Transplanted tree maintenance monitoring method and system based on big data

Abstract

The invention discloses a transplanted tree maintenance monitoring method and system based on big data, and the method comprises the steps: collecting image data, environment data and historical monitoring data of a transplanted tree, carrying out the segmentation and feature extraction of an image, and inputting an environment parameter into a health evaluation model, thereby obtaining a tree health state evaluation result; and constructing a health state sequence based on a health state evaluation result and historical monitoring data, performing risk prediction on a health change trend in a period of time in the future by using a time sequence prediction model, and dynamically optimizing the health evaluation model by taking a risk prediction result as a feedback signal. According to the method and the system, on the basis of multi-source data fusion, continuous monitoring and prospective risk early warning of the health state of the transplanted tree can be realized, and the adaptability and the evaluation precision of the model in a complex environment and a long-period maintenance scene can be improved through feedback learning.

Description

Technical Field

[0001] This invention relates to the field of forestry informatization and intelligent maintenance technology, specifically to a method and system for monitoring the maintenance of transplanted trees based on big data. Background Technology

[0002] Transplanted trees are widely used in forestry planting, ecological restoration, landscaping, and urban renewal. Because the transplanting process significantly impacts the tree's root structure, water balance, and physiological metabolism, transplanted trees are more susceptible to problems such as weakened growth, pest and disease infestation, and decreased adaptability to environmental stresses for a certain period.

[0003] Existing methods for monitoring the maintenance of transplanted trees largely rely on manual inspections or experience-based judgments based on data from a single sensor. For example, this involves visually observing leaf color and branch growth, or using simple threshold judgments based on environmental parameters such as soil moisture and temperature. These methods suffer from low monitoring frequency, high subjectivity, and difficulty in timely detection of hidden diseases or early risks, making them unsuitable for large-scale, long-term, and meticulous maintenance management.

[0004] With the development of sensor technology, image acquisition equipment, and data storage and computing capabilities, some existing technologies have begun to incorporate image processing and data analysis methods to monitor plant health. For example, image segmentation or feature extraction techniques are used to identify leaf lesions and insect damage, or environmental data is combined to assess plant growth. However, most existing solutions only focus on health identification results at a single point in time, lacking modeling of the evolution of tree health status and making it difficult to effectively predict potential risks in the future.

[0005] Some studies employ time-series analysis or prediction methods to assess plant growth trends, but these often focus on structured monitoring data, failing to fully depict the coupling relationship between image information and environmental factors, thus resulting in biases in reflecting actual health changes. Furthermore, existing schemes often isolate different analytical stages, lacking a linkage correction mechanism based on subsequent observations. When analytical conclusions differ from actual conditions, timely iterative adjustments are difficult, affecting the overall stability and reliability of monitoring. On the other hand, abnormal states in transplanted trees exhibit significant low-frequency and diverse characteristics, and insufficient training samples limit the ability to identify and adapt to rare situations. While attempts have been made to expand the sample through generative or augmentative methods, these often lack domain knowledge constraints, resulting in insufficient rationality and usability of the expanded data. Moreover, many related schemes rely on offline centralized updates, making rapid iteration during continuous operation difficult and hindering responses to the dynamic impacts of environmental changes and adjustments to maintenance measures.

[0006] In conclusion, how to continuously monitor the health status of trees after transplantation and identify potential risks in a timely manner is a technical problem that urgently needs to be solved in the field of forestry maintenance. Summary of the Invention

[0007] To address the shortcomings of existing transplanted tree maintenance and monitoring technologies in areas such as multi-source data collaborative analysis, health status evolution characterization, prediction result feedback utilization, and adaptability to anomalous situations with few samples, this invention proposes a big data-based transplanted tree maintenance and monitoring method and system. By comprehensively processing image data, environmental data, and historical monitoring data, it achieves continuous assessment of the health status of transplanted trees, prediction of future risks, and dynamic optimization of models based on feedback learning, thereby improving the efficiency of information utilization and the reliability of decision-making in the transplanted tree maintenance and monitoring process.

[0008] To achieve the above objectives, the present invention provides a method for monitoring and maintaining transplanted trees based on big data, the method comprising: Acquire current image data, environmental data, and historical monitoring data of transplanted trees. The historical monitoring data includes at least the historical health tags of transplanted trees, historical sensor data, and historical maintenance records. The image data is segmented to generate structured image features; Based on structured image features and environmental data, a health assessment model is used to assess the health status of trees and generate health status assessment results. Based on the health status assessment results and historical monitoring data, a time-series prediction model is used to predict the future health status of trees and generate risk prediction results. The risk prediction results are used as feedback signals to dynamically optimize the health assessment model through a feedback learning mechanism.

[0009] As a further preferred embodiment of the above scheme, the image segmentation process includes: A multi-scale convolutional neural network was used to segment the target region of transplanted trees in image data, separating the target regions of tree leaves and branches. Feature extraction is performed on the target region, and the structural features of the trees are obtained through deep feature fusion. Morphological operations are used to optimize the segmentation results, eliminate small noise points, improve the boundaries of segmented regions, and generate structured image features.

[0010] As a further preferred embodiment of the above scheme, the health status assessment of trees using a health assessment model includes: Structured image features are aligned and fused with environmental data to obtain fused features; The fused features are input into the health assessment model, which outputs the tree's health status category and health status score as the health status assessment result. The health assessment model encodes image features and environmental features separately and performs feature interaction so that the health status assessment results simultaneously reflect the characteristics of tree morphology and the impact of changes in environmental conditions on tree growth.

[0011] As a further preferred embodiment of the above scheme, the risk prediction of the future health status of trees using a time-series prediction model includes: The health status assessment results are constructed into a health status sequence in chronological order and time-aligned with the historical sensor data and historical maintenance records in the historical monitoring data to obtain the time-series input information. Input the time-series input information into the time-series prediction model to predict the changing trend of tree health status and the corresponding risk value within a preset window, which is used as the risk prediction result; The preset window is determined at least based on the sampling period of historical monitoring data, and the risk value is the risk score output by the time series prediction model, which is used to characterize the probability or degree of the tree's health status evolving from healthy to sub-healthy or unhealthy.

[0012] As a further preferred embodiment of the above scheme, the dynamic optimization of the health assessment model through a feedback learning mechanism includes: After obtaining the risk prediction results, the subsequent monitoring results within the time period corresponding to the risk prediction results are obtained, and a supervision signal for feedback learning is generated based on the subsequent monitoring results. Based on the difference between the risk prediction results and the monitoring signals, a feedback loss is constructed and the parameters of the health assessment model are updated to reduce the health status assessment error and further reduce the risk prediction error caused by the propagation of the health status assessment error. The feedback learning mechanism updates the health assessment model by at least the following: when constructing the feedback loss, assigning higher weights to samples that lead to higher risk prediction errors.

[0013] As a further preferred embodiment of the above scheme, the method further includes: A generative adversarial network is constructed, and expert knowledge related to transplanted trees is used as a conditional input to constrain the generation process. The expert knowledge includes at least one of the following: tree species category, tree age category, pest and disease type, climate and seasonal factors. Based on the conditional input, synthetic data matching the anomaly category is generated. The synthetic data includes at least synthetic image data or synthetic time-series data that is isomorphic to the sensor-acquired data. Consistency constraints are applied to the synthesized data to ensure that the synthesized data meets preset ecological rationality conditions, which include at least physiological indicator range constraints or environmental variable coupling constraints.

[0014] As a further preferred embodiment of the above scheme, the feedback learning mechanism further includes: Based on feedback from manual inspections and maintenance records, the feedback results are compared with the predicted events corresponding to the risk prediction results to identify abnormal states that the model missed or predicted incorrectly. When the number of samples corresponding to an abnormal state in the training dataset is less than a threshold, it is judged as a few-sample abnormal category, triggering the generative adversarial network to generate synthetic data corresponding to the few-sample abnormal category. Synthetic data is combined with existing training data, and incremental training is performed on the health assessment model and time series prediction model to optimize model parameters.

