Cunninghamia lanceolata forest-oriented periodic tending scheduling method, device and equipment and storage medium

Abstract

The invention discloses a cedarwood forest-oriented periodic tending scheduling method, device and equipment and a storage medium, and relates to the technical field of intelligent forestry driven by artificial intelligence, and the method comprises the steps: optimizing a preset initial tending period through a tending period adjustment model based on LSTM training, and obtaining a target tending period; when the period is started, collecting multi-mode cedarwood forest data, and preprocessing the multi-mode cedarwood forest data to form target data; inputting the target data into a cunninghamia lanceolata forest growth prediction model constructed by CNN and PointNet, and outputting a breast diameter growth prediction value, a species richness prediction value and a tending standard reaching probability; and an equipment control strategy is generated, and intelligent felling equipment, an automatic replanting device or mechanical brush cutting equipment is scheduled to execute tending operation. According to the method, closed-loop scheduling from period setting, state sensing and intelligent decision making to automatic execution is realized, and periodic intelligent tending scheduling can be realized on the premise of avoiding excessive interference on the Chinese fir forest.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent forestry driven by artificial intelligence, in particular to a periodic tending scheduling method and device for Chinese fir forests, equipment and a storage medium. BACKGROUND

[0002] Chinese fir is a core fast-growing timber forest species in southern China. Periodic tending (thinning, brush cutting, and re-planting, etc.) is a key technical measure to ensure forest stand quality and timber yield. The current technical path for periodic tending scheduling of Chinese fir forests mainly falls into two categories: one is the traditional experience-driven technical mode, in which forestry personnel collect data such as Chinese fir diameter at breast height and stand density through field surveys, and then manually schedule equipment operation by determining tending parameters based on experience; the other is the preliminary intelligent exploration technical mode, in which remote sensing equipment such as laser radar and hyperspectral are introduced to collect forest stand data, and a simple time series analysis model is used to determine the tending time window, and a single intelligent device is used to carry out batch operation according to fixed parameters.

[0003] However, the traditional experience-driven technical mode has strong subjectivity, limited data collection coverage, and delayed scheduling response, and cannot accurately match the heterogeneous growth state of large-area Chinese fir forests. The preliminary intelligent exploration technical mode has the technical shortcomings of fixed scheduling logic, lack of forest stand difference zoning adaptation, and lack of dynamic adjustment mechanism, which cannot guarantee the accuracy and ecological compatibility of tending operation. Therefore, how to realize periodic intelligent tending scheduling without excessive interference with Chinese fir forests has become a problem to be solved. SUMMARY

[0004] The present application aims to provide a periodic tending scheduling method and device for Chinese fir forests, equipment and a storage medium, which aims to solve the technical problem of how to realize periodic intelligent tending scheduling without excessive interference with Chinese fir forests.

[0005] To achieve the above-mentioned purpose, the present application provides a periodic tending scheduling method for Chinese fir forests, which comprises: adjusting a preset initial tending period through a tending period adjustment model to obtain a target tending period, wherein the tending period adjustment model is obtained by training an LSTM model; collecting multi-modal Chinese fir forest data and preprocessing the multi-modal Chinese fir forest data to obtain target data when the target tending period is started, wherein the multi-modal Chinese fir forest data includes hyperspectral images, laser radar point clouds, and understory microenvironment data; inputting the target data into a Chinese fir forest growth prediction model to obtain diameter growth prediction values, species richness prediction values, and tending compliance probabilities, wherein the Chinese fir forest growth prediction model is constructed according to a CNN network and a PointNet network architecture; According to the breast diameter growth prediction value, the species richness prediction value and the tending up-to-standard probability, a device control strategy is generated; According to the device control strategy, intelligent cutting equipment, automatic replanting devices or mechanical brush cutting equipment are dispatched to perform tending operations.

[0006] In an embodiment, the Chinese fir forest growth prediction model comprises a spectral feature extraction branch, a structural feature extraction branch, a global average pooling layer, a maximum pooling layer, a dual-channel attention fusion layer, a fully connected prediction layer and a Softmax output layer; The step of inputting the target data into the Chinese fir forest growth prediction model to obtain a breast diameter growth prediction value, a species richness prediction value and a tending up-to-standard probability comprises: The target data is split into a spectral data subset and a point cloud data subset, which are respectively input into the spectral feature extraction branch and the structural feature extraction branch; The spectral data subset is subjected to convolution operation, batch normalization and activation processing by the CNN network of the spectral feature extraction branch, and after dimension reduction by the global average pooling layer, a spectral feature vector is output; The point cloud data subset is subjected to coordinate transformation and multi-layer perceptron feature learning by the PointNet network of the structural feature extraction branch, and after processing by the maximum pooling layer, a structural feature vector is output; The spectral feature vector and the structural feature vector are input into the dual-channel attention fusion layer to obtain attention weight coefficients thereof; The spectral feature vector and the structural feature vector are weighted and fused according to the attention weight coefficients to obtain a fused feature; The fused feature is input into the fully connected prediction layer, and after calculation by the Softmax output layer, a breast diameter growth prediction value and a species richness prediction value are obtained; According to the breast diameter growth prediction value, the species richness prediction value, a preset breast diameter growth threshold and a preset species richness reduction threshold, a comprehensive tending up-to-standard probability is calculated.

[0007] In an embodiment, the step of constructing the Chinese fir forest growth prediction model comprises: A spectral feature extraction branch is constructed, which adopts a CNN network comprising a preset number of convolution blocks, each of which comprises a Conv2D layer, a BatchNorm layer and a ReLU activation function in sequence, and a global average pooling layer is arranged at the end of the spectral feature extraction branch; The structure feature extraction branch is constructed by using a PointNet network architecture, which includes a T-Net coordinate transformation module and a multi-layer MLP network, and a max-pooling layer is arranged at the end of the structure feature extraction branch. According to the spectral feature extraction branch and the structure feature extraction branch, a dual-feature extraction branch architecture is constructed. After the dual-feature extraction branch architecture, a dual-channel attention fusion layer, a full connection prediction layer and a Softmax output layer are built to obtain a Chinese fir forest growth prediction model.

[0008] In an embodiment, the step of generating a device control strategy according to the diameter growth prediction value, the species richness prediction value and the tending compliance probability includes: Based on the terrain slope data derived from the laser radar point cloud and the understory microenvironment data, growth environment zoning is performed to obtain a zoning result including a growth zone, a transition zone and an edge sparse zone; In combination with the zoning result, the diameter growth prediction value and the species richness prediction value in each zone are respectively compared to determine the tending priority of each zone, wherein the tending priority of the growth zone is higher than that of the transition zone and the edge sparse zone; According to the interference factors in each zone and the intervention level corresponding to the tending compliance probability, the device type and the basic operation parameters of each zone are matched; According to the tending priority, the device type, the basic operation parameters and the device operation timing, a device control strategy with a zoning label is generated.

[0009] In an embodiment, the step of scheduling intelligent felling equipment, automatic replanting devices or mechanical brush cutting equipment for tending operation according to the device control strategy includes: The device control strategy is parsed to extract the device type, operation parameters, operation timing and tending priority corresponding to each zone; Based on the tending priority, the operation tasks of each zone are sorted to generate a zoning operation sorting table; According to the zoning operation sorting table, the device type and the operation parameters, a zoning operation scheduling instruction is generated; The zoning operation scheduling instruction is sent to the corresponding tending operation equipment, which includes intelligent felling equipment, automatic replanting devices and / or mechanical brush cutting equipment; When receiving the instruction confirmation signal fed back by the tending operation equipment, an operation start instruction is sent according to the zoning operation sorting table to schedule the tending operation equipment to perform tending operation according to the operation parameters.

[0010] In an embodiment, the training step of the Chinese fir forest growth prediction model comprises: Collecting Chinese fir forest hyperspectral images, point cloud data, meteorological monitoring data and Chinese fir forest field investigation data to construct a training data set; The training data set is labeled and feature selected, and divided into a training set and a validation set according to a predetermined proportion; The mean square error and cross entropy are combined as a loss function, and the Chinese fir forest growth prediction model is iteratively trained according to a second predetermined learning rate, an AdamW optimizer, a predetermined weight decay coefficient and the training set until the value of the loss function converges, obtaining a reference prediction model; The hyperparameters of the reference prediction model are adjusted using the cross-validation method, and the prediction accuracy of the reference prediction model is verified according to the validation set; When the prediction accuracy is less than a predetermined accuracy threshold, return to the step of combining the mean square error and the cross entropy as the loss function, and iteratively training the Chinese fir forest growth prediction model according to the second predetermined learning rate, the AdamW optimizer, the predetermined weight decay coefficient and the training set until the value of the loss function converges, obtaining a reference prediction model; When the prediction accuracy is greater than or equal to the predetermined accuracy threshold, the reference prediction model is used as the trained Chinese fir forest growth prediction model.

[0011] In an embodiment, the step of adjusting the predetermined initial tending cycle by the tending cycle adjustment model to obtain a target tending cycle comprises: Obtain historical growth time series data, understory vegetation annual difference data and historical tending cycle effect data of Chinese fir forest to form a cycle adjustment input data set; The cycle adjustment input data set is normalized to obtain standardized time series data; The standardized time series data is input into the tending cycle adjustment model to obtain an initial adjustment coefficient; The initial adjustment coefficient is corrected according to a predetermined understory vegetation annual difference constraint threshold to obtain a target adjustment coefficient; According to the target adjustment coefficient and the predetermined initial tending cycle, the target tending cycle is calculated.

