A field trait detection method based on crop breeding

Abstract

The present application relates to the technical field of field trait detection, and particularly relates to a field trait detection method based on crop breeding, comprising the following steps: determining an environment penalty value according to detected field crop color data and soil data, determining a yield reward value according to detected field crop growth data and nutrient data stored in a crop management platform, and determining a fertilizer utilization value according to detected field soil element data; performing weighted summation to generate a fertilization adjustment reward function, wherein a dynamic coefficient is determined according to crop breeding stage big data fitting of the crop management platform; updating a fertilization amount reinforcement learning model based on the fertilization adjustment reward function at regular time intervals, and determining a breeding multi-element recommended fertilization amount based on the fertilization amount reinforcement learning model by combining a Critic network with an Actor network. The present application realizes collaborative closed-loop detection of stress resistance, yield increase and fertilizer saving and field detection data mining by combining a reinforcement learning model with breeding.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of field trait detection, and particularly relates to a field trait detection method based on crop breeding. BACKGROUND

[0002] Crop breeding is a core link for improving the yield, quality and stress resistance of agricultural products, and field trait detection as a key supporting technology in the breeding process directly affects the breeding efficiency and the final variety performance. With the development of agricultural modernization, how to dynamically optimize the breeding management strategy based on real-time field data has become a research hotspot in the field of crop breeding.

[0003] In recent years, the application of machine learning technology in the field of agriculture provides a new idea for solving the above problems, and in particular, reinforcement learning shows potential through dynamic decision optimization. However, existing researches mostly focus on a single target, and do not build a multi-dimensional reward mechanism considering environmental cost, yield benefit and resource utilization, and lack of target function design balancing short-term fertilization effect and long-term soil health.

[0004] Therefore, the traditional field trait detection has the following problems:

[0005] Firstly, the single data collection dimension leads to the staticization of the fertilization strategy. The existing methods mostly use single sensors, which are difficult to comprehensively judge the data for adjusting the fertilization amount, and the traditional fertilization scheme is based on fixed formula. Therefore, the traditional field trait detection does not dynamically adjust in combination with real-time field data and breeding goals, which easily causes fertilizer waste or insufficient nutrient supply, affecting the breeding effect.

[0006] Secondly, the utilization efficiency of big data is low. Although part of the technology introduces agricultural big data, it does not deeply integrate it with real-time detection data, and cannot dynamically adjust key parameters through data fitting, which limits the adaptability and generalization ability of the model.

[0007] Therefore, how to combine multi-dimensional reward item design, dynamic coefficient weighting and reinforcement learning model to realize the precise recommendation of fertilization amount and the comprehensive detection of field traits in the breeding process is a technical problem to be solved at present. SUMMARY

[0008] Therefore, the present application provides a field trait detection method based on crop breeding, which realizes the precise recommendation of fertilization amount and the comprehensive detection of field traits in the breeding process through the environmental crop damage penalty item, the yield optimization reward item and the fertilizer waste penalty item, realizes the comprehensive evaluation of the crop growth environment, the stress resistance state and the nutrient demand, realizes the collaborative closed-loop detection of the stress resistance, yield increase and fertilizer saving of the reinforcement learning model in combination with different stages of breeding, and realizes the precise field detection data mining.

[0009] To achieve the above object, the present application provides a field trait detection method based on crop breeding, comprising:

[0010] An environmental penalty value is determined according to the detected field crop color data and soil data, a yield reward value is determined according to the detected field crop growth data and the nutrient data stored in the crop management platform, and a fertilizer utilization value is determined according to the detected field soil element data;

[0011] A dynamic coefficient is determined according to the crop breeding stage big data fitting of the crop management platform, and the environmental penalty value, the yield reward value and the fertilizer utilization value are weighted and summed with the corresponding dynamic coefficient to generate a fertilization adjustment reward function;

[0012] In the process of iteratively updating the fertilization amount reinforcement learning model based on the fertilization adjustment reward function, the fertilization amount reinforcement learning model determines a breeding multi-element recommended fertilization amount based on a Critic network combined with an Actor network, wherein the Critic network is used to generate a target action value that balances short-term fertilization effect and long-term soil health.

