Intelligent water distribution system and method for intercropping system root space-time competition

Abstract

The invention relates to the field of agricultural irrigation, in particular to an intelligent water distribution system and method for root space-time competition of an intercropping system, and is applied to agricultural precise irrigation, efficient utilization of water resources and management of the intercropping system. According to the system, the ground penetrating radar and the multispectral laser technology are fused, and crop root system distribution, activity and nutrient data are accurately obtained. Based on transfer learning and multi-agent reinforcement learning algorithms, a root system competition model is constructed, and an optimal water distribution strategy is intelligently decided. The deformable drip irrigation execution module achieves precise irrigation, the water consumption is greatly reduced by 20%-35%, and waste is effectively avoided. Meanwhile, the system balances the growth requirements of intercropping crops, relieves root competition, improves the overall yield and stability, and the amplification reaches 15%-25%. According to the system, the utilization efficiency of water resources is remarkably improved, and powerful support is provided for sustainable development of modern agriculture.

Description

Technical Field

[0001] The present invention relates to the field of agricultural irrigation, and particularly to an intelligent water distribution system and method for the spatio-temporal competition of roots in an intercropping system, which is applied to agricultural precision irrigation, efficient utilization of water resources, and management of the intercropping system. Background Art

[0002] Intercropping is an agricultural technique of simultaneously or alternately planting two or more crops on the same piece of land, which can make full use of land resources, increase the yield per unit area, and enhance ecological stability. However, in an intercropping system, there are spatial and resource competition relationships among the roots of different crops, and this competition relationship has an important impact on crop growth and water and fertilizer absorption.

[0003] Traditional agricultural irrigation systems generally adopt uniform or regional irrigation methods, and cannot accurately distribute water according to the complex competition relationships among the roots of different crops in the intercropping system. This leads to three main problems: First, the water resource utilization efficiency is low, and over-irrigation may occur in some areas, while water shortage exists in other areas; second, the crop growth is unbalanced, and the dominant crops often obtain too much water and fertilizer resources, while the growth of the weak crops is restricted; third, there is a lack of effective means to monitor and predict the development of the roots of different crops in real time, and it is difficult to adjust the irrigation strategy according to the crop growth dynamics.

[0004] In the prior art, there have been some studies on root monitoring and irrigation optimization. For example, some methods use a single ground penetrating radar technology to monitor the root distribution, but they can often only obtain two-dimensional plane data and cannot comprehensively reflect the three-dimensional spatial structure and competition relationship of the roots. There are also some intelligent irrigation systems that use simple machine learning algorithms to optimize the water distribution strategy, but these systems are usually designed for single crops and cannot effectively cope with the complex root competition problems in the intercropping system.

[0005] Therefore, there is an urgent need to develop an intelligent system that can monitor the spatial distribution and competition relationship of the roots of different crops in the intercropping system in real time, and optimize the water distribution strategy accordingly, so as to improve the water resource utilization efficiency, balance the growth requirements of different crops, and achieve high yield and high efficiency in the intercropping system. Summary of the Invention

[0006] The purpose of the present invention is to provide an intelligent water distribution system and method for the spatio-temporal competition of roots in an intercropping system, which can monitor the spatial distribution and competition relationship of the roots of different crops in the intercropping system in real time through multi-scale perception technology, establish a root interaction growth prediction model based on the transfer learning algorithm, and generate an optimal water distribution strategy by using multi-agent reinforcement learning to achieve precision irrigation in the intercropping system.

[0007] The present invention proposes an intelligent water distribution system for the spatio-temporal competition of roots in an intercropping system, including:

[0008] A ground penetrating radar array module for scanning and acquiring three-dimensional spatial distribution data of the roots of different crops in an intercropping system;

[0009] A multi-spectral laser detection module for monitoring the root activity and nutrient content of crops through laser-induced fluorescence technology;

[0010] A root competition model module, connected to the ground penetrating radar array module and the multi-spectral laser detection module, for constructing a root interaction growth prediction model based on a transfer learning algorithm;

[0011] An intelligent water distribution decision-making module, connected to the root competition model module, for generating an optimal water distribution strategy based on multi-agent reinforcement learning;

[0012] A deformable drip irrigation execution module, connected to the intelligent water distribution decision-making module, for precisely irrigating the intercropping system according to the water distribution strategy.

[0013] Preferably, the ground penetrating radar array module includes:

[0014] Multiple ground penetrating radar units, interconnected through a wireless data communication module, for collaborative scanning in different regions;

[0015] A three-dimensional reconstruction processing unit for reconstructing the three-dimensional structure of the roots of different crops based on the scanning data;

[0016] A time-series data acquisition unit for continuously scanning and acquiring dynamic change data of root growth.

[0017] Preferably, the multi-spectral laser detection module includes:

[0018] A main crop root laser probe, using multi-spectral laser fluorescence technology, for extracting main crop root biomass data;

[0019] A cover crop root laser probe, using hyperspectral laser fluorescence technology, for extracting cover crop root biomass data;

[0020] An optoelectronic conversion unit for converting the detected fluorescence signal into root activity data.

