Method and device for dynamic decision of field irrigation and fertilization system

Abstract

The application provides a field irrigation and fertilization system dynamic decision method and device, the method comprising: obtaining the environmental characteristics of the growth period of the crop to be decided, the characteristics including meteorological characteristics and soil characteristics; inputting the environmental characteristics into a trained reinforcement learning network model to output a water and fertilizer strategy corresponding to the maximum water and fertilizer utilization efficiency; wherein the reinforcement learning network model is trained according to a water and fertilizer strategy expansion dataset; the water and fertilizer strategy expansion dataset is obtained after performing water and fertilizer strategy characteristic optimization to determine the maximum water and fertilizer utilization efficiency based on the water and fertilizer utilization efficiency of the water and fertilizer strategy dataset by using a Bayesian optimization algorithm; and the water and fertilizer strategy dataset is a dataset obtained after planting based on different water and fertilizer strategies, including the corresponding relationship of environmental characteristics, water and fertilizer strategy characteristics and water and fertilizer utilization efficiency. The method can improve the confidence of the optimal water and fertilizer strategy, obtain the water and fertilizer strategy corresponding to the maximum water and fertilizer utilization efficiency, and improve the real water and fertilizer utilization efficiency of the water and fertilizer strategy obtained by the final decision.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of crop planting, and in particular to a method and device for dynamically determining a field irrigation and fertilization system. BACKGROUND

[0002] In field planting, water and fertilizer are key factors that determine crop yield and quality. However, in actual production, water and fertilizer management strategies are usually implemented based on the experience of local planting managers. For example, the formulation of a water and fertilizer system is usually based on long-term planting experiments over many years, combining crop growth and yield to develop a simple irrigation and fertilization system that takes into account the growth period, including irrigation frequency, irrigation volume, fertilization frequency, and fertilization volume.

[0003] However, under field planting conditions, water and fertilizer conditions are significantly affected by weather, for example, precipitation is an important source of soil moisture, and the precipitation difference between wet years and dry years can reach 50%, making it difficult for experimental results in a few years to have obvious representativeness and applicability for decades of planting processes. In addition, crop water and fertilizer requirements are also affected by planting varieties, planting density, and agronomic management measures. It can be imagined that the accuracy of irrigation and fertilization using a set of empirical parameters will be affected, and it is a huge engineering task to consider the above factors and conduct experimental research to establish a comprehensive fertilization and irrigation system.

[0004] It is difficult to cover regions, crops, and meteorological differences by conducting experiments only for specific regions, crops, and time. In existing research, the water and fertilizer system usually comes from multi-year actual planting experiments. Considering the differences in crop growth time, variety, and agronomic measures, it is necessary to conduct a large number of experiments of different types and different periods. However, it takes a long time to complete a complete planting process, which is time-consuming and small in size, and it is difficult to accumulate data to cover the weak. In addition, the microclimate changes significantly in different regions, and it is necessary to conduct experiments under different types of climate conditions to obtain typical water and fertilizer consumption characteristics. This makes existing research under limited conditions only focus on a few major crops in some regions, making it difficult to cover vast agricultural land and rich crop varieties. SUMMARY

[0005] To solve the problems in the prior art, the present application provides a method and device for dynamically determining a field irrigation and fertilization system.

[0006] The application provides a field irrigation and fertilization system dynamic decision-making method, comprising the following steps: obtaining environmental characteristics of a growth period of a crop to be decided, wherein the environmental characteristics comprise meteorological characteristics and soil characteristics; inputting the environmental characteristics into a trained reinforcement learning network model to output a water and fertilizer strategy corresponding to maximum water and fertilizer utilization efficiency; wherein the reinforcement learning network model is trained according to a water and fertilizer strategy expansion dataset; the water and fertilizer strategy expansion dataset is obtained by performing water and fertilizer strategy feature optimization to determine maximum water and fertilizer utilization efficiency based on a Bayesian optimization algorithm on water and fertilizer utilization efficiency of a water and fertilizer strategy dataset; and the water and fertilizer strategy dataset is a dataset obtained by planting under different meteorological conditions and different soil environments based on different water and fertilizer strategies, and comprising a corresponding relationship among environmental characteristics, water and fertilizer strategy features and water and fertilizer utilization efficiency.

[0007] According to the field irrigation and fertilization system dynamic decision-making method of one embodiment of the application, before the step of inputting the environmental characteristics into the trained reinforcement learning network model to output the water and fertilizer strategy corresponding to the maximum water and fertilizer utilization efficiency, the method further comprises the following steps: obtaining environmental characteristics from historical meteorological data and soil data of crops planted under different meteorological conditions and different soil conditions according to a determined water and fertilizer strategy, combining water and fertilizer strategy features and calculated water and fertilizer utilization efficiency, and constructing the water and fertilizer strategy dataset; taking water and fertilizer utilization efficiency as an optimization target, determining a proxy model according to a Bayesian algorithm, and adjusting the proxy model based on a tree structure probability density estimation algorithm to obtain a final proxy model; determining a sampling function based on an expected improvement algorithm, performing multiple iterations of sampling on the water and fertilizer strategy corresponding to each environmental characteristic in the water and fertilizer strategy dataset, and selecting a next evaluation point that increases a target function value for each iteration to update the water and fertilizer strategy and the water and fertilizer utilization efficiency, and obtaining the water and fertilizer strategy expansion dataset with the maximum water and fertilizer utilization efficiency and corresponding water and fertilizer strategy features.

