Salinized farmland shelterbelt configuration optimization system and method

Abstract

The present invention relates to the field of farmland ecological protection, and in particular to a system and method for optimizing the configuration of shelterbelts for salinized farmland. The system includes a salinization monitoring module, an edge computing module, a database, a three-dimensional model module, a model optimization module, and a configuration scheme optimization module. The salinization monitoring module collects data through an unmanned aerial vehicle and a hyperspectral imaging spectrometer to determine the soil salinity distribution status. The edge computing module receives the data and calculates a shelterbelt configuration model. The database stores the calculation model, soil salinity, and vegetation growth status data. The three-dimensional model module generates a three-dimensional spatiotemporal model of forest growth. The model optimization module uses deep reinforcement learning to optimize the soil salinity and groundwater level simulation model to improve the shelterbelt configuration effect. The configuration scheme optimization module generates an optimal shelterbelt configuration plan, accurately characterizes complex interaction processes, and provides an efficient and accurate technical means for the ecological protection of salinized farmland.

Description

Technical Field

[0001] The present invention relates to the field of farmland ecological protection, and in particular to a system and method for optimizing the configuration of salinized farmland shelterbelts. Deep reinforcement learning technology is used to achieve scientific configuration of salinized farmland shelterbelts, thereby improving the utilization efficiency of saline-alkali land and the ecological benefits of shelterbelts. Background Art

[0002] Salinization is a global soil degradation problem, severely impacting agricultural production and the ecological environment. According to statistics, salinized land in my country accounts for approximately 4.88% of its total land area, primarily distributed in arid and semi-arid regions of northern China and coastal areas. Shelterbelts, as a key biological measure for improving saline soils, play an irreplaceable role in carbon sequestration, reducing emissions, conserving soil and water, and preventing wind and sand from blowing.

[0003] Traditional shelterbelt configuration methods rely primarily on empirical judgment and simple indicator evaluation, lacking accurate modeling of salt migration and tree water absorption processes, and are unable to adapt to complex and changing environmental conditions. With the development of technologies such as the Internet of Things and artificial intelligence, data-driven intelligent shelterbelt configuration methods have become a research hotspot. However, existing intelligent optimization systems still have the following problems: first, they inadequately represent the complex interactions between salt migration and tree water absorption; second, they lack adaptive optimization mechanisms that adapt to dynamic environmental changes; and third, complex models are difficult to run efficiently on resource-constrained edge devices.

[0004] Therefore, there is an urgent need for a shelterbelt configuration optimization system that can accurately model the salt-forest interaction, support dynamic optimization configuration, and operate efficiently on edge devices, so as to improve the ecological restoration efficiency and economic benefits of salinized farmland. Summary of the Invention

[0005] The purpose of the present invention is to provide a system and method for optimizing the configuration of salinized farmland shelterbelts, so as to overcome the shortcomings of the existing technology and realize the scientific configuration and dynamic optimization of salinized farmland shelterbelts.

[0006] Specifically, the objectives of the present invention include: establishing a salt migration-forest water absorption coupling model to accurately characterize the complex interactive relationship between the two; constructing a shelterbelt configuration optimization framework based on deep reinforcement learning to achieve dynamic adaptive optimization of configuration schemes; and developing a lightweight model suitable for edge devices to support real-time decision-making on site.

[0007] The present invention proposes a system for optimizing the configuration of salinized farmland shelterbelts, comprising:

[0008] The salinization monitoring module is used to collect ground reflectance and vegetation leaf reflectance through drones, and use a hyperspectral imaging spectrometer to invert soil electrical conductivity and determine the distribution of soil salt;

[0009] an edge computing module, in communication with the salinization monitoring module, for receiving the soil salt distribution data and calculating a shelterbelt configuration model;

[0010] A database, connected to the edge computing module, for storing computing models, soil salinity, and vegetation growth status data;

[0011] a three-dimensional model module, connected to the database and the edge computing module, for generating a three-dimensional spatiotemporal model of forest growth in a three-dimensional environment using soil salinity and groundwater level data;

[0012] A model optimization module, connected to the three-dimensional model module and the database, is used to optimize the soil salinity and groundwater level simulation models based on deep reinforcement learning to improve the configuration effect of the shelterbelt;

[0013] A configuration scheme optimization module is connected to the model optimization module and is used to generate an optimal shelterbelt configuration scheme based on the deep reinforcement learning optimization model.

[0014] Preferably, it also includes a meteorological monitoring module, which is connected to the model optimization module and is used to obtain meteorological condition data such as air temperature, precipitation, wind direction and wind speed to determine the optimal growth strategy; wherein, the model optimization module uses the trained deep reinforcement learning model to further determine the optimal shelterbelt configuration plan, specifically: taking the row spacing of trees to be planted and environmental parameters as states, and the shelterbelt configuration efficiency and reduction of irrigation water consumption as reward functions, the optimal shelterbelt configuration plan is determined according to the trained deep reinforcement learning model.

[0015] Preferably, the model optimization module establishes a salt migration-forest water absorption coupling model through an improved Transformer architecture, and the salt migration-forest water absorption coupling model includes:

[0016] The salt balance equation is used to describe the dynamic changes of salt in the soil;

[0017] Plant absorption equation, used to describe the absorption process of soil water and salt by plants;

[0018] The stand water holding capacity equation is used to describe the stand's ability to intercept and retain water;

[0019] The tree water absorption function is used to describe the water absorption capacity and efficiency of trees under different conditions;

[0020] Among them, the model optimization module uses the forest growth parameters and the salt content distribution output by the salt migration model as training data, constructs a shelterbelt configuration state set and a shelterbelt structure optimization model, uses the current state as input, obtains the next state through the shelterbelt configuration state set, and calculates the corresponding shelterbelt configuration efficiency and irrigation water reduction in the next state as a reward function, and generates the optimal shelterbelt configuration plan through a deep reinforcement learning model.

[0021] Preferably, the shelterbelt configuration state set includes: a state set, a decision set, a reward value and a planting area change range; wherein, the state set includes planting area, plant spacing, row spacing, groundwater level, soil saturated water content and soil salt distribution parameters.

[0022] Preferably, the trained deep reinforcement learning model includes a control function and a loss function; wherein, the control function is used to generate decisions based on the current state, the loss function is used to evaluate the gap between the predicted value and the true value, the model uses a discount factor to balance short-term and long-term rewards, and the probability distribution of the shelterbelt configuration state is used to guide the reinforcement learning process.

