A plant water utilization strategy analysis method and system, and a storage medium

Abstract

The present application relates to the technical field of wisdom agriculture, in particular to a plant water utilization strategy analysis method and system and a storage medium, multi-modal data of environment time series, plant physiological time series and hyperspectral images are collected and feature-level fusion is performed; a first prediction model based on an attention mechanism is constructed, transpiration rate prediction values are output and environmental factor contribution weights are extracted; a second prediction model based on ecological consistency constraints is constructed, soil volume water content and matric potential are predicted to complete the missing of environmental data; through alternating iterative collaborative coupling training, the two models promote each other, and the robustness of transpiration prediction under the condition of data missing is significantly enhanced; based on the transpiration rate time series clustering, the water utilization strategy type is divided; the influence degree of each environmental factor on different strategy types is quantified by using the attention weight. The present application can alleviate the precision defects caused by the missing of key data in wetland environment, and enhance the generalization ability and ecological consistency of the model.

Description

Technical Field

[0001] This invention relates to the field of smart agriculture and ecological monitoring technology, specifically to a method, system, and storage medium for analyzing plant water use strategies. Background Technology

[0002] Plant water use strategies are a cutting-edge and hot topic in ecology, agronomy, and global change ecology. They reveal how plants coordinate the trade-off between carbon assimilation and water loss in specific habitats to adapt to environmental changes. Accurately analyzing plant water use strategies under different environmental conditions has significant theoretical and applied value for assessing ecosystem responses to climate change, guiding precision irrigation, and optimizing crop water management.

[0003] Traditional plant water use research mainly relies on process-based physiological and ecological models (such as the Jarvis model and the Ball-Berry model) or empirical statistical methods, as well as stable isotope tracing technology, Granier thermal diffusion probes for trunk sap flow monitoring, and tree growth meter observations.

[0004] With the development of the Internet of Things (IoT) and remote sensing technologies, we can acquire massive amounts of environmental and plant physiological time-series data, as well as hyperspectral image data. These vast amounts of data can be used to train artificial intelligence models to predict plant water usage. However, in actual field monitoring, the completeness of environmental data is often difficult to guarantee. In particular, the two key parameters, soil volumetric water content and matrix potential, are highly susceptible to loss due to factors such as flooding, animal activity, and electrode corrosion in complex environments like wetlands. These factors can cause gaps, drift, or even long-term data loss, making them the components with the highest loss rate in environmental data. In contrast, meteorological data such as air temperature, relative humidity, and photosynthetically active radiation are continuously collected from ground weather stations, resulting in a lower probability of loss. Soil volumetric water content and matrix potential, as core variables describing the soil moisture state in the rhizosphere, directly determine the availability of water absorption by plants and are a crucial link between environmental conditions and plant physiological responses. The loss of these two data points leads to severely incomplete environmental information, making it impossible for existing artificial intelligence models to accurately perceive soil water supply status, thus significantly reducing the accuracy of plant transpiration rate predictions.

[0005] Therefore, existing technologies are unable to solve the problem of information link breakage caused by the lack of key soil parameters, which directly affects the accurate analysis and classification of plant water use strategies. Summary of the Invention

[0006] The purpose of this invention is to provide a method, system, and storage medium for analyzing plant water use strategies, in order to solve the problem of information link breakage caused by the lack of key soil parameters in the prior art, which directly affects the technical problem of accurately analyzing and classifying plant water use strategies.

[0007] To solve the above-mentioned technical problems, the present invention specifically provides the following technical solution: A method for analyzing plant water use strategies includes the following steps: Step S1: Collect multimodal data of the target plant, perform feature-level fusion on the multimodal data to obtain multimodal fusion features for characterizing the interaction state between the plant and the environment. The multimodal data includes environmental time series data, plant physiological time series data and hyperspectral image time series data. Step S2: Construct and train a first prediction model based on an attention mechanism. The first prediction model is used to predict the transpiration rate of the target plant at future times based on the multimodal fusion features. Step S3: Construct and train a second prediction model based on ecological consistency constraints. The second prediction model is used to predict the soil volumetric water content and matrix potential of the target plant in time and space according to the multimodal fusion features and the spatial coordinates of the target plant, so as to complete the missing parts of soil volumetric water content and matrix potential in the environmental time series data. Step S4: Co-couple the first prediction model and the second prediction model for training to enhance the robustness of the first prediction model in predicting plant transpiration rate under conditions of missing environmental time series data. Step S5: The trained first prediction model predicts the transpiration rate of all target plants at future times. A clustering algorithm is used to cluster all target plants based on the plant transpiration rate at the future times, and each target plant is divided into different water use strategy types. Step S6: Apply the interpretability analysis algorithm to extract the weights of the attention mechanism in the first prediction model, obtain the contribution weights of each environmental factor included in the multimodal fusion feature to the plant transpiration rate at different sampling times, calculate the contribution value of each environmental factor to the output of the first prediction model based on the contribution weights, and quantify the degree of influence of each environmental factor on the plant transpiration water consumption behavior in different water use strategy types.

