A tree sap flow and growth synchronous monitoring method and system and a storage medium

Abstract

The present application relates to the field of agricultural forestry informatization technology, in particular to a tree sap flow and growth synchronous monitoring method, system and storage medium, through deploying multi-modal sensor nodes to synchronously collect multi-modal data; after preprocessing the original data, constructing a tree ecological and physiological data set; training a neural network model based on the data set, realizing end-to-end mapping from original multi-modal data to sap flow density and net growth, and embedding the model in a digital twin; the digital twin collects the predicted values at each time to construct a historical sequence, and calculates the sap flow density and net growth at future time. The present application organically integrates ecological and physiological mechanisms with deep learning, improves the accuracy, anti-interference ability and automation level of tree physiological monitoring in complex field environment, and improves the defects of environmental noise and physiological signal aliasing in original monitoring data.

Description

Technical Field

[0001] This invention relates to the field of agricultural and forestry information technology, specifically to a method, system, and storage medium for synchronous monitoring of tree sap flow and growth. Background Technology

[0002] In the research and practice of forest ecology, global change ecology, and precision forestry, continuous and accurate monitoring of tree physiological status has significant scientific importance and application value. Tree sap flow is a key indicator characterizing the efficiency of water transport and use in trees, directly reflecting transpiration and water stress. Trunk radial variation includes elastic contraction and expansion caused by water storage and release, as well as irreversible growth information resulting from xylem cell division and expansion; it is an important parameter for assessing tree carbon sequestration capacity and growth vitality. Simultaneously acquiring high-precision sap flow and growth data plays a crucial guiding role in deeply understanding the carbon-water coupling mechanism of trees, assessing forest health, and responding to climate change.

[0003] Currently, data acquisition of tree sap flow mainly utilizes trunk sap flow probes based on the Granier thermal diffusion principle, while obtaining radial changes in the trunk primarily relies on high-precision point-based trunk radial change recorders. However, existing monitoring technologies still have the following shortcomings in data acquisition, processing, and analysis: First, the raw data collected by sensors is highly susceptible to interference from the field environment. Drastic changes in solar radiation, temperature drift, wind-induced tree trunk swaying, and poor sensor contact or power supply fluctuations can all lead to a large amount of spike noise, outliers, and missing data in the raw data. Traditional data processing methods often rely on manual experience for cleaning, which is inefficient and lacks standardization.

[0004] Secondly, sap flow data are significantly affected by meteorological factors. Changes in sap flow density simultaneously include the direct driving effects of meteorological factors such as solar radiation and saturated vapor pressure difference, as well as the regulatory effects of the tree's own water state. These two factors are highly colinear and difficult to separate effectively. Similarly, radial changes in the trunk contain both reversible water deficit fluctuations and irreversible net growth components, which traditional methods often struggle to accurately distinguish.

[0005] Furthermore, existing monitoring methods mostly involve independent analysis of single indicators, failing to organically integrate and collaboratively model sap flow and growth data. This makes it difficult to leverage the coupling relationship between the two for mutual constraint and enhanced prediction, and also makes it impossible to effectively extrapolate the tree's moisture status and growth trend in the future based on historical monitoring data.

[0006] Therefore, in complex field environments, it is currently difficult to achieve automated anti-interference cleaning of sap flow and growth data, effective decoupling of meteorological and physiological drivers, accurate differentiation between net growth and reversible water deficit, and multimodal collaborative modeling and future trend prediction of sap flow and growth, thus affecting the monitoring effect of tree sap flow and growth. Summary of the Invention

[0007] The purpose of this invention is to provide a method, system and storage medium for synchronous monitoring of tree sap flow and growth, in order to solve the technical problems in the prior art that make it difficult to achieve automated anti-interference cleaning of sap flow and growth data, effective decoupling of meteorological and physiological driving forces, accurate differentiation of net growth and reversible water deficit, and multimodal collaborative modeling and future trend prediction of sap flow and growth.

[0008] To solve the above-mentioned technical problems, the present invention specifically provides the following technical solution: A method for simultaneously monitoring tree sap flow and growth includes the following steps: Step S1: Simultaneously collect raw sap flow temperature difference data, trunk radial change data, soil moisture data, and micro-meteorological data of trees through multiple self-powered multimodal sensor nodes deployed in the monitoring area. The multimodal sensor nodes include a trunk sap flow probe based on the Granier thermal diffusion principle, a high-precision point-shaped trunk radial change recorder, a soil moisture sensor, and a micro-meteorological station. Step S2: Preprocess the collected raw sap flow temperature difference data, trunk radial variation data, soil moisture data and micro-meteorological data, and decouple the preprocessed raw sap flow temperature difference data from the environment to obtain sap flow density data driven by the tree's internal water state after removing micro-meteorological factors. Decompose the preprocessed trunk radial variation data into a zero-growth model to obtain the tree's net growth data. Step S3: Combine the preprocessed raw sap flow temperature difference data, trunk radial variation data, soil moisture data, and micrometeorological data with the sap flow density data and the net growth data to form a tree ecology and physiology dataset. Step S4: Based on the tree ecology and physiology dataset, construct and train a neural network model to predict sap flow density data and net growth data based on the original sap flow temperature difference data, trunk radial variation data, soil moisture data and micrometeorological data. Embed the neural network model into a digital twin constructed based on the real-world tree growth scene. The digital twin and the multimodal sensor have a data interaction channel for the digital twin to receive data collected in real time by the multimodal sensor. Step S5: Use a digital twin to collect sap flow density data and net growth data predicted by the neural network model at all times, and perform time series extrapolation based on the sap flow density data and net growth data at all times to obtain sap flow density data and net growth data in the future time series, so as to characterize the tree's water trend and growth trend.

[0009] As a preferred embodiment of the present invention, the environmental decoupling method for the fluid density data includes: The initial sap flow density, obtained by Granier empirical formula after preprocessing, was used as the result variable. The trunk water deficit, obtained by decomposing the trunk radial variation data through a zero-growth model, was used as the treatment variable. Solar radiation and saturated vapor pressure difference were used as confounding variables. A dual machine learning framework was used to decouple the initial sap flow density from the environment. In the first stage of the dual machine learning framework, the processing variable and the result variable are fitted with the confusion variable respectively to obtain the residuals of the processing variable and the residuals of the result variable; The second stage of the dual machine learning framework regresses and estimates the causal effect of the processing variable on the outcome variable in the residual space to separate the portion of the initial sap flow density affected by solar radiation and saturated vapor pressure difference from the portion driven by the tree's intrinsic moisture state, thus obtaining the sap flow density data driven only by the tree's intrinsic moisture state.

[0010] As a preferred embodiment of the present invention, the net growth data and the trunk water deficit are obtained by decomposing using a zero-growth model, the decomposition method comprising: The preprocessed radial variation data of the tree trunk is used as a time series of tree trunk diameters, with the maximum tree trunk diameter in the time series as the reference benchmark, where: When the trunk diameter at the current moment in the trunk diameter time series is greater than or equal to the maximum trunk diameter, the excess portion is recorded as net growth. When the trunk diameter at the current moment in the time sequence is less than the maximum trunk diameter, the difference is recorded as trunk water deficit.

[0011] As a preferred embodiment of the present invention, the neural network model includes a multimodal feature fusion layer, a two-stream feature extraction network, a cross-modal attention interaction module, and a multi-task prediction head; The preprocessed raw sap flow temperature difference data, trunk radial variation data, soil moisture data, and micrometeorological data are input into the multimodal feature fusion layer. Through embedding mapping and layer normalization processing, the data of each modality are aligned to the same feature space to obtain multimodal fusion features. The multimodal fusion features are respectively input into the first and second flow branches of the dual-flow feature extraction network. The first flow branch is composed of stacked convolutional modules and bidirectional recurrent network layers, and is used to extract fluid flow-related features from the multimodal fusion features. The second flow branch is composed of stacked convolutional modules and bidirectional recurrent network layers, and is used to extract growth-related features from the multimodal fusion features. The fluid flow feature vector output from the first flow branch and the growth feature vector output from the second flow branch are input into the cross-modal attention interaction module. The cross-modal attention interaction module uses a multi-head attention mechanism to perform attention calculation by using the fluid flow feature vector as a query and the growth feature vector as a key and value to obtain the fluid flow enhancement feature that integrates growth information. Symmetrically, the growth feature vector is used as a query and the fluid flow feature vector is used as a key and value to obtain the growth enhancement feature that integrates fluid flow information. The sap flow enhancement features and the growth enhancement features are respectively input into the first and second prediction heads of the multi-task prediction head. The predicted values ​​of sap flow density data and net growth data corresponding to the preprocessed original sap flow temperature difference data, trunk radial variation data, soil moisture data and micrometeorological data are obtained by mapping through the fully connected layer.