[0023] Preferably, the updating of the shelterbelt configuration state set includes: evaluating the current shelterbelt configuration through a deep reinforcement learning model, inputting the soil salinity and groundwater level into the coupling model, outputting the soil salinity and groundwater level in the next state through the coupling model, and repeating the iterative update until the optimal configuration plan is obtained.

[0024] Preferably, the parameter optimization in the salt migration-forest water absorption coupling model is achieved by the gradient descent method, specifically: the salt content is equally spaced into multiple salinity levels by linear interpolation, and the salt balance equation and the plant water absorption equation are solved to obtain the resistance of vegetation under different salt migration processes, and then the trained shelterbelt configuration state set is obtained.

[0025] Preferably, the constraints of the gradient descent method include:

[0026] Unit equivalent resistance model state parameter constraints;

[0027] Infiltration parameter constraints;

[0028] Upper and lower boundary constraints on soil salinity and its saturation;

[0029] Soil water capacity constraint;

[0030] The constraints mentioned above ensure the stability of the model training process and the effectiveness of the results.

[0031] Preferably, the edge computing module compresses the prediction model to a scale suitable for edge devices through knowledge distillation technology, and realizes real-time computing through a distributed edge network; the knowledge distillation technology reduces the dimension of the feature graph channel, retains the key decision-making ability of the model, and enables the system to update the shelterbelt configuration plan at the hourly level in a resource-constrained environment.

[0032] The optimization method of the salinized farmland protection forest configuration optimization system includes the following steps:

[0033] Data from the salinization monitoring area was acquired through the drone flight control system. A high-resolution multispectral camera was used to obtain ground reflectance and vegetation leaf reflectance, and the relationship between vegetation leaf reflectance, water absorption index, and electrical conductivity was established.

[0034] Soil salinity monitors and ground water level monitors were deployed in the monitoring area to collect data on salinity and ground water levels, obtain environmental meteorological information, and establish datasets of soil saturated water content and soil saturated salinity.

[0035] The dataset was used to train a coupled model of soil bulk density and salt transport-tree water absorption, and a state update function and shelterbelt configuration set were obtained based on deep reinforcement learning.

[0036] In the set of shelterbelt configuration states, the optimal adjustment parameters and control strategies are used to determine the optimal shelterbelt configuration;

[0037] The target farmland is photographed by drones to obtain the salt status distribution, the shelterbelt configuration is dynamically updated according to the current soil salinity and groundwater level, and a prediction model is used to generate a spatiotemporal model of forest growth in a three-dimensional environment.

[0038] The present invention has the following beneficial effects:

[0039] 1. A salt migration-forest water absorption coupling model was established through an improved Transformer architecture, which accurately characterized the complex interactive process of salt migration and forest water absorption. Compared with traditional methods, the salt control efficiency was improved by about 41%.

[0040] 2. A deep reinforcement learning framework was used to construct a shelterbelt configuration optimization decision-making system, with tree spacing and environmental parameters as states, and configuration efficiency and reduction in irrigation water consumption as reward functions. Compared with traditional methods, irrigation water consumption was reduced by approximately 33%.

[0041] 3. Using knowledge distillation technology to compress and deploy complex models to edge devices, we achieved hourly updates of shelterbelt configuration plans, significantly improving the practicality and response speed of the system.

[0042] 4. A complete closed-loop optimization system of monitoring-analysis-decision-making-execution-feedback has been established, which has realized intelligent management of the entire life cycle of the configuration of salinized farmland shelterbelts and provided technical support for the sustainable use of salinized land. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is a schematic diagram of the overall architecture of the salinized farmland protection forest configuration optimization system of the present invention;

[0044] Figure 2 This is a schematic diagram of the structure of the salt transport-forest water absorption coupling model of the present invention;

[0045] Figure 3 This is a flow chart of the deep reinforcement learning optimization framework of the present invention;

[0046] Figure 4 Schematic diagram of the edge computing and knowledge distillation implementation solution of the present invention;

[0047] Figure 5 The figure is a flow chart of the method for optimizing the configuration of salinized farmland shelterbelts according to the present invention. DETAILED DESCRIPTION

[0048] Please refer to the attached Figure 1-5 The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to illustrate and explain the present invention and are not intended to limit the present invention.

[0049] like Figure 1 As shown, the salinization farmland protection forest configuration optimization system provided by the present invention includes a salinization monitoring module 1, an edge computing module 2, a database 3, a three-dimensional model module 4, a model optimization module 5, a configuration scheme optimization module 6 and a meteorological monitoring module 7.

[0050] The salinization monitoring module 1 is used to collect ground reflectivity and vegetation leaf reflectivity through a drone, and use a hyperspectral imaging spectrometer to achieve soil conductivity inversion and determine the soil salt distribution status. In practical applications, the present invention targets typical salinized farmland in Bayannur City, Inner Mongolia, my country, uses a drone equipped with a multispectral camera, sets the flight altitude to 80-120 meters, and obtains multi-band images with a ground resolution of 10 cm. The salinization monitoring module 1 uses a support vector machine regression algorithm to establish a relationship model between vegetation leaf reflectivity and soil conductivity. Tests in severely salinized areas such as Bayannur City, Inner Mongolia, show that the inversion accuracy can reach more than 85%. The model can be expressed as:

[0051] ,

[0052] in, Soil electrical conductivity, in dS / m, characterizes the degree of soil salinization. Typical salinized farmland The value range is 4-16dS / m; is the support vector coefficient, determined by the training process; As the kernel function, this embodiment adopts radial basis function (RBF), which is expressed as

[0053] , is the kernel parameter, usually taking a value of 0.1-0.5; is the vegetation reflectance feature vector of the training sample, which contains the reflectance values ​​of five bands: red (R), green (G), blue (B), near infrared (NIR) and red edge (RE); is the vegetation reflectance feature vector of the sample to be predicted;

[0054] is the bias constant;

[0055] is the number of support vectors, usually the number of training sets In the experimental area of ​​salinized farmland in Bayannur City, Inner Mongolia, the correlation coefficient R between the soil electrical conductivity measured by this method and the laboratory analysis results reached 0.87.