[0021] Preferably, the root competition model module includes:

[0022] A feature extraction unit for extracting key features of root spatial distribution through a double convolutional neural network;

[0023] A time-series recalibration unit for recalibrating the feature data according to the crop growth cycle;

[0024] A transfer learning unit for constructing a root interaction growth prediction model that can be transferred between different intercropping systems.

[0025] Preferably, the intelligent water distribution decision-making module includes:

[0026] A competition evaluation unit, configured to calculate the volume of the spatial intersection of the roots of the main crop and the cover crop as a competition evaluation index;

[0027] A multi-agent reinforcement learning unit, configured to generate a water distribution strategy based on the competition evaluation index;

[0028] A strategy adjustment unit, configured to dynamically optimize the strategy in each water distribution cycle.

[0029] Preferably, the deformable drip irrigation execution module includes:

[0030] A deformable multi-functional drip irrigation pipe, capable of adjusting the position and drip irrigation direction;

[0031] An intelligent drip head system, capable of adjusting fertilization according to the nutrient content;

[0032] A partition irrigation control unit, configured to achieve independent irrigation control for different root regions.

[0033] Preferably, the transfer learning unit adopts a high-order ensemble learning framework, including:

[0034] A binary decision tree generation unit, configured to screen decision trees based on the principle of minimizing the cross-entropy of positive and negative pairs;

[0035] A low-order ensemble learning machine, configured to combine binary decision trees;

[0036] A high-order ensemble learning machine, configured to integrate the prediction results of multiple low-order ensemble learning machines.

[0037] Preferably, the multi-agent reinforcement learning unit works based on a hierarchical analysis model, including:

[0038] A biomass enrichment layer division sub-unit, configured to divide the biomass enrichment layer according to the root biomass distribution;

[0039] A water distribution strategy optimization sub-unit, configured to optimize the water distribution strategy according to the biomass enrichment layer information;

[0040] A reward signal calculation sub-unit, configured to convert the root competition evaluation index into a reinforcement learning reward signal.

[0041] Preferably, the system further includes:

[0042] A data management module, configured to store and manage the data collected by each module;

[0043] A self-learning control module, configured to continuously optimize the system parameters according to historical data and water distribution effects;

[0044] The water stress warning module is used to monitor and warn the water stress status of crops.

[0045] An intelligent water distribution method for the spatio-temporal competition of roots in an intercropping system includes the following steps:

[0046] Obtain the dataset of the spatial distribution of roots in the intercropping system for the current growth cycle;

[0047] Verify the spatio-temporal competition simulation model of the intercropping roots for the previous growth cycle;

[0048] If the model verification is passed, predict the interactive growth of roots for the next growth cycle based on the verified model;

[0049] If the model verification fails, reconstruct the spatio-temporal competition model of the intercropping system roots based on deep learning and transfer learning;

[0050] Use deep learning algorithms to extract the characteristics of the intercropping plots;

[0051] Use transfer learning algorithms to establish a spatio-temporal competition prediction model for intercropping roots;

[0052] Precisely distribute water to the intercropping system using a deformable drip irrigation system based on multi-agent reinforcement learning;

[0053] Monitor the water status of crops and determine whether the growth termination condition is reached;

[0054] If the termination condition is not reached, return to the first step and continue to loop.

[0055] Through the construction of an intelligent water distribution system with multi-module collaborative work, the present invention realizes the precise monitoring, prediction, and response to the spatio-temporal competition relationship of the intercropping system roots, and has the following beneficial effects:

[0056] 1. Greatly improve the water resource utilization efficiency, conduct targeted irrigation according to the spatial distribution and competition relationship of the roots of different crops, avoid water resource waste, and it can be measured that the irrigation water consumption can be reduced by 20%-35%;

[0057] 2. Balance the growth requirements among intercropping crops, relieve root competition through precise water distribution, improve the overall yield and stability of the intercropping system, and the crop yield can be increased by 15%-25%;

[0058] 3. Enhance the environmental adaptability of the system, continuously optimize the model and strategy through transfer learning and reinforcement learning, enable the system to adapt to different soils, climates, and crop combinations, and has a wide range of applications;

[0059] 4. Realize the intelligent management of the entire irrigation process, form a closed-loop system from data collection, model prediction to strategy execution, reduce manual intervention, and improve management efficiency. Description of the Drawings

[0060] Figure 1 Schematic diagram of the overall architecture of the intelligent water distribution system for the spatio-temporal competition of root systems in the intercropping system according to an embodiment of the present invention;

[0061] Figure 2 Schematic diagram of the structure of the ground penetrating radar array module according to an embodiment of the present invention;

[0062] Figure 3 Schematic diagram of the working principle of the multi-spectral laser detection module according to an embodiment of the present invention;

[0063] Figure 4 Data processing flow chart of the root system competition model module according to an embodiment of the present invention;

[0064] Figure 5 Decision flow chart of the intelligent water distribution decision module according to an embodiment of the present invention;

[0065] Figure 6 Schematic diagram of the structure of the deformable drip irrigation execution module according to an embodiment of the present invention;

[0066] Figure 7 Schematic diagram of the high-order ensemble learner framework of the transfer learning unit according to an embodiment of the present invention;

[0067] Figure 8 Schematic diagram of the architecture of the multi-agent reinforcement learning unit according to an embodiment of the present invention;

[0068] Figure 9 Flow chart of the intelligent water distribution method for the spatio-temporal competition of root systems in the intercropping system according to an embodiment of the present invention. Detailed implementation manners

[0069] Please refer to the attached Figures 1-9 , and the present invention will be further described in detail below with reference to the drawings and embodiments.