[0008] According to the field irrigation and fertilization system dynamic decision-making method of one embodiment of the application, before the step of inputting the environmental characteristics into the trained reinforcement learning network model to output the water and fertilizer strategy corresponding to the maximum water and fertilizer utilization efficiency, the method further comprises the following step: training the reinforcement learning network model based on a strategy learning method DDPG according to the water and fertilizer strategy expansion dataset.

[0009] The field irrigation and fertilization system dynamic decision method according to one embodiment of the present application, the water and fertilizer strategy expansion data set is trained based on the strategy learning method DDPG to the reinforcement learning network model, including: the combination of each environment feature and water and fertilizer strategy feature in the water and fertilizer strategy expansion data set is regarded as the action space, the water and fertilizer utilization efficiency corresponding to each environment feature and water and fertilizer strategy feature combination under the condition is regarded as the state space, and the corresponding loss function is set; in the Actor part of DDPG, the action is selected based on the current state through the Eval network, and the strategy network parameter θ is updated according to the loss function; the corresponding optimal action is selected according to the sampled next state through the Target network, and the network parameter θ' of the Target is updated according to the parameter θ of the Eval network; in the Critic part of DDPG, the current Q value is calculated according to the current state and the current action through the Eval network, and the network parameter ω is updated according to the loss function; the Q value of the next state is calculated according to the next state and the next action through the Target network, and the network parameter ω' of the Target is updated according to the parameter ω of the Eval network.

[0010] The field irrigation and fertilization system dynamic decision method according to one embodiment of the present application, the water and fertilizer utilization efficiency is determined according to the crop yield, the fertilization amount and the irrigation amount.

[0011] The field irrigation and fertilization system dynamic decision method according to one embodiment of the present application, the crop yield is obtained based on the water and fertilizer efficiency model AquaCrop model according to the environment feature.

[0012] The field irrigation and fertilization system dynamic decision method according to one embodiment of the present application, the water and fertilizer utilization efficiency determination method includes the following:

[0013]

[0014] Crop yield represents the crop yield, WU total represents the total water amount in the growth period, NU total represents the total fertilization amount in the growth period, FA WU , FA NU respectively represent the irrigation and fertilization rated values in the place where the crops are planted.

[0015] The present application also provides a field irrigation and fertilization system dynamic decision device, including: an input module, used for acquiring environment features of a crop growth period to be decided, the environment features including meteorological features and soil features; a processing module, used for inputting the environment features into a trained reinforcement learning network model, and outputting a water and fertilizer strategy corresponding to the maximum water and fertilizer utilization efficiency;

[0016] The reinforcement learning network model is trained according to a water and fertilizer strategy expansion data set; the water and fertilizer strategy expansion data set is obtained by performing water and fertilizer strategy feature optimization to determine the maximum water and fertilizer utilization efficiency based on a Bayesian optimization algorithm on the water and fertilizer utilization efficiency of a water and fertilizer strategy data set; and the water and fertilizer strategy data set is a data set including the corresponding relationship of environmental features, water and fertilizer strategy features and water and fertilizer utilization efficiency obtained after planting based on different water and fertilizer strategies under different meteorological conditions and different soil environments.

[0017] The field irrigation and fertilization system dynamic decision method and device provided by the application are based on a historical sample water and fertilizer strategy data set, combine meteorological features and soil features to configure a water and fertilizer strategy, and can improve the confidence of the optimal water and fertilizer strategy. In addition, the water and fertilizer strategy data set based on history is optimized to obtain the water and fertilizer strategy corresponding to the maximum water and fertilizer utilization efficiency, so that the model is trained based on the environmental features and water and fertilizer strategy data of the maximum water and fertilizer efficiency, and the real water and fertilizer utilization efficiency of the water and fertilizer strategy obtained by the final decision is improved. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or the prior art description. Obviously, the drawings in the following description are some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor.

[0019] Figure 1 is a flowchart of the field irrigation and fertilization system dynamic decision method provided by the application;

[0020] Figure 2 is a structural schematic diagram of the field irrigation and fertilization system dynamic decision device provided by the application;

[0021] Figure 3 is a structural schematic diagram of the electronic device provided by the application. DETAILED DESCRIPTION

[0022] In order to make the purpose, technical scheme and advantages of the application more clear, the technical scheme in the application will be clearly and completely described below in combination with the drawings in the application. Obviously, the described embodiments are some embodiments of the application, not all embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor belong to the protection scope of the application.

[0023] The current water and fertilizer system lacks standardized processes and is difficult to form dynamic decisions for special areas and weather conditions. Compared with the traditional manual experience of water and fertilizer strategy, the water and fertilizer management system developed by combining meteorological, soil and other factors can improve the yield by up to 40%, while reducing more than 20% of the irrigation amount and reducing 30% of the fertilizer application amount. Therefore, combining meteorological, soil, crop and agricultural management information to develop reasonable irrigation and fertilization strategies can effectively improve the growth conditions and yield of crops, reduce water and fertilizer consumption and labor intensity, and achieve the "three high" goal of high yield, high quality and efficient use of water and fertilizer.

[0024] The following will be combined Figures 1-3 The field irrigation and fertilization system dynamic decision method and device are described. Figure 1 The field irrigation and fertilization system dynamic decision method and device are described. Figure 1 As shown in the flowchart of the field irrigation and fertilization system dynamic decision method provided by the present application, the present application provides a field irrigation and fertilization system dynamic decision method, which comprises:

[0025] 101, obtaining the environmental characteristics of the growth period of the crop to be decided, the environmental characteristics including meteorological characteristics and soil characteristics.

