JP2024064151A - Information processing device, information processing method, and program - Google Patents

Abstract

[Problem] To provide an information processing device, information processing method, and program that enable yield prediction for a desired target plot without requiring sampling surveys on the ground. [Solution] An information processing device comprising at least an acquisition unit that acquires time-series images of farmland taken from above and time-series weather information, and a derivation unit that derives a predicted yield of crops in a target plot by inputting data representing the time-series state of vegetation obtained from the time-series images and the time-series weather information into a trained model, the trained model having been trained in advance so as to reduce a loss function that includes an error between the total value of the derived predicted yields for each target plot for a wide-area plot that includes a plurality of the target plots and the value indicated by the yield data acquired for the wide-area plot. [Selected Figure] Figure 3

Description

The present invention relates to an information processing device, an information processing method, and a program.

The Ministry of Agriculture, Forestry and Fisheries provides crop yield data for each city, town and village. The technologies described in Non-Patent Documents 1 and 2 both predict yields at the city, town and village level.

In reality, there is a need to obtain yield prediction data for smaller plots (for example, fields). In land-use crop cultivation, differences in growing environments such as weather and soil, as well as differences in cultivation management exist between fields, which leads to variations in crop growth and results in differences in final yields. Therefore, if it were possible to predict the yield for each field during growth, it would be useful for setting yield targets and reviewing cultivation management.

"Study on the applicability of remote sensing technology to agricultural land surveys - Survey method for distribution of cool weather damage to rice using Landsat MSS data", Kazuya Miyama, Hiroshi Sato, Yoshizumi Yasuda, Yasufumi Emori, Journal of the Japan Society of Irrigation, Drainage and Reclamation Engineers, Vol. 1983 (1983) No. 105 "Creating rice yield maps using Landsat TM data and MOS-1/MESSR data", Hiroyuki Shiga, Daiji Asaka, Japanese Journal of Soil and Plant Nutrition, Vol. 66 (1995) No. 6 "Classification of cropping patterns in farmland in the Osaki region of Miyagi Prefecture using multi-temporal ASTER images," Kazumasa Osawa, Daisuke Kunii, and Motoya Saito, Systems Agriculture, Vol. 26 (2010) No. 2

However, currently known methods for predicting crop yields on a field-by-field basis require on-the-ground sampling surveys to prepare training data, which is extremely labor-intensive.

The present invention was made in consideration of these circumstances, and one of its objectives is to provide an information processing device, information processing method, and program that enable yield prediction for a desired target plot without requiring sampling surveys on the ground.

An information processing device according to one aspect of the present invention includes at least an acquisition unit that acquires time-series images of farmland taken from above and time-series weather information, and a derivation unit that derives a predicted yield of crops in a target plot by inputting data representing the time-series state of vegetation obtained from the time-series images and the time-series weather information into a trained model, the trained model being trained in advance so as to reduce a loss function that includes an error between the sum of the derived predicted yields for each target plot and the value indicated by the yield data acquired for the wide-area plot, for a wide-area plot that includes a plurality of the target plots.

An information processing method according to another aspect of the present invention includes an information processing device that acquires at least time-series images of farmland taken from above and time-series weather information, and inputs data representing the time-series state of vegetation obtained from the time-series images and the time-series weather information into a trained model to derive a predicted yield of crops in a target plot, the trained model being trained in advance so as to reduce a loss function including an error between the sum of the derived predicted yields for each target plot and the value indicated by the yield data acquired for the wide-area plot, for a wide-area plot including a plurality of the target plots.

A program according to another aspect of the present invention is a program for causing a computer to acquire at least time-series images of farmland taken from above and time-series weather information, and inputting data representing the time-series state of vegetation obtained from the time-series images and the time-series weather information into a trained model to derive a predicted yield of crops in a target plot, the trained model being trained in advance so as to reduce a loss function including an error between the sum of the predicted yields derived for each target plot and the value indicated by the yield data acquired for the wide-area plot, for a wide-area plot including a plurality of the target plots.

