Crowdsourced plant protection data driven spatio-temporal dynamic prediction method for rice sheath blight

Abstract

The application discloses a crowd-sourced plant protection data-driven rice sheath blight spatiotemporal dynamic prediction method; the process is as follows: 1, constructing a model data set; 2, constructing a rice sheath blight dynamic prediction model with meteorological data and growth period data as input and a disease dynamic development curve as output; 3, calculating the disease curve area of the historical disease dynamic development curve of different regions according to the verification auxiliary data set; the average of the calculated multiple disease curve areas is recorded as the average of the disease AUDPC of the years; 4, taking the deviation acceptance rate of the disease curve area as the precision evaluation index, calibrating the rice sheath blight dynamic prediction model constructed in step 2; 5, using the rice sheath blight dynamic prediction model trained in step 4 to predict the rice sheath blight of the research area, and obtaining the predicted disease dynamic development curve.

Description

Technical Field

[0001] This invention belongs to the field of biological disaster prediction technology, specifically involving a spatiotemporal dynamic prediction method for rice sheath blight based on crowdsourced plant protection data. Background Technology

[0002] Achieving large-scale dynamic disease prediction is crucial for accurate and scientific guidance of disease control. In recent years, disease prediction models have been continuously evolving towards temporal dynamics and spatial continuity, leading to the emergence of dynamic disease mechanism prediction models such as SEIR, which can effectively describe the disease epidemic process and meet the information needs for accurate and scientific guidance of disease control regarding the temporal process and spatial distribution of large-scale disease epidemics. Disease occurrence data sources are key to dynamic disease prediction research. Currently, most disease data used in prediction models are obtained through fixed-point observations by professional departments or field trials, which suffers from drawbacks such as high workload, small data volume, and insufficient representativeness. This makes it increasingly difficult to meet the needs of practical plant protection prediction work for production management.

[0003] In recent years, with the continuous development and maturation of mobile internet big data and image deep learning technologies, some smart agriculture apps have begun to provide services such as disease identification. The identified data, recorded in the system backend, naturally forms crowdsourced plant protection data with geographic location information. This data has advantages such as ease of acquisition and large quantity, providing rich spatiotemporal data resources for crop disease prediction and representing a future trend in the field of disease prediction. However, because crowdsourced plant protection data lacks continuous temporal disease observation data, it is difficult to drive dynamic disease prediction models. Therefore, this invention proposes a calibration method for dynamic disease prediction models driven by crowdsourced plant protection data, which will make the dynamic model predictions more closely reflect actual conditions. Summary of the Invention

[0004] The purpose of this invention is to provide a spatiotemporal dynamic prediction method for rice sheath blight based on crowdsourced plant protection data.

[0005] The present invention provides a spatiotemporal dynamic prediction method for rice sheath blight based on crowdsourced plant protection data, comprising the following steps:

[0006] Step 1: Construct the model dataset. The model dataset includes a model prediction dataset and a validation auxiliary dataset. The model prediction dataset includes plant protection data on rice sheath blight from multiple sources, as well as corresponding meteorological and growth stage data. The validation auxiliary dataset includes historical disease dynamic development curves for the study area.

[0007] Step 2: Construct a dynamic prediction model for rice sheath blight, using meteorological data and growth period data as inputs and the dynamic development curve of the disease as output.

[0008] Step 3: Calculate the area under the disease curve (AUDPC) of the historical disease dynamic development curves for the study area using the validation auxiliary dataset. The average of the calculated AUDPCs is denoted as the historical AUDPC average.

[0009] Step 4: Use the rice sheath blight dynamic prediction model to predict the dynamic development curve of rice sheath blight corresponding to the mass plant protection data; and calculate the area under the curve (AUC) of the dynamic development curve. Calculate the deviation value (DV) between the obtained AUC and the historical AUDPC average obtained in Step 3. Compare this deviation value with the acceptable deviation standard (STD). If DV is less than STD, the prediction result for that sample point is considered acceptable. Calculate the proportion of samples with acceptable prediction results out of all samples; this proportion is the AUDPC deviation acceptability rate. Use the AUDPC deviation acceptability rate as the evaluation index to calibrate the rice sheath blight dynamic prediction model constructed in Step 2, completing the training of the rice sheath blight dynamic prediction model.