[0012] In addition, in order to achieve the above purpose, the present application also provides a periodic tending scheduling device for Chinese fir forest, which comprises: The cycle adjustment module is used for adjusting the predetermined initial tending cycle by the tending cycle adjustment model to obtain a target tending cycle, and the tending cycle adjustment model is obtained by training the LSTM model; A data processing module is configured to collect multi-modal Chinese fir forest data when the target nurturing cycle is started, and to preprocess the multi-modal Chinese fir forest data to obtain target data, wherein the multi-modal Chinese fir forest data includes hyperspectral images, laser radar point clouds, and understory microenvironment data. A growth prediction module is configured to input the target data into a Chinese fir forest growth prediction model to obtain a diameter at breast height growth prediction value, a species richness prediction value, and a nurturing standard achievement probability, wherein the Chinese fir forest growth prediction model is constructed according to a CNN network and a PointNet network architecture. A strategy generation module is configured to generate a device control strategy according to the diameter at breast height growth prediction value, the species richness prediction value, and the nurturing standard achievement probability. A device scheduling module is configured to schedule intelligent felling devices, automatic replanting devices, or mechanical brush cutting devices to perform nurturing operations according to the device control strategy.

[0013] In addition, to achieve the above-mentioned purpose, the present application also proposes a periodic nurturing scheduling device for Chinese fir forest, which comprises a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the computer program is configured to implement the steps of the periodic nurturing scheduling method for Chinese fir forest as described above.

[0014] In addition, to achieve the above-mentioned purpose, the present application also proposes a storage medium, which is a computer-readable storage medium, and the storage medium stores a computer program, wherein the computer program is executed by a processor to implement the steps of the periodic nurturing scheduling method for Chinese fir forest as described above.

[0015] The one or more technical solutions proposed in the present application have at least the following technical effects: Firstly, the preset initial tending cycle is dynamically corrected by a tending cycle adjustment model, and a target tending cycle more suitable for the actual succession state of the stand is output, so as to avoid tending too early or lag caused by a fixed cycle; when the target tending cycle starts, high-spectrum images, laser radar point clouds and understory micro-environment data are collected synchronously and preprocessed to form target data with unified structure, providing high-quality multi-modal input for subsequent accurate prediction; then, the target data is input into a Chinese fir forest growth prediction model constructed by a CNN network and a PointNet network architecture, which jointly analyzes the crown spectral characteristics, three-dimensional stand structure and micro-environment information, and outputs the diameter growth prediction value, species richness prediction value and tending compliance probability, realizing comprehensive evaluation of the growth potential and ecological state of the stand; then, the equipment control strategy is generated according to the three prediction indexes, so that the tending measures can take into account the timber production and biodiversity goals; finally, the intelligent harvesting equipment, automatic replanting device or mechanical brush cutting equipment are dispatched according to the strategy to perform differentiated operations, ensuring accurate and efficient operation. The present application realizes a closed-loop scheduling from cycle setting, state sensing, intelligent decision-making to automatic execution, which can realize periodic intelligent tending scheduling while avoiding excessive disturbance to Chinese fir forests. BRIEF DESCRIPTION OF DRAWINGS

[0016] The accompanying drawings, which are incorporated into and form a part of the specification, illustrate an embodiment consistent with the present application and, together with the description, serve to explain the principles of the application.

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the accompanying drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, for those skilled in the art, other drawings can also be obtained based on these drawings without creative labor.

[0018] Figure 1 The flowchart provided by the present application for the periodic tending scheduling method for Chinese fir forests embodiment one; Figure 2 The flowchart provided by the present application for the periodic tending scheduling method for Chinese fir forests embodiment two; Figure 3 The module structure diagram of the periodic tending scheduling device for Chinese fir forests in the embodiment of the present application; Figure 4 The device structure diagram of the hardware running environment involved in the periodic tending scheduling method for Chinese fir forests in the embodiment of the present application.

[0019] The purpose realization, functional characteristics and advantages of the present application will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0020] It should be understood that the specific embodiments described herein are merely intended to explain the technical solutions of the present application, and are not intended to limit the present application.

[0021] In order to better understand the technical solutions of the present application, the following will be described in detail in combination with the drawings of the specification and specific embodiments.

[0022] It should be noted that the execution subject of the embodiments of the present application can be a computing service device with data processing, network communication and program running functions, such as a tablet computer, a personal computer, a mobile phone, etc., or an electronic device, a tending scheduling system, etc. capable of realizing the above functions. The following will take the tending scheduling system as an example to describe the embodiments and the following embodiments.

[0023] Based on this, the present application provides a periodic tending scheduling method for Chinese fir forest, which refers to Figure 1 , Figure 1 is a flowchart of the first embodiment of the periodic tending scheduling method for Chinese fir forest of the present application.

[0024] In the present embodiment, the periodic tending scheduling method for Chinese fir forest comprises steps S10-S50: Step S10, adjusting a preset initial tending period by a tending period adjustment model to obtain a target tending period, wherein the tending period adjustment model is obtained by training an LSTM model; Step S20, collecting multi-modal Chinese fir forest data when the target tending period is started, and pre-processing the multi-modal Chinese fir forest data to obtain target data, wherein the multi-modal Chinese fir forest data comprises hyperspectral images, laser radar point clouds and understory microenvironment data; Step S30, inputting the target data into a Chinese fir forest growth prediction model to obtain a diameter growth prediction value, a species richness prediction value and a tending standard reaching probability, wherein the Chinese fir forest growth prediction model is constructed according to a CNN network and a PointNet network architecture; Step S40, generating a device control strategy according to the diameter growth prediction value, the species richness prediction value and the tending standard reaching probability; Step S50, scheduling intelligent felling equipment, automatic replanting devices or mechanical brush cutting equipment to perform tending operations according to the device control strategy.

[0025] It should be noted that the tending cycle adjustment model refers to a time series prediction model constructed and trained based on a long short-term memory (LSTM) network, which is used to dynamically optimize the tending time interval of Chinese fir forests according to historical stand conditions and tending records. The preset initial tending cycle refers to the reference tending time interval set according to forestry regulations or expert experience before specific stand dynamic information is lacking, which is 5 years in this embodiment. The target tending cycle refers to the optimal tending time interval that better fits the actual growth state and environmental conditions of the current stand after being calculated and output by the tending cycle adjustment model. The LSTM model refers to a recurrent neural network (RNN) variant suitable for processing and predicting time series data, which can effectively capture long-term dependencies. The target data refers to the structured input data formed after standardization, denoising, registration and feature extraction of the collected multi-modal Chinese fir forest original data, which is used to supply the Chinese fir forest growth prediction model. The hyperspectral image refers to the remote sensing image data containing tens to hundreds of continuous narrow bands obtained by a hyperspectral imaging device, which can finely reflect the physiological state and biochemical component information of the Chinese fir canopy. The laser radar point cloud refers to a set of three-dimensional spatial coordinate points generated after a light detection and ranging (LiDAR) system scans the forest area, which is used to reconstruct the height, crown width and vertical structure of the forest. The understory microenvironment data refers to the local ecological environment parameters under the Chinese fir forest canopy, mainly including temperature, humidity, light intensity, etc., which are used to represent the influence of the local climate in the forest on the growth and regeneration of the trees. The Chinese fir forest growth prediction model refers to a multi-modal deep learning model constructed by fusing convolutional neural network (CNN) and PointNet (Point Network) network architecture, which is used to jointly analyze images, point clouds and environmental data to predict key ecological indicators of the stand.

[0026] The breast diameter growth prediction value refers to the estimation result of the model for the average breast diameter (usually the diameter of the trunk at 1.3 meters from the ground) of Chinese cypress in a future tending cycle (for example, 5 centimeters). The species richness prediction value refers to the prediction result of the model for the number of species in the understory plant community, used to measure the impact of tending measures on biodiversity. In this embodiment, the prediction of species richness (the total number of species in a specific area) is a trend-based and interval-based prediction based on feature correlation, and the prediction value takes the middle value of the interval (which is an estimated value), rather than an accurate numerical prediction. The tending compliance probability refers to the possibility that the current stand meets the preset management objectives (such as density control, growth rate, ecological function, etc.) at the end of the next tending cycle, and the value range is between 0 and 1. The CNN network refers to a convolutional neural network (CNN), a deep learning architecture that is good at processing grid-like data (such as images) and automatically extracts spatial features through convolutional layers. The PointNet network architecture refers to a deep neural network structure specifically designed for processing unordered three-dimensional point cloud data, which can directly learn geometric shape and spatial distribution features from point clouds. The device control strategy refers to a set of operation instructions automatically generated based on the prediction results, used to guide different intelligent tending devices to perform corresponding work tasks at specific spatiotemporal locations. The intelligent felling device refers to an automated logging machine with autonomous positioning, target recognition and decision-making capabilities, which can selectively fell target individual trees or areas according to the control strategy. The automatic replanting device refers to an integrated robotic device that integrates seedling transportation, positioning, hole opening and planting functions, and can automatically complete replanting operations based on stand gap information. The mechanical brush cutting device refers to an automated or remote-controlled work machine for removing competing vegetation such as understory shrubs and weeds, which can perform precise brush cutting operations in designated areas according to the control strategy. In this embodiment, as an implementation manner, manually operated felling devices, replanting devices or mechanical brush cutting devices can be used instead of intelligent felling devices, automatic replanting devices or mechanical brush cutting devices. When manually operated, the operation is also based on the device control strategy.