[0008] As a preferred embodiment of the present invention, the environmental time series data is environmental factor data collected continuously over time, including lake water level elevation, groundwater depth, soil volumetric water content, matrix potential, air temperature, relative humidity, photosynthetically active radiation, precipitation, and saturated vapor pressure difference. The plant physiological time series data are plant physiological response data collected continuously over time. The plant physiological response data includes trunk sap flow density continuously monitored by the Granier thermal diffusion probe, trunk water deficit extracted from trunk radial changes continuously monitored by a high-precision growth instrument, leaf water potential at dawn and noon measured according to the hydrological cycle, and leaf stable carbon isotopes. The hyperspectral image time series data is multi-temporal hyperspectral remote sensing data repeatedly collected at preset time intervals. The multi-temporal hyperspectral remote sensing data includes the spatial distribution of canopy temperature, normalized vegetation index, and photochemical reflectance index.

[0009] As a preferred embodiment of the present invention, the fusion method of the multimodal fusion features includes: A feature fusion network is constructed for feature-level fusion of multimodal data. The feature fusion network consists of a dual encoder-decoder structure with a gated fusion mechanism, wherein the first encoder is used to extract the spatial-spectral features of the hyperspectral image time series data. The second encoder is used to extract the temporal features of the environmental time series data and the plant physiological time series data. ; The spatial-spectral features are dynamically calculated using a gating fusion mechanism. and the aforementioned time-series features The fusion weights α are used to generate multimodal fusion features F, where: ; ; In the formula, σ is the sigmoid activation function. and Here are the learnable weights and bias parameters in the feature fusion network, and [:] represents the feature concatenation operation.

[0010] As a preferred embodiment of the present invention, the first prediction model is an LSTM-Transformer hybrid model; The multimodal fusion features are input into the LSTM layer of the LSTM-Transformer hybrid model. The LSTM layer performs temporal encoding on the multimodal fusion features, extracts the short-term temporal dependencies contained therein, and outputs latent features with short-term temporal memory. The latent features are input into the Transformer encoder in the LSTM-Transformer hybrid model. The multi-head self-attention mechanism in the Transformer encoder is used to capture the long-term dependencies and global interaction features between variables in the multimodal fusion features. The output of the Transformer encoder is mapped through a fully connected layer to obtain the predicted value of the plant transpiration rate at future times for the target plant; Specifically, the attention weights of the multi-head self-attention mechanism in the Transformer encoder are aggregated and normalized to obtain the contribution weights of various environmental factors to the transpiration rate of the plant at different sampling times.

[0011] As a preferred embodiment of the present invention, the second prediction model adopts a fully connected neural network structure; The spatial coordinates of the target plant and the multimodal fusion features are input into the input layer of a fully connected neural network structure. The input layer concatenates the spatial coordinates of the target plant and the multimodal fusion features to obtain a fusion vector. The fused vector is input into the hidden layer of a fully connected neural network structure, and the hidden layer performs a nonlinear transformation on the fused vector to extract the spatiotemporal correlation features contained therein. The spatiotemporal correlation features are mapped through the output layer to obtain the soil volumetric water content and matrix potential of the target plant in the spatiotemporal space.

[0012] In a preferred embodiment of the present invention, the training loss of the second prediction model is composed of a weighted sum of two parts: data fitting loss and ecological consistency constraint loss. The formula for calculating the training loss is as follows: In the formula, For data fitting loss, Loss due to ecological consistency constraints. For hyperparameters; The data fitting loss is the mean square error between the predicted soil volumetric water content and matrix potential output by the second prediction model and the corresponding measured values. The formula for calculating the data fitting loss is as follows: In the formula, N is the total number of measured samples. For the first The measured volumetric water content of the soil in each sample For the first Predict soil volumetric water content for each sample For the first Measured matrix potential of each sample For the first Predicting matrix potential for each sample The ecological consistency constraint loss is the residual norm of the water flow equation in the soil-plant-atmosphere continuum, and the formula for calculating the ecological consistency constraint loss is as follows: In the formula For soil hydraulic conductivity, The rate of change of soil volumetric water content over time. The gradient of the matrix potential along the vertical depth is given by the soil hydraulic conductivity, which is a function of the soil volumetric water content.