[0010] Step 5: Using the dynamic prediction model for rice sheath blight obtained in Step 4, predict the rice sheath blight in the region and obtain the spatiotemporal prediction curve of the disease's dynamic development.

[0011] As a preferred option, the rice sheath blight mass-source plant protection data refers to data obtained voluntarily by a large number of non-professionals and provided through the Internet, which has advantages such as low acquisition cost, timeliness, and high representativeness of the disease occurrence area.

[0012] Preferably, the meteorological data includes temperature data and precipitation data. Specifically, the temperature data refers to the daily average temperature during the rice growing season; the precipitation data refers to the daily cumulative precipitation during the rice growing season.

[0013] Preferably, the expression for the area under the disease epidemic curve (AUDPC) is as follows:

[0014]

[0015] Among them, y i t represents the disease incidence rate observed at the i-th time phase on the disease dynamic development curve. i This represents the i-th time phase. n is the total number of time phases.

[0016] Preferably, the expression for the deviation value DV is as follows:

[0017] DV = |AUDPC - AUDPC'|

[0018] Where AUDPC is the area under the disease curve of the dynamic development curve of rice sheath blight predicted by the dynamic prediction model of rice sheath blight from multiple plant protection data sources, and AUDPC' is the average area under the disease curve over the years.

[0019] Preferably, the deviation standard STD is taken as N times the average value of AUDPC of diseases in the corresponding area over the years, where N is a constant in (0,1).

[0020] Preferably, the expression for the AUDPC deviation acceptable rate is as follows:

[0021]

[0022] Where Pr represents the acceptable rate of AUDPC bias, U' is the number of samples for which the prediction results are acceptable, and U is the total number of samples.

[0023] As a preferred option, the dynamic prediction model for rice sheath blight in step two is constructed based on the SEIR model.

[0024] As a preferred option, the specific model calibration process in step four is as follows: A fitness function and encoding are constructed based on a genetic algorithm, and model parameters are optimized for the dynamic prediction model of rice sheath blight. During the model parameter optimization process, the calculated AUDPC bias acceptable rate is used as the fitness function of the genetic algorithm.

[0025] The beneficial effects of this invention are as follows:

[0026] 1. This invention is based on the statistical method of AUDPC deviation acceptability rate, which can make dynamic prediction of diseases without relying on continuous temporal plant protection observation data that is difficult to obtain in practice, and use crowdsourced plant protection data to evaluate the consistency between dynamic prediction of diseases and actual disease occurrence trends.

[0027] 2. Based on the AUDPC deviation acceptability rate, this invention calibrates a dynamic prediction model for rice sheath blight driven by mass-sourced plant disease data, providing a key method for constructing dynamic prediction models for diseases.

[0028] 3. This invention proposes a method for establishing a dynamic prediction model of crop diseases using crowdsourced plant protection data, which can greatly expand the scenarios for dynamic prediction of diseases and provide technical support for understanding the dynamic process of disease occurrence and development and carrying out effective prevention and control in more regions. Attached Figure Description

[0029] Figure 1 This is a schematic diagram illustrating the response relationship between temperature data and the temperature-related influence module characterizing rice sheath blight in the dynamic prediction model for rice sheath blight constructed in this invention.

[0030] Figure 2This is a schematic diagram illustrating the response relationship between precipitation data and the precipitation impact module characterizing rice sheath blight in the dynamic prediction model for rice sheath blight constructed in this invention.

[0031] Figure 3 This is a trend chart showing the predicted incidence of rice sheath blight in different locations in Hunan and Jiangxi provinces in 2021, based on the present invention.

[0032] Figure 4 This is a statistical graph showing the probability of the area under the disease curve and its deviation within the corresponding time phase of the dynamic development curve of rice sheath blight predicted for Hunan and Jiangxi provinces in this invention. Detailed Implementation

[0033] The present invention will be further described below with reference to the accompanying drawings.