[0026] In order to fully consider the water and fertilizer growth mechanism combined with real-time climate change, a climate intelligent irrigation and fertilization method is formed, and the present application constructs a dynamic water and fertilizer system decision method based on massive prior knowledge driven by meteorology.

[0027] Firstly, for the crop to be decided for water and fertilizer strategy, the relevant feature data of the growth period is collected, including meteorological data and soil data, and feature extraction is performed to obtain corresponding meteorological characteristics and soil characteristics, which are part of the above-mentioned environmental characteristics.

[0028] In one embodiment, the above-mentioned environmental characteristics further include growth period characteristics, and the growth period characteristics include growth period start date and growth period days.

[0029] In one embodiment, the meteorological characteristics include air temperature (T), air humidity (H), precipitation (R), sunshine (S), atmospheric pressure (P) and wind speed (U), maximum temperature (Tx) and minimum temperature (Tn) characteristics. Soil characteristics include longitude X (decimal system), latitude Y (decimal system), elevation E (m) characteristics and initial soil water content.

[0030] Specifically, if water and fertilizer decision is made before the growth period, meteorological forecast data of the growth period can be obtained to obtain meteorological characteristics.

[0031] 102, inputting the environmental characteristics into the trained reinforcement learning network model, and outputting the water and fertilizer strategy corresponding to the maximum water and fertilizer utilization efficiency.

[0032] The reinforcement learning network model is trained according to a water and fertilizer strategy expansion data set; the water and fertilizer strategy expansion data set is obtained by performing water and fertilizer strategy feature optimization to determine the maximum water and fertilizer utilization efficiency based on a Bayesian optimization algorithm on the water and fertilizer utilization efficiency of a water and fertilizer strategy data set; and the water and fertilizer strategy data set is a data set including the corresponding relationship of environmental features, water and fertilizer strategy features and water and fertilizer utilization efficiency obtained after planting under different meteorological conditions and different soil environments based on different water and fertilizer strategies.

[0033] Before 101, the application first collects experimental data that has been planted, constructs a water and fertilizer efficiency estimation model based on a meteorological driving mechanism model using historical long-time series meteorological data as a basic database, and determines water and fertilizer utilization efficiency (WNUE). The determination method of water and fertilizer utilization efficiency includes determining according to the relationship between water and fertilizer use amount and crop yield.

[0034] For example, the evolution simulation of crop water and fertilizer response can be realized based on the water and fertilizer efficiency model AquaCrop model of the meteorological driving mechanism model. The crop yield estimation under the relevant conditions is obtained, and based on the water irrigation amount and the fertilizer amount corresponding to the crop yield and the water and fertilizer strategy, the corresponding water and fertilizer utilization efficiency WNUE can be obtained.

[0035] In the AquaCrop model, it is assumed that the yield (Y) of crops is the response result of the size of crop evapotranspiration (ET):

[0036]

[0037] In the formula, Y x and Y a are the maximum yield (kg / m 2 ) and the actual yield (kg / m 2 ) of crops, respectively. ET x and ET a are the potential evapotranspiration (mm) and the actual evapotranspiration (mm) of crops, respectively. ET x refers to the evapotranspiration under sufficient water supply conditions, and is calculated by multiplying the reference evapotranspiration ET0 calculated according to the Penman-Monteith formula by the corresponding crop coefficient. ET a refers to the evapotranspiration under actual water supply conditions, and is affected by soil water stress caused by insufficient water supply. k y is the sensitive coefficient of crop yield to soil water deficit, and changes with the growth period of crops.

[0038] The reference crop evapotranspiration ET0 is calculated according to the following formula:

[0039]

[0040] Where ET0 is the reference crop evapotranspiration, mm / d; Δ is the tangent slope of the temperature-saturation vapor pressure relationship curve at temperature T, kPa / ℃; R n is the net radiation, MJ / (m2·d); G is the soil heat flux, MJ / (m2·d); T is the average temperature, °C; r is the psychrometric constant; U2 is the wind speed at 2 m height, m / s; e s is the average saturated water vapor pressure, kPa; e a is the actual water vapor pressure, kPa.

[0041] The AquaCrop model uses a daily simulation step to simulate the evolution of crop photosynthesis, respiration, transpiration, and other processes by inputting data on climate, crops, soil, and field management such as irrigation, fertilization, and weeding. It outputs canopy cover (CC), biomass (Bx), and yield (Crop yield ) and other results. Crop yield is obtained by multiplying the biomass (Bx) of the crop at maturity by the harvest index (HI). Crop biomass is obtained by multiplying the normalized water productivity (WP*), which is the accumulated biomass per unit water consumption, by the ratio of crop transpiration to reference evapotranspiration:

[0042]

[0043] Where B i is the daily aboveground biomass (t / hm2); T ri is daily transpiration (mm), ET o,i WP is the daily reference evapotranspiration (mm). * To normalize water productivity (gram / m2), crop water productivity is divided by crop evapotranspiration under standard evapotranspiration conditions, and parameters are adjusted according to actual atmospheric CO2 concentration, crop type, and crop growth period and yield maturity period.