[0013] Further, the process of determining the environmental penalty value comprises:

[0014] The overall electrical conductivity in the field is calculated according to the soil electrical conductivity data detected by multiple electrical conductivity sensors in the field;

[0015] The crop disease state value is determined according to the root stem color data detected by the visual sensor in the field;

[0016] The environmental penalty value is comprehensively calculated according to the overall electrical conductivity in the field, the crop disease state value and the soil nitrate nitrogen leaching amount detected by the water quality sensor of the field drainage ditch.

[0017] Further, the process of determining the yield reward value comprises:

[0018] The crop management platform stores the historical crop plant volume detected by the laser radar in the field;

[0019] The plant volume increment is calculated according to the historical crop plant volume;

[0020] The nitrogen application penalty term is calculated based on the difference between the actual phosphorus application amount recorded by the crop management platform and the recommended value of the crop phosphorus application amount;

[0021] The leaf area index fluctuation coefficient is calculated based on the multiple root stem color data detected by the visual sensor in the field;

[0022] The yield reward value is calculated based on the plant volume increment, the nitrogen application penalty term and the leaf area index fluctuation coefficient.

[0023] Further, the process of determining the fertilizer utilization value comprises:

[0024] determining a potassium element conversion value according to the biological yield stored in the crop management platform and the actual potassium application amount;

[0025] calculating the fertilizer utilization value according to the weighted difference between the potassium element conversion value and the nitrate potassium leaching amount detected by the water quality sensor.

[0026] In particular, the calculation of the environmental penalty value integrates the key indicators of soil conductivity, root color, and nitrate nitrogen leaching amount, providing a quantitative basis for avoiding environmental risks in the breeding process. The calculation of the yield reward value realizes the whole-process monitoring of the key links of yield formation in crop breeding through the plant volume increment, nitrogen application penalty term, and leaf area index fluctuation coefficient. The calculation of the fertilizer utilization value quantifies the loss degree of potassium fertilizer through the weighted difference between the potassium element conversion value and the nitrate potassium leaching amount.

[0027] Further, the process of determining the breeding multi-element recommended fertilization amount by the fertilization amount reinforcement learning model comprises:

[0028] constructing a state space including the soil conductivity, crop disease state value, soil nitrate nitrogen leaching amount, plant volume increment, actual phosphorus application amount, leaf area index fluctuation coefficient, and nitrate potassium leaching amount;

[0029] constructing an action space including the recommended nitrogen element application amount, recommended phosphorus element application amount, and recommended potassium element application amount;

[0030] the state space and the action space generate a target action value through the Critic network;

[0031] the state space generates the action space through the Actor network, and determines the breeding multi-element recommended fertilization amount according to the action space.

[0032] Further, the process of generating the target action value comprises: the Critic network determines the target action value based on the fertilization adjustment reward function, discount factor, Critic network output value, state space, and Actor network output value, wherein the discount factor is a balance value of short-term fertilization effect and long-term soil health determined according to the field type.

[0033] Further, the process of generating the action space comprises: the Actor network updates once through the back propagation gradient ascent algorithm when the Critic network is updated for a set number of times, wherein the set number of times is determined according to the normal fertilization frequency of the crop variety.

[0034] Further, the Critic network outputs a network prediction value through a ReLU activation function of a second hierarchical full connection layer.

[0035] The Actor network increases the nonlinear expression ability of the target action value and the state space representation feature through the ReLU activation functions of the first full connection layer and the second full connection layer, and after updating the action space through the output value of the Tanh activation function of the third full connection layer, the action space is multiplied by the standard fertilization amount to determine the recommended fertilization amount of the breeding multi-element.

[0036] Further, the training process of the Critic network is:

[0037] The Critic network is optimized and trained based on an improved mean square error loss function of the target action value and the network prediction value, wherein the improved mean square error loss function is provided with a plurality of period action space smoothing penalty terms to reduce the change amplitude of the target action value of the plurality of periods.

[0038] Especially, the Actor-Critic network combines the interaction effect of disease treatment and fertilization amount, the Critic network evaluates the action value, the Actor network generates the optimal fertilization action, realizes the adaptive optimization in the complex scene, realizes the balance between the short-term fertilization effect such as rapid yield increase and the long-term soil health such as salinization prevention in the crop breeding process, realizes the conversion of agricultural experience into a quantitative model through the reinforcement learning framework, realizes the intelligent, precise and ecological of the fertilization strategy, and the timeliness and effectiveness of the breeding trait detection process, effectively balances yield, resource utilization and environmental protection.

