Camellia oleifera forest growth and growth environment monitoring method

Abstract

The application discloses a kind of camellia forest growth and growth environment monitoring method, belong to wisdom forestry monitoring technical field.In the camellia forest land layout monitoring terminal, continuously synchronously collect growth environment factor and camellia plant growth characteristic data, construct with time as index environment-growth coupling time series dataset;Utilize unsupervised learning algorithm to construct environment baseline model, real-time detection environment deviates and exports stress early warning;Utilize causal inference algorithm to establish growth response model, learn the lag mapping relationship between environment factor sequence and growth rate of change;When growth is abnormal, the contribution of each environmental factor is calculated by characteristic disturbance method, and the attribution analysis result is output;Using reinforcement learning algorithm, with environment state and growth state as input, with preset growth target as reward function, generate optimal management action instruction.The application solves the problem that the lagging effect of environmental changes on camellia growth cannot be quantitatively modeled, and the cause of growth abnormalities cannot be traced to specific environmental factors.

Description

Technical Field

[0001] This invention relates to the field of smart agricultural and forestry monitoring technology, and in particular to a method for monitoring the growth and growth environment of camellia oleifera forests. Background Technology

[0002] The growth and development of Camellia oleifera are significantly influenced by environmental factors. Changes in air temperature and humidity, light, soil moisture, and nutrients affect its growth. However, these effects are not immediate but exhibit a clear lag: prolonged drought may only manifest as a decrease in fruit enlargement rate several days later, and the adverse effects of low temperatures during flowering on fruit set often take several weeks to fully materialize. A complex nonlinear temporal correlation exists between environmental factors and plant growth.

[0003] Existing monitoring technologies for camellia oleifera forests are insufficient to effectively address this characteristic. CN118864141A discloses an intelligent system for integrated management of camellia oleifera forest stands, which collects environmental and soil data through patrol vehicles, calculates a growth index using a geometric mean formula, and compares it with a threshold. This scheme employs a discrete patrol method, with time intervals between adjacent data collections, making it impossible to obtain a continuous sequence of environmental factor changes and construct a complete process of environmental change. When abnormal growth is subsequently detected, the key environmental event causing the abnormality may have occurred several days prior, making it impossible to trace the specific cause. Furthermore, this scheme compresses multidimensional environmental information into a single scalar index, losing independent change information for each environmental factor. It cannot determine whether the abnormal growth is caused by drought, waterlogging, nutrient deficiency, or insufficient sunlight, nor can it quantify the contribution of each factor.

[0004] Therefore, existing technologies have technical problems such as the inability to quantitatively model the lagged effects of environmental changes on the growth of camellia oleifera, and the inability to trace the causes of abnormal growth to specific environmental factors. Summary of the Invention

[0005] The embodiments of the present invention provide a method for monitoring the growth and growth environment of Camellia oleifera forests, aiming to solve the problem that the lagged effects of environmental changes on the growth of Camellia oleifera cannot be quantitatively modeled, resulting in the inability to trace the cause of abnormal growth to specific environmental factors, and thus the inability to achieve accurate attribution and differentiated intervention.

[0006] To achieve the above objectives, the present invention provides a method for monitoring the growth and growth environment of Camellia oleifera forests, comprising the following steps: Monitoring terminals were deployed in the camellia oleifera forest to continuously and synchronously collect data on growth environment factors and growth characteristics of camellia oleifera plants at predetermined intervals. The growth environment factors and growth characteristics of camellia oleifera plants collected at the same time were aligned on the time axis to construct a time-coupled environment-growth data series indexed by time. Using the historical environmental factor data sequence in the environment-growth coupled time series dataset, an environmental baseline model is constructed by training a predetermined unsupervised learning algorithm. The environmental baseline model is used to characterize the normal environmental dynamic range of the camellia oleifera forest at different time series, and to predict the normal fluctuation range of environmental factors at the current moment based on the real-time input environmental factor data sequence. Using the environment-growth coupled time series dataset, a growth response model is established through a predetermined causal inference algorithm; the growth response model is used to characterize the nonlinear mapping relationship and hysteresis effect between the environmental factor data sequence and the rate of change of growth characteristic data. The real-time collected growth environment factor data sequence is input into the environmental baseline model. If the real-time growth environment factor data sequence deviates from the normal environmental dynamic range, environmental stress early warning information is output. When abnormal growth characteristic data is detected, the current and historical environmental factor data sequences are input into the growth response model to identify the key environmental factors that cause abnormal growth and their contribution, and to form attribution analysis results. Based on the environmental stress early warning information and / or attribution analysis results, the optimal management action instructions are generated using a reinforcement learning algorithm. The reinforcement learning algorithm takes the current environmental state and growth status as input, uses the preset growth target trajectory as the reward function, and outputs differentiated agricultural management actions. The optimal management action instruction is executed, and the environmental and growth change data after execution are fed back to the environmental baseline model and the growth response model for iterative updates.

[0007] Furthermore, the unsupervised learning algorithm employs a variational autoencoder or a long short-term memory network-autoencoder. The environmental baseline model constructs a normal manifold in a high-dimensional environmental feature space by learning the temporal distribution characteristics of historical growth environmental factor data sequences. When the reconstruction error of the real-time growth environmental factor data sequence in the high-dimensional environmental feature space exceeds a set threshold, it is determined to deviate from the normal environmental dynamic range.

[0008] Furthermore, the growth response model employs a long short-term memory network or a temporal convolutional network, using the growth environment factor data sequence within a sliding time window as input features and the change rate of growth feature data within a preset lag time period after the window as output labels for training, thereby learning the delayed effect of environmental changes on growth.

