Forest grass ecological restoration system, method, equipment and medium

Abstract

The invention discloses a forest grass ecological restoration method, system and device and a medium. A multi-source data acquisition module acquires static space data and dynamic time sequence data through a sensor network and an unmanned aerial vehicle; the spatio-temporal feature fusion module preprocesses data, extracts key features through a spatio-temporal attention mechanism, calculates dynamic contribution degrees, and unifies the key features and the dynamic contribution degrees into preset resolution grids; the multi-model collaborative decision-making module generates plant suitability evaluation and optimizes a planting scheme by means of reinforcement learning; the dynamic restoration execution module plans an unmanned aerial vehicle path to execute planting, receives feedback of a sensor and updates model parameters through incremental learning; and the visualization and storage module generates a dynamic graph, stores data to a distributed database according to a timestamp and a space coordinate, and synchronizes the data to the cloud. According to the method, the accuracy and the adaptability of a restoration scheme are improved through a forest grass ecological restoration technology which fuses space-time dynamic characteristics and multi-model collaborative decision-making and has a real-time feedback capability.

Description

Technical Field

[0001] This invention belongs to the field of ecological restoration and intelligent decision-making technology, and relates to a forest and grassland ecological restoration system, method, equipment and medium. Background Technology

[0002] Currently, forest and grassland ecological restoration faces challenges such as high environmental complexity, diverse data dimensions, and insufficient dynamic adaptability of restoration schemes. Existing technologies for degraded grassland restoration largely rely on static environmental feature analysis, such as calculating the contribution of environmental features through random forests and optimizing SVM models to select plant species. However, they lack dynamic modeling of time-series data (such as seasonal weather changes and vegetation growth cycles). Ecological restoration in arid regions uses semi-supervised learning to handle small sample problems, but struggles to integrate high-resolution spatial data with real-time monitoring data. While forest and grassland planting suitability assessment introduces Long Short-Term Memory (LSTM) networks to process time series data, it is not deeply coupled with spatial features (such as topographic gradients and soil heterogeneity).

[0003] In addition, existing technologies have the following limitations: environmental feature extraction is mostly a static snapshot, ignoring the impact of temporal fluctuations in parameters such as weather and soil moisture on plant suitability; plant planting schemes and path planning lack dynamic adjustment mechanisms and cannot cope with environmental mutations (such as extreme precipitation and pests) during the restoration process; and the ability of multi-model collaboration is insufficient, with single models (such as SVM and random forest) struggling to handle the nonlinear relationships of spatiotemporal cross features.

[0004] Therefore, there is an urgent need for a forest and grassland ecological restoration technology that integrates spatiotemporal dynamic characteristics, multi-model collaborative decision-making, and real-time feedback capabilities to improve the accuracy and adaptability of restoration solutions. Summary of the Invention

[0005] This invention aims to provide a precise forest and grassland ecological restoration system and method that integrates spatiotemporal dynamic features. By integrating multi-source data, spatiotemporal feature modeling, multi-model collaborative decision-making, and dynamic feedback optimization, it solves the problems of static analysis, poor adaptability, and insufficient accuracy in existing technologies, and realizes intelligent and precise forest and grassland ecological restoration.

[0006] The first aspect of this application provides a forest and grassland ecological restoration system, comprising: The multi-source data acquisition module is used to collect environmental data through sensor networks and drones. The collected data includes static spatial data and dynamic time-series data. The dynamic time-series data includes hourly recorded temperature, humidity, light intensity, and weekly recorded vegetation coverage. The spatiotemporal feature fusion module is used to preprocess static spatial data and dynamic temporal data. It extracts key features and calculates dynamic contribution through a spatiotemporal attention mechanism. The preprocessing includes removing outliers using the IQR method, smoothing temporal data using the sliding window method, and completing spatial data using Kriging interpolation, and uniformly converting it into a raster format with a preset resolution. The multi-model collaborative decision-making module is used to generate dynamic evaluation results of plant suitability and optimize planting plans through reinforcement learning; The dynamic repair execution module is used to plan the planting path of the drone and execute the planting task, receive sensor feedback data in real time, and update the model parameters through incremental learning; The visualization and storage module is used to generate a visualization map of the spatiotemporal dynamic repair scheme, store the data in a distributed database according to timestamps and spatial coordinates, and synchronize it to the cloud.

[0007] Optionally, in the spatiotemporal feature fusion module, the spatiotemporal attention mechanism weights the features using temporal attention weights and spatial attention weights, wherein the weighting formula is as follows: Where X(i,t) is the feature vector of grid cell i at time t, α(t) is the temporal attention weight, obtained by modeling the temporal data using LSTM, and β(i) is the spatial attention weight, calculated by a spatial autoencoder. This is the weighted eigenvector.

[0008] Optionally, the multi-model collaborative decision-making module uses an improved hybrid model of random forest and Transformer to output the fitness S(k,i,t) of the k-th plant at grid i and time t, calculated using the following formula: Where σ is the sigmoid function, W and b are model parameters, and Transformer is the Transformer model. This is the weighted eigenvector.

[0009] Optionally, the dynamic repair execution module performs the planting task using an objective function derived from an improved A* algorithm, wherein the objective function is: , where dist(i,j) is the distance from grid i to grid j, λ is the fitness penalty coefficient, and S(k,i,t) is the fitness of the k-th plant at grid i and time t.

[0010] A second aspect of this application provides a method for forest and grassland ecological restoration, comprising: S1. Multi-source data acquisition and preprocessing: Static spatial data and dynamic time-series data are acquired through sensor networks and UAVs, and the data is preprocessed and converted into a raster format with a preset resolution; S2. Spatiotemporal Feature Fusion and Contribution Calculation: Construct a spatiotemporal feature matrix, weight the features through a spatiotemporal attention mechanism, and calculate the dynamic contribution of each feature in different spatiotemporal units based on an improved random forest; S3. Multi-model collaborative decision-making: The weighted features are input into the improved random forest-transformer hybrid model to obtain plant suitability, and the planting plan is optimized through reinforcement learning; S4. Dynamic Execution and Feedback Optimization: Based on the improved A* algorithm, the drone planting path is planned, sensor feedback data is received in real time, and the model parameters are updated through incremental learning; S5. Data storage and visualization: Data is stored in a distributed database by timestamp and spatial coordinates, and a repair effect prediction map is generated and overlaid with real-time monitoring data for display.