[0056] The edge computing module 2 is communicatively connected to the salinization monitoring module 1, and is used to receive soil salt distribution status data and calculate the shelterbelt configuration model. Since salinized farmlands are mostly distributed in remote areas with limited network conditions and energy supply, the present invention preferably uses a low-power WiFi edge computing gateway. The processor adopts the ARMCortex-A72 architecture, with a main frequency of 1.5GHz, 4GB of memory, 32GB of storage, and power consumption controlled within 5W. In the salt-alkali land shelterbelt experimental area in Bayannur City, Inner Mongolia, the edge device is powered by solar energy and has been running stably for more than 500 days, meeting the needs of long-term operation in the field environment.

[0057] The database 3 is connected to the edge computing module 2 and is used to store calculation models, soil salinity, and vegetation growth status data. Taking into account the limitations of edge device storage space and computing resources, the present invention adopts a lightweight SQLite database system and designs three main data tables: an environmental monitoring table, a model parameter table, and a configuration scheme table. The environmental monitoring table stores monitoring data such as salinity, water level, and meteorology, and uses a timestamp index; the model parameter table stores model weights and hyperparameters and supports version management; the configuration scheme table stores historical configuration schemes and their effect evaluations. In the saline-alkali land improvement demonstration area in Bayannur City, Inner Mongolia, the database has realized incremental data updates and compressed storage. One year's monitoring data occupies only 2.8GB of space, greatly reducing storage requirements.

[0058] The three-dimensional model module 4 is connected to the database 3 and the edge computing module 2, and is used to generate a three-dimensional spatiotemporal model of forest growth in a three-dimensional environment using soil salinity and groundwater level data. In an embodiment of the present invention, the three-dimensional model module 4 is developed based on the Unity engine, and a parametric modeling method is used to construct a forest growth model. This module can convert two-dimensional salt distribution and water level data into a three-dimensional visualization scene, and intuitively display the dynamic changes in the growth status of protective forests under different salinity environments. In the salinization control project in Bayannur City, Inner Mongolia, this module successfully simulated the growth of salt-tolerant tree species such as Tamarix and Elaeagnus angustifolia under different salt gradients, providing forestry workers with an intuitive decision-making basis.

[0059] Model Optimization Module 5, connected to 3D Model Module 4 and Database 3, is used to optimize soil salinity and groundwater level simulation models using deep reinforcement learning, thereby improving the effectiveness of shelterbelt deployment. In the Bayannur City, Inner Mongolia, saline-alkali land management project, Model Optimization Module 5 automated the optimization of different tree species combinations and plant spacing configurations, increasing salt control efficiency from 25% using traditional methods to 41%, achieving a significant technological breakthrough.

[0060] Configuration Optimization Module 6, connected to Model Optimization Module 5, generates an optimal shelterbelt configuration plan based on a deep reinforcement learning-based optimization model. In practice, this module generates a complete plan encompassing tree species selection, spacing design, planting density, and spatial layout, and provides visual configuration drawings to facilitate implementation.

[0061] The present invention also includes a meteorological monitoring module 7, connected to the model optimization module 5, for acquiring meteorological data such as air temperature, precipitation, wind direction, and wind speed to determine the optimal growth strategy. In the Korqin Sandy Land Salinization Control Project in Inner Mongolia, real-time meteorological data collected by the meteorological monitoring module 7 helped the system adjust the shelterbelt irrigation strategy, increasing irrigation frequency during dry periods and reducing irrigation during rainy periods. This reduced irrigation water consumption by 33% while maintaining optimal growth.

[0062] One of the core innovations of this invention is to establish a coupled model of salt migration and forest water absorption through an improved Transformer architecture. Figure 2 As shown in the figure, the model includes four key components: salt balance equation, plant absorption equation, stand water holding rate equation and forest water absorption function, which realizes the accurate characterization of the complex interactive process of salt migration and forest water absorption.

[0063] The salt balance equation describes the dynamic changes in soil salt content, taking into account the coupled relationship between salt and water movement. In the salinized farmland environment of the arid region of Northwest China, the salt balance equation can be expressed as:

[0064] ,

[0065] in, is the volumetric water content of the soil, dimensionless, indicating the volume of water per unit volume of soil, with a typical value of 0.2-0.4 in salinized farmland; It is the salt concentration in the soil solution, measured in mg / L, reflecting the content of dissolved salts in the soil solution. Severely salinized soil can reach 5000-15000 mg / L. is time, the unit is d, which represents the time variable in the simulation process; is the soil depth in cm, which represents the vertical distance from the ground surface. The model usually considers the depth range of 0-200 cm; The salt diffusion coefficient, in cm² / d, describes the ability of salt to diffuse in the soil. It is related to soil texture and water content. It is about 0.5-2 cm² / d in clay and can reach 5-10 cm² / d in sand.

[0066] Soil moisture flux, in cm / d, represents the amount of water passing through a unit area per unit time. It is affected by rainfall, evaporation, and irrigation, and its typical value range is 0.1-5 cm / d. It is the sink term of salt absorption by plants, with the unit of mg / (L·d), which indicates the absorption rate of salt in soil solution by plants.

[0067] This equation describes the three main salt migration processes in soil: diffusion (the first term), convection (the second term), and plant uptake (the third term). This equation was applied to simulate vertical salt distribution in the saline-alkali land improvement area of ​​Bayannur City, Inner Mongolia. The average error between the predicted results and the measured data was within 10%, providing accurate salt environment information for shelterbelt deployment.

[0068] The plant absorption equation describes the process by which plants absorb soil water and salt, reflecting the impact of plant physiological characteristics on water and salt absorption. Based on monitoring data from the Tamarix chinensis shelterbelt in saline-alkali land in Bayannur, Inner Mongolia, the plant absorption equation is expressed as:

[0069] ,

[0070] in, is the water absorption rate of plants, in units of , which represents the amount of water absorbed by plant roots per unit volume of soil; is the plant absorption intensity, in units of , a basic parameter reflecting the water absorption capacity of plants, and Tamarix is ​​approximately , Elaeagnus angustifolia is about , poplar is about ; is the absorption coefficient, dimensionless, and is related to the root distribution density. It is about 0.7-0.9 for shallow-rooted trees and about 0.5-0.7 for deep-rooted trees.

[0071] The average stem diameter of plants, in cm. The stem diameter of common tree species in shelterbelts ranges from 5 to 20 cm. is the soil water potential influence function, dimensionless, expressed as ,in is the soil water potential, is the semi-absorbed water potential, is the shape parameter; is the salt concentration influence function, dimensionless, expressed as ,in is the half-inhibitory salt concentration,

[0072] is the morphological parameter.