[0015] As a further preferred embodiment of the above scheme, the incremental training adopts a rolling update method, which includes at least: When new monitoring data or maintenance feedback data is received based on the abnormal state identification results, the data is written into the historical monitoring data and an incremental sample batch is formed. Based on the incremental sample batch, local parameters of the health assessment model or time series prediction model are updated, and the model version is recorded. Once the model version is updated, the updated model will be used for subsequent health status assessments and risk predictions.

[0016] As a further preferred embodiment of the above scheme, the method further includes: Input the health status assessment results and risk prediction results into the decision tree model or rule tree model; The decision tree model or rule tree model outputs a set of maintenance strategies corresponding to the tree health status category and risk value range. The set of maintenance strategies includes at least one of irrigation, fertilization, pruning and pest and disease treatment. The set of maintenance strategies is matched with historical maintenance records to determine the target maintenance strategy and generate executable maintenance task parameters.

[0017] This invention also provides a big data-based monitoring system for the maintenance of transplanted trees, the system comprising: The data acquisition unit is used to acquire current image data, environmental data and historical monitoring data of transplanted trees. The historical monitoring data includes at least the historical health tags of transplanted trees, historical sensor data, and historical maintenance records. The image processing unit is used to perform image segmentation processing on the image data and generate structured image features; The health assessment unit is used to assess the health status of trees based on the structured image features and the environmental data using a health assessment model, and generate a health status assessment result. The risk prediction unit is used to predict the future health status of trees based on the health status assessment results and the historical monitoring data, and generate risk prediction results by using a time-series prediction model. The feedback optimization unit is used to use the risk prediction result as a feedback signal to dynamically optimize the health assessment model through a feedback learning mechanism.

[0018] Compared with the prior art, the present invention has at least the following technical effects: By combining image data, environmental data, and historical monitoring data, the assessment process of the health status of transplanted trees can comprehensively reflect the tree morphological information and changes in the growth environment. By performing time-series modeling on the health status assessment results, it is possible to predict the future health trends of transplanted trees, which is beneficial for identifying potential risks in advance. By introducing a feedback learning mechanism based on prediction results, the model can be dynamically optimized according to subsequent monitoring results, reducing the cumulative impact of evaluation errors in subsequent prediction processes. In cases of few-sample anomalies, the training samples are expanded by combining generated data constrained by expert knowledge, thereby improving the model's adaptability to low-frequency anomaly states. By using a rolling incremental update method, the model can be iteratively adjusted during continuous operation to adapt to changes in environmental conditions and maintenance strategies. Attached Figure Description

[0019] Figure 1 This is a flowchart of the big data-based transplanted tree maintenance and monitoring method in an embodiment of the present invention; Figure 2 This is an architecture diagram of a big data-based transplanted tree maintenance monitoring system according to an embodiment of the present invention. Detailed Implementation

[0020] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, but the present invention is not limited to the following embodiments. Equivalent substitutions or modifications made by those skilled in the art without departing from the spirit and substance of the present invention shall fall within the protection scope of the present invention.

[0021] As explained in the background section, transplanted trees are more prone to problems such as weakened growth, pest and disease infestation, or accumulation of latent stresses over a prolonged period after transplantation, especially when faced with a combination of factors including root damage, sudden environmental changes, and decreased adaptability. Current monitoring methods largely focus on observable anomalies and the current state, lacking a systematic portrayal of the health evolution process and risk accumulation. Given the continuously expanding scale of forestry planting and limited maintenance resources, this is gradually becoming a key bottleneck restricting transplant survival rates and long-term stability.

[0022] While existing research has incorporated image analysis, environmental sensing, and historical records to monitor plant status, it often relies on a single data source or processes different types of data in isolation. This makes it difficult to organically integrate tree morphological characteristics with environmental changes within a unified framework. Consequently, health assessments often remain static results at a single moment, failing to support maintenance decisions based on evolutionary trends. Furthermore, time-series-based growth trend modeling heavily emphasizes sensor data, with insufficient integration of image information and maintenance records. Moreover, the lack of feedback mechanisms between analysis stages to correct earlier assessments based on prediction results makes it difficult to automatically converge prediction biases in a timely manner, potentially leading to the accumulation and amplification of assessment errors in subsequent predictions.

[0023] In view of the above problems, it is necessary to construct a transplanted tree maintenance monitoring technology solution that can uniformly utilize image data, environmental data, and historical monitoring data, and form a closed-loop linkage between current status assessment, future risk judgment, and model adaptive updates. Based on this, the present invention provides a big data-based transplanted tree maintenance monitoring method, such as... Figure 1 As shown, the method specifically includes: Acquire current image data, environmental data, and historical monitoring data of transplanted trees. The historical monitoring data includes at least the historical health tags of transplanted trees, historical sensor data, and historical maintenance records. The image data is segmented to generate structured image features; Based on structured image features and environmental data, a health assessment model is used to assess the health status of trees and generate health status assessment results. Based on the health status assessment results and historical monitoring data, a time-series prediction model is used to predict the future health status of trees and generate risk prediction results. The risk prediction results are used as feedback signals to dynamically optimize the health assessment model through a feedback learning mechanism.

[0024] Based on the aforementioned transplanted tree maintenance and monitoring method, under a unified data processing framework, structured feature extraction is performed on images of transplanted trees, which are then fused with environmental parameters and historical monitoring information to form a graded assessment of their current health status. On this basis, time-series modeling is used to predict future health trends and corresponding risks, and the prediction results are combined with subsequent monitoring data, manual inspections, and maintenance records. A feedback learning and incremental update mechanism is introduced to drive continuous optimization of the front-end model. For anomalies with a small number of samples, a synthetic data generation and training strategy incorporating domain knowledge constraints is further introduced to enhance the overall monitoring method's adaptability to long-tailed anomalies.

[0025] Before implementing the aforementioned big data-based monitoring method for transplanted tree maintenance, it is typically necessary to complete a series of tasks related to the deployment of the monitoring infrastructure and data preparation to ensure the stable operation of subsequent steps and the formation of a complete data loop. These tasks may include, but are not limited to, the following:

[0026] (1) Determination and identification of monitoring objects and monitoring areas In forestry planting, ecological restoration, or landscaping projects, the transplanted trees planned to be incorporated into the methods of this invention are identified, and each transplanted tree is assigned a unique identifier, such as a number, QR code, or electronic tag. Basic attributes such as tree species, age, planting location, transplanting time, diameter at breast height (DBH), and crown width are recorded. For concentrated, contiguous planting areas, monitoring plots can be divided according to the land parcels, establishing a correlation between trees and plots, facilitating batch management and data aggregation analysis.

[0027] (2) Hardware deployment for monitoring In a typical implementation, image acquisition and environmental monitoring equipment are configured around the target transplanted trees. The image acquisition equipment can be a fixed surveillance camera, a retractable shooting terminal, or a mobile acquisition device deployed along the work path, used to periodically acquire image data of the tree as a whole and in specific areas. The environmental monitoring equipment may include soil moisture sensors, soil temperature sensors, air temperature and humidity sensors, light intensity sensors, etc., used to collect environmental data related to tree growth. These devices can be connected to a unified monitoring network through a centralized acquisition node or acquisition gateway, and support time synchronization and data tagging for subsequent association by tree identifier and timestamp.

[0028] (3) Multimodal data acquisition and basic processing After the aforementioned hardware deployment is completed, the multimodal data acquisition process is initiated. Image data is acquired according to a preset acquisition cycle or triggering conditions, and basic preprocessing is performed on the edge device or server side, such as format conversion, resolution unification, and writing timestamps and tree identifiers. Environmental data and historical sensor acquisition data are written to the monitoring database according to a preset sampling cycle and bound to the corresponding tree identifiers and spatial location information. Historical monitoring data and historical maintenance records can be imported from existing information systems or supplemented by the maintenance unit through manual entry to form a multi-source historical dataset containing historical health tags, maintenance operation records, and key environmental variables.