[0027] It can be understood that first, the tending scheduling system calls the trained tending cycle adjustment model, takes the preset initial tending cycle as input, calculates the target tending cycle that better fits the current stand succession rhythm, to avoid tending too early or too late due to fixed cycle. Second, when the target tending cycle reaches the starting time point, the tending scheduling system triggers the multi-source data acquisition task, synchronously acquires hyperspectral images, laser radar point clouds and understory microenvironment data covering the target forest area, and performs preprocessing procedures such as spatiotemporal alignment, radiation correction, point cloud filtering, missing value interpolation and normalization on these heterogeneous data, to generate target data with uniform format and complete features, providing high-quality input for subsequent accurate prediction.

[0028] Then, the tending scheduling system sends the target data into a Chinese fir forest growth prediction model, which extracts the crown physiological characteristics in the hyperspectral image using a CNN network, analyzes the three-dimensional stand structure represented by the laser radar point cloud through a PointNet network, and fuses the understory microenvironment data to jointly infer the diameter at breast height growth prediction value, species richness prediction value and tending compliance probability in the future period, so as to comprehensively evaluate the stand state and tending demand. Next, the tending scheduling system generates a specific device control strategy according to the three prediction results, combined with the preset management target (such as giving priority to timber growth, considering biodiversity or ensuring ecological function compliance), and determines which areas, when and with what intensity to perform logging, re-planting or hedge cutting operations. Finally, the tending scheduling system issues the device control strategy to the operation layer, automatically schedules intelligent logging equipment, automatic re-planting devices or mechanical hedge cutting equipment to the designated location to perform corresponding tending operations, realizes a closed-loop scheduling from period determination, state perception, intelligent prediction to precise execution, and ensures that tending intervention is both scientific and effective and minimizes unnecessary interference to the stand.

[0029] As an example, the step of generating a device control strategy according to the diameter at breast height growth prediction value, the species richness prediction value and the tending compliance probability includes: based on the terrain slope data derived from the laser radar point cloud and the understory microenvironment data, performing growth environment partitioning to obtain a partitioning result including a growth zone, a transition zone and an edge sparse zone; combining the partitioning result, respectively comparing the diameter at breast height growth prediction value and the species richness prediction value in each partition to determine the tending priority of each partition, wherein the tending priority of the growth zone is higher than that of the transition zone and the edge sparse zone; according to the interference factors in each partition and the intervention level corresponding to the tending compliance probability, matching the device type and the basic operation parameters of each partition; according to the tending priority, the device type, the basic operation parameters and the device operation timing sequence, generating a device control strategy with partition markers.

[0030] It should be noted that topographic slope data refers to the degree of forest surface inclination calculated using a Digital Elevation Model (DEM) based on lidar point clouds. This data is used to delineate areas of operational feasibility and ecological sensitivity, and is categorized according to preset thresholds into gentle slope areas (slope ≤ 5°), mild slope areas (slope 6°-15°), sloping slope areas (slope 16°-25°), steep slope areas (slope 26°-35°), rapid slope areas (slope 36°-45°), and dangerous slope areas (slope > 46°). The growth zone refers to a core area identified based on comprehensive topographic slope, light intensity ≥ 800 lux, and soil moisture within the 30%-60% range, indicating good growth of Chinese fir, stable stand structure, and high productivity potential. The transition zone refers to an intermediate area where environmental conditions partially meet suitable growth standards (e.g., light or humidity slightly deviates from the ideal range, or it is located in a gentle slope area), with moderate tree growth and a community structure in a dynamic adjustment phase. Sparse marginal areas refer to forest edges or degraded areas located on steep slopes, with insufficient sunlight, excessively dry or wet soil, or with significantly low forest coverage. They are typically characterized by restricted growth of Chinese fir, difficulty in regeneration, and sparse vegetation.

[0031] The zoning result refers to the geographic zoning map output by the tending and scheduling system after spatial clustering of the target forest area based on topographic slope data and understory microenvironment data. This map includes three semantic labels: growth zone, transition zone, and marginal sparse zone. In this embodiment, the zoning result can also be a geographic zoning map including semantic labels such as high-density zone, relatively dense zone, moderate zone, and sparse zone. A high-density zone refers to an area with a tree density significantly higher than the management standard (e.g., above 2500 trees / hectare), high canopy closure (usually >0.8), severely insufficient light within the forest, and intense competition, requiring thinning to alleviate growth inhibition. A relatively dense zone refers to an area with a tree density slightly higher than the ideal management density (e.g., 1800-2500 trees / hectare) and canopy closure between 0.6 and 0. Areas with a density between 8 and 8 exhibit moderate competitive pressure and can be lightly thinned or pruned to optimize the structure; moderately dense areas refer to areas with a tree density close to the target management density (e.g., 1200-1800 trees / hectare), moderate canopy closure (approximately 0.4-0.6), stable stand structure, and good growth, requiring only maintenance tending or no intervention for the time being; sparsely dense areas refer to areas with significantly low tree density (e.g., below 1200 trees / hectare), canopy closure <0.4, exhibiting forest gaps or poor regeneration, requiring priority for replanting or promoting natural regeneration.

[0032] Tending priority refers to the order of operations determined by a comprehensive assessment of predicted diameter at breast height (DBH) growth and species richness within each zone. Growth zones, due to their high growth potential and ecological value, are given the highest priority, followed by transition zones, and then sparsely populated marginal zones. Disturbance factors refer to the actual distribution of obstacles within the forest that affect equipment access or tending effectiveness, including deadwood density, the extent of fallen tree obstruction, and the degree of shrub entanglement, used to adjust equipment selection and operational paths. Intervention level refers to the tending response intensity category based on the probability of achieving tending targets. For example, "standard maintenance level" indicates that the stand is close to the target state and requires only minimal maintenance; "light intervention level" indicates that moderate thinning or clearing is needed to promote recovery. Equipment type refers to the type of operational equipment matched to the characteristics of different zones. For example, small-scale harvesting equipment is used in growth zones to protect seedlings, while efficient brush cutters are used in sparsely populated marginal zones to remove competing vegetation. Basic operational parameters refer to the operational control variables corresponding to the equipment type and intervention level, including specific performance indicators such as thinning intensity (e.g., 15%), brush cut height, and travel speed. Partition markers are spatial semantic identifiers embedded in the generated device control strategy to indicate the geographic partition type corresponding to each control command, ensuring that the device accurately performs differentiated operations according to regional attributes.

[0033] Understandably, the nurturing and scheduling system first classifies the ground points in the lidar point cloud, then uses filtering algorithms (such as cloth simulation filtering or progressive triangulation encryption) to separate the surface point cloud, and then constructs a continuous DEM based on the surface point cloud using interpolation methods (such as inverse distance weighting), and then calculates the terrain slope data of each grid cell, and divides the slope into gentle areas, mild slope areas, sloping slope areas, steep slope areas, rapid slope areas and dangerous slope areas according to preset thresholds. Subsequently, combined with the synchronously collected forest understory microenvironment data (including light intensity and soil moisture), spatial overlay analysis was performed according to preset rules: areas with gentle slopes (i.e., slope ≤ 15°) and simultaneously satisfying light intensity ≥ 800 lux and soil moisture between 30% and 60% were classified as growth zones; areas that only partially met the above conditions (such as slopes of 16°-25° but with suitable microenvironment, or slopes ≤ 15° but with light or moisture slightly deviating from the ideal range) were classified as transition zones; and the remaining areas that did not meet the core conditions (such as slopes ≥ 26°, insufficient light, excessively dry or wet soil, or significantly low forest cover) were classified as marginal sparse zones. This generated zoning results containing three semantic labels: growth zone, transition zone, and marginal sparse zone, to achieve refined tending zoning based on site conditions and ecological status. Secondly, the system extracts the predicted values ​​of diameter at breast height (DBH) growth and species richness for all sample points or grid units within each partition, calculates the partition mean and performs a weighted evaluation (e.g., DBH growth as the main objective and species richness as a constraint). If both indicators of a partition are significantly higher than those of other areas, a higher nurturing priority is given, and the priority of the growth zone is always set higher than that of the transition zone and the marginal sparse zone, so as to ensure that resources are given priority to high-potential areas.

[0034] Then, the system identifies the interfering factors in each zone (by detecting the spatial distribution density and coverage of dead trees, fallen trees, and shrubs through point cloud and image fusion). At the same time, it determines the required intervention level based on the success rate of the zone's tending. If the success rate is high (e.g., ≥0.85), it is designated as "standard maintenance level," requiring only low-intensity operations. If the success rate is moderate (e.g., 0.6-0.85), it is designated as "light intervention level," requiring moderate clearing or thinning. Based on this, the system matches the equipment type (e.g., in densely vegetated growth areas, small harvesting equipment with a small turning radius and low ground pressure is selected to avoid crushing; in sparsely vegetated edges, high-power mechanical brush cutters are selected to improve efficiency) and sets basic operating parameters (e.g., light intervention level corresponds to 15% thinning intensity and 50 cm brush cutting height, while standard maintenance level uses 5% thinning intensity and preservation of native shrubs and grasses). Finally, the system uses nurturing priority as the primary scheduling order, and combines the equipment type, basic operating parameters, and equipment operation paths and timing constraints (such as avoiding conflicts between multiple devices in the same sub-area) allocated to each zone. It binds each control command to the original zone label and outputs a structured, executable equipment control strategy with zone markings, ensuring that subsequent nurturing operations are both accurately adapted to the regional ecological characteristics and efficiently and orderly complete the full-cycle scheduling.

