Forestry optimization management and control method and system based on artificial intelligence

Abstract

The invention relates to the technical field of forestry intelligent management and control, and discloses a forestry optimization management and control method and system based on artificial intelligence. The method comprises the following steps: collecting multi-source environment monitoring data such as soil humidity, illumination intensity and vegetation growth indexes in a forestry area, aligning a time sequence through a dynamic time warping algorithm, and separating steady-state and transient components; detecting anomalies through an isolated forest algorithm and generating a thermodynamic diagram, and obtaining a potential disease and pest outbreak area through causal reasoning in combination with historical records; constructing a dynamic growth reference curve, and identifying a disturbance sensitive area by using a generative adversarial network and a convolutional neural network; and integrating and generating a management and control priority partition map, optimizing resource allocation through deep reinforcement learning, outputting a task sequence, and driving equipment to execute irrigation, fertilization or pest control operation. According to the method, multiple artificial intelligence algorithms are fused, accurate control over the whole forestry process is achieved, and the method adapts to the dynamically-changing forestry environment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of forestry intelligent management and control, in particular to a forestry optimization management and control method and system based on artificial intelligence. BACKGROUND

[0002] As an important part of the ecological system, the healthy and stable development of forestry plays an irreplaceable role in maintaining ecological balance, ensuring timber resource supply and responding to climate change. With the promotion of large-scale forestry operation, the traditional forestry management and control mode has gradually exposed many limitations and has been difficult to meet the needs of modern forestry management.

[0003] Traditional forestry management and control relies on manual patrol and experience judgment, which not only consumes a lot of manpower and material resources, but also has the problems of incomplete data collection and delayed response. In terms of environmental data monitoring, the forestry area often has wide span and complex terrain, and single type of monitoring data cannot fully reflect the ecological conditions in the region. Due to the differences in collection equipment and environmental interference, the time series of multi-source data often exists out of sync, which makes it difficult to effectively mine the correlation between data, and further affects the accuracy of subsequent analysis and decision-making.

[0004] In the aspect of anomaly detection and risk prediction, the traditional method sets a fixed threshold for a single environmental index, which is difficult to capture transient abnormal signals in data and cannot effectively distinguish between environmental natural fluctuations and potential risk-induced anomalies, which is prone to misjudgment or omission. For disease and pest control, existing technologies mostly rely on historical outbreak records for statistical analysis, lack of deep mining of the causal relationship between abnormal environmental data and disease and pest outbreaks, and often take measures after large-scale disease and pest outbreaks, missing the best prevention opportunity and causing serious ecological and economic losses.

[0005] In terms of vegetation growth management, the traditional mode usually uses a unified growth standard, ignoring the differences in micro-environment in the forestry area and the dynamic change characteristics of vegetation growth, resulting in lack of pertinence of irrigation, fertilization and other control measures. At the same time, the existing resource allocation scheme is mostly planned based on static zoning, without fully considering the risk level and growth sensitivity of different regions, which is prone to resource waste or insufficient control of key areas. In addition, the forestry environment has strong dynamics and complexity, and the adaptability and flexibility of the traditional control method are insufficient, which is difficult to cope with the impact of sudden changes such as climate change and extreme weather, restricting the overall efficiency and quality of forestry management. SUMMARY

[0006] The purpose of the present application is to provide a forestry optimization management and control method and system based on artificial intelligence to solve the problems raised in the background.

[0007] To achieve the above object, the present application provides a forestry optimization management method based on artificial intelligence, which comprises: Collecting multi-source environmental monitoring data of the forestry area, which includes soil humidity, light intensity and vegetation growth index; Aligning the time series of the multi-source environmental monitoring data by dynamic time warping algorithm, extracting time sequence characteristics and separating steady-state components and transient components; Performing anomaly detection on the transient components based on the isolated forest algorithm, marking abnormal data points and generating an abnormal distribution heat map; Combining historical pest records with the abnormal distribution heat map, using a causal inference model to infer potential pest outbreak areas; According to the correlation between the steady-state components and the vegetation growth index, a dynamic growth benchmark curve is constructed; Using a generative adversarial network to generate perturbation samples of the dynamic growth benchmark curve, and identifying perturbation sensitive areas through a convolutional neural network; Integrating the potential pest outbreak areas and the perturbation sensitive areas to generate a forestry management priority zoning map; Using a deep reinforcement learning model to optimize resource allocation for the forestry management priority zoning map, and outputting a management task sequence; Based on the management task sequence, driving the forestry equipment to perform irrigation, fertilization or pest control operations.

[0008] Preferably, the dynamic time warping algorithm is used to align the time series of the multi-source environmental monitoring data, extract time sequence characteristics and separate steady-state components and transient components, which comprises: Performing sliding window segmentation on the multi-source environmental monitoring data to obtain windowed time sequence segments; Calculating the dynamic time warping distance between each windowed time sequence segment and a preset standard template, and normalizing to obtain an alignment offset; According to the alignment offset, correcting the phase deviation of the windowed time sequence segment to generate aligned time sequence data; Applying empirical mode decomposition to process the aligned time sequence data to decompose high-frequency transient components and low-frequency steady-state components.

[0009] Preferably, the isolated forest algorithm is used to perform anomaly detection on the transient components, mark abnormal data points and generate an abnormal distribution heat map, which comprises: Inputting the high-frequency transient components into the isolated forest algorithm to construct a random partition tree set; Statistically analyzing the average path length of each data point in the tree set to calculate the anomaly score; Screening abnormal data points with anomaly scores exceeding a dynamic threshold and mapping them to forestry area spatial coordinates; The abnormal distribution heat map is generated by clustering the spatial coordinates through kernel density estimation.

[0010] Preferably, the history of pest and disease records is combined with the abnormal distribution heat map, and a potential pest and disease outbreak area is inferred using a causal inference model, including: Matching the spatial overlap area of the abnormal data points in the abnormal distribution heat map and the history of pest and disease records; Extracting the environmental feature vector of the overlap area as a training sample for the causal inference model; The trained causal inference model outputs the pest and disease outbreak probability of each area; The area with a probability exceeding a preset threshold is marked as a potential pest and disease outbreak area.

[0011] Preferably, the dynamic growth benchmark curve is constructed according to the correlation between the steady-state component and the vegetation growth index, including: The correlation degree between the low-frequency steady-state component and the vegetation growth index is calculated using grey correlation analysis method; Selecting the steady-state component with a correlation degree exceeding a set threshold as a feature input; The mapping relationship between the feature input and the vegetation growth index is fitted through a long short-term memory network to generate a dynamic growth benchmark curve.

[0012] Preferably, the disturbance sample of the dynamic growth benchmark curve is generated using a generative adversarial network, and a convolutional neural network is used to identify the disturbance sensitive area, including: The dynamic growth benchmark curve is input into the generator of the generative adversarial network, and a synthetic disturbance curve is output; The discriminator distinguishes between real growth curves and synthetic disturbance curves, and iteratively optimizes the generator parameters; The trained generator generates a variety of disturbance samples; The convolutional neural network performs semantic segmentation on the difference area between the disturbance sample and the real sample, and outputs the disturbance sensitive area.