[0029] (4) Establishment of network communication and data transmission channels In typical application scenarios, on-site monitoring equipment connects to a data acquisition gateway via wired or wireless communication. The gateway then transmits the acquired data to a central server or cloud platform via a local area network, 4G / 5G cellular network, or other data links. Data compression and encryption mechanisms can be employed during transmission to ensure efficient bandwidth utilization and data security. The central side deploys data storage and computing modules to receive and persist image data, environmental data, and historical monitoring data, providing a unified data source for subsequent segmentation, feature extraction, sequence construction, and model training.

[0030] (5) Interaction and integration with transplanting and maintenance processes During transplantation and subsequent maintenance, maintenance personnel can use mobile or desktop terminals to record actual work performed, such as watering, fertilizing, pruning, support reinforcement, and pest and disease control, and associate these records with the corresponding tree identifiers to create standardized historical maintenance records. Optionally, after the trees are transplanted, one or more baseline monitoring sessions are conducted to collect initial image and environmental data and generate initial health tags, which serve as reference benchmarks for subsequent health status assessments and risk predictions. Integration with on-site work processes allows the monitoring data in this invention to form a closed loop with actual maintenance activities, providing a reliable data source for subsequent time-series analysis and feedback learning based on historical maintenance records.

[0031] (6) Configuration of Big Data Management and Model Running Environment In another embodiment, a big data management and computing environment can be configured on the central platform to uniformly store, index, and control access to data from different time periods, regions, and individual trees. The model runtime environment is pre-configured according to business needs, including computing resources for image segmentation, time series analysis, and generative model training and incremental updates. Through unified planning of data storage structures, computing resources, and task scheduling strategies, batch historical data mining and online monitoring inference can be supported in subsequent implementations of this invention.

[0032] By completing the above preparatory work, transplanted trees at different growth stages and under different site conditions can have the basic conditions to continuously acquire image data, environmental data, and historical monitoring data. This provides complete data support and an operating environment for subsequent implementation of health status assessment based on structured image features, risk prediction based on time series information, and model optimization based on feedback learning and synthetic data.

[0033] After completing the deployment of the aforementioned monitoring infrastructure and the acquisition of multimodal data, in a preferred embodiment of the present invention, the image segmentation processing performed on the current image data of the transplanted trees is specifically implemented in the following steps.

[0034] In one embodiment, for transplanted tree image data acquired through an image acquisition device, the images are first preprocessed, including: uniformly scaling the image size to standardize images with inconsistent original resolutions to a preset size; normalizing brightness and contrast according to the imaging characteristics of the acquisition device; and optionally performing noise suppression to reduce the interference of background noise on subsequent segmentation results. The preprocessed image is used as input to a multi-scale convolutional neural network.

[0035] The image data may include at least one color image with a resolution not lower than a preset threshold, preferably an RGB image in the visible light band; it may also further include images in other bands such as near-infrared, to help distinguish the health or moisture status of leaves. To ensure consistency in subsequent processing, each frame of the image is appended with metadata such as a unique tree identifier, acquisition time, acquisition device number, and shooting angle or location when the image is stored, so as to associate it with the corresponding environmental data and historical monitoring data.

[0036] Subsequently, a multi-scale convolutional neural network is employed to segment the transplanted tree target region in the image data. Specifically, a segmentation network can be constructed based on an encoder-decoder structure. In the encoding stage, multi-level features from local details to overall structure are extracted through convolutional and pooling layers of different scales. In the decoding stage, spatial resolution is gradually restored, and feature information from different scales is fused through skip connections or feature concatenation. The network output includes pixel-level category predictions for tree leaf regions, branch regions, and non-tree background regions, thereby separating the target regions of tree leaves and branches from the background and forming preliminary target segmentation results.

[0037] After segmenting the target region, feature extraction is performed on it. In one implementation, feature descriptions can be constructed separately for the leaf and branch regions. Deep convolutional feature maps are pooled, reduced in dimensionality, or subjected to region statistical processing to obtain feature vectors representing leaf texture, color distribution, shape contours, and branch distribution structure. Simultaneously, multi-scale feature maps are fused using channel concatenation, weighted fusion, or feature transformation to generate a structured feature representation for subsequent health status assessment. This structured feature can correspond to multiple dimensions of information for each tree, such as leaf density, branch integrity, and crown morphology, and is stored or passed to subsequent models in a fixed-dimensional format.

[0038] To further improve the accuracy of the target region segmentation results, this embodiment also applies morphological operations to optimize the preliminary segmentation results. Specifically, opening, closing, dilation, and erosion operations can be performed on the binary mask of the leaf and branch target regions to eliminate isolated small noise points, fill gaps at the leaf edges, and smooth the target contours. For local missing regions caused by lighting, occlusion, or shooting angle, connected component analysis and region filling operations can be used to merge or complete scattered regions that clearly belong to the overall tree structure, thereby obtaining segmentation results with more complete boundaries and a more reasonable topological structure.

[0039] After morphological optimization, the optimized target region mask is combined with the aforementioned depth feature fusion results to generate structured image features corresponding to each transplanted tree. These features are then associated with the tree's identification information and timestamp, serving as one of the inputs to the health assessment model. Through this implementation method, image segmentation not only separates the target regions of leaves and branches but also extracts and fuses multi-scale structural features based on segmentation, providing fine-grained and structured image features for subsequent health status assessment and temporal risk prediction.

[0040] After completing the aforementioned image segmentation process and obtaining structured image features to characterize tree structural information, in a preferred embodiment of the present invention, a health assessment model is used to assess the health status of the trees. The specific implementation process includes: Structured image features are aligned and fused with environmental data to obtain fused features; The fused features are input into the health assessment model, which outputs the tree's health status category and health status score as the health status assessment result. The health assessment model encodes image features and environmental features separately and performs feature interaction so that the health status assessment results simultaneously reflect the characteristics of tree morphology and the impact of changes in environmental conditions on tree growth.

[0041] First, for each transplanted tree, its current-time structured image features and environmental data are compiled under a unified data identifier. The environmental data may include one or more of the following: soil moisture, soil temperature, air temperature, air humidity, light intensity, wind speed, and precipitation. Optionally, environmental derivative indicators calculated from the original measurements may be included, such as average values, extreme values, or rates of change within a certain time window. To align with the image features temporally, in practice, the closest sampling point in time can be selected from the environmental data based on the image acquisition timestamp, or the sampled data can be interpolated or statistically summarized within a preset time range to obtain an environmental feature vector corresponding one-to-one with the image sample.

[0042] The health assessment model can adopt a multi-branch neural network structure that includes image feature encoding branch, environmental feature encoding branch, and health determination branch.

[0043] The image feature encoding branch takes a structured image feature vector as input and encodes the image features independently. In one implementation, the image feature encoding branch can adopt a structure with a convolutional neural network (CNN) as the backbone. For example, a residual network structure based on several convolutional layers and residual blocks can be used to perform layer-by-layer feature transformation and abstraction on the structured image features; after feature extraction, the high-dimensional features are mapped to a latent image representation vector of a preset dimension through fully connected layers and nonlinear activation functions. In another implementation, the image feature encoding branch can adopt a network structure containing a self-attention mechanism, such as a visual Transformer structure, to perform block encoding and multi-head self-attention calculation on the input structured image features to obtain richer global relevant information, and then obtain the latent image representation through a projection layer. Furthermore, a hybrid structure combining convolution and attention can be constructed based on the above to simultaneously utilize local receptive fields and global dependency modeling capabilities.

[0044] The environmental feature encoding branch takes environmental feature vectors as input and is used to independently encode environmental data. In one embodiment, the environmental feature encoding branch can adopt a multilayer perceptron (MLP) structure, including two or more fully connected layers and nonlinear activation units, to perform nonlinear mapping on features such as soil moisture, temperature, air humidity, and light intensity to obtain an environmental latent representation vector. In another embodiment, if the environmental features have significant temporal correlation, a recurrent neural network structure, such as a long short-term memory network (LSTM) or a gated recurrent unit (GRU), can be introduced into the environmental feature encoding branch to encode the environmental change sequence within a short time window, thereby forming an environmental latent representation that takes into account both the current state and short-term change trends.