[0039] Further, the determination process of the dynamic coefficient comprises:

[0040] After the improved multilayer perceptron model is optimized and trained through the improved mean square error loss function, the dynamic coefficient is determined through the improved multilayer perceptron model corresponding to the crop breeding stage big data of the state space, wherein the improved multilayer perceptron model is provided with a nitrogen element increase super limit constraint.

[0041] Further, the improved multilayer perceptron model is optimized and trained through the improved mean square error loss function.

[0042] Compared with the prior art, the present application has the beneficial effects that through the environment crop damage penalty term, the yield optimization reward term and the fertilizer waste penalty term, the present application realizes the accurate recommendation of the fertilization amount and the comprehensive detection of the field traits in the breeding process, realizes the comprehensive evaluation of the crop growth environment, the stress resistance state and the nutrient demand, realizes the collaborative closed loop detection of stress resistance, yield increase and fertilizer saving in different breeding stages, and realizes the precise field detection data mining.

[0043] Especially, the application integrates the key indicators of soil conductivity, root color, and nitrate leaching amount through the calculation of environmental penalty value, provides a quantitative basis for avoiding environmental risks in the breeding process, realizes the whole-process monitoring of the key links of yield formation in crop breeding through the calculation of yield reward value by plant volume increment, nitrogen application penalty term, and leaf area index fluctuation coefficient, and quantifies the loss degree of potassium fertilizer through the weighted difference value of potassium element conversion value and nitrate potassium leaching amount.

[0044] Especially, the application realizes self-adaptive optimization in a complex scene by combining the interaction effect of disease and fertilizer amount through Actor-Critic network, evaluates the action value through Critic network, and generates the optimal fertilization action through Actor network, realizes the balance between rapid yield increase and long-term soil health such as salinization prevention in the crop breeding process, realizes the conversion of agricultural experience into a quantitative model through the reinforcement learning framework, realizes the intelligent, precise, and ecological fertilization strategy, and the timeliness and effectiveness of the breeding trait detection process, and effectively balances yield, resource utilization, and environmental protection. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 It is a flowchart of the crop breeding-based field trait detection method of the embodiment of the application;

[0046] Figure 2 It is a calculation flowchart of the fertilization adjustment reward function of the crop breeding-based field trait detection method of the embodiment of the application;

[0047] Figure 3 It is a use flowchart of the fertilization amount reinforcement learning model of the crop breeding-based field trait detection method of the embodiment of the application;

[0048] Figure 4 It is a fertilization amount reinforcement learning model structure diagram of the crop breeding-based field trait detection method of the embodiment of the application. DETAILED DESCRIPTION

[0049] In order to make the purpose and advantages of the application more clear and apparent, the application will be further described below in combination with embodiments; it should be understood that the specific embodiments described herein are only used to explain the application, and do not limit the application.

[0050] The preferred embodiments of the application will be described below with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are only used to explain the technical principles of the application, and are not intended to limit the protection scope of the application.

[0051] It should be noted that in the description of the present application, the terms of direction or position relationship indicated by "upper", "lower", "left", "right", "inner", "outer" and the like are based on the direction or position relationship shown in the drawings, which is only for the convenience of description, and does not indicate or imply that the device or element must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application.

[0052] In addition, it should also be noted that in the description of the present application, unless otherwise explicitly specified and limited, the terms "mounting", "connecting", "connecting" should be understood broadly, for example, it can be fixed connection, or detachable connection, or integral connection; it can be mechanical connection, or electrical connection; it can be directly connected, or indirectly connected through intermediate medium, or the communication inside two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0053] As shown in Figures 1 to 4 The present application provides a field trait detection method based on crop breeding, which realizes the accurate recommendation of fertilizer amount and the comprehensive detection of field traits in the breeding process through environmental crop damage penalty term, yield optimization reward term and fertilizer waste penalty term, realizes the comprehensive evaluation of crop growth environment, stress resistance state and nutrient demand, realizes the collaborative closed loop detection of stress resistance, yield increase and fertilizer saving in different stages of breeding, and realizes the precise field detection data mining.