[0012] As a preferred embodiment of the present invention, the loss function of the neural network model is composed of a weighted sum of three parts: liquid flow density prediction loss, net growth prediction loss, and consistency constraint loss. The expression for the loss function is: ; In the formula, For the total loss, The mean square error for fluid flow prediction. Mean square error for trunk growth prediction For time-series consistency loss, For hyperparameters; The fluid flow density prediction loss is the mean square error between the fluid flow density prediction value output by the first prediction head and the fluid flow density data, which is used to constrain the prediction accuracy of fluid flow density. The expression for the fluid density prediction loss is: In the formula, This represents the model's predicted fluid density at time t. This corresponds to the true value of the fluid density. This represents the number of time steps; The net growth prediction loss is the mean square error between the net growth prediction value output by the second prediction head and the net growth data, which is used to constrain the prediction accuracy of net growth. The expression for the net growth prediction loss is: In the formula, This represents the model's predicted net growth at time t. This is the true value of the corresponding net growth. The consistency constraint loss is used to constrain the synergy between the predicted sap flow density and the predicted net growth in terms of their changing trends, so that the two remain in sync with the coupling relationship between tree water transport and radial growth. The expression for the consistency constraint loss is: In the formula, The Pearson correlation coefficient is used. This is a sequence of predicted fluid density values ​​at various time points. It is a net growth prediction sequence composed of the predicted net growth values ​​at each time point.

[0013] As a preferred embodiment of the present invention, a digital twin is used to collect fluid density data and net growth data predicted by a neural network model at all times, and time-series extrapolation is performed based on the fluid density data and net growth data at all times. The method includes: The predicted fluid density values ​​output by the neural network model at each historical moment are arranged in chronological order to construct a historical sequence of fluid density. Arrange the predicted net growth values ​​output by the neural network model at each historical moment in chronological order to construct a historical sequence of net growth. The Holt-Winters exponential smoothing method is used to extrapolate the trend of the historical sequence of liquid flow density, extract the horizontal component, trend component and seasonal component of liquid flow density, and recursively extrapolate based on the three components to obtain the predicted value of liquid flow density at future time. The historical sequence of net growth is accumulated day by day to construct a cumulative net growth sequence. The cumulative net growth sequence is fitted with a Logistic growth model to obtain three growth parameters: growth upper limit, growth rate, and inflection point time. Based on the fitted Logistic growth model, the cumulative net growth at future time is extrapolated, and then the net growth at future time is restored by difference. When the predicted sap flow density at future times shows a continuous downward trend and the predicted net growth at future times shows a slowing or stagnant growth trend, it is determined that the trees are at potential risk of water stress.

[0014] To address the aforementioned technical problems, the present invention further provides the following technical solution: A tree sap flow and growth synchronous monitoring system, applied to a method for synchronous monitoring of tree sap flow and growth, the system comprising: Multiple self-powered multimodal sensor nodes are used to synchronously collect raw sap flow temperature difference data, trunk radial change data, soil moisture data and micro-meteorological data of trees. The multimodal sensor nodes include a trunk sap flow probe based on the Granier thermal diffusion principle, a high-precision point-like trunk radial change recorder, a soil moisture sensor and a micro-meteorological station. An edge computing gateway, communicatively connected to the multimodal sensor nodes, is used to preprocess the collected raw sap flow temperature difference data, trunk radial variation data, soil moisture data, and micrometeorological data. It then decouples the preprocessed raw sap flow temperature difference data from the environment to obtain sap flow density data driven only by the tree's intrinsic water state, removing micrometeorological factors. The gateway also performs zero-growth model decomposition on the preprocessed trunk radial variation data to obtain the tree's net growth data. Finally, it combines the preprocessed raw sap flow temperature difference data, trunk radial variation data, soil moisture data, and micrometeorological data with the sap flow density data and the net growth data to form a tree ecology and physiology dataset. The digital twin platform is communicatively connected to the edge computing gateway and is used to construct and train a neural network model based on the tree ecology and physiology dataset to predict sap flow density data and net growth data based on the original sap flow temperature difference data, trunk radial change data, soil moisture data and micrometeorological data, and to embed the neural network model into the digital twin constructed based on the real-world tree growth scene. The digital twin has a data interaction channel with the multimodal sensor node, used to receive data collected in real time by the multimodal sensor node; and to collect sap flow density data and net growth data predicted by the neural network model at all times using the digital twin, and to perform time-series extrapolation based on the sap flow density data and net growth data at all times to obtain sap flow density data and net growth data in the future time series, so as to characterize the tree's water trend and growth trend, and at the same time generate early warning information based on potential risks and push it to the user terminal.

[0015] To address the aforementioned technical problems, the present invention further provides the following technical solution: A computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement a method for synchronously monitoring tree sap flow and growth.

[0016] Compared with the prior art, the present invention has the following advantages: This invention effectively eliminates the collinear interference of meteorological factors on sap flow signals by introducing environmental decoupling and zero-growth model decomposition into the data preprocessing stage. It also achieves accurate differentiation between irreversible net growth and reversible water deficit in the radial changes of tree trunks, solving the technical bottleneck of traditional methods in extracting pure physiological signals from raw monitoring data. On this basis, a neural network model with multimodal feature fusion, dual-flow feature extraction, and cross-modal attention interaction as its core is constructed. It utilizes the coupling relationship between sap flow and growth to achieve multi-task collaborative prediction, and constrains the synchronous changes of the two prediction values ​​through temporal consistency loss, significantly improving the prediction accuracy and physiological rationality of sap flow density and net growth. Furthermore, the trained neural network model is embedded in a digital twin. By collecting the prediction values ​​at each time point and using exponential smoothing and growth models for temporal extrapolation, it achieves a leap from current state monitoring to future trend prediction, providing a reliable technical means for early warning of tree water stress and forest health management. Attached Figure Description

[0017] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0018] Figure 1 This is a flowchart of the tree sap flow and growth synchronous monitoring method disclosed in an embodiment of the present invention.

[0019] Figure 2 This is a schematic diagram illustrating the deployment of sensor nodes in different water level gradient zones of a typical wetland forest, as disclosed in an embodiment of the present invention.

[0020] Figure 3 This is a schematic diagram of the installation of the Granier thermal diffusion probe and high-precision growth instrument disclosed in an embodiment of the present invention.

[0021] Figure 4 The time series curves of trunk diameter change and trunk sap flow disclosed in the embodiments of the present invention are shown. Detailed Implementation

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

[0024] like Figure 1 As shown, this invention provides a method for synchronously monitoring tree sap flow and growth. First, at the sensing layer, self-powered multimodal sensor nodes deployed in the monitoring area synchronously collect raw sap flow temperature difference, trunk radial variation, soil moisture, and micrometeorological data, providing a multimodal data foundation for subsequent analysis. At the data processing layer, after preprocessing the raw data, an environment decoupling method based on a dual machine learning framework is used to separate the meteorological driving component from the initial sap flow density, obtaining sap flow density data driven solely by the tree's internal water state. Furthermore, a zero-growth model is used to decompose the trunk radial variation into irreversible net growth and reversible trunk water deficit, thereby solving the technical challenge of extracting pure physiological indicators from raw signals using traditional methods.

[0025] At the modeling layer, based on a tree ecology and physiology dataset composed of the aforementioned pure physiological indicators and multimodal environmental data, a neural network model is constructed and trained, with multimodal feature fusion, dual-flow feature extraction, and cross-modal attention interaction as its core. This model utilizes the coupling relationship between sap flow and growth to achieve end-to-end synchronous mapping from raw monitoring data to sap flow density and net growth, and is embedded into a digital twin constructed based on real-world tree growth scenarios. At the extrapolation layer, the digital twin collects the model's predicted values ​​at various times and constructs a historical sequence. Exponential smoothing and a growth model are used for temporal extrapolation to obtain future sap flow density and net growth, thereby characterizing the tree's water and growth trends and providing a basis for decision-making regarding early warning of water stress and forest health management.

[0026] Specifically, the first step involves simultaneously collecting raw sap flow temperature difference data, trunk radial change data, soil moisture data, and micro-meteorological data of trees through multiple self-powered multimodal sensor nodes deployed in the monitoring area. The multimodal sensor nodes include a trunk sap flow probe based on the Granier thermal diffusion principle, a high-precision point-like trunk radial change recorder, a soil moisture sensor, and a micro-meteorological station.

