 12, an input / output unit 13, and a communication unit 14.

[0022] (Storage Unit 12) Data referred to by the control unit 11 is stored in the storage unit 12. Examples of the data stored in the storage unit 12 include the first machine learning model TM1, the second machine learning model TM2, and the teacher data TD. That the first machine learning model TM1 and the second machine learning model TM2 are stored in the storage unit 12 indicates that the parameters defining each of the first machine learning model TM1 and the second machine learning model TM2 are stored in the storage unit 12.

[0023] (First Machine Learning Model TM1) The first machine learning model TM1 is a model that has been trained to output information indicating the locations of rice blast disease in the rice ears included as subjects in the input image. Furthermore, if the input image includes multiple rice ears, the first machine learning model TM1 is a model that has been trained to output information indicating the locations of rice blast disease in each of the multiple rice ears. The method for training the first machine learning model TM1 will be described later.

[0024] An example of "information indicating areas affected by rice blast disease" is a score corresponding to the affected area. An example of a score will be explained with reference to Figure 2. Figure 2 is a diagram showing an example of a score in this embodiment.

[0025] Each of the multiple panicles shown in Figure 2 has been treated with a cut flower dye. The rice blast fungus has the characteristic of blocking water supply and causing the panicle tissue to wither (whitening phenomenon). Therefore, when a panicle is treated with a cut flower dye, healthy panicles that are not infected with rice blast absorb the dye and become colored. On the other hand, when a panicle infected with rice blast is treated with a cut flower dye, the part of the panicle beyond the infected area cannot absorb the dye and does not become colored. Thus, by treating a panicle with a cut flower dye, it is possible to identify the parts of the panicle that are infected with rice blast.

[0026] For example, in cut flower dyeing, in the field, the dyeing process is carried out in the early morning (on sunny days, after sunrise). Also, since the dyeing will not work if the flower spikes are wet, the dyeing process is carried out in a sunny and easily drying location (such as a greenhouse with natural light) for 2-3 hours. In the case of a greenhouse, the dyeing process is carried out in the early morning (after sunrise). On sunny days, if the flower spikes are left in a greenhouse with natural light without being covered with plastic, they will be dyed in 2-3 hours. On rainy days, if they are left under light irradiation in an artificial climate chamber, they will be dyed in 4-6 hours. In both the field and greenhouse, cut flower dyeing should not be carried out until sunset.

[0027] In Figure 2, the white spikes represent spikes that have not been colored by the cut flower dye during the cut flower dye treatment. It is preferable to dye the spikes in an environment where there is no significant difference between the field environment and the temperature during the cut flower dye treatment.

[0028] As shown in Figure 2, a score is assigned to each ear of rice depending on the location of the blast disease. The scores are set as follows: Score 0: Not infected with rice blast disease (no damage) Score 1: Disease present from the glenoid to the twig (dotted line in the spike for score 1 in Figure 2) (1 glenoid) Score 2: Disease is present from the twig to the main stem (solid line in the spike with score 2 in Figure 2) (two adjacent glumes). Score 3: Disease present in the primary branch (dotted line in the spike with score 3 in Figure 2) (at least 3 adjacent glumes). Score 4: The entire primary branch (marked with an "X" in the spike with score 4 in Figure 2) is diseased, but there is no damage to the rachis. Score 5: Disease is present from the primary branch to the rachis (squares in the ears with scores 5 and 6 in Figure 2). Score 6: Diseased in the rachis between the branches (marked with an "X" in the spikes with scores 5 and 6 in Figure 2). Score 7: The ear of grain is affected (the dotted rectangle for score 7 in Figure 2). In this embodiment, when an image containing a rice panicle during the panicle-turning stage is input to the first machine learning model TM1, it outputs one of the scores from 0 to 7 depending on the location of the panicle affected by rice blast disease.

[0029] Furthermore, the first machine learning model TM1 outputs the maximum score when a single ear of rice is affected by blast disease in multiple locations (in other words, when there are multiple scores). This configuration will be explained with reference to Figure 3. Figure 3 is a diagram showing the scores of an ear of rice affected by blast disease in multiple locations in this embodiment.