[0009] Furthermore, the identification of key environmental factors leading to abnormal growth and their contribution includes: keeping the parameters of the growth response model fixed, perturbing each dimension of the input growth environmental factor data sequence, calculating the difference in the growth change rate of the model output before and after the perturbation of each dimension, and using the difference value as the contribution of the growth environmental factor.

[0010] Furthermore, the reinforcement learning algorithm employs a deep Q-network; the state space includes real-time values ​​of current growth environment factors, real-time values ​​of growth characteristics, and the growth and development stage of the camellia oleifera; the action space includes the application dosage, application timing, and pruning intensity level of different types of water and fertilizer; the reward function is configured as a monotonically decreasing function of the negative deviation between the actual growth trajectory and the preset optimal growth trajectory after the current action is executed.

[0011] Furthermore, the monitoring terminals are deployed in a fixed or semi-fixed manner. Each monitoring terminal uploads the collected growth environment factor data and growth characteristic data to the edge computing node or cloud server in real time through low-power wireless self-organizing network communication. The environmental baseline model and growth response model are trained, updated, and reinforcement learning decisions are made on the edge computing node or cloud server.

[0012] Furthermore, the growth environment factors include air temperature and humidity, light intensity, wind speed, soil temperature and humidity, soil nitrogen, phosphorus and potassium content, soil pH value and / or precipitation; the growth characteristics data of the camellia oleifera plant include plant height, crown width, leaf area index, number of fruits and / or fruit enlargement rate.

[0013] Furthermore, the environmental stress early warning information includes the type of environmental stress and the expected stress intensity; the attribution analysis results include the dominant environmental factors leading to abnormal growth, the ranking of the contribution of each environmental factor, and the confidence interval.

[0014] Furthermore, the environmental baseline model and the growth response model are automatically retrained according to a preset period or after the accumulated new data reaches a set threshold.

[0015] Furthermore, it also includes: displaying the environmental stress warning information, attribution analysis results, and generated optimal management action instructions through a visual interface, and receiving confirmation or correction instructions from the manager, using the corrected management actions as positive or negative reward samples for reinforcement learning, and using them to optimize the decision boundary of the reinforcement learning policy network.

[0016] The above technical solution has the following technical effects: First, a growth response model was constructed, quantifying the lag mapping relationship between environmental factor data sequences and the rate of change of growth characteristics into calculable model parameters. Existing technologies cannot determine how long after an environmental change occurs and how much it will affect growth. This invention, through a sliding time window and lag label settings, enables the model to learn the specific lag duration and response magnitude.

[0017] Secondly, it enables the attribution and source tracing of abnormal growth. Employing the feature perturbation method, it calculates the contribution of each environmental factor to the abnormal growth by successively changing the environmental factors in the input and observing the changes in the model output, outputting the dominant factor and its contribution value. Existing technologies can only determine whether the growth is below a threshold, but cannot distinguish whether it is caused by drought, waterlogging, or nutrient deficiency. This invention fills this gap.

[0018] Third, by continuously and synchronously collecting data, an environment-growth coupled time series dataset was constructed, providing a complete data foundation for hysteresis modeling and attribution analysis, which is different from the discrete inspection method of existing technologies. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating a method for monitoring the growth and growth environment of camellia oleifera forests according to an embodiment of the present invention. Detailed Implementation

[0020] To further illustrate the various embodiments, the present invention provides accompanying drawings. These drawings are part of the disclosure of the present invention and are mainly used to illustrate the embodiments, and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementations and the advantages of the present invention. Components in the drawings are not drawn to scale, and similar component symbols are generally used to represent similar components.

[0021] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments.

[0022] Example 1: This embodiment uses a general cross-media operation scenario as the application object to describe the specific implementation of the present invention.

[0023] Figure 1 This is a flowchart illustrating a method for monitoring the growth and growth environment of camellia oleifera forests according to an embodiment of the present invention. Figure 1 As shown, this method achieves continuous and synchronous collection of environmental and growth data through monitoring terminals deployed in camellia oleifera forests, and realizes intelligent decision-making for environmental stress early warning, abnormal growth attribution analysis, and management actions based on machine learning models, including: Within the target camellia oleifera forest, integrated monitoring terminals were deployed at a density of 0.5 to 1 hectare. The monitoring terminals were fixed in place, with a 2.5-meter-high pole buried at the center of each plot, and the terminal equipment secured to the pole. The terminal equipment collected data at three levels: near-surface microclimate data at 0.3 meters above the ground, environmental data from the middle of the canopy at 1.5 meters above the ground, and environmental data from above the canopy at 2.0 meters above the ground. Soil sensors were buried in areas with dense root distribution among the camellia oleifera plants, specifically at soil depths of 10 cm, 30 cm, and 50 cm, 0.5 to 1.0 meters from the trunk.

[0024] The monitoring terminal's hardware consists of: an environmental sensor group, specifically including an air temperature and humidity sensor, a light intensity sensor, a wind speed sensor, a rainfall sensor, a soil temperature and humidity sensor, a soil nutrient sensor, and a soil pH sensor; a plant growth acquisition module, specifically including an industrial-grade high-definition camera and a laser rangefinder; a data acquisition and transmission unit, including a low-power microcontroller and a LoRa wireless communication module; and a power supply unit, including a 30W solar panel and a 60Ah battery.