[0027] Among these, sap flow temperature difference data was acquired using a trunk sap flow probe based on the Granier thermal diffusion principle. This technology utilizes the temperature difference between a heated probe and a reference probe to invert the sap flow rate in the trunk, offering advantages such as simple structure, low power consumption, and suitability for long-term field deployment. Trunk radial variation data was acquired using a high-precision point-based trunk radial variation recorder with micron-level accuracy, capable of capturing subtle elastic deformations in the trunk caused by water storage and release, as well as irreversible growth due to xylem cell division. Soil moisture sensors and micro-weather stations simultaneously collected data on soil volumetric water content, solar radiation, saturated vapor pressure difference, air temperature, precipitation, and other environmental variables. The simultaneous acquisition of these four types of data provides a complete data foundation for subsequent analysis of tree physiological processes from an atmosphere-plant-soil continuum perspective, effectively overcoming the inherent limitation of single-indicator monitoring in distinguishing between internal and external driving factors.

[0028] Specifically, in the second step, the collected raw sap flow temperature difference data, trunk radial variation data, soil moisture data, and micrometeorological data are preprocessed. The preprocessed raw sap flow temperature difference data is then decoupled from the environment to obtain sap flow density data driven only by the tree's internal water state after removing micrometeorological factors. The preprocessed trunk radial variation data is then decomposed using a zero-growth model to obtain the tree's net growth data.

[0029] Environmental decoupling methods for fluid density data include: The initial sap flow density, obtained by Granier empirical formula after preprocessing, was used as the result variable. The trunk water deficit, obtained by decomposing the trunk radial variation data through a zero-growth model, was used as the treatment variable. Solar radiation and saturated vapor pressure difference were used as confounding variables. A dual machine learning framework was used to decouple the initial sap flow density from the environment. In the first stage of the dual machine learning framework, the processing variable and the outcome variable are fitted with confounding variables respectively to obtain the residuals of the processing variable and the residuals of the outcome variable. The second stage of the dual machine learning framework regresses and estimates the causal effects of the processing variables on the outcome variables in the residual space to separate the portion of the initial sap flow density affected by solar radiation and saturated vapor pressure difference from the portion driven by the tree’s intrinsic water state, thus obtaining sap flow density data driven only by the tree’s intrinsic water state.

[0030] The net growth data and trunk water deficit were obtained by decomposing the data using a zero-growth model. The decomposition methods included: The preprocessed radial variation data of the tree trunk is used as a time series of tree trunk diameters, with the maximum tree trunk diameter in the time series as the reference benchmark, where: When the trunk diameter at the current moment in the time series is greater than or equal to the maximum trunk diameter, the excess portion is recorded as net growth. When the trunk diameter at the current moment in the time series is less than the maximum trunk diameter, the difference is recorded as trunk water deficit.

[0031] The change in the original sap flow density has two sources: a passive response driven by meteorological factors such as enhanced solar radiation and increased saturated vapor pressure difference, and an active regulation generated by the tree's own changes in water state to adjust stomatal conductance. The two are highly collinear.

[0032] If the raw sap flow density is directly used as a representation of the tree's water state, meteorologically driven fluctuations will be misinterpreted as physiological changes in the tree itself. This invention employs a causal inference method based on a dual machine learning framework, using solar radiation and saturated vapor pressure difference as confounding variables, trunk water deficit as a treatment variable, and the initial sap flow density obtained by transforming it using the Granier empirical formula as the outcome variable. The decoupling is performed in two stages: In the first stage, a machine learning model is used to fit the treatment variable and the outcome variable with the confounding variables respectively, and the residuals are taken to eliminate the linear and nonlinear effects of meteorological factors.

[0033] The second stage involves regression estimation in the residual space to determine the causal effect of the processing variables on the outcome variables. This separates the meteorologically driven component of the initial sap flow density from the component driven by the tree's intrinsic water state, resulting in pure sap flow density data determined solely by the tree's physiological state. This process effectively eliminates the interference of environmental collinearity on the sap flow signal, ensuring that the sap flow label data relied upon for subsequent modeling accurately reflects the tree's water use status, rather than being mixed with meteorological noise.

[0034] The temporal variation in trunk diameter comprises two distinct components: irreversible radial growth due to xylem cell division and expansion, and reversible elastic contraction and expansion due to the release or replenishment of water by the trunk's water-storing tissues. Traditional methods, which directly analyze changes in the original trunk diameter, cannot distinguish between genuine growth and temporary water fluctuations.

[0035] This invention employs a zero-growth model, using the maximum diameter recorded before the current moment in the tree trunk diameter time series as the growth baseline: when the trunk diameter exceeds the historical maximum, the excess is determined as irreversible net growth, and the maximum diameter baseline is updated simultaneously; when the trunk diameter is lower than the historical maximum, the difference is recorded as trunk water deficit, characterizing the degree of water storage deficit caused by transpiration. This model, with its extremely simple judgment rules, achieves online automatic separation of two physiological processes without relying on additional sensors or complex parameters. Thus, it provides a processing variable characterizing the tree's internal water state for environmental decoupling, and provides training labels for net growth for the neural network model, enabling the model to learn to directly predict irreversible growth signals from raw multimodal data.

[0036] Through the two processing steps described above, the original monitoring data, which was mixed with environmental noise and had overlapping components, was transformed into a pure dataset with clear physiological and ecological indicative significance—sap flow density data reflects the intrinsic transpiration driving force of trees, net growth data reflects the structural accumulation of carbon assimilation products, and trunk water deficit data reflects the water buffering status of trees. This provides high-quality supervision labels for the subsequent training of neural network models. In some specific implementations, the purpose of the first stage is to eliminate the influence of confounding variables on processing variables and outcome variables, obtain their respective residuals, and thus cut off the confounding path.

[0037] To confuse variables As the independent variable, to process the variable Using the dependent variable as the input variable, train the first machine learning model to obtain the predicted value of the input variable. Calculate the residuals of the treatment variables: .

[0038] The residuals of the treated variables represent the portion of trunk water deficit that cannot be explained by solar radiation and VPD (Vacuum-Produced Water Deficit), i.e., the pure intrinsic water state change determined by the tree's own water storage, release, or replenishment behavior after excluding the influence of meteorological factors. For example, under the same high VPD conditions, the true value of trunk water deficit differs between trees with good water status and trees that have experienced water deficit. Different, but predicted by meteorological factors Similarity means that residuals reflect these individual differences.

[0039] To confuse variables As the independent variable, and as the outcome variable Using the dependent variable as the input variable, train a second machine learning model to obtain the predicted value of the outcome variable. Calculate the residuals of the resulting variables: .

[0040] The residuals of the outcome variable represent the portion of the initial sap flow density that cannot be explained by solar radiation and VPD. In other words, after excluding the transpiration response directly driven by meteorological factors, the remaining signal in the sap flow density that is unrelated to meteorological conditions is represented. This signal reflects the physiological regulatory role of trees.

[0041] Both machine learning models use LightGBM gradient boosting trees as base learners. Their advantage lies in their ability to automatically capture the non-linear relationship between confounding variables and target variables, and to select hyperparameters through cross-validation to avoid overfitting.

[0042] In the residual space, the effects of confounding variables on treatment and outcome variables have been eliminated, and the residuals only reflect the direct relationship between treatment and outcome variables. Regression in the residual space can then be performed to estimate the causal effect of treatment variables on outcome variables: ; In the formula, For causal effect parameters, This is the random error term.

[0043] The physiological meaning of is: the purely causal change in sap density caused by a one-unit change in trunk water deficit. This refers to the sap flow density component driven by the tree's internal moisture state, separated from the initial sap flow density.

[0044] Finally, the fluid density data after environmental decoupling Obtained through the following methods: ; This data represents sap flow density driven solely by the tree's intrinsic water state, excluding meteorological drivers such as solar radiation and VPD. It characterizes the tree's active transpiration regulation level under a given intrinsic water state, rather than a passive physical evaporation response imposed by the environment.

[0045] Fluid density data obtained after environmental decoupling Together with the net growth data obtained from the zero-growth model decomposition, the trunk water deficit data, and the preprocessed soil moisture and micrometeorological data, this constitutes a tree ecology and physiology dataset. This dataset serves as a training and supervision label for the neural network model, enabling the model to learn the ability to directly output pure physiological indicators from the raw multimodal data, thereby achieving an end-to-end mapping from raw sensor signals to indicators with clear physiological significance.