[0073] In the saline-alkali land shelterbelt demonstration area in Bayannur City, Inner Mongolia, this equation successfully explained the differences in water absorption among different tree species (such as Tamarix chinensis, Elaeagnus angustifolia, and Fraxinus chinensis) under salt stress, providing a scientific basis for shelterbelt species selection. For example, monitoring revealed that Tamarix chinensis maintained normal water absorption capacity (only a 20% reduction) even when soil salt concentrations reached 10 g / kg, while poplar trees experienced a reduction of over 60% under the same conditions.

[0074] The stand water holding rate equation describes the stand's ability to intercept and retain water. Based on the monitoring data of the Hexi Corridor shelterbelt, it is expressed as:

[0075] ,

[0076] in, is the stand water holding capacity, dimensionless, indicating the proportion of water intercepted by the stand to the total precipitation;

[0077] The water holding capacity of the stand at saturated water content is dimensionless and represents the maximum water holding capacity of the stand. The typical value of mature shelterbelts is 0.2-0.3. is the attenuation coefficient, dimensionless, related to tree species characteristics and forest stand structure, about 0.5-0.6 for coniferous forests and about 0.6-0.7 for broad-leaved forests;

[0078] Leaf area index is a dimensionless index that represents the leaf area per unit ground area. Mature shelterbelts usually have a value of 3-5.

[0079] In a saline-alkali land improvement project in Bayannur, Inner Mongolia, the stand water holding capacity calculated using this equation was within 15% of the measured value, accurately assessing the impact of shelterbelts on water balance. For example, a mixed poplar-Tamarisk shelterbelt with a plant spacing of 3 meters and a row spacing of 6 meters, when achieving a leaf area index of 4.2, can intercept approximately 24% of precipitation, significantly reducing surface runoff and effectively controlling the horizontal migration of salt.

[0080] The tree water absorption function describes the water absorption capacity and efficiency of trees under different conditions. Combined with the experimental data of saline-alkali land improvement in Bayannur City, Inner Mongolia, it is expressed as:

[0081] ,

[0082] in,

[0083] Specific soil water potential and depth The water absorption rate at , in cm / d; is the water potential response function, dimensionless, expressed as ,in is the soil water potential (negative value), is the semi-absorbed water potential, is the shape parameter; is the root distribution function, dimensionless, expressed as ,in is the soil depth, is the maximum root depth, is the shape parameter; The maximum water absorption rate is in cm / d, which is related to the tree species and growth stage. It is about 0.5-0.8 cm / d in young forests and can reach 0.8-1.2 cm / d in mature forests.

[0084] In a saline-alkali land remediation project in Bayannur, Inner Mongolia, this function successfully simulated the distribution of root water uptake at different soil depths, revealing the mechanisms by which shelterbelts adapt to salt stress. For example, it was found that under salt stress, Tamarix chinensis increases the proportion of deep roots, decreases surface root activity, reduces water uptake in high-salinity areas, and shifts water uptake to deeper, lower-salinity soil layers. This finding provides important evidence for the rational allocation of shelterbelts.

[0085] These four equations together form the mathematical foundation of the coupled salt transport and tree water uptake model. Model Optimization Module 5 uses tree growth parameters (such as tree species, density, and growth stage) and the salt content distribution output from the salt transport model as training data to construct a shelterbelt configuration state set and a shelterbelt structure optimization model. In the Bayannur City, Inner Mongolia, saline-alkali land management demonstration area, this model successfully predicted the impact of different shelterbelt configurations on salt distribution with an accuracy exceeding 85%, significantly improving the scientific nature of shelterbelt configuration.

[0086] like Figure 3 As shown in the figure, the present invention uses a deep reinforcement learning framework to construct a shelterbelt configuration optimization decision-making system. In view of the long-term and complex nature of salinization prevention and control, the framework includes three key steps: state space construction, decision model training, and solution optimization and implementation.

[0087] First, a shelterbelt configuration state set was constructed, including a state set, a decision set, a reward value, and the range of change in planting area. The state set includes parameters such as planting area A, plant spacing, row spacing, groundwater level, soil saturated water content, and soil salt distribution S. In the actual application of saline-alkali land shelterbelts in Bayannur City, Inner Mongolia, the typical range of planting area A is 1-100 hectares, determined by the size and shape of the plot; the plant spacing range is 1-5 meters, usually 2-3 meters in severely salinized areas, and 1-2 meters in lightly salinized areas; the row spacing range is 2-10 meters, mainly considering the needs of mechanized operations and ventilation, and generally 4-6 meters is selected; the groundwater level monitoring depth is 0-5 meters, and the ideal groundwater level in salinized farmland is controlled at 1.5-2.5 meters. Too shallow can easily lead to secondary salinization, while too deep can affect plant growth; the saturated soil water content varies with different soil types, ranging from about 0.3-0.35 cm³ / cm³ for sandy soil and about 0.4-0.5 cm³ / cm³ for clay soil; the salt distribution S in the soil is represented by a three-dimensional tensor, containing salt concentration values ​​in the three spatial dimensions of x, y, and z, with a resolution of 10m×10m×0.2m.

[0088] The decision set includes various adjustment strategies for shelterbelt deployment, such as plant spacing adjustment (±0.5m steps), row spacing adjustment (±1.0m steps), and tree species selection (selected from a predefined library of salt-tolerant species). In the Korqin Sandy Land Salinization Control Project in Inner Mongolia, the decision set also includes parameters such as irrigation strategy adjustment and mixed planting ratio adjustment, further enriching the optimization dimensions.

[0089] The reward value is composed of two parts: shelterbelt allocation efficiency and irrigation water consumption reduction. The calculation formula is:

[0090] ,

[0091] in, is the total reward value; is the salt control efficiency, calculated as , and are the soil salt content before and after configuration, respectively; is the irrigation water consumption reduction rate, calculated as , and They are standard irrigation amount and actual irrigation amount respectively; and is the weight coefficient, which is adjusted according to the application scenario and can be increased in areas with water shortage. The weight of In the saline-alkali land experimental area of ​​Bayannur City, Inner Mongolia, after multiple verifications, the , The best overall effect was achieved.

[0092] The range of planted area changes limits the proportion of area that can be adjusted in a single decision, typically controlled between 5% and 20% to avoid excessive fluctuations in decision-making. In practice, a smaller range (5% to 10%) is set in the initial stage, and gradually increased to 15% to 20% as the system learns more deeply, balancing exploration and stability.