[0030] In the ears of grain shown on the left side of Figure 3, scores of 1 and 2 are mixed. In this case, the first machine learning model TM1 outputs a score of 2 for these ears of grain.

[0031] Furthermore, the ear of grain shown in the center of Figure 3 has a mixture of scores of 1, 3, 4, and 5. In this case, the first machine learning model TM1 outputs a score of 5 for that ear of grain.

[0032] Furthermore, in the ears of grain shown on the right side of Figure 3, scores of 5 and 7 are mixed. In this case, the first machine learning model TM1 outputs a score of 7 for those ears of grain.

[0033] Furthermore, the first machine learning model TM1 outputs a score for each of the multiple ears of grain in an image if the image contains multiple ears of grain.

[0034] (Second machine learning model TM2) The second machine learning model, TM2, is a model that has been trained to take information indicating the locations affected by rice blast disease as input and output the resistance strength of the rice panicle.

[0035] Specifically, when data showing the proportion of panicles with each score among multiple panicles in the early stages after heading (hereinafter referred to as the "score pattern") is input to the second machine learning model TM2, the second machine learning model TM2 outputs the resistance strength of multiple panicles. The score pattern is calculated specifically by the following formula. • Number of ears with a score of 0 / Total number of ears × 100 (%) • Number of ears with score 1 / Total number of ears × 100 (%) • Number of ears with score 2 / Total number of ears × 100 (%) • Number of ears with a score of 3 / Total number of ears × 100 (%) • Number of ears with a score of 4 / Total number of ears × 100 (%) • Number of ears with a score of 5 / Total number of ears × 100 (%) • Number of ears with a score of 6 / Total number of ears × 100 (%) • Number of ears with a score of 7 / Total number of damaged ears × 100 (%) In other words, the score pattern is represented by (percentage of score 0, percentage of score 1, percentage of score 2, percentage of score 3, percentage of score 4, percentage of score 5, percentage of score 6, percentage of score 7).

[0036] Here, when the score patterns of multiple spikes of a certain variety are input into the second machine learning model TM2, the resistance strength output from the second machine learning model TM2 is the resistance strength of that particular variety. The method for training the second machine learning model TM2 will be described later.

[0037] (Training data TD) The training data TD is the data used to train the first machine learning model TM1. As shown in Figure 1, the training data TD contains multiple pairs of images PI and scores SC.

[0038] Image PI is an image that includes a panicle as the subject during the heading stage. Score SC, as mentioned above, is information indicating the areas of the panicle that are infected with rice blast disease. In the training data TD, Image PI and Score SC, which indicates the areas of the panicle included as the subject in Image PI that are infected with rice blast disease, are paired together.

[0039] The ears of grain included as subjects in the image PI are, for example, ears of grain that have been cut approximately 10 cm below the neck. Alternatively, for example, ears of grain that have been cut and placed upright in a container of water may be included as subjects in the image PI in order to maintain the shape of the ears as they were before cutting. Furthermore, the image PI may be a three-dimensional image (or multiple images from multiple directions, such as images from the side or from above).

[0040] Furthermore, among the training data TD, a pair of image PI that includes a rice panicle treated with cut flower dye as the subject, and a score SC that indicates the areas on the panicle that are infected with rice blast disease, is referred to as the first training data TD1. Furthermore, among the training data TD, a pair of image PI that includes a rice panicle that has not been treated with cut flower dye as the subject, and a score SC that indicates the areas on the panicle that are infected with rice blast disease, is referred to as the second training data TD2.

[0041] (Input / output section 13) The input / output unit 13 is an interface to input devices that accept data input and output devices that output data. Examples of input devices include, but are not limited to, microphones, cameras, eye-tracking devices, keyboards, and touchpads. Examples of output devices include, but are not limited to, speakers and liquid crystal displays.

[0042] (Communications Section 14) The communication unit 14 is an interface for sending and receiving data over a network. Examples of the communication unit 14 include, but are not limited to, communication chips in various communication standards such as Ethernet®, Wi-Fi®, and wireless communication standards for mobile data communication networks, as well as USB-compliant connectors.