[0054] As shown in Figure 1 The present embodiment provides a field trait detection method based on crop breeding, which comprises:

[0055] Step S1, determining an environmental penalty value according to the detected field crop color data and soil data, determining a yield reward value according to the detected field crop growth data and the nutrient data stored in the crop management platform, and determining a fertilizer utilization value according to the detected field soil element data;

[0056] Step S2, determining a dynamic coefficient according to the crop breeding stage big data fitting of the crop management platform, and performing weighted summation on the environmental penalty value, the yield reward value and the fertilizer utilization value and the corresponding dynamic coefficient to generate a fertilizer adjustment reward function;

[0057] Step S3, in the process of iterative updating the fertilizer amount reinforcement learning model based on the fertilizer adjustment reward function, the fertilizer amount reinforcement learning model determines the recommended fertilizer amount of multiple elements and the next nutrient detection period based on the Critic network combined with the Actor network, wherein the Critic network is used to generate the target action value of balancing the short-term fertilizer effect and the long-term soil health.

[0058] Specifically, the crop management platform is connected with the field sensors through Internet of Things networking to realize storage and mining application of field trait detection data of crop breeding. The crop management platform includes a sensor network as a perception layer of a field data portal, an LPWAN edge gateway as a transport layer, a TimescaleDB time series database as a platform layer, a MySQL / PostgreSQL relational database, and a fertilization amount reinforcement learning model as an application layer.

[0059] As shown in Figure 2 , the process of determining the environmental penalty value includes:

[0060] calculating the overall field conductivity according to the soil conductivity data detected by the multiple conductivity sensors in the field;

[0061] determining the crop disease state value according to the root stem color data detected by the visual sensor in the field;

[0062] comprehensively calculating the environmental penalty value according to the overall field conductivity, the crop disease state value, and the soil nitrate nitrogen leaching amount detected by the water quality sensor of the field drainage ditch.

[0063] Specifically, after removing abnormal data in the multiple soil conductivity data by adopting the 3 times standard deviation principle, the Lagrange algorithm is used to complete the missing data, and the area weighted average method (preferably, the weight is the proportion of the area of each sub-region to the total area of the field) is used to calculate the overall field conductivity. After determining whether the disease and insect pests occur according to the proportion of green color of root stem in RGB format and HSV format (preferably, in RGB format and HSV format, the green component of healthy root stem is significantly higher than the red and blue colors. The proportion of green color of root stem in the total pixels is calculated based on the proportion of pixels satisfying the condition in ROI, and whether the root stem is at risk of losing green, i.e., whether the disease and insect pests occur, is determined based on the comparison between the proportion and a threshold value), the yellowing index is calculated as the crop disease state value, and the yellowing index is R-B / R+G+B.

[0064] Specifically, the process of calculating the environmental penalty value is:

[0065]

[0066] In the formula, is the environmental penalty value, Disease is the crop disease state value, N_Leaching is the soil nitrate nitrogen leaching amount, is the overall field conductivity. Wherein is determined according to the breeding stage of the crop, for example, is greater than . The nitrate nitrogen leaching amount is obtained by a soil solution sensor.

[0067] It should be noted that the formulas of the functions used to calculate the reinforcement learning model in this embodiment are all subjected to Min-Max normalization or Z-score standardization processing to avoid dimensional conflicts.

[0068] It can be understood that the breeding stage includes: seedling stage (nutritive organ construction), tillering stage (population structure formation), flowering stage (reproductive growth initiation), grain filling stage (grain material accumulation), and maturation stage (harvest quality setting). The breeding stage can be determined by detecting the stem elongation rate by laser radar. For example, the seedling stage is 0.7, 0.2, and 0.1, the tillering stage is 0.3, 0.6, and 0.1, and the grain filling stage is 0.2, 0.5, and 0.3. Therefore, in the seedling stage, the control of pest threats is biased, and in the tillering stage, the nitrogen fertilizer is biased.

[0069] Therefore, the environmental penalty value can reduce fertilizer waste, avoid environmental pollution, and avoid excessive fertilization to cause salt damage or diseases of crops according to the breeding stage.