[0013] As a preferred embodiment of the present invention, the collaborative coupling training of the first prediction model and the second prediction model includes: Step S401: Using an alternating optimization strategy, in each training round, first fix the parameters of the first prediction model, and use the second prediction model to generate predicted values ​​at the missing time points of soil volumetric water content and matrix potential in the environmental time series data to complete the environmental time series data; Step S402: Fix the parameters of the second prediction model, reconstruct new multimodal data using the completed environmental time series data, plant physiological time series data and hyperspectral image time series data, extract new multimodal fusion features from the new multimodal data using a feature fusion network, train the first prediction model based on the new multimodal fusion features, and calculate the prediction error of the first prediction model. Step S403: The prediction error of the first prediction model is used as an additional loss term and backpropagated to the second prediction model to update the parameters of the second prediction model so that the soil volume water content and matrix potential data of the missing segment generated by the second prediction model are more conducive to improving the prediction accuracy of the first prediction model. Step S404: Repeat steps S401 to S403 until the loss functions of the first prediction model and the second prediction model converge, thus obtaining the trained first prediction model and the second prediction model.

[0014] As a preferred embodiment of the present invention, the method of clustering all target plants based on the plant transpiration rate at the future time using a clustering algorithm includes: The transpiration rates of all target plants output by the first prediction model after training are used to form the time series feature vector of the transpiration rate of each target plant at future times. Set the number of clusters K, and randomly select K transpiration rate time series feature vectors of target plants as the initial cluster centers; Calculate the Euclidean distance between the transpiration rate time series feature vector of each target plant and each cluster center, and assign each target plant to the category of the nearest cluster center; Update each cluster center based on the allocation results, and repeat the allocation and update steps until the cluster centers no longer change or the preset number of iterations is reached, so that each target plant is classified into different water use strategy types.

[0015] To address the aforementioned technical problems, the present invention further provides the following technical solution: A system for analyzing plant water use strategies, the system comprising: The data acquisition and fusion module is used to acquire multimodal data of the target plant, perform feature-level fusion on the multimodal data, and obtain multimodal fusion features to characterize the interaction state between the plant and the environment. The multimodal data includes environmental time series data, plant physiological time series data, and hyperspectral image time series data. The model building module is used to build and train a first prediction model based on an attention mechanism. The first prediction model is used to predict the transpiration rate of the target plant at future times based on the multimodal fusion features. The module also builds and trains a second prediction model based on ecological consistency constraints. The second prediction model is used to predict the soil volumetric water content and matrix potential of the target plant in the space-time of the target plant based on the multimodal fusion features and the spatial coordinates of the target plant, so as to complete the missing parts of soil volumetric water content and matrix potential in the environmental time series data. The module also performs collaborative coupling training of the first prediction model and the second prediction model to enhance the robustness of the first prediction model in predicting the transpiration rate of the plant under the condition of missing environmental time series data. The water use strategy classification module is used to predict the transpiration rate of all target plants at future times by the first prediction model that has been trained. The clustering algorithm is used to cluster all target plants based on the plant transpiration rate at the future time, and classify each target plant into different water use strategy types. The contribution quantification module is used to apply an interpretability analysis algorithm to extract the weights of the attention mechanism in the first prediction model, obtain the contribution weights of each environmental factor included in the multimodal fusion feature to the plant transpiration rate at different sampling times, calculate the contribution value of each environmental factor to the output of the first prediction model based on the contribution weights, and quantify the degree of influence of each environmental factor on the plant transpiration water consumption behavior in different water use strategy types.

[0016] 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 analyzing plant water use strategies.

[0017] Compared with the prior art, the present invention has the following advantages: This invention integrates multimodal data such as environmental time series, plant physiological time series, and hyperspectral images. It uses soil moisture movement equations to ecologically complete soil volumetric water content and matrix potential, which are easily missing in monitoring. Then, through collaborative coupling training, the completed data is more conducive to improving the prediction accuracy of transpiration rate, thereby significantly enhancing the prediction robustness under the condition of missing data. Finally, based on the transpiration rate time series, plants are clustered to automatically classify water use strategy types and improve the degree of automation. Attached Figure Description

[0018] 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.

[0019] Figure 1 This is a flowchart illustrating a method for analyzing plant water use strategies, provided in an embodiment of the present invention.

[0020] Figure 2 This is a schematic diagram illustrating the installation of the Granier thermal diffusion probe and high-precision growth instrument provided in an embodiment of the present invention.

[0021] Figure 3 This is a graph showing the raw temperature difference data of sap flow density in tree trunks collected by a Granier thermal diffusion probe, provided for an embodiment of the present invention.

[0022] Figure 4 The time series curves of trunk radius change, water deficit and radial growth provided for embodiments of the present invention.

[0023] Figure 5 This is a schematic diagram of sensor deployment in different water level gradient zones of a typical wetland forest, provided as an embodiment of the present invention.