[0046] Specifically, in the third step, the preprocessed raw sap flow temperature difference data, trunk radial variation data, soil moisture data, and micrometeorological data are combined with sap flow density data and net growth data to form a tree ecology and physiology dataset.

[0047] Specifically, in the fourth step, based on tree ecology and physiology datasets, a neural network model is constructed and trained to predict sap flow density and net growth data based on raw sap flow temperature difference data, trunk radial variation data, soil moisture data, and micrometeorological data. The neural network model is then embedded into a digital twin constructed based on real-world tree growth. The digital twin and the multimodal sensors have a data interaction channel for the digital twin to receive data collected in real time by the multimodal sensors.

[0048] The neural network model includes a multimodal feature fusion layer, a two-stream feature extraction network, a cross-modal attention interaction module, and a multi-task prediction head.

[0049] The preprocessed raw sap flow temperature difference data, trunk radial variation data, soil moisture data, and micrometeorological data are input into the multimodal feature fusion layer. Through embedding mapping and layer normalization, the data of each modality are aligned to the same feature space to obtain multimodal fusion features.

[0050] The raw multimodal data collected in this invention comes from different types of sensors, with significant differences in physical dimensions, numerical ranges, and sampling characteristics. For example, liquid flow temperature difference data typically fluctuates within the range of 0–15°C, tree trunk radial variation is measured in micrometers and changes slowly, soil moisture is expressed as a volume percentage, and solar radiation in micrometeorological data can reach hundreds of watts per square meter. If these data with different physical meanings are directly concatenated and input into subsequent networks, the difference in dimensions will cause the gradient update direction to be dominated by large numerical features during model training, while the role of small numerical features will be overwhelmed. By embedding mapping to project the data of each modality into a unified vector space, and then eliminating the influence of dimensions through layer normalization, the multimodal data can participate in subsequent feature extraction on an equal footing, laying the foundation for the full fusion of multimodal information.

[0051] The multimodal fusion features are input into the first and second flow branches of the dual-flow feature extraction network, respectively. The first flow branch consists of stacked convolutional modules and bidirectional recurrent network layers, and is used to extract fluid flow-related features from the multimodal fusion features. The second flow branch consists of stacked convolutional modules and bidirectional recurrent network layers, and is used to extract growth-related features from the multimodal fusion features.

[0052] The two branches have the same structure, both consisting of stacked convolutional modules and bidirectional recurrent network layers, but they undertake different feature extraction tasks: the first-stream branch focuses on extracting fluid flow-related features from the fused features, while the second-stream branch focuses on extracting features related to radial growth.

[0053] Although sap flow and growth are coupled, they belong to different physiological processes. Sap flow mainly reflects water transport and transpiration, belonging to hydraulic processes; radial growth involves the distribution of carbon assimilation products and xylem cell division, belonging to carbon metabolism processes. While they influence each other, their response characteristics and sensitive variables are not entirely the same. If a single-flow network is used to extract features from both types simultaneously, the network tends to learn the average representation of both tasks, resulting in insufficient specificity for either task. A two-flow architecture allows each branch to perform its specific function, which is beneficial for the network to learn more discriminative modality-specific features.

[0054] Within each branch, stacked convolutional modules expand the receptive field layer by layer, capturing local interaction patterns between variables from multimodal fusion features; bidirectional recurrent network layers further capture forward and backward dependencies in the sequence, ensuring that the extracted feature vectors contain contextual information. The combination of these two approaches allows the branch outputs to reflect both instantaneous responses and encode cumulative effects.

[0055] The fluid flow feature vector output from the first flow branch and the growth feature vector output from the second flow branch are input into the cross-modal attention interaction module. The cross-modal attention interaction module uses a multi-head attention mechanism to perform attention calculation with the fluid flow feature vector as the query and the growth feature vector as the key and value to obtain the fluid flow enhancement feature that integrates growth information. Symmetrically, the growth feature vector is used as the query and the fluid flow feature vector is used as the key and value to obtain the growth enhancement feature that integrates fluid flow information.

[0056] This module uses a multi-head attention mechanism for bidirectional information interaction: on the one hand, it uses the fluid flow feature vector as a query and the growth feature vector as a key and value for attention calculation to obtain fluid flow enhancement features that incorporate growth information; on the other hand, it symmetrically uses the growth feature vector as a query and the fluid flow feature vector as a key and value to obtain growth enhancement features that incorporate fluid flow information.

[0057] A clear physiological coupling exists between tree sap flow and radial growth. Water is the raw material for photosynthesis, and the sap flow rate determines the canopy's transpiration cooling capacity and carbon assimilation rate. Photosynthetic products, in turn, form the material basis for xylem cell division and radial growth. Simultaneously, intense diurnal transpiration leads to the release of water stored in the trunk, resulting in reversible contraction, while nighttime water replenishment causes the trunk to expand and recover. This coupling manifests as a synergistic change and phase lag between the two over time. Traditional independent modeling or simple splicing and fusion methods cannot capture this cross-modal correlation.

[0058] The cross-modal attention interaction module utilizes a query-key-value attention mechanism, allowing flow features to proactively query information relevant to themselves within growth features, and vice versa. For example, when a flow feature vector is used as a query, the attention mechanism automatically focuses on the part of the growth feature most relevant to the current flow state—such as whether recent changes in growth rate predict changes in the sapwood's water-conducting area, thus affecting flow density. The symmetrical bidirectional design ensures full information exchange between the two modalities, rather than a unidirectional flow. Multi-head attention simultaneously calculates attention weights from multiple subspaces, enabling the model to simultaneously focus on different types of cross-modal correlations.

[0059] The sap flow enhancement features and growth enhancement features are input into the first and second prediction heads of the multi-task prediction head, respectively. Through the fully connected layer, the predicted values ​​of sap flow density data and net growth data corresponding to the preprocessed original sap flow temperature difference data, trunk radial variation data, soil moisture data and micrometeorological data are obtained.

[0060] The two tasks share the underlying feature extraction and cross-modal interaction modules, allowing them to benefit from each other during training. The water state features learned by the sap flow prediction task help the growth prediction task determine the water constraints of carbon assimilation, while the structural change features learned by the growth prediction task help the sap flow prediction task perceive the impact of sapwood area changes on water conductivity. The loss function used during training consists of a weighted sum of three parts: sap flow density prediction loss, net growth prediction loss, and consistency constraint loss. The consistency constraint loss uses the Pearson correlation coefficient to measure the coordinated change trend of the two prediction sequences, forcing the model to learn a prediction pattern that conforms to the physiological law of tree sap flow decline accompanied by slowed growth, thereby improving the prediction accuracy and physiological rationality of both tasks.

[0061] The trained neural network model is embedded into a digital twin built based on real-world tree growth. A data interaction channel is established between the digital twin and multimodal sensor nodes to receive the latest data collected by the sensors in real time, and the embedded neural network model is used to synchronously output predicted values ​​for sap flow density and net growth.

[0062] A digital twin is not only a container for running models, but also a real-time mapping of real trees in virtual space. It unifies and integrates scattered sensor data, offline data preprocessing logic, and online model inference capabilities, enabling the entire process from sensor acquisition to physiological indicator output to be automated. At the same time, the digital twin continuously collects predictive outputs at various times, accumulating a continuous data foundation for subsequent time-series extrapolation and trend prediction, achieving a leap from single-point monitoring to continuous twinning.

[0063] The loss function of the neural network model is composed of a weighted sum of three parts: liquid flow density prediction loss, net growth prediction loss, and consistency constraint loss. The expression for the loss function is: ; In the formula, For the total loss, The mean square error for fluid flow prediction. Mean square error for trunk growth prediction For time-series consistency loss, This is a hyperparameter.

[0064] hyperparameters The relative weights used to adjust the consistency constraint loss can achieve a balance between prediction accuracy and physiological consistency through hyperparameter tuning methods such as grid search.

[0065] and Each prediction head must be guaranteed to have its own numerical accuracy. As a bridge across tasks, it encodes the physiological coupling relationship between fluid flow and growth as an optimization objective, so that while pursuing their own accuracy, the two prediction heads are always constrained and guided by the other's information.

[0066] The fluid flow density prediction loss is the mean square error between the fluid flow density prediction value output by the first prediction head and the fluid flow density data, which is used to constrain the prediction accuracy of fluid flow density.