[0045] After completing image feature encoding and environmental feature encoding, a feature interaction and fusion structure is set up to realize feature interaction between image features and environmental features. In a preferred embodiment, the feature interaction structure can take the form of feature concatenation plus a multilayer perceptron, concatenating the image latent representation vector and the environmental latent representation vector along the feature dimension to obtain a fused input vector; the fused input vector is then input into an interaction network containing one or more fully connected layers and nonlinear activation functions to jointly transform the two types of features, thereby characterizing the combination relationship between tree morphological features and environmental conditions. In another embodiment, the feature interaction structure can also employ an attention mechanism or gating mechanism, such as constructing a cross-modal attention module, using the image latent representation as the query and the environmental latent representation as the key and value, or vice versa, using the environmental latent representation as the query, to weight the importance of features from different modalities, generating a fused feature representation containing intermodal correlation information.

[0046] In this embodiment, the aforementioned fused features are used as input to the health assessment model. The model processes these features and generates at least two types of results at the output. One type is a discrete health status category label, such as a preset one of healthy, sub-healthy, and unhealthy; the other type is a continuous health status score, used to quantify the current health status. Therefore, the health determination branch can include a classification branch and a scoring branch set in parallel. The classification branch uses one or more fully connected layers and a Softmax output layer to output the category probability of each health level; the scoring branch uses one or more fully connected layers and a linear or bounded activation function to output the health status score. The health status category and the health status score together constitute the health status assessment result.

[0047] During the model training phase, a supervisory signal can be constructed based on health level labels and / or manually labeled health scores recorded in historical monitoring data to supervise the training of the health assessment model. The classification branch uses cross-entropy loss, and the scoring branch uses mean squared error loss; these two losses are weighted and summed to form the total loss function. Gradient descent-based optimization algorithms are used to jointly train the parameters of the image feature encoding branch, environmental feature encoding branch, feature interaction structure, and health determination branch. During the online execution phase, the fused features (after feature alignment and fusion) are input into the health assessment model to obtain the corresponding health status category and health status score. This score is then associated with tree identifiers and timestamps, providing the data for subsequent time-series prediction models to construct health status sequences, serving as the input basis for predicting the future health status risk of trees.

[0048] After completing the assessment of the tree's health status at the current moment in the aforementioned embodiments, in another preferred embodiment of the present invention, a time-series prediction model is used to predict the risk of the tree's future health status. The specific implementation process is as follows: The health status assessment results are constructed into a health status sequence in chronological order and time-aligned with the historical sensor data and historical maintenance records in the historical monitoring data to obtain the time-series input information. Input the time-series input information into the time-series prediction model to predict the changing trend of tree health status and the corresponding risk value within a preset window, which is used as the risk prediction result; The preset window is determined at least based on the sampling period of historical monitoring data, and the risk value is the risk score output by the time series prediction model, which is used to characterize the probability or degree of the tree's health status evolving from healthy to sub-healthy or unhealthy.

[0049] First, for each transplanted tree, a health status sequence is constructed chronologically based on the health status category and score output by the health assessment model at different monitoring times. Specifically, the health status score at each monitoring time can be used as a one-dimensional or multi-dimensional numerical feature. The health status category is then one-hot encoded or embedded encoded and concatenated with the score to form a health status feature vector corresponding to that time. Next, historical sensor data and maintenance records from the same time period are combined to align various data types by timestamp. For example, based on the monitoring timeline, environmental data such as soil moisture, soil temperature, air temperature, air humidity, and light intensity are interpolated or windowed between adjacent sampling points to obtain environmental features aligned with the health status assessment results. Simultaneously, maintenance operations performed on the tree during that time period, such as watering, fertilization, pruning, and pest and disease treatment, are encoded as maintenance features, such as operation type, intensity, and frequency, and mapped to the corresponding time steps. Through this method, health status features, environmental features, and maintenance features are aligned and concatenated along the time dimension to form multi-dimensional temporal input information.

[0050] Subsequently, in this embodiment, a time-series prediction model is used to model the aforementioned time-series input information. In one specific implementation, the time-series prediction model can employ a Long Short-Term Memory (LSTM) network structure. The time-series input information arranged in chronological order is used as the input sequence of the LSTM network. One or more LSTM units are used to jointly model the short-term fluctuations and long-term trends in the sequence, outputting the corresponding hidden state representation at the end of the sequence or at each time step. The hidden state is then input into subsequent prediction layers, and through fully connected layers and nonlinear activation functions, health status change trend characteristics and risk scores are generated within a preset prediction window. The risk score can be designed as the probability or severity value of the tree's health status evolving from healthy to sub-healthy or unhealthy within a preset future time window.

[0051] In another embodiment, the time-series prediction model can employ a gated recurrent unit (GRU) network to reduce the model parameter size and improve inference efficiency; or it can employ a temporal convolutional network (TCN) structure, which uses one-dimensional convolutional kernels to perform causal convolution and multi-scale receptive field modeling on the temporal input information along the time dimension to extract multi-scale temporal dependencies. Furthermore, a time-series prediction network based on a self-attention mechanism can be introduced on top of the above recurrent or convolutional structures, such as a time series Transformer, to explicitly model the correlations between different time steps, thereby enhancing the model's ability to perceive complex fluctuation patterns and abrupt events.

[0052] In a preferred embodiment, the preset window can be set according to the sampling period of historical monitoring data and the characteristics of the tree species. For example, it can be set to the time length corresponding to several sampling periods. During the model training phase, samples are constructed by "inputting time-series input information from the past period, predicting the trend of health status changes and risk scores within the preset window in the future". In terms of the prediction output format, a multi-task output structure can be adopted. On the one hand, it outputs a sequence of health status scores at different time points within the future window to characterize the trend of health changes; on the other hand, it outputs one or more aggregated risk scores to characterize whether there is a significant risk of deterioration in health status within the window and the degree of risk.

[0053] During model training, a realistic health status change trajectory can be constructed based on subsequent health status labels and scores recorded in historical monitoring data, along with corresponding environmental data and maintenance records. This trajectory serves as a supervisory signal for training the time-series prediction model. During training, a sequence-to-sequence loss function can be used to measure the mean squared error between the predicted future health status score sequence and the actual score sequence. Simultaneously, cross-entropy or other classification losses are calculated between the predicted risk score and the risk label derived from the actual health status label. By weighted combining these losses, the parameters of network structures such as LSTM, GRU, TCN, or Transformer are jointly optimized.

[0054] During the online operation phase, once a new health status assessment result is generated, this result, along with the latest sensor data and maintenance records, is appended to the corresponding tree's time-series input information. A recent time-series segment is extracted as model input, and the aforementioned time-series prediction model outputs the future health change trend and risk value within a preset window. The risk value output by the model can be used as part of the risk prediction results to drive subsequent feedback learning mechanisms and maintenance strategy generation.

[0055] After completing the aforementioned risk prediction of the future health status of trees, in a preferred embodiment of the present invention, the health assessment model is dynamically optimized through a feedback learning mechanism, specifically including: After obtaining the risk prediction results, the subsequent monitoring results within the time period corresponding to the risk prediction results are obtained, and a supervision signal for feedback learning is generated based on the subsequent monitoring results. Based on the difference between the risk prediction results and the monitoring signals, a feedback loss is constructed and the parameters of the health assessment model are updated to reduce the health status assessment error and further reduce the risk prediction error caused by the propagation of the health status assessment error. The feedback learning mechanism updates the health assessment model by at least the following: when constructing the feedback loss, assigning higher weights to samples that lead to higher risk prediction errors.

[0056] First, based on the risk prediction results output by the time-series prediction model, the corresponding prediction time period is determined, and subsequent monitoring results within this time period are collected. These subsequent monitoring results may include: the subsequent health status assessment results of the same transplanted tree given by the health assessment model within the prediction time period; annotation information regarding tree growth changes and pest and disease conditions in on-site manual inspection records; and new environmental data and maintenance records collected within this time period. Based on these subsequent monitoring results, a supervisory signal for feedback learning can be constructed. For example, the actual health status category at the end of the time period can be used as the target label, or the change in health status score within the time period can be calculated based on multiple monitoring results and transformed into a target risk indicator reflecting the degree of health deterioration.