[0044] After the reference evapotranspiration was calculated using the Penman-Monteith formula, the crop transpiration (T) during the crop growth process was calculated based on the crop transpiration coefficient when the canopy coverage was maximum and the canopy coverage (CC) during the actual growth process. r ) is separated from evapotranspiration to eliminate non-productive consumptive water use interference from soil evaporation. In addition, as the canopy expands, CC needs to be revised to reduce the effects of shading and canopy air advection on crop transpiration. The formula is:

[0045] T r =CC * ×Kcbx ×ET0

[0046] Kcb x =K aer ×KS sto

[0047] In the formula, CC * is the canopy coverage (%) considering the canopy shading and air flow influence; Kcb x is the transpiration coefficient of the reference crop under sufficient water supply condition. The actual crop transpiration is affected by the soil water stress, which is expressed as the stomatal conductance stress under water shortage condition, KS sto represents the stress coefficient and the soil aeration stress under water abundance condition, K aer represents the stress coefficient. Both of them are dimensionless parameters with the value range of [0, 1], and the value of 1 means no stress.

[0048] The canopy coverage changes with the growth of the crop, which directly affects the photosynthesis of the crop. The exponential form is used to express the increase and decline of CC, including the exponential growth period, the stable growth period and the decline period, and the expressions are respectively:

[0049]

[0050]

[0051] In the formula, CC is the canopy coverage (%); t is the cumulative time from emergence; CC0 is the initial canopy coverage (%), which is generally taken as the average seedling coverage at 90% emergence; CC x is the value of the canopy coverage when it reaches the maximum (%); CGC is the canopy growth coefficient, which represents the increase of the canopy coverage per unit length of day (%); CDC is the canopy decline coefficient, which represents the decrease of the canopy coverage per unit length of day (%). The correction formula of CGC and CDC under soil water stress is as follows:

[0052] CGC adj =Ks exp ×CGC

[0053]

[0054] In the formula, CGC adj and CDC adj represent the CGC and CDC affected by water stress, respectively, Ks exp is the stress coefficient of water on the increase of canopy coverage; is the stress coefficient of water on the decrease of canopy coverage.

[0055] In the AquaCrop model, the root growth rate calculation formula when calculating crop yield is as follows:

[0056]

[0057] In the formula, Z is the effective root depth at time t (m); Z ini is the initial root depth (m); Z x is the maximum effective root depth (m); t0 is the time of 90% emergence; t x is the time when the root reaches the maximum effective root depth; t is the root growth time.

[0058] Considering the constructed water and fertilizer strategy data set, the water and fertilizer strategy is not the optimal water and fertilizer strategy, in the application, based on the Bayesian optimization algorithm, the water and fertilizer utilization efficiency of the water and fertilizer strategy data set is optimized to obtain the water and fertilizer strategy corresponding to the maximum water and fertilizer utilization efficiency, and the water and fertilizer strategy is updated to obtain a water and fertilizer strategy expansion data set. Based on the generated water and fertilizer strategy expansion data set as a total experience pool, a reinforcement learning algorithm is used to learn the water and fertilizer strategy, and a trained reinforcement learning network model is obtained. The trained reinforcement learning network model can configure the optimal water and fertilizer strategy for the environmental characteristics of the future growth period, so that the water and fertilizer utilization efficiency is maximized. The water and fertilizer strategy includes irrigation amount, fertilizer amount, irrigation days, fertilizer days, etc.

[0059] The field irrigation and fertilization system dynamic decision-making method of the application is based on the water and fertilizer strategy data set of historical samples, and the water and fertilizer strategy is configured in combination with meteorological characteristics and soil characteristics, so that the confidence of the optimal water and fertilizer strategy can be improved. In addition, optimization is performed based on the historical water and fertilizer strategy data set to obtain the water and fertilizer strategy corresponding to the maximum water and fertilizer utilization efficiency, so that the model is trained based on the environmental characteristics and water and fertilizer strategy data of the maximum water and fertilizer efficiency, the real water and fertilizer utilization efficiency of the water and fertilizer strategy obtained by the final decision is improved, and the powerful learning and decision-making ability of reinforcement learning is combined to realize high-confidence water and fertilizer strategy decision-making.

[0060] In one embodiment, the water and fertilizer utilization efficiency is determined according to crop yield, fertilizer amount and irrigation amount. In one embodiment, the crop yield is obtained based on the water and fertilizer efficiency model AquaCrop model according to environmental characteristics. The above embodiments have been exemplified and will not be repeated here.

[0061] In one embodiment, the water and fertilizer utilization efficiency determination method comprises the following steps:

[0062]

[0063] Crop yield represents crop yield, WU totalNU represents the total water consumption during the growth period. total FA represents the total fertilizer application amount during the growth period. WU FA represents the total fertilizer application amount during the growth period. NU respectively represent the local irrigation and fertilization rating values.

[0064] It can be seen that the embodiment of the present application combines the local irrigation and fertilization rating values, i.e. the local recommended irrigation and fertilization rating, to realize accurate quantification of water and fertilizer utilization efficiency and avoid excessive irrigation and fertilization.

[0065] In one embodiment, before the inputting the environmental characteristics into the trained reinforcement learning network model and outputting the water and fertilizer strategy corresponding to the maximum water and fertilizer utilization efficiency, the method further comprises: obtaining historical meteorological data and soil data of crops planted under different meteorological conditions and different soil conditions according to the determined water and fertilizer strategy, combining the water and fertilizer strategy characteristics and the calculated water and fertilizer utilization efficiency to construct the water and fertilizer strategy data set; determining the proxy model according to the Bayesian algorithm with the water and fertilizer utilization efficiency as the optimization objective, and adjusting based on the tree structure probability density estimation algorithm to obtain the final proxy model; determining the sampling function based on the expected improvement algorithm, and performing multiple iterations on the water and fertilizer strategy corresponding to each environmental characteristic in the water and fertilizer strategy data set, and selecting the next evaluation point that increases the value of the objective function in each iteration to update the water and fertilizer strategy and the water and fertilizer utilization efficiency, to obtain the water and fertilizer strategy expansion data set with the maximum water and fertilizer utilization efficiency and the corresponding water and fertilizer strategy characteristics.