[0067] The expression for predicting loss based on fluid flow density is: In the formula, This represents the model's predicted fluid density at time t. This corresponds to the true value of the fluid density. This represents the time step number.

[0068] The net growth prediction loss is the mean square error between the net growth prediction value output by the second prediction head and the net growth data, which is used to constrain the prediction accuracy of net growth.

[0069] The expression for predicting loss based on net growth is: In the formula, This represents the model's predicted net growth at time t. This is the true value of the corresponding net growth.

[0070] The two losses mentioned above independently monitor the output accuracy of the two prediction heads, but they cannot guarantee that the two predicted values ​​conform to the physiological laws of trees in terms of temporal variation trends. There is a clear coupling relationship between tree water transport and radial growth: when diurnal transpiration intensifies, sap flow density increases, and the release of water stored in the trunk leads to reversible contraction; under drought stress, sap flow density continues to decrease, and net growth also tends to stagnate. If only two independent mean squared error losses are used, the model may learn results with acceptable prediction accuracy but physiological inconsistencies. For example, it might predict a decrease in sap flow density while net growth is abnormally accelerated, which violates basic tree physiology.

[0071] Consistency constraint loss is used to constrain the synergy between the predicted sap flow density and the predicted net growth in terms of their changing trends, so that the two remain in sync with the coupling relationship between tree water transport and radial growth. The expression for the consistency constraint loss is: In the formula, The Pearson correlation coefficient is used. This is a sequence of predicted fluid density values ​​at various time points. It is a net growth prediction sequence composed of the predicted net growth values ​​at each time point.

[0072] The range of values ​​for the consistency constraint loss is: When two sequences are perfectly positively correlated, the absolute value of the correlation coefficient is 1, and the loss is 0; when two sequences are completely unrelated, the correlation coefficient is 0, and the loss is 1. By minimizing this loss, the model is forced to learn to output two predicted values ​​with a co-changing trend.

[0073] It is worth noting that this invention uses a neural network model to replace the step-by-step environmental decoupling and zero-growth model decomposition to obtain sap flow density and net growth data. This is not an overcomplication of a simple problem. Specifically, environmental decoupling and zero-growth model decomposition obtain high-quality physiological indicators, but this process relies on stringent intermediate computational conditions. The Granier formula requires the maximum daily temperature difference as a baseline, and the zero-growth model requires the historical maximum trunk diameter as a reference. If data is lost due to sensor failure or data transmission interruption, the entire processing chain cannot operate. In this case, using complete historical data as training labels, the model learns an end-to-end mapping from raw multimodal data to physiological indicators. After deployment, it no longer relies on those stringent intermediate variables; it can directly output results by inputting available data at the current moment. At the same time, the model compresses multi-step serial processing into a single matrix operation, meeting the real-time inference requirements of digital twins. Through a cross-modal attention interaction module, data from each modality serves as a backup for each other. Even if the signal quality of one sensor deteriorates, reasonable results can still be inferred from other modalities.

[0074] Specifically, in the fifth step, a digital twin is used to collect sap flow density data and net growth data predicted by the neural network model at all times. Based on the sap flow density data and net growth data at all times, a time series extrapolation is performed to obtain sap flow density data and net growth data in the future time series, so as to characterize the tree's water trend and growth trend.

[0075] The aforementioned neural network model runs continuously within the digital twin, outputting predicted values ​​for sap flow density and net growth at each moment. However, these discrete current-moment predictions cannot directly represent future trends—a crucial issue. Therefore, this step utilizes the digital twin to collect the model's prediction outputs across all historical moments. A time-series extrapolation algorithm extends the discrete current-moment prediction sequence into the future, enabling advanced predictions of tree moisture and growth trends.

[0076] The method involves using a digital twin to collect fluid density and net growth data predicted by a neural network model at all time points, and then performing time-series extrapolation based on this data. The method includes: The predicted values ​​of fluid flow density output by the neural network model at each historical moment are arranged in chronological order to construct a historical sequence of fluid flow density. Arrange the predicted net growth values ​​output by the neural network model at each historical moment in chronological order to construct a historical sequence of net growth. The Holt-Winters exponential smoothing method was used to extrapolate the trend of the historical series of liquid flow density, extracting the horizontal, trend, and seasonal components of the liquid flow density. Based on these three components, the predicted values ​​of liquid flow density for future times were obtained.

[0077] The historical fluid density series was extrapolated using the Holt-Winters exponential smoothing method. This algorithm decomposes the time series into three interpretable components: a horizontal component representing the basic magnitude of the series; a trend component representing the long-term upward or downward direction of the series; and a seasonal component representing the periodic recurring fluctuation patterns in the series.

[0078] The Holt-Winters algorithm was chosen instead of simple linear extrapolation or moving averages because the liquid flow density time series exhibits a significant diurnal rhythm. Under natural conditions, liquid flow density shows a regular fluctuation with a 24-hour cycle: liquid flow is close to zero in the early morning, rises rapidly after sunrise with increased solar radiation, reaches its peak around noon, gradually declines in the afternoon, and remains at a very low level at night. Simultaneously, the peak height and overall level of this diurnal cycle change with trends under different seasons or water conditions. The Holt-Winters algorithm can effectively model these three dimensions simultaneously: the seasonal component captures the waveform characteristics of the diurnal rhythm, the trend component captures the daily changes in the peak and mean, and the level component adapts to the overall shift in the series.

[0079] Based on the three extracted components, the algorithm extrapolates forward according to the length of the seasonal cycle to generate predicted values ​​of sap flow density for future times. The rationale for this extrapolation is that, in the absence of sudden extreme weather or irrigation intervention, the diurnal rhythm and daily variation of tree transpiration have strong inertia, and short-term sap flow patterns can be effectively extended from recent observation patterns.

[0080] The historical sequence of net growth is accumulated daily to construct a cumulative net growth sequence. The cumulative net growth sequence is then fitted using a Logistic growth model to obtain three growth parameters: growth upper limit, growth rate, and inflection point time. Based on the fitted Logistic growth model, the cumulative net growth at future times is extrapolated, and then the predicted net growth value at future times is restored through difference.

[0081] Specifically, the historical net growth sequence is accumulated daily to construct a cumulative net growth sequence. Accumulation is performed first because net growth fluctuates significantly on a daily scale and is greatly influenced by random factors; directly extrapolating the original net growth sequence would introduce excessive noise. The cumulative net growth, on the other hand, exhibits a smooth, monotonically increasing trend, reflecting the overall process of radial growth in trees from initiation to acceleration and then to saturation, making it more suitable for trend modeling.

[0082] The cumulative net growth sequence was fitted using a Logistic growth model. The mathematical form of the Logistic growth model is an S-shaped curve, describing a growth process divided into three stages: slow initial growth, near-linear rapid growth in the middle stage, and a tendency towards the upper growth limit in the later stage. This curve shape closely matches the phenological rhythm of tree radial growth: slow growth begins during the spring leaf unfolding period, growth accelerates during the summer rapid growth period, and growth slows down until it stops in autumn and winter. After fitting, three growth parameters with clear physiological meanings were obtained: the upper growth limit represents the maximum cumulative radial growth of the tree during the growing season, determined by tree species characteristics, stand age, and environmental carrying capacity; the growth rate represents the growth speed during the rapid growth period, reflecting the tree's growth vitality; and the inflection point time represents the point at which the growth rate reaches its maximum, which can be used to determine the stage of the growing season.

[0083] The cumulative net growth at future times is extrapolated based on the fitted Logistic growth model, and then the predicted daily net growth at future times is obtained by difference reduction. The necessity of difference reduction is that the cumulative value only reflects the total growth and cannot reflect the daily growth rhythm, while the daily net growth is a direct indicator for assessing the growth vitality of trees on that day and judging growth abnormalities.

[0084] The key to ensuring the reliability of extrapolation results lies in the targeted selection of extrapolation algorithms that match the characteristics of the data.

[0085] When the predicted sap flow density at future time points shows a continuous downward trend and the predicted net growth at future time points shows a slowing or stagnant growth trend, the trees are considered to be at potential risk of water stress.

[0086] A sustained decrease in sap flow density indicates a weakening of the tree's water transport capacity and unmet transpiration demands; a slowdown or stagnation in net growth indicates obstruction in the allocation of carbon assimilation products to structural growth. The simultaneous occurrence of both is a reliable signal that the tree is transitioning from a normal state to a state of water stress, rather than a random fluctuation of a single indicator. Compared to single-indicator warnings relying solely on sap flow or growth, the combined assessment using these two indicators significantly reduces the false alarm rate and can detect early physiological disturbances before visible wilting or abnormal growth appears in the tree, directly transmitting abnormal risk information to the user terminal for early warning.