[0013] Preferably, the potential pest and disease outbreak area and the disturbance sensitive area are integrated to generate a forestry management priority zoning map, including: The potential pest and disease outbreak area and the disturbance sensitive area are subjected to spatial overlay analysis, the fusion weight is calculated according to the area and center distance of the overlap area, and the forestry management priority zoning map containing priority levels one to three is generated based on the fusion weight.

[0014] Preferably, the forestry management priority zoning map is optimized for resource allocation using a deep reinforcement learning model, and a control task sequence is output, including: The state vector is encoded by the state of the forestry equipment, the resource inventory, and the priority zoning map; The resource consumption efficiency is defined as a reward function together with a region priority weighting; The deep reinforcement learning model explores an optimal resource allocation strategy through a policy gradient algorithm; The policy output is decoded to obtain a management and control task sequence.

[0015] Preferably, the forestry equipment is driven to perform irrigation, fertilization or pest control operations based on the management and control task sequence, which comprises: The device instructions and parameters in the management and control task sequence are parsed, and the instructions are issued to the corresponding forestry equipment through an Internet of Things protocol, and the device execution feedback is monitored in real time and the task state is updated.

[0016] Preferably, the present application further comprises an artificial intelligence-based forestry optimization management and control system, which comprises a memory, a processor and a computer program stored in the memory and running on the processor, wherein the processor implements the steps of the artificial intelligence-based forestry optimization management and control method as described above when executing the computer program.

[0017] Compared with the prior art, the present application has the following advantages: By collecting multi-source environmental monitoring data, key ecological factors such as soil moisture, light intensity and vegetation growth indicators are comprehensively covered, breaking the limitations of traditional single data monitoring, and providing more comprehensive basic data support for forestry management and control. With the help of dynamic time warping algorithm, the time series of multi-source data are aligned, which can effectively solve the problem of inconsistent time series of different source data, accurately extract the time series characteristics and separate the steady-state component and the transient component, so that the internal law and fluctuation characteristics of the data are clearly presented, providing a high-quality data basis for subsequent analysis and decision-making.

[0018] Based on the isolated forest algorithm, the transient component is detected for abnormality, which can accurately capture abnormal signals in the data, mark abnormal data points and generate an abnormal distribution heat map, and intuitively display the spatial distribution characteristics of the abnormal area. Compared with the traditional fixed threshold detection method, it is more suitable for the complex fluctuations of the forestry environment, and can reduce the misjudgment rate and omission rate of abnormal judgment. Combined with historical pest records and abnormal distribution heat maps, a causal inference model is used to infer potential pest outbreak areas, which can reveal the internal relationship between environmental abnormalities and pest outbreaks, realize the forward prediction of pest risks, change the passive situation of traditional post-response, and create conditions for deploying prevention and control measures in advance.

[0019] The dynamic growth benchmark curve is constructed according to the correlation between the steady-state components and the vegetation growth indicators, which discards the rigid mode of the traditional fixed growth standard, can adapt to the dynamic changes of vegetation growth and the differences of regional micro-environment in real time, and makes the growth evaluation more targeted and scientific. The disturbance sample of the benchmark curve is generated through the generative adversarial network, and then the convolutional neural network is used to identify the disturbance sensitive area, which can accurately locate the vegetation area with strong reaction to external interference and clearly define the growth vulnerable point of different areas, providing accurate basis for differentiated management and control.

[0020] The forestry management and control priority zoning map is generated by integrating the potential pest outbreak area and the disturbance sensitive area, which clearly defines the management and control demand level of different areas, so that the management and control resources can be tilted to the high-risk and high-sensitive areas. The deep reinforcement learning model is used to optimize the resource allocation of the priority zoning map, which can dynamically adjust the resource allocation scheme according to the real-time environmental changes and management and control targets, avoiding the irrationality of traditional static resource allocation, and realizing the maximization of resource utilization efficiency.

[0021] The intelligent analysis results are converted into specific operation instructions by driving the forestry equipment to perform irrigation, fertilization or pest control operation through the management and control task sequence, realizing the whole-process intelligent closed loop of forestry management and control from data collection, analysis, decision-making to execution. The whole method fully integrates the advantages of various artificial intelligence algorithms, can flexibly cope with the dynamic changes of forestry environment, realizes the fine and accurate management and control of forestry resources, effectively guarantees the healthy growth of vegetation, and maintains the stability of forestry ecological system. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 is an abnormal distribution thermal map of forestry area; Figure 2 is a flowchart of time alignment and component separation; Figure 3 is a flowchart of potential pest outbreak area inference; Figure 4 is a vegetation growth disturbance sample difference analysis diagram based on generative adversarial network. DETAILED DESCRIPTION

[0023] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0024] Please refer to Figure 1The application provides an artificial intelligence-based forestry optimization management method, which comprises: through the integration of multiple links such as multi-source data collection, time series analysis, anomaly detection, causal reasoning, growth modeling, disturbance generation, regional integration and resource optimization, intelligent and accurate forestry management is realized. The method starts with collecting multi-source environmental monitoring data of the forestry area, which includes soil moisture, light intensity and vegetation growth indicators. These data are obtained in real time through the sensor network deployed in the forestry area, ensuring the comprehensiveness and timeliness of the data. The collected multi-source environmental monitoring data have inconsistent time series. The dynamic time warping algorithm is used to align the time series of the multi-source environmental monitoring data. The dynamic time warping algorithm can process data points with different sampling frequencies, and align them by calculating the similarity between sequences. The aligned time series is convenient for extracting time series features, including trend and periodic components. The empirical mode decomposition technique is applied to separate the steady-state component and the transient component. The steady-state component represents long-term stable environmental changes, and the transient component captures short-term fluctuations. The transient component is input into the isolation forest algorithm for anomaly detection. The isolation forest algorithm constructs a random partition tree set and calculates the anomaly score of each data point. Points with high anomaly scores are marked as abnormal data points. The abnormal data points are mapped to the spatial coordinates of the forestry area. The kernel density estimation method is used to generate an abnormal distribution heat map. The abnormal distribution heat map visualizes the abnormal aggregation area. Combine the historical pest records from the forestry management database with the abnormal distribution heat map. The historical pest records contain the location and intensity of past pest events. The causal reasoning model analyzes the spatial correlation between the abnormal distribution heat map and the historical records to infer the potential pest outbreak area. The causal reasoning model uses a graph-based reasoning method to evaluate the causal impact of environmental anomalies on pest outbreaks.