[0011] Optionally, in step S2, the spatiotemporal feature matrix is ​​X(i,t)=[DEM(i), salinity(i),T(t),H(t),L(t),C(i,t)], where X(i,t) is the data of raster cell i at time t, DEM(i) is the digital elevation model data of raster i, salinity(i) is the soil salinity data of raster i, T(t) is the air temperature data at time t, H(t) is the humidity data at time t, L(t) is the light intensity data at time t, and C(i,t) is the vegetation cover data of raster i at time t.

[0012] Optionally, in step S2, the dynamic contribution is determined by the formula... Calculate, where D(i,t) is the feature dataset of raster i at time t, and Var(D) is the variance of dataset D. For the subset D k The variance of the feature is R(i,t), which is the reduction in variance of the feature in the spatiotemporal unit (i,t).

[0013] Optionally, in step S3, reinforcement learning uses a reward function. Optimize planting ratios, where M is the Shannon diversity index. Let p(k,t) be the planting proportion of the k-th plant at time t, and S(k,i,t) be the suitability of the k-th plant at grid i and time t. , is the weighting coefficient, T is the total number of time steps, and K is the total number of plant species.

[0014] A third aspect of this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described forest and grassland ecological restoration method.

[0015] In a fourth aspect, this application provides a computer-readable medium storing a computer program that, when executed by a processor, implements the above-described forest and grassland ecological restoration method.

[0016] Compared with the prior art, the present invention has the following beneficial effects: For the first time, a spatiotemporal attention mechanism is introduced into forest and grassland restoration, dynamically weighting environmental characteristics to solve the problem that static analysis cannot cope with temporal changes; By integrating the random forest-transformer hybrid model with reinforcement learning, we can achieve spatiotemporal dynamic evaluation of plant suitability and adaptive optimization of planting schemes; A path planning algorithm with suitability penalty is proposed to balance planting efficiency and vegetation survival rate for ecological restoration.

[0017] By overcoming the limitations of a single model, the accuracy of suitability assessment is obtained through multi-model collaborative processing of spatiotemporal cross-features.

[0018] Incremental learning and real-time feedback mechanisms are introduced to enable the repair scheme to dynamically adapt to environmental changes, thereby improving the survival rate.

[0019] Multi-source data fusion technology (static + dynamic) improves the accuracy of repair in small sample areas. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of a forest and grassland ecological restoration system according to one embodiment of the present invention; Figure 2 This is a flowchart of a forest and grassland ecological restoration method according to an embodiment of the present invention; Figure 3 This is a schematic diagram of a robot electronic device according to an embodiment of the present invention. Detailed Implementation

[0021] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0022] In one embodiment, such as Figure 1 As shown, a forest and grassland ecological restoration system is provided, which corresponds one-to-one with the forest and grassland ecological restoration methods in the following embodiments. The forest and grassland ecological restoration system includes: a multi-source data acquisition module, a spatiotemporal feature fusion module, a multi-model collaborative decision-making module, a dynamic restoration execution module, and a visualization and storage module. The detailed descriptions of each functional module are as follows: The multi-source data acquisition module is used to collect environmental data through sensor networks and drones. The collected data includes static spatial data and dynamic time-series data. The dynamic time-series data includes hourly recorded temperature, humidity, light intensity, and weekly recorded vegetation coverage. Specifically, the multi-source data acquisition module collects data from the forest and grassland ecological precision restoration system, while integrating ground sensor networks and aerial drone remote sensing technology to achieve comprehensive, multi-dimensional, and time-series data collection of the restoration area environment. By acquiring basic data reflecting regional spatial heterogeneity (such as topographic differences and soil property distribution) and temporal dynamic changes (such as meteorological fluctuations and vegetation growth cycles), the multi-source data acquisition module provides high-quality input for subsequent spatiotemporal feature fusion and model decision-making.

[0023] The multi-source data acquisition module includes two acquisition methods: a sensor network, which collects soil and near-surface meteorological data in real time by deploying ground devices such as soil moisture sensors, temperature sensors, and light sensors in the restoration area to capture micro-dynamic changes in the environment; and a UAV remote sensing, in which UAVs equipped with LiDAR (Light Detection and Ranging) and multispectral cameras acquire macro-data of the area from the air. The LiDAR is used to generate high-precision terrain data (such as digital elevation models), and the multispectral camera is used to retrieve soil properties (such as salinity and organic matter) and vegetation status (such as cover and growth vitality).

[0024] The data collected by the multi-source data acquisition module is divided into two categories: static spatial data and dynamic temporal data. The static spatial data reflects the stable spatial attributes of the region, such as topography (digital elevation model) and soil type (salt content, organic matter content), etc. This type of data changes slowly over time and is usually collected periodically (e.g., once a month). The dynamic temporal data reflects real-time or periodic changes in the environment, such as hourly recorded temperature, humidity, and light intensity (meteorological data), and weekly monitored vegetation cover (vegetation growth data), etc. This type of data is used to capture the impact of temporal fluctuations on plant fitness.

[0025] In this embodiment, the multi-source data acquisition module can realize comprehensive monitoring of forest and grassland at both the "micro-macro" and "static-dynamic" levels.

[0026] The spatiotemporal feature fusion module is used to preprocess static spatial data and dynamic temporal data. It extracts key features and calculates dynamic contribution through a spatiotemporal attention mechanism. The preprocessing includes removing outliers using the IQR method, smoothing temporal data using the sliding window method, and completing spatial data using Kriging interpolation, and uniformly converting it into a raster format with a preset resolution.

[0027] Specifically, the spatiotemporal feature fusion module integrates and optimizes the static spatial data and dynamic temporal data acquired by the multi-source data acquisition module, and extracts key features through intelligent algorithms to quantify the importance (dynamic contribution) of different features in the spatiotemporal dimension, providing high-quality feature input for subsequent model decision-making.