[0066] The water and fertilizer efficiency model based on the meteorological driving mechanism model in the embodiment of the present application simulates the yield and water and fertilizer utilization efficiency improvement results under different water and fertilizer system configurations under real meteorological conditions, uses the Bayesian optimization algorithm as the core algorithm of random optimization to explore the optimal configuration of the water and fertilizer system, i.e. to find the maximum water and fertilizer utilization efficiency by dynamically configuring the water and fertilizer strategy, to improve the final WNUE, and finally to integrate all input and output combinations of the model to construct the water and fertilizer strategy expansion data set.

[0067] In specific embodiments, the model input parameters can include: initial soil water content, growth period start date, growth period duration, irrigation frequency, irrigation date, irrigation amount, fertilization frequency, fertilization date, fertilization amount, and daily meteorological data corresponding to the growth period stage. Appropriate value ranges are set for each parameter, and then the input strategy is dynamically generated by the intelligent optimization algorithm of the embodiment to realize the maximum optimization of WNUE.

[0068] Specifically, the Bayesian optimization algorithm is used as the basic global parameter optimization algorithm. The optimization process uses the Bayesian theorem to fit the target function using a probability proxy model, and selects the next evaluation point according to the previous sampling results to quickly reach the optimal solution, and the expression is:

[0069]

[0070] H i = {(x1, f(x1)),..., (x i , f(x i ))}

[0071] where p(f) and p(H i | f) are the prior and likelihood of f, respectively, and p(H i | f) represents the conditional probability distribution of the parameters f given the observation data set H i , i.e., the posterior distribution.

[0072] Further, the present application adopts a tree-structured Parzen estimator (TPE) algorithm as the probabilistic surrogate model, where p(H i | f) is defined as:

[0073]

[0074] where y * = min{(x1, f(x1)),..., (x i , f(x i ))} is the optimal value on the observation domain, and l(x) is the density estimate of the observation value x whose loss function is less than y * , and g(x) is the density estimate of the observation value x whose loss function is greater than or equal to y * .

[0075] The present application adopts expected improvement (EI) as the sampling function to select the next evaluation point that has an optimization effect on the target function value, and the expression is:

[0076]

[0077] When p(y | x) is integrated as positive at y < y * , setting the hyperparameter x to model the algorithm will produce better results than the optimal value on the observation domain.

[0078] Let γ = p(y < y*), and construct:

[0079]

[0080] Substituting the sampling function of the formula can obtain:

[0081]

[0082] When the hyperparameter x has the maximum probability l(x) and the minimum probability g(x), the maximum EI value is obtained. The sample hyperparameter set is constructed by l(x) and g(x), and x is evaluated in the form of l(x) / g(x), and in each iteration process, the algorithm returns the point x with the maximum EI value * .

[0083] Through the above algorithm, the dynamic generation and simulation of water and fertilizer strategy are realized, and the feedback result is updated constantly, and a large amount of 'weather-soil-water and fertilizer strategy + WNUE' data combination pairs are generated in the process. All data combinations can be integrated in the form of json to form a water and fertilizer system expansion data set.

[0084] The field irrigation and fertilization system dynamic decision-making method of the embodiment of the application optimizes based on historical water and fertilizer strategy data sets to obtain a water and fertilizer strategy corresponding to the maximum water and fertilizer utilization efficiency, so that the model is trained based on the environmental characteristics and water and fertilizer strategy data of the maximum water and fertilizer efficiency, and the real water and fertilizer utilization efficiency of the water and fertilizer strategy obtained by the final decision is improved.

[0085] In one embodiment, before the environmental characteristics are input into the trained reinforcement learning network model and the water and fertilizer strategy corresponding to the maximum water and fertilizer utilization efficiency is output, the method further includes: training the reinforcement learning network model based on the strategy learning method DDPG according to the water and fertilizer system expansion data set.

[0086] The water and fertilizer system expansion data set is used as a total experience replay pool to learn the water and fertilizer system strategy by using a reinforcement learning algorithm, and the optimal water and fertilizer system strategy is configured for future meteorological data of the growth period.

[0087] The reinforcement learning model adopts a deep learning neural network to fuse the strategy learning method (Deep Deterministic Policy Gradient, DDPG) of DPG as a core learning decision algorithm. The algorithm adopts the 'actor-critic' algorithm framework of different strategies, and a double network structure based on a double deep Q-learning network (DDQN) is used to solve the slow convergence problem. The agent interacts with the environment to obtain the state, obtains the action strategy through the neural network, and executes the strategy. After the agent executes the strategy, the feedback of the environment is obtained, and the decision ability of the reinforcement learning is used to evaluate each strategy, and then the neural network is updated.

[0088] In one embodiment, the training of the reinforcement learning network model based on the strategy learning method DDPG according to the water and fertilizer strategy expansion data set includes:

[0089] Each combination of environmental features and water and fertilizer strategy features in the extended water and fertilizer strategy dataset is taken as an action space, and the corresponding water and fertilizer utilization efficiency under each combination of environmental features and water and fertilizer strategy features is taken as a state space, and a corresponding loss function is set;

[0090] In the Actor part of DDPG, the action is selected based on the current state through the Eval network, and the policy network parameter θ is updated according to the loss function; the optimal action corresponding to the sampled next state is selected through the Target network, and the network parameter θ' of the Target is updated according to the parameter θ of the Eval network;

[0091] In the Critic part of DDPG, the current Q value is calculated according to the current state and the current action through the Eval network, and the network parameter ω is updated according to the loss function; the Q value of the next state is calculated according to the next state and the next action through the Target network, and the network parameter ω' of the Target is updated according to the parameter ω of the Eval network.