[0057] After acquiring the supervision signal, this embodiment constructs a feedback loss function to measure the difference between the prediction result and the supervision signal. Specifically, several prediction-result pairs can be selected from the historical records as feedback samples. Each sample consists of the health status assessment result at the input time and the corresponding time-series input information, the risk value output by the time-series prediction model, and the supervision signal generated based on subsequent monitoring results. For each feedback sample, the feedback error of the sample is obtained by calculating the deviation between the risk prediction value and the supervision signal, for example, through mean squared error or cross-entropy loss. Since risk prediction in this invention depends on the output of the front-end health assessment model, the health assessment model and the time-series prediction model can be regarded as a series structure. Through backpropagation, the feedback loss constructed based on the risk prediction error is propagated back along the time-series prediction model to the health assessment model, updating only the parameters of the health assessment model, keeping the parameters of the time-series prediction model unchanged in this round of feedback learning, or updating them only to a limited extent under a preset strategy. In this way, the error exposed at the risk prediction level can act back on the health assessment stage, correcting the mapping relationship between the health assessment model and structured image features and environmental features, thereby reducing the health status assessment error and reducing the cumulative impact of this error in the subsequent risk prediction process.

[0058] In a preferred embodiment, to emphasize learning from problem samples, the present invention assigns higher weights to samples that lead to higher risk prediction errors when constructing the feedback loss. Specifically, in each batch of feedback samples, sample weights can be calculated based on the magnitude of the risk prediction error of each individual sample. For example, the error value can be normalized or transformed using a nonlinear function and used as a weight coefficient, or higher weight factors can be pre-assigned to samples predicted as low risk but actually experiencing significant health deterioration (i.e., underreported samples). When calculating the overall feedback loss, the loss values ​​of each sample are weighted and summed according to their corresponding weights, so that the model update process pays more attention to samples with significant risk prediction errors and prioritizes correcting the judgment bias of the health assessment model on these samples.

[0059] In another implementation, feedback learning can employ either a periodic or triggered strategy. In a periodic strategy, after the system has run for a period of time, the deviation between the risk prediction results and subsequent monitoring results is periodically analyzed to form a batch of feedback samples, and a model update is performed. In a triggered strategy, when a high-error sample count exceeding a threshold occurs within a certain period, a round of feedback learning is automatically initiated to specifically adjust the parameters of the health assessment model. Regardless of the strategy used, the updated health assessment model can be applied to subsequent health status assessments, making the new assessment results closer to the actual observed health changes, thereby improving the overall assessment accuracy and risk prediction stability of the method of this invention.

[0060] After completing the aforementioned dynamic optimization of the health assessment model based on the feedback learning mechanism, in another preferred embodiment of the present invention, for anomaly categories that have a small number of samples in historical monitoring data but are significant in practical application scenarios, a generative adversarial network can be further introduced to generate synthetic data, thereby improving the model's adaptability to anomaly scenarios with few samples. This process further includes: A generative adversarial network is constructed, and expert knowledge related to transplanted trees is used as a conditional input to constrain the generation process. The expert knowledge includes at least one of the following: tree species category, tree age category, pest and disease type, climate and seasonal factors. Based on the conditional input, synthetic data matching the anomaly category is generated. The synthetic data includes at least synthetic image data or synthetic time-series data that is isomorphic to the sensor-acquired data. Consistency constraints are applied to the synthesized data to ensure that the synthesized data meets preset ecological rationality conditions, which include at least physiological indicator range constraints or environmental variable coupling constraints.

[0061] In one embodiment, the abnormality categories of transplanted trees are first identified and defined based on historical monitoring data and subsequent monitoring results. These abnormality categories may include, but are not limited to: severe pest and disease types within a specific tree species or age range, extreme high or low temperatures combined with drought, and root damage caused by prolonged waterlogging. For each type of abnormality, relevant expert knowledge information is extracted, such as tree species, age, common lesion morphology, typical environmental condition combinations, and peak seasons. This expert knowledge is encoded into a condition vector, which serves as the conditional input for the generative adversarial network (GAN).

[0062] In this embodiment, the generative adversarial network (GAN) can adopt a conditional GAN ​​structure, including a generative subnetwork and a discriminative subnetwork. Taking synthetic image data as an example, the generative subnetwork takes a combination of random noise vectors and expert knowledge condition vectors as input, and gradually generates a synthetic image that matches the size of the original tree image through multi-layer transposed convolutions or upsampling structures. The discriminative subnetwork takes a real image or synthetic image and the corresponding condition vector as input, and determines whether the input image is a real sample under given conditions through convolutional layers and fully connected layers. During training, real anomalous samples (anomalies with a small number of samples) and approximate scenes selected from normal samples are used as positive samples, and randomly generated synthetic samples are used as negative samples. The parameters of the generative and discriminative subnetworks are optimized through adversarial training, enabling the generative subnetwork to gradually learn to generate tree images that conform to the characteristics of the target anomaly category under given expert knowledge constraints.

[0063] In another implementation, for time-series data isomorphic to sensor-acquired data, a conditional generative adversarial network (GAN) structure with time series as output can be constructed. The generation subnetwork takes random noise and expert knowledge condition vectors as input, and generates a sequence of environmental variables or physiological indicators over a preset time period through a one-dimensional convolutional network, a recurrent neural network (e.g., LSTM or GRU), or a combination thereof. The discrimination subnetwork takes real or generated time-series segments and corresponding condition vectors as input, and outputs a true / false judgment result through temporal convolution or a recurrent structure. During training, the overall shape of the generated sequence under given conditions is constrained by adversarial loss, and distance constraints between the generated and real anomalous time-series samples can be introduced to improve the similarity between the generated time-series data and the real data.

[0064] To ensure the ecological rationality of the synthesized data, this embodiment introduces consistency constraints on both synthesized image data and synthesized time-series data. Specifically, for synthesized time-series data, during the generation or post-processing stage, it can be checked whether each environmental variable or physiological indicator falls within a preset reasonable range. For example, whether soil moisture, air temperature, and trunk moisture content meet the physiological indicator range constraints obtained based on expert knowledge or historical statistics. Samples that clearly exceed the reasonable range are discarded or a penalty term is added to the training loss. For samples reflecting the relationship between multiple environmental variables, environmental variable coupling constraints can be set based on empirical formulas or statistical correlations, such as the relationship between soil moisture and precipitation, or the relationship between temperature and evapotranspiration intensity. Samples that do not meet the coupling constraints are screened out or their penalty weight in the adversarial loss is increased. For synthesized image data, simple geometric and color priors can be combined, such as the lesion distribution not covering the entire tree crown, the range of color changes in the diseased area, and the continuity of branch structure, to design morphological or texture consistency indicators. Synthetic samples that clearly do not conform to the true morphology of trees are filtered out.

[0065] Through the above methods, under the constraint of expert knowledge, the synthetic image data and synthetic time-series data generated by the generative adversarial network not only closely approximate real abnormal samples in statistical distribution, but also meet the preset ecological rationality conditions in terms of physiological indicator range and environmental variable coupling relationship. The synthetic data, after screening and consistency constraint processing, can be merged with the original training dataset for incremental training of the health assessment model and time-series prediction model. When combined with the abnormal state recognition and triggering mechanism, when the number of samples of a certain type of abnormal state is detected to be below a threshold, the generative adversarial network can be scheduled to generate synthetic data of the corresponding abnormal category to supplement the training samples of that category, thereby enhancing the overall ability of the method of this invention to identify and predict abnormal scenarios with few samples.