[0028] In this embodiment, the suitability of forest and grassland ecological restoration depends not only on stable spatial attributes (static characteristics) such as topography and soil type, but also on dynamic factors (temporal characteristics) such as temperature and precipitation that fluctuate over time. Furthermore, the weight of both factors changes dynamically with time and spatial location (for example, rainfall during the rainy season has a more significant impact on low-lying areas than during the dry season). Therefore, fusion needs to be achieved through the following steps: Standardization processing is performed on static spatial data (such as topography and soil type) and dynamic time-series data (such as temperature, precipitation, and vegetation NDVI index), including outlier removal (to ensure data accuracy), missing value completion (such as using kriging interpolation for spatial data and sliding window filling for time-series data), and format unification (converting to raster data of the same resolution, such as 10m×10m), to eliminate differences in data dimensionality and scale.

[0029] The algorithm automatically learns "temporal weights" and "spatial weights" to dynamically adjust the importance of features in different spatiotemporal units.

[0030] It is used to measure the impact of temporal features (such as temperature and precipitation) at different points in time (for example, precipitation has a higher weight during critical periods of plant growth than during non-critical periods), and is obtained by modeling temporal data through a Long Short-Term Memory (LSTM) network. It is used to measure the influence of static features (such as terrain) at different spatial locations (e.g., terrain has a higher weight on steep slopes than on flat land), and is calculated by reducing the dimensionality of spatial data through a spatial autoencoder. Applying time and spatial weights to the original features yields key features that integrate spatiotemporal information, namely, spatiotemporal attention features.

[0031] The dynamic contribution calculation is based on an improved random forest algorithm, quantifying the degree of influence (i.e., contribution) of each fused feature on plant fitness in different spatiotemporal units (a grid at a certain time point). The higher the contribution, the more significant the influence of the feature on vegetation growth in that spatiotemporal unit, providing a priority basis for subsequent model decisions.

[0032] In the spatiotemporal feature fusion module, the spatiotemporal attention mechanism weights features using temporal attention weights and spatial attention weights, wherein the weighting formula is as follows: X(i,t) is the feature vector of grid cell i at time t, α(t) is the temporal attention weight, obtained by modeling the temporal data using LSTM, and β(i) is the spatial attention weight, calculated by a spatial autoencoder. This is the weighted eigenvector.

[0033] For example, based on the improved random forest algorithm, fusion features are calculated. The contribution of each dimension (e.g., DEM, precipitation) to raster i (slope 18°) and time t (growing season) is calculated when the fused features of the raster over the past 3 months and the corresponding vegetation survival rate data are input. Then, through a random forest splitting process, the variance reduction of each feature to "vegetation survival rate" is quantified (using the formula). From this, we can see that the contribution of precipitation (P) is R(i,t)=0.72 (the highest), and the contribution of DEM is R(i,t)=0.65. This indicates that precipitation has the most significant impact on vegetation survival rate during the growing season of this grid, and should be used as the core feature for subsequent model decisions.

[0034] In this embodiment, the spatiotemporal feature fusion module can output key features that "integrate spatiotemporal information + dynamic weights", which solves the limitation of the "one-size-fits-all" approach in traditional static analysis (such as ignoring the differences in the influence of different seasons and terrains), and provides more accurate input for subsequent dynamic evaluation results of plant suitability and optimization of planting schemes.

[0035] The multi-model collaborative decision-making module is used to generate dynamic evaluation results of plant suitability and optimize planting plans through reinforcement learning; Specifically, the multi-model collaborative decision-making module integrates an improved random forest-transformer hybrid model with reinforcement learning algorithms to achieve closed-loop decision-making for "dynamic assessment of plant suitability" and "optimization of planting schemes." Its core logic is: first, to accurately assess the plant suitability of different plants in different spatiotemporal units through the hybrid model; and then, based on the assessment results, to dynamically optimize plant species, planting ratios, and timing arrangements through reinforcement learning to maximize the ecological restoration effect (such as vegetation survival rate and biodiversity).

[0036] The improved random forest-Transformer hybrid model consists of two parts: a random forest component and a Transformer component. The random forest component excels at handling high-dimensional nonlinear data and can quickly extract the basic correlations between environmental features (such as topography, soil, and weather) and plant growth, outputting a preliminary fitness score. The Transformer component uses a self-attention mechanism to capture long-range dependencies in spatiotemporal data (such as the impact of precipitation changes in a certain area on vegetation growth over the next three months), compensating for the random forest's deficiency in capturing temporal correlations. It is worth noting that the hybrid logic of the random forest-transformer hybrid model is that the basic features extracted by the random forest are used as the input of the Transformer. After processing by the Transformer, the final dynamic evaluation result of plant fitness (i.e., the fitness of a certain plant in a certain grid and at a certain time) is output, taking into account both static feature association and spatiotemporal dynamic dependence.

[0037] The multi-model collaborative decision-making module uses an improved hybrid model of random forest and Transformer to output the fitness S(k,i,t) of the k-th plant at grid i and time t, calculated using the following formula: Where σ is the sigmoid function, W and b are model parameters, and Transformer is the Transformer model. This is the weighted eigenvector.

[0038] In this embodiment, the reinforcement learning-optimized planting scheme is based on the dynamic evaluation results of plant suitability output by the hybrid model. Through the interaction of "agent (decision model) - environment (remediation area) - reward (ecological effect)," iteratively adjusts the plant species, planting ratio (the proportion of different plants), and timing arrangement (such as phased planting time) to ultimately find the scheme that optimizes the ecological effect. Here, the state refers to the environmental characteristics of the current spatiotemporal unit (such as soil moisture and temperature of grid i in month t) and the status of the planted plants; the action refers to selecting the plant species to be planted, the planting ratio of that species, and the planting time (such as month t or month t+1); the reward is a comprehensive indicator centered on plant diversity (Shannon diversity index) and dynamic suitability (hybrid model evaluation results). The higher the reward value, the better the scheme.