[0035] As an example, the step of scheduling intelligent logging equipment, automatic replanting devices, or mechanical brush cutters for tending operations according to the equipment control strategy includes: parsing the equipment control strategy and extracting the equipment type, operation parameters, operation sequence, and tending priority corresponding to each zone; sorting the operation tasks of each zone based on the tending priority to generate a zone operation sorting table; generating a zone operation scheduling instruction according to the zone operation sorting table, the equipment type, and the operation parameters; sending the zone operation scheduling instruction to the corresponding tending operation equipment, which includes intelligent logging equipment, automatic replanting devices, and / or mechanical brush cutters; and when receiving an instruction confirmation signal from the tending operation equipment, sending an operation start instruction according to the zone operation sorting table to schedule the tending operation equipment to perform tending operations according to the operation parameters.

[0036] It should be noted that operational parameters refer to the quantitative control indicators used by silvicultural equipment when performing specific silvicultural tasks. These include, but are not limited to, thinning intensity, shrub clearing height, replanting density (e.g., 800 trees per hectare), travel speed, working depth, or robotic arm working angle. These parameters are used to precisely control the intensity and method of silvicultural operations to ensure compliance with forest stand management objectives. Operational sequence refers to the temporal arrangement or execution window of silvicultural tasks in each zone. This includes the relative order of task initiation (e.g., tasks in growing areas take precedence over those in sparsely populated areas), designated operation time periods (e.g., "shrub clearing in gentle slope areas from 9:00 AM to 11:30 AM"), or temporal constraints with other tasks (e.g., replanting must be carried out within 24 hours of logging completion). These parameters are used to coordinate multi-equipment collaborative operations, avoid resource conflicts, and adapt to changes in the forest environment (e.g., avoiding rainy days or periods of high temperatures).

[0037] The zonal task sequencing table is a task sequence list generated by the mulching scheduling system after arranging all pending tasks in ascending or descending order according to the mulching priority of each zone. It clearly records the task order, equipment type, and associated parameters for each zone, ensuring that high-priority areas are processed first. The zonal task scheduling instruction is a structured control command generated by the system based on the zonal task sequencing table, equipment type, and task parameters, targeting specific equipment. It includes detailed execution information such as the target zone coordinate range, operation type (e.g., logging, replanting, or brush clearing), task intensity, travel path, and safety constraints. The instruction confirmation signal is a digital communication signal returned by the mulching operation equipment to the mulching scheduling system after receiving the zonal task scheduling instruction, completing self-checks, path planning, and status readiness assessments, indicating "received and ready to execute." This confirms that the equipment has the conditions to execute and prevents misoperation due to instruction loss or equipment malfunction. The task start instruction is a formal execution trigger command issued by the mulching scheduling system according to the zonal task sequencing table after receiving the instruction confirmation signal from the corresponding mulching operation equipment. It instructs the equipment to carry out actual mulching operations within the designated zone according to preset task parameters.

[0038] Understandably, the afforestation scheduling system first parses the equipment control strategy field by field, extracting the structured equipment type, operation parameters, operation sequence, and afforestation priority value bound to each partition. Second, the system uses afforestation priority as the primary sorting criterion, dynamically ranking all tasks in all pending partitions globally—higher priority partitions are placed at the front of the queue; if priorities are the same, they are supplemented by operation sequence or equipment readiness status, generating a strictly ordered partition operation sorting table. The aim is to ensure that high-potential areas receive afforestation intervention first, avoiding resource misallocation or operation conflicts. Then, the system iterates through each task in the sorting table, combining its specified equipment type and operation parameters to generate a partition operation scheduling instruction containing the target geographical range, operation command code, operation intensity setting, safety obstacle avoidance rules, and estimated operation duration. These instructions are then pushed to the corresponding afforestation operation equipment, including intelligent logging equipment, automatic replanting devices, and / or mechanical brush cutters, via a wireless communication network. When a piece of equipment successfully receives the instruction and completes internal path planning, mechanical self-checking, and environmental perception, it sends a digital instruction confirmation signal back to the afforestation scheduling system. Finally, after receiving the confirmation signal, the system sends an encrypted operation start command to the device according to the order in the partition operation sorting table, triggering it to accurately perform tending operations such as logging, replanting, or shrub clearing in the designated partition according to the preset operation parameters. This achieves a closed-loop scheduling mechanism of task issuance, status confirmation, and orderly start, ensuring safe, efficient, and ecologically sound operations.

[0039] As an example, the training steps of the nurturing cycle adjustment model include: collecting multi-source time-series sample data of Chinese fir forests of different age groups, wherein the multi-source time-series sample data includes sample growth data, sample vegetation difference data, sample nurturing cycle and sample nurturing effect data; cleaning and interpolating the multi-source time-series sample data to obtain target sample data; constructing an initial LSTM model based on the basic LSTM model architecture, the preset number of neurons in the LSTM layer, the preset number of hidden layers, and the fully connected output layer; using the mean squared error as the loss function, iteratively training the initial LSTM model according to the first preset learning rate, the Adam optimizer, and the target sample data to obtain the nurturing cycle adjustment model.

[0040] It should be noted that different age groups of Chinese fir forests refer to multiple plantation communities of Chinese fir classified by age, divided into commercial forests and public welfare forests. Commercial forests include age groups such as young forests (e.g., 0-10 years), middle-aged forests (e.g., 11-20 years), near-mature forests (e.g., 21-25 years), mature forests (e.g., 26-35 years), and over-mature forests (≥36 years). Public welfare forests include age groups such as young forests (e.g., 0-15 years), middle-aged forests (e.g., 16-25 years), near-mature forests (e.g., 26-30 years), mature forests (e.g., 31-40 years), and over-mature forests (≥41 years), used to cover the tending response characteristics throughout the entire growth cycle of Chinese fir. Multi-source time-series sample data refers to a long-term, continuous collection of serialized data from various observation methods and recording systems from different age groups of Chinese fir forests, organized chronologically, used to reflect the dynamic evolution of stand status over time and with tending interventions. Sample growth data refers to the time series of tree growth indicators recorded in each age group of Chinese fir forest during various monitoring sessions, including quantitative observations such as diameter at breast height (DBH), tree height, crown width, single-tree volume, or stand volume. Sample vegetation difference data refers to time-series information obtained through remote sensing or ground surveys that characterizes changes in understory vegetation composition or community structure, such as NDVI (Normalized Difference Vegetation Index), Shannon diversity index, shrub and grass cover, or records of dominant species succession, used to characterize the impact of tending on the ecosystem. Sample tending cycle refers to the time interval records of actual tending operations performed on each age group of Chinese fir forest in history (e.g., if a Chinese fir forest is thinned in the 3rd, 8th, and 13th years, then its sample tending cycle is 5 years), serving as one of the input labels for the model to learn "when to tend." Sample tending effect data refers to the stand response results observed within a certain period of time after each tending operation, including growth acceleration rate, mortality rate change, degree of improvement in stand evenness or ecological function recovery indicators, etc., which are used to evaluate the effectiveness of tending measures.

[0041] The target sample data refers to a complete, consistent, and standardized time-series dataset that can be used for model training, formed after cleaning and interpolation processes such as missing value imputation, outlier removal, time alignment, and normalization of the original multi-source time-series sample data. The basic LSTM model architecture refers to the LSTM computational unit structure composed of input gates, forget gates, output gates, and cell states, possessing the basic ability to handle time-series dependencies. The preset number of neurons in each LSTM layer refers to the number of memory units (i.e., neurons) manually set before model construction; in this embodiment, it is 128, used to control the model's expressive power and computational complexity. The preset number of hidden layers refers to the total number of recurrent layers stacked in the LSTM part; in this embodiment, it is 2 layers, determining the model's depth of abstraction of time-series features. The fully connected output layer refers to one or more densely connected neural network layers following the LSTM layers, ultimately outputting a scalar value (i.e., the adjusted nurturing cycle), used to map high-dimensional time-series features to specific nurturing cycle prediction results. The initial LSTM model refers to the untrained, raw neural network model built from the basic LSTM model architecture, a preset number of neurons in the LSTM layers, a preset number of hidden layers, and a fully connected output layer. The first preset learning rate is the optimization step size parameter set in the early stages of training (e.g., 0.001), used to control the magnitude of the Adam optimizer's updates to the model weights in each iteration. The Adam optimizer is a stochastic gradient descent optimization algorithm based on adaptive moment estimation. It can automatically adjust the learning rate of each parameter based on historical gradients, and has the advantages of fast convergence and insensitivity to hyperparameters.

[0042] Understandably, firstly, the tending scheduling system extracts historical records from multiple representative Chinese fir forest plots, organized annually or at fixed monitoring intervals. It then meticulously compiles the sample growth data, vegetation difference data, actual tending cycles, and subsequent observed tending effect data for each plot at various time points. This constructs a multi-source time-series sample dataset covering the complete growth stage and tending response process, aiming to provide the model with sufficient and diverse learning support. Secondly, the system performs standardized preprocessing on this raw data: unifying the timelines of all plots, filling in missing yearly observations using linear interpolation or collaborative interpolation based on neighboring plots, removing obviously abnormal outliers (such as negative diameter at breast height growth), and normalizing variables of different dimensions. Finally, it generates structurally aligned and numerically reliable target sample data to improve the stability and generalization ability of the model training.