An information processing device according to another aspect of the present invention includes an acquisition unit that acquires at least time-series images of farmland taken from above, time-series weather information, and yield data for a wide-area block including a plurality of target blocks, and a learning unit that generates a trained model that derives a predicted yield of a crop in the target block by inputting at least data representing the time-series state of vegetation obtained from the time-series images and the time-series weather information, and the learning unit learns parameters of the trained model so as to reduce a loss function that includes an error between the sum of the derived predicted yields for each of the target blocks for the wide-area block and the value indicated by the yield data acquired for the wide-area block.

An information processing method according to another aspect of the present invention is a method in which an information processing device acquires at least time-series images of farmland taken from above, time-series weather information, and yield data for a wide-area block including a plurality of target blocks, and generates a trained model that derives a predicted yield of a crop in the target block by inputting at least data representing the time-series state of vegetation obtained from the time-series images and the time-series weather information, and when generating the trained model, learns parameters of the trained model so as to reduce a loss function that includes an error between the sum of the derived predicted yields for each of the target blocks for the wide-area block and the value indicated by the yield data acquired for the wide-area block.

A program according to another aspect of the present invention is a program for causing a computer to acquire at least time-series images of farmland taken from above, time-series weather information, and yield data for a wide-area block including a plurality of target blocks, and for generating a trained model for deriving a predicted yield of a crop in the target block by inputting at least data representing the time-series state of vegetation obtained from the time-series images and the time-series weather information, and for causing the computer to learn parameters of the trained model so that, when generating the trained model, a loss function including an error between the sum of the predicted yields derived for each of the target blocks for the wide-area block and the value indicated by the yield data acquired for the wide-area block is reduced.

According to each of the above aspects, it is possible to predict yields for desired target plots without requiring sampling surveys on the ground.

FIG. 1 is a diagram illustrating an example of a usage environment of an information processing device 100. FIG. 1 is a diagram illustrating a configuration of an information processing device. 10 is a diagram for explaining the processing at the inference stage executed by the information processing device 100. FIG. FIG. 11 is a diagram (part 1) for explaining the processing of the derivation unit 130. FIG. 2 is a diagram (part 2) for explaining the processing of the derivation unit 130. FIG. 3 is a diagram (part 3) for explaining the processing of the derivation unit 130. FIG. 13 is a diagram (part 1) showing an example of information provided by an information providing unit 140. FIG. 2 is a diagram (part 2) showing an example of information provided by the information providing unit 140. 11 is a diagram for explaining the processing of a learning unit 150. FIG. FIG. 1 is a diagram (part 1) illustrating the characteristics of a penalty function g. FIG. 2 is a diagram (part 2) illustrating the characteristics of a penalty function g.

Hereinafter, an embodiment of an information processing device, an information processing method, and a program of the present invention will be described with reference to the drawings. Note that in the drawings, vectors are shown in bold, but in the description, a symbol with a "#" after it or in the middle of it indicates a vector.

[composition]
FIG. 1 is a diagram showing an example of a usage environment of the information processing device 100. For example, the information processing device 100 acquires images (satellite images) captured by the satellite 10 from one or more satellites 10 via a satellite image provider server 20 and a network NW, and acquires various information such as brush polygon data, yield data, weather information, and soil characteristics from one or more public information providing servers 30 via the network NW. The satellite images are images of farmland F on the ground repeatedly captured. A time-series satellite image is an example of "time-series images of farmland captured from above". Instead of satellite images, images of the farmland F captured by a drone or a camera attached to a steel tower may be used. The information processing device 100 provides information to the terminal device 50 via the network NW. The network NW is, for example, any network such as a LAN, a WAN, or an Internet line, and may be wired or wireless. Note that the method of acquiring satellite images is not limited to this, and satellite images may be acquired by mounting an image stored in a storage medium on a drive device of the information processing device 100. Furthermore, the satellite images may be subjected to pre-processing before being passed to the information processing device 100 .

The satellite 10, for example, orbits the Earth and periodically captures images of the ground. The satellite image provider server 20 provides the images captured by the satellite 10 to the information processing device 100, etc.