[0025] The steady-state components are associated with vegetation growth indicators, including leaf area index and biomass. The grey correlation analysis method is used to calculate the correlation degree of the steady-state components and the vegetation growth indicators. The steady-state components with high correlation degree are taken as characteristic inputs. The long short-term memory network learns the mapping relationship between the characteristic inputs and the vegetation growth indicators to generate a dynamic growth benchmark curve. The dynamic growth benchmark curve reflects the normal growth trajectory of the vegetation. The generative adversarial network processes the dynamic growth benchmark curve. The generator generates a synthetic disturbance curve, and the discriminator distinguifies the real curve and the disturbance curve. Through adversarial training, the generator is optimized to generate diverse disturbance samples. The convolutional neural network analyzes the differences between the disturbance samples and the real samples to identify disturbance sensitive areas, which indicate the parts of the vegetation growth susceptible to interference. The potential pest outbreak areas and the disturbance sensitive areas are integrated. The spatial superposition analysis calculates the area overlap degree. Based on the area and distance, the fusion weight is calculated to generate a forestry management control priority zoning map. The forestry management control priority zoning map divides the areas into first to third priority areas. The deep reinforcement learning model processes the forestry management control priority zoning map. The model encodes the device status and resource inventory into a state vector. The reward function is defined based on the resource consumption efficiency and priority weighting. The policy gradient algorithm explores the optimal resource allocation strategy, and outputs a management control task sequence. The management control task sequence is parsed into device instructions, which are issued to forestry devices through Internet of Things protocols to drive irrigation, fertilization or pest control operations. The execution feedback updates the task state in real time.

[0026] Embodiment 1: refer to Figure 2 After the multi-source environmental monitoring data is collected, there is a problem of inconsistent time stamps. The multi-source environmental monitoring data comes from a heterogeneous sensor network deployed in a forestry area. Soil moisture sensors measure the volume water content using the frequency domain reflection principle. Light intensity sensors record the photosynthetically active radiation flux density based on the photoelectric effect. Vegetation growth indicators are represented by the normalized vegetation index captured by a multispectral camera. These sensors operate at different sampling frequencies. Soil moisture data is recorded every minute, light intensity data is sampled every second, and vegetation growth indicators are collected once a day, resulting in a serious time desynchronization phenomenon in the original data stream. The sliding window segmentation technique is used to process the time series of multi-source environmental monitoring data. The sliding window size is determined according to the optimal time resolution. The window width is set to twenty-four hours to cover the entire day cycle. The window sliding step is set to one hour to achieve ninety-five percent data overlap. Each windowed time series segment contains a fixed number of data points. The soil moisture segment contains one thousand four hundred and forty points. The light intensity segment contains eighty-six thousand four hundred points. The vegetation growth indicator segment retains twenty-four points. Missing values are completed by linear interpolation method.

[0027] The windowed time series is matched with a preset standard template, which is derived from historical typical growing season data, and the template sequence length is consistent with the windowed time series. The dynamic time warping distance calculation adopts a dynamic programming algorithm to construct a cumulative cost matrix to find the minimum bending path, and the path bending weight is set in a symmetric form. The distance value is normalized by using the minimum-maximum scaling method to map the result to the interval of zero to one, and the normalized alignment offset reflects the time drift degree of the windowed time series relative to the preset standard template. The alignment offset corrects the phase deviation of the windowed time series, and the correction process adopts a phase shift algorithm, and the data points are moved by the sampling point number corresponding to the offset on the time axis. The moved data points maintain the smoothness of the curve through cubic spline interpolation. The generated aligned time series data realizes the accurate synchronization of multi-source data on a unified time axis, and the time alignment error of soil moisture data and light intensity data is controlled within hundreds of milliseconds, and the vegetation growth index data is aligned to a fixed daily timestamp. The aligned time series data is processed by empirical mode decomposition, and the empirical mode decomposition iteratively extracts intrinsic mode functions through a screening process. The screening number is automatically determined according to the standard deviation threshold. Each intrinsic mode function satisfies the condition that the number of extreme points and the number of zero-crossing points differ by no more than one. Six intrinsic mode functions and a residual term are finally generated by decomposition. The high-frequency transient component is superimposed by the first three intrinsic mode functions, and the frequency range is above 0.5 Hz, which contains short-term events such as rainfall impact and mechanical vibration. The low-frequency steady-state component is composed of the last three intrinsic mode functions and the residual term, and the frequency is below 0.1 Hz, which reflects the slow change of soil moisture and the long-term growth trend of vegetation. The instantaneous frequency of the intrinsic mode function is calculated by Hilbert transform, which ensures that the time-frequency distribution has clear physical meaning.

[0028] The high-frequency transient component is input into the Isolation Forest algorithm for anomaly detection. The Isolation Forest algorithm constructs one hundred randomly partitioned trees, and each tree is constructed using a random sample of two hundred data points. The randomly partitioned tree randomly selects the feature dimension during construction, and the split value is randomly generated within the range of the feature value. The tree growth depth is limited to the logarithmic level. The path length of the data point in each tree is counted from the root node to the leaf node. The path length of the isolated data point is shorter, and the path length of the normal data point is longer. The average path length is calculated by taking the arithmetic mean of the path lengths of all trees. The anomaly score is calculated based on the ratio of the average path length to the theoretical average path length. The dynamic threshold is determined using the boxplot method. The upper limit is the third quartile plus one and a half times the interquartile range. Data points with anomaly scores exceeding the dynamic threshold are marked as abnormal data points. The abnormal data points are mapped to the forestry area spatial coordinates, which are derived from the differential GPS positioning system with a plane positioning accuracy of centimeters. Each abnormal data point is associated with a three-dimensional geographic coordinate, with longitude and latitude recording the plane position and altitude recording the vertical position. The kernel density estimation method uses a Gaussian kernel function for spatial clustering. The kernel function bandwidth is optimized to fifty meters using the cross-validation method. The set of spatial coordinate points generates a two-dimensional probability density distribution. The density value is visualized through a color gradient. The red area represents an abnormal point density of more than ten points per hectare, and the blue area represents an abnormal point density of less than one point per hectare. The grid resolution of the abnormal distribution heat map is set to five meters by five meters. Each grid cell stores the kernel density estimation value. The heat map file is stored in GeoTIFF format and overlaid with a satellite base map.