[0043] Then, the system builds a neural network according to the basic LSTM model architecture, specifically setting the number of neurons in the LSTM layer to 128 and the number of stacked hidden layers to 2. A fully connected output layer is added at the end to output the predicted value for a single tending cycle, thus completing the initial LSTM model construction. Finally, the system uses mean squared error as the loss function, selects the first preset learning rate and the Adam optimizer, and divides the target sample data into training and validation sets. Through multiple rounds of iterative training, the model weights are continuously adjusted—each round inputs a piece of historical time-series data, the model outputs a predicted tending cycle, the system calculates the squared error between this predicted cycle and the actual sample tending cycle, and uses the Adam optimizer to automatically adjust the learning step size for backpropagation updates until the validation loss no longer decreases significantly (or the decrease is less than a certain preset value). Ultimately, a tending cycle adjustment model is obtained that can dynamically recommend the optimal tending time based on the current stand status, thereby avoiding insufficient or excessive tending caused by relying on a fixed cycle.

[0044] As an example, the step of adjusting a preset initial tending cycle using a tending cycle adjustment model to obtain a target tending cycle includes: acquiring historical growth time-series data of Chinese fir forests, annual difference data of understory vegetation, and historical tending cycle effect data to form a cycle adjustment input dataset; normalizing the cycle adjustment input dataset to obtain standardized time-series data; inputting the standardized time-series data into the tending cycle adjustment model to obtain initial adjustment coefficients; correcting the initial adjustment coefficients according to a preset annual difference constraint threshold for understory vegetation to obtain target adjustment coefficients; and calculating the target tending cycle based on the target adjustment coefficients and the preset initial tending cycle.

[0045] It should be noted that historical growth time-series data for Chinese fir forests refers to a sequence of tree growth indicators that are regularly observed and recorded over multiple consecutive years for a specific Chinese fir stand. These include annual average diameter at breast height (DBH), tree height, crown width, or single-tree volume, arranged chronologically to reflect the stand's own growth dynamics. Annual understory vegetation variation data refers to quantitative indicators obtained through remote sensing imagery or ground quadrat surveys that characterize the degree of change in understory plant communities between adjacent years. Examples include the interannual NDVI variation rate, Shannon diversity index difference, or the increase or decrease in shrub and grass cover. These are used to measure ecosystem stability or the impact of tending disturbances. Historical tending cycle effect data refers to records of stand response results observed after each past tending operation, including growth increments, mortality rates, changes in stand density, or the degree of ecological function recovery within 1-3 years after tending. This data is used to assess the rationality of previous tending cycle settings. The periodic adjustment input dataset refers to the set of multivariate time series data formed by aligning the historical growth time series data of the Chinese fir forest, the annual difference data of understory vegetation, and the historical tending cycle effect data along the same time axis, and using it as the original input of the tending cycle adjustment model.

[0046] Standardized time-series data refers to the normalization of each feature dimension in the periodically adjusted input dataset (e.g., min-max scaling to the [0,1] interval or Z-score standardization) to eliminate dimensional differences, making data from different sources comparable and facilitating stable model training and inference. The initial adjustment coefficient is a dimensionless value calculated by the periodic adjustment model after receiving standardized time-series data, through its internal LSTM layer extracting long-term temporal dependencies and then through a fully connected output layer. It is used for initial scaling of the preset initial periodic adjustment cycle. The preset annual variation constraint threshold for understory vegetation refers to a human-set upper limit for tolerance of interannual variation in understory vegetation (e.g., 5%). When the annual variation data of understory vegetation exceeds this threshold, it indicates that the ecosystem may be significantly disturbed, requiring a conservative correction of the initial adjustment coefficient to avoid over-nurturing. The target adjustment coefficient refers to the final adjustment factor obtained by combining the initial adjustment coefficient with the preset annual difference constraint threshold of forest understory vegetation for logical judgment and numerical correction. For example, if the annual difference of forest understory vegetation is ≤5%, the initial adjustment coefficient is retained; if it is >5%, it is reduced towards 1.0 to reduce the tending frequency, thereby balancing the needs of production and ecological protection.

[0047] Understandably, firstly, the tending scheduling system extracts historical growth time-series data, annual understory vegetation difference data, and historical tending cycle effect data of the target Chinese fir forest from the monitoring database by year. These three types of data are then time-aligned and horizontally concatenated using the year as the key to form a multi-channel time series, constituting the periodic adjustment input dataset to comprehensively reflect the historical trajectory of stand status and tending response. Secondly, the system performs min-max normalization on each feature column in the periodic adjustment input dataset, linearly mapping the original values ​​to the 0-1 interval to obtain standardized time-series data. This eliminates the interference of differences in the dimensions and numerical ranges of different indicators on model inference, improving prediction stability.

[0048] Then, the system inputs the standardized time-series data into a pre-trained tending cycle adjustment model. The model internally processes the sequence information step-by-step using LSTM layers and extracts long-term dependencies, ultimately outputting a scalar value as the initial adjustment coefficient. Next, the system reads the latest annual variation data of understory vegetation and determines whether its absolute value exceeds a preset annual variation constraint threshold. If it does, the initial adjustment coefficient is compressed towards 1.0 by a fixed proportion (e.g., multiplied by 0.9). If it does not exceed the threshold, it remains unchanged, thus obtaining the target adjustment coefficient. This correction aims to suppress over-adjustment during periods of significant vegetation fluctuation and avoid ecological disturbance caused by frequent tending. Finally, the system multiplies the target adjustment coefficient by the preset initial tending cycle to calculate the target tending cycle, which is then used as the time basis for starting the next stage of tending, realizing the transformation of cycle setting from "empirical fixed" to "dynamic adaptation."

[0049] As an example, the construction steps of the Chinese fir forest growth prediction model include: constructing a spectral feature extraction branch, which adopts a CNN network containing a preset number of convolutional blocks, each of which consists of a Conv2D layer, a BatchNorm layer, and a ReLU activation function in sequence, and a global average pooling layer is set at the end of the spectral feature extraction branch; constructing a structural feature extraction branch, which adopts a PointNet network architecture, which includes a T-Net coordinate transformation module and a multi-layer MLP network, and a max pooling layer is set at the end of the structural feature extraction branch; constructing a dual feature extraction branch architecture based on the spectral feature extraction branch and the structural feature extraction branch; and building a dual-channel attention fusion layer, a fully connected prediction layer, and a Softmax output layer after the dual feature extraction branch architecture to obtain the Chinese fir forest growth prediction model.

[0050] It should be noted that the spectral feature extraction branch refers to the sub-network used to automatically extract spectral features related to the physiological state of Chinese fir forests from hyperspectral images. Its core structure is a CNN network containing multiple convolutional blocks, ultimately outputting a fixed-dimensional spectral feature vector. The preset number refers to the number of convolutional blocks manually set before model construction; in this embodiment, it is 4, used to control the depth and complexity of spectral feature extraction. The Conv2D layer refers to a two-dimensional convolutional layer (Convolutional2D Layer), which extracts local features in the spatial-spectral dimension of the hyperspectral image by sliding convolutional kernels. It is the basic unit for feature perception in CNN networks. The BatchNorm layer refers to a batch normalization layer, used to normalize the mean and variance of the feature maps output by the Conv2D layer, accelerating training convergence and improving model stability. The ReLU activation function refers to a rectified linear unit, used to introduce non-linear expressive power while avoiding the gradient vanishing problem. Global average pooling refers to averaging all spatial locations of each feature channel, compressing a feature map of arbitrary size into a 1×1 scalar, thereby generating a fixed-dimensional spectral feature vector that is independent of the input size.

[0051] The structural feature extraction branch is a sub-network specifically designed for processing LiDAR point cloud data and extracting 3D forest stand structure information. Built on the PointNet network architecture, it ultimately outputs a fixed-dimensional structural feature vector. The T-Net (Transformation Network) coordinate transformation module is a small neural network sub-module within PointNet used for affine transformations of the input point cloud. It learns a 3×3 transformation matrix to align or normalize the point cloud coordinates, improving the model's robustness to rotation and translation. The multi-layer MLP network is a feedforward neural network composed of multiple fully connected layers (Multi-Layer Perceptrons), used to abstract the high-dimensional feature representation of each point in the point cloud layer by layer. The max-pooling layer is an operation at the end of PointNet that maximizes the feature vectors of all points according to their dimensions, achieving "symmetric function" aggregation and outputting a global structural feature vector independent of the point cloud order. The dual feature extraction branch architecture refers to the overall network skeleton that combines the spectral feature extraction branch and the structural feature extraction branch in parallel, processing hyperspectral images and LiDAR point clouds independently, preserving the specific information of the two modalities. A dual-channel attention fusion layer is a weighted fusion mechanism that dynamically adjusts the contribution ratio of spectral and structural feature vectors in the fusion process through learnable attention weights, allowing the model to focus more on modal information that is more important to the current prediction task. A fully connected prediction layer is a network component consisting of one or more densely connected layers, used to map the fused feature vectors to the intermediate representation space required for the prediction task, preparing for the final output. A softmax output layer is a softmax activation function layer added after the fully connected prediction layer, transforming the output into a probability distribution.

[0052] Understandably, the process begins with two main steps. First, a spectral feature extraction branch is constructed by stacking convolutional blocks according to a predetermined number. Each convolutional block is connected strictly in the order of Conv2D layer → BatchNorm layer → ReLU activation function. A global average pooling layer is added at the end of this branch to compress the feature map extracted from the hyperspectral image through multiple convolutions into a fixed-length spectral feature vector, thus eliminating input image size differences and preserving key spectral response information. Second, a structural feature extraction branch is constructed. The LiDAR point cloud is first input into the T-Net coordinate transformation module for spatial affine correction, then fed into a feature encoder composed of a multi-layer MLP network to extract geometric features point by point. Finally, a max pooling layer is used to globally aggregate all point features, generating a structural feature vector independent of the point cloud arrangement, thereby effectively capturing the three-dimensional spatial structure of the forest stand. Then, these two branches are set up in parallel to form a dual feature extraction branch architecture, ensuring that hyperspectral and point cloud data are processed independently in their respective optimal paths, avoiding modal interference. Finally, a dual-channel attention fusion layer is connected to the dual-branch output end. The fusion ratio of spectral and structural features is dynamically adjusted through learnable weights. The information is then integrated by a fully connected prediction layer, and the Softmax output layer generates prediction results such as the probability of achieving the cultivation target. This allows the model to make full use of the complementarity of multimodal data and accurately output the predicted values ​​of diameter at breast height (DBH), species richness, and cultivation target achievement for the Chinese fir forest cultivation scenario.