[0070] Embodiment 1

[0071] As Figure 1 shown, the intelligent water distribution system for the spatio-temporal competition of root systems in the intercropping system provided by the present invention includes: a ground penetrating radar array module 1, a multi-spectral laser detection module 2, a root system competition model module 3, an intelligent water distribution decision module 4, and a deformable drip irrigation execution module 5.

[0072] The ground penetrating radar array module 1 is used to scan and obtain the three-dimensional spatial distribution data of different crop root systems in the intercropping system. Preferably, as Figure 2 shown, the ground penetrating radar array module 1 includes a plurality of ground penetrating radar units 11, a three-dimensional reconstruction processing unit 12, and a time series data acquisition unit 13.

[0073] Multiple ground penetrating radar units 11 are interconnected through a wireless data communication module for collaborative scanning in sub - regions. In a specific embodiment of the present invention, the ground penetrating radar unit 11 can adopt a ground penetrating radar with a frequency range of 1 GHz - 2.5 GHz, which has good soil penetration ability and spatial resolution. For different soil types, the radar frequency can be dynamically adjusted. For example, in sandy soil, a higher frequency (about 2 GHz) can be adopted to obtain higher resolution; while in clay soil, a lower frequency (about 1.2 GHz) can be adopted to enhance the penetration ability.

[0074] The 3D reconstruction processing unit 12 is used to reconstruct the 3D structure of different crop roots according to the scanning data. The present invention adopts a voxel reconstruction algorithm for 3D structure reconstruction. This algorithm first converts the radar reflection signal into 3D point cloud data, and then extracts the root structure through voxel filtering and density clustering algorithms. The voxel reconstruction algorithm can be expressed by the following formula:

[0075]

[0076] where V(x,y,z) represents the voxel value at the position (x,y,z), S i (x,y,z) represents the signal intensity detected by the i - th ground penetrating radar unit at this position, w i is the weight coefficient, and n is the number of ground penetrating radar units. The weight coefficient w i can be dynamically adjusted according to the position and signal quality of the ground penetrating radar unit 11. For example, for a radar unit with a higher signal - to - noise ratio, a larger weight (usually in the range of 0.6 - 0.8) can be given.

[0077] The time - series data acquisition unit 13 is used to continuously scan to obtain the dynamic change data of root growth. In practical applications, the system will conduct a comprehensive scan at a fixed time interval (such as every 12 hours) to generate time - series data of the root 3D structure. These time - series data are crucial for constructing a dynamic model of root growth.

[0078] The multi - spectral laser detection module 2 is used to monitor the root activity and nutrient content of crops through laser - induced fluorescence technology. As Figure 3 shown, the multi - spectral laser detection module 2 includes a main - crop root laser probe 21, a cover - crop root laser probe 22, and a photoelectric conversion unit 23.

[0079] The main - crop root laser probe 21 adopts multi - spectral laser fluorescence technology to extract the biomass data of the main - crop roots. In the embodiment of the present invention, the main - crop root laser probe 21 adopts multi - spectral laser with a wavelength range of 450 nm - 550 nm, which can effectively excite pigments such as chlorophyll and carotenoids in the roots, and judge the root activity according to the characteristics of the reflected fluorescence spectrum.

[0080] The laser probe 22 for cover crop roots uses hyperspectral laser fluorescence technology to extract biomass data of cover crop roots. The laser probe 22 for cover crop roots has a wider wavelength range (450nm - 660nm), which can capture specific fluorescence signals of different cover crops.

[0081] The photoelectric conversion unit 23 is used to convert the detected fluorescence signal into root activity data. The photoelectric conversion unit 23 includes a high - sensitivity photodiode array, which can detect weak fluorescence signals and convert them into electrical signals. The photoelectric conversion efficiency is usually between 70% - 85%. According to different crop types and root conditions, the system will automatically adjust the gain parameter (typical range is 5 - 20dB).

[0082] The root competition model module 3 is connected to the ground - penetrating radar array module 1 and the multi - spectral laser detection module 2, and is used to construct a root interaction growth prediction model based on the transfer learning algorithm. As Figure 4 shown, the root competition model module 3 includes a feature extraction unit 31, a temporal recalibration unit 32, and a transfer learning unit 33.

[0083] The feature extraction unit 31 is used to extract key features of the root spatial distribution through a dual convolutional neural network. The dual convolutional neural network includes two parallel convolutional network branches, which process spatial distribution data and temporal change data respectively, and then integrate the two parts of features through a feature fusion layer. The convolutional kernel size is usually 3×3×3 (for three - dimensional data). The first - layer convolutional network uses 32 convolutional kernels, the second layer uses 64 convolutional kernels, and the activation function uses ReLU. The feature extraction process can be expressed by the following formula:

[0084] F = σ(W2·σ(W1b + b1)+b2)

[0085] Among them, F represents the extracted feature, X represents the input data, W1 and W2 represent the weight matrices of the first - layer and second - layer convolutional kernels respectively, b1 and b2 represent the bias terms, and σ represents the ReLU activation function.