[0024] Figure 6 This is a structural diagram of a plant water use strategy analysis system provided in an embodiment of the present invention. Detailed Implementation

[0025] 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.

[0026] like Figure 1As shown, this invention provides a method for analyzing plant water use strategies. First, environmental time-series data, plant physiological time-series data, and hyperspectral image time-series data of the target plant are collected. Multimodal fusion features representing the interaction between the plant and the environment are generated through feature-level fusion. Then, a first prediction model based on an attention mechanism is constructed, using the multimodal fusion features as input, to predict the transpiration rate of the target plant at future times.

[0027] To address the issue of missing soil volumetric water content and matrix potential in field monitoring, a second prediction model based on ecological consistency constraints was constructed. By combining the spatial coordinates and fusion characteristics of the target plant, ecological consistency was achieved in these two missing parameters.

[0028] Based on this, the two models are trained in a synergistic coupling with alternating optimization. Backpropagation of prediction errors guides the completion of data, which is more conducive to transpiration rate prediction and enhances the robustness of the first prediction model under the condition of missing environmental data. After the prediction is completed, clustering is performed based on the transpiration rate time series of each target plant to classify them into different water use strategy types. Finally, by extracting the attention weight of the first prediction model, the influence of each environmental factor on transpiration water consumption behavior in different strategy types is quantified, generating interpretable analysis results with clear ecological significance.

[0029] Specifically, the first step involves collecting multimodal data of the target plant and performing feature-level fusion on the multimodal data to obtain multimodal fusion features that characterize the interaction between the plant and the environment. The multimodal data includes environmental time-series data, plant physiological time-series data, and hyperspectral image time-series data. By organically integrating these three types of heterogeneous data, the limitations of traditional single data sources in comprehensively reflecting plant-environment interactions are overcome.

[0030] Environmental time series data are environmental factor data collected continuously over time. These environmental factor data include lake water level elevation, groundwater depth, soil volumetric water content, matrix potential, air temperature, relative humidity, photosynthetically active radiation, precipitation, and saturated vapor pressure difference.

[0031] Soil volumetric water content and matrix potential are core variables describing the soil moisture state in the rhizosphere, directly determining the availability of water absorption by plants. However, in actual field monitoring, these two parameters are prone to data loss due to factors such as flooding, animal activity, and electrode aging caused by long-term underground sensor installations, making them the components with the highest missing data rate in environmental data. Lake water level elevation and groundwater depth reflect the unique hydrological fluctuation characteristics in wetland ecosystems and are key driving factors influencing plant water sources and rhizosphere hypoxia stress. Air temperature, relative humidity, photosynthetically active radiation, and saturated vapor pressure difference constitute the main characteristics of atmospheric evaporation demand, while precipitation determines the flux of water input. These meteorological factors are obtained through continuous automatic monitoring, resulting in high data integrity and providing a reliable data foundation for analyzing the atmospheric driving forces of plant transpiration water consumption.

[0032] Plant physiological time series data are plant physiological response data collected continuously over time. The plant physiological response data include trunk sap flow density continuously monitored by Granier thermal diffusion probe, trunk water deficit extracted from trunk radial changes continuously monitored by high-precision growth instrument, leaf water potential at dawn and noon measured according to hydrological cycle, and leaf stable carbon isotopes.

[0033] Among these parameters, trunk sap flow density is a core indicator for quantifying plant transpiration water consumption rates. Based on the Granier heat diffusion principle, it is calculated using empirical formulas by measuring the temperature difference between the heated probe and the reference probe as heat dissipates through sap flow. Trunk water deficit extracts reversible fluctuations in tree water content from high-precision radial variations, reflecting the accumulation and recovery process of water deficit on diurnal and seasonal scales. Leaf morning and noon water potentials characterize the maximum water recovery state and the degree of water stress during peak transpiration conditions, respectively. Leaf stable carbon isotopes integrate water use efficiency information over longer timescales. These physiological parameters characterize plant responses to water conditions at different timescales and tissue levels, providing rich physiological information for subsequent transpiration rate prediction and water use strategy classification.

[0034] The hyperspectral image time series data consists of multi-temporal hyperspectral remote sensing data repeatedly collected at preset time intervals. The multi-temporal hyperspectral remote sensing data includes the spatial distribution of canopy temperature, normalized vegetation index, and photochemical reflectance index.

[0035] The hyperspectral image time series data consists of multi-temporal hyperspectral remote sensing data repeatedly acquired at preset time intervals, including the spatial distribution of canopy temperature, normalized difference vegetation index (NDVI), and photochemical reflectance index. Canopy temperature is obtained through inversion from the thermal infrared band and is closely related to the plant transpiration cooling effect; the NDVI reflects canopy green biomass and photosynthetic capacity; and the photochemical reflectance index is sensitive to changes in leaf photosynthetic pigment composition and light energy use efficiency. Hyperspectral data provides continuous spatial information on the physiological state of plant canopies in a planar, non-contact manner, compensating for the spatial representativeness limitations of single-point ground observations.