[0039] The dynamic repair execution module is used to plan the planting path of the drone and execute the planting task, receive sensor feedback data in real time, and update the model parameters through incremental learning; Specifically, the dynamic repair execution module is responsible for transforming the optimized planting plan output by the multi-model collaborative decision-making module into actual planting actions, and ensuring dynamic adaptation of the repair effect through real-time feedback and model update mechanisms. In this invention, the dynamic repair execution module plans the UAV planting path based on the improved A* algorithm to ensure efficient and accurate execution of the planting task; at the same time, it monitors the vegetation growth status and environmental changes after planting in real time through a sensor network. When the monitoring data shows that the deviation between the actual effect and the expectation exceeds a threshold, incremental learning is triggered to update the model parameters, making subsequent decisions more in line with the actual environment, forming a closed loop of "execution-monitoring-optimization".

[0040] Traditional A* algorithm only plans paths based on the shortest distance, while the improved A* algorithm introduces a "plant suitability penalty" in addition to distance, prioritizing high-suitability grid areas to reduce ineffective drone flights in low-suitability areas, while balancing path length and ecological restoration effectiveness. Its objective function is: Where dist(i,j) is the distance from grid i to grid j, λ is the fitness penalty coefficient, and S(k,i,t) is the fitness of the k-th plant at grid i and time t.

[0041] This invention uses a sensor network (soil moisture and vegetation coverage sensors) and drones for regular inspections to collect real-time data after planting, including: vegetation survival rate (e.g., number of surviving plants in a certain area 3 weeks after planting / total number of plants); soil moisture changes (e.g., moisture difference between planted and unplanted areas); and vegetation growth rate (e.g., weekly plant height increase). When the vegetation survival rate in a certain area is lower than a preset threshold (e.g., 60%), or the soil moisture deviates from the suitable range (e.g., below the critical humidity for plant growth), it is determined as an "effect deviation," triggering the model update process.

[0042] It is worth noting that traditional model updates require retraining on all data, which is time-consuming and inefficient. Incremental learning updates model parameters only with newly collected biased data (such as environmental characteristics and plant status in low-survival-rate areas), retaining historical training results, thus achieving "lightweight updates." The specific process is as follows: Extract the spatiotemporal characteristics of the biased area (such as soil salinity and precipitation data of raster i at time t).

[0043] By associating features with actual survival rates, an incremental training set is formed.

[0044] Fine-tuning the parameters of the random forest-Transformer hybrid model in the multi-model collaborative decision-making module using incremental training sets (such as adjusting the attention weights of the Transformer) makes the model's suitability assessment of the region more accurate.

[0045] The updated model outputs a new suitability S(k,i,t), providing a basis for subsequent optimization of planting schemes.

[0046] In this embodiment, the dynamic repair execution module realizes the precise execution and dynamic optimization of planting tasks. By improving the A* algorithm, the planting efficiency has been significantly improved. Incremental learning has significantly improved the repair effect of low survival rate areas and effectively solved the problem of "disconnect between planning and actual environment".

[0047] The visualization and storage module is used to generate a visualization map of the spatiotemporal dynamic repair scheme, store the data in a distributed database according to timestamps and spatial coordinates, and synchronize it to the cloud.

[0048] Specifically, the visualization and storage module undertakes two key functions: first, it transforms complex spatiotemporal dynamic repair solutions into intuitive and easy-to-understand visual maps, helping decision-makers quickly understand the spatiotemporal distribution and dynamic changes of the solutions; second, it stores all data throughout the process (from raw data collection to the final repair solution) according to standardized rules, ensuring data traceability and reusability, and enabling cross-scenario access and backup through cloud synchronization. Its core logic is: to lower the decision-making threshold through visualization technology, and to ensure data integrity and security through structured storage and cloud synchronization, forming a full lifecycle management system of "data-solution-feedback".

[0049] The spatiotemporal dynamic repair scheme visualization map generation is based on the planting scheme output by the multi-model collaborative decision-making module, the path planning of the dynamic repair execution module, and real-time monitoring data. Through geographic information system (GIS) and dynamic visualization technology, a multi-dimensional map is generated, including: spatiotemporal suitability dynamic map, planting path and progress map, repair effect comparison map, and scheme optimization dynamic trajectory map. The spatiotemporal suitability dynamic map uses a grid as the unit and color gradients (e.g., blue-green-red) to represent the spatial distribution of plant suitability at different time points (e.g., monthly). Overlaying a timeline allows for a visual view of suitability changes with the seasons (e.g., the expansion of high suitability areas during the rainy season). The planting path and progress map uses dashed lines or arrows to mark the drone planting path, combined with different colors (e.g., yellow indicates completed, gray indicates pending) to display the planting progress of each area, updating the ratio of planted area to planned area in real time. The restoration effect comparison map shows the vegetation coverage grid map before planting on the left and the data from the same period after planting on the right. The restoration effect is visually displayed through color differences (e.g., brown for bare land, green for vegetation coverage), and a survival rate heatmap can be overlaid (red indicates low survival rate, blue indicates high survival rate). The scheme optimization dynamic trajectory map uses broken lines or bubbles to mark the adjustment trajectory of the planting ratio during the reinforcement learning iteration process (e.g., the process of a certain plant ratio increasing from 30% to 50%), assisting in the analysis of key nodes in scheme optimization.

[0050] Data storage and cloud synchronization employ a structured storage strategy with a dual-dimensional index of "timestamp-spatial coordinates." Data throughout the entire process is categorized and stored in a distributed database (such as HBase), and synchronized to the cloud (such as a private or public cloud platform) at a fixed frequency, ensuring data security and accessibility. Specifically, this includes four steps: data classification, storage structure, distributed database storage, and cloud synchronization.

[0051] The data is categorized into raw acquisition data (raw sensor values, UAV imagery), preprocessed data (cleaned raster data, spatiotemporal fusion features), decision data (suitability assessment results, planting plans), and feedback data (survival rate, soil change monitoring values). The storage structure consists of each data record containing a timestamp (accurate to the second), spatial coordinates (raster center point latitude and longitude), data type label, and feature value, for example, "2025-05-10 08:00:00, E112.3°N38.5°, soil moisture, 25%". The distributed database storage supports petabyte-level data storage, distributing data by spatial region (e.g., 1km×1km) using sharding technology to improve query efficiency (e.g., querying precipitation data for a certain area within 3 months can return within 10 seconds). Cloud synchronization uses an incremental synchronization mechanism, synchronizing only newly added or updated data to the cloud, with automatic backup every morning, supporting access from multiple terminals (computers, mobile devices) through permission verification, facilitating cross-team collaboration (e.g., researchers remotely accessing data, decision-makers viewing repair progress in real time).