[0053] This embodiment provides a periodic tending scheduling method for Chinese fir forests. First, a tending cycle adjustment model dynamically corrects the preset initial tending cycle, outputting a target tending cycle that better reflects the actual succession state of the forest stand, thus avoiding premature or delayed tending due to a fixed cycle. When the target tending cycle begins, hyperspectral images, lidar point clouds, and understory microenvironment data are simultaneously collected and preprocessed to form structurally unified target data, providing high-quality multimodal input for subsequent accurate predictions. Then, the target data is input into a Chinese fir forest growth prediction model constructed using a CNN and PointNet network architecture. This model jointly analyzes canopy spectral characteristics, three-dimensional stand structure, and microenvironment information, outputting predicted values ​​for diameter at breast height (DBH), species richness, and the probability of achieving the tending target, achieving a comprehensive assessment of the stand's growth potential and ecological status. Next, an equipment control strategy is generated based on these three prediction indicators, ensuring that tending measures balance timber production and biodiversity goals. Finally, based on this strategy, intelligent logging equipment, automatic replanting devices, or mechanical brush cutters are scheduled to perform differentiated operations, ensuring precise and efficient operation. This embodiment realizes a closed-loop scheduling from cycle setting, state perception, intelligent decision-making to automatic execution, which can achieve periodic intelligent tending scheduling while avoiding excessive interference with Chinese fir forests.

[0054] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in Embodiment 1 above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 2 , Figure 2 This is a flowchart illustrating the second embodiment of the periodic tending and scheduling method for Chinese fir forests according to this application. The Chinese fir forest growth prediction model includes a spectral feature extraction branch, a structural feature extraction branch, a global average pooling layer, a max pooling layer, a dual-channel attention fusion layer, a fully connected prediction layer, and a Softmax output layer. Step S30 of the periodic tending and scheduling method for Chinese fir forests includes steps S31 to S37: Step S31: Split the target data into a subset of spectral data and a subset of point cloud data, and input them into the spectral feature extraction branch and the structural feature extraction branch respectively; Step S32: The spectral data subset is subjected to convolution, batch normalization and activation processing through the CNN network of the spectral feature extraction branch, and after dimensionality reduction by the global average pooling layer, the spectral feature vector is output. Step S33: The PointNet network of the structural feature extraction branch performs coordinate transformation and multilayer perceptron feature learning on the subset of point cloud data, and outputs the structural feature vector after processing by the max pooling layer. Step S34: Input the spectral feature vector and the structural feature vector into the dual-channel attention fusion layer to obtain the attention weight coefficients of the two. Step S35: The spectral feature vector and the structural feature vector are weighted and fused according to the attention weight coefficients to obtain the fused features; Step S36: Input the fused features into the fully connected prediction layer, and after calculation by the Softmax output layer, obtain the predicted value of diameter at breast height growth and the predicted value of species richness. Step S37: Calculate the probability of achieving the overall breeding target based on the predicted value of diameter at breast height (DBH) growth, the predicted value of species richness, the preset DBH growth threshold, and the preset species richness reduction threshold.

[0055] It should be noted that the spectral data subset refers to the hyperspectral image portion separated from the target data, containing pixel matrices of multiple consecutive bands, and serves as the input to the spectral feature extraction branch. The point cloud data subset refers to the lidar point cloud portion separated from the target data, consisting of a large number of points with three-dimensional coordinates (x, y, z), and serves as the input to the structural feature extraction branch. The spectral feature vector is a fixed-dimensional one-dimensional vector output by the spectral feature extraction branch after sequentially performing convolution, batch normalization, ReLU activation, and global average pooling on the spectral data subset, used to characterize the physiological and biochemical state of the Chinese fir canopy. The structural feature vector is a fixed-dimensional one-dimensional vector output by the structural feature extraction branch after processing the point cloud data subset through T-Net coordinate transformation, multi-layer MLP feature learning, and max pooling, used to characterize the three-dimensional spatial structural features of the forest stand, such as tree height, canopy distribution, and stand density.

[0056] Attention weight coefficients refer to two scalar weights (usually summed to 1) adaptively calculated by the dual-channel attention fusion layer based on the contents of the spectral feature vector and structural feature vector, respectively representing the relative importance of spectral and structural information in the current prediction task. Fusion features are weighted composite vectors obtained by multiplying the spectral and structural feature vectors by their corresponding attention weight coefficients and then summing them, preserving multimodal complementary information while highlighting the role of the dominant mode. Predicted diameter at breast height (DBH) growth refers to the model's numerical prediction of the average annual DBH growth of Chinese fir trees in the next tending cycle (unit: cm / year), reflecting timber production potential. Predicted species richness refers to the model's prediction of the number of species that can be maintained or are expected to appear in the understory plant community (unit: species), used to measure the level of biodiversity. Preset DBH growth threshold refers to an artificially set minimum acceptable standard for annual DBH growth of Chinese fir trees (e.g., 0.5 cm / year); values ​​below this indicate limited stand growth, requiring intervention. The preset species richness reduction threshold refers to the maximum allowable decline in the number of understory species (e.g., 3 species) set by humans. If the predicted species richness decreases more than this threshold from the current level, it is considered an increased risk of ecological function damage.

[0057] Understandably, firstly, the nurturing scheduling system splits the target data into spectral data subsets and point cloud data subsets according to data type, and sends them to the spectral feature extraction branch and the structural feature extraction branch respectively to achieve parallel processing of multimodal data. Secondly, in the spectral feature extraction branch, the system sequentially performs two-dimensional convolution, batch normalization, and ReLU activation operations on the spectral data subsets to extract canopy spectral response features layer by layer, and calculates the average of the spatial dimensions of each feature channel through a global average pooling layer to output a fixed-dimensional spectral feature vector.

[0058] Then, in the structural feature extraction branch, the system first inputs a subset of point cloud data into the T-Net coordinate transformation module. This module predicts a 3×3 affine transformation matrix through a small MLP network and multiplies the original point cloud coordinates by this matrix to achieve spatial alignment. The purpose is to eliminate rotation / translation differences caused by the acquisition angle or forest stand orientation, thereby improving the stability of structural features. The aligned point cloud is then fed into a multi-layer MLP network to extract high-dimensional features point by point. Finally, a max pooling layer is used to take the maximum value of the feature dimension at all points to generate a structural feature vector that is independent of the point order. Next, the system concatenates the spectral feature vector and the structural feature vector and inputs them into a dual-channel attention fusion layer. This layer first performs linear transformations on the two vectors and calculates their similarity to the learnable query vector. Then, it is normalized by the Softmax function to obtain the attention weight coefficients of the two modalities (e.g., spectral weight 0.65, structural weight 0.35), thereby dynamically measuring which modal information is more reliable in the current task.

[0059] Subsequently, the system multiplies the two weighting coefficients by their corresponding feature vectors and adds them together to obtain the fused features. Then, the fused features are fed into a fully connected prediction layer for nonlinear mapping, and the predicted values ​​for diameter at breast height (DBH) growth and species richness are output through a Softmax output layer. Finally, the system uses weighted logic to determine the overall success probability of nurturing based on whether the predicted DBH growth is greater than or equal to a preset DBH growth threshold and whether the decrease in the predicted species richness compared to the current value is less than a preset species richness decrease threshold. The probability is calculated as follows: if both criteria are met, the probability is above 0.9; if only one criterion is met, the probability is 0.6-0.8; and if neither criterion is met, the probability is below 0.4. This probability is then used for subsequent decision-making.

[0060] As an example, the training steps of the Chinese fir forest growth prediction model include: collecting hyperspectral images, point cloud data, meteorological monitoring data, and field survey data of the Chinese fir forest to construct a training dataset; labeling and feature-selecting the training dataset, and dividing it into a training set and a validation set according to a preset ratio; using the mean squared error and cross-entropy as a loss function, iteratively training the Chinese fir forest growth prediction model according to a second preset learning rate, AdamW optimizer, preset weight decay coefficient, and the training set until the value of the loss function converges to obtain a reference prediction model; and adjusting the training dataset using cross-validation. The hyperparameters of the reference prediction model are described, and the prediction accuracy of the reference prediction model is verified according to the validation set. When the prediction accuracy is less than a preset accuracy threshold, the process returns to the step of iteratively training the Chinese fir forest growth prediction model using a loss function consisting of mean squared error and cross-entropy, according to a second preset learning rate, AdamW optimizer, preset weight decay coefficient, and the training set, until the value of the loss function converges to obtain the reference prediction model. When the prediction accuracy is greater than or equal to the preset accuracy threshold, the reference prediction model is taken as the trained Chinese fir forest growth prediction model.