[0025] The specific implementation method of continuous synchronous data acquisition is as follows: the data acquisition cycle is set to once every 30 minutes. At the beginning of each acquisition cycle, the microcontroller wakes up each sensor and acquisition module. First, it triggers the environmental sensor group to collect air temperature and humidity, light intensity, wind speed, soil temperature and humidity, soil nitrogen, phosphorus and potassium content, soil pH value, and rainfall. The above environmental factor data are collected sequentially within 100 milliseconds. Then, the plant growth acquisition module is triggered. The distance from the monitoring terminal to the top and crown edge of the selected camellia oleifera plant is measured by a laser rangefinder. Combined with the known terminal installation height and angle, the plant height and crown width are calculated. At the same time, a camera captures one image each from the front and side of the camellia oleifera plant. The image resolution is 1920×1080 pixels. The image data is processed by a local embedded image processing algorithm to extract leaf area index, number of fruits, and fruit enlargement rate. The method for extracting the number of fruits is as follows: the YOLOv8 target detection model is used to identify and count the camellia oleifera fruits in the image. This model is pre-trained with 3000 labeled camellia oleifera fruit images, and the detection accuracy reaches over 92%. The method for extracting the fruit expansion rate is as follows: the pixel area of ​​the same fruit in multiple consecutively acquired images is tracked, and the actual transverse diameter is converted by combining the known object distance and focal length, and the change in the transverse diameter of the fruit over time is recorded.

[0026] All collected data is packaged into a single data record, with each record containing a collection timestamp, measurement point number, values ​​of various environmental factors, and values ​​of various growth characteristics. The data is stored locally on the terminal in JSON format and simultaneously uploaded to the forest edge computing node via a LoRa wireless communication module using a multi-hop relay method. The edge computing node is deployed within the forest management station, equipped with a GPU computing unit and 256GB of solid-state storage, for receiving and managing data uploaded from various monitoring terminals.

[0027] After receiving data uploaded by each monitoring terminal, the edge computing node performs data preprocessing and time-series dataset construction. The preprocessing steps are as follows: First, missing value handling is performed. For time points with no more than 3 consecutive missing periods, linear interpolation is used to fill in the missing values. For time periods with more than 3 consecutive missing periods, the data is marked as invalid and discarded. Second, outlier removal is performed using a method based on interquartile range (IQR). For each univariate time series data dimension, the 25th percentile Q1 and 75th percentile Q3 are calculated, and the IQR is calculated as IQR = Q3 - Q1. Data points with IQR values ​​less than Q1 - 1.5 × IQR or greater than Q3 + 1.5 × IQR are marked as outliers and replaced with the mean of the two valid data points before and after the outlier.

[0028] The preprocessed environmental factor data and growth characteristic data were precisely aligned according to the acquisition timestamp. Due to slight differences in response time between different sensors, the time when the air temperature and humidity sensor completed its acquisition was used as the reference timestamp during alignment, and the acquisition values ​​of other sensors within 1 second before and after this reference time were grouped into the same time point.

[0029] The aligned data is organized into a two-dimensional table structure with time as the row index and measurement points and variables as column indexes. Specifically, each monitoring terminal independently constructs a data matrix, where rows correspond to collection time points, arranged in ascending chronological order. The first m columns correspond to m environmental factor variables, and the last n columns correspond to n growth characteristic variables. For subsequent model training, the data matrices from each monitoring terminal are concatenated along the time dimension to form an environment-growth coupled time-series dataset covering multiple Camellia oleifera growth cycles (including at least two complete spring shoot growth periods, budding and flowering periods, fruit enlargement periods, and oil conversion periods).

[0030] An environmental baseline model is used to characterize the normal environmental dynamic range of camellia oleifera forests at different time series. This embodiment employs a variational autoencoder as an unsupervised learning algorithm. A variational autoencoder is a generative model capable of learning the probability distribution of high-dimensional data in a low-dimensional latent space and determining whether the input data belongs to a normal distribution by reconstructing the probabilities.

[0031] The variational autoencoder network structure is as follows: The encoder consists of three convolutional layers and one fully connected layer. The first convolutional layer has a 3×3 kernel size, a stride of 1, and 16 output channels, using the ReLU activation function. The second convolutional layer has a 3×3 kernel size, a stride of 1, and 32 output channels, also using the ReLU activation function. The third convolutional layer has a 3×3 kernel size, a stride of 1, and 64 output channels, also using the ReLU activation function. The fully connected layer flattens the feature maps output from the convolutional layers and maps them to the latent space dimension, which is set to 8. The encoder outputs the mean vector μ and the log-variance vector logσ² of the latent variables. The decoder structure is symmetrical to the encoder, containing a fully connected layer and three transposed convolutional layers, reconstructing the latent variable sample values ​​into data with the same dimensions as the input.

[0032] The input data for the environmental baseline model is prepared as follows: All historical environmental factor data sequences are extracted from the environment-growth coupling time-series dataset. Each training sample represents environmental factor data within a sliding time window, with a window length of 24 hours. The sliding step size is 12 hours, and adjacent windows overlap by 50%. Each sample has a dimension of 48×E, where E is the number of environmental factor variables.

[0033] Model training process: Unsupervised learning is performed on the above training samples using a variational autoencoder. The loss function consists of two parts: reconstruction loss and KL divergence loss. The reconstruction loss uses mean squared error to calculate the difference between the input data and the decoder output data. The KL divergence loss measures the difference between the latent variable distribution of the encoder output and the standard normal distribution. The total loss function is the reconstruction loss plus the KL divergence loss multiplied by a coefficient of 0.001. Training uses the Adam optimizer, with an initial learning rate of 0.001, a batch size of 64, and 200 training epochs. After each training epoch, the loss value is calculated on the validation set, which consists of 20% of the samples randomly selected from the training set. Training stops when the validation set loss does not decrease for 10 consecutive epochs, and the model parameters of the epoch with the lowest loss are restored.