[0052] In this embodiment, the visualization and storage module realizes "complex data visualization" and "full lifecycle data management". Decision-makers can grasp the overall restoration through the map. Distributed storage and cloud synchronization ensure that massive amounts of data during the restoration period can be efficiently queried and securely retained, providing data support for subsequent ecological restoration research and solution iteration.

[0053] Specific limitations regarding the forest and grassland ecological restoration system can be found in the limitations of the forest and grassland ecological restoration methods below, and will not be repeated here. Each module in the aforementioned forest and grassland ecological restoration system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0054] In one embodiment, such as Figure 2 As shown, a method for forest and grassland ecological restoration is provided, which is applied to... Figure 2 Taking China as an example, the following specific steps will be used: S1: Multi-source data acquisition and preprocessing: Static spatial data and dynamic time-series data are acquired through sensor networks and drones, and the data is preprocessed and converted into a raster format with a preset resolution.

[0055] Specifically, multi-source data acquisition and preprocessing integrates ground sensors and UAV remote sensing technology to obtain full-dimensional data of the restoration area. Standardization processes are then used to eliminate data noise and format differences, ultimately converting the data into raster data of uniform resolution. This provides high-quality input for subsequent spatiotemporal feature fusion and model decision-making. The core logic is: first, data reflecting the spatial heterogeneity (static features) and temporal dynamics (temporal series features) of the region is collected through multiple methods; then, through cleaning, completion, and format standardization, the raw data is transformed into a structured, computable raster dataset, ensuring the accuracy, completeness, and consistency of the data.

[0056] The data acquisition methods are divided into two categories: static spatial data and dynamic temporal data. Static spatial data refers to data that changes slowly over time and primarily reflects stable spatial attributes of the region. This data is acquired through UAV remote sensing and includes topographic data and soil property data. Topographic data refers to digital elevation models (DEMs) acquired by UAVs equipped with LiDAR, reflecting topographic features such as slope and aspect. Soil property data refers to soil salinity and organic matter content acquired by UAVs equipped with multispectral cameras, reflecting the physicochemical properties of the soil. The static spatial data is acquired at a low frequency (e.g., once a month), mainly focusing on spatial distribution differences. Dynamic temporal data refers to data that changes rapidly over time and reflects the real-time state of the environment. This data is acquired through a ground-based sensor network and includes meteorological data and vegetation growth data.

[0057] The meteorological data refers to the hourly records of soil moisture, air temperature, and light intensity from soil moisture sensors, temperature sensors, and light sensors; the vegetation growth data refers to the weekly vegetation cover data collected by the drone (derived through NDVI index inversion), reflecting the dynamics of vegetation growth. Dynamic time-series data is collected frequently (hourly or weekly), primarily focusing on fluctuations over time.

[0058] After multi-source data collection, the raw data needs to be standardized, such as noise removal, missing value filling, and format unification. Specific steps include outlier removal, missing value filling, format conversion, and resolution unification.

[0059] The outlier removal method uses IQR (interquartile range) to identify and remove data that exceeds a reasonable range (such as soil moisture > 100% and temperature < -40℃) to avoid interference from extreme values. The missing value completion method is as follows: for spatial data (such as missing soil salinity in some raster cells), Kriging interpolation is used (based on the spatial correlation of known surrounding raster data); for time-series data (such as missing light data for a certain hour), a sliding window method is used (using the average value of the three hours before and after to complete the data). The format conversion and resolution unification are to convert all data into raster format with uniform raster resolution (e.g., 10m×10m), ensuring that data from different sources and of different types can be aligned in spatial scale (i.e., static and dynamic features can be associated at the same raster location).

[0060] In this embodiment, the multi-source data acquisition and preprocessing steps transform the raw, heterogeneous multi-source data into a structured, standardized raster dataset, which effectively reduces outliers and missing values, providing a clean, aligned, and computable input foundation for the spatiotemporal feature fusion module.

[0061] S2: Spatiotemporal Feature Fusion and Contribution Calculation: Construct a spatiotemporal feature matrix, weight the features through a spatiotemporal attention mechanism, and calculate the dynamic contribution of each feature in different spatiotemporal units based on an improved random forest.

[0062] Specifically, spatiotemporal feature fusion and contribution calculation integrate static spatial features and dynamic temporal features to extract the most critical spatiotemporal information affecting vegetation growth and quantify the importance (dynamic contribution) of different features in specific spatiotemporal units, providing feature inputs that "focus on key influencing factors" for subsequent model decisions. Since forest and grassland growth is jointly influenced by spatial attributes (such as topography and soil) and temporal dynamics (such as weather and growth stage), and the intensity of both influences varies with time and space (e.g., rainfall during the rainy season has a more significant impact on low-lying areas), it is necessary to integrate features through matrix construction, highlight key factors through attention mechanisms, and quantify the influence intensity through improved random forests, ultimately achieving "precise spatiotemporal feature fusion + dynamic contribution quantification."

[0063] A spatiotemporal feature matrix is ​​a structured data carrier that integrates static spatial features and dynamic temporal features, used to uniformly describe the comprehensive environmental characteristics of a specific raster at a specific time. Its construction logic includes: spatial dimension, temporal dimension, and spatiotemporal feature matrix. The spatial dimension uses preprocessed raster cells as basic units (e.g., 10m × 10m), with each raster containing static spatial features (e.g., Digital Elevation Model (DEM), soil salinity); the temporal dimension uses fixed time intervals (e.g., hours, weeks) as time units, with each time unit containing dynamic temporal features (e.g., temperature, precipitation, vegetation cover). The spatiotemporal feature matrix is: X(i,t)=[DEM(i), salinity(i),T(t),H(t),L(t),C(i,t)], where X(i,t) is the feature vector of raster cell i at time t, DEM(i) is the digital elevation model data of raster i, salinity(i) is the soil salinity data of raster i, T(t) is the air temperature data at time t, H(t) is the humidity data at time t, L(t) is the light intensity data at time t, and C(i,t) is the vegetation cover data of raster i at time t.