[0086] The temporal recalibration unit 32 is used to recalibrate the feature data according to the crop growth cycle. The core of temporal recalibration is to establish a time mapping relationship and map the feature data at different growth stages to a standard time scale. The recalibration process can be expressed by the following formula:

[0087] F s (t)=M(F(t),g(t))

[0088] Among them, F s(t) represents the recalibrated features, F(t) represents the original features, g(t) represents the growth stage function, and M represents the mapping function. The growth stage function g(t) is defined according to the physiological characteristics of the crop. For example, for corn, the growth cycle can be divided into six stages: emergence stage, seedling stage, jointing stage, heading stage, filling stage, and maturity stage. Appropriate weights are assigned to each stage on the standard time scale, such as 0.1 for the emergence stage, 0.15 for the seedling stage, 0.2 for the jointing stage, 0.25 for the heading stage, 0.2 for the filling stage, and 0.1 for the maturity stage.

[0089] The transfer learning unit 33 is used to construct a root interaction growth prediction model that can be transferred between different intercropping systems. The transfer learning unit 33 adopts domain adaptation technology to transfer the knowledge of the source domain (known intercropping system) to the target domain (new intercropping system). The domain adaptation process is achieved by minimizing the difference in feature distributions between the source domain and the target domain, which can be expressed by the following formula:

[0090] L transfer = L task + λ·d(P s , P t )

[0091] Among them, L transfer represents the total loss function of transfer learning, L task represents the loss function of the prediction task (usually mean square error), d(P s , P t ) represents the distance metric between the source domain feature distribution P s and the target domain feature distribution P t (usually maximum mean discrepancy MMD), and λ is the weight coefficient used to balance the relationship between task learning and domain adaptation, and the typical value range is 0.1 - 1.0.

[0092] The intelligent water distribution decision-making module 4 is connected to the root competition model module 3 and is used to generate an optimal water distribution strategy based on multi-agent reinforcement learning. As Figure 5 shown, the intelligent water distribution decision-making module 4 includes a competition evaluation unit 41, a multi-agent reinforcement learning unit 42, and a policy adjustment unit 43.

[0093] The competition evaluation unit 41 is used to calculate the volume of the spatial intersection of the roots of the main crop and the cover crop as the competition evaluation index. The competition evaluation index C can be calculated by the following formula:

[0094] C = ∫(R m ∩R c )dV

[0095] Among them, R m represents the spatial distribution of the roots of the main crop, R c represents the spatial distribution of the roots of the cover crop, Rm ∩R c represents the spatial intersection of the two, and ∫(·)dV represents the integral of the volume. In the actual calculation, the system discretizes the three-dimensional space into voxel grids, and then calculates the number of voxels that contain the roots of the main crop and the cover crop at the same time as the approximate value of the competition evaluation index.

[0096] The multi-agent reinforcement learning unit 42 is used to generate a water distribution strategy based on the competitive evaluation index. The present invention adopts a two-layer multi-agent reinforcement learning architecture, including a main crop agent and a cover crop agent, and each agent is responsible for optimizing the water distribution strategy of the corresponding crop. Reinforcement learning is based on a state-action-reward mechanism, where the state includes information such as root distribution, soil moisture, and crop growth stage, and the action represents the irrigation amount in different areas. The reward function is designed based on water resource utilization efficiency and competitive evaluation indicators. The core of reinforcement learning is the Q learning algorithm, and the update rules are as follows:

[0097]

[0098] Among them, Q(s t ,a t ) indicates state s t Next, perform action a t The value function, r t represents the reward obtained, α represents the learning rate (usually set between 0.01 and 0.1), γ represents the discount factor (usually set to around 0.9), and max a Q(s t+1 ,a) represents the maximum action value of the next state.

[0099] The strategy adjustment unit 43 is used to dynamically optimize the strategy in each water distribution cycle. The strategy adjustment is based on historical irrigation effects and real-time environmental changes, and adopts an incremental strategy improvement algorithm. On the basis of retaining the original strategy main framework, the algorithm performs local optimization for areas with poor effects to improve the adaptability and stability of the system.

[0100] The deformable drip irrigation execution module 5 is connected to the intelligent water distribution decision module 4 to accurately irrigate the intercropping system according to the water distribution strategy. Figure 6 As shown, the deformable drip irrigation execution module 5 includes a deformable multifunctional drip irrigation pipe 51 , an intelligent dripper system 52 and a zoned irrigation control unit 53 .

[0101] The deformable multifunctional drip irrigation pipe 51 can adjust the position and drip irrigation direction. In the embodiment of the present invention, the deformable multifunctional drip irrigation pipe 51 is made of a highly elastic polymer material, and has a built-in micro motor drive device, which can adjust the bending angle and direction according to the control signal. The adjustment range is generally ±30° in the horizontal direction and ±15° in the vertical direction, which can cover most of the root distribution area.