[0066] In the foregoing embodiments, the current health status and future risks of transplanted trees can be analyzed using a health assessment model and a time-series prediction model, and the health assessment model can be dynamically optimized through a feedback learning mechanism. Building upon this, in another preferred embodiment of the invention, to focus on learning about anomaly categories with a small sample size but significant practical value, the feedback learning mechanism can also combine manual inspections and maintenance records to perform post-hoc comparisons of the prediction results, triggering a generative adversarial network to generate synthetic data, and incrementally training the model. Specific implementation steps include: Based on feedback from manual inspections and maintenance records, the feedback results are compared with the predicted events corresponding to the risk prediction results to identify abnormal states that the model missed or predicted incorrectly. When the number of samples corresponding to an abnormal state in the training dataset is less than a threshold, it is judged as a few-sample abnormal category, triggering the generative adversarial network to generate synthetic data corresponding to the few-sample abnormal category. Synthetic data is combined with existing training data, and incremental training is performed on the health assessment model and time series prediction model to optimize model parameters.

[0067] In one embodiment, the system continuously records two types of information during normal operation. One type is the risk prediction results output by the time-series prediction model in each monitoring period, including the prediction time, target tree identifier, corresponding prediction window, and risk score within that window. The other type is feedback information obtained from subsequent actual monitoring, including descriptions of tree health status in manual inspection records, such as whether there is extensive yellowing of leaves, leaf drop, branch dieback, or spread of lesions; the type of pests and diseases confirmed by professionals when necessary; and the maintenance operations and monitoring results actually performed within the corresponding time period. When storing data, the system attaches a unique tree identifier and timestamp to each record for association during post-processing.

[0068] In this embodiment, a feedback analysis task can be set up to periodically scan historical prediction records and subsequent feedback records. For each risk prediction result, the task retrieves the monitoring results at the end of the prediction window or within that window, comparing the predicted risk score with the health changes reflected in the actual feedback. For example, if the prediction result shows a low risk score, but subsequent inspection records or health assessment results indicate a significant deterioration in the tree's health, or it is marked as being affected by a specific pest or disease, the sample can be marked as a missed detection sample. If the prediction result gives a high risk value, but subsequent monitoring results show that the tree's health remains stable or without significant abnormalities, the sample can be marked as a false alarm sample. According to preset judgment rules, the missed detection samples and false alarm samples are statistically analyzed and classified to form an abnormal state sample set.

[0069] Subsequently, for each type of abnormal state in the aforementioned abnormal state sample set, the number of samples in the existing training dataset is counted. In a preferred embodiment, abnormal categories divided by tree species, tree age, pest / disease type, and typical environmental combination can be used as statistical units. When the number of valid samples corresponding to a certain abnormal category is lower than a preset threshold, the category is marked as a low-sample abnormal category. This threshold can be set according to factors such as the sample size, target recognition accuracy, and training stability at the initial stage of system deployment, for example, it can be set to no less than a number of complete time-series samples or a number of valid image samples.

[0070] Upon detecting a few-sample anomaly categories, the system invokes a generative adversarial network (GAN) and constructs conditional inputs based on expert knowledge corresponding to that anomaly category. For example, for the anomaly category of a specific type of pest or disease occurring in a specific season for a particular tree species, the tree species category, age range, pest or disease type, seasonal information, and typical environmental elements can be encoded as conditional vectors and input into the GAN to generate synthetic image data matching that anomaly category or synthetic time-series fragments isomorphic to sensor time-series data. After generation, the synthetic data undergoes the aforementioned consistency constraint processing to filter out samples that do not conform to the range of physiological indicators or the coupling relationship of environmental variables, resulting in a set of synthetic anomaly samples for training.

[0071] After obtaining the synthetic anomaly sample set, this embodiment fuses the synthetic data with existing training data to construct sample batches for incremental training. Specifically, for the health assessment model, training samples can be constructed with synthetic image data and its corresponding environmental features as input, and the target health state category and score as output; for the time series prediction model, training samples can be constructed with time series input information including synthetic anomaly time series segments as input, and future health state changes and risk indicators as output. To avoid causing significant perturbation to the converged parameters, incremental training can use a small learning rate and a small batch size, updating the model parameters only within a limited number of rounds; if necessary, higher loss weights can be applied to synthetic samples and real anomaly samples, allowing the model to prioritize learning feature patterns of fewer anomaly categories with limited training resources.

[0072] In another implementation, the system can configure the above-described anomaly detection—few-sample identification—synthetic data generation—incremental training process as a triggered procedure. When the number of missed samples or the cumulative prediction error of a certain anomaly category exceeds a preset threshold within a certain time window, a synthetic data generation and incremental training task based on that category is automatically initiated. After the incremental training is completed, the updated health assessment model and time series prediction model will replace the original model version for subsequent health status assessment and risk prediction.

[0073] In the foregoing embodiments, feedback based on manual inspections and maintenance records has been described, along with post-hoc comparison of prediction results to identify abnormal states, and triggering a generative adversarial network to generate synthetic data and perform incremental training based on this data in cases of few abnormal categories. Building upon this, in a preferred embodiment of the present invention, the incremental training can be implemented using a rolling update method, allowing the health assessment model or time-series prediction model to continuously absorb new monitoring data and feedback information during the system's ongoing operation. The specific operation process includes: When new monitoring data or maintenance feedback data is received based on the abnormal state identification results, the data is written into the historical monitoring data and an incremental sample batch is formed. Based on the incremental sample batch, local parameters of the health assessment model or time series prediction model are updated, and the model version is recorded. Once the model version is updated, the updated model will be used for subsequent health status assessments and risk predictions.

[0074] In one embodiment, the system continuously receives new monitoring data and maintenance feedback data related to transplanted trees during operation. The new monitoring data may include the latest image data and environmental data collected by image acquisition devices and environmental sensors, as well as structured image features, health status assessment results, and risk prediction results generated through the aforementioned processing flow. The maintenance feedback data may include manual inspection records, maintenance operation records, and confirmation results of abnormal states by professionals. When the abnormal state identification results determine that there are missed samples, false alarm samples, or few sample abnormalities within a certain time period, the system writes the health status assessment results related to these events, time-series input information, subsequent monitoring results, and any potentially generated synthetic abnormal samples into the historical monitoring data storage, and divides them into incremental sample batches according to a preset time window or sample quantity threshold.

[0075] After forming the incremental sample batch, in this embodiment, the training module extracts the incremental sample set from the historical monitoring data within the most recent period as the training data source for the current rolling update round. Specifically, for the health assessment model, samples containing structured image features, corresponding environmental features, and health status labels or health status scores corrected by subsequent monitoring can be selected from the incremental sample batch to construct the training data for the health assessment model; for the time series prediction model, samples containing updated health status sequences, historical sensor data sequences, historical maintenance records, and corresponding real future health changes and risk indicators can be selected to construct the training data for the time series prediction model. In a preferred embodiment, to reduce the disturbance of model updates to existing performance, incremental training can use a smaller learning rate and fewer iteration rounds, and can combine the strategy of assigning higher weights to high-error samples in the aforementioned feedback learning, so that the rolling update focuses more on recently exposed error samples and abnormal category samples.

[0076] In this embodiment, after each round of model parameter updates based on incremental sample batches, the system can assign a new model version identifier to the updated health assessment model or time series prediction model, and record the training time, incremental sample batch range, main training parameters, and differences from the previous version. The version record can be stored in the model management module for subsequent backtracking analysis or, if necessary, rollback to a historical version. After registering the model version update, the system switches the new model version to the online service version for health status assessment and risk prediction of subsequently collected image and environmental data. Optionally, in some implementations, the performance of the old and new models can be compared and verified on a set of validation data before model switching, and the formal version switch is only performed if the verification results meet preset performance conditions.

[0077] Through the aforementioned rolling update method, the system can continuously utilize newly collected monitoring data, maintenance feedback information, and synthetic anomaly samples obtained based on generative adversarial networks to update the local parameters of the health assessment model and time series prediction model without relying on large-scale offline retraining. This allows the model to gradually adapt to long-term changes in the monitored objects, environmental conditions, and maintenance strategies.