[0034] A spatiotemporal dynamic prediction method for rice sheath blight based on crowdsourced plant protection data includes the following steps:

[0035] Step 1: Obtain data on plant diseases and their habitat factors from various sources, and construct a model dataset.

[0036] The model dataset includes a model prediction dataset and a validation auxiliary dataset.

[0037] The model prediction dataset includes mass-source plant protection data for rice sheath blight, along with corresponding meteorological (temperature, precipitation) and growth stage data. Mass-source plant protection data for rice sheath blight refers to data acquired voluntarily by a large number of non-professionals and provided via the internet, offering advantages such as low acquisition cost, timeliness, and high representativeness of disease-affected areas.

[0038] In this embodiment, the rice sheath blight crowdsourced plant protection data is provided by the domestic crowdsourced plant protection data platform, Huizhi Nongdangjia APP (Ruikun Technology). The acquisition method involves providing a mobile app to rice growers or managers; when users discover rice sheath blight, they take a photo of the diseased rice and upload the photo along with the geographic location information of the diseased rice. Therefore, the rice sheath blight crowdsourced plant protection data has high real-time performance, strong targeting (all data corresponds to diseased rice), high representativeness of the disease occurrence area, and low acquisition cost, providing rich data resources for disease prediction. 60% of the rice sheath blight crowdsourced data is used as the training set, and 40% as the validation set. In this embodiment, the rice sheath blight crowdsourced plant protection data includes rice sheath blight identification records from Hunan and Jiangxi provinces from 2019 to 2021, totaling 549 samples. The temperature and precipitation data used are from the European Centre for Medium-Range Weather Forecasts (ECMWF) ERA5 reanalysis meteorological dataset. The rice growth period data from 2019 to 2021 comes from the Key Laboratory of Digital Earth, Chinese Academy of Sciences.

[0039] The validation auxiliary dataset includes information on rice crop type, growth stage, survey time, and disease incidence rate from previous years' rice disease nurseries. The survey frequency is 5 days, starting 15 days after transplanting and ending 70 days after transplanting. Data from consecutive time phases are grouped together, resulting in 106 groups of continuous time phase disease incidence rate observation data. The validation auxiliary dataset was provided by the National Agricultural Technology Extension Center and includes multi-temporal fixed-point observation data from rice sheath blight nurseries in Hunan and Jiangxi provinces from 2010 to 2015. The survey method follows the national agricultural industry standard (NY / T613-2002) "Specifications for Rice Sheath Blight Monitoring and Survey".

[0040] Step 2: Construct a dynamic prediction model for rice sheath blight and predict the occurrence of rice sheath blight.

[0041] Dynamic prediction models can meet the information needs of precise scientific guidance for disease control regarding the temporal process and spatial distribution of large-scale disease outbreaks. They are models that simulate and predict the occurrence and development of crop diseases, possessing stronger interpretability and versatility. Therefore, this embodiment uses the SEIR model as an example to construct a dynamic prediction model for rice sheath blight, and combines it with datasets related to plant protection, meteorology, and growth stages of various diseases to predict the occurrence of rice sheath blight.

[0042] The dynamic prediction model for rice sheath blight uses four state variables, S, E, I, and R, to describe the disease occurrence and development process. The expression for the disease incidence rate Y is: Y = I + R.

[0043] The relationship between the four state variables S, E, I, and R is as follows:

[0044]

[0045] Where S, E, I, and R correspond to the proportions of disease-susceptible, exposed, infected, and recovered plants in the observation field, respectively, out of the total number of plants in the observation field; 1 / ω is the average incubation period of the disease; 1 / μ is the average infection period of the disease; t is time; and β is the disease infection rate.

[0046] The expression for the disease infection rate β is shown in equation (2):

[0047] β=ka×β0×T×P×A+kb Formula (2)

[0048] Where ka and kb are buffer coefficients; β0 is the basic infection rate; T is the temperature influence module; P is the precipitation influence module; and A is the rice growth period module.

[0049] The expression for the temperature-affected module T is as follows:

[0050]

[0051] Among them, OptimumTEM is the center of the optimal temperature range, which is set to 28℃ based on the research on the disease mechanism of rice sheath blight; δ is the variance of temperature variation; TEM is the daily average temperature.