[0036] This invention performs feature-level fusion of the above three types of multimodal data. By using a gated fusion mechanism, it dynamically adjusts the fusion weights of spatial-spectral features and environmental-physiological temporal features of hyperspectral images. This enables the model to adaptively decide whether to rely on image spectral features or temporal features based on the content of the input data, thereby generating multimodal fusion features with richer information and stronger representation capabilities. This provides high-quality unified input for the subsequent evaporation rate prediction of the first prediction model and the ecological completion of the second prediction model.

[0037] The fusion method for multimodal fusion features is as follows: A feature fusion network is constructed for feature-level fusion of multimodal data. The feature fusion network consists of a dual encoder-decoder structure with a gated fusion mechanism, wherein the first encoder is used to extract the spatial-spectral features of the hyperspectral image time series data. The second encoder is used to extract the temporal features of environmental time-series data and plant physiological time-series data. ; Dynamic computation of spatial-spectral features using a gated fusion mechanism and time series characteristics The fusion weights α are used to generate multimodal fusion features F, where: ; ; In the formula, σ is the sigmoid activation function. and Here are the learnable weights and bias parameters in the feature fusion network, and [:] represents the feature concatenation operation.

[0038] The multimodal fusion feature generation of this invention employs a dual-encoder structure based on a gated fusion mechanism. The first encoder extracts spatial-spectral features from hyperspectral image time-series data, while the second encoder extracts temporal features from environmental and plant physiological time-series data. The two features are input into a gated fusion unit, and after concatenation, linear transformation, and mapping using a sigmoid activation function, the fusion weight α is dynamically calculated, and then... A weighted summation is performed to generate a multimodal fusion feature F. This gating mechanism enables the fusion weights to adaptively adjust according to the content of the input data: when the time-series data contains significant trends, the network automatically increases the weights of the time-series features to enhance the perception of key dynamics; when the hyperspectral image captures canopy anomalies related to water stress, the network increases the weights of the image features to utilize the discriminative advantages of spatial-spectral information. Compared to fixed weights or simple concatenation methods, this adaptive fusion strategy effectively addresses the differences in information density and sampling frequency between multi-source heterogeneous data, providing a richer and more unified input foundation for subsequent models.

[0039] Specifically, the second step is to build and train a first prediction model based on an attention mechanism. This first prediction model is used to predict the transpiration rate of the target plant at future times based on multimodal fusion features.

[0040] The first prediction model is an LSTM-Transformer hybrid model; Multimodal fusion features are input into the LSTM layer of the LSTM-Transformer hybrid model. The LSTM layer performs temporal encoding on the multimodal fusion features, extracting the short-term temporal dependencies contained therein, and outputting latent features with short-term temporal memory. LSTM can effectively capture the rapid response of plant transpiration to instantaneous changes in environmental factors, such as the instantaneous increase in transpiration caused by a sudden increase in light intensity or the decrease in transpiration caused by stomatal closure in the afternoon. These short-term dynamics are crucial for accurately predicting the transpiration rate at future moments.

[0041] The latent features are input into the Transformer encoder in the LSTM-Transformer hybrid model. The multi-head self-attention mechanism in the Transformer encoder is used to capture the long-term dependencies and global interaction features between variables in the multimodal fusion features.

[0042] Unlike LSTM, which can only propagate information unidirectionally along the time axis, the self-attention mechanism allows the model to directly focus on the correlation between any two time steps in the sequence. This effectively models the cumulative response of transpiration rate to environmental changes over a longer timescale, such as the gradual suppression of transpiration due to prolonged drought or the slow recovery process after water levels rise. Furthermore, the multi-head self-attention mechanism learns diverse interaction patterns from different subspaces through multiple parallel attention heads, further enhancing the model's ability to express complex coupling relationships.

[0043] The output of the Transformer encoder is mapped through a fully connected layer to obtain the predicted transpiration rate of the target plant at future time points. Plant transpiration rate, as a quantitative indicator of plant water consumption, directly reflects the rate at which the plant uses water under current environmental conditions and is the core basis for subsequent cluster analysis and strategy classification.

[0044] Furthermore, during training, the attention weights of the multi-head self-attention mechanism in the Transformer encoder were aggregated and normalized to obtain the contribution weights of various environmental factors to the plant transpiration rate at different sampling times. This operation provides a quantitative basis for subsequent interpretability analysis, enabling the model not only to provide predicted values ​​of transpiration rate but also to reveal the key environmental factors affecting the predicted results. This transforms the black-box weights of the attention mechanism into a metric of environmental factor contributions with clear ecological significance.