[0078] In the foregoing embodiments, health status assessment results and future risk prediction results for each transplanted tree have been obtained through health assessment models and time-series prediction models, respectively, and the models can be continuously optimized by combining feedback learning and incremental training. Based on this, in a preferred embodiment of the present invention, in order to transform the above assessment and prediction results into executable maintenance operations, this embodiment generates a set of maintenance strategies through a decision tree model or rule tree model, and determines the target maintenance strategy by combining historical maintenance records. The specific implementation process includes the following steps: Input the health status assessment results and risk prediction results into the decision tree model or rule tree model; The decision tree model or rule tree model outputs a set of maintenance strategies corresponding to the tree health status category and risk value range. The set of maintenance strategies includes at least one of irrigation, fertilization, pruning and pest and disease treatment. The set of maintenance strategies is matched with historical maintenance records to determine the target maintenance strategy and generate executable maintenance task parameters.

[0079] In one embodiment, an input feature vector for decision-making is first constructed for each transplanted tree included in the monitoring. This input feature vector includes at least: the health status category and health status score output by the health assessment model, the risk value or risk level output by the time-series prediction model, environmental characteristics corresponding to the current time or prediction window, basic tree attribute characteristics (such as tree species category, age category, etc.), and a summary of major maintenance operations performed in the period preceding the prediction window. In a specific implementation, the health status category can be encoded as a discrete feature, the health status score and risk value as continuous features, and the tree species category, age range, and maintenance operation type can be converted into numerical features through embedding or one-hot encoding, together forming the input feature vector for strategy decision-making.

[0080] In this embodiment, the decision tree model or rule tree model can be constructed based on historical monitoring data and historical maintenance records. In one implementation, a supervised learning decision tree model is constructed, taking health status assessment results, risk prediction results, and environmental and tree attribute characteristics from historical data as input, and using proven maintenance strategy combinations, such as watering frequency and volume, fertilizer type and application amount, pruning time and intensity, and pest and disease control methods, as output labels. The model is then trained to split at tree nodes based on health status category, risk value range, tree species, and tree age, and output a set of maintenance strategies matching different health-risk combinations at the leaf nodes. In another embodiment, a rule tree model can be pre-constructed based on the experience of forestry and maintenance experts, encoding decision rules in an "if-then" format. For example, when the health status category is sub-healthy and the risk value is in the medium-high range, while the soil moisture is below the threshold, the priority of irrigation operations in the strategy set is increased; when the risk value for a certain type of pest or disease is high and the peak season has arrived, pest and disease treatment operations are added to the target strategy set, etc.

[0081] After generating an initial set of maintenance strategies using a decision tree or rule tree model, this embodiment further matches this set of strategies with historical maintenance records to select target maintenance strategies for the current situation. Specifically, historical samples with similar tree species, age, and health-risk characteristics to the current tree can be retrieved from a historical database. Evaluation indicators of the effectiveness of different maintenance strategy combinations on these samples are statistically analyzed, such as the degree of improvement in health status and the recurrence of pests and diseases in subsequent monitoring. Based on these statistical results, the candidate maintenance operations in the current strategy set are ranked or screened. Strategy combinations with historically stable effects and significant improvement can be assigned higher selection weights; strategy combinations with poor effects or those that have caused adverse results can have their priority reduced or be eliminated in the current situation. After matching and screening with historical maintenance records, target maintenance strategies suitable for the current tree health status, risk range, and environmental conditions can be obtained.

[0082] In a preferred embodiment, for the finally determined target maintenance strategy, this embodiment further generates executable maintenance task parameters. Specifically, this includes: determining the maintenance task category based on the strategy type output by the decision tree or rule tree, such as irrigation, fertilization, pruning, or pest and disease control; and quantifying the task execution time (e.g., specific date and time period), execution frequency (e.g., weekly frequency or execution within a prediction window), operation intensity (e.g., irrigation water volume range, fertilizer dosage range, pruning ratio, or pesticide dosage range), and related resource allocation (e.g., required equipment or personnel quantity) based on empirical parameters from the tree's location and historical maintenance records, combined with current environmental conditions, to form a structured set of maintenance task parameters. These maintenance task parameters can be stored in a task scheduling system and distributed to the mobile terminals of maintenance personnel or automated control equipment to guide subsequent maintenance operations.

[0083] In the foregoing embodiments, the methodological process has been explained as to how the present invention sequentially completes data acquisition, image segmentation and feature extraction, health status assessment, temporal risk prediction, and dynamic model optimization based on feedback learning. In another embodiment of the present invention, the above method can be specifically implemented through a big data-based transplanted tree maintenance and monitoring system divided into functional modules, such as... Figure 2 As shown, the transplanted tree maintenance and monitoring system includes: a data acquisition unit, an image processing unit, a health assessment unit, a risk prediction unit, and a feedback optimization unit.

[0084] In one embodiment, the big data-based transplanted tree maintenance monitoring system may include front-end acquisition equipment deployed at the monitoring site and a data processing and analysis platform deployed on the central side or cloud platform. Each functional unit may be implemented by a software module or hardware circuit executed by a processor.

[0085] In one specific embodiment, the data acquisition unit may include: an image acquisition submodule, used to receive tree images uploaded by devices such as fixed cameras, patrol terminals, or drones, and store them according to tree identification and acquisition time; an environmental data acquisition submodule, used to receive environmental data collected by front-end soil moisture sensors, soil temperature sensors, air temperature and humidity sensors, light intensity sensors, etc., and write them into an environmental database according to a preset sampling period; and a historical data management submodule, used to read historical health tags, historical sensor acquisition data, and historical maintenance records related to the target tree from the historical database, and establish a correlation between them and the currently acquired data on the tree identification and timeline. The above-mentioned data acquisition unit can communicate with the on-site acquisition equipment via wired or wireless networks, uniformly storing multi-source data into the system's data storage module, providing input for subsequent image processing and model inference.

[0086] In one embodiment, the image processing unit may include an image preprocessing submodule and a segmentation and feature extraction submodule: the image preprocessing submodule is used to perform size unification, brightness and contrast normalization, noise suppression, and quality screening on the acquired tree images; the segmentation and feature extraction submodule is used to invoke a multi-scale convolutional neural network to perform pixel-level segmentation of the tree leaf and branch regions in the image, and to perform feature extraction and multi-scale feature fusion on the segmented target regions. After eliminating small noise points and correcting boundaries through morphological optimization operations, it outputs structured image features for subsequent analysis. The output of the image processing unit is transmitted to the health assessment unit through an internal bus or software interface.

[0087] In a preferred embodiment, the health assessment unit incorporates an inference module for the health assessment model, which can employ a multi-branch neural network structure, including an image feature encoding branch, an environmental feature encoding branch, and a feature interaction and decision branch. The image feature encoding branch receives structured image features output by the image processing unit and encodes them using structures such as convolutional neural networks, residual networks, or visual Transformers to obtain an image latent representation. The environmental feature encoding branch receives environmental features aligned with the image at the current moment and encodes them using a multilayer perceptron or a short-time LSTM / GRU structure to obtain an environmental latent representation. The feature interaction and decision branch concatenates, interacts with, or fuses the image latent representation and the environmental latent representation through attention, and outputs the tree's health status category and health status score through a classification subnetwork and a scoring subnetwork, constituting the health status assessment result. The health assessment unit can periodically or as needed perform assessments on trees at each monitoring point and store the results in historical monitoring data.

[0088] In one embodiment, the risk prediction unit may include a temporal feature construction submodule and a temporal prediction model inference submodule. The temporal feature construction submodule constructs a health status sequence based on the health status categories and scores output by the health assessment unit at multiple monitoring times, and aligns it with the historical sensor data and historical maintenance records for the corresponding time period in the historical monitoring data to generate multi-dimensional temporal input information. The temporal prediction model inference submodule receives the aforementioned temporal input information, calls a pre-trained temporal prediction model (e.g., LSTM, GRU, Temporal Convolutional Network TCN, or a self-attention-based temporal network), outputs the health status change trend and risk score within a preset prediction window, stores the risk score as part of the risk prediction result, and provides it to the feedback optimization unit and the maintenance decision module for use.