[0070] As shown in Figure 2 , in this embodiment, the process of determining the yield reward value includes:

[0071] The crop management platform stores the historical crop plant volume detected by the field laser radar;

[0072] The plant volume increment is calculated according to the historical crop plant volume;

[0073] The nitrogen application penalty term is calculated based on the difference between the actual phosphorus application amount and the crop phosphorus application recommended value recorded by the crop management platform;

[0074] The leaf area index fluctuation coefficient is calculated according to the plurality of root stem color data detected by the field visual sensor;

[0075] The yield reward value is calculated based on the plant volume increment, the nitrogen application penalty term, and the leaf area index fluctuation coefficient.

[0076] Specifically, the process of calculating the yield reward value is:

[0077]

[0078] In the formula, Y is the yield reward value, a, b, and c are three weights, V is the plant volume increment, P is the actual phosphorus application amount, and P' is the crop phosphorus application recommended value, and the difference between them is the phosphorus amount difference, ​is the leaf area index fluctuation coefficient. The leaf area index fluctuation coefficient is estimated by multispectral imaging.

[0079] It can be understood that there is a hierarchical influence among the plant volume increment, the phosphorus amount difference, and the leaf area index fluctuation coefficient, and the effects of the three variables on yield are logically clear, and the corresponding weights are set by expert experience and structural equation. Specifically, the plant volume increment directly reflects the accumulation of crop biomass and is the basis for yield formation, and the corresponding weight is 0.5 to 0.7, which is determined according to the crop variety, for example, the biomass at the tasseling stage of corn is highly correlated with the final yield, so the weight is 0.7. Phosphorus is a key nutrient for crops, and if the phosphorus amount difference is too small, the crop will be deficient in phosphorus, or if it is too large, the seedling will be burned, which will inhibit yield, and the corresponding weight is 0.2 to 0.3, which is determined according to the breeding stage of the crop, for example, if the crop is in the seedling stage which is sensitive to phosphorus, the weight is 0.3. The leaf area index fluctuation coefficient reflects the stability of photosynthetic efficiency, which promotes biomass growth through photosynthetic efficiency, so the leaf area index fluctuation coefficient directly drives the accumulation of plant volume increment and indirectly affects the yield reward value, so the corresponding weight is determined by structural equation , in which is the plant volume increment, is the leaf area index fluctuation coefficient, is the corresponding weight, is the intercept, is the residual error, and reflect the influence of unobserved variables, which are determined by least squares or maximum likelihood estimation fitting of historical data of the same type of crop.

[0080] As shown in Figure 2 , in this embodiment, the process of determining the fertilizer utilization value includes:

[0081] determining the potassium element conversion value according to the biological yield stored by the crop management platform divided by the actual potassium application amount;

[0082] calculating the fertilizer utilization value according to the weighted difference value of the potassium element conversion value and the nitrate potassium leaching amount detected by the water quality sensor.

[0083] Specifically, the nitrate nitrogen leaching amount and the nitrate potassium leaching amount are monitored using ion selective electrodes, such as nitrate sensors.

[0084] Specifically, the process of calculating the fertilizer utilization value is:

[0085]

[0086] wherein, is the fertilizer utilization value, are two weights, is the biological yield, is the the first redundancy term of, avoid the denominator to take zero, is the actual potassium application amount, is the nitrate potassium leaching amount.

[0087] It can be understood that the biological yield directly reflects the effect of potassium fertilizer, and the yield is high when the potassium fertilizer is sufficient, but there is no gain or even yield reduction when the amount is excessive. The actual potassium application amount reflects the potassium fertilizer input cost, and needs to be set based on the input-output relationship to determine the fertilizer utilization value. Therefore, is a set value according to the field environment, for example, the wheat crop is planted in the medium potassium content soil, and its are 0.6 and 0.4 respectively, and the corn crop is planted in the potassium deficient soil, and the potassium fertilizer input has more significant effect on the yield, and its are 0.5 and 0.5 respectively.

[0088] Specifically, the fertilizer adjustment reward function is:

[0089]

[0090] wherein, is the fertilizer adjustment reward function, is the environmental penalty value, is the yield reward value, is the fertilizer utilization value, are three dynamic coefficients.

[0091] In the embodiment, the determination process of the dynamic coefficient includes:

[0092] After the improved multi-layer perception machine model is optimized and trained by improving the mean square error loss function, the dynamic coefficient is determined by the improved multi-layer perception machine model through the state space corresponding to the crop breeding stage big data, wherein the improved multi-layer perception machine model is provided with a nitrogen element increase application overrun constraint.