[0045] Specifically, the third step involves constructing and training a second prediction model based on ecological consistency constraints. This second prediction model is used to predict the soil volumetric water content and matrix potential of the target plant in the time and space based on multimodal fusion features and the spatial coordinates of the target plant, so as to complete the missing parts of soil volumetric water content and matrix potential in the environmental time series data.

[0046] The second prediction model uses a fully connected neural network structure; The spatial coordinates and multimodal fusion features of the target plant are input into the input layer of a fully connected neural network structure. The input layer then concatenates the spatial coordinates and multimodal fusion features of the target plant to obtain a fusion vector. The fused vector is input into the hidden layer of a fully connected neural network structure, and the hidden layer performs a nonlinear transformation on the fused vector to extract the spatiotemporal correlation features contained therein. The spatiotemporal correlation features are mapped through the output layer to obtain the soil volumetric water content and matrix potential of the target plant in the spatiotemporal space.

[0047] This invention constructs a second predictive model to supplement the missing portions of soil volumetric water content and matrix potential in environmental time series data. In actual field monitoring, these two parameters are highly susceptible to interruptions or omissions due to factors such as flooding, animal activity, and electrode aging caused by the long-term underground burial of sensors, making them the components with the highest missing rates in environmental data. However, soil volumetric water content and matrix potential are precisely the core ecological variables describing soil moisture transport, and the quantitative relationship between them follows the Richards equation. This characteristic of being both easily lost and having ecological patterns makes it possible to supplement these two data by introducing ecological constraints.

[0048] In terms of model structure, the second prediction model employs a fully connected neural network. The spatial coordinates of the target plant (i.e., the vertical depth of the sensor installation) and the multimodal fusion features generated in the first step are input into the input layer, and then concatenated to form a fusion vector. This fusion vector contains both environmental and plant physiological information extracted from the multimodal data and embeds spatial location identifiers, enabling the model to distinguish the ecological environment differences between soil layers at different depths. The fusion vector is then passed through a hidden layer consisting of four to six fully connected layers, undergoing a layer-by-layer nonlinear transformation to extract the spatiotemporal correlation features. Finally, the output layer maps the predicted values ​​of soil volumetric water content and matrix potential at the target plant's location in the specified spatiotemporal space.

[0049] The training loss of the second prediction model consists of a weighted sum of two parts: data fitting loss and ecological consistency constraint loss. The formula for calculating the training loss is as follows: In the formula, For data fitting loss, Loss due to ecological consistency constraints. For hyperparameters; The data fitting loss is the mean square error between the predicted soil volumetric water content and matrix potential output by the second prediction model and the corresponding measured values. The formula for calculating the data fitting loss is: In the formula, N is the total number of measured samples. For the first The measured volumetric water content of the soil in each sample For the first Predict soil volumetric water content for each sample For the first Measured matrix potential of each sample For the first Predicting matrix potential for each sample The ecological consistency constraint loss is the residual norm of the water flow equation in the soil-plant-atmosphere continuum. The formula for calculating the ecological consistency constraint loss is: In the formula For soil hydraulic conductivity, The rate of change of soil volumetric water content over time. The gradient of the matrix potential along the vertical depth is given by the soil hydraulic conductivity, which is a function of the soil volumetric water content.

[0050] Furthermore, in terms of the training strategy, the total loss of the second prediction model in this invention consists of a weighted sum of two parts: data fitting loss and ecological consistency constraint loss. The data fitting loss adopts the form of mean squared error, which calculates the sum of the squares of the differences between the predicted values ​​of soil volumetric water content and matrix potential output by the model and their corresponding measured values. Then, the average is taken over all measured samples to ensure that the model's predicted values ​​can approximate the actual observed values ​​at spatiotemporal points where measured data is available.

[0051] However, if only data fitting loss is used, the model will rely entirely on data-driven interpolation in areas with missing data, easily leading to predictions that violate ecological principles. Therefore, an ecological consistency constraint loss is introduced. The core of this loss function is to incorporate the residuals of the Richards equation, which describes one-dimensional vertical water flow in the soil-plant-atmosphere continuum, as a penalty term. This equation describes the quantitative relationship between the rate of change of soil volumetric water content over time, matrix potential along the vertical depth gradient, and soil hydraulic conductivity, representing the fundamental ecological laws governing soil water transport. By minimizing the residual norm of this equation, even in spatiotemporal points lacking measured data, the model's output soil volumetric water content and matrix potential must satisfy the ecological constraints stipulated by the Richards equation, thus ensuring the ecological consistency of the completed values.