[0089] In a preferred embodiment, the feedback optimization unit may include a feedback sample construction submodule and a model update submodule. The feedback sample construction submodule receives risk prediction results and extracts subsequent monitoring results corresponding to the prediction window time period from historical monitoring data maintained by the data acquisition unit. These results include subsequent health assessment results, manual inspection feedback, and actual maintenance records. Based on the differences between these two data, it generates a supervision signal and feedback error information. The model update submodule constructs a feedback loss function based on the feedback error, weights the feedback samples, assigning higher weights to samples that cause significant risk prediction errors, and calls the training interface of the health assessment model to perform local parameter updates or incremental training. The updated model parameters are then written to the model storage and management module.

[0090] In an implementation combining generative adversarial networks and incremental training mechanisms, the feedback optimization unit can also collaborate with the module used to generate synthetic anomalous samples to focus on learning anomalous categories with few samples, thereby further improving the system's adaptability to long-tailed anomalous scenarios.

[0091] During system operation, the data acquisition unit continuously collects multi-source monitoring data, the image processing unit and health assessment unit analyze the current health status of the trees, the risk prediction unit predicts the health change trend and risk within the future time window, and the feedback optimization unit dynamically optimizes the health assessment model based on the deviation between the prediction results and subsequent monitoring results, so as to realize the function of the transplanted tree maintenance monitoring system.

[0092] The technical solution of the present invention has been described above with reference to specific embodiments. It should be understood that the above embodiments are only used to explain the technical concept of the present invention, and not to limit the scope of protection of the present invention; those skilled in the art can make equivalent substitutions or modifications to the specific implementation methods, parameter settings, network structure details and combination relationships of each unit without departing from the concept of the present invention, and all such substitutions or modifications should fall within the scope of protection of the present invention.

[0093] In summary, the embodiments described in this specification are merely illustrative of the technical solutions of this invention and are not intended to limit the scope of protection of this invention. Unless explicitly stated that they are mutually exclusive or incompatible, the technical features described in the embodiments of this invention can be combined, replaced, or deleted by those skilled in the art according to specific needs to form other implementation methods. Any equivalent substitutions or obvious modifications made to the above technical solutions without departing from the spirit and substance of this invention should be considered to fall within the scope of protection of this invention, which is defined by the appended claims.

Claims

1. A method for monitoring the maintenance of transplanted trees based on big data, characterized in that, The method includes: Acquire current image data, environmental data, and historical monitoring data of transplanted trees. The historical monitoring data includes at least the historical health tags of transplanted trees, historical sensor data, and historical maintenance records. The image data is segmented to generate structured image features; Based on structured image features and environmental data, a health assessment model is used to assess the health status of trees and generate health status assessment results. Based on the health status assessment results and historical monitoring data, a time-series prediction model is used to predict the future health status of trees and generate risk prediction results. The risk prediction results are used as feedback signals to dynamically optimize the health assessment model through a feedback learning mechanism.

2. The method for monitoring and maintaining transplanted trees based on big data according to claim 1, characterized in that, The image segmentation process includes: A multi-scale convolutional neural network was used to segment the target region of transplanted trees in image data, separating the target regions of tree leaves and branches. Feature extraction is performed on the target region, and the structural features of the trees are obtained through deep feature fusion. Morphological operations are used to optimize the segmentation results, eliminate small noise points, improve the boundaries of segmented regions, and generate structured image features.

3. The method for monitoring and maintaining transplanted trees based on big data according to claim 1, characterized in that, The method of using a health assessment model to assess the health status of trees includes: Structured image features are aligned and fused with environmental data to obtain fused features; The fused features are input into the health assessment model, which outputs the tree's health status category and health status score as the health status assessment result. The health assessment model encodes image features and environmental features separately and performs feature interaction so that the health status assessment results simultaneously reflect the characteristics of tree morphology and the impact of changes in environmental conditions on tree growth.

4. The method for monitoring and maintaining transplanted trees based on big data according to claim 1, characterized in that, The method of using a time-series prediction model to predict the future health status of trees includes: The health status assessment results are constructed into a health status sequence in chronological order and time-aligned with the historical sensor data and historical maintenance records in the historical monitoring data to obtain the time-series input information. Input the time-series input information into the time-series prediction model to predict the changing trend of tree health status and the corresponding risk value within a preset window, which is used as the risk prediction result; The preset window is determined at least based on the sampling period of historical monitoring data, and the risk value is the risk score output by the time series prediction model, which is used to characterize the probability or degree of the tree's health status evolving from healthy to sub-healthy or unhealthy.

5. The method for monitoring and maintaining transplanted trees based on big data according to claim 1, characterized in that, The dynamic optimization of the health assessment model through a feedback learning mechanism includes: After obtaining the risk prediction results, the subsequent monitoring results within the time period corresponding to the risk prediction results are obtained, and a supervision signal for feedback learning is generated based on the subsequent monitoring results. Based on the difference between the risk prediction results and the monitoring signals, a feedback loss is constructed and the parameters of the health assessment model are updated to reduce the health status assessment error and further reduce the risk prediction error caused by the propagation of the health status assessment error. The feedback learning mechanism updates the health assessment model by at least the following: when constructing the feedback loss, assigning higher weights to samples that lead to higher risk prediction errors.

6. The method for monitoring and maintaining transplanted trees based on big data according to claim 1 or 5, characterized in that, The method further includes: A generative adversarial network is constructed, and expert knowledge related to transplanted trees is used as a conditional input to constrain the generation process. The expert knowledge includes at least one of the following: tree species category, tree age category, pest and disease type, climate and seasonal factors. Based on the conditional input, synthetic data matching the anomaly category is generated. The synthetic data includes at least synthetic image data or synthetic time-series data that is isomorphic to the sensor-acquired data. Consistency constraints are applied to the synthesized data to ensure that the synthesized data meets preset ecological rationality conditions, which include at least physiological indicator range constraints or environmental variable coupling constraints.

7. The method for monitoring and maintaining transplanted trees based on big data according to claim 6, characterized in that, The feedback learning mechanism further includes: Based on feedback from manual inspections and maintenance records, the feedback results are compared with the predicted events corresponding to the risk prediction results to identify abnormal states that the model missed or predicted incorrectly. When the number of samples corresponding to an abnormal state in the training dataset is less than a threshold, it is judged as a few-sample abnormal category, triggering the generative adversarial network to generate synthetic data corresponding to the few-sample abnormal category. Synthetic data is combined with existing training data, and incremental training is performed on the health assessment model and time series prediction model to optimize model parameters.

8. The method for monitoring and maintaining transplanted trees based on big data according to claim 7, characterized in that, The incremental training adopts a rolling update method and includes at least: When new monitoring data or maintenance feedback data is received based on the abnormal state identification results, the data is written into the historical monitoring data and an incremental sample batch is formed. Based on the incremental sample batch, local parameters of the health assessment model or time series prediction model are updated, and the model version is recorded. Once the model version is updated, the updated model will be used for subsequent health status assessments and risk predictions.

9. The method for monitoring and maintaining transplanted trees based on big data according to claim 1, characterized in that, The method further includes: Input the health status assessment results and risk prediction results into the decision tree model or rule tree model; The decision tree model or rule tree model outputs a set of maintenance strategies corresponding to the tree health status category and risk value range. The set of maintenance strategies includes at least one of irrigation, fertilization, pruning and pest and disease treatment. The set of maintenance strategies is matched with historical maintenance records to determine the target maintenance strategy and generate executable maintenance task parameters.

10. A big data-based monitoring system for the maintenance of transplanted trees, characterized in that, The system includes: The data acquisition unit is used to acquire current image data, environmental data and historical monitoring data of transplanted trees. The historical monitoring data includes at least the historical health tags of transplanted trees, historical sensor data, and historical maintenance records. The image processing unit is used to perform image segmentation processing on the image data and generate structured image features; The health assessment unit is used to assess the health status of trees based on the structured image features and the environmental data using a health assessment model, and generate a health status assessment result. The risk prediction unit is used to predict the future health status of trees based on the health status assessment results and the historical monitoring data, and generate risk prediction results by using a time-series prediction model. The feedback optimization unit is used to use the risk prediction result as a feedback signal to dynamically optimize the health assessment model through a feedback learning mechanism.

Images