[0087] In some specific implementations, the system utilizes a trunk sap flow probe (TDP) based on the Granier thermal diffusion principle, a high-precision point-based trunk radial variation recorder (such as the Milanger TR4-10MM model, with an accuracy of ±2μm), a soil moisture sensor (such as the SM926 model, monitoring depths of 10 / 30 / 50cm), and a micro-weather station (such as the VantagePro2 model, monitoring air temperature and humidity, photosynthetically active radiation, wind speed, precipitation, etc.). All sensors synchronously collect data at a preset high frequency (e.g., once every 10 or 30 minutes), record the data using a CR1000X or DL800 data acquisition unit to ensure timeline alignment, and transmit the data to an edge computing gateway via wired or wireless means.

[0088] In some specific implementations... Figure 2 This diagram illustrates the deployment of sensor nodes across different water level gradient zones in a typical wetland forest, using a 1-hectare long-term monitoring plot in the Poyang Lake wetland forest near Leigongling in Lushan City, Jiangxi Province. Figure 2 As shown in Figure a, three water level gradient zones were divided according to the degree of flooding during the high-water season (lakeside forest G1, wetland transition forest G2, and lakeside terrestrial forest G3). Within each gradient zone, five 20m × 20m quadrats were selected and numbered G11-G15, G21-G25, and G31-G35. Four standard trees of each tree species were selected, for a total of 60 sample trees. Figure 2 Figure b shows the deployment of sensor nodes in G11-G15, G21-G25, and G31-G35, respectively. For example, G15 is equipped with a growth fluid system, which is composed of a Granier thermal diffusion probe and a high-precision growth instrument.

[0089] A sap flow system was installed at a height of 1.3m on each sample tree, including the following sensors: Figure 3 As shown: Granier thermal diffusion probes: a pair of probes (heating probe on top, reference probe on the bottom), each probe tube is 20mm long and 2mm in outer diameter, internally encapsulated with a copper-nickel alloy thermocouple. The probes are connected to the data acquisition unit via a four-core shielded cable.

[0090] High-precision tree growth meter: Milang TR4-10MM spring self-resetting displacement sensor, the measuring rod makes point contact with the tree trunk, with an accuracy of ±2μm. During installation, a small amount of bark is removed at 1020cm above breast height, and the growth meter is fixed in place using a special fixture. The outside is wrapped and protected with plastic foam and reflective film, which is not shown in the picture.

[0091] Other sensor nodes include a soil moisture sensor: SM926 type, which is horizontally inserted at depths of 10cm, 30cm, and 50cm in each sample plot to continuously monitor soil volumetric water content. A micro-weather station: VantagePro2 type, installed in an open area of ​​the sample plot to monitor air temperature, relative humidity, photosynthetically active radiation, wind speed, wind direction, and precipitation.

[0092] All sensors acquire data via a CR1000X data acquisition unit (32 channels, capable of connecting 15 trees (with two additional channels for signal transmission)) or a CR1000X+RM10 / 20 expansion board (62 channels, capable of connecting 30 trees), with a sampling frequency set to 30 seconds and recording a 10-minute average. The data acquisition unit and power supply system (12V battery + solar panel) are housed in a waterproof enclosure. The raw data is periodically uploaded to the edge computing gateway via a 4G router.

[0093] Fluid flow density data were obtained from the raw temperature difference measured by the Granier thermal diffusion probe. According to Granier's empirical formula The initial sap flow density was converted, and after anomaly removal and environmental decoupling, sap flow density data driven solely by the tree's intrinsic moisture state was obtained and used as training labels. The specific steps included: Raw temperature difference acquisition: Granier thermal diffusion probe (TDP) continuously measures the temperature difference between the heating probe and the reference probe. (Unit: ℃) The data acquisition device (such as CR1000X) records every 30 seconds and stores it as a 10-minute average.

[0094] Empirical formula for calculating fluid density: The temperature difference is converted into fluid density using Granier's classic formula. (unit: ): ; in The maximum temperature difference within a day (or within a set time window) is usually obtained when the liquid flow is close to zero in the early morning. The determination can be achieved using Baseliner software for automatic identification or the sliding window method.

[0095] Data collection on changes in original trunk diameter: Trunk diameter was continuously recorded using a high-precision growth meter (e.g., Milanger TR4-10MM, accuracy ±2μm). The sampling frequency is synchronized with the liquid flow temperature difference data (e.g., once every 10 minutes).

[0096] Zero-growth model decomposition: Using the zero-growth model, the change in trunk diameter is considered... It is broken down into irreversible net growth (GRO) and reversible trunk water deficit (TWD): ; ; in This represents the maximum trunk diameter recorded up to the current time. When the trunk diameter exceeds the historical maximum, the excess portion is recorded as net growth. When the trunk diameter is lower than the historical maximum, the difference is recorded as water deficit. .

[0097] The Zweifel zero-growth model decomposes the trunk diameter variation (SRV) recorded by a high-precision growth meter into irreversible net growth (GRO) and reversible trunk water deficit (TWD). Specifically, the DendRoAnalyst or treenetproc packages in R are used to calculate GRO and TWD daily. After decomposition using the zero-growth model, the following results are obtained: Figure 4 The time series curves showing the changes in trunk diameter and sap flow are shown, in which... Figure 4 (a) Average change in trunk diameter relative to seasonal changes in Toona sinensis; (b) Monthly average daily change in trunk diameter after difference correction; (c) Monthly average change in trunk sap flow relative to seasonal changes in Toona sinensis; (d) Monthly average daily change in trunk sap flow. The red curve represents irreversible net radial growth (GRO), the blue curve represents reversible trunk water deficit (TWD), and the gray curve represents the original trunk radius change (SRV). The figures clearly distinguish between elastic shrinkage / expansion caused by water fluctuations and true xylem growth.

[0098] And from Figure 4 The data shows that the changes in sap flow density and trunk diameter exhibit a clear diurnal rhythm and seasonal dynamics, with a phase lag of approximately 1-2 hours between the two, consistent with the physiological patterns of water transport in trees. This sap flow density data, together with net growth (GRO), constitutes an ecological and physiological dataset.

[0099] To verify the technical effectiveness of the neural network model of this invention, two baseline models were selected and compared on the same dataset. The results are shown in Tables 1 and 2. All models adopted the same input-output settings as this invention: the input was a multimodal vector composed of preprocessed raw sap flow temperature difference data, trunk radial variation data, soil moisture data, and micrometeorological data at the current moment; the output was the predicted sap flow density and net growth at the current moment. The training labels were sap flow density data obtained after environmental decoupling and net growth data obtained after zero-growth model decomposition.

[0100] Compared to Model 1, a single-stream convolutional network: A single-stream one-dimensional convolutional network is used as the backbone for feature extraction. The input vector, after layer normalization, is fed into a network consisting of four one-dimensional convolutional layers with a kernel size of 3 and channel numbers of 64, 128, 128, and 256 per layer, respectively. Each layer is followed by a ReLU activation function. The outputs of the convolutional layers are then subjected to global average pooling and fed into two independent fully connected heads, mapping to predicted fluid flow density and net growth, respectively. Both outputs are supervised using a mean squared error loss function, and the two losses are directly added together. The model uses the Adam optimizer with an initial learning rate of 0.001, a batch size of 64, and 200 training epochs, with an early stopping mechanism. The limitation of the single-stream convolutional network is that all input features share the same convolutional path, resulting in the mixed extraction of fluid flow-related and growth-related features, making it impossible to differentiate modeling for the two tasks.

[0101] Compared to Model 2, a two-stream convolutional network, there is no cross-modal interaction: A two-stream architecture similar to that of this invention is adopted, but without a cross-modal attention interaction module. The input vector is fed into two structurally identical but parameter-independent branches. Each branch consists of three one-dimensional convolutional layers with a kernel size of 3 and channel numbers of 64, 128, and 128 respectively, followed by a ReLU activation function after each layer. After global average pooling, the outputs of the two branches are connected to a fully connected head to output the predicted fluid flow density, and the second branch is also connected to a fully connected head to output the predicted net growth. Both outputs are supervised using a mean squared error loss function, and the two losses are directly added together. The optimizer configuration is consistent with that of the single-stream convolutional network. The two-stream convolutional network achieves independent extraction of fluid flow and growth features through branch separation, providing a targeted improvement compared to the single-stream convolutional network. However, due to the lack of information interaction between the two branches, it cannot leverage the coupling relationship between fluid flow and growth for mutual enhancement.