[0064] The spatiotemporal attention mechanism automatically learns "temporal weights" and "spatial weights" to dynamically adjust the influence intensity of different features in specific spatiotemporal units, making the weighted features more focused on key influencing factors. The specific process is as follows: Temporal attention weights (α(t)): These measure the intensity of the influence of dynamic temporal features (such as precipitation and temperature) at different points in time, and are obtained by modeling temporal data using a Long Short-Term Memory (LSTM) network. For example, precipitation has a higher weight during the growing season (May-September) than during the non-growing season because precipitation during the growing season has a more direct impact on vegetation growth.

[0065] Spatial attention weight (β(i)): Measures the influence intensity of static spatial features (such as terrain and soil) in different rasters, and is calculated by reducing the dimensionality of spatial data through a spatial autoencoder. For example, the terrain weight of a steep slope raster is higher than that of a flat land raster because steep slopes are prone to soil erosion, and the influence of terrain is more significant.

[0066] Weighted fusion: Multiplying the temporal weights and spatial weights and applying the result to the original feature matrix yields a feature vector that fuses key spatiotemporal information. Where X(i,t) is the feature vector of grid cell i at time t, α(t) is the temporal attention weight, obtained by modeling the temporal data using LSTM, and β(i) is the spatial attention weight, calculated by a spatial autoencoder. The weighted feature vector is denoted as , and the larger its value, the more critical the influence of the feature on vegetation growth in that spatiotemporal unit.

[0067] The dynamic contribution is used to quantify the impact of each weighted feature on vegetation growth in a specific spatiotemporal unit (grid i, time t). It is calculated using an improved random forest algorithm, and its core function is to measure the reduction in variance of vegetation growth indicators (such as survival rate) by the feature (the higher the contribution, the more significant the variance reduction). The dynamic contribution is expressed by the formula... Calculate, where D(i,t) is the feature dataset of raster i at time t, and Var(D) is the variance of dataset D. For the subset D k The variance, R(i,t), is the reduction in variance of the feature in the spatiotemporal unit (i,t), where D is the feature dataset of that spatiotemporal unit. k This is the split subset of the dataset.

[0068] In this embodiment, the spatiotemporal feature fusion and contribution calculation steps achieve the following: 1. Integrating static and dynamic features into a structured matrix. 2. Highlighting key spatiotemporal influencing factors through an attention mechanism. 3. Quantifying the influence intensity of features in different spatiotemporal regions. The final output weighted features and dynamic contribution provide the multi-model collaborative decision-making module with precise input for "focusing on key factors and quantifying influence intensity".

[0069] S3: Multi-model collaborative decision-making: Weighted features are input into an improved random forest-transformer hybrid model to obtain plant suitability, and planting schemes are optimized through reinforcement learning.

[0070] Specifically, multi-model collaborative decision-making achieves synergy between "accurate plant suitability assessment" and "dynamic optimization of planting schemes" by coupling an improved random forest-Transformer hybrid model with a reinforcement learning algorithm. The multi-model collaborative decision-making first uses the hybrid model to predict the growth suitability of different plants in each spatiotemporal unit based on spatiotemporal features. Then, reinforcement learning is used to find planting strategies that maximize ecological restoration effects (such as vegetation survival rate and biodiversity), forming a closed loop of "prediction-decision-optimization." This method overcomes the limitations of single models—random forests excel at handling high-dimensional features but ignore temporal correlations, Transformers can capture temporal dynamics but lack feature selection capabilities, while reinforcement learning compensates for the difficulty of traditional optimization methods in dealing with the uncertainties of complex ecosystems.

[0071] The improved random forest-transformer hybrid model's workflow involves processing the weighted feature vector output by the spatiotemporal feature fusion module. The input consists of spatial features such as slope and soil salinity, and temporal features such as temperature and precipitation. The input features are ranked by importance, and key features are selected (e.g., in arid areas, soil moisture is far more important than sunlight). Multiple decision trees are constructed, each trained based on a different subset of features, and a preliminary fitness score S(k,i,t) is output (k is the plant species, i is the grid, and t is the time). The dependency relationship between different spatiotemporal units is captured through a self-attention mechanism (e.g., the influence of upstream precipitation on downstream soil moisture). The preliminary score output by the random forest is modeled temporally to output the final plant fitness S(k,i,t), which ranges from [0,1]. A higher value indicates that the plant k has a better plant fitness in grid i and time t.

[0072] The reinforcement learning uses the following reward function: The reinforcement learning described in this invention optimizes the planting ratio through a reward function, where M is the Shannon diversity index. Let p(k,t) be the planting proportion of the k-th plant at time t, and S(k,i,t) be the suitability of the k-th plant at grid i and time t. , is the weighting coefficient, T is the total number of time steps, and K is the total number of plant species.

[0073] In this embodiment, multi-model collaborative decision-making uses the feature importance ranking of random forest to clarify the impact of each factor on the decision, and as the monitoring data is updated, the model can continuously adjust the scheme to adapt to environmental changes, so as to provide a precise solution for forest and grassland restoration.

[0074] S4: Dynamic Execution and Feedback Optimization: Based on the improved A* algorithm, the drone planting path is planned, sensor feedback data is received in real time, and the model parameters are updated through incremental learning.

[0075] Specifically, dynamic execution and feedback optimization, through a linkage mechanism of "precise execution - real-time monitoring - model iteration," transforms the planting plan output by multi-model collaborative decision-making into actual action, and continuously optimizes the decision-making model based on environmental feedback to ensure that the restoration effect adapts to the dynamically changing ecological environment. Its core logic is: based on an improved A* algorithm, it plans the drone planting path, efficiently executing the planting task while simultaneously capturing the vegetation growth status and environmental changes in real time through a sensor network; when the deviation between the actual effect and the expected result exceeds a threshold, it triggers incremental learning to update the model parameters, making subsequent decisions more realistic and solving the traditional problem of "disconnect between planning and the actual situation."