[0034] After training, the environmental baseline model can learn the temporal variation patterns of various environmental factors under normal environmental conditions. For any input environmental factor data sequence window, the model calculates the reconstruction error through an encoder-decoder structure. The reconstruction error is defined as the root mean square error between the input window and the reconstruction window. To obtain the normal environmental dynamic range, the reconstruction error is calculated on all samples in the training set, the distribution of the reconstruction error is statistically analyzed, and the reconstruction error value corresponding to the 95th percentile is set as the threshold. When the real-time collected environmental factor data sequence is input into the model, if its reconstruction error exceeds this threshold, it is determined that the current environmental state deviates from the normal dynamic range, and environmental stress warning information is output. The warning information includes a preliminary judgment of the stress type: the stress type is initially judged based on the environmental factor dimension that contributes the most to the reconstruction error. For example, if the reconstruction error of the soil moisture dimension accounts for more than 60%, it is judged as drought stress or waterlogging stress; if the reconstruction error of the light intensity dimension accounts for more than 50%, it is judged as insufficient light stress.

[0035] The growth response model is used to characterize the nonlinear mapping relationship and lag effect between environmental factor data sequences and the rate of change of growth characteristic data. This embodiment uses a Long Short-Term Memory (LSTM) network as the basic model for the causal inference algorithm. LTM networks, through their gating mechanism, can effectively learn long-term dependencies in time series, making them suitable for modeling the delayed effects of environmental changes on plant growth.

[0036] The network structure of the Long Short-Term Memory (LSTM) network is as follows: The input layer receives a sequence of environmental factor data with a length of T=96, corresponding to 48 hours. The input features are environmental factor variables. The main body of the network contains two LSM layers: the first layer has 128 hidden units, and the second layer has 64 hidden units. Both layers have a dropout rate of 0.2 to prevent overfitting. The output layer is a fully connected layer that maps the output of the last time step of the second LSM layer to the output dimension, where the output is the rate of change of the growth feature to be predicted.

[0037] The training data is constructed as follows: training samples are extracted from the environment-growth coupled time-series dataset. The input feature X of each sample is a T×E matrix, representing environmental factor data for T consecutive collection periods. The output label Y of each sample is a vector of length 5, representing the growth change rate within a preset lag time period starting from the end of the Tth collection period. The length of the lag time period is set according to different growth and development stages of Camellia oleifera: the lag time for the spring shoot growth period is set to 7 days, because the impact of environmental changes on plant height and canopy width takes about a week to manifest; the lag time for the budding and flowering period is set to 10 days; and the lag time for the fruit enlargement period is set to 15 days. The growth change rate is calculated as follows: at the end of the lag time period, the value of the growth index is subtracted from the value at the beginning of the lag time period, and then divided by the length of the lag time period.

[0038] The sample construction adopts a sliding window method. The window starts from the first time point of the dataset and slides forward in steps of 24 collection cycles (12 hours) until the end of the window plus the lag time period exceeds the range of the dataset. Each monitoring terminal constructs samples independently, and the samples from all terminals are merged to form the training set.

[0039] Model training process: The loss function used is mean squared error, which is the sum of squares of the differences between the predicted and actual rates of change in crop growth. The optimizer used is Adam, with an initial learning rate of 0.0005, a batch size of 32, and 300 training epochs. An early stopping strategy is employed: training stops when the validation set loss does not decrease for 20 consecutive epochs. The validation set consists of 20% randomly selected samples from the training set, and the time range of the validation set samples does not overlap with that of the training set to ensure the model's temporal generalization ability.

[0040] After training, the growth response model can predict the rate of change in growth over a future period based on the input environmental factor data sequence. This predictive ability forms the basis of causal inference: by perturbing specific dimensions of the input environmental factors and observing the changes in the model output, the contribution of each environmental factor to the growth change can be quantified.

[0041] When abnormalities are detected in growth characteristic data, attribution analysis is performed. The criteria for determining abnormal growth are: in the currently collected growth characteristic data, any indicator is lower than the mean of the historical normal range minus 1.5 standard deviations for this growth and development period, or higher than the mean plus 1.5 standard deviations. The historical normal range is obtained by statistically analyzing data from at least two complete growth cycles.

[0042] The attribution analysis employs the feature perturbation method, with the following specific steps. First, environmental factor data sequences for L time windows preceding the current anomaly are extracted from the environment-growth potential coupled time series dataset. L is set to 3, each window has a length of T=96, and the step size between adjacent windows is 48. For each window, it is used as the baseline input. X base Then, keeping the model parameters fixed, for the i-th environmental factor variable, construct a perturbation input. X perturbed_i :Will X base The values ​​of the i-th variable at all time points are multiplied by a perturbation coefficient, with the perturbation coefficient set to 0.9 and 1.1 respectively, for two perturbations. X base and X perturbed_i Input the data into the long-response model and calculate the output growth rate of change. The contribution of the i-th environmental factor variable is calculated using the following formula: Contribution_i=0.5×| f ( X perturbed_i 0.9 )- f ( X base )| / f ( X base )+0.5×| f ( X perturbed_i 1.1 )- f ( X base )| / f ( X base ), in, f (·) indicates the output of the growth response model; X perturbed_i Indicates a disturbance input; X base Use as the baseline input.

[0043] After calculating the contribution of each environmental factor for each of the three windows, the average value is taken as the final contribution of that factor. The contributions of all environmental factors are normalized to a sum of 100%, and then sorted by contribution from largest to smallest. The top three environmental factors in terms of contribution are identified as the key environmental factors causing the current abnormal growth, and their contribution values ​​are the contribution percentages.

[0044] The output attribution analysis results include: the dominant environmental factor (i.e., the environmental factor with the largest contribution), a list of the contribution rankings of each environmental factor, and confidence intervals. The confidence interval is calculated as follows: the standard deviation of the contribution calculation results for the three windows is calculated, and the mean plus or minus 1.96 times the standard deviation is used as the 95% confidence interval.

[0045] This embodiment uses a deep Q-network algorithm to generate optimal management action instructions. Deep Q-networks are reinforcement learning algorithms that combine deep learning with Q-learning, capable of handling high-dimensional continuous state spaces and suitable for complex decision-making problems in camellia oleifera forest management.