[0029] The implementation of dynamic time warping algorithm adopts multi-thread parallel computing, and the windowed time series fragments are allocated to multiple processor cores for synchronous processing. The empirical mode decomposition algorithm adds an ensemble average technique, which averages multiple decomposition results with white noise to suppress mode mixing. The tree construction process of the isolation forest algorithm introduces a subsampling strategy, which only uses a subset of the dataset to build a single tree each time, reducing the time complexity of the algorithm. The calculation of kernel density estimation uses a graphics processing unit to accelerate, divides the spatial grid into multiple calculation blocks for parallel processing. The entire processing flow is embedded in a distributed computing framework, using the Apache Spark platform to manage data partitioning and task scheduling. The storage of multi-source environmental monitoring data uses a time series database structure, and each sensor's data stream is segmented into hourly data blocks for compressed storage. In the data alignment process, a quality control system is added to mark and exclude data segments with a signal missing rate of more than ten percent. The adaptive update mechanism is set in the anomaly detection module, and the isolation forest model is retrained every twenty-four hours to adapt to the temporal evolution of environmental patterns. The spatial visualization component integrates WebGIS technology, supporting interactive browsing of anomaly distribution heat maps through the browser, and users can adjust the transparency and color scheme of the heat map. The system operation monitoring interface displays the status of the data processing pipeline in real time, including data reception rate, computing resource usage, and result output delay. The boundary processing of sliding window segmentation uses the mirror extension method, which replicates the symmetric paragraph at both ends of the data sequence to avoid edge effects. The slope constraint is added in the dynamic time warping calculation to limit the inclination angle of the curved path to less than forty-five degrees, preventing excessive distortion of the time series shape. The boundary processing of empirical mode decomposition uses the characteristic wave extension method, which predicts the envelope line trend outside the boundary according to the extreme point distribution characteristics. The isolation forest algorithm adds a feature selection step when processing high-dimensional data, retaining only the key dimensions with a variance greater than the threshold for anomaly detection. The boundary correction of kernel density estimation uses the reflection method to generate mirror points outside the calculation region boundary to eliminate boundary bias.

[0030] Example 2: see Figure 3The historical pest record is derived from the monitoring archives of forestry management department in the past ten years, and the record format contains four fields of pest occurrence timestamp, geographical boundary coordinates, pest species code, and damage degree level. The anomaly distribution heat map is from the output results of the above embodiment, and the heat map is stored in a raster data format, each raster cell contains an anomaly density value and geographical reference information. The spatial overlap area matching adopts the overlay analysis tool of geographic information system to perform overlay processing on the polygon boundary layer of the historical pest record and the raster layer of the anomaly distribution heat map. The overlay analysis uses the intersection operation method to calculate the spatial intersection of the geometric elements of the two layers, and generates a set of planar elements of the overlapping area. Each overlapping area is associated with the original attribute data, and the historical pest record field and the anomaly density value field are combined to form a temporary attribute table. The process of extracting the environmental feature vector of the overlapping area involves multi-source data fusion, and the environmental feature vector contains fifteen dimensional indexes. The soil moisture anomaly value is taken from the standardized anomaly score of the corresponding raster cell in the anomaly distribution heat map, and the light intensity anomaly value is obtained by calculating the deviation of the current value from the historical same period mean value. The historical pest frequency index counts the number of pest occurrences in the past five years, and records the occurrence frequency according to the pest species. The environmental feature vector also includes terrain feature parameters, the elevation data is derived from the digital elevation model, and the slope and aspect are calculated by deriving the elevation data. The vegetation coverage index uses the normalized vegetation index product of the Sentinel-2 satellite image, and the phenological period index is extracted according to the vegetation index time series. The soil type data comes from the national soil survey result map, and the climate factors include monthly mean temperature and precipitation data.

[0031] The environmental feature vector is used as a training sample for the causal inference model. The training sample set contains 3,000 historical case data. The causal inference model uses a dual machine learning method based on the latent outcome framework. The dual machine learning method eliminates confounding variable bias through orthogonalization processing. The training sample set is divided into a training set and a test set in a ratio of seven to three. The training set is used for model parameter estimation, and the test set is used for model performance evaluation. The causal inference model structure includes two-stage machine learning models. The first-stage model predicts the latent relationship between the treatment variable and the outcome variable, and the second-stage model estimates the conditional average treatment effect. The model training uses cross-validation techniques, and the five-fold cross-validation cycle optimizes the model hyperparameters. The feature engineering process includes missing value filling, outlier truncation, and numerical standardization steps. Categorical variables are converted using one-hot encoding. The trained causal inference model outputs the probability of pest and disease outbreaks in each region. The probability calculation is based on the linear combination of the environmental feature vector and the model weight matrix. The weight matrix is learned through the model training process, and the matrix dimension is fifteen by one, corresponding to the fifteen environmental feature dimensions. The probability value calculation uses the logistic function to map the linear output to the interval between zero and one. The slope parameter of the logistic function is determined by maximum likelihood estimation. The probability output results of each region are stored as probability distribution maps, and the distribution map grid resolution is consistent with the abnormal distribution heat map. The probability distribution map is spatially smoothed using a 300m x 300m moving window for median filtering to eliminate local noise interference. The preset threshold is determined based on the receiver operating characteristic curve, which depicts the true positive rate and false positive rate at different threshold values. The threshold selection principle is to balance the false alarm risk and the false alarm risk. Finally, 0.65 is set as the probability threshold. The grid cell values in the probability distribution map are compared with the preset threshold. Grid cells with values greater than 0.65 are marked as potential pest and disease outbreak areas. The potential pest and disease outbreak areas are morphologically processed. First, the erosion algorithm is used to remove isolated patches with an area less than one hectare. Then, the dilation algorithm is used to fill the internal holes of the region. The processed region is vectorized into polygon elements, and each polygon is associated with the average outbreak probability, area perimeter, and pest and disease species prediction information.

[0032] The historical pest record database establishes strict quality control procedures, and the data is checked for format, spatial topology, and logical consistency before being stored. The spatial reference system of the anomaly distribution heat map is unified as the UTM projection coordinate system to ensure consistency of the coordinate system when superimposing the analysis. The training process of the causal reasoning model records complete experimental logs, including data version, parameter setting, training loss curve, and verification index changes. Uncertainty quantification is added during the model inference stage, and the self-sampling method is used to generate the confidence interval of the probability prediction. The output results of the potential pest outbreak area are visualized and rendered, with different probability levels represented by gradient color systems, and the contour line background is added to enhance the expression effect of the terrain. The spatial overlap area matching algorithm optimizes the calculation efficiency, and the spatial indexing technology is used to speed up the geometric query operation. The environmental feature vector extraction process is automated and pipelined, and the feature data table is generated in batches through spatial query language. The causal reasoning model is deployed as a microservice architecture, supporting high-concurrency probability calculation requests. The potential pest outbreak area results are integrated into the forestry management decision-making platform, and the data is synchronized with the mobile terminal of the patrol personnel. A feedback mechanism is established to return the actual occurrence of pest events to the historical pest record database, forming a model optimization closed loop. The historical pest record data update mechanism stipulates that the monthly incremental synchronization is performed within 72 hours of the occurrence of new pest events. The anomaly distribution heat map generation cycle is set to once a day to ensure the timeliness of environmental anomaly information. The causal reasoning model is retrained quarterly to continuously improve the model's prediction accuracy. The potential pest outbreak area release process includes three levels of review, and the model output results are manually reviewed by domain experts. Detailed operation procedure documents are established throughout the implementation process of the entire technical solution, including data preprocessing standards, model training steps, and result verification methods. The overlap area matching considers the time dimension association, and the historical pest records and anomaly distribution heat map are weighted according to the time difference of occurrence. The environmental feature vector adds time series features, calculating the monthly change rate and annual fluctuation amplitude of environmental indicators. The causal reasoning model introduces an attention mechanism, giving higher weights to key environmental features. The potential pest outbreak area analysis adds a transmission risk factor, considering the potential path of pest spread through biological vectors. The system output results provide multiple data interfaces, supporting seamless integration with existing forestry monitoring platforms.