[0092] In the Actor-Critic framework, both the Actor and the Critic contain two neural networks: the Eval network and the Target network. In the Actor algorithm framework, the Eval network selects the action a according to the current state s t Select action a t , and is responsible for updating the policy network parameter θ; the Target network selects the next optimal action a t+1 according to the next state s t+1 sampled from the experience pool, and the network parameter θ' is updated from θ. The evaluation mean of the state and the behavior is used as the loss function:

[0093] L oss = -m ean (Q(s t , a t , θ))

[0094] In the Critic, the Eval network calculates the current Q value Q(st, at) according to the state s t and the action a t , and is responsible for updating the value network parameter ω; the Target network calculates the actual Q value of the next state according to s t+1 and a t+1 , and the network parameter ω' is updated from ω, and the expression is:

[0095] Q Target = r t+1 + γ·Q'(s t+1 , a t+1 , ω')

[0096] Where: γ is the discount factor, which indicates the degree of influence of time distance on the reward. The smaller the factor, the more emphasis is placed on the current reward.

[0097] Furthermore, the mean square error is used to construct a loss function that can be optimized by the network:

[0098] L oss =E[(Q Target -Q(s t , a t ,ω)) 2 ]

[0099] This method introduces random noise during training, transforming decision-making from a deterministic process to a stochastic one. The ε-greedy strategy is simple to implement, but suffers from low exploration efficiency. Because it uses random selection, there is no memory, which can lead to repeated exploration. The Ornstein-Uhlenbeck (OU) process is introduced in DDPG. The OU process is a time-dependent process that is more efficient in inertial systems. Its continuous form is as follows:

[0100] dx t =θ(μ-x t )dt+σdW t

[0101] Its discrete form is:

[0102] x t -x t-1 =θ(μ-x t )+σW t

[0103] Where: μ is the mean, θ is the rate of change, σ is the range of noise, and x t is the state, W t It is a Wiener process.

[0104] The present invention uses the experience replay pool to eliminate the correlation between input samples, that is, randomly select a small batch of samples from it each time to update the neural network. Since the larger the temporal-difference (TD) error in the samples in the experience replay pool, the greater the effect on back propagation, the faster the convergence of the algorithm will be, and the stability of the training process will also be improved. Therefore, the priority of experience replay is introduced, and the TD error E is used to update the neural network. rror =r t +γ·Q′(s t+1 , a t+1 ,ω′)-Q(s t , a t ,ω) to measure the learning value of each sample, and samples with high priority are more likely to be selected.

[0105] In the implementation process, the constructed water and fertilizer strategy decision model of reinforcement learning can make dynamic decisions in combination with different input strategies, the input parameters of the model are: initial soil water content, growth period start date, growth period days, irrigation frequency, irrigation date, irrigation amount, fertilization frequency, fertilization date, fertilization amount and corresponding meteorological forecast data after the start of the growth period, the output is a WNUE prediction value, the early stopping method is used to determine the optimal water and fertilizer strategy, and when the model is iterated for 200 times, if the WNUE value does not continue to decrease, the water and fertilizer strategy corresponding to the highest WNUE is the recommended optimal configuration.

[0106] The field irrigation and fertilization system dynamic decision device provided by the present application is described below, and the field irrigation and fertilization system dynamic decision device described below can be correspondingly referred to the field irrigation and fertilization system dynamic decision method described above.

[0107] Figure 2 The field irrigation and fertilization system dynamic decision device provided by the present application is described below, and the field irrigation and fertilization system dynamic decision device described below can be correspondingly referred to the field irrigation and fertilization system dynamic decision method described above. Figure 2 As shown in the figure, the field irrigation and fertilization system dynamic decision device comprises: an input module 201 and a processing module 202. The input module 201 is used to obtain the environmental characteristics of the growth period of the crop to be decided, and the environmental characteristics include meteorological characteristics and soil characteristics. The processing module 202 is used to input the environmental characteristics into the trained reinforcement learning network model, and output the water and fertilizer strategy corresponding to the maximum water and fertilizer utilization efficiency. The reinforcement learning network model is trained according to the water and fertilizer strategy expansion data set. The water and fertilizer strategy expansion data set is obtained by performing water and fertilizer strategy feature optimization to determine the maximum water and fertilizer utilization efficiency based on the water and fertilizer utilization efficiency of the water and fertilizer strategy data set by using the Bayesian optimization algorithm. The water and fertilizer strategy data set is a data set obtained by planting based on different water and fertilizer strategies under different meteorological conditions and different soil environments, and the data set includes the corresponding relationship between the environmental characteristics, the water and fertilizer strategy characteristics and the water and fertilizer utilization efficiency.

[0108] The device embodiment provided by the embodiment of the present application is used to realize the above-mentioned method embodiments, and the specific process and detailed content are referred to the above-mentioned method embodiments, which will not be described here.

[0109] The field irrigation and fertilization system dynamic decision device provided by the embodiment of the present application is based on the water and fertilizer strategy data set of historical samples, and the water and fertilizer strategy is configured in combination with the meteorological characteristics and the soil characteristics, so that the confidence of the optimal water and fertilizer strategy can be improved. In addition, the water and fertilizer strategy data set based on history is optimized to obtain the water and fertilizer strategy corresponding to the maximum water and fertilizer utilization efficiency, so that the model is trained based on the environmental characteristics and the water and fertilizer strategy data of the maximum water and fertilizer efficiency, and the real water and fertilizer utilization efficiency of the water and fertilizer strategy obtained by the final decision is improved.