[0052] It is worth noting that the partial derivatives in the ecological consistency constraint loss (including the rate of change of soil volumetric water content with time, the gradient of matrix potential along vertical depth, etc.) are not approximated by numerical difference, but are precisely calculated using the automatic differentiation mechanism of the neural network framework. This technique not only avoids the discrete errors introduced by traditional numerical methods, but also enables ecological constraints to directly participate in model training in an end-to-end manner, significantly improving the accuracy of ecological constraint application and training efficiency.

[0053] Therefore, even with missing measured data, the second prediction model can still generate complete values ​​that conform to the laws of soil moisture transport, providing complete and ecologically self-consistent environmental data input support for the first prediction model, fundamentally solving the problem of information link breakage caused by the lack of key soil parameters.

[0054] Specifically, in the fourth step, the first prediction model and the second prediction model are trained in a synergistic coupling manner to enhance the robustness of the first prediction model in predicting plant transpiration rate under conditions of missing environmental time series data.

[0055] The collaborative training of the first and second prediction models includes: Step S401: Using an alternating optimization strategy, in each training round, first fix the parameters of the first prediction model, and use the second prediction model to generate predicted values ​​at the missing time points of soil volumetric water content and matrix potential in the environmental time series data to complete the environmental time series data. Step S402: Fix the parameters of the second prediction model, reconstruct new multimodal data using the completed environmental time series data, plant physiological time series data and hyperspectral image time series data, extract new multimodal fusion features from the new multimodal data using a feature fusion network, train the first prediction model based on the new multimodal fusion features, and calculate the prediction error of the first prediction model. Step S403: The prediction error of the first prediction model is used as an additional loss term and backpropagated to the second prediction model to update the parameters of the second prediction model so that the soil volume water content and matrix potential data of the missing segment generated by the second prediction model are more conducive to improving the prediction accuracy of the first prediction model. Step S404: Repeat steps S401 to S403 until the loss functions of the first prediction model and the second prediction model converge, and obtain the trained first prediction model and the second prediction model.

[0056] Traditional data completion and prediction tasks are usually performed independently: missing data is completed first, and then a prediction model is trained based on the completed data. This two-step approach has significant drawbacks: the second prediction model, when generating completed values, only aims to fit the measured data and does not consider the impact of the completed results on the downstream prediction task. This may result in the completed data being numerically reasonable but failing to effectively improve or even harming the prediction accuracy of the first prediction model. This invention, through a collaborative coupling training mechanism, deeply binds the optimization processes of the two models, making data completion no longer an independent step unrelated to the prediction task, but an organically linked component serving the overall goal of evapotranspiration rate prediction.

[0057] Specifically, the collaborative coupling training employs an alternating optimization strategy. In each training round, the parameters of the first prediction model are fixed, and the second prediction model is used to generate predicted values ​​for the missing time points of soil volumetric water content and matrix potential in the environmental time series data, thus completing the environmental time series data. At this point, the second prediction model, relying on the ecological constraints of the Richards equation embedded in its training, can generate complete values ​​that conform to the laws of soil moisture transport, ensuring the ecological rationality of the completed data.

[0058] Next, with the parameters of the second prediction model fixed, the completed environmental time-series data, plant physiological time-series data, and hyperspectral image time-series data are reconstructed into new multimodal data. After extracting the new multimodal fusion features through a feature fusion network, this data is input into the first prediction model for training, and its prediction error is calculated. Since the environmental data has been fully restored, the first prediction model can predict transpiration rates based on the complete input information, thus fully perceiving the impact of soil water supply on plant transpiration.

[0059] The key to collaborative training lies in the third step: using the prediction error of the first prediction model as an additional loss term, backpropagating it to the second prediction model to update its parameters. Essentially, this operation feeds back the requirements of the prediction task to the completion model, enabling the second prediction model, when generating completion data, to not only strive for a close match with measured values ​​but also to focus on whether the completion results can help the first prediction model improve the accuracy of evaporation rate prediction. Through this "task-oriented" error backpropagation, the second prediction model gradually learns to generate optimal completion values ​​for missing segments that both conform to ecological principles and contribute to downstream prediction tasks.

[0060] Repeat the above steps until the loss functions of both models converge. Through this iterative and mutually reinforcing collaborative training, the first and second prediction models form a tightly coupled whole: the second prediction model provides the first prediction model with complete and ecologically consistent environmental input, and the prediction error of the first prediction model guides the second prediction model to optimize the completion strategy. The two improve together in collaboration, and ultimately maintain high-precision prediction of plant transpiration rate even under conditions of missing environmental data.

[0061] Specifically, in the fifth step, the trained first prediction model predicts the transpiration rate of all target plants at future times, and a clustering algorithm is used to cluster all target plants based on the transpiration rate at future times, classifying each target plant into different water use strategy types.