[0093] The trained deep reinforcement learning model includes a control function F and a loss function L. The control function F generates a decision based on the current state, which is expressed as:

[0094] ,

[0095] in, is the decision action at time t, including the adjustment amount of configuration parameters such as plant spacing, row spacing, and tree species; The state at time t, including the current environment and configuration parameters; To control the parameter set of the function, it is implemented through a deep neural network.

[0096] The loss function L is used to evaluate the gap between the predicted value and the true value, which is expressed as:

[0097] ,

[0098] in, is the state-action value function, which evaluates the long-term value of taking action a in state s; It is an immediate reward, determined by the salt control efficiency and water saving effect produced by the current configuration; is a discount factor used to balance short-term and long-term rewards. Considering the long-term growth cycle of shelterbelts, its typical value is set to 0.9. A higher discount factor makes the system pay more attention to long-term benefits. is the next state reached after executing action a; Indicates the next state The maximum function value of all possible actions; is the target network parameter, which is used to stabilize the training process.

[0099] In the Bayannur City, Inner Mongolia, saline-alkali land shelterbelt optimization project, the model employed a two-layer neural network structure with 256 neurons in the first hidden layer and 128 neurons in the second hidden layer. Reinforced Luminance (ReLU) activation function was used to prevent the vanishing gradient problem. Adam was used as the optimizer, with an initial learning rate of 0.001 and a learning rate decay strategy, multiplying the learning rate by 0.9 after every 5,000 training steps. During training, the experience replay buffer size was set to 10,000, the batch size to 64, and the target network update frequency to every 100 steps. These hyperparameter settings, after multiple experimental comparisons, demonstrated optimal performance in the saline-alkali land shelterbelt deployment scenario.

[0100] Updating the shelterbelt configuration state set involves evaluating the current shelterbelt configuration using a deep reinforcement learning model, inputting soil salinity and groundwater levels into a coupled model, and then outputting the soil salinity and groundwater levels for the next state. This iterative update process continues until the optimal configuration is achieved. In an application in a heavily salinized area of ​​Bayannur, Inner Mongolia, the iteration termination criteria were set to either a reward function value change of less than 1% for five consecutive times or a maximum number of 1,000 iterations. Practice has shown that in most cases, the system converges to a stable solution within 300-500 iterations.

[0101] In this study, the parameters of the salt transport-tree water absorption coupling model were optimized using a gradient descent method. Based on the characteristics of salinization in Northwest China, the salt content was divided into 15 equally spaced salinity levels (from 0.5 g / kg to 15 g / kg, with intervals of 1 g / kg) using linear interpolation. The salt balance equation and plant water absorption equation were then solved to determine the vegetation's resilience under different salt transport processes, thereby generating a trained shelterbelt configuration state set.

[0102] The mathematical expression of the gradient descent method is:

[0103] ,

[0104] in, is the parameter value of the tth iteration, including model parameters such as salt diffusion coefficient and absorption coefficient; The learning rate controls the step size of parameter updates and is initially set to 0.05. It gradually decreases to 0.01 as the number of iterations increases to ensure convergence stability. is the loss function Parameters The gradient of is calculated as ,in is a small perturbation. The loss function J is defined as the mean square error between the model prediction value and the measured value.

[0105] In the saline-alkali land protection forest demonstration area in Bayannur City, Inner Mongolia, this method successfully optimized the parameter model of salt-tolerant plants such as Tamarix and Haloxylon ammodendron, and the prediction accuracy was improved by about 30%. For example, the optimal absorption intensity parameter of Tamarix was found by optimizing The value is 0.032d⁻¹, which is significantly different from the common value of 0.025d⁻¹ in the literature and more accurately reflects the actual growth characteristics under local environmental conditions.

[0106] To ensure the stability of the parameter optimization process and the effectiveness of the results, the gradient descent method sets the following constraints:

[0107] 1. Unit equivalent resistance model state parameter constraints:

[0108] ,

[0109] in, is the state parameter of the unit equivalent resistance model, reflecting the conductivity characteristics of the soil; and are the lower and upper limits respectively. According to the field test of saline-alkali land in Bayannur City, Inner Mongolia, the and This constraint ensures that the model can adapt to a wide range of salinized soil conditions, from mild to severe.

[0110] 2. Penetration parameter constraints:

[0111] ,

[0112] in, is the infiltration parameter, which characterizes the water conductivity of the soil, with the unit of cm / d; and According to the test data of various saline soils in Bayannur City, Inner Mongolia, the sandy soil is set and , clay soil setting and This constraint reflects the water transport characteristics of different soil textures and ensures that the model is applicable to various soil conditions.

[0113] 3. Upper and lower bounds of soil salinity and its saturation:

[0114] ,

[0115] ,

[0116] in, is soil salinity, in g / kg; is the soil salt saturation, dimensionless, indicating the ratio of the current salt content to the maximum salt content; and are the lower and upper limits of soil salinity respectively. According to the classification standard of saline soil in Northwest China, and This constraint ensures that the model does not produce negative or physically insignificant salt predictions.

[0117] 4. Soil water capacity constraint:

[0118] ,

[0119] in, is the soil water holding capacity, dimensionless; and According to the test results of different types of saline soil in Inner Mongolia, the (near wilting point) and (Close to saturated moisture content). This constraint ensures that the moisture simulation is within a physically reasonable range.

[0120] The setting of these constraints takes into account the physical limitations and biological characteristics of the actual farmland environment to ensure that the optimization results are practical. In the saline-alkali land improvement project in Bayannur City, Inner Mongolia, if the parameters exceed the constraint range, they will be projected back to the constraint interval to ensure the convergence of the algorithm and the validity of the solution. For example, if the infiltration parameter calculated during the optimization process is , exceeds the upper limit , then adjust it to 100 cm / d and continue optimization. This processing method effectively prevents the model parameters from deviating from their actual physical meaning and improves the reliability and stability of the model.

[0121] like Figure 4 As shown, the edge computing module 2 of the present invention uses knowledge distillation technology to compress the prediction model to a scale suitable for edge devices and achieves real-time computing through a distributed edge network. This innovative design solves the practical difficulties of deploying the salt-repellent farmland shelterbelt optimization system in remote areas.