The public information providing server 30 is a server operated by public institutions such as the Ministry of Agriculture, Forestry and Fisheries or the Japan Meteorological Agency, network distributors requested by these institutions, or private businesses that secondarily provide information provided by public institutions.

The terminal device 50 is, for example, a computer device such as a personal computer, a smartphone, or a tablet terminal. The terminal device 50 communicates with the information processing device 100 via the network NW, and displays information received from the information processing device 100.

FIG. 2 is a configuration diagram of the information processing device 100. The information processing device 100 has, for example, the function of a web server. The information processing device 100 includes, for example, an acquisition unit 110, a crop classification unit 120, a derivation unit 130, an information provision unit 140, a learning unit 150, and a memory unit 180. The memory unit 180 stores information such as a trained model 182. The components other than the memory unit 180 are realized by, for example, a processor such as a CPU (Central Processing Unit) executing a program (software). Some or all of these components may be realized by hardware (including circuitry) such as an LSI (Large Scale Integration), an ASIC (Application Specific Integrated Circuit), an FPGA (Field-Programmable Gate Array), or a GPU (Graphics Processing Unit), or may be realized by collaboration between software and hardware. The program may be stored in advance in a storage device (a storage device with a non-transient storage medium) such as a hard disk drive (HDD) or flash memory, or may be stored in a removable storage medium (non-transient storage medium) such as a DVD or CD-ROM, and installed in the storage device by inserting the storage medium into a drive device.

The functions of each unit of the information processing device 100 will be described below with reference to FIG. 3. FIG. 3 is a diagram for explaining the processing of the inference stage executed by the information processing device 100. The acquisition unit 110 acquires information such as satellite images, brush polygon data, and weather information as described above using a communication interface (not shown). The acquisition unit 110 also acquires yield data for the processing of the learning stage. The acquisition unit 110 may acquire this information automatically, or may acquire the information by accepting an operation by an operator as necessary. For example, the data may be acquired by a procedure in which the operator downloads data from a website, stores the data in a specified address in the storage unit 180 of the information processing device 100, and the acquisition unit 110 reads the downloaded data from the storage unit 180.

[Inference stage]
The acquisition unit 110 performs processing to divide the satellite image into fields using the satellite image and the brush polygon data to acquire the field image. The brush polygon data is data on plots of farmland provided by the Ministry of Agriculture, Forestry and Fisheries. The satellite image divided into fields is used for subsequent processing. This field is an example of a "target plot."

Furthermore, the acquisition unit 110 generates data representing the state of vegetation over time from the satellite images in time series. The data representing the state of vegetation is data of an index indicating the characteristics of light reflection by plants, and includes data such as NDVI (Normalized Difference Vegetation Index) and reflectance. In the following description, the data representing the state of vegetation is assumed to be NDVI. The NDVI can be calculated by using the reflectance (light intensity) Ref ^Red of the wavelength band (hereinafter, band) of red light in the satellite image and the reflectance Ref ^NIR of the near-infrared band according to formula (1). The acquisition unit 110 may perform a process of excluding areas where it is estimated that no crops exist from the field image. In this case, the acquisition unit 110 excludes pixels in which the NDVI indicates an abnormal value (such as a value outside the range of plus or minus 2σ with respect to the average of the entire image).

NDVI=(Ref ^NIR -Ref ^Red )/(Ref ^NIR +Ref ^Red ) ... (1)

The crop classification unit 120 performs crop classification (classifying which crops are planted) on the NDVI image by applying, for example, the method described in Non-Patent Document 3. The crop classification may be performed based on input from an operator. For example, a trained model 182 is prepared for each crop, and the derivation unit 130 selects the trained model 182 based on the classification result of the crop classification unit 120.

The derivation unit 130 derives the predicted yield of crops in a target plot by inputting the time-series NDVI obtained from time-series satellite images and the time-series weather information into the trained model 182. The trained model 182 is trained in advance for a wide-area plot (e.g., a city, town, or village, which is assumed below) that includes multiple fields, so as to reduce the error between the total value of the derived predicted yields for each field and the value indicated by the yield data acquired for each city, town, or village. In this embodiment, the derivation unit 130 further inputs soil properties into the trained model 182.