[0102] The intelligent drip head system 52 can adjust fertilization according to the nutrient content. The intelligent drip head system 52 integrates a micro valve and a flow sensor, and can accurately control the output of water and fertilizer. According to the needs of different crops at different growth stages, the system will automatically adjust the N / Li nutrient content. For example, for corn, a higher nitrogen content (usually 15-20 mg / L) is required during the jointing stage, while a higher potassium content (usually 12-18 mg / L) is required during the filling stage.

[0103] The zoned irrigation control unit 53 is used to achieve independent irrigation control for different root areas. The zoned irrigation control is based on the spatial distribution of root competition, divides the irrigation area into multiple functional zones, and each functional zone can independently control the irrigation amount and irrigation time. In practical applications, the system usually divides the irrigation area into 3-5 main functional zones, such as the dominant area of the main crop, the dominant area of the cover crop, and the highly competitive area, and adopts different irrigation strategies for different functional zones.

[0104] Embodiment 2

[0105] As Figure 7 shown, in another embodiment of the present invention, the transfer learning unit 33 adopts a high-order ensemble learning framework, including a binary decision tree generation unit 331, a low-order ensemble learning machine 332, and a high-order ensemble learning machine 333.

[0106] The binary decision tree generation unit 331 is used to screen decision trees based on the principle of minimizing the cross-entropy of positive pairs and negative pairs. Positive pairs represent the positive effects of root interactive growth, and negative pairs represent the negative effects of root competition. The principle of minimizing cross-entropy can be expressed by the following formula:

[0107]

[0108] where L CE represents the cross-entropy loss, P represents the number of positive pair samples, N represents the number of negative pair samples, p i represents the prediction probability of the i-th positive pair sample, and n j represents the prediction probability of the j-th negative pair sample. The system will select the combination of decision trees that minimizes the cross-entropy loss, and usually screens out 10-15 basic decision trees with better performance.

[0109] The low-order ensemble learning machine 332 is used to combine binary decision trees. The low-order ensemble adopts the random forest algorithm and combines the screened decision trees with weights. The weight assignment is based on the performance of each decision tree, and the better the performance of the decision tree, the greater the weight it obtains. The output of the low-order ensemble learning machine can be expressed as:

[0110]

[0111] Among them, F low (x) represents the output of the low-order ensemble learner, and T i (x) represents the prediction result of the i-th decision tree, and w i represents the corresponding weight, and m represents the number of decision trees. The weight w i is usually determined by cross-validation and ranges between 0.05 - 0.2.

[0112] The high-order ensemble learner 333 is used to integrate the prediction results of multiple low-order ensemble learners. The high-order ensemble adopts the Stacking algorithm and uses a meta-learner (usually logistic regression or support vector machine) to integrate the prediction results of the low-order learners. The output of the high-order ensemble learner can be expressed as:

[0113]

[0114] Among them, F high (x) represents the output of the high-order ensemble learner, and F l ow i (x) represents the output of the i-th low-order ensemble learner, M represents the meta-learner, and n represents the number of low-order ensemble learners. In practical applications, the system usually uses 3 - 5 low-order ensemble learners to balance the prediction performance and computational complexity.

[0115] Embodiment III

[0116] As Figure 8 shown, in another embodiment of the present invention, the multi-agent reinforcement learning unit 42 works based on a hierarchical analysis model, including a biomass enrichment layer division subunit 421, a water distribution strategy optimization subunit 422, and a reward signal calculation subunit 423.

[0117] The biomass enrichment layer division subunit 421 is used to divide the biomass enrichment layer according to the root biomass distribution. The division of the biomass enrichment layer is based on the root biomass density, and the root space is divided into a high-density area, a medium-density area, and a low-density area by using a clustering algorithm (such as K-means or DBSCAN). For different crops, the biomass density threshold will be different. For example, for corn, the biomass density in the high-density area is usually greater than 0.5 g / cm 3 , the medium-density area is between 0.2 - 0.5 g / cm 3 , and the low-density area is less than 0.2 g / cm 3 .

[0118] The water distribution strategy optimization subunit 422 is used to optimize the water distribution strategy according to the biomass enrichment layer information. The optimization of the water distribution strategy is based on a multi-objective optimization framework, considering three objectives: water resource utilization efficiency, crop growth requirements, and root competition mitigation. The optimization problem can be expressed as:

[0119] minf(x) = [f1(x), f2(x), f3(x)]

[0120] s.t. g i (x) ≤ 0, i = 1, 2,..., m

[0121] h j (x) = 0, j = 1, 2,..., n

[0122] Among them, f1(x) represents water resource consumption, f2(x) represents the negative value of the crop growth index, f3(x) represents the root competition index, g i (x) and h j (x) represent constraint conditions, such as the maximum irrigation amount, the minimum soil water content, etc. The system uses the NSGA-II algorithm to solve this multi-objective optimization problem, generates the Pareto optimal solution set in each water distribution cycle, and then selects the most suitable solution according to the current environmental conditions and crop growth stages.