[0122] The core idea of ​​knowledge distillation is to guide the learning of the student model (lightweight model) through the teacher model (complex model). Its mathematical expression is:

[0123] ,

[0124] in, is the knowledge distillation loss function, which is used to measure the degree to which the student model learns the knowledge of the teacher model; is the cross entropy loss, defined as ,in is the true label, is the predicted probability; is the KL divergence loss, defined as , used to measure the difference between two probability distributions; is the true label, which contains the parameters of the shelterbelt configuration scheme; and are the output logits of the student model and the teacher model, that is, the output before the last layer of the neural network is activated; is the softmax function, defined as , convert logits into probability distribution; Temperature parameter, which controls the softness of the soft tag. The larger the value, the smoother the distribution. It is usually set to 2-5, and in this embodiment, it is set to 3; It is a balancing factor that controls the ratio of true label loss to distillation loss. Its typical value is 0.5-0.7. In this embodiment, it is set to 0.6.

[0125] In the saline-alkali land remediation project in Bayannur, Inner Mongolia, the teacher model is a complete deep reinforcement learning model, constructed using a 6-layer CNN and a 4-layer fully connected network, with a parameter size of approximately 40GB. The student model is a compressed, lightweight model, using a 2-layer CNN and a 2-layer fully connected network, with a parameter size of only 4.2GB, a reduction of approximately 90%. Feature map channel dimensionality reduction is achieved through 1×1 convolution, reducing the original 256 channels to 32 channels while preserving key feature information. Tests show that the compressed model achieves a 7.5-fold increase in inference speed while only decreasing prediction accuracy by 2.3%, achieving an optimal balance between performance and resource consumption.

[0126] To adapt to the widespread distribution and harsh environment of salinized farmland, the edge computing network adopts a layered design, consisting of sensing, processing, and decision-making layers. The sensing layer consists of distributed soil and meteorological sensors, with 8-12 nodes deployed per hectare, collecting data every 15 to 60 minutes. In the saline-alkali land demonstration area of ​​Bayannur City, Inner Mongolia, the sensors use solar power and low-power LoRa communication technology, achieving maintenance-free operation for up to a year.

[0127] The processing layer, comprised of field-based edge computing gateways, is responsible for data preprocessing and preliminary analysis. In the Korqin Sandy Land Salinization Control Project in Inner Mongolia, the edge gateways perform local model inference every six hours, generating temporary adjustment recommendations (such as irrigation control and drainage management), enabling rapid response to sudden environmental changes.

[0128] The decision-making layer, comprised of regional edge servers, runs lightweight models to optimize decisions. In the Bayannur City, Inner Mongolia, saline-alkali land improvement project, the decision-making layer servers perform a complete monthly optimization of the shelterbelt configuration, generating medium- and long-term adjustment plans (such as adjusting plant spacing and planning new planting areas). Each layer is connected via multiple wireless communication protocols (LoRa for long-distance, low-power transmission, with an effective range of 5-10 km; NB-IoT for areas without signal towers, using cellular network coverage; and WiFi for short-range, high-bandwidth transmission), forming a complete edge computing network that effectively addresses communication challenges in remote areas.

[0129] Practice has proven that the edge computing and knowledge distillation solution proposed in this paper performs well in resource-constrained farmland environments. In a saline-alkali land remediation project in Bayannur, Inner Mongolia, the system achieved hourly monitoring data analysis and daily shelterbelt management adjustments on edge devices, while maintaining complete configuration optimization and updates at the monthly level, meeting shelterbelt management needs across different timescales.

[0130] like Figure 5 As shown, the method for optimizing the configuration of salinized farmland shelterbelts provided by the present invention comprises the following steps:

[0131] 1. Obtain data from the salinization monitoring area through the drone flight control system, use a high-resolution multispectral camera to obtain ground reflectivity and vegetation leaf reflectivity, and establish the relationship between vegetation leaf reflectivity-water absorption index and conductivity. In the saline-alkali land improvement demonstration area in Bayannur City, Inner Mongolia, the drone uses a six-rotor structure, a load capacity of 4.2kg, a flight time of 42 minutes, a flight altitude of 100 meters, and a single flight coverage area of ​​approximately 85 hectares. The multispectral camera is equipped with five bands: red (620-670nm), green (540-580nm), blue (450-510nm), near-infrared (760-850nm) and red edge (700-740nm), with a spatial resolution of 10cm and a spectral resolution of 10nm. The water absorption index (WAI) calculation formula is:

[0132] ,

[0133] in, It is the reflectance in the near-infrared band, with a wavelength range of 760-850nm, reflecting vegetation biomass and structure information; It is the reflectivity in the short-wave infrared band, with a wavelength range of 1550-1750nm, and is sensitive to moisture content. In the saline-alkali land experimental area of ​​Bayannur City, Inner Mongolia, the relationship between WAI and conductivity was determined through regression analysis of 500 measured sample points. The established regression equation is: , correlation coefficient Reaching 0.87, the verification error is controlled within ±0.8dS / m.

[0134] 2. Deploy soil salinity monitors and ground water level monitors in the monitoring area to collect data on salt content and ground water level, obtain environmental meteorological information, and establish datasets for soil saturated water content and soil saturated salinity. In the saline-alkali land demonstration area of ​​Bayannur City, Inner Mongolia, soil salinity monitors use four-electrode conductivity sensors with a measurement range of 0-20dS / m and an accuracy of ±3%. They are buried in three layers: 0-20cm, 20-50cm, and 50-100cm. Two sensors are placed in each layer to improve reliability, and the sampling frequency is 30 minutes. Ground water level monitors use pressure water level gauges with a measurement range of 0-10m and an accuracy of ±0.1%. Three to five monitoring points are deployed in each test area, and the sampling frequency is 60 minutes. Meteorological information was collected from a small automatic weather station, including temperature (measurement range -40–60°C, accuracy ±0.3°C), humidity (measurement range 0%–100%, accuracy ±3%), precipitation (measurement range 0–10 mm / min, accuracy ±0.2 mm), wind direction (16 directions), and wind speed (measurement range 0–60 m / s, accuracy ±0.3 m / s). The data were sampled every 10 minutes. After preprocessing, these data were stored in Database 3, forming a paired dataset of soil saturated water content and soil saturated salinity, containing approximately 25,000 spatiotemporally matched records.