[0102] This invention's model, based on a two-stream convolutional network, introduces a cross-modal attention interaction module and a temporal consistency constraint loss. The key difference from a two-stream convolutional network is that the output feature vectors of the two branches are not directly fed into their respective fully connected heads, but instead first enter the cross-modal attention interaction module. Through a multi-head attention mechanism, bidirectional information exchange occurs, enabling the features of the fluid flow branch to fuse with growth information, and the features of the growth branch to fuse with fluid flow information. Furthermore, a consistency constraint loss term based on the Pearson correlation coefficient is added to the loss function, forcing the two predicted values ​​to maintain a consistent trend in their changes.

[0103] Table 1 Comparison of Liquid Flow Density Prediction Performance As shown in Table 1, the performance of sap flow density prediction improves stepwise from single-stream convolutional networks to dual-stream convolutional networks and then to this invention. The RMSE of the single-stream convolutional network is 4.53 because its shared convolutional path cannot distinguish between sap flow and growth features, making it difficult to effectively capture the specific variation patterns of sap flow density. The dual-stream convolutional network reduces the RMSE to 4.12 through branch separation, but the two branches work independently, and the sap flow branch cannot obtain clues about the water state contained in the radial changes of the tree trunk. After introducing a cross-modal attention interaction module, the RMSE of this invention drops significantly from 4.12 to 2.96, a reduction of 28.2%. 2 The gain jumped from 0.74 to 0.86. This significant gain highlights the core value of cross-modal information interaction: the sap flow branch actively queries the radial variation features of the trunk in the growth branches through the attention mechanism, and obtains cross-modal cues of the tree's internal water state, thereby enabling a more accurate distinction between meteorological and physiological driving components.

[0104] Table 2 Comparison of Net Growth Prediction Performance As shown in Table 2, single-stream convolutional networks perform poorly in the net growth prediction task, R 2 With an R value of only 0.59, it is evident that single-stream convolutional structures struggle to effectively extract subtle features related to radial growth from multimodal raw data. Two-stream convolutional networks, by independently setting growth branches, achieve a higher R value. 2 The RSE was improved to 0.67, but due to the lack of cross-modal information interaction, its growth prediction could not refer to the water constraint information contained in the fluid flow characteristics, resulting in a lag in the prediction of growth inhibition under water stress. The model of this invention, through bidirectional information flow of the cross-modal attention interaction module, enables the growth branch to actively query the water status of the fluid flow branch, thereby promptly sensing growth constraints when hydraulic conditions change. The RSE decreased from 5.48 to 3.67, a reduction of 33.0%. 2 It jumped to 0.83; at the same time, the consistency constraint loss forced the two outputs to maintain coordinated changes, further improving the physiological rationality of net growth prediction.

[0105] Therefore, the model in this invention, by introducing a cross-modal attention interaction module and temporal consistency constraint loss on the basis of dual-stream CNN, achieves a significant performance improvement: fluid density R 2 The net growth R jumped from 0.74 to 0.86. 2The RMSE jumped from 0.67 to 0.83, with reductions of 28.2% and 33.0% for the two tasks, respectively. This gain far exceeds the improvement from single-flow to dual-flow, indicating that cross-modal interaction is the core source of efficiency enhancement in this scheme. The underlying mechanism is that the cross-modal attention interaction module prevents the two branches from operating in isolation. Instead, they bidirectionally query complementary information in each other's features through a multi-head attention mechanism—the sap flow branch obtains clues about the inherent water state implied by the radial changes in the trunk from the growth branch, thus more accurately distinguishing between meteorological and physiological driving components; the growth branch obtains real-time constraints on water transport capacity from the sap flow branch, thereby adjusting growth predictions promptly when hydraulic conditions change. Furthermore, the temporal consistency constraint loss explicitly encodes the physiological coupling law between tree water transport and radial growth as an optimization objective, forcing the two predictions to maintain synergy in trends, further improving the physiological rationality of the predictions.

[0106] A tree sap flow and growth synchronous monitoring system, applied to a method for synchronous monitoring of tree sap flow and growth, the system comprising: Multiple self-powered multimodal sensor nodes are used to synchronously collect raw sap flow temperature difference data, trunk radial change data, soil moisture data and micro-meteorological data of trees. The multimodal sensor nodes include a trunk sap flow probe based on the Granier thermal diffusion principle, a high-precision point-like trunk radial change recorder, a soil moisture sensor and a micro-meteorological station. An edge computing gateway communicates with multimodal sensor nodes to preprocess the collected raw sap flow temperature difference data, trunk radial variation data, soil moisture data, and micrometeorological data. It then decouples the preprocessed raw sap flow temperature difference data from the environment to obtain sap flow density data driven only by the tree's intrinsic water state, removing micrometeorological factors. The preprocessed trunk radial variation data is decomposed using a zero-growth model to obtain the tree's net growth data. Finally, the preprocessed raw sap flow temperature difference data, trunk radial variation data, soil moisture data, and micrometeorological data are combined with the sap flow density data and net growth data to form a tree ecology and physiology dataset. The digital twin platform communicates with the edge computing gateway and is used to build and train a neural network model based on tree ecology and physiology datasets. This model is used to predict sap flow density data and net growth data based on raw sap flow temperature difference data, trunk radial change data, soil moisture data and micrometeorological data. The neural network model is then embedded into a digital twin built based on the real-world tree growth scene. The digital twin has a data interaction channel with the multimodal sensor nodes to receive data collected in real time by the multimodal sensor nodes; and uses the digital twin to collect sap flow density data and net growth data predicted by the neural network model at all times, and performs time series extrapolation based on the sap flow density data and net growth data at all times to obtain sap flow density data and net growth data in the future time series, so as to characterize the tree's water trend and growth trend. At the same time, it generates early warning information based on potential risks and pushes it to the user terminal.

[0107] A computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement a method for synchronously monitoring tree sap flow and growth.

[0108] This invention effectively eliminates the collinear interference of meteorological factors on sap flow signals by introducing environmental decoupling and zero-growth model decomposition into the data preprocessing stage. It also achieves accurate differentiation between irreversible net growth and reversible water deficit in the radial changes of tree trunks, solving the technical bottleneck of traditional methods in extracting pure physiological signals from raw monitoring data. On this basis, a neural network model with multimodal feature fusion, dual-flow feature extraction, and cross-modal attention interaction as its core is constructed. It utilizes the coupling relationship between sap flow and growth to achieve multi-task collaborative prediction, and constrains the synchronous changes of the two prediction values ​​through temporal consistency loss, significantly improving the prediction accuracy and physiological rationality of sap flow density and net growth. Furthermore, the trained neural network model is embedded in a digital twin. By collecting the prediction values ​​at each time point and using exponential smoothing and growth models for temporal extrapolation, it achieves a leap from current state monitoring to future trend prediction, providing a reliable technical means for early warning of tree water stress and forest health management.

[0109] The above embodiments are merely exemplary embodiments of this application and are not intended to limit this application. The scope of protection of this application is defined by the claims. Those skilled in the art can make various modifications or equivalent substitutions to this application within its substance and scope of protection, and such modifications or equivalent substitutions should also be considered to fall within the scope of protection of this application.

Claims

1. A method for simultaneously monitoring tree sap flow and growth, characterized in that, Includes the following steps: Step S1: Simultaneously collect raw sap flow temperature difference data, trunk radial change data, soil moisture data, and micro-meteorological data of trees through multiple self-powered multimodal sensor nodes deployed in the monitoring area. The multimodal sensor nodes include a trunk sap flow probe based on the Granier thermal diffusion principle, a high-precision point-shaped trunk radial change recorder, a soil moisture sensor, and a micro-meteorological station. Step S2: Preprocess the collected raw sap flow temperature difference data, trunk radial variation data, soil moisture data and micro-meteorological data, and decouple the preprocessed raw sap flow temperature difference data from the environment to obtain sap flow density data driven by the tree's internal water state after removing micro-meteorological factors. Decompose the preprocessed trunk radial variation data into a zero-growth model to obtain the tree's net growth data. Step S3: Combine the preprocessed raw sap flow temperature difference data, trunk radial variation data, soil moisture data, and micrometeorological data with the sap flow density data and the net growth data to form a tree ecology and physiology dataset. Step S4: Based on the tree ecology and physiology dataset, construct and train a neural network model to predict sap flow density data and net growth data based on the original sap flow temperature difference data, trunk radial variation data, soil moisture data and micrometeorological data. Embed the neural network model into a digital twin constructed based on the real-world tree growth scene. The digital twin and the multimodal sensor have a data interaction channel for the digital twin to receive data collected in real time by the multimodal sensor. Step S5: Use a digital twin to collect sap flow density data and net growth data predicted by the neural network model at all times, and perform time series extrapolation based on the sap flow density data and net growth data at all times to obtain sap flow density data and net growth data in the future time series, so as to characterize the tree's water trend and growth trend.