The yield data is statistical yield data for each city, town, village, and crop provided by the Ministry of Agriculture, Forestry, and Fisheries. In the information processing device 100, it is treated as training data for machine learning. Because the yield data corresponds to a wide area, it cannot be treated as training data as is when trying to find the expected yield for each field. Therefore, the information processing device 100 is able to derive the expected yield for each field by using the following techniques. The soil characteristics are based on data provided by the National Agriculture and Food Research Organization as the Japan Soil Inventory, for example.

The parameters used in the following processing are defined and explained as follows. The NDVI for each field is a statistical value such as the average, median, or mode obtained for the NDVI obtained from the pixels included in the field. The weather information _wl , which is an element of the vector W(X, k)#, may be a scalar or a vector. In other words, the weather information may include only one event such as temperature, or may include multiple events such as temperature, solar radiation, and rainfall, or may be a result (scalar) of some calculation performed on multiple events. Similarly, the soil properties may include only one event, the content (or content rate) of a certain element, or may include multiple events, the content (or content rate) of multiple elements, or may be a result (scalar) of some calculation performed on multiple events.

R# _kX = [ _r1 , _r2 , ..., _rm ]: NDVI of the time series of the kth field in X town ⁽ or city, village; the same below)
W# _kX = [ _w1 , _w2 , ^... , _wn ]: Time-series meteorological information for _{the kth field in X town S#kX} ⁼ [ _s1 , _s2 , ^... , _s0 ]: Soil properties of the kth field in X town _ykX : Yield (expected yield per unit area) of the kth field in X town
Y _k ^X : Yield of the kth field in X town (expected yield = field yield)

Figures 4 to 6 are diagrams for explaining the processing of the derivation unit 130. In the explanations related to Figures 4 to 6, it is assumed that processing is performed focusing on a certain farm field. First, the derivation unit 130 performs processing to make the number of elements of the input data input to the trained model 182 suitable for prediction and to match the number of input nodes of the trained model 182. The parameters shown in Figure 4 have the following meanings.

r _p ^s : NDVI obtained from a satellite image taken at timing t _p (p = 1, 2, ...)
r _q : NDVI corresponding to sampling timing t _q (q = 1, 2, ..., m)
w _l : Weather information corresponding to sampling timing t _l (l = 1, 2, ..., n)

As shown in the upper diagram of FIG. 4, the derivation unit 130 performs a process of interpolating between observation points for r _p ^s (p=1 to 5) which are NDVI values obtained at each of times p=1 to 5, for example, to generate a curve showing the NDVI corresponding to an arbitrary time. For example, the derivation unit 130 performs a process of interpolating using a method such as linear interpolation, a spline function, or kernel regression. Then, the derivation unit 130 obtains r _q which is an NDVI value corresponding to a desired sampling timing q=1 to m from the curve obtained by the interpolation. In addition, the derivation unit 130 obtains w _l which is weather information corresponding to a desired sampling timing l=1 to n from the obtained weather information, as shown in the lower diagram of FIG. These are expressed in vector form to become the above-mentioned R# _k ^X and W# _k ^X.

5, the derivation unit 130 derives the yield y k X in the kth field in town X by inputting R# _k ^X , W# _k ^X , and S# _k ^X as input data into the trained model 182. The derivation unit 130 then multiplies the yield y _k ^X by the area ^of the field (which can be derived from the brush polygon data) or the planted area of the target crop in the field to derive the field yield _{Y k} ^X.

As an application example, as shown in FIG. 6, the derivation unit 130 may derive the predicted yield for the entire X town. In this case, the derivation unit 130 derives the predicted yield for as many fields as there are fields, Nx, and can derive the predicted yield for the entire X town by summing these.