[0052] The expression for the precipitation impact module P is as follows:

[0053]

[0054] Where OptimumPRE is the precipitation threshold, τ is the adjustment parameter, and PRE is the daily cumulative precipitation.

[0055] The expression for module A of the reproductive period is as follows:

[0056]

[0057] Here, DACE (DaysAfterCropEstablishment) represents the time elapsed after rice transplanting, where length is the predicted time period in days. On the day of rice transplanting, DACE = 0.

[0058] Step 3: Calculate the average AUDPC of diseases over the years using the validation auxiliary dataset.

[0059] Since the calculation of AUDPC value is closely related to the time phase length of the disease development curve, in order to more accurately calculate the average AUDPC value of the disease over the years in the region and to perform probability statistics on the prediction results of the dynamic prediction model for rice sheath blight based on mass plant protection data, this embodiment will calculate the average AUDPC value of the disease over the years in the corresponding region within the time phase length of the professional plant protection survey, based on the start and end points of the survey time phase of the regional professional plant protection survey data.

[0060] The expression for the area under the disease epidemic curve AUDPC in any region is shown in equation (6).

[0061]

[0062] Where y is the disease incidence rate, t is the time, n is the total number of time phases, and i is the current time phase number.

[0063] Statistics show that from 2010 to 2015, the average AUDPC value of rice sheath blight in Hunan and Jiangxi provinces was 540, occurring 15-70 days after transplanting.

[0064] Step 4: Based on the AUDPC bias acceptable rate corresponding to the mass source plant protection data of rice sheath blight, calibrate the dynamic prediction model for rice sheath blight.

[0065] Rice sheath blight crowdsourced plant protection data has advantages such as large quantity, easy accessibility, and high representativeness of disease occurrence areas, effectively reflecting the disease's occurrence trend over a large area and providing an ideal data source for regional-scale disease prediction modeling. However, because it only contains disease incidence identification records and lacks continuous temporal disease epidemic observation data, it is difficult to directly use common model evaluation methods to calibrate dynamic prediction models driven by crowdsourced plant protection data, and is currently unsuitable for training dynamic prediction models. The area under the curve (AUDPC) is a commonly used quantitative indicator for fitting disease growth models and can provide information about the dynamics of disease development. Considering the characteristics of crowdsourced plant protection data, this embodiment proposes a probabilistic statistical method based on the AUDPC deviation acceptability rate to calibrate dynamic prediction models.

[0066] This embodiment calculates the deviation between the AUDPC value predicted by the dynamic prediction model for rice sheath blight and the historical average AUDPC value of rice sheath blight in the corresponding region, using the acceptable deviation rate as an evaluation index to train the dynamic prediction model for rice sheath blight constructed in step two. A higher acceptable AUDPC deviation rate indicates better model performance.

[0067] 4.1 Calculate the AUDPC value and bias value of the predicted occurrence of rice sheath blight.

[0068] Based on the predicted occurrence of rice sheath blight in step two, the AUDPC value of the predicted occurrence of rice sheath blight within the starting and ending range of the professional plant protection survey data (15-70 days after transplanting) is calculated. The deviation value (DV) between the area under the disease curve and the average AUDPC value of the disease over the years obtained in step three is calculated. The expression for the deviation value DV is shown in equation (7):

[0069] DV = |AUDPC - AUDPC'| Equation (7)

[0070] Wherein, AUDP is the area under the disease curve of the dynamic development curve of rice sheath blight predicted by the dynamic prediction model of rice sheath blight from multiple plant protection data sources, and AUDPC' is the average area under the disease curve over the years.

[0071] 4.2 Determination of Acceptable AUDPC Deviation Criteria

[0072] Since disease dynamic prediction models are unlikely to perfectly match the actual occurrence of diseases, a certain deviation is allowed between the AUDPC of the disease dynamic development curve predicted by the model and the actual occurrence of diseases. The acceptable AUDPC deviation standard STD is taken as N times the average AUDPC of diseases in the corresponding region over the years, where N is a constant in (0,1).