[0076] Traditional A* algorithm only plans paths with the objective of "shortest distance," while the improved A* algorithm introduces a "plant suitability penalty" in addition to distance. It prioritizes high-suitability grid areas, reducing invalid drone flights in low-suitability areas, while balancing path length and planting accuracy. Its objective function is: where dist(i,j) is the distance from grid i to grid j, λ is the suitability penalty coefficient, and S(k,i,t) is the suitability of the k-th plant at grid i and time t.

[0077] This invention collects post-planting data in real time through a sensor network (soil moisture and vegetation cover sensors) and regular drone inspections, including: The data includes: vegetation survival rate (e.g., number of surviving plants in a certain area 3 weeks after planting / total number of plants); soil moisture change (e.g., moisture difference between planted and unplanted areas); and vegetation growth rate (e.g., weekly plant height increase). When the vegetation survival rate in a certain area is lower than a preset threshold (e.g., 60%), or the soil moisture deviates from the suitable range (e.g., below the critical humidity for plant growth), it is judged as "effect deviation" and the model update process is triggered.

[0078] It is worth noting that traditional methods require retraining on all data for updates, which is time-consuming and inefficient. Incremental learning updates model parameters only with newly collected biased data (such as environmental characteristics and plant status in low-survival-rate areas), retaining historical training results, thus achieving "lightweight updates." The specific process is as follows: Extract the spatiotemporal characteristics of the biased area (such as soil salinity and precipitation data of raster i at time t).

[0079] By associating features with actual survival rates, an incremental training set is formed.

[0080] Fine-tuning the parameters of the random forest-Transformer hybrid model in the multi-model collaborative decision-making module using incremental training sets (such as adjusting the attention weights of the Transformer) makes the model's suitability assessment of the region more accurate.

[0081] The updated model outputs a new suitability S(k,i,t), providing a basis for subsequent optimization of planting schemes.

[0082] In this embodiment, dynamic execution and feedback optimization achieve precise execution and dynamic optimization of planting tasks. By improving the A* algorithm, planting efficiency is significantly improved, and incremental learning significantly improves the repair effect of low survival rate areas. It also effectively solves the problem of "disconnect between planning and actual environment".

[0083] S5: Data storage and visualization: Store data to a distributed database by timestamp and spatial coordinates, generate a prediction map of the repair effect and overlay it with real-time monitoring data for display.

[0084] Specifically, data storage and visualization ensure traceability and reusability through structured storage of the entire data process, while dynamic visualization technology transforms abstract data into intuitive graphs, achieving full-link transparency from "data to decision-making to results." Its core logic is: data is indexed by both timestamps and spatial coordinates, stored in a distributed database to support efficient management of massive amounts of data; a predictive model generates a repair effect prediction graph, which is overlaid with real-time monitoring data for display, helping decision-makers intuitively compare expected and actual results, adjust repair strategies promptly, and solve the problems of "dispersed and disordered" traditional data management and "static and singular" visualization.

[0085] The spatiotemporal dynamic repair scheme visualization map generation is based on the planting scheme output by the multi-model collaborative decision-making module, the path planning of the dynamic repair execution module, and real-time monitoring data. Through geographic information system (GIS) and dynamic visualization technology, a multi-dimensional map is generated, including: spatiotemporal suitability dynamic map, planting path and progress map, repair effect comparison map, and scheme optimization dynamic trajectory map. The spatiotemporal suitability dynamic map uses a grid as the unit and color gradients (e.g., blue-green-red) to represent the spatial distribution of plant suitability at different time points (e.g., monthly). Overlaying a timeline allows for a visual view of suitability changes with the seasons (e.g., the expansion of high suitability areas during the rainy season). The planting path and progress map uses dashed lines or arrows to mark the drone planting path, combined with different colors (e.g., yellow indicates completed, gray indicates pending) to display the planting progress of each area, updating the ratio of planted area to planned area in real time. The restoration effect comparison map shows the vegetation coverage grid map before planting on the left and the data from the same period after planting on the right. The restoration effect is visually displayed through color differences (e.g., brown for bare land, green for vegetation coverage), and a survival rate heatmap can be overlaid (red indicates low survival rate, blue indicates high survival rate). The scheme optimization dynamic trajectory map uses broken lines or bubbles to mark the adjustment trajectory of the planting ratio during the reinforcement learning iteration process (e.g., the process of a certain plant ratio increasing from 30% to 50%), assisting in the analysis of key nodes in scheme optimization.

[0086] Data storage and cloud synchronization employ a structured storage strategy with a dual-dimensional index of "timestamp-spatial coordinates." Data throughout the entire process is categorized and stored in a distributed database (such as HBase), and synchronized to the cloud (such as a private or public cloud platform) at a fixed frequency, ensuring data security and accessibility. Specifically, this includes four steps: data classification, storage structure, distributed database storage, and cloud synchronization.

[0087] The data is categorized into raw acquisition data (raw sensor values, UAV imagery), preprocessed data (cleaned raster data, spatiotemporal fusion features), decision data (suitability assessment results, planting plans), and feedback data (survival rate, soil change monitoring values). The storage structure consists of each data record containing a timestamp (accurate to the second), spatial coordinates (raster center point latitude and longitude), data type label, and feature value, for example, "2025-05-10 08:00:00, E112.3°N38.5°, soil moisture, 25%". The distributed database storage supports petabyte-level data storage, distributing data by spatial region (e.g., 1km×1km) using sharding technology to improve query efficiency (e.g., querying precipitation data for a certain area within 3 months can return within 10 seconds). Cloud synchronization uses an incremental synchronization mechanism, synchronizing only newly added or updated data to the cloud, with automatic backup every morning, supporting access from multiple terminals (computers, mobile devices) through permission verification, facilitating cross-team collaboration (e.g., researchers remotely accessing data, decision-makers viewing repair progress in real time).

[0088] In this embodiment, the visualization and storage module realizes "complex data visualization" and "full lifecycle data management". Decision-makers can grasp the overall restoration through the map. Distributed storage and cloud synchronization ensure that massive amounts of data during the restoration period can be efficiently queried and securely retained, providing data support for subsequent ecological restoration research and solution iteration.

[0089] In one embodiment, Figure 3 As shown, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements a forest and grassland ecological restoration method.