[0061] It should be noted that meteorological monitoring data refers to environmental and climatic parameters closely related to stand growth that are continuously recorded during the growth period of Chinese fir forests, including average daily temperature, cumulative precipitation, relative humidity, solar radiation, and wind speed, used to reflect the impact of external climatic conditions on tree growth and vegetation dynamics. Field survey data of Chinese fir forests refers to ground-based measured information obtained through manual plot surveys or long-term monitoring of fixed plots, including individual tree diameter at breast height (DBH), tree height, crown width, stand density, understory plant species and quantity, and litter thickness, serving as true labels or auxiliary features for model training. The preset ratio refers to the ratio of training set to validation set samples (e.g., 8:2) set manually before splitting the dataset, used to ensure sufficient model training and reliable validation results. The second preset learning rate refers to the initial learning step size parameter of the optimizer set during model training (e.g., 0.0005), controlling the magnitude of model weight updates, and is usually smaller than the first preset learning rate to achieve finer convergence.

[0062] The AdamW optimizer is an improved adaptive moment estimation optimization algorithm (Adam with Weight Decay). It explicitly decouples the weight decay term from Adam, more effectively preventing overfitting and improving generalization performance. It is suitable for training deep neural networks with regularization. The preset weight decay coefficient refers to the L2 regularization strength parameter (e.g., 0.01) set in the AdamW optimizer to constrain model complexity and suppress overfitting. The reference prediction model is a provisional version of the Chinese fir forest growth prediction model obtained after one complete iteration of training and when the loss function converges. It has not yet undergone hyperparameter tuning and accuracy verification and needs further evaluation to determine if it meets performance requirements. Prediction accuracy refers to the degree of matching between the predicted values ​​of diameter at breast height (DBH) growth and species richness output by the reference prediction model on the validation set and the actual values. It is usually expressed as the R-squared value of the regression task. 2 Alternatively, the MAE compliance rate, the F1-score for the classification task, or the probability calibration degree can be used as a comprehensive measure. The preset accuracy threshold refers to the pre-set standard for model performance (e.g., 90%). Only when the prediction accuracy on the validation set reaches or exceeds this threshold is the model considered successfully trained and can be used as the final deployment model for Chinese fir forest growth prediction.

[0063] Understandably, the system first simultaneously collects hyperspectral images, lidar point clouds, meteorological monitoring data, and field survey data of Chinese fir forests from multiple typical Chinese fir forest plots. This data is then aligned with geographical coordinates according to observation time to form a training dataset where each sample contains multimodal inputs and corresponding ground truth labels (such as annual diameter at breast height (DBH) increase, species richness, and whether the target is met). Secondly, the training dataset is labeled—the field survey results are used as a supervisory signal, and redundant or low-relevance features are removed (such as removing bands unrelated to growth or abnormal meteorological records). Then, the training and validation sets are randomly divided according to a preset ratio to ensure consistent data distribution between training and evaluation.

[0064] Then, a joint loss function is constructed by weighting and summing the mean squared error (regression loss for diameter at breast height growth and species richness) and cross-entropy (classification loss for the probability of reaching the target for cultivation). Using a second preset learning rate and the AdamW optimizer, iterative training is performed on the Chinese fir forest growth prediction model on the training set: each round traverses all training samples, forward propagation calculates the predicted value and loss, back propagation updates the network weights, and training continues until the loss no longer decreases after several consecutive rounds of verification, resulting in a reference prediction model.

[0065] Next, k-fold cross-validation (e.g., 5-fold) is used to perform grid search or random search adjustments on the key hyperparameters of the reference prediction model (including the number of CNN convolutional blocks, the number of PointNet MLP layers, the number of attention heads, and the loss weighting coefficients). Each time, the model is retrained with a different combination of hyperparameters, and the prediction accuracy is calculated on the validation set (defined as the percentage of samples with a diameter at breast height error ≤ 0.1 cm, a species error ≤ 1, and a correct probability of meeting the standard). If the prediction accuracy is lower than the preset accuracy threshold, the training step is returned, and the model is re-initialized and trained using the newly selected hyperparameter combination. When the prediction accuracy of the reference prediction model obtained from a certain training is greater than or equal to the preset accuracy threshold, the tuning is stopped, and the model is determined as the final trained Chinese fir forest growth prediction model, thereby ensuring that it has stable and reliable prediction capabilities in practical applications.

[0066] This embodiment first splits the target data into a spectral data subset and a point cloud data subset, which are then input into a spectral feature extraction branch and a structural feature extraction branch, respectively. This allows the hyperspectral image and the LiDAR point cloud to be processed independently in their respective optimal paths, preserving modality-specific information. Subsequently, the spectral feature extraction branch performs convolution operations, batch normalization, and activation processing on the spectral data subset using a CNN network, followed by dimensionality reduction through a global average pooling layer, outputting a spectral feature vector characterizing the physiological state of the Chinese fir canopy, effectively extracting growth-related spectral response features. Simultaneously, the structural feature extraction branch uses a PointNet network to first perform T-Net coordinate transformation on the point cloud data subset to correct its spatial pose, then learns geometric features through a multilayer perceptron, and finally performs maximum... Pooling layers aggregate and output structural feature vectors describing the three-dimensional structure of forest stands, enhancing the ability to model forest morphology. Next, the two feature vectors are input into a dual-channel attention fusion layer, where attention weight coefficients are adaptively calculated to dynamically measure the contribution of different modalities in the current prediction. Then, the two vectors are weighted and fused according to these weights to obtain fused features, enhancing key information and suppressing redundancy. The fused features are then passed through a fully connected prediction layer and a Softmax output layer, simultaneously outputting predicted values ​​for diameter at breast height (DBH) growth and species richness, achieving joint prediction of growth and ecological indicators. Finally, combining these two predicted values ​​with preset DBH growth thresholds and preset species richness reduction thresholds, the probability of achieving comprehensive silvicultural standards is calculated, providing a quantitative basis for silvicultural decision-making.

[0067] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the periodic tending and scheduling method for Chinese fir forests. Any simple modifications based on this technical concept are within the scope of protection of this application.

[0068] This application also provides a periodic tending and scheduling device for Chinese fir forests, please refer to... Figure 3 The periodic tending and scheduling device for Chinese fir forests includes: The cycle adjustment module 10 is used to adjust the preset initial nurturing cycle through the nurturing cycle adjustment model to obtain the target nurturing cycle. The nurturing cycle adjustment model is obtained by training an LSTM model. The data processing module 20 is used to collect multimodal Chinese fir forest data when the target tending cycle starts, and to preprocess the multimodal Chinese fir forest data to obtain target data. The multimodal Chinese fir forest data includes hyperspectral images, lidar point clouds and understory microenvironment data. The growth prediction module 30 is used to input the target data into the Chinese fir forest growth prediction model to obtain the predicted value of diameter at breast height (DBH) growth, the predicted value of species richness, and the probability of achieving the cultivation target. The Chinese fir forest growth prediction model is constructed based on the CNN network and PointNet network architecture. Strategy generation module 40 is used to generate equipment control strategies based on the predicted value of diameter at breast height growth, the predicted value of species richness, and the probability of achieving breeding standards. The equipment scheduling module 50 is used to schedule intelligent logging equipment, automatic replanting devices, or mechanical brush cutters to perform tending operations according to the equipment control strategy.

[0069] The periodic tending and scheduling device for Chinese fir forests provided in this application adopts the periodic tending and scheduling method for Chinese fir forests in the above embodiments, which can solve the technical problem of how to achieve periodic intelligent tending and scheduling while avoiding excessive interference with Chinese fir forests. Compared with the prior art, the beneficial effects of the periodic tending and scheduling device for Chinese fir forests provided in this application are the same as those of the periodic tending and scheduling method for Chinese fir forests provided in the above embodiments, and other technical features in the periodic tending and scheduling device for Chinese fir forests are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0070] This application provides a periodic tending and scheduling device for Chinese fir forests. The periodic tending and scheduling device for Chinese fir forests includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the periodic tending and scheduling method for Chinese fir forests in the above embodiment 1.

[0071] like Figure 4As shown, the periodic tending and scheduling equipment for Chinese fir forests may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in ROM (Read Only Memory) 1002 or a program loaded from storage device 1003 into RAM (Random Access Memory) 1004. RAM 1004 also stores various programs and data required for the operation of the periodic tending and scheduling equipment for Chinese fir forests. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via bus 1005. Input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, LCDs (Liquid Crystal Displays), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the periodic tending and scheduling equipment for Chinese fir forests to exchange data wirelessly or via wired communication with other devices. Although the figure shows a periodic tending and scheduling equipment for Chinese fir forests with various systems, it should be understood that it is not required to implement or possess all the systems shown. More or fewer systems can be implemented alternatively.

[0072] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0073] The periodic tending and scheduling device for Chinese fir forests provided in this application adopts the periodic tending and scheduling method for Chinese fir forests in the above embodiments, which can solve the technical problem of how to achieve periodic intelligent tending and scheduling while avoiding excessive interference with Chinese fir forests. Compared with the prior art, the beneficial effects of the periodic tending and scheduling device for Chinese fir forests provided in this application are the same as the beneficial effects of the periodic tending and scheduling method for Chinese fir forests provided in the above embodiments, and other technical features of the periodic tending and scheduling device for Chinese fir forests are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0074] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to execute the periodic tending and scheduling method for Chinese fir forests in the above embodiments.