[0073] Since disease dynamic prediction models are unlikely to perfectly match the actual occurrence of diseases, and according to relevant research, the error in professional field surveys and records of rice sheath blight incidence is approximately 20%. Considering the quality control standards for professional plant protection survey disease data and the lack of continuous temporal records of disease epidemic processes in mass-sourced plant protection disease data, this embodiment sets the acceptable standard deviation (STD) to 40% of the historical AUDPC average for the corresponding region, i.e., N is set to 0.4.

[0074] 4.3 AUDPC Deviation Acceptability Statistics

[0075] Compare the DV obtained in step 4.1 with the STD obtained in step 4.2. If the DV is less than the STD, the prediction result for that sample point is considered acceptable. The proportion of samples with acceptable prediction results out of all samples is the AUDPC bias acceptable rate. The expression for the AUDPC bias acceptable rate is shown in equation (8):

[0076]

[0077] Where Pr represents the acceptable rate of AUDPC bias, U' is the number of samples for which the prediction results are acceptable, and U is the total number of samples.

[0078] 4.4 Based on the AUDPC deviation acceptability rate statistics in step 4.3, the dynamic prediction model for rice sheath blight is calibrated.

[0079] Using rice growth stage, temperature, and precipitation as inputs, and the dynamic development curve of rice sheath blight as output, the AUDPC bias acceptability rate was used as the evaluation index to calibrate the dynamic prediction model for rice sheath blight in training step two. It was concluded that the higher the AUDPC bias acceptability rate, the better the performance of the dynamic prediction model for rice sheath blight.

[0080] The specific model calibration process in this embodiment is as follows: Based on the genetic algorithm, the fitness function and encoding are constructed, and the model parameters are optimized for eight parameters in the SEIR mechanism of rice sheath blight dynamic prediction model in step two: the average latent period 1 / ω of the disease, the average infection period 1 / μ of the disease, the basic infection rate β0, the buffer coefficient ka, kb, the variance δ of the normal distribution in the temperature influence module, the optimal precipitation amount optinumPRE in the precipitation influence module, and the adjustment parameter τ.

[0081] In this study, the AUDPC bias acceptance rate of the dynamic prediction results of rice sheath blight is used as the fitness function for parameter optimization in the genetic algorithm. The optimal parameter set is considered to be the parameter set obtained when the AUDPC bias acceptance rate is maximized, and the number of iterations is fixed at 100. Considering the model calibration time and efficiency, this embodiment uses a fixed number of iterations (100 iterations) to find an approximate optimal solution that satisfies the fitness function.

[0082] In this embodiment, after model calibration, the average incubation period ω of SEIR is 9.6, the average incubation period μ is 55, the basic infection rate β0 parameter is 0.69, the buffer coefficient ka is 1.47, the buffer coefficient kb is 0.15, the variance δ is 47.5, the optimal precipitation optinumPRE and the adjustment parameter τ are 12 and 15, respectively.

[0083] Step 5: Based on the dynamic prediction model of rice sheath blight obtained from the model calibration in Step 4, input the meteorological and growth period data corresponding to the study area to obtain the prediction results of the occurrence of rice sheath blight in the tested area.

[0084] The results obtained by executing this embodiment are as follows: Figure 3 , Figure 4 As shown. Figure 3 The model predicts the incidence of rice sheath blight at different locations and time periods in 2021. The disease incidence at different time periods is shown in the data. Figure 3 Rice remains largely disease-free for 40 days after transplanting (before tillering). Symptoms begin to appear around 50 days after transplanting (late tillering stage), with the disease incidence gradually increasing, rising rapidly around 70 days after transplanting (jointing to booting stage), and reaching its peak around 80 days after transplanting (around heading stage). The overall epidemic trend of rice sheath blight roughly follows an S-shaped curve. Looking at the peak incidence... Figure 3 The significant differences in peak disease incidence rates across different regions indicate variations in disease occurrence due to local weather and other factors. By combining the probabilistic statistical method of AUDPC deviation acceptability rate proposed in this embodiment with the dynamic prediction model calibration, the rice sheath blight dynamic prediction model driven by crowdsourced plant protection data can better reflect these factors, thereby achieving the expected prediction results. Figure 4The AUDPC values ​​for the predicted incidence of rice sheath blight in Hunan and Jiangxi provinces within the corresponding time periods are displayed. In the validation set data sample, the acceptable AUDPC bias rate was 85%, indicating that the prediction model reflects the epidemic trend of rice sheath blight to a certain extent, providing scientific and technological support for disease control. Compared with traditional methods of dynamic disease prediction based on professional plant protection survey data, the dynamic prediction model for rice sheath blight driven by crowdsourced plant protection data can more effectively reflect the occurrence and epidemic status of the disease in a regional area, as well as the processes of hotspot formation and decline.