2. The method for synchronous monitoring of tree sap flow and growth according to claim 1, characterized in that, The environmental decoupling method for the fluid flow density data includes: The initial sap flow density, obtained by Granier empirical formula after preprocessing, was used as the result variable. The trunk water deficit, obtained by decomposing the trunk radial variation data through a zero-growth model, was used as the treatment variable. Solar radiation and saturated vapor pressure difference were used as confounding variables. A dual machine learning framework was used to decouple the initial sap flow density from the environment. In the first stage of the dual machine learning framework, the processing variable and the result variable are fitted with the confusion variable respectively to obtain the residuals of the processing variable and the residuals of the result variable; The second stage of the dual machine learning framework regresses and estimates the causal effect of the processing variable on the outcome variable in the residual space to separate the portion of the initial sap flow density affected by solar radiation and saturated vapor pressure difference from the portion driven by the tree's intrinsic moisture state, thus obtaining the sap flow density data driven only by the tree's intrinsic moisture state.

3. The method for synchronous monitoring of tree sap flow and growth according to claim 2, characterized in that, The net growth data and the trunk water deficit are obtained by decomposing the data using a zero-growth model. The decomposition method includes: The preprocessed radial variation data of the tree trunk is used as a time series of tree trunk diameters, with the maximum tree trunk diameter in the time series as the reference benchmark, where: When the trunk diameter at the current moment in the trunk diameter time series is greater than or equal to the maximum trunk diameter, the excess portion is recorded as net growth. When the trunk diameter at the current moment in the time sequence is less than the maximum trunk diameter, the difference is recorded as trunk water deficit.

4. The method for synchronous monitoring of tree sap flow and growth according to claim 3, characterized in that, The neural network model includes a multimodal feature fusion layer, a two-stream feature extraction network, a cross-modal attention interaction module, and a multi-task prediction head; The preprocessed raw sap flow temperature difference data, trunk radial variation data, soil moisture data, and micrometeorological data are input into the multimodal feature fusion layer. Through embedding mapping and layer normalization processing, the data of each modality are aligned to the same feature space to obtain multimodal fusion features. The multimodal fusion features are respectively input into the first and second flow branches of the dual-flow feature extraction network. The first flow branch is composed of stacked convolutional modules and bidirectional recurrent network layers, and is used to extract fluid flow-related features from the multimodal fusion features. The second flow branch is composed of stacked convolutional modules and bidirectional recurrent network layers, and is used to extract growth-related features from the multimodal fusion features. The fluid flow feature vector output from the first flow branch and the growth feature vector output from the second flow branch are input into the cross-modal attention interaction module. The cross-modal attention interaction module uses a multi-head attention mechanism to perform attention calculation by using the fluid flow feature vector as a query and the growth feature vector as a key and value to obtain the fluid flow enhancement feature that integrates growth information. Symmetrically, the growth feature vector is used as a query and the fluid flow feature vector is used as a key and value to obtain the growth enhancement feature that integrates fluid flow information. The sap flow enhancement features and the growth enhancement features are respectively input into the first and second prediction heads of the multi-task prediction head. The predicted values ​​of sap flow density data and net growth data corresponding to the preprocessed original sap flow temperature difference data, trunk radial variation data, soil moisture data and micrometeorological data are obtained by mapping through the fully connected layer.

5. The method for synchronous monitoring of tree sap flow and growth according to claim 4, characterized in that, The loss function of the neural network model is composed of a weighted sum of three parts: liquid flow density prediction loss, net growth prediction loss, and consistency constraint loss. The expression for the loss function is: ； In the formula, For the total loss, The mean square error of fluid flow prediction. The mean square error for trunk growth prediction. For time-series consistency loss, For hyperparameters; The fluid flow density prediction loss is the mean square error between the fluid flow density prediction value output by the first prediction head and the fluid flow density data, which is used to constrain the prediction accuracy of fluid flow density. The expression for the fluid density prediction loss is: In the formula, This represents the model's predicted fluid density at time t. This corresponds to the true value of the fluid density. This represents the number of time steps. The net growth prediction loss is the mean square error between the net growth prediction value output by the second prediction head and the net growth data, which is used to constrain the prediction accuracy of net growth. The expression for the net growth prediction loss is: In the formula, This represents the model's predicted net growth at time t. This is the true value of the corresponding net growth. The consistency constraint loss is used to constrain the synergy between the predicted sap flow density and the predicted net growth in terms of their changing trends, so that the two remain in sync with the coupling relationship between tree water transport and radial growth. The expression for the consistency constraint loss is: In the formula, The Pearson correlation coefficient is used. This is a sequence of predicted fluid density values ​​at various time points. This is a net growth prediction sequence composed of the predicted net growth values ​​at each time point.

6. The method for synchronous monitoring of tree sap flow and growth according to claim 5, characterized in that, The method involves using a digital twin to collect fluid density and net growth data predicted by a neural network model at all time points, and then performing time-series extrapolation based on this data. The method includes: The predicted fluid density values ​​output by the neural network model at each historical moment are arranged in chronological order to construct a historical sequence of fluid density. Arrange the predicted net growth values ​​output by the neural network model at each historical moment in chronological order to construct a historical sequence of net growth. The Holt-Winters exponential smoothing method is used to extrapolate the trend of the historical sequence of liquid flow density, extract the horizontal component, trend component and seasonal component of liquid flow density, and recursively extrapolate based on the three components to obtain the predicted value of liquid flow density at future time. The historical sequence of net growth is accumulated day by day to construct a cumulative net growth sequence. The cumulative net growth sequence is fitted with a Logistic growth model to obtain three growth parameters: growth upper limit, growth rate, and inflection point time. Based on the fitted Logistic growth model, the cumulative net growth at future time is extrapolated, and then the net growth at future time is restored by difference. When the predicted sap flow density at future times shows a continuous downward trend and the predicted net growth at future times shows a slowing or stagnant growth trend, it is determined that the trees are at potential risk of water stress.

7. A tree sap flow and growth synchronous monitoring system, characterized in that, The system applied to the tree sap flow and growth synchronous monitoring method according to any one of claims 1 to 6, the system comprising: Multiple self-powered multimodal sensor nodes are used to synchronously collect raw sap flow temperature difference data, trunk radial change data, soil moisture data and micro-meteorological data of trees. The multimodal sensor nodes include a trunk sap flow probe based on the Granier thermal diffusion principle, a high-precision point-like trunk radial change recorder, a soil moisture sensor and a micro-meteorological station. An edge computing gateway, communicatively connected to the multimodal sensor nodes, is used to preprocess the collected raw sap flow temperature difference data, trunk radial variation data, soil moisture data, and micrometeorological data. It then decouples the preprocessed raw sap flow temperature difference data from the environment to obtain sap flow density data driven only by the tree's intrinsic water state, removing micrometeorological factors. The gateway also performs zero-growth model decomposition on the preprocessed trunk radial variation data to obtain the tree's net growth data. Finally, it combines the preprocessed raw sap flow temperature difference data, trunk radial variation data, soil moisture data, and micrometeorological data with the sap flow density data and the net growth data to form a tree ecology and physiology dataset. The digital twin platform is communicatively connected to the edge computing gateway and is used to construct and train a neural network model based on the tree ecology and physiology dataset to predict sap flow density data and net growth data based on the original sap flow temperature difference data, trunk radial change data, soil moisture data and micrometeorological data, and to embed the neural network model into the digital twin constructed based on the real-world tree growth scene. The digital twin has a data interaction channel with the multimodal sensor node, used to receive data collected in real time by the multimodal sensor node; and to collect sap flow density data and net growth data predicted by the neural network model at all times using the digital twin, and to perform time-series extrapolation based on the sap flow density data and net growth data at all times to obtain sap flow density data and net growth data in the future time series, so as to characterize the tree's water trend and growth trend, and at the same time generate early warning information based on potential risks and push it to the user terminal.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by the processor, implement a method for synchronous monitoring of tree sap flow and growth as described in any one of claims 1 to 6.

Images