[0046] The state space of a deep Q-network is defined as follows: The state vector consists of three parts. The first part is the current environmental state, including the real-time values ​​of environmental factors and their trends over the past 24 hours (increasing, decreasing, or remaining stable), encoded as +1, -1, and 0, respectively. The second part is the current growth state, including the real-time values ​​of five growth features and their rates of change over the past 7 days. The third part is the time state, including the current growth and development stage of the camellia (spring shoot growth stage encoded as 1, budding and flowering stage encoded as 2, fruit enlargement stage encoded as 3, oil conversion stage encoded as 4), the time interval since the last management action, and the seasonal factor.

[0047] The action space of a deep Q-network is defined as follows: actions are discretized combinations of management operations. Water and fertilizer application actions include: irrigation amount, nitrogen fertilizer application amount, phosphorus fertilizer application amount, and potassium fertilizer application amount. Forest management actions include: pruning intensity and pest and disease control measures. The permutations and combinations of these actions constitute the action space. To reduce combinatorial explosion, action masking technology is used: based on the current growth stage and growth status, unreasonable actions are prohibited, such as prohibiting heavy pruning during the fruiting period.

[0048] The deep Q-network architecture is as follows: The input layer receives a 40-dimensional state vector. The hidden layers consist of three fully connected layers: the first layer has 256 nodes, the second layer has 128 nodes, and the third layer has 64 nodes. The ReLU activation function is used for all layers. The output layer has the same number of nodes as the action space and outputs the Q-value estimate for each action. The network weights are initialized using Xavier.

[0049] The configuration of the reward function is the core of reinforcement learning decision-making. The reward function is designed as follows: after the current action is executed, and after an observation period (e.g., 15 days for spring shoot growth, 10 days for budding and flowering, 20 days for fruit enlargement, and 30 days for oil conversion), the negative deviation between the actual growth trajectory and the preset optimal growth trajectory is calculated. The preset optimal growth trajectory is determined by expert experience and specifically includes: the optimal curve for plant height growth rate, the optimal cumulative curve for fruit quantity, and the optimal curve for fruit enlargement rate. For each growth indicator, the absolute value of the deviation between the actual value and the optimal value is calculated. The reward value is calculated using the following formula: Reward=-Σ w j ×| Actual j - Optimal j | / Optimal j , in, j Traversing growth indicators, w The weights of each indicator, Actual j This represents the actual growth value. Optimalj The optimal growth value is preset. If any indicator in the actual growth trajectory exceeds the optimal value by a certain range without causing a significant decline in other indicators, an additional positive reward of +10 is given. If the implemented management actions cause environmental stress, such as waterlogging due to over-irrigation, a negative reward of -20 is given.

[0050] The training process of a deep Q-network is as follows: An experience replay mechanism is used, where the experience tuples obtained from each interaction, including state, action, reward, and next state, are stored in an experience pool with a capacity of 10,000 samples. During each training iteration, a batch of samples is randomly sampled from the experience pool, and the target Q-value and the current Q-value are calculated. The network parameters are updated using mean squared error as the loss function. The formula for calculating the target Q-value is: TargetQ=Reward+γ×max_a'Q_target(s',a'), Here, γ is the discount factor, set to 0.95, representing the model's emphasis on future rewards. Q_target is the output of the target network, which has the same structure as the main network, with its parameters copied from the main network every 100 steps. The exploration strategy adopts an ε-greedy strategy, with an initial ε=1.0, decreasing exponentially by 0.995 with each training step, and a minimum ε=0.01.

[0051] During the actual deployment and operation phase, the system collects the current state vector for each management decision cycle, inputs it into a deep Q-network, and the network outputs Q-value estimates for each action. The action with the largest Q-value is selected as the optimal management action instruction. The output management action instruction is concretized into executable agricultural operations, such as applying 100 grams of nitrogen fertilizer, 30 grams of phosphorus fertilizer, and 40 grams of potassium fertilizer per plant to the camellia oleifera plants in area 3 of the measuring point, irrigating 10 liters per plant, and performing light pruning.

[0052] The environmental baseline model and the growth response model are automatically retrained according to a preset cycle. The retraining cycle is set to once every 30 days. The conditions for triggering retraining are: more than 30 days have passed since the last training, or the accumulated amount of new data reaches a set threshold of 10,000 valid samples.

[0053] The specific retraining process is as follows: Newly accumulated monitoring data since the last training is extracted from the database and merged with the existing historical data to form an expanded training dataset. The model is retrained using the same parameter configuration as the initial training, such as window length, sliding step size, network structure, and learning rate. After model training, its performance is evaluated on the latest validation set. For the environmental baseline model, the evaluation metrics are the stability of the reconstruction error and the anomaly detection accuracy. For the growth response model, the evaluation metric is the root mean square error between the predicted growth change rate and the actual growth change rate. If the new model's performance metrics are better than the old model, the new model is deployed to replace the old model; if the new model's performance degrades, the old model is retained, and the reason for the performance degradation is logged for manual review.

[0054] This embodiment also includes a visualization and interaction module. This module runs on the server of the forest management station in the form of a web dashboard, which can be accessed by managers through a browser. The dashboard displays the following content: a forest map and the distribution of each monitoring terminal, real-time environmental factor data curves for each measuring point, real-time growth characteristic data curves for each measuring point, a list of environmental stress early warning information, charts of attribution analysis results such as pie charts or bar charts of the contribution of key environmental factors, and optimal management action instructions recommended by the deep Q-network.

[0055] Managers can confirm or modify the system-recommended actions on the dashboard. A confirmation action indicates that the manager approves the recommended action, which will be sent to the execution terminal (such as a smart irrigation valve or automatic fertilizer applicator) or to the maintenance personnel's mobile application as work instructions. A modification action indicates that the manager adjusts the recommended action, for example, changing the system-recommended nitrogen fertilizer from 100 grams per plant to 80 grams per plant, or refusing to execute a certain action.