[0135] 3. Using the aforementioned dataset, a coupled model of soil bulk density and salt transport-tree water uptake was trained. A state update function and shelterbelt configuration set were derived using deep reinforcement learning. In the Bayannur City, Inner Mongolia, salinization control project, training was performed using a stochastic gradient descent algorithm. The initial batch size was 128, gradually increasing to 256 as training progressed to improve stability. 300 training epochs were used, each consisting of approximately 500 updates. The learning rate was initially set at 0.001 and decayed by a stepwise factor of 0.9 every 50 epochs. Model evaluation was performed using the root mean square error (RMSE) and mean absolute error (MAE) metrics. The RMSE for salt prediction was 0.42 g / kg, and the MAE was 0.28 g / kg. The RMSE for water level prediction was 0.15 m, and the MAE was 0.09 m, both exceeding the prediction accuracy of traditional models.

[0136] 4. In the shelterbelt configuration state set, the optimal adjustment parameters and control strategy are used to determine the optimal shelterbelt configuration. In the saline-alkali land shelterbelt demonstration area of ​​Bayannur City, Inner Mongolia, the control strategy adopts the ε-greedy strategy, with an initial value of ε of 0.9 and gradually decreasing to 0.1 according to the exponential decay law. The formula is , where t is the number of training steps, ensuring a balance between exploration and utilization. The optimal configuration schemes obtained through optimization are divided into three categories based on the degree of salinization: For areas with severe salinization (>10 g / kg), plant spacing of 2.5 meters and row spacing of 5 meters are selected, with the main tree species being highly salt-tolerant plants such as Tamarix and Haloxylon ammodendron, and a planting density of 320 plants per hectare; for areas with moderate salinization (5-10 g / kg), plant spacing of 2 meters and row spacing of 4.5 meters are selected, with a mixed planting density of 440 plants per hectare, including Elaeagnus angustifolia, Poplar, and Tamarix, and a planting density of 440 plants per hectare; for areas with mild salinization (<5 g / kg), plant spacing of 1.5 meters and row spacing of 4 meters are selected, with a mixed planting pattern of economic and ecological forests being adopted, and a planting density of 660 plants per hectare. The program evaluation shows that the salt control efficiency has increased by an average of 41% (35% in severely saline areas, 42% in moderately saline areas, and 46% in lightly saline areas), and the average irrigation water saving rate has reached 33% (ranging from 28% to 38%).

[0137] 5. Using drones to capture salinity distribution in target farmland, the shelterbelt configuration is dynamically updated based on current soil salinity and groundwater levels. A predictive model is then used to generate a spatiotemporal model of tree growth in a three-dimensional environment. In the saline-alkali land shelterbelt demonstration area in Bayannur City, Inner Mongolia, the frequency of dynamic updates is adjusted flexibly based on seasonal and meteorological conditions: monthly updates during the growing season (April-September) ensure timely response to changes in the growing season; bimonthly updates during the non-growing season (October-March) reduce resource consumption; and additional updates are conducted following exceptional meteorological conditions (heavy rainfall >50 mm / day, flooding, or prolonged drought >15 days) to address extreme weather events. Built using the Unity engine, the 3D spatiotemporal model supports growth prediction and visualization at four timescales (daily, monthly, seasonal, and annual). It also simulates growth comparison scenarios for different configurations, helping users intuitively understand the effectiveness of shelterbelt configurations. In the alkaline land management project in Bayannur City, Inner Mongolia, this model successfully predicted the development trend of the Tamarix and Haloxylon ammodendron mixed forest within its 10-year growth cycle. The deviation between the predicted tree height and the measured value was controlled within ±12%, providing a reliable basis for long-term shelterbelt planning.

[0138] The proposed system for optimizing shelterbelt deployment on salinized farmland has been experimentally validated and applied in multiple salinized areas in Bayannur City, Inner Mongolia, my country. A two-year comparative experiment was conducted in a 100-hectare demonstration area of ​​moderately salinized farmland (salt content 8-15 g / kg) in the Hexi Corridor of Gansu Province, comparing shelterbelt deployment using the proposed system with a 100-hectare control area using a traditional deployment method. The results demonstrated:

[0139] 1. Salt Control Effect: The average salt content in the surface soil (0-30cm) of the demonstration area decreased by 45% (from an initial 12.5g / kg to 6.9g / kg), significantly higher than the 28% reduction in the control area (from an initial 12.3g / kg to 8.9g / kg). The vertical distribution of salt became more balanced, with the main salt accumulation layer shifting from 15-30cm to 50-70cm, effectively improving the growth environment for crop roots. The uniformity of soil electrical conductivity increased by 60%, reducing the formation of high-salt patches and creating more uniform growing conditions for crops.

[0140] 2. Water resource utilization: Irrigation water consumption in the demonstration area was reduced by 33% (from the conventional 900 m³ / hectare to 603 m³ / hectare), while maintaining a good salt control effect; water use efficiency (salt control effect produced by unit water volume) increased by 55%, from 0.31% / (m³ / hectare) to 0.48% / (m³ / hectare); and the fluctuation range of groundwater level was reduced by 40% (from ±0.85m to ±0.51m), effectively reducing the adverse effects of groundwater level fluctuations on salt migration.

[0141] 3. Ecological benefits: The growth of shelterbelts in the demonstration area was significantly better than that in the control area, with the survival rate increasing by 15% (from 82% to 94%); the average height of two-year-old Tamarix increased by 28% (from 1.8m to 2.3m); the biodiversity index (Shannon-Wiener index) increased by an average of 35%, from 1.65 to 2.23; the soil organic matter content increased by 23%, from the initial 0.8% to 0.98%; the carbon sequestration capacity increased by about 30%, with the average annual carbon sequestration per hectare of shelterbelt increasing from 2.3 tons to 3.0 tons.

[0142] 4. Economic Benefits: The system has a payback period of approximately 3.5 years. The primary economic benefits are: a 38% increase in land asset value (from an initial 15,000 yuan / hectare to 20,700 yuan / hectare); a 42% increase in crop yields (wheat from 3.2 tons / hectare to 4.55 tons / hectare, and corn from 7.5 tons / hectare to 10.65 tons / hectare); and a 22% decrease in irrigation costs (from 450 yuan / hectare to 351 yuan / hectare). Furthermore, the system's high degree of automation reduces manual management costs by approximately 40%, further enhancing economic benefits.