[0033] In Example 3, the steady-state components are derived from the low-frequency signal components obtained by empirical mode decomposition in Example 1, which represent the long-term trends of environmental factors such as soil moisture and light intensity. Vegetation growth indicators are obtained through field surveys, including tree diameter growth, leaf area index measurements, and tree height annual ring width measurements. The grey relational analysis method is used to calculate the correlation between the steady-state components and the vegetation growth indicators. The grey relational analysis method determines the degree of correlation based on the geometric similarity of the sequences, and the calculation process involves mean value processing, correlation coefficient solving, and correlation degree synthesis. The mean value processing converts the time series of steady-state components and the sequence of vegetation growth indicators into dimensionless values. The correlation coefficient reflects the closeness of the two sequences at each time. The correlation degree is finally taken as the average of the correlation coefficients. The correlation degree threshold is set to 0.75, and the steady-state components with a correlation degree greater than 0.75 are selected as the feature input. The feature input includes the daily variation sequence of the soil moisture steady-state value and the seasonal trend item of the light intensity steady-state value. The long short-term memory network is used to fit the mapping relationship between the feature input and the vegetation growth indicators. The long short-term memory network structure includes an input layer, a hidden layer, and an output layer. The number of input layer neurons is consistent with the dimension of the feature input. The hidden layer uses thirty-two LSTM units, and the output layer corresponds to the predicted value of the vegetation growth indicators. The network training uses a sliding window with a time step of thirty days, and each time window contains the feature input sequence and the corresponding real value of the vegetation growth indicators. The loss function uses the mean square error to calculate the deviation between the predicted value and the true value, and the optimizer selects the Adam algorithm to update the network weights. The early stopping mechanism is added to prevent overfitting during the training process. When the validation set loss does not decrease for ten consecutive rounds, the training is terminated. The dynamic growth benchmark curve is generated by multi-step prediction using the trained long short-term memory network, covering the entire growing season.

[0034] The mathematical expression of the dynamic growth benchmark curve is in the form of a growth function, and the curve shape follows the physiological laws of vegetation growth. The following formula is introduced to describe the relationship between the growth benchmark and the feature input during the curve generation process:

[0035] wherein: represents the predicted value of the vegetation growth indicator at time t, represents the state vector of the long short-term memory network hidden layer at time t, is the weight matrix from the hidden layer to the output layer, is the output layer bias term, represents the S-shaped activation function. Both sides of the formula have dimensionless growth indicator values, maintaining dimensional consistency. The vector contains thirty-dimensional feature information, with each dimension corresponding to an influence factor of a steady-state component. The matrix has a dimension of thirty-two by one, and the optimal weight value is learned through the training process. The bias term adjusts the output reference value to adapt to the growth characteristics of different vegetation types.

[0036] The steady-state component data preprocessing includes outlier rejection and trend decomposition steps, using robust regression methods to remove abnormal fluctuations caused by measurement errors. The vegetation growth index measurement establishes a standardization procedure, the diameter at breast height measurement is unified at 1.3 meters, and the leaf area index is measured using the LAI-2200 canopy analyzer. The gray correlation analysis method is used to calculate the data standardization processing, which converts the sequence of different dimensions into a comparable dimensionless sequence. The long short-term memory network training data is divided into training set, validation set and test set, with a ratio of 6:2:2. The network hyperparameters are determined by grid search optimization, with a learning rate of 0.001 and a batch size of 64. After generating the dynamic growth reference curve, smoothing processing is performed to eliminate high-frequency noise using the Savitzky-Golay filter. The curve verification uses the cross-validation method, dividing the historical data into multiple subsets to be used as the test set in turn. The curve prediction performance evaluation uses the root mean square error and the determination coefficient index, the root mean square error reflects the prediction accuracy, and the determination coefficient represents the model explanation ability. The dynamic growth reference curve is stored in the time series database format, each timestamp is associated with the predicted value and its confidence interval. The curve updating mechanism is set to retrain the model every month, incorporating the latest monitoring data to maintain prediction accuracy.

[0037] The time series alignment of steady-state components and vegetation growth indicators uses dynamic time warping method to solve the problem of inconsistent sampling frequency. The grey correlation analysis method adds a weight adjustment mechanism, giving higher weight to the correlation degree of the key growth period. The long short-term memory network introduces an attention mechanism to enhance the model's attention to important time steps. The dynamic growth benchmark curve integrates phenological information to distinguish the benchmark patterns of different growth stages such as leaf expansion, flowering, and fruiting. The curve output results are visualized, including historical growth trajectory, current prediction curve, and future trend line. The steady-state component data sources are expanded to remote sensing inversion products, adding microwave remote sensing soil moisture data and photosynthetically active radiation products. The vegetation growth indicators increase unmanned aerial vehicle remote sensing observations, including canopy height model and three-dimensional distribution of vegetation index. The grey correlation analysis method is improved to multi-resolution analysis, calculating the correlation degree at different time scales. The long short-term memory network architecture is upgraded to bidirectional LSTM, considering the influence of previous and subsequent time steps. The dynamic growth benchmark curve applies difference analysis to compare the deviation of actual growth values from benchmark values. The data quality control establishes an automated process, using multiple imputation methods to fill in missing values of steady-state component data. The vegetation growth indicator measurement implements indoor and outdoor calibration to ensure consistency of measurement results from different instruments. The grey correlation analysis method records detailed logs during the calculation process, including data preprocessing parameters and intermediate results of correlation degree calculation. The long short-term memory network monitors the loss curve and gradient changes during training to prevent gradient explosion or disappearance problems. The dynamic growth benchmark curve is deployed as a real-time service interface, supporting dynamic queries by forestry management departments.

[0038] The steady-state component data acquisition frequency is optimized according to the vegetation growth rate. The vegetation growth indicator measurement network layout considers spatial representativeness, setting a stratified sampling scheme according to site conditions. The grey correlation analysis method introduces time lag effect to analyze the delay characteristics of environmental factors on growth. The long short-term memory network adds a regularization term to improve the model's generalization ability. The dynamic growth benchmark curve is coupled with weather forecast data to realize extended prediction of growth trends. The steady-state component sensor network is calibrated regularly to maintain the accuracy and comparability of measurement data. The vegetation growth indicator database establishes metadata standards to record measurement time, location, method, and instrument information. The grey correlation analysis method realizes parallel computing to accelerate large-scale data correlation analysis. The long short-term memory network training uses GPU acceleration to improve model training efficiency. The dynamic growth benchmark curve system is integrated into the forestry Internet of Things platform, sharing data with other monitoring modules.