[0123] 1. Regarding the problem of the negative value of the crop growth index

[0124] In the multi-objective optimization problem expression:

[0125] minf(x) = [f1(x), f2(x), f3(x)] f2(x) represents the negative value of the crop growth index. The reason for using the negative value is that in the multi-objective optimization framework, all objectives are usually set in the minimization form. Therefore, for the indicators that need to be maximized (such as crop growth amount, yield, etc.), by taking their negative values and then minimizing them, it is actually equivalent to maximizing the original indicators. For example, if the original crop growth index is the biomass growth rate G(x), then f2(x) = -G(x) can be set, so minimizing f2(x) is equivalent to maximizing G(x). This is the standard processing method in multi-objective optimization.

[0126] 2. Regarding the specific definition of the constraint conditions

[0127] The optimization constraint conditions "g i (x) ≤ 0" and "h j (x) = 0" have clear physical meanings in this system:

[0128] The inequality constraint g i (x)) ≤ 0 includes:

[0129] The maximum irrigation amount constraint: g1(x) = W(x) - W max ≤ 0, where W(x) is the total irrigation amount under the given water distribution strategy x, and W ma is the maximum irrigation amount allowed by the system (usually 30 - 50 mm / time)

[0130] Minimum soil water content constraint: g2(x) = θ min -θ(x) ≤ 0, where θ(x) is the expected soil water content after water allocation, and θ min is the minimum soil water content required for the safe growth of crops (usually 50% of the field capacity);

[0131] Root competition threshold constraint: g3(x) = C(x) - C max ≤ 0, where C(x) is the expected root competition index under the water allocation strategy x, and C max is the maximum allowable competition threshold;

[0132] Equality constraint h j (x) = 0 includes:

[0133] Water balance constraint: h1(x) = W in(x) -W out(x) -W soil(x) =0, where W in is the input water volume, W out is the drainage volume, W soil is the increment of soil water storage;

[0134] Regional water allocation ratio constraint: h2(x) = ∑ i R i(x) -1 = 0, where R i(x) is the proportion of the irrigation water volume allocated to the i-th region, ensuring that the total proportion is 1.

[0135] The reward signal calculation subunit 423 is used to convert the root competition evaluation index into a reward signal for reinforcement learning. The design of the reward signal comprehensively considers the water resource utilization efficiency, crop growth status, and root competition degree, and can be expressed as:

[0136] r t =w1·e water +w2·g crop -w3·C

[0137] where, r t represents the reward signal at time step t, e water represents the water resource utilization efficiency (expressed by the water saving rate), g crop represents the crop growth index (such as the biomass increase rate), C represents the root competition evaluation index, and w1, w2, and w3 are weight coefficients. In practical applications, the weight coefficients are usually adjusted according to user requirements and environmental conditions. For example, in arid regions, w1 can be increased (such as set to 0.5), in high-yield agricultural areas, w2 can be increased (such as set to 0.6), and in a highly competitive intercropping system, w3 can be increased (such as set to 0.4).

[0138] The specific physical meanings of each weight coefficient are:

[0139] w1: The weight coefficient of water resource utilization efficiency, which reflects the degree of emphasis of the system on water resource conservation. Setting a relatively high value (e.g., 0.5) in arid regions means taking water conservation as the primary goal, and the system will preferentially select more water-saving irrigation strategies, even if it may have a slight impact on yield.

[0140] w2: The weight coefficient of crop growth indicators, which reflects the degree of emphasis of the system on crop yield. Setting a relatively high value (e.g., 0.6) in high-yield agricultural areas indicates that the system is more inclined to select strategies that can significantly promote crop growth and may appropriately increase water use to obtain higher yields.

[0141] w3: The weight coefficient of root competition indicators, which reflects the degree of emphasis of the system on root competition management. Setting a relatively high value (e.g., 0.4) in an intensively competitive intercropping system means that the system will pay special attention to irrigation strategies that reduce root competition, even if it may require an increase in total water use.

[0142] The sum of these three weight coefficients is usually set to 1.0, representing the complete consideration factors for system decision-making. The specific values of the weights can be dynamically adjusted according to different environmental conditions and agricultural goals. For example, w1 can be reduced during the rainy season and increased during the dry season; w2 can be increased during the critical growth period of crops; and w3 can be increased during the stage of intensified competition in the intercropping system. The system continuously adjusts these weights through the Bayesian optimization algorithm to maximize the satisfaction of the comprehensive optimization goal with the water distribution strategy.

[0143] Example 4

[0144] The present invention also provides an intelligent water distribution method for the spatial and temporal root competition in an intercropping system, as Figure 9 shown, the method includes the following steps:

[0145] Step S1, obtaining the dataset of the spatial distribution of the roots of the intercropping system in the current growth cycle. In this step, the ground penetrating radar array module 1 and the multi-spectral laser detection module 2 are used to collect the spatial distribution and physiological activity data of the roots of different crops in the intercropping system. The data collection frequency is usually once every 12 hours, and can be increased to once every 8 hours during the rapid growth period of crops (such as the jointing stage).