[0056] When a manager modifies or rejects a recommended action, the system uses this interaction record as a correction sample for reinforcement learning. Specifically, the system-recommended action is denoted as `a_recommended`, and the action actually performed by the manager is denoted as `a_actual`. The state `s` at the time of the interaction, the recommended reward value `R`, and the modification record are stored as a correction sample in a dedicated correction experience pool. Samples in the correction experience pool participate in training with higher sampling weights during the training of the deep Q-network, enabling the model to learn the constraints of the manager's professional experience and gradually optimize the decision boundary towards a direction that better aligns with actual production needs.

[0057] Example 2 Based on Example 1, this embodiment replaces the variational autoencoder with a long short-term memory network-autoencoder for the environmental baseline model, and replaces the long short-term memory network with a temporal convolutional network for the growth response model, in order to adapt to different data processing scenarios.

[0058] The Long Short-Term Memory (LSTM) network-autoencoder is a temporal anomaly detection model that combines a LSM network with an autoencoder structure. The encoder consists of two LSM network layers: the first layer has 64 hidden units, and the second layer has 32 hidden units. The encoder receives the input sequence X=[x1,x2,...,x...]. T The encoder updates the hidden state step by step over time, ultimately outputting a fixed-dimensional encoding vector h that captures the temporal features of the entire input sequence. The decoder also consists of two layers of long short-term memory networks, with a structure symmetrical to the encoder. It receives the encoding vector h as the initial hidden state and gradually reconstructs the output sequence X'=[x'1,x'2,...,x'...]. T The goal of model training is to minimize the reconstruction error, which is the mean squared error between the input sequence and the reconstructed sequence.

[0059] After training, for any input sequence, the model calculates the reconstruction error at each time point. Unlike variational autoencoders, Long Short-Term Memory (LSTM) autoencoders do not output reconstruction probabilities; instead, they output a time series of reconstruction errors. The normal dynamic range is determined as follows: For all samples in the training set, the reconstruction error distribution at each time point is statistically analyzed, and the 90th percentile of the reconstruction error at each position is calculated, forming a dynamic threshold curve of length T. When the reconstruction error at any time point of the real-time input sequence exceeds the threshold at the corresponding position, that time point is marked as abnormal. If more than three consecutive abnormal time points are identified, the environmental state is determined to deviate from the normal dynamic range, and a warning is issued.

[0060] Temporal convolutional networks employ causal convolution and dilated convolution structures, achieving a large receptive field while maintaining computational efficiency. The temporal convolutional network structure used in this embodiment is as follows: The input layer receives a T×E input sequence. The network contains four residual blocks, each containing two dilated convolution layers with dilation coefficients set to 1, 2, 4, and 8 in the four residual blocks, respectively. Each convolutional layer has a kernel size of 3, a stride of 1, and uses causal padding, meaning it only pads the left side of the sequence to ensure that future information is not utilized. Each convolutional layer is followed by a batch normalization layer, a ReLU activation function, and a dropout layer with a dropout rate of 0.2. Residual connections sum the input and output of each residual block before passing it to the next residual block. The output of the last residual block passes through a global average pooling layer to compress the temporal dimension, and then through a fully connected layer to map to a 5-dimensional output, corresponding to the change rates of five growth features.

[0061] Temporal convolutional networks (TCNNs) offer advantages over long short-term memory (LSM) networks, including faster training speeds, more stable gradients, and the ability to process data in parallel, making them particularly suitable for scenarios with large datasets. In this embodiment, when the number of monitoring terminals exceeds 50 and the daily number of newly added samples exceeds 2400, the system automatically switches to the temporal convolutional network model to reduce training time. Parameter transfer between the two models can be achieved through a model converter: the hidden state sequence of the LSM network is used as temporal features to initialize some of the convolutional kernel parameters of the temporal convolutional network, thereby enabling rapid transfer learning.

[0062] This embodiment also introduces a dynamic model selection strategy. During runtime, the system automatically selects the optimal model combination based on real-time data characteristics. The specific mechanism is as follows: The system maintains a model performance evaluation module, which evaluates the anomaly detection accuracy of the variational autoencoder model and the long short-term memory network-autoencoder model, as well as the root mean square error of the prediction of the long short-term memory network model and the temporal convolutional network model, on the validation data of the day, every 24 hours. The model with the higher accuracy is selected as the active environment baseline model for the current period, and the model with the smaller prediction error is selected as the active growth response model for the current period. This dynamic selection mechanism enables the system to adapt to the seasonal and interannual variations of the ecological conditions in camellia oleifera forests, consistently maintaining optimal monitoring and prediction performance.

[0063] The remaining steps not detailed in this embodiment, such as data acquisition, time series dataset construction, attribution analysis, reinforcement learning decision-making, model iterative update, and visualization interaction, are the same as in Embodiment 1 and will not be repeated here.

[0064] Example 3 This embodiment uses a camellia oleifera forest in a certain region as an application scenario to explain in detail the complete execution process of the method of the present invention.

[0065] I. Scene Setting This camellia oleifera forest is located in the heart of the Wuling Mountains. The forest covers a total area of ​​120 mu (approximately 8 hectares), with an average altitude of 450 meters and a gentle slope ranging from 5° to 15°. About 2,000 camellia oleifera trees, all eight years old and in their stable production period, are planted within the forest. The region has a mid-subtropical humid monsoon climate. Influenced by the Wuling Mountains' topography, the average annual temperature is 16.5℃, and the annual precipitation is 1350 mm, concentrated between May and July. July and August often experience drought, and winters are characterized by frequent fog and relatively insufficient sunlight.