[0075] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, RAM (Random Access Memory), ROM (Read Only Memory), EPROM (Erasable Programmable Read Only Memory or Flash Memory), optical fibers, CD-ROM (CD-Read Only Memory), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0076] The aforementioned computer-readable storage medium carries one or more programs. When these programs are executed by a periodic tending scheduling device for Chinese fir forests, the device performs the following actions: Adjusts a preset initial tending cycle using a tending cycle adjustment model to obtain a target tending cycle, wherein the tending cycle adjustment model is obtained by training an LSTM model; when the target tending cycle is initiated, collects multimodal Chinese fir forest data and preprocesses the multimodal Chinese fir forest data to obtain target data, wherein the multimodal Chinese fir forest data includes hyperspectral images, lidar point clouds, and understory microenvironment data; inputs the target data into a Chinese fir forest growth prediction model to obtain predicted values ​​for diameter at breast height (DBH), species richness, and the probability of achieving the tending target, wherein the Chinese fir forest growth prediction model is constructed based on a CNN network and a PointNet network architecture; generates an equipment control strategy based on the predicted DBH, the predicted species richness, and the probability of achieving the tending target; and schedules intelligent logging equipment, automatic replanting devices, or mechanical brush trimming equipment to perform tending operations according to the equipment control strategy.

[0077] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including LAN (Local Area Network) or WAN (Wide Area Network)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0078] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the aforementioned periodic tending and scheduling method for Chinese fir forests. This solves the technical problem of how to achieve periodic intelligent tending and scheduling while avoiding excessive interference with the Chinese fir forest. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the periodic tending and scheduling method for Chinese fir forests provided in the above embodiments, and will not be repeated here.

[0079] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A method for periodic tending and scheduling of Chinese fir forests, characterized in that, The method includes: The target nurturing cycle is obtained by adjusting the preset initial nurturing cycle using a nurturing cycle adjustment model, which is obtained by training an LSTM model. When the target tending cycle is started, multimodal Chinese fir forest data is collected and preprocessed to obtain target data. The multimodal Chinese fir forest data includes hyperspectral images, lidar point clouds and understory microenvironment data. The target data is input into the Chinese fir forest growth prediction model to obtain the predicted value of diameter at breast height (DBH) growth, the predicted value of species richness, and the probability of achieving the cultivation target. The Chinese fir forest growth prediction model is constructed based on the CNN network and PointNet network architecture. Based on the predicted value of diameter at breast height (DBH) growth, the predicted value of species richness, and the control strategy of the breeding success rate generation device; According to the equipment control strategy, intelligent logging equipment, automatic replanting devices, or mechanical brush cutters are scheduled to perform tending operations.

2. The method as described in claim 1, characterized in that, The Chinese fir forest growth prediction model includes a spectral feature extraction branch, a structural feature extraction branch, a global average pooling layer, a maximum pooling layer, a dual-channel attention fusion layer, a fully connected prediction layer, and a Softmax output layer. The step of inputting the target data into the Chinese fir forest growth prediction model to obtain the predicted value of diameter at breast height (DBH), the predicted value of species richness, and the probability of achieving the cultivation target includes: The target data is split into a spectral data subset and a point cloud data subset, and then input into the spectral feature extraction branch and the structural feature extraction branch, respectively. The CNN network of the spectral feature extraction branch performs convolution, batch normalization and activation processing on the spectral data subset, and outputs the spectral feature vector after dimensionality reduction by the global average pooling layer. The PointNet network of the structural feature extraction branch performs coordinate transformation and multilayer perceptron feature learning on the subset of point cloud data, and outputs structural feature vectors after processing by the max pooling layer. The spectral feature vector and the structural feature vector are input into the dual-channel attention fusion layer to obtain their attention weight coefficients; The spectral feature vector and the structural feature vector are weighted and fused according to the attention weight coefficients to obtain the fused feature; The fused features are input into the fully connected prediction layer, and after calculation by the Softmax output layer, the predicted values ​​of diameter at breast height (DBH) growth and species richness are obtained. The probability of achieving the overall breeding target is calculated based on the predicted value of diameter at breast height (DBH) growth, the predicted value of species richness, the preset DBH growth threshold, and the preset species richness reduction threshold.

3. The method as described in claim 1, characterized in that, The steps for constructing the Chinese fir forest growth prediction model include: A spectral feature extraction branch is constructed. The spectral feature extraction branch adopts a CNN network containing a preset number of convolutional blocks. Each convolutional block is composed of a Conv2D layer, a BatchNorm layer and a ReLU activation function in sequence. A global average pooling layer is set at the end of the spectral feature extraction branch. A structural feature extraction branch is constructed, which adopts the PointNet network architecture. The PointNet network architecture includes a T-Net coordinate transformation module and a multi-layer MLP network. A max pooling layer is set at the end of the structural feature extraction branch. Based on the spectral feature extraction branch and the structural feature extraction branch, a dual feature extraction branch architecture is constructed; Following the dual feature extraction branch architecture, a dual-channel attention fusion layer, a fully connected prediction layer, and a Softmax output layer are constructed to obtain the Chinese fir forest growth prediction model.

4. The method as described in claim 1, characterized in that, The step of generating a control strategy for the equipment based on the predicted value of diameter at breast height (DBH), the predicted value of species richness, and the probability of achieving breeding standards includes: Based on the terrain slope data derived from the lidar point cloud and the forest understory microenvironment data, the growth environment is divided into zones, resulting in a zoning result that includes a growth zone, a transition zone, and a sparse edge zone. Based on the zoning results, the predicted values ​​of diameter at breast height (DBH) growth and species richness within each zoning zone are compared to determine the nurturing priority of each zoning zone. The nurturing priority of the growth zone is higher than that of the transition zone and the marginal sparse zone. Based on the interference factors in each partition and the intervention level corresponding to the probability of achieving the breeding standard, the equipment type and basic operating parameters of each partition are matched; Based on the nurturing priority, the equipment type, the basic operating parameters, and the equipment operating sequence, a device control strategy with partition markings is generated.

5. The method as described in claim 1, characterized in that, The step of scheduling intelligent logging equipment, automatic replanting devices, or mechanical brush cutters to perform tending operations according to the equipment control strategy includes: The equipment control strategy is analyzed to extract the equipment type, operating parameters, operating sequence, and brooding priority corresponding to each zone; Based on the nurturing priority, the tasks of each zone are sorted to generate a zone task sorting table; Based on the partition job sorting table, the device type, and the job parameters, a partition job scheduling instruction is generated; The zoning operation scheduling instructions are sent to the corresponding tending operation equipment, which includes intelligent logging equipment, automatic replanting device and / or mechanical brush cutting equipment; When a confirmation signal is received from the nurturing operation equipment, a work start command is sent according to the partitioned work sorting table, and the nurturing operation equipment is scheduled to perform the nurturing operation according to the work parameters.

6. The method as described in claim 1, characterized in that, The training steps for the Chinese fir forest growth prediction model include: Hyperspectral images, point cloud data, meteorological monitoring data, and field survey data of Chinese fir forests were collected to construct a training dataset; The training dataset is labeled and its features are filtered, and then divided into a training set and a validation set according to a preset ratio; Using the mean squared error and cross-entropy as the loss function, the Chinese fir forest growth prediction model is iteratively trained according to the second preset learning rate, AdamW optimizer, preset weight decay coefficient and the training set until the value of the loss function converges, and a reference prediction model is obtained. The hyperparameters of the reference prediction model are adjusted using cross-validation, and the prediction accuracy of the reference prediction model is verified based on the validation set. When the prediction accuracy is less than the preset accuracy threshold, the process returns to the step of using the mean square error and cross entropy as the loss function, iteratively training the Chinese fir forest growth prediction model according to the second preset learning rate, AdamW optimizer, preset weight decay coefficient and the training set until the value of the loss function converges to obtain the reference prediction model. When the prediction accuracy is greater than or equal to the preset accuracy threshold, the reference prediction model is used as the trained Chinese fir forest growth prediction model.

7. The method according to any one of claims 1 to 6, characterized in that, The step of adjusting the preset initial nurturing cycle using a nurturing cycle adjustment model to obtain the target nurturing cycle includes: Acquire historical growth time-series data of Chinese fir forest, annual difference data of understory vegetation, and historical pruning cycle effect data to form a periodic adjustment input dataset; The periodically adjusted input dataset is normalized to obtain standardized time-series data; The standardized time-series data is input into the nurturing cycle adjustment model to obtain the initial adjustment coefficients; The initial adjustment coefficient is corrected based on the preset annual difference constraint threshold of forest understory vegetation to obtain the target adjustment coefficient; The target nurturing period is calculated based on the target adjustment coefficient and the preset initial nurturing period.

8. A periodic tending and scheduling device for Chinese fir forests, characterized in that, The device includes: The cycle adjustment module is used to adjust the preset initial nurturing cycle through the nurturing cycle adjustment model to obtain the target nurturing cycle. The nurturing cycle adjustment model is obtained by training an LSTM model. The data processing module is used to collect multimodal Chinese fir forest data when the target tending cycle starts, and to preprocess the multimodal Chinese fir forest data to obtain target data. The multimodal Chinese fir forest data includes hyperspectral images, lidar point clouds and forest understory microenvironment data. The growth prediction module is used to input the target data into the Chinese fir forest growth prediction model to obtain the predicted value of diameter at breast height (DBH) growth, the predicted value of species richness, and the probability of achieving the cultivation target. The Chinese fir forest growth prediction model is constructed based on the CNN network and PointNet network architecture. The strategy generation module is used to generate a device control strategy based on the predicted value of diameter at breast height growth, the predicted value of species richness, and the probability of achieving the breeding target. The equipment scheduling module is used to schedule intelligent logging equipment, automatic replanting devices, or mechanical brush cutters to perform tending operations according to the equipment control strategy.

9. A periodic tending and scheduling device for Chinese fir forests, characterized in that, The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the periodic tending and scheduling method for Chinese fir forests as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, it implements the steps of the periodic tending and scheduling method for Chinese fir forests as described in any one of claims 1 to 7.

Images