[0093] Specifically, the nitrogen element increase application overrun constraint is:

[0094]

[0095] wherein, is the nitrogen element increase application overrun constraint, are the actual nitrogen application amount and the crop nitrogen application recommended value respectively.

[0096] Specifically, the improved multi-layer perception model converts the crop breeding stage big data corresponding to the state space into a vector through Z-score standardization and One-Hot encoding, and learns features from the original state and identifies the priority of the optimization target in different growth periods through a 3-layer fully connected layer (FC) with 128 nodes in each layer.

[0097] In this embodiment, the improved multi-layer perception model is optimized and trained through the improved mean square error loss function.

[0098] Specifically, the improved mean square error loss function is:

[0099]

[0100] In the formula, L is the improved mean square error loss function, N is the number of samples, is the network prediction value, y is the target action value, is a smoothing penalty term.

[0101] In particular, the calculation of the environmental penalty value integrates the key indicators of soil conductivity, root color, and nitrate leaching amount, providing a quantitative basis for avoiding environmental risks during breeding. The calculation of the yield reward value realizes the whole-process monitoring of the key links of yield formation in crop breeding through the plant volume increment, nitrogen application penalty term, and leaf area index fluctuation coefficient. The calculation of the fertilizer utilization value quantifies the loss degree of potassium fertilizer through the weighted difference value of the potassium element conversion value and the nitrate potassium leaching amount.

[0102] As shown in Figure 3 and 4 In this embodiment, the process of determining the recommended fertilization amount of multiple elements for breeding by the fertilization amount reinforcement learning model includes:

[0103] constructing a state space including the soil conductivity, crop disease state value, soil nitrate nitrogen leaching amount, plant volume increment, actual phosphorus application amount, leaf area index fluctuation coefficient, and nitrate potassium leaching amount;

[0104] constructing an action space including the recommended nitrogen element application amount, recommended phosphorus element application amount, and recommended potassium element application amount;

[0105] The state space and the action space generate a target action value through the Critic network;

[0106] The state space generates the action space through the Actor network, and determines the recommended fertilization amount of multiple elements for breeding according to the action space.

[0107] Specifically, the fertilization amount reinforcement learning model is as shown in Figure 4 wherein the action space is, for example =[N,P,K], which is, for example =[N=150kg / ha, P=50kg / ha, K=80kg / ha].

[0108] The state space is .

[0109] It can be understood that in the fertilizer application amount reinforcement learning model, the Critic network is used to evaluate the action value of the target action, and the Actor network is used to generate the action space of the next state, and the action space of the next state is input into the field environment to realize closed-loop fertilizer application amount control and closed-loop detection of field traits.

[0110] As shown in Figure 3 and 4 , in this embodiment, the process of generating the target action value includes: the Critic network determines the target action value based on the fertilizer adjustment reward function, the discount factor, the Critic network output value, the state space and the Actor network output value, wherein the discount factor is a balance value of short-term fertilizer effect and long-term soil health determined according to the field type.

[0111] As shown in Figure 3 and 4 , in this embodiment, the process of generating the action space includes: the Actor network updates once through the back propagation gradient ascent algorithm when the Critic network is updated for a set number of times, wherein the set number of times is determined according to the normal fertilization frequency of the crop variety.

[0112] In this embodiment, the Critic network outputs the network prediction value through the ReLU activation function of the second hierarchical fully connected layer;

[0113] The Actor network increases the nonlinear expression ability of the target action value and the state space representation features through the ReLU activation functions of the first and second fully connected layers, and after updating the action space through the output value of the Tanh activation function of the third fully connected layer, the action space is multiplied by the standard fertilizer application amount to determine the recommended fertilizer application amount of the breeding multi-element.

[0114] Specifically, the input data dimension of the Actor is 8, and the first fully connected layer to the third fully connected layer are respectively provided with 128 nodes, 64 nodes and 64 nodes. The standard fertilizer application amount, for example, nitrogen fertilizer is 150kg / ha, phosphorus fertilizer is 50kg / ha, and potassium fertilizer is 80kg / ha, which can also be determined according to the crop variety.