[0039] The dynamic growth benchmark curve is derived from the long short-term memory network output of the above example, and the curve data is stored in a time series format, containing vegetation index prediction values for two hundred and forty growth days. The generative adversarial network adopts a deep convolutional generative adversarial network architecture, and the input layer of the generator network receives a one-hundred-dimensional random noise vector, which is sampled from a standard normal distribution. The generator network contains four transpose convolutional layers, each using batch normalization and ReLU activation function, and the output layer uses the hyperbolic tangent function to generate a synthetic disturbance curve with the same length as the dynamic growth benchmark curve. The discriminator network contains four convolutional layers, each using a leaky ReLU activation function and a dropout layer, and finally outputs a true or false probability value through a fully connected layer. The adversarial network training process uses the Wasserstein distance loss function, and adds a gradient penalty term to enhance the stability of training. The generator network and the discriminator network are trained alternately, and each training batch contains sixty-four real dynamic growth benchmark curves and sixty-four generated curves. The number of training iterations is set to twenty thousand rounds, and the initial learning rate is zero. zero zero zero one, and is halved every five thousand rounds. After training, the generator network generates a variety of disturbance samples, which simulate vegetation growth curves under different intensity environmental stresses, including drought stress curves, waterlogging stress curves, and shade stress curves. The discriminator network can accurately distinguish between real growth curves and synthetic disturbance curves with an accuracy of more than eighty-five percent. The convolutional neural network uses a U-Net architecture for semantic segmentation, and the network input is the difference matrix of the real growth curve and the disturbance sample. The difference matrix calculates the absolute difference between the dynamic growth benchmark curve and each disturbance curve at each growth day, and the rows correspond to the growth day sequence and the columns correspond to different disturbance types. The U-Net encoder part contains five down-sampling stages, each using two convolutional layers and a max-pooling layer, and the number of feature maps increases from sixty-four layers to one thousand and twenty-four layers. The decoder part contains five up-sampling stages, each using transpose convolution for up-sampling and concatenating with the corresponding encoder feature maps, and finally the output layer uses the Sigmoid activation function to generate a segmentation mask.

[0040] The disturbance sensitive area identification is based on connected component analysis of segmentation mask, and the mask threshold is set to 0.5. The area with a pixel value greater than 0.5 is marked as a sensitive area. The sensitive area is subjected to morphological opening operation to remove noise points and closed operation to fill internal holes. The spatial range, sensitive intensity and duration index of each sensitive area are calculated. The spatial range is measured in hectares, the sensitive intensity takes the maximum difference value in the difference matrix, and the duration is measured in growth days. The disturbance sensitive area output is in the format of vector polygon, and the attribute table records the index values. The potential pest outbreak area is derived from the causal reasoning model output of Example 2, and the area boundary is stored in the form of polygon elements. The spatial overlay analysis uses the geographic information system overlay tool to jointly analyze the potential pest outbreak area layer and the disturbance sensitive area layer. The spatial intersection of the two layers is calculated, and the original attribute information of each intersection area is recorded. The fusion weight is calculated based on the area and the center distance of the overlaid area. The area weight takes the proportion of the intersection area to the original area, and the distance weight takes the reciprocal of the distance between the two region center points. The total fusion weight calculation formula is the product of the area weight and the distance weight. Referring to Table 1, the forestry management priority zoning map is generated based on the hierarchical rules of the fusion weight.

[0041] Table 1: Forestry management priority zoning table Fusion weight range Control priority level Color coding Response time requirement 0.8-1.0 Primary priority Red Within 24 hours 0.6-0.8 Secondary priority Orange Within 72 hours 0.4-0.6 Tertiary priority Yellow Within 168 hours <0.4 Routine monitoring zone Green Routine patrol The zoning map is created using a grid-to-vector method, which converts the weight distribution map into polygon features based on threshold ranges. The feature attribute table includes weight values, priority codes, area and perimeter fields. The map decoration adds scale, legend and north arrow elements, and the output format supports Shapefile and GeoJSON standards. The adversarial generative network training data enhancement uses time series distortion techniques to randomly scale and translate the dynamic growth benchmark curve. The convolutional neural network training uses a focal loss function to address class imbalance issues, with sensitive area pixel weights set to three times those of non-sensitive areas. The spatial overlay analysis optimization algorithm uses STR tree spatial indexing to accelerate large-scale polygon intersection operations. The fusion weight calculation introduces a terrain adjustment factor, with a coefficient of 1.2 for areas with a slope greater than 15 degrees. The zoning map update mechanism is set to generate a new version every week to reflect the latest monitoring data changes. The dynamic growth benchmark curve preprocessing includes outlier removal and trend smoothing, using the median absolute deviation method to detect outliers. The adversarial generative network generates samples that pass the authenticity test and are compared to historical abnormal growth curves for similarity. The convolutional neural network segmentation results are post-processed using morphological methods to remove small fragments with an area less than 0.1 hectares. The spatial overlay analysis considers time matching, only overlaying the same period of monitoring pest and disease areas with disturbance areas. The fusion weight calculation adds an ecological importance coefficient, with a one-level increase in the weight of areas within natural reserves. The adversarial generative network model is deployed as a distributed service, supporting multi-user concurrent generation of disturbance samples. The convolutional neural network inference uses TensorRT acceleration, with a single difference matrix processing time of less than 50 milliseconds. The spatial overlay analysis engine is based on a PostGIS database, supporting SQL spatial query interfaces. The fusion weight calculation module provides RESTful API services, receiving GeoJSON format input data. The forestry management and control priority zoning map is published as a Web map service, supporting hierarchical access control. The dynamic growth benchmark curve data is stored using a time series database structure, with separate storage for each forest plot. The adversarial generative network training implements model version management, recording hyperparameters and performance indicators for each training. The convolutional neural network segmentation results are quantified for uncertainty, outputting the prediction confidence of each pixel. The spatial overlay analysis results are topologically verified to ensure that the polygon boundaries meet the geometric validity rules. The forestry management and control priority zoning map is integrated into the mobile patrol system, supporting offline map browsing.

[0042] The disturbance sample generation scale is set to one thousand curves per growing season, covering major environmental stress scenarios. The convolutional neural network training data annotation uses a semi-automatic method combined with expert interpretation and automatic segmentation results. The spatial overlay analysis considers the scale effect and performs scale conversion on input data of different scales. The fusion weight calculation supports custom rules, allowing forestry management departments to adjust weight parameters. The zoning map symbol system uses national forestry standard patterns to ensure the standardization of map expression. The generative adversarial network generates samples for diversity evaluation, calculating the Fréchet distance between different samples. The convolutional neural network segmentation accuracy is evaluated through cross-validation, calculating the intersection over union and Dice coefficient indicators. The spatial overlay analysis performs error propagation analysis to evaluate the impact of input data accuracy on the results. The fusion weight calculation results are subjected to sensitivity analysis to test the degree of influence of weight parameter changes on zoning results. The forestry control priority zoning map mapping process is ISO quality system certified to ensure controllable results quality.