[0043] (Control Unit 11) The control unit 11 controls each component of the prediction device 1. As shown in Figure 1, the control unit 11 includes an acquisition unit 111, a determination unit 112, a prediction unit 113, a learning unit 114, and an output unit 115. Specific examples of the processing of each unit will be described later.

[0044] (Acquisition part 111) The acquisition unit 111 acquires data output from the input / output unit 13 or the communication unit 14. The acquisition unit 111 stores the acquired data in the storage unit 12.

[0045] As an example, the acquisition unit 111 acquires an image that includes ears of grain in the ear-turning stage as the subject. As another example, the acquisition unit 111 acquires training data TD (first training data TD1 and second training data TD2).

[0046] (Judgment unit 112) The determination unit 112 determines which parts of the panicles are infected with rice blast disease. For example, the determination unit 112 inputs an image of panicles acquired by the acquisition unit 111 during the panicle-turning stage as its subject to the first machine learning model TM1, and determines which parts of the panicles included as subjects in the image are infected with rice blast disease. In this case, the determination unit 112 may store the determination result, including the score output from the first machine learning model TM1, in the storage unit 12.

[0047] Furthermore, if the image acquired by the acquisition unit 111 includes multiple ears of rice in the ear-turning stage as its subject, the determination unit 112 determines the location of the diseased area for each of the multiple ears. For example, the determination unit 112 inputs an image containing multiple ears of rice in the ear-turning stage as its subject into the first machine learning model TM1, and obtains a score for each of the multiple ears included as subjects in the image. The determination unit 112 then stores the determination result, including the score output from the first machine learning model TM1, into the storage unit 12.

[0048] Furthermore, the determination unit 112 refers to the score for each of the multiple ears output from the first machine learning model TM1 and generates a score pattern. In this case, the determination unit 112 stores the determination result, including the score pattern, in the storage unit 12.

[0049] (Prediction unit 113) The prediction unit 113 makes a prediction about the state of the ears of grain included as subjects in the image after a predetermined period of time. For example, the prediction unit 113 refers to the determination result by the determination unit 112 and makes a prediction about the state of the ears of grain included as subjects in the image acquired by the acquisition unit 111 after a predetermined period of time. As described above, in this embodiment, the prediction unit 113 makes a prediction about the state of the ears of grain included as subjects in the image during the observation period. The prediction unit 113 stores the prediction result in the storage unit 12.

[0050] Furthermore, the prediction unit 113 refers to the determination results for each of the multiple ears of grain made by the determination unit 112 and predicts the resistance strength of the ears of grain included as subjects in the image as a prediction of the state of the ears of grain included as subjects in the image acquired by the acquisition unit 111 after a predetermined period of time. As an example, the prediction unit 113 predicts the resistance strength by inputting the determination results from the determination unit 112 into the second machine learning model TM2.

[0051] (Learning Section 114) The learning unit 114 trains a machine learning model. For example, the learning unit 114 trains a first machine learning model TM1 using the training data TD.

[0052] As another example, the learning unit 114 trains the first machine learning model TM1 using the first training data TD1, and then further trains the first machine learning model TM1 using the second training data TD2.

[0053] As yet another example, the learning unit 114 trains a second machine learning model TM2 using training data of score patterns and resistance strength pairs.

[0054] Here, the prediction device 1 can also be described as a learning device comprising an acquisition unit 111 and a learning unit 114. That is, the prediction device 1 is a learning device comprising an acquisition unit 111 that acquires training data TD, which consists of an image including a rice panicle in the panicle stage as the subject and information indicating the location of the panicle that is infected with rice blast disease, and a learning unit 114 that trains a first machine learning model TM1 using the training data TD.

[0055] (Output section 115) The output unit 115 outputs data via the input / output unit 13 or the communication unit 14. For example, the output unit 115 outputs the prediction results predicted by the prediction unit 113 via the input / output unit 13 or the communication unit 14.

[0056] (Learning process flow 1) An example of the learning process flow (learning method) executed by the prediction device 1 will be explained with reference to Figure 4. Figure 4 is a flowchart showing an example of the learning process flow executed by the prediction device 1 according to this embodiment. Below, the process flow when training the first machine learning model TM1 will be explained. In this embodiment, the ears of grain included as subjects in the image PI may be ears grown in either a field or a greenhouse environment.