The information providing unit 140 provides the terminal device 50 with information based on the processing result of the derivation unit 130. The information providing unit 140 may provide the expected yield as is, or may provide the information after processing it as follows. Figures 7 and 8 are diagrams showing an example of information provided by the information providing unit 140. As shown in Figure 7, the information providing unit 140 may display on the terminal device 50 an image showing the yield growth potential for each field, which is obtained by subtracting the expected yield for each field from the potential yield, superimposed on a map. In Figure 7, the darker the hatching, the higher the yield or the greater the growth potential for the yield. The potential yield is the maximum yield when there is no influence of, for example, water or nutrient stress, pests, etc., and is derived based on an arbitrary crop model. When using agricultural meteorological data provided by the National Agriculture and Food Research Organization, the potential yield can be derived in units of 1 km square in accordance with the units of the agricultural meteorological data.

Also, as shown in FIG. 8, the information providing unit 140 may cause the terminal device 50 to display, in a comparative manner, the trends in vegetation indices such as NDVI in a field in a certain region that is predicted to have a higher yield than surrounding fields, the trends in the average vegetation indices in the region, and the trends in the vegetation indices of the target field.

[Learning stage]
The learning unit 150 will be described below. In this embodiment, the learning unit 150 is a part of the information processing device 100, but the learning process and the inference process in machine learning may be realized by a separate device. Therefore, the learning unit 150 may be a function of a device separate from the derivation unit 130 and the like.

Figure 9 is a diagram for explaining the processing of the learning unit 150. Learning is performed based on data on multiple cities, towns, and villages. Here, Town A, City B, and Village C are shown as representative examples. The parameters in the explanation of the learning stage are defined as follows. k is the field number.

E: City/town label (E=A, B, C, ...)
N _E, : Number of farm fields in municipality E T ^E : Yield data for municipality E (teaching data)

The learning unit 150 inputs input data similar to that in the inference stage for a plurality of different municipalities (A, B, C, ...) into a machine learning model (having the same connection structure as the trained model ¹⁸² and with parameters provisionally determined), learns parameters of the machine learning model by backpropagation or the like so as to reduce (approach zero) a loss function loss including an error between each of the predicted yields Y ^A , Y ^B , Y ^C , ... obtained by summing up the field yields obtained from the derived yields for each municipality and the yield data T ^A , T B , T ^C , ... for each municipality, and sets the machine learning model that satisfies the convergence condition or completes a certain number of learning rounds as the trained model 182. The trained model 182 is, for example, a model such as a neural network.

The loss function loss is defined, for example, as in Equation (2). That is, the loss function loss is the sum of the squared errors between the predicted yield Y ^E and the yield data T ^E for each city, town, or village E, for the cities, towns, or villages, and the penalty term G.

loss = [(1/2)Σ _E ^{A,B,C ...} {(Y ^E -T ^E ) ² }] + G ... (2)

The penalty term G will be explained. The penalty term G is the sum of the values of the penalty function g for each city, town, and village. The penalty function g for each city, town, and village is a function that returns a positive penalty value if the correlation function r ^{E between the N E-dimensional vector y#* E} _{, which} ^combines the expected yield value for _each field, calculated based on the rules for the input data R# k ^E , W# _k ^E , and S# k ^E for each field, and the N _E -dimensional vector z# ^E , which combines specific vector elements of the input data R# _k ^E , W# _k ^E , and S# _k ^E for each field, is negative (positive) when it should be positive (negative), that is, if it is inversely correlated with the knowledge obtained about the relationship between the yield and specific input. The penalty function g for each city, town, and village is a function that returns a value equal to or greater than zero, for example. r ^E is expressed by formula (3), and y#* ^E and z# ^E are expressed by formulas (4) and (5), respectively.

r ^E =〈y # * ^E , z # ^E 〉 ... (3)
y # * ^E = [y ₁ ^E , y ₂ ^E , ..., y _{N E} ^E ] ... (4)
z # ^E = [z ₁ ^E , z ₂ ^E , ..., z _N ^{E E} ] ... (5)