[0110] Figure 3Schematic diagram of the structure of the electronic device provided by the present invention, such as Figure 3 As shown, the electronic device may include: a processor (processor) 301, a communication interface (Communications Interface) 302, a memory (memory) 303 and a communication bus 304, wherein the processor 301, the communication interface 302, and the memory 303 communicate with each other through the communication bus 304. The processor 301 can call the logic instructions in the memory 303 to execute a dynamic decision-making method for field irrigation and fertilization schedules, the method comprising: obtaining environmental characteristics of the growth period of the crop to be decided, the environmental characteristics including meteorological characteristics and soil characteristics; inputting the environmental characteristics into a trained reinforcement learning network model, and outputting a water and fertilizer strategy corresponding to the maximum water and fertilizer utilization efficiency; wherein the reinforcement learning network model is trained based on an extended water and fertilizer strategy data set; the extended water and fertilizer strategy data set is obtained by optimizing the water and fertilizer strategy characteristics of the water and fertilizer strategy data set based on a Bayesian optimization algorithm to determine the maximum water and fertilizer utilization efficiency; the water and fertilizer strategy data set is a data set including the correspondence between environmental characteristics, water and fertilizer strategy characteristics, and water and fertilizer utilization efficiency obtained after planting based on different water and fertilizer strategies under different meteorological conditions and different soil environments.

[0111] In addition, the logic instructions in the above-mentioned memory 303 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when sold or used as an independent product. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0112] In another aspect, the present application also provides a computer program product, which comprises a computer program stored on a non-transitory computer-readable storage medium, and the computer program comprises program instructions, when the program instructions are executed by a computer, the computer can execute the field irrigation and fertilization system dynamic decision-making method provided by the above-mentioned methods, and the method comprises: obtaining environmental characteristics of a growth period of a crop to be decided, the environmental characteristics comprising meteorological characteristics and soil characteristics; inputting the environmental characteristics into a trained reinforcement learning network model to output a water and fertilizer strategy corresponding to a maximum water and fertilizer utilization efficiency; wherein the reinforcement learning network model is trained according to a water and fertilizer strategy expansion data set; the water and fertilizer strategy expansion data set is obtained after water and fertilizer strategy characteristics optimization to determine the maximum water and fertilizer utilization efficiency based on a Bayesian optimization algorithm to water and fertilizer utilization efficiency of a water and fertilizer strategy data set; and the water and fertilizer strategy data set is a data set comprising a corresponding relationship among environmental characteristics, water and fertilizer strategy characteristics and water and fertilizer utilization efficiency, which is obtained after planting based on different water and fertilizer strategies under different meteorological conditions and different soil environments.

[0113] In another aspect, the present application also provides a non-transitory computer-readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the field irrigation and fertilization system dynamic decision-making method provided by the above-mentioned embodiments, and the method comprises: obtaining environmental characteristics of a growth period of a crop to be decided, the environmental characteristics comprising meteorological characteristics and soil characteristics; inputting the environmental characteristics into a trained reinforcement learning network model to output a water and fertilizer strategy corresponding to a maximum water and fertilizer utilization efficiency; wherein the reinforcement learning network model is trained according to a water and fertilizer strategy expansion data set; the water and fertilizer strategy expansion data set is obtained after water and fertilizer strategy characteristics optimization to determine the maximum water and fertilizer utilization efficiency based on a Bayesian optimization algorithm to water and fertilizer utilization efficiency of a water and fertilizer strategy data set; and the water and fertilizer strategy data set is a data set comprising a corresponding relationship among environmental characteristics, water and fertilizer strategy characteristics and water and fertilizer utilization efficiency, which is obtained after planting based on different water and fertilizer strategies under different meteorological conditions and different soil environments.

[0114] The device embodiments described above are only schematic, wherein the units illustrated as separate components can or can not be physically separate, and the components illustrated as units can or can not be physical units, i.e., can be located in one place, or can be distributed on a plurality of network units. Part or all of the modules can be selected according to actual needs to achieve the purpose of the present embodiment scheme. Those skilled in the art can understand and implement without creative labor.