[0057] (Step S11: Acquisition process) In step S11, the acquisition unit 111 acquires training data TD, which consists of an image PI and a score SC indicating the areas of the rice grains included in the image PI that are affected by rice blast disease. The acquisition unit 111 stores the acquired training data TD in the storage unit 12.

[0058] (Step S12: Learning process) In step S12, the learning unit 114 trains the first machine learning model TM1 using the training data TD acquired by the acquisition unit 111 in step S11. In other words, the learning unit 114 trains the first machine learning model TM1 so that when an image PI is input to the first machine learning model TM1, the output score is the score SC that is paired with the image PI in the training data.

[0059] With this configuration, the prediction device 1 can take an image containing rice ears as input and train the first machine learning model TM1 to output a score indicating the parts of the rice ears that are infected with rice blast disease.

[0060] The number of ears of grain included as subjects in the image PI in the training data TD is not limited; there may be one or multiple ears. If the number of ears of grain included as subjects in the image PI in the training data TD is one, then the image PI and one score SC are included in the training data TD.

[0061] If the image PI included in the training data TD contains multiple ears of grain, the training data TD will contain the same number of images PI and the same number of scores SC as the number of ears of grain. In this case, the training data TD will also associate the multiple ears of grain included in the image PI with the score SC corresponding to each of the multiple ears of grain.

[0062] Alternatively, the learning unit 114 may train a first machine learning model TM1 using training data TD that includes image PI containing one ear of grain, and then train the first machine learning model TM1 again using training data TD that includes image PI containing multiple ears of grain.

[0063] This configuration allows the first machine learning model TM1 to be trained to output a score for each of the multiple ears of grain when an image containing multiple ears of grain is input.

[0064] (Learning process flow 2) Another example of the learning process in which the prediction device 1 trains the first machine learning model TM1 will be described with reference to Figure 5. Figure 5 is a flowchart showing another example of the flow of the learning process executed by the prediction device 1 according to this embodiment.

[0065] (Step S11A) In step S11A, the acquisition unit 111 acquires training data TD, which includes first training data TD1 consisting of an image PI containing a cut flower dye-treated ear of grain and a score SC indicating areas affected by rice blast disease, and second training data TD2 consisting of an image PI containing a cut flower dye-treated ear of grain and a score SC indicating areas affected by rice blast disease. The acquisition unit 111 stores the acquired training data TD in the storage unit 12.

[0066] (Step S12A) In step S12A, the learning unit 114 executes the following steps S121A and S122A.

[0067] (Step S121A) In step S121A, the learning unit 114 trains the first machine learning model TM1 using the first training data TD1, which consists of a pair of images PI including a cut flower dyed ear as the subject and a score SC indicating the parts of the ear affected by rice blast disease.

[0068] (Step S122A) In step S122A, the learning unit 114 further trains the first machine learning model TM1 using second training data TD2, which consists of an image PI containing a flower head that has not been treated with a cut flower dye, and a score SC indicating the parts of the flower head that are infected with rice blast disease.

[0069] For example, using an image PI containing a flower spike that has been treated with a cut flower dye and has a score of 1 as the subject, and a first training data TD1 which is a pair of the image PI and the score 1, the learning unit 114 trains a first machine learning model TM1 in step S121A.

[0070] Next, using the first training data TD1, which consists of an image PI containing a flower spike with a score of 1 that has not been treated with a cut flower dye and has a score of 1, the learning unit 114 trains the first machine learning model TM1 in step S122A.

[0071] In image PIs that include ears of grain as the subject, the ears may overlap or be obscured by other ears. When the first machine learning model TM1 is trained using such image PIs, the first machine learning model TM1 cannot output a score with high accuracy.

[0072] On the other hand, as in this example, when the first machine learning model TM1 is trained using image PI which includes a rice ear treated with cut flower dye as the subject, the colored and uncolored areas of the rice ear treated with the cut flower dye are clearly distinguishable. Therefore, when the first machine learning model TM1 is trained using the first training data TD1 which consists of image PI which includes a rice ear treated with cut flower dye as the subject and a score SC which indicates the areas of the rice ear affected by blast disease, it is possible to generate a first machine learning model TM1 that outputs a highly accurate score.