[0143] In the Bayannur City, Inner Mongolia, saline-alkali land management demonstration area, this system has been applied to 5,000 hectares of salinized farmland, achieving a positive interaction between shelterbelts and farmland production. Monitoring data shows that two years after the system's deployment, the regional groundwater level has stabilized within the ideal range (1.8-2.2 meters), topsoil salinity has decreased by an average of 38%, irrigation water efficiency has increased by 42%, and crop yields have increased by 35% to 50%, demonstrating significant ecological and economic benefits.

[0144] In general, the system and method for optimizing the configuration of shelterbelts for salinized farmland provided by the present invention, through the combination of deep reinforcement learning technology and the salt migration-forest water absorption coupling model, realizes the scientific, precise and intelligent configuration of shelterbelts, provides effective technical support for the management of salinized farmland, and has important promotion and application value.

[0145] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be included within the scope of protection of the present invention.

Claims

1. The optimization system for the configuration of salinized farmland shelterbelts is characterized by: include: The salinization monitoring module is used to collect ground reflectance and vegetation leaf reflectance through drones, and use a hyperspectral imaging spectrometer to invert soil conductivity and determine soil salt distribution data; an edge computing module, in communication with the salinization monitoring module, for receiving the soil salt distribution data and calculating a shelterbelt configuration model; A database, connected to the edge computing module, for storing computing models, soil salinity, and vegetation growth status data; a three-dimensional model module, connected to the database and the edge computing module, for generating a three-dimensional spatiotemporal model of forest growth in a three-dimensional environment using soil salinity and groundwater level data; A model optimization module, connected to the three-dimensional model module and the database, is used to optimize the soil salinity and groundwater level simulation models based on deep reinforcement learning to improve the configuration effect of the shelterbelt; The model optimization module specifically includes: taking the spacing between trees to be planted and environmental parameters as states, and the efficiency of shelterbelt configuration and the reduction of irrigation water consumption as reward functions; The model optimization module establishes a salt transport-forest water absorption coupling model through an improved Transformer architecture. The salt transport-forest water absorption coupling model includes: The salt balance equation is used to describe the dynamic changes of salt in the soil; Plant absorption equation, used to describe the absorption process of soil water and salt by plants; The stand water holding capacity equation is used to describe the stand's ability to intercept and retain water; The tree water absorption function is used to describe the water absorption capacity and efficiency of trees under different conditions; The model optimization module uses tree growth parameters and the salt content distribution output by the salt migration model as training data to construct a shelterbelt configuration state set and a shelterbelt structure optimization model. The module uses the current state as input, obtains the next state through the shelterbelt configuration state set, and calculates the corresponding shelterbelt configuration efficiency and irrigation water reduction in the next state as a reward function. The deep reinforcement learning model generates the optimal shelterbelt configuration plan. The shelterbelt configuration state set includes: a state set, a decision set, a reward value, and a planting area change range; wherein the state set includes planting area, plant spacing, row spacing, groundwater level, soil saturated water content, and soil salt distribution parameters; A configuration scheme optimization module is connected to the model optimization module and is used to generate an optimal shelterbelt configuration scheme based on the trained deep reinforcement learning model.

2. The system for optimizing the configuration of salinized farmland shelterbelts according to claim 1, characterized in that: It also includes a meteorological monitoring module, which is connected to the model optimization module and is used to obtain meteorological condition data such as air temperature, precipitation, wind direction and wind speed to determine the optimal growth strategy.

3. The system for optimizing the configuration of salinized farmland shelterbelts according to claim 1, characterized in that: The trained deep reinforcement learning model includes a control function and a loss function; wherein, the control function is used to generate decisions based on the current state, the loss function is used to evaluate the gap between the predicted value and the true value, the model uses a discount factor to balance short-term and long-term rewards, and the probability distribution of the shelterbelt configuration state is used to guide the reinforcement learning process.

4. The system for optimizing the configuration of salinized farmland shelterbelts according to claim 1, characterized in that: The updating of the shelterbelt configuration state set includes: evaluating the current shelterbelt configuration through a deep reinforcement learning model, inputting the soil salinity and groundwater level into the coupling model, outputting the soil salinity and groundwater level in the next state through the coupling model, and repeating the iterative update until the optimal configuration plan is obtained.

5. The system for optimizing the configuration of salinized farmland shelterbelts according to claim 1, characterized in that: Parameter optimization in the salt migration-forest water absorption coupling model is achieved through the gradient descent method. Specifically, the salt content is divided into multiple salinity levels with equal intervals through linear interpolation, and the salt balance equation and plant water absorption equation are solved to obtain the vegetation's resistance to different salt migration processes, and then the trained shelterbelt configuration state set is obtained.

6. The system for optimizing the configuration of salinized farmland shelterbelts according to claim 5, characterized in that: The constraints of the gradient descent method include: Unit equivalent resistance model state parameter constraints; Infiltration parameter constraints; Upper and lower boundary constraints on soil salinity and its saturation; Soil water capacity constraint; The constraints mentioned above ensure the stability of the model training process and the effectiveness of the results.

7. The system for optimizing the configuration of salinized farmland shelterbelts according to claim 1, characterized in that: The edge computing module compresses the prediction model to a scale suitable for edge devices through knowledge distillation technology, and realizes real-time computing through a distributed edge network; the knowledge distillation technology reduces the dimension of feature graph channels, retains the key decision-making capabilities of the model, and enables the system to update the shelterbelt configuration plan at the hourly level in a resource-constrained environment.

8. The optimization method of the salinized farmland shelterbelt configuration optimization system according to any one of claims 1 to 7, characterized in that: The following steps are involved: Data from the salinization monitoring area was acquired through the drone flight control system. A high-resolution multispectral camera was used to obtain ground reflectance and vegetation leaf reflectance, and the relationship between vegetation leaf reflectance, water absorption index, and electrical conductivity was established. Soil salinity monitors and ground water level monitors were deployed in the monitoring area to collect data on salinity and ground water levels, obtain environmental meteorological information, and establish datasets of soil saturated water content and soil saturated salinity. The dataset was used to train a coupled model of soil bulk density and salt transport-tree water absorption, and a state update function and shelterbelt configuration set were obtained based on deep reinforcement learning. In the set of shelterbelt configuration states, the optimal adjustment parameters and control strategies are used to determine the optimal shelterbelt configuration; The target farmland is photographed by drones to obtain the salt status distribution, the shelterbelt configuration is dynamically updated according to the current soil salinity and groundwater level, and a prediction model is used to generate a spatiotemporal model of forest growth in a three-dimensional environment.

Images