[0090] For specific limitations on electronic devices, please refer to the limitations on forest and grassland ecological restoration methods mentioned above, which will not be repeated here.

[0091] In one embodiment, a computer-readable medium is provided, the computer-readable medium storing a computer program, wherein when a processor executes the computer program, it performs the following steps: S1. Multi-source data acquisition and preprocessing: Static spatial data and dynamic time-series data are acquired through sensor networks and UAVs, and the data is preprocessed and converted into a raster format with a preset resolution; S2. Spatiotemporal Feature Fusion and Contribution Calculation: Construct a spatiotemporal feature matrix, weight the features through a spatiotemporal attention mechanism, and calculate the dynamic contribution of each feature in different spatiotemporal units based on an improved random forest; S3. Multi-model collaborative decision-making: The weighted features are input into the improved random forest-transformer hybrid model to obtain plant suitability, and the planting plan is optimized through reinforcement learning; S4. Dynamic Execution and Feedback Optimization: Based on the improved A* algorithm, the drone planting path is planned, sensor feedback data is received in real time, and the model parameters are updated through incremental learning; S5. Data storage and visualization: Data is stored in a distributed database by timestamp and spatial coordinates, and a repair effect prediction map is generated and overlaid with real-time monitoring data for display.

[0092] For specific limitations on computer-readable media, please refer to the limitations on forest and grassland ecological restoration methods mentioned above, which will not be repeated here.

[0093] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0094] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A forest and grassland ecological restoration system, characterized in that, Including: The multi-source data acquisition module is used to collect environmental data through sensor networks and drones. The collected data includes static spatial data and dynamic time-series data. The dynamic time-series data includes hourly recorded temperature, humidity, light intensity, and weekly recorded vegetation coverage. The spatiotemporal feature fusion module is used to preprocess static spatial data and dynamic temporal data. It extracts key features and calculates dynamic contribution through a spatiotemporal attention mechanism. The preprocessing includes removing outliers using the IQR method, smoothing temporal data using the sliding window method, and completing spatial data using Kriging interpolation, and uniformly converting it into a raster format with a preset resolution. The multi-model collaborative decision-making module is used to generate dynamic evaluation results of plant suitability and optimize planting plans through reinforcement learning; The dynamic repair execution module is used to plan the planting path of the drone and execute the planting task, receive sensor feedback data in real time, and update the model parameters through incremental learning; The visualization and storage module is used to generate a visualization map of the spatiotemporal dynamic repair scheme, store the data in a distributed database according to timestamps and spatial coordinates, and synchronize it to the cloud.

2. The forest and grassland ecological restoration system according to claim 1, characterized in that, In the spatiotemporal feature fusion module, the spatiotemporal attention mechanism weights the features using temporal attention weights and spatial attention weights, wherein the weighting formula is as follows: Where X(i,t) is the feature vector of grid cell i at time t, α(t) is the temporal attention weight, obtained by modeling the temporal data using LSTM, and β(i) is the spatial attention weight, calculated by a spatial autoencoder. This is the weighted eigenvector.

3. The forest and grassland ecological restoration system according to claim 1, characterized in that, The multi-model collaborative decision-making module uses an improved hybrid model of random forest and Transformer to output the fitness S(k,i,t) of the k-th plant at grid i and time t, calculated using the following formula: Where σ is the sigmoid function, W and b are model parameters, and Transformer is the Transformer model. This is the weighted eigenvector.

4. The forest and grassland ecological restoration system according to claim 1, characterized in that, The dynamic repair execution module performs the planting task using the objective function of the improved A* algorithm. The objective function is: , where dist(i,j) is the distance from grid i to grid j, λ is the fitness penalty coefficient, and S(k,i,t) is the fitness of the k-th plant at grid i and time t.

5. A method for forest and grassland ecological restoration, characterized in that, Includes the following steps: S1. Multi-source data acquisition and preprocessing: Static spatial data and dynamic time-series data are acquired through sensor networks and UAVs, and the data is preprocessed and converted into a raster format with a preset resolution; S2. Spatiotemporal Feature Fusion and Contribution Calculation: Construct a spatiotemporal feature matrix, weight the features through a spatiotemporal attention mechanism, and calculate the dynamic contribution of each feature in different spatiotemporal units based on an improved random forest; S3. Multi-model collaborative decision-making: The weighted features are input into the improved random forest-transformer hybrid model to obtain plant suitability, and the planting plan is optimized through reinforcement learning; S4. Dynamic Execution and Feedback Optimization: Based on the improved A* algorithm, the drone planting path is planned, sensor feedback data is received in real time, and the parameters of the random forest-transformer hybrid model are updated through incremental learning; S5. Data storage and visualization: Data is stored in a distributed database by timestamp and spatial coordinates, and a repair effect prediction map is generated and overlaid with real-time monitoring data for display.

6. The method according to claim 5, characterized in that, In step S2, the spatiotemporal feature matrix is ​​X(i,t)=[DEM(i), salinity(i),T(t),H(t),L(t),C(i,t)], where X(i,t) is the data of raster cell i at time t, DEM(i) is the digital elevation model data of raster i, salinity(i) is the soil salinity data of raster i, T(t) is the air temperature data at time t, H(t) is the humidity data at time t, L(t) is the light intensity data at time t, and C(i,t) is the vegetation cover data of raster i at time t.

7. The method according to claim 5, characterized in that, In step S2, the dynamic contribution is determined by the formula... Calculate, where D(i,t) is the feature dataset of raster i at time t, and Var(D) is the variance of dataset D. For the subset D k The variance of the feature is R(i,t), which is the reduction in variance of the feature in the spatiotemporal unit (i,t).

8. The forest and grassland ecological restoration method according to claim 5, characterized in that, In step S3, reinforcement learning uses a reward function. Optimize planting ratios, where M is the Shannon diversity index. Let p(k,t) be the planting proportion of the k-th plant at time t, and S(k,i,t) be the suitability of the k-th plant at grid i and time t. , is the weighting coefficient, T is the total number of time steps, and K is the total number of plant species.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the forest and grassland ecological restoration method according to any one of claims 5 to 8.

10. A computer-readable medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the forest and grassland ecological restoration method according to any one of claims 5 to 8.

Images