[0073] Furthermore, by training the first machine learning model TM1 using a second training data set TD2, which consists of images PI containing untreated cut flower ears and scores SC indicating areas affected by rice blast disease, it is possible to generate a first machine learning model TM1 that outputs a highly accurate score even when inputting images containing untreated cut flower ears.

[0074] In other words, through the learning process in this example, the learning unit 114 can generate a first machine learning model TM1 that outputs a highly accurate score.

[0075] Furthermore, in this example as well, the number of ears of grain included as subjects in the image PI contained in the first training data TD1 and the second training data TD2 is not limited; there may be one or multiple ears. Also, the learning unit 114 may train the first machine learning model TM1 using the first training data TD1 and the second training data TD2 which contain image PI with one ear of grain as subjects, and then train the first machine learning model TM1 using the first training data TD1 and the second training data TD2 which contain image PI with multiple ears of grain as subjects.

[0076] (Learning process flow 3) The learning process flow when prediction device 1 trains the second machine learning model TM2 will be explained again with reference to Figure 4.

[0077] (Step S11) In step S11, the acquisition unit 111 acquires training data consisting of a score pattern and resistance strength. An example of the training data acquired by the acquisition unit 111 in step S11 is shown in Figure 6. Figure 6 is an example of a graph of training data used when training the second machine learning model TM2 according to this embodiment.

[0078] In the graph shown in Figure 6, the horizontal axis represents resistance strength, and the vertical axis represents the proportion of score patterns. For example, in the graph shown in Figure 6, score patterns with strong resistance strength have a high proportion of scores of 0 and 1, while score patterns with weak resistance strength have a high proportion of scores between 5 and 7.

[0079] (Step S12) In step S12, the learning unit 114 trains the second machine learning model TM2 using training data such as the graph shown in Figure 4. In other words, the learning unit 114 trains the second machine learning model TM2 so that when a certain score pattern (a certain column) in the graph shown in Figure 4 is input to the second machine learning model TM2, the output resistance strength becomes the resistance strength of that score pattern in the graph shown in Figure 4.

[0080] With this configuration, the prediction device 1 can train a second machine learning model TM2 to output resistance strength using a score pattern as input.

[0081] (Prediction processing flow 1) An example of the prediction process flow (prediction method) executed by the prediction device 1 will be explained with reference to Figure 7. Figure 7 is a flowchart showing an example of the prediction process flow executed by the prediction device 1 according to this embodiment.

[0082] (Step S21: Acquisition process) In step S21, the acquisition unit 111 acquires an image that includes the ear of grain in the ear-turning stage as the subject. The acquisition unit 111 stores the acquired image in the storage unit 12.

[0083] The ears of grain included as subjects in the images acquired by the acquisition unit 111 are, for example, ears of grain that have been cut approximately 10 cm below the neck. Alternatively, for example, ears of grain that have been cut and placed upright in a container of water to maintain the same shape as before cutting may be included as subjects in the images. Furthermore, the images may be three-dimensional images (or multiple images from multiple directions, such as images from the side or from above).

[0084] (Step S22: Determination process) In step S22, the determination unit 112 inputs the image acquired by the acquisition unit 111 in step S21 to the first machine learning model TM1 to determine the areas in the image that are affected by rice blast disease. In other words, the determination unit 112 determines the areas affected by rice blast disease during the panicle-turning stage.

[0085] More specifically, the determination unit 112 refers to the score output from the first machine learning model TM1 and determines that the location indicated by the score is affected by rice blast disease. The determination unit 112 stores the determination result in the storage unit 12. The determination unit 112 may include the score in the determination result. Also, if the first machine learning model TM1 outputs a confidence level, the determination unit 112 may include the score and confidence level in the determination result.

[0086] Furthermore, if the image acquired by the acquisition unit 111 in step S21 includes multiple ears of grain as the subject, the unit refers to the score for each of the multiple ears of grain and generates a score pattern. In this case, the determination unit 112 stores the determination result, including the score pattern, in the storage unit 12.