[0043] Referring to Figure 4 The graph takes the number of growth days as the horizontal axis and the vegetation growth index as the vertical axis, showing the differences in the growth trajectories of forestry vegetation under different environmental disturbances. The baseline curve is generated by fitting the steady-state component and vegetation growth index using a long short-term memory network, reflecting the theoretical trajectory of normal vegetation growth. The drought disturbance curve simulates the deviation of vegetation growth under drought stress, reflecting the inhibitory effect of water shortage on growth. The waterlogging disturbance curve presents the growth changes under waterlogging stress, reflecting the impact of waterlogged environments on vegetation growth. The value of this graph lies in providing a visual reference for forestry management and control: by comparing the growth deviations under different disturbances, we can accurately identify the sensitive growth periods of vegetation to drought and waterlogging, and then guide irrigation strategies and pest control timing, ultimately assisting in the scientific formulation of forestry control priority zoning.

[0044] Example 5: The deep reinforcement learning model is built based on the actor-critic framework, and a 3000-hectare management area of a provincial state-owned forest farm is taken as the implementation scenario. The forestry management priority zoning map input shows that there are seven first-level priority areas, thirteen second-level priority areas, and twenty third-level priority areas. The resource inventory state contains 5.5 tons of biological pesticide inventory, 800 cubic meters of irrigation water quota, and 35 intelligent spraying devices available. The state vector encoding converts the above information into a 384-dimensional feature vector, including the priority encoding of each grid cell, the normalized value of resource inventory, and the Boolean flag of device working status. The environment simulator of the deep reinforcement learning model simulates a 30-day management cycle, with each day divided into 24 decision-making time steps. The reward function design includes two components: the resource consumption efficiency component is calculated as the resource consumption ratio of the completed management area, and the region priority weighted component is calculated as the weighted sum of the completion degree of each priority area. The action space is defined as a discrete action set, including 16 basic operation actions such as "dispatch three devices to grid A01 for pesticide spraying" and "adjust the irrigation volume of grid B05 to 20 cubic meters per hour". The policy gradient algorithm uses the proximal policy optimization, and 5000 experience samples are used for each update, with a learning rate of 0.000025. The deep reinforcement learning model training process runs 150,000 iterations, each iteration including an environment interaction and data collection phase. Model parameters are stored in a distributed parameter server, and each worker node is responsible for policy evaluation of part of the state space. After training, the policy network outputs the resource allocation strategy, and the policy network includes three fully connected layers with 512 neurons in each layer, using residual connections to accelerate convergence. In the policy execution phase, the model reads the current state vector, calculates the probability distribution of each action through forward propagation, and selects the specific execution action using random sampling.

[0045] The task sequence generation is achieved by decoding the output action sequence of the policy network, and each action is converted into a specific task description. The task sequence format adopts the JSON standard, including task number, target area geographic coordinates, task type code, resource allocation amount, and execution time window fields. For example, the task sequence may include: "task ID: 20240520007, area: grid C12, operation: biological pesticide spraying, dose: zero point five liters per hectare, equipment: unmanned aerial vehicle group 03, time window: 09:00-11:00". The task sequence is prioritized, with first-priority area tasks occupying the first thirty sequence positions, second-priority area tasks occupying the middle forty positions, and third-priority area tasks arranged at the end. The control task sequence is issued to the forestry equipment control system, which is designed based on microservices. The task analysis service converts the JSON task description into device instructions, the irrigation task generates water quantity control signals, the fertilization task calculates fertilizer proportioning parameters, and the pest control task sets pesticide concentration. The Internet of Things communication uses the MQTT protocol to transmit instructions, and the communication payload is encrypted using AES-256. Each data packet is attached with a digital signature. After receiving the instructions, the forestry equipment returns an acknowledgment signal, and the device state monitoring service tracks the execution progress in real time.

[0046] An example of the intelligent irrigation system execution process is as follows: the system receives the instruction "grid D08 area irrigation eighty cubic meters", the central controller adjusts the pump station output pressure, the electromagnetic valve opens according to the preset program, and the flow sensor monitors the actual water consumption. When the irrigation volume reaches seventy-eight cubic meters, the system sends a status report request for subsequent instructions. An example of unmanned aerial vehicle plant protection operation is as follows: the flight control station loads the task route, the unmanned aerial vehicle group cooperates in flight, the multispectral camera monitors the spraying coverage in real time, and the obstacle avoidance program is automatically triggered when obstacles are encountered. An example of ground mobile equipment is as follows: the automatic fertilizer applicator dynamically adjusts the nitrogen, phosphorus, and potassium proportion according to the soil detection results, and the closed-loop control of travel speed and fertilizer quantity is achieved. The deep reinforcement learning model continuously optimizes the strategy through online learning mechanism, and the device execution feedback data is re-injected into the experience replay buffer. The model is incrementally trained every twenty-four hours to adapt to dynamic conditions such as resource inventory changes and equipment failures. The abnormal handling module monitors the task execution deviation, and when the actual execution effect deviates from the expected value by more than fifteen percent, the strategy recalculation process is triggered. The system maintains a digital twin model, and the state data of the physical space and the virtual space are synchronized in real time. The state vector encoding scheme performs feature engineering, the category variables are one-hot encoded, and the continuous variables are quantile normalized. The action space design considers the actual operation constraints to avoid assigning conflicting tasks to the same area at the same time. The reward function sets a penalty term to deduct reward points for resource waste behavior. The policy network parameter initialization uses the Xavier method to ensure the stability of forward propagation. The experience replay buffer uses priority sampling, and higher sampling weights are given to high learning value samples.

[0047] The device instruction transmission establishes a retry mechanism, and automatically selects a backup communication channel when the network is interrupted. The execution feedback data includes device operation logs, sensor readings, field images, and other multi-modal information. The state update service aggregates various feedback data to generate a new state vector for the deep reinforcement learning model to make decisions. The digital twin model performs real-time simulation and prediction to identify potential task conflict problems in advance. The resource allocation optimization considers seasonal factors, reducing irrigation task weights in the rainy season and increasing fertilization task priority in the growing season. The device scheduling considers mobile path optimization and uses the ant colony algorithm to calculate the shortest service path. The task sequence generation introduces a preemptive scheduling algorithm, and high-priority tasks can interrupt the execution of low-priority tasks. The communication protocol defines standard data formats to ensure interoperability between devices from different manufacturers.