[0146] Step S2, verifying the simulation model of the spatial and temporal root competition of the intercropping system in the previous growth cycle. The verification process is based on the comparison between the model prediction values and the actual observation values, and multiple performance indicators are used to evaluate the accuracy of the model, including the mean absolute error (MAE), the root mean square error (RMSE), and the coefficient of determination (R 2 ). When MAE is less than 10%, RMSE is less than 15%, and R 2 is greater than 0.85, it is considered that the model verification is passed.

[0147] Step S3, if the model passes the verification, then predict the root interaction growth in the next growth cycle based on the verified model. The prediction content includes the spatial expansion trend of different crop roots, the change in biomass distribution, and possible competition hotspots. The prediction time span is usually 3 - 7 days in the future, and it is dynamically adjusted according to the crop growth rate and environmental conditions.

[0148] Step S4, if the model fails to pass the verification, then reconstruct the spatio-temporal competition model of the intercropping system roots based on deep learning and transfer learning. The model reconstruction process includes three stages: feature extraction, parameter optimization, and model verification, and usually requires 4 - 8 hours of computing time (based on a standard computing server).

[0149] Step S5, use deep learning algorithms to extract the features of the intercropping plot. The feature extraction adopts a double convolutional neural network architecture, and the extracted features include soil properties, root distribution patterns, biomass density gradients, etc. The main feature parameters include root depth (typical range is 10 - 120 cm), root density (typical range is 0.1 - 1.0 g / cm 3 ), soil water content (typical range is 15% - 35%), etc.

[0150] Step S6, use transfer learning algorithms to establish a spatio-temporal competition prediction model for intercropping roots. The transfer learning process includes three links: source domain selection, domain adaptation, and model fine-tuning. The source domain usually selects historical data similar to the target intercropping system, such as data under the same crop combination or similar environmental conditions. Domain adaptation is mainly achieved by minimizing the maximum mean discrepancy (MMD), and the typical MMD threshold is set to 0.05 - 0.1.

[0151] Step S7, the deformable drip irrigation system based on multi-agent reinforcement learning conducts precise water distribution for the intercropping system. The water distribution process considers various factors such as crop growth stage, root distribution, soil moisture status, and weather forecast. The system will automatically adjust the position of the drip irrigation pipe and the parameters of the drip heads according to the irrigation requirements of different regions to achieve precise irrigation. Typical irrigation parameters include the single irrigation amount (usually in the range of 15 - 40 mm according to different crops and growth stages) and the irrigation interval (usually 2 - 7 days).

[0152] Step S8, monitor the crop water status and judge whether the growth termination condition is reached. The water status monitoring is achieved through various methods, including soil moisture sensors, crop transpiration rate monitoring, and leaf temperature monitoring, etc. The growth termination condition is usually when the crop is physiologically mature or the harvest index is met. For example, for corn, when the grain water content drops to 25% - 30%, it can be considered that the harvest suitable period is reached; for wheat, when the water content drops to 14% - 16%, it can be considered mature. The system will also monitor the water stress index. When the relative water content of the leaves (RWC) is lower than 65% or the soil water potential is lower than -0.8 MPa, a water stress warning will be triggered.

[0153] Step S9, if the termination condition is not reached, return to the first step to continue the loop. The system will adjust the data acquisition frequency and water distribution strategy according to the previous round of water distribution results and crop response conditions, continuously optimizing the irrigation effect. The cycle period is usually 3 - 7 days, and can be shortened to 1 - 2 days during critical growth stages (such as the flowering stage and filling stage) to improve the response speed.

[0154] In this embodiment, the system further includes a data management module 6, a self-learning control module 7, and a water stress warning module 8.

[0155] The data management module 6 is used to store and manage the data collected by each module. The data management adopts a hierarchical storage architecture, storing the original data, processed data, and model data separately, and establishing an index relationship. The storage period varies according to the data type. For example, the root system spatial distribution data is usually saved throughout the growing season, while the real-time monitoring data may only retain recent data (such as the most recent 30 days).

[0156] The self-learning control module 7 is used to continuously optimize the system parameters according to historical data and water distribution effects. The self-learning control is based on the Bayesian optimization algorithm, and by continuously adjusting the detection parameters, model hyperparameters, and water distribution parameters, the overall performance of the system is improved. The optimization goal is a comprehensive index, including three aspects: water resource utilization efficiency, crop yield, and system stability, and the weight ratio is usually 3:4:3, which can be adjusted according to user needs.

[0157] The water stress warning module 8 is used to monitor and warn the crop water stress status. The warning mechanism is based on multi-parameter threshold judgment, including indicators such as soil water content, leaf temperature, transpiration rate, and root activity. The warning is divided into three levels: mild warning (soil water content is lower than 65% of the field capacity or leaf temperature is 2 - 3℃ higher than the surrounding environment), moderate warning (soil water content is lower than 50% of the field capacity or leaf temperature is 3 - 5℃ higher than the surrounding environment), and severe warning (soil water content is lower than 35% of the field capacity or leaf temperature is more than 5℃ higher than the surrounding environment). For different levels of warnings, the system will take corresponding countermeasures, such as adjusting the irrigation plan, increasing the irrigation frequency, or starting emergency irrigation.