[0087] (Step S23: Prediction process) In step S23, the prediction unit 113 makes a prediction about the state of the ears of grain included as subjects in the image acquired by the acquisition unit 111 in step S21 after a predetermined period. As described above, in step S23, the prediction unit 113 makes a prediction about the state of the ears of grain included as subjects in the image during the observation period. The prediction unit 113 stores the prediction result in the storage unit 12.

[0088] For example, the prediction unit 113 refers to the determination result from the determination unit 112 and predicts that the areas of the rice panicles included as subjects in the image that have been determined to be infected with rice blast disease will also be infected with rice blast disease during the observation period.

[0089] As another example, if the judgment result includes a degree of certainty, the unit predicts the locations affected by rice blast disease during the observation period, according to the degree of certainty. For example, as shown on the left side of Figure 3, if there is a mix of scores 1 and 2, a score of 2 is assigned to the panicle in question. Here, for example, if the degree of certainty for score 1 is 80% and the degree of certainty for score 2 is 10%, the prediction unit 113 predicts that the locations affected by rice blast disease in the panicle during the observation period will be "from the glumes to the twigs," which corresponds to a score of 1.

[0090] Furthermore, if the image acquired by the acquisition unit 111 in step S21 includes multiple ears of grain as subjects, the prediction unit 113 makes predictions about the state of each of the multiple ears of grain during the observation period.

[0091] (Step S24) In step S24, the prediction unit 113 determines whether or not the image acquired by the acquisition unit 111 in step S21 includes multiple ears of grain as subjects.

[0092] (Step S25) If it is determined in step S24 that multiple ears of grain are included as subjects (step S24: YES), then in step S25 the prediction unit 113 predicts the resistance strength of the multiple ears of grain.

[0093] As an example, the prediction unit 113 predicts the resistance strength of multiple ears of grain by inputting the score patterns included in the judgment result into a second machine learning model TM2. This configuration will be explained with reference to Figure 8. Figure 8 is a diagram showing an example of a score pattern in this embodiment.

[0094] Figure 8 shows score patterns SP1 to SP3. As shown in Figure 8, in score pattern SP1, the percentage of scores of 0 or 1 is approximately 63%, in score pattern SP2, the percentage of scores of 0 or 1 is approximately 68%, and in score pattern SP3, the percentage of scores of 0 or 1 is 100%.

[0095] The prediction unit 113 inputs score patterns SP1 to SP3, respectively, into the second machine learning model TM2. When the second machine learning model TM2 is trained using the training data shown in Figure 6, the higher the proportion of scores 0 or 1 in the score patterns, the stronger the output resistance strength. That is, when score patterns SP1 to SP3 are input into the second machine learning model TM2, the second machine learning model TM2 outputs resistance strengths in the order of score pattern SP3, score pattern SP2, and score pattern SP1, respectively.

[0096] Furthermore, for example, in the graph shown in Figure 4, the relationship between the proportion of score pattern 0 or score pattern 1 and the resistance strength is derived, and based on the derived relationship, the prediction unit 113 predicts the resistance strength. For example, if the relationship between the proportion of score pattern 0 or score pattern 1 and the resistance strength is linear, the prediction unit 113 calculates the resistance strength based on the proportion of score pattern 0 or score pattern 1 included in the judgment result and the said linearity. Note that when the relationship between the proportion of score pattern 0 or score pattern 1 and the resistance strength is graphed, if the relationship between the proportion of score pattern 0 or score pattern 1 and the resistance strength is linear, the straight line showing the relationship between the proportion of score pattern 0 or score pattern 1 and the resistance strength is called the "calibration curve".

[0097] The accuracy of the resistance strength of multiple ears predicted by the prediction unit 113 increases as the number of ears included in the image acquired by the acquisition unit 111 increases. To ensure the accuracy of the resistance strength of multiple ears predicted by the prediction unit 113, it is preferable that the image acquired by the acquisition unit 111 includes at least 10 ears as subjects. More preferably, it is preferable that the image acquired by the acquisition unit 111 includes at least 40 ears as subjects.