[0066] Eight integrated monitoring terminals were deployed in the forest area, distributed in a grid pattern, with adjacent terminals spaced approximately 120 meters apart. The monitoring terminals began operation on March 1, 2024, and have been collecting data continuously for four months, covering two growth and development periods: the spring shoot growth period (March to April) and the budding and flowering period (May to June).

[0067] II. Implementation Examples of Environmental Baseline Models Take an environmental anomaly event from June 15th to 18th, 2024 as an example. From 00:00 on June 15th to 24:00 on June 17th, the area experienced continuous rainfall, with a cumulative rainfall of 158 mm within 72 hours. At 8:30 AM on June 18th, the environmental baseline model detected an anomaly while processing real-time data uploaded from monitoring point 3.

[0068] The specific process is as follows: The model receives a 24-hour environmental factor data sequence (48 collection periods) from 00:00 on June 17th to 00:00 on June 18th as input. The reconstruction error of this sequence is calculated to be 0.247, while the threshold for this monitoring point is set at 0.118 (based on statistical analysis of normal data from the previous two months). The reconstruction error exceeds the threshold by 109%, triggering an environmental stress warning. Further analysis of the contribution of the reconstruction error to each environmental factor dimension reveals that the reconstruction error in the soil moisture dimension accounts for 72% of the total error, the rainfall dimension accounts for 18%, and other dimensions account for a combined 10%. Based on this, the system determines the stress type to be waterlogging stress, the warning level to be moderate, and outputs the warning information: "In monitoring point 3 area, continuous rainfall has led to excessively high soil moisture, posing a risk of waterlogging, and the impact is expected to last for 3 to 5 days."

[0069] III. Implementation Examples of Abnormal Growth and Attribution Analysis Following the aforementioned environmental stress events, on June 25th, when analyzing the growth data of Camellia oleifera at monitoring point 3, the system detected anomalies in the growth characteristics. Specifically, compared to monitoring point 5 within the same forest area, which was not affected by waterlogging, the leaf area index of Camellia oleifera at monitoring point 3 decreased from 3.2 seven days prior to 2.8, a drop of 12.5%; the number of fruits remained unchanged, but the fruit enlargement rate decreased from 0.18 mm / day to 0.11 mm / day, a drop of 38.9%. Both exceeded the historical normal range by 1.5 times the standard deviation.

[0070] The system initiated attribution analysis. Environmental factor data sequences from June 18th to June 24th were extracted, with three 96-period windows ending on June 20th, June 22nd, and June 24th respectively serving as baseline inputs. The characteristic perturbation method was used to calculate the contribution of each environmental factor to the decrease in fruit enlargement rate. The results showed that soil moisture contributed an average of 58%, soil temperature contributed 21%, light intensity contributed 12%, and other environmental factors contributed a total of 9%. The attribution analysis output was as follows: the dominant environmental factor was excessively high soil moisture, contributing 58% with a confidence interval of [52%, 64%]; the secondary environmental factor was low soil temperature, contributing 21%; and the tertiary environmental factor was insufficient light intensity, contributing 12%.

[0071] IV. Examples of Reinforcement Learning Decision-Making After the attribution analysis is completed, the system automatically triggers the reinforcement learning decision-making module to generate the optimal management action instruction. The current status is: the growth and development period is in the fruit enlargement stage, the environmental status is soil volumetric moisture content of 38%, the growth status is fruit enlargement rate of 0.11 mm / day, and the time interval since the last management action is 5 days.

[0072] After receiving the state vector, the deep Q-network outputs Q-value estimates for each action. The action with the highest Q-value is: drainage treatment, which involves reducing soil moisture content by opening ditches for drainage, suspending irrigation, applying 80 grams of potassium fertilizer per plant to promote fruit enlargement and recovery, and temporarily refraining from pruning. This action has a Q-value of 12.6, significantly higher than other actions.

[0073] The system outputs the management action command to the maintenance personnel's mobile application, and simultaneously displays the attribution analysis results and the basis for the recommended actions through a visual dashboard. After reviewing the actual site conditions, the maintenance personnel confirmed the system's recommended actions and supplemented them with soil loosening and aeration. The system stores this interaction record as a correction sample in the correction experience pool for subsequent deep Q-network training.

[0074] V. Execution Results and Model Updates On June 26, maintenance personnel carried out drainage, soil loosening, and potassium supplementation. For the following week, the system continuously monitored environmental and growth data in the area. On June 27, soil moisture content decreased to 32%; on June 29, it decreased to 27%, returning to the normal range. On July 2, the fruit enlargement rate rebounded to 0.15 mm / day; on July 5, it rebounded to 0.17 mm / day, approaching the target value.

[0075] On July 1st, the system reached its pre-set monthly retraining time. The system added approximately 11,520 newly accumulated valid data samples from June to the training set, retraining the environmental baseline model and the growth response model. Performance evaluation results on the validation set showed that the anomaly detection accuracy of the environmental baseline model improved from 90.2% to 92.5%, and the root mean square error of prediction for the growth response model decreased from 0.023 to 0.019. The system automatically deployed the new model to replace the old one.

[0076] VI. Overall Effectiveness Evaluation Since the system's implementation, the above methods have achieved the following results: the average lead time for environmental stress warnings is 24 hours; the consistency rate between attribution analysis results and manual field diagnosis reached 86.7%; and the adoption rate of management actions recommended by the deep Q network was 71.2%. Compared with adjacent camellia oleifera forests without this system, the forests with this system showed an 18% increase in the average transverse diameter growth rate of fruits in the first half of 2024, an improvement in the evenness of fruit quantity (measured by the coefficient of variation) from 0.32 to 0.24, and an expected increase in camellia oil yield per unit area of ​​12% to 15%. These data demonstrate that the method of this invention can effectively achieve intelligent monitoring and precise management of camellia oleifera forest growth and its growing environment.