[0158] In summary, the intelligent water distribution system and method for the spatio-temporal competition of the root systems in the intercropping system provided by the present invention combines multi-scale sensing technology, transfer learning algorithms, and multi-agent reinforcement learning to achieve precise monitoring, prediction, and response to the root system competition relationship in the intercropping system, greatly improving water resource utilization efficiency, balancing the growth requirements between intercropped crops, enhancing the environmental adaptability of the system, and realizing intelligent management of the entire irrigation process. This system is particularly suitable for precise irrigation of complex intercropping systems, and is of great significance for promoting the development of precision agriculture, improving agricultural water resource utilization efficiency, and promoting sustainable agriculture.

[0159] It should be noted that the above are only the preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. An intelligent water distribution system for the spatio-temporal competition of root systems in an intercropping system, characterized in that, include: Ground penetrating radar array module, used to scan and obtain three-dimensional spatial distribution data of the roots of different crops in the intercropping system; Multispectral laser detection module for monitoring crop root activity and nutrient content using laser-induced fluorescence technology; A root competition model module, connected to the ground penetrating radar array module and the multi-spectral laser detection module, for building a root interaction growth prediction model based on a transfer learning algorithm; An intelligent water distribution decision module, connected to the root competition model module, for generating an optimal water distribution strategy based on multi-agent reinforcement learning; The deformable drip irrigation execution module is connected to the intelligent water distribution decision module and is used to accurately irrigate the intercropping system according to the water distribution strategy.

2. The system according to claim 1, characterized in that, The ground penetrating radar array module comprises: Multiple ground penetrating radar units are interconnected through wireless data communication modules for coordinated scanning in different areas; A three-dimensional reconstruction processing unit, used to reconstruct the three-dimensional structure of the root system of different crops based on the scan data; The time series data acquisition unit is used to continuously scan and obtain the dynamic change data of root growth.

3. The system according to claim 1, wherein The multi-spectral laser detection module comprises: The main crop root laser probe uses multi-spectral laser fluorescence technology to extract the root biomass data of the main crops; Cover crop root laser probe, which uses hyperspectral laser fluorescence technology to extract cover crop root biomass data; The photoelectric conversion unit is used to convert the detected fluorescence signal into root activity data.

4. The system according to claim 1, wherein The root competition model module includes: A feature extraction unit, used to extract key features of the spatial distribution of the root system through a dual convolutional neural network; A time-series recalibration unit, used to recalibrate characteristic data according to the crop growth cycle; Transfer learning unit, used to build a root interaction growth prediction model that can be transferred between different intercropping systems.

5. The system according to claim 1, characterized in that, The intelligent water distribution decision module includes: A competition evaluation unit is used to calculate the volume of the spatial intersection of the root systems of the main crop and the cover crop as a competition evaluation index; A multi-agent reinforcement learning unit for generating water allocation strategies based on competitive evaluation metrics; The strategy adjustment unit is used to dynamically optimize the strategy in each water distribution cycle.

6. The system according to claim 1, wherein The deformable drip irrigation execution module comprises: Deformable multifunctional drip irrigation pipe, capable of adjusting position and drip irrigation direction; Intelligent dripper system that can adjust fertilization according to nutrient content; Zone irrigation control unit is used to achieve independent irrigation control of different root zones.

7. The system according to claim 4, characterized in that, The transfer learning unit adopts a high-order integrated learner framework, including: A binary decision tree generation unit, used to select decision trees based on the principle of minimizing the cross entropy of positive and negative pairs; Low-order ensemble learners for combining binary decision trees; A high-order ensemble learner is used to integrate the prediction results of multiple low-order ensemble learners.

8. The system according to claim 5, wherein The multi-agent reinforcement learning unit works based on a hierarchical analysis model, including: A biomass enrichment layer division subunit is used to divide the biomass enrichment layer according to the root biomass distribution; A water distribution strategy optimization subunit is used to optimize the water distribution strategy according to the biomass enrichment layer information; The reward signal calculation subunit is used to convert the root competition evaluation index into a reward signal for reinforcement learning.

9. The system according to claim 1, wherein The system further comprises: Data management module, used to store and manage the data collected by each module; A self-learning control module for continuously optimizing system parameters according to historical data and water distribution effects; A water stress warning module for monitoring and warning the water stress status of crops.

10. An intelligent water distribution method for the spatio-temporal competition of roots in an intercropping system, using the system according to any one of claims 1-9, comprising the following steps: Obtain the spatial distribution data set of the roots of the intercropping system in the current growth period; Verify the spatio-temporal competition simulation model of the intercropping roots in the previous growth period; If the model verification is passed, predict the interactive growth of roots in the next growth period based on the verified model; If the model verification fails, reconstruct the spatio-temporal competition model of the intercropping system roots based on deep learning and transfer learning; Use a deep learning algorithm to extract the characteristics of the intercropping plot; Use a transfer learning algorithm to establish a spatio-temporal competition prediction model for intercropping roots; Precisely distribute water to the intercropping system using a deformable drip irrigation system based on multi-agent reinforcement learning; Monitor the water status of the crops and determine whether the growth termination condition is reached; If the termination condition is not reached, return to the first step and continue the loop.

Images