[0098] (Step S26) If it is determined in step S24 that multiple ears of grain are not included in the subject (step S24: NO), or after step S25 is performed, in step S26 the output unit 115 outputs the prediction result.

[0099] (Effect of prediction device 1) Thus, the prediction device 1 includes an acquisition unit 111 that acquires an image containing a panicle as the subject during the panicle-turning stage; a determination unit 112 that determines which parts of the panicle contained in the image are affected by rice blast disease by inputting the image acquired by the acquisition unit 111 to a first machine learning model TM1 that has been trained to output information indicating which parts of the panicle contained in the image are affected by rice blast disease; and a prediction unit that makes a prediction about the state of the panicle contained in the image acquired by the acquisition unit 111 after a predetermined period of time by referring to the determination result by the determination unit 112.

[0100] The prediction device 1 uses a first machine learning model TM1 to predict the state of the panicles at a predetermined period later (for example, during the observation period) from images of panicles taken in the early stages after heading. Therefore, it can accurately predict the future state of the panicles in relation to blast disease.

[0101] Furthermore, the prediction device 1 predicts the resistance strength of the ears of grain included in the image as a prediction regarding the state of the ears of grain included in the image after a predetermined period of time.

[0102] Traditionally, evaluating resistance strength has presented several challenges, including: (1) the need for experts proficient in panicle blast testing to conduct multiple surveys in dedicated test fields; (2) the need for repeated experiments because disease progression is easily influenced by environmental factors; (3) the requirement for dedicated test fields for panicle blast; (4) the inability to objectively evaluate the differences in the effects of resistance genes; and (5) the inability to evaluate resistance between lines with different heading dates, as it is a comparative study comparing lines with heading dates at the same time.

[0103] To address the above issues, a method for evaluating resistance from images of panicles taken during the observation period was also considered. However, since the color of panicles during the observation period is similar to that of panicles infected with rice blast, there was a challenge in evaluating resistance from images of panicles taken during the observation period.

[0104] On the other hand, the prediction device 1 can predict the resistance strength of the panicle from images of the panicle in the early stages after heading, thus solving the above-mentioned problems and predicting the resistance strength.

[0105] Furthermore, the prediction device 1 predicts resistance strength using a second machine learning model TM2 that has already been trained. Therefore, the prediction device 1 can predict resistance strength with high accuracy.

[0106] [Example 1] One embodiment of the present invention is described below.

[0107] The inventors analyzed the relationship between resistance strength in the early panicles after heading and resistance strength at the observation period for each of the seven lines with different resistance strengths, as follows. • Score patterns for panicles two weeks after heading (early post-heading) were created for each panicle. • Score patterns for the panicles were created for each panicle four weeks after heading (the observation period). The percentage of score 0 or score 1 in the early panicle score pattern after heading, and the percentage of score 5, score 6, or score 7 (resistance strength) in the panicle score pattern during the advanced observation period were graphed. The generated graph is shown in Figure 9. Figure 9 is the graph in this example.

[0108] As shown in Figure 9, it was found that the higher the proportion of scores of 0 or 1 in the score pattern of early panicles after heading, the lower the proportion of scores of 5, 6, or 7 in the score pattern of panicles at the advanced observation stage (in other words, the higher the resistance strength). Furthermore, it was found that the resistance strength of early panicles after heading and at the advanced observation stage is represented by the calibration curve shown by the dotted line in Figure 9.

[0109] [Examples of implementation using software] The function of the prediction device (learning device) 1 (hereinafter referred to as "the device") is a program that causes the device to function as a computer, and can be realized by a program (prediction program, learning program) that causes the computer to function as each control block of the device (especially each part included in the control unit 11).

[0110] In this case, the device includes a computer having at least one control device (e.g., a processor) and at least one storage device (e.g., memory) as hardware for executing the program. By executing the program using this control device and storage device, the functions described in each of the embodiments are realized.

[0111] The above program may be recorded on one or more computer-readable recording media, not temporary ones. These recording media may or may not be provided by the above device. In the latter case, the program may be supplied to the above device via any wired or wireless transmission medium.