[0115] Specifically, the input data dimension of the Critic network is 11, and the first hierarchical fully connected layer and the second hierarchical fully connected layer are respectively provided with 256 nodes and 128 nodes.

[0116] In this embodiment, the training process of the Critic network is as follows:

[0117] The Critic network is optimized and trained based on an improved mean square error loss function of the target action value and the network prediction value, wherein the improved mean square error loss function is provided with a smoothing penalty term of the action space of multiple periods, so as to reduce the change range of the target action value of multiple periods.

[0118] Specifically, the fertilization adjustment reward function is as follows:

[0119]

[0120] wherein, is a smoothing penalty term, is an adjustment coefficient, is the action space of the current period t and the previous period t-1 respectively.

[0121] Especially, the Actor-Critic network combines the interaction effect of disease treatment and fertilization amount, the Critic network evaluates the action value, the Actor network generates the optimal fertilization action, realizes the adaptive optimization under the complex scene, realizes the balance between the short-term fertilization effect such as rapid yield increase and the long-term soil health such as salinization prevention in the crop breeding process, realizes the conversion of agricultural experience into a quantitative model through the reinforcement learning framework, realizes the intelligentization, precision and ecology of the fertilization strategy, and the timeliness and effectiveness of the breeding trait detection process, effectively balances the yield, resource utilization and environmental protection.

[0122] In this embodiment, through the environmental crop damage penalty term, the yield optimization reward term and the fertilizer waste penalty term, the comprehensive evaluation of the crop growth environment, the stress resistance state and the nutrient demand is realized, the collaborative closed loop detection of stress resistance, yield increase and fertilizer saving in different stages of breeding is realized, and the precise field detection data mining is realized. The calculation of the environmental penalty value combines the key indicators of soil conductivity, root color and nitrate nitrogen leaching amount, which provides a quantitative basis for avoiding environmental risks in the breeding process. The calculation of the yield reward value realizes the whole process monitoring of the key links of yield formation in crop breeding through the plant volume increment, the nitrogen application penalty term and the leaf area index fluctuation coefficient. The calculation of the fertilizer utilization value quantifies the loss degree of potassium fertilizer through the weighted difference value of the potassium element conversion value and the nitrate potassium leaching amount. The Actor-Critic network combines the interaction effect of disease and fertilizer amount, the Critic network evaluates the action value, the Actor network generates the optimal fertilization action, realizes the self-adaptive optimization in the complex scene, realizes the balance between the short-term fertilization effect such as rapid yield increase and the long-term soil health such as salinization prevention in the crop breeding process, realizes the conversion of agricultural experience into a quantitative model through the reinforcement learning framework, realizes the intelligentization, precision and ecology of the fertilization strategy, and the timeliness and effectiveness of the breeding trait detection process, and effectively balances yield, resource utilization and environmental protection.

[0123] So far, the technical solutions of the present application have been described in combination with the preferred embodiments shown in the drawings, but it is easy for those skilled in the art to understand that the protection scope of the present application is obviously not limited to these specific embodiments. Those skilled in the art can make equivalent changes or replacements to related technical features without departing from the principles of the present application, and the technical solutions after these changes or replacements will fall within the protection scope of the present application.

[0124] The above description is only the preferred embodiments of the present application and is not intended to limit the present application; for those skilled in the art, the present application can have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for detecting field traits based on crop breeding, characterized by, The method comprises the following steps: determining an environmental penalty value according to detected field crop color data and soil data, determining a yield reward value according to detected field crop growth data and nutrient data stored in a crop management platform, and determining a fertilizer utilization value according to detected field soil element data; determining a dynamic coefficient according to big data fitting of a crop breeding stage of the crop management platform, and performing weighted summation of the environmental penalty value, the yield reward value and the fertilizer utilization value and the corresponding dynamic coefficient to generate a fertilization adjustment reward function; in the process of iteratively updating the fertilization amount reinforcement learning model based on the fertilization adjustment reward function, the fertilization amount reinforcement learning model determines a breeding multi-element recommended fertilization amount based on a Critic network combined with an Actor network, wherein the Critic network is used to generate a target action value balancing short-term fertilization effect and long-term soil health; wherein the process of calculating the fertilization adjustment reward function is: ; wherein, is an environmental penalty value, are three weights, Disease is a crop disease state value, N_Leaching is a soil nitrate nitrogen leaching amount, is a field overall conductivity, is a yield reward value, are three weights, is a plant volume increment, are an actual phosphorus application amount and a crop phosphorus application recommended value respectively, is a leaf area index fluctuation coefficient, is a fertilizer utilization value, are two weights, is a biological yield, is a first redundancy term for avoiding denominator zero, is an actual potassium application amount, is a nitrate potassium leaching amount, R is a fertilization adjustment reward function, are three dynamic coefficients; The determination process of the dynamic coefficient comprises: after optimization training of an improved multilayer perception machine model through an improved mean square error loss function, determining the dynamic coefficient through the improved multilayer perception machine model for the state space corresponding to the crop breeding stage big data, wherein the improved multilayer perception machine model is provided with a nitrogen element increase super limit constraint; the nitrogen element increase super limit constraint is: ; wherein, is a nitrogen element increase super limit constraint, respectively, is an actual nitrogen application amount and a crop nitrogen application recommended value.