[0048] The deep reinforcement learning model is deployed on edge computing nodes to reduce cloud communication delay. The task sequence version management uses the Git mechanism to support task rollback and version traceability. The device execution process is monitored for security, and abnormal operations trigger immediate emergency stop protection. The system records complete audit trails in the system operation logs to meet forestry management compliance requirements. The model inference performance reaches 200 state decisions per second, meeting real-time scheduling requirements. The task sequence visualization interface displays a Gantt chart to visually display the time arrangement and dependency of each task. The device resource utilization statistics function generates a daily resource consumption report. The system supports manual intervention mode, and administrators can temporarily adjust the automatically generated task sequence. The deep reinforcement learning model training uses historical job data to preheat and accelerate the initial convergence process. The state encoding adds spatiotemporal features to capture the periodic variation of resource demand. The action selection strategy adds exploration noise to promote the discovery of better resource allocation schemes. The reward function design considers long-term benefits and introduces a discount factor to balance immediate and long-term returns. The device control instructions include security check codes to prevent illegal instruction injection. The execution process performs double confirmation, and important operations require manual review and approval. The system establishes a disaster recovery mechanism, and the slave node automatically takes over the service when the master node fails. Data persistence uses multi-copy storage to ensure that task records are not lost. The deep reinforcement learning model is regularly evaluated for fairness to avoid long-term resource allocation in certain areas. The task sequence generation considers human resource allocation and coordinates with the patrol personnel's work plan. The system interface supports integration with the weather forecasting system to dynamically adjust task scheduling based on weather changes.

[0049] It is to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting; it is not intended to exclude myriad other embodiments of the present application that other inventors can develop based on the same general inventive concepts embodied by the described embodiments. That is, although the present application is described in terms of particular embodiments and illustrative figures, it should be apparent that the scope of the present application is not limited to these specific embodiments.

[0050] While the embodiments of the application have been shown and described herein, it will be understood by those skilled in the art that many changes, modifications, substitutions and alterations to these embodiments can be made without departing from the principles and spirits of the application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A forestry optimization management method based on artificial intelligence, characterized in that, include: Collect multi-source environmental monitoring data of forestry areas, including soil moisture, light intensity and vegetation growth indicators; The time series of the multi-source environmental monitoring data are aligned by a dynamic time warping algorithm to extract time series features and separate steady-state and transient components. Anomaly detection of the transient components is performed based on the isolated forest algorithm, abnormal data points are marked, and an anomaly distribution heatmap is generated. By combining historical pest and disease records with the aforementioned abnormal distribution heat map, a causal reasoning model is used to infer potential pest and disease outbreak areas. Based on the correlation between the steady-state components and vegetation growth indicators, a dynamic growth baseline curve is constructed. Perturbation samples of the dynamic growth baseline curve are generated using a generative adversarial network, and perturbation-sensitive regions are identified using a convolutional neural network. By integrating the potential pest and disease outbreak areas with the disturbance-sensitive areas, a forestry management priority zoning map is generated; A deep reinforcement learning model is used to optimize resource allocation in the forestry management priority zoning map and output a management task sequence. Based on the aforementioned control task sequence, forestry equipment is driven to perform irrigation, fertilization, or pest and disease control operations.

2. The forestry optimization management method based on artificial intelligence according to claim 1, characterized in that, The process of aligning the time series of the multi-source environmental monitoring data using a dynamic time warping algorithm, extracting time series features, and separating steady-state and transient components includes: Perform sliding window segmentation on the multi-source environmental monitoring data to obtain windowed time-series segments; Calculate the dynamic time warping distance between each windowed time segment and the preset standard template, and obtain the alignment offset after normalization; The phase deviation of the windowed timing segment is corrected based on the alignment offset to generate aligned timing data; The aligned time series data is processed by empirical mode decomposition to decompose high-frequency transient components and low-frequency steady-state components.

3. The forestry optimization management method based on artificial intelligence according to claim 2, characterized in that, The method of detecting anomalies in the transient components based on the isolated forest algorithm, marking anomalous data points, and generating anomaly distribution heatmaps includes: The high-frequency transient components are input into the isolated forest algorithm to construct a set of random split trees; Calculate the average path length of each data point in the tree set and then calculate the anomaly score. Filter out abnormal data points with scores exceeding a dynamic threshold and map them to spatial coordinates of forestry areas; Spatial coordinates are clustered using kernel density estimation to generate anomaly distribution heatmaps.

4. The forestry optimization management method based on artificial intelligence according to claim 3, characterized in that, The method of combining historical pest and disease records with the abnormal distribution heatmap, and using a causal reasoning model to infer potential pest and disease outbreak areas, includes: Match the spatial overlap area between the abnormal data points in the abnormal distribution heatmap and the historical pest and disease records; Environmental feature vectors of overlapping regions are extracted as training samples for the causal inference model; The trained causal reasoning model outputs the probability of pest and disease outbreaks in each region. Areas with a probability exceeding a preset threshold are marked as potential pest and disease outbreak areas.

5. The forestry optimization management method based on artificial intelligence according to claim 4, characterized in that, The step of constructing a dynamic growth baseline curve based on the correlation between the steady-state components and vegetation growth indicators includes: The correlation between the low-frequency steady-state components and vegetation growth indicators was calculated using grey relational analysis. Steady-state components with correlation exceeding a set threshold are selected as feature inputs; By fitting the mapping relationship between feature inputs and vegetation growth indicators through a long short-term memory network, a dynamic growth baseline curve is generated.

6. The forestry optimization management method based on artificial intelligence according to claim 5, characterized in that, The process of generating perturbation samples of the dynamic growth baseline curve using a generative adversarial network and identifying perturbation-sensitive regions using a convolutional neural network includes: The dynamic growth baseline curve is input into the generator of the adversarial generative network, and the synthesized perturbation curve is output. The discriminator distinguishes between the real growth curve and the synthetic perturbation curve, and iteratively optimizes the generator parameters; The trained generator produces diverse perturbation samples; Convolutional neural networks perform semantic segmentation on the difference regions between perturbed samples and real samples, and output perturbed sensitive regions.

7. The forestry optimization management method based on artificial intelligence according to claim 6, characterized in that, The process of integrating potential pest and disease outbreak areas with disturbance-sensitive areas to generate a forestry management priority zoning map includes: Spatial overlay analysis is performed on the potential pest and disease outbreak areas and the disturbance-sensitive areas. The fusion weight is calculated based on the area and center distance of the overlay area. Based on the fusion weight, a forestry management priority zoning map containing first- to third-level priorities is generated.

8. The forestry optimization management method based on artificial intelligence according to claim 7, characterized in that, The method employs a deep reinforcement learning model to optimize resource allocation in the forestry management priority zoning map, outputting a management task sequence, including: Encode the status of forestry equipment, resource inventory, and priority zoning map into a status vector; Define the reward function as a weighted sum of resource consumption efficiency and regional priority. Deep reinforcement learning models explore optimal resource allocation strategies through policy gradient algorithms; The decoding strategy outputs a sequence of control tasks.

9. The forestry optimization management method based on artificial intelligence according to claim 8, characterized in that, The process of driving forestry equipment to perform irrigation, fertilization, or pest and disease control operations based on the control task sequence includes: The system analyzes and controls the equipment commands and parameters in the task sequence, sends the commands to the corresponding forestry equipment through the Internet of Things protocol, and monitors the equipment execution feedback and updates the task status in real time.

10. A forestry optimization and management system based on artificial intelligence, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the artificial intelligence-based forestry optimization management method as described in any one of claims 1 to 9.

Images