[0112] Furthermore, some or all of the functions of each of the above control blocks can also be realized by logic circuits. For example, an integrated circuit in which logic circuits functioning as each of the above control blocks are formed is also included in the scope of the present invention. In addition, it is also possible to realize the functions of each of the above control blocks by, for example, a quantum computer.

[0113] Furthermore, each process described in the above embodiments may be performed by AI (Artificial Intelligence). In this case, the AI ​​may operate on the control device described above, or it may operate on other devices (for example, an edge computer or a cloud server).

[0114] The present invention is not limited to the embodiments described above, and various modifications are possible within the scope of the claims. Embodiments obtained by appropriately combining the technical means disclosed in different embodiments are also included in the technical scope of the present invention. [Explanation of Symbols]

[0115] 1. Prediction device 111 Acquisition Department 112 Judgment section 113 Prediction Section 114 Learning Department TM1 First Machine Learning Model TM2, the second machine learning model TD training data PI image SC Score

Claims

1. An acquisition unit that acquires images that include ears of grain as the subject during the ear-turning stage, A determination unit determines which parts of the rice ears included in the image are affected by rice blast disease by inputting the image to a first machine learning model that has been trained to output information indicating which parts of the rice ears included in the image are affected by rice blast disease. A prediction unit that, referring to the determination result by the determination unit, makes a prediction about the state of the ear of grain included as a subject in the image after a predetermined period of time, A prediction device equipped with the following features.

2. The prediction unit makes predictions regarding the state of the ear of grain included as a subject in the image during the observation period. The prediction device according to claim 1.

3. The aforementioned image includes multiple ears of grain in the ear-turning stage as subjects. The determination unit determines the location of the blast disease in each of the plurality of ears of rice, The prediction unit refers to the determination result for each of the plurality of ears of grain made by the determination unit and predicts the resistance strength of the ears of grain included as subjects in the image as a prediction regarding the state of the ears of grain included as subjects in the image after a predetermined period of time. The prediction device according to claim 2.

4. The prediction unit takes information indicating the locations affected by rice blast disease as input and predicts the resistance strength by inputting the determination result from the determination unit into a second machine learning model that has been trained to output the resistance strength. The prediction device according to claim 3.

5. The system further includes a learning unit that trains a first machine learning model using first training data consisting of a pair of images including a cut flower head treated with a dye and information indicating the location of the affected area of ​​the said head, and then trains the first machine learning model using second training data consisting of a pair of images including a cut flower head that has not been treated with a dye and information indicating the location of the affected area of ​​the said head. A prediction device according to any one of claims 1 to 4.

6. An acquisition process to obtain images that include ears of grain as the subject during the ear-turning stage, A first machine learning model, which has been trained to output information indicating the locations of rice blast disease in the rice ears included as subjects in the input image, is input to the image, and a determination process is performed to determine the locations of rice blast disease in the rice ears included as subjects in the image. A prediction process that refers to the determination result in the determination process and makes a prediction about the state of the ear of grain included as a subject in the image after a predetermined period of time, A prediction method that includes this.

7. A program for causing a computer to function as a prediction device according to claim 1, the prediction program for causing the computer to function as the acquisition unit, the determination unit, and the prediction unit.

8. An acquisition unit acquires training data consisting of an image containing a rice panicle during the panicle-turning stage and information indicating the location of the panicle affected by rice blast disease. A learning unit that trains a machine learning model using the aforementioned training data, A learning device equipped with the following features.

9. The acquisition unit is, A first training data set consisting of an image containing a rice stalk treated with a cut flower dye and information indicating the location of the rice stalk affected by blast disease, A second training data set consists of an image containing a rice stalk that has not been treated with a cut flower dye, and information indicating the location of the rice stalk that is infected with rice blast disease. Obtain training data that includes, The learning unit trains the machine learning model using the first training data, and then trains the machine learning model using the second training data. The learning device according to claim 8.

10. An acquisition process to obtain training data consisting of images containing rice panicles during the panicle-turning stage and information indicating the areas of the panicles affected by rice blast disease, Using the aforementioned training data, a learning process is performed to train a machine learning model. Learning methods that include this.

11. A program for causing a computer to function as a learning device according to claim 8, the learning program for causing the computer to function as the acquisition unit and the learning unit.

Images