2. The field trait detection method based on crop breeding according to claim 1, characterized in that, The process of determining the environmental penalty value comprises: calculating the overall electrical conductivity in the field according to the soil electrical conductivity data detected by multiple electrical conductivity sensors in the field; determining a crop disease state value according to root stem color data detected by a visual sensor in the field; determining the amount of soil nitrate nitrogen leaching loss through a water quality sensor in a field drainage ditch.

3. The field trait detection method based on crop breeding according to claim 2, characterized in that, The process of determining the yield reward value comprises: the crop management platform stores historical crop plant volume detected by a laser radar in the field; calculating the plant volume increment according to the historical crop plant volume; obtaining the actual phosphorus application amount and the crop phosphorus application amount recommended value recorded by the crop management platform; calculating the leaf area index fluctuation coefficient according to the multiple root stem color data detected by the visual sensor in the field.

4. The field trait detection method based on crop breeding according to claim 3, characterized in that, The process of determining the fertilizer utilization value comprises: obtaining the biological yield and the actual potassium application amount stored in the crop management platform; obtaining the amount of nitrate potassium leaching loss detected by the water quality sensor.

5. The field trait detection method based on crop breeding according to claim 4, characterized in that, The process of determining the breeding multi-element recommended fertilization amount by the fertilization amount reinforcement learning model comprises: constructing a state space comprising the soil electrical conductivity, the crop disease state value, the soil nitrate nitrogen leaching loss, the plant volume increment, the actual phosphorus application amount, the leaf area index fluctuation coefficient and the amount of nitrate potassium leaching loss; constructing an action space comprising the recommended nitrogen element application amount, the recommended phosphorus element application amount and the recommended potassium element application amount; the state space and the action space generate a target action value through the Critic network; the target action value and the state space update the action space through the Actor network, and determine the breeding multi-element recommended fertilization amount according to the action space.

6. The field trait detection method based on crop breeding according to claim 5, characterized in that, The process of generating the target action value comprises: the Critic network determines the target action value based on the fertilization adjustment reward function, a discount factor, a Critic network output value, the state space and an Actor network output value, wherein the discount factor is a balance value of short-term fertilization effect and long-term soil health determined according to the field type.

7. The field trait detection method based on crop breeding according to claim 5, characterized in that, The process of generating the action space comprises: the Actor network performs an update through a back propagation gradient ascent algorithm once when the Critic network is updated for a set number of times, wherein the set number of times is determined according to the normal fertilization frequency of the crop variety.

8. The field trait detection method based on crop breeding according to claim 5, characterized in that, The Critic network outputs a network prediction value through a ReLU activation function of a second hierarchical full connection layer; The Actor network increases the nonlinear expression ability of the target action value and the state space representation features through the ReLU activation functions of the first full connection layer and the second full connection layer, and after updating the action space through the output value of the Tanh activation function of the third full connection layer, the action space is multiplied by the standard fertilization amount to determine the breeding multi-element recommended fertilization amount.

9. The field trait detection method based on crop breeding according to claim 8, characterized in that, The training process of the Critic network is as follows: The Critic network is optimized and trained based on an improved mean square error loss function of the target action value and the network prediction value, wherein the improved mean square error loss function is provided with a plurality of period action space smoothing penalty terms to reduce the change range of the target action value of a plurality of periods.

Images