Claims

1. A spatiotemporal dynamic prediction method for rice sheath blight driven by crowdsourced plant protection data, characterized in that: Includes the following steps: Step 1: Construct the model dataset; the model dataset includes a model prediction dataset and a validation auxiliary dataset; the model prediction dataset includes rice sheath blight source protection data with geolocation information, as well as corresponding meteorological data and growth period data; the validation auxiliary dataset includes historical disease dynamic development curves for different regions; the rice sheath blight source protection data is obtained by providing a mobile terminal to rice growers or managers; when a user of the mobile terminal discovers rice sheath blight, they upload geolocation information. Step 2: Construct a dynamic prediction model for rice sheath blight, using meteorological data and growth period data as inputs and the dynamic development curve of the disease as output; the dynamic prediction model for rice sheath blight is based on the SEIR model. Step 3: Calculate the area under the disease curve for the historical dynamic development curve of different regions based on the validation auxiliary dataset; the average of the calculated area under the disease curves is denoted as the average AUDPC of diseases over the years. Step 4: Use the rice sheath blight dynamic prediction model to predict the dynamic development curve of rice sheath blight corresponding to the mass plant protection data; and calculate the area under the curve of the disease; calculate the deviation value DV between the obtained area under the curve of the disease and the average AUDPC of the disease over the years obtained in Step 3; compare the deviation value with the preset deviation standard STD. If the deviation value DV is less than the preset deviation standard STD, the prediction result is considered acceptable; and count the proportion of samples with acceptable prediction results to all samples as the AUDPC deviation acceptable rate. Using the AUDPC deviation acceptable rate as an evaluation index, the rice sheath blight dynamic prediction model constructed in step two was calibrated to complete the training of the rice sheath blight dynamic prediction model; the deviation standard STD was taken as the average value of AUDPC for the corresponding region over the years. times, Let be a constant within the interval (0, 1); The model calibration process is as follows: a fitness function and encoding are constructed based on a genetic algorithm, and the model parameters are optimized for the dynamic prediction model of rice sheath blight. During the optimization process, the calculated AUDPC bias acceptable rate is used as the fitness function of the genetic algorithm. Step 5: Using the dynamic prediction model for rice sheath blight obtained in Step 4, predict the rice sheath blight in the region and obtain the spatiotemporal prediction curve of the disease's dynamic development.

2. The spatiotemporal dynamic prediction method for rice sheath blight driven by crowdsourced plant protection data according to claim 1, characterized in that: The meteorological data mentioned includes temperature data and precipitation data.

3. The spatiotemporal dynamic prediction method for rice sheath blight driven by crowdsourced plant protection data according to claim 1, characterized in that: The expression for the area under the disease curve (AUDPC) is as follows: ； wherein y i is the disease incidence observed at the i-th time phase on the dynamic development curve of the disease, t i is the i-th time phase; n is the total number of time phases.

4. The spatiotemporal dynamic prediction method for rice sheath blight driven by crowdsourced plant protection data according to claim 1, characterized in that: The expression for the deviation value DV is as follows: ； in, The area under the disease curve is the dynamic development curve of rice sheath blight based on the multi-source plant protection data predicted by the dynamic prediction model for rice sheath blight. This represents the average AUDPC value over the years.

5. The spatiotemporal dynamic prediction method for rice sheath blight driven by crowdsourced plant protection data according to claim 1, characterized in that: AUDPC Deviation Acceptability Rate The expression is as follows: ； in, For samples where the prediction results are acceptable, This represents the total number of samples.

Images