In order to derive the vector z# ^E , the learning unit 150 first ^defines a vector f( ^R#kE _, W# _kE , S# _kE ⁾ that combines ^the elements of the input data R# _kE , W# _kE ^{, and S#kE. As described above, R#kE is m-dimensional time series data, W#kE} _{is n} ^{- dimensional} time series data, _and S ^#kE _is o-dimensional data, so f(R# _kE , W# _kE , S# _kE ) is expressed by formula (7). Next, the learning unit 150 multiplies the vector f(R# ^{kE ,} W# ^kE , S _#kE ⁾ from the left by ^a previously prepared (m+n+o)-dimensional weight vector D#, and derives the ^expected yield (scalar) _{zkE for} each field ⁽ formula ₍ 8)). The weight vector D# is expressed by the following equation (9): The learning unit 150 derives an N _E -dimensional vector z# ^E by arranging and combining the expected yield values (scalars) z _k ^E for each field obtained in this way from k = 1 to N _E.

f(R# _kE , W# _kE ^, S# _kE ) = ^[ _r1 , _r2 , ..., _rm ^, _w1 , _w2 , ..., _wn , _s1 , _s2 , ..., _s0 ] ^T ... (7)
_zkE = D# f ( ^R # _kE ^, W# _kE ^, S# _kE ) ^... (8)
D# = [ _d1 , _d2 , ..., _{dm + n + o} ] ... (9)

Here, each element of the weight vector D# is a specified value based on knowledge and is not a learning target. For example, knowledge has been obtained to the extent that it is possible to predict to some extent how much the NDVI and meteorological information at a certain stage of growth will affect the yield at the time of harvest. Therefore, by multiplying the weight vector D# based on this knowledge, it is possible to obtain an expected yield value with a certain degree of accuracy, although not with high accuracy. For example, in a case where a field where a high NDVI value is observed at a certain stage tends to have a high yield, a positive correlation should be observed between the expected yield value and the input NDVI, but a negative correlation indicates that the characteristics of the machine learning model are not appropriate. Therefore, the learning unit 150 can improve the robustness of machine learning and speed up processing by introducing a penalty term G into the loss function loss(E).

The penalty function g ( ^rE ) for each city, town, and village is a function that uses ^rE as an input value and exhibits the characteristics shown in, for example, Figs. 10 and 11. Figs. 10 and 11 are diagrams illustrating the characteristics of the penalty function g. The penalty function g may be a function that monotonically increases as the absolute value of rE increases in a region where ^rE is less than zero, as shown in Fig. 10, or may be a function that exponentially increases as the absolute value of ^rE increases in a region where rE is less than zero, as shown in Fig. 11.

The learning unit 150 calculates the penalty term G by summing up the values of the penalty function g(r ^E ) for each city, town, and village, as shown in equation (10).

G = Σ _E ^{A, B, C, ...} {g(r ^E )} ... (10)

By performing such processing, it becomes possible to predict yields for each field using a trained model generated by combining satellite images and yield data for each city, town, or village, without the need for sampling surveys on the ground. This is also useful in that it allows detailed predicted yields for each field to be derived based on relatively coarse-grained information for each city, town, or village. Furthermore, predicted yields can be derived for each city, town, or village using a common trained model. Furthermore, the information obtained as a result of processing by the information processing device 100 can be used by producers to set yield targets, and can also be used as basic information for agricultural guidance provided by agricultural extension workers.

According to each of the embodiments described above, it is possible to predict yields for desired target plots without requiring sampling surveys on the ground.

The above describes the form for carrying out the present invention using an embodiment, but the present invention is not limited to such an embodiment, and various modifications and substitutions can be made without departing from the spirit of the present invention.

REFERENCE SIGNS LIST 10 Satellite 20 Satellite image provider server 30 Public information provider server 50 Terminal device 100 Information processing device 110 Acquisition unit 120 Crop classification unit 130 Derivation unit 140 Information provision unit 150 Learning unit 180 Storage unit 182 Learned model