[0077] Although the invention has been specifically shown and described in conjunction with preferred embodiments, those skilled in the art should understand that various changes in form and detail may be made to the invention without departing from the spirit and scope of the invention as defined in the appended claims, all of which shall be within the scope of protection of the invention.

Claims

1. A method for monitoring the growth and growth environment of Camellia oleifera forests, characterized in that, Includes the following steps: Monitoring terminals were deployed in the camellia oleifera forest to continuously and synchronously collect data on growth environment factors and growth characteristics of camellia oleifera plants at predetermined intervals. The growth environment factors and growth characteristics of camellia oleifera plants collected at the same time were aligned on the time axis to construct a time-coupled environment-growth data series indexed by time. Using the historical environmental factor data sequence in the environment-growth coupled time series dataset, an environmental baseline model is constructed by training a predetermined unsupervised learning algorithm. The environmental baseline model is used to characterize the normal environmental dynamic range of the camellia oleifera forest at different time series, and to predict the normal fluctuation range of environmental factors at the current moment based on the real-time input environmental factor data sequence. Using the environment-growth coupled time series dataset, a growth response model is established through a predetermined causal inference algorithm; the growth response model is used to characterize the nonlinear mapping relationship and hysteresis effect between the environmental factor data sequence and the rate of change of growth characteristic data. The real-time collected growth environment factor data sequence is input into the environmental baseline model. If the real-time growth environment factor data sequence deviates from the normal environmental dynamic range, environmental stress early warning information is output. When abnormal growth characteristic data is detected, the current and historical environmental factor data sequences are input into the growth response model to identify the key environmental factors that cause abnormal growth and their contribution, and to form attribution analysis results. Based on the environmental stress early warning information and / or attribution analysis results, the optimal management action instructions are generated using a reinforcement learning algorithm. The reinforcement learning algorithm takes the current environmental state and growth status as input, uses the preset growth target trajectory as the reward function, and outputs differentiated agricultural management actions. The optimal management action instruction is executed, and the environmental and growth change data after execution are fed back to the environmental baseline model and the growth response model for iterative updates.

2. The method for monitoring the growth and growth environment of Camellia oleifera forests according to claim 1, characterized in that, The unsupervised learning algorithm employs a variational autoencoder or a long short-term memory network-autoencoder. The environmental baseline model constructs a normal manifold in a high-dimensional environmental feature space by learning the temporal distribution characteristics of historical growth environmental factor data sequences. When the reconstruction error of the real-time growth environmental factor data sequence in the high-dimensional environmental feature space exceeds a set threshold, it is determined to deviate from the normal environmental dynamic range.

3. The method for monitoring the growth and growth environment of Camellia oleifera forests according to claim 1, characterized in that, The growth response model employs a long short-term memory network or a temporal convolutional network, using the growth environment factor data sequence within a sliding time window as input features and the change rate of growth feature data within a preset lag period after the window as output labels for training, thereby learning the delayed effect of environmental changes on growth.

4. The method for monitoring the growth and growth environment of Camellia oleifera forests according to claim 1, characterized in that, The identification of key environmental factors leading to abnormal growth and their contribution includes: keeping the parameters of the growth response model fixed, perturbing each dimension of the input growth environmental factor data sequence, calculating the difference in the growth change rate of the model output before and after the perturbation of each dimension, and using the difference value as the contribution of the growth environmental factor.

5. The method for monitoring the growth and growth environment of Camellia oleifera forests according to claim 1, characterized in that, The reinforcement learning algorithm employs a deep Q-network; the state space includes real-time values ​​of current growth environment factors, real-time values ​​of growth characteristics, and the growth and development stage of the camellia oleifera; the action space includes the application dosage, application timing, and pruning intensity level of different types of water and fertilizer; the reward function is configured as a monotonically decreasing function of the negative deviation between the actual growth trajectory and the preset optimal growth trajectory after the current action is executed.

6. The method for monitoring the growth and growth environment of Camellia oleifera forests according to claim 1, characterized in that, The monitoring terminals are deployed in a fixed or semi-fixed manner. Each monitoring terminal uploads the collected growth environment factor data and growth characteristic data to the edge computing node or cloud server in real time through low-power wireless self-organizing network communication. The environmental baseline model and growth response model are trained, updated, and reinforcement learning decisions are made on the edge computing node or cloud server.

7. The method for monitoring the growth and growth environment of Camellia oleifera forests according to claim 1, characterized in that, The growth environment factors include air temperature and humidity, light intensity, wind speed, soil temperature and humidity, soil nitrogen, phosphorus and potassium content, soil pH value and / or precipitation; the growth characteristics data of the camellia oleifera plant include plant height, crown width, leaf area index, number of fruits and / or fruit enlargement rate.

8. The method for monitoring the growth and growth environment of Camellia oleifera forests according to claim 1, characterized in that, The environmental stress early warning information includes the type of environmental stress and the expected stress intensity; the attribution analysis results include the dominant environmental factors that cause abnormal growth, the ranking of the contribution of each environmental factor, and the confidence interval.

9. The method for monitoring the growth and growth environment of Camellia oleifera forests according to claim 1, characterized in that, The environmental baseline model and growth response model are automatically retrained according to a preset period or after the accumulated new data reaches a set threshold.

10. The method for monitoring the growth and growth environment of Camellia oleifera forests according to claim 1, characterized in that, Also includes: The environmental stress warning information, attribution analysis results, and generated optimal management action instructions are displayed through a visual interface, and confirmation or correction instructions from the manager are received. The corrected management actions are used as positive or negative reward samples for reinforcement learning to optimize the decision boundary of the reinforcement learning policy network.

Images