[0115] Those skilled in the art can clearly understand the technical solutions of the various embodiments from the above description of the embodiments, and the various embodiments can be implemented by means of software with the necessary general hardware platforms, and of course, can also be implemented by hardware. Based on such understanding, the above technical solutions, essentially or in other words, the part of the prior art that makes a contribution, can be embodied in the form of a software product, which can be stored in a computer readable storage medium, such as a ROM / RAM, a magnetic disk, an optical disk, and the like, and includes a number of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0116] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, rather than limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for some technical features therein; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A dynamic decision-making method for field irrigation and fertilization schedule, characterized in that: include: Obtaining environmental characteristics of the growth period of the crop to be determined, wherein the environmental characteristics include meteorological characteristics and soil characteristics; Input the environmental characteristics into the trained reinforcement learning network model, and output the water and fertilizer strategy corresponding to the maximum water and fertilizer utilization efficiency; The reinforcement learning network model is trained based on an extended water and fertilizer strategy dataset; the extended water and fertilizer strategy dataset is obtained by performing water and fertilizer strategy feature optimization on the water and fertilizer utilization efficiency of the water and fertilizer strategy dataset based on a Bayesian optimization algorithm to determine the maximum water and fertilizer utilization efficiency; the water and fertilizer strategy dataset is a dataset of the corresponding relationships between environmental characteristics, water and fertilizer strategy characteristics, and water and fertilizer utilization efficiency obtained after planting under different water and fertilizer strategies under different meteorological conditions and soil environments; Before inputting the environmental characteristics into the trained reinforcement learning network model and outputting the water and fertilizer strategy corresponding to the maximum water and fertilizer utilization efficiency, the method further includes: Acquire historical meteorological data and soil data of crops planted under different meteorological conditions and soil conditions according to the determined water and fertilizer strategy to obtain environmental characteristics, and construct the water and fertilizer strategy dataset by combining the water and fertilizer strategy characteristics and the calculated water and fertilizer use efficiency; Taking water and fertilizer use efficiency as the optimization goal, the proxy model is determined according to the Bayesian algorithm and adjusted based on the tree structure probability density estimation algorithm to obtain the final proxy model. Determine a sampling function based on an expected improvement algorithm, perform multiple iterative sampling on the water-fertilization strategy corresponding to each environmental feature in the water-fertilization strategy dataset, and select the next evaluation point that increases the objective function value in each iteration to update the water-fertilization strategy and water-fertilization utilization efficiency, thereby obtaining the water-fertilization strategy extended dataset with the maximum water-fertilization utilization efficiency and corresponding water-fertilization strategy characteristics; Before inputting the environmental characteristics into the trained reinforcement learning network model and outputting the water and fertilizer strategy corresponding to the maximum water and fertilizer utilization efficiency, the method further includes: Expand the dataset based on water and fertilizer strategies and learn from the strategy The reinforcement learning network model is trained.

2. The dynamic decision-making method for field irrigation and fertilization schedule according to claim 1, characterized in that: The data set is expanded according to the water and fertilizer strategy, based on the strategy learning method Training the reinforcement learning network model includes: The combination of each environmental feature and water-fertilizer strategy feature in the water-fertilizer strategy extended dataset is used as the action space, and the corresponding water and fertilizer utilization efficiency under each combination of environmental feature and water-fertilizer strategy feature is used as the state space, and the corresponding loss function is set. exist of Part, through The network selects an action based on the current state and updates the policy network parameters according to the loss function ;pass The network selects the corresponding optimal action according to the next state of the sample, and Network parameters renew Network parameters ; exist of Part, through The network calculates the current state and current action value, and update the network parameters ω according to the loss function; by The network calculates the next state based on the next state and the next action value, and according to The network's parameter ω is updated Network parameters .

3. The dynamic decision-making method for field irrigation and fertilization schedule according to claim 1, characterized in that: The water and fertilizer utilization efficiency is determined based on crop yield, fertilizer application amount, and irrigation amount.

4. The dynamic decision-making method for field irrigation and fertilization schedule according to claim 3, characterized in that: The crop yield is based on environmental characteristics and the water-fertilizer efficiency model Model obtained.

5. A dynamic decision-making device for field irrigation and fertilization schedule, characterized in that: include: An input module is used to obtain environmental characteristics of the crop growth period to be determined, wherein the environmental characteristics include meteorological characteristics and soil characteristics; a processing module for inputting the environmental characteristics into a trained reinforcement learning network model and outputting a water and fertilizer strategy corresponding to the maximum water and fertilizer utilization efficiency; The reinforcement learning network model is trained based on an extended water and fertilizer strategy dataset; the extended water and fertilizer strategy dataset is obtained by performing water and fertilizer strategy feature optimization on the water and fertilizer utilization efficiency of the water and fertilizer strategy dataset based on a Bayesian optimization algorithm to determine the maximum water and fertilizer utilization efficiency; the water and fertilizer strategy dataset is a dataset of the corresponding relationships between environmental characteristics, water and fertilizer strategy characteristics, and water and fertilizer utilization efficiency obtained after planting under different water and fertilizer strategies under different meteorological conditions and soil environments; Before inputting the environmental characteristics into the trained reinforcement learning network model and outputting the water and fertilizer strategy corresponding to the maximum water and fertilizer utilization efficiency, the method further includes: Acquire historical meteorological data and soil data of crops planted under different meteorological conditions and different soil conditions according to the determined water and fertilizer strategy, obtain environmental characteristics, and construct the water and fertilizer strategy dataset by combining the water and fertilizer strategy characteristics and the calculated water and fertilizer use efficiency; Taking water and fertilizer use efficiency as the optimization goal, the proxy model is determined according to the Bayesian algorithm and adjusted based on the tree structure probability density estimation algorithm to obtain the final proxy model. Determine a sampling function based on an expected improvement algorithm, perform multiple iterative sampling on the water-fertilization strategy corresponding to each environmental feature in the water-fertilization strategy dataset, and select the next evaluation point that increases the objective function value in each iteration to update the water-fertilization strategy and water-fertilization utilization efficiency, thereby obtaining the water-fertilization strategy extended dataset with the maximum water-fertilization utilization efficiency and corresponding water-fertilization strategy characteristics; Before inputting the environmental characteristics into the trained reinforcement learning network model and outputting the water and fertilizer strategy corresponding to the maximum water and fertilizer utilization efficiency, the method further includes: Expand the dataset based on water and fertilizer strategies and learn from the strategy The reinforcement learning network model is trained.

6. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the steps of the dynamic decision-making method for field irrigation and fertilization schedule as claimed in any one of claims 1 to 4 are implemented.

7. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the dynamic decision-making method for field irrigation and fertilization schedule as claimed in any one of claims 1 to 4 are implemented.

Images