Claims (11)

an acquisition unit that acquires time-series images of farmland taken from above and time-series weather information;
A derivation unit that derives a predicted yield of crops in a target plot by inputting at least data representing a time-series state of vegetation obtained from the time-series images and the time-series weather information into a trained model;
Equipped with
The trained model is trained in advance so that a loss function including an error between a total value of the predicted yield for each of the derived target sections and a value indicated by yield data obtained for the wide area section is small for the wide area section.
Information processing device. The derivation unit further derives a predicted yield of a crop in the target plot by inputting soil properties of the target plot into the trained model.
2. The information processing device according to claim 1. The trained model derives an expected yield per unit area,
The deriving unit derives the predicted yield of the crop in the target plot by multiplying the predicted yield per unit area by the area of the target plot or the cultivated area of the target crop in the target plot.
2. The information processing device according to claim 1. The loss function includes a penalty term that generates a penalty if a correlation between an expected value of yield per unit area calculated based on the data representing the time series of vegetation conditions and the time series of meteorological information for each target plot on a rule basis and a value obtained by combining specific elements of the data representing the time series of vegetation conditions and the time series of meteorological information is inversely correlated with the obtained knowledge.
4. The information processing device according to claim 3. and a learning unit that learns parameters of the machine learning model to generate the trained model so that a loss function including an error between a value obtained by inputting at least data representing a time-series vegetation state obtained from the time-series images and the time-series weather information into a machine learning model for a wide-area section including a plurality of the target sections and a value indicated by yield data acquired for the wide-area section is reduced.
2. The information processing device according to claim 1. The trained model derives an expected yield per unit area,
The deriving unit derives an expected yield of the crop in the target plot by multiplying the expected yield per unit area by the area of the target plot or the cultivated area of the target crop in the target plot;
The loss function includes a penalty term that generates a penalty if a correlation between an expected value of yield per unit area calculated based on the data representing the time series of vegetation conditions and the time series of meteorological information for each target plot on a rule basis and a value obtained by combining specific elements of the data representing the time series of vegetation conditions and the time series of meteorological information is inversely correlated with the obtained knowledge.
6. The information processing device according to claim 5. An information processing device,
Obtaining time-series images of farmland taken from above and time-series weather information;
Deriving a predicted yield of crops in a target plot by inputting at least data representing a time-series state of vegetation obtained from the time-series images and the time-series weather information into a trained model;
The trained model is trained in advance so that a loss function including an error between a total value of the predicted yield for each of the derived target sections and a value indicated by yield data obtained for the wide area section is small for the wide area section.
Information processing methods. On the computer,
Obtaining time-series images of farmland taken from above and time-series weather information;
A program for deriving a predicted yield of a crop in a target plot by inputting at least data representing a time-series state of vegetation obtained from the time-series images and the time-series weather information into a trained model,
The trained model is trained in advance so that a loss function including an error between a total value of the predicted yield for each of the derived target sections and a value indicated by yield data obtained for the wide area section is small for the wide area section.
program. An acquisition unit that acquires at least a time series of images of farmland taken from above, time series weather information, and yield data regarding a wide area block including a plurality of target blocks;
A learning unit that generates a trained model that derives a predicted yield of crops in the target plot by inputting at least data representing a time-series state of vegetation obtained from the time-series images and the time-series weather information;
Equipped with
The learning unit learns parameters of the learned model so that a loss function including an error between a total value of the predicted yields for the derived target sections and a value indicated by the yield data acquired for the wide area section is reduced.
Information processing device. An information processing device,
At least a time series of images of farmland taken from above, time series meteorological information, and yield data for a wide area including a plurality of target areas are acquired;
By inputting at least data representing a time-series state of vegetation obtained from the time-series images and the time-series weather information, a trained model is generated that derives a predicted yield of crops in the target plot;
When generating the trained model, the parameters of the trained model are trained so that a loss function including an error between a total value of the predicted yields for the derived target sections and a value indicated by yield data acquired for the wide area section is reduced.
Information processing methods. On the computer,
At least a time series of images of farmland taken from above, time series meteorological information, and yield data for a wide area including a plurality of target areas are acquired;
A program for generating a trained model that derives a predicted yield of a crop in the target plot by inputting at least data representing a time-series state of vegetation obtained from the time-series images and the time-series weather information,
The computer includes:
When generating the trained model, the trained model parameters are trained so that a loss function including an error between a total value of the predicted yields for the derived target sections and a value indicated by yield data acquired for the wide area section is reduced.
program.

Images