[0062] Methods that use clustering algorithms to cluster all target plants based on future plant transpiration rates include: The transpiration rates of all target plants output by the first prediction model after training are used to form the time series feature vector of transpiration rate of each target plant at future times. Set the number of clusters K, and randomly select K transpiration rate time series feature vectors of target plants as the initial cluster centers; Calculate the Euclidean distance between the transpiration rate time series feature vector of each target plant and each cluster center, and assign each target plant to the category of the nearest cluster center; Update each cluster center based on the allocation results, and repeat the allocation and update steps until the cluster centers no longer change or the preset number of iterations is reached, so that each target plant is classified into different water use strategy types.

[0063] The first prediction model, after training, outputs predicted transpiration rates for all target plants at future times, forming a time-series feature vector of transpiration rate for each target plant. Unlike grouping plants directly using instantaneous transpiration rate values ​​at a single moment, the transpiration rate time series encodes the complete response pattern of plants to environmental changes, such as the rate of transpiration decline during drought and the recovery rate after precipitation. These time-series patterns are key to distinguishing different water use strategies.

[0064] The K-means clustering algorithm used in this invention is simple to calculate and converges quickly. It can automatically classify target plants into different water use strategy types based on the similarity of dynamic patterns of transpiration rate. The classification results have a clear ecological correspondence with the water level gradient zone where the target plant is located, realizing an effective transformation from a single physiological indicator to a water use strategy type.

[0065] The sixth step involves applying an interpretability analysis algorithm to extract the weights of the attention mechanism in the first prediction model, obtaining the contribution weights of various environmental factors included in the multimodal fusion features to the plant transpiration rate at different sampling times, and calculating the contribution value of each environmental factor to the output of the first prediction model based on the contribution weights, thereby quantifying the degree of influence of each environmental factor on the plant transpiration water consumption behavior in different water use strategy types.

[0066] This invention integrates multimodal data such as environmental time series, plant physiological time series, and hyperspectral images. It uses soil moisture movement equations to ecologically complete soil volumetric water content and matrix potential, which are easily missing in monitoring. Then, through collaborative coupling training, the completed data is more conducive to improving the prediction accuracy of transpiration rate, thereby significantly enhancing the prediction robustness under the condition of missing data. Finally, based on the transpiration rate time series, plants are clustered to automatically classify water use strategy types and improve the degree of automation.

[0067] In some specific implementations, after acquiring the aforementioned multimodal data, feature-level fusion is performed based on a gated fusion mechanism. Two independent encoders are constructed: the first encoder (based on the ConvNeXtV2 backbone network) is used to process hyperspectral image data and extract spatial-spectral features. The second encoder (based on LSTM or TCN) is used to process environmental and physiological time series data and extract temporal features. .

[0068] This invention employs an LSTM-Transformer hybrid model, first fusing the feature sequence F... t In one or more LSTM layers, the LSTM unit captures short-term temporal dependencies through input gates, forget gates, and output gates, and outputs a new feature sequence H. t Then H tThe input to the Transformer encoder, and the calculation process of its multi-head self-attention mechanism, are as follows: ; Wherein, Q, K, and V are respectively derived from H t It is obtained through three different linear transformations. The Transformer encoder captures the long-range dependencies between any two time steps in the sequence.

[0069] The training loss of the LSTM-Transformer hybrid model includes the mean squared error between the predicted and measured values ​​of plant transpiration rate, where the measured value of plant transpiration rate (TR) is determined by the trunk sap flow density. Multiply by sapwood area get.

[0070] In some specific implementations, the trunk sap flow density Calculated using Granier's empirical formula, In the formula, This represents the maximum temperature difference when the liquid flow is zero. For instantaneous temperature difference; Tree trunk water deficit Obtained through decomposition using the zero-growth model, In the formula Let t be the measured diameter of the tree trunk. The measured trunk diameter is the actual diameter of the tree for all time periods before time t.

[0071] In practice, multimodal data is collected through a combination of drones, IoT sensor networks, and manual sampling. Specifically: A small weather station (such as VantagePro2) was installed at the center of the study plot to provide meteorological data with a 10-minute resolution, including air temperature, relative humidity, photosynthetically active radiation, wind speed, precipitation, and to calculate the saturated vapor pressure difference (VPD). At the center of the intermediate quadrats in the three water level gradient zones (G1, G2, G3), three layers of soil volumetric water content sensors were vertically installed (monitoring depths of 10cm, 30cm, and 50cm), recording data every 10 minutes. Pressure-type self-recording water level gauges (CTD-Diver) were deployed along the elevation near the lakeside forest (G1) to automatically record water level elevation and flooding duration every 10 minutes. Groundwater observation wells (2.5m deep) were drilled in each gradient zone, and miniature automatic groundwater level recorders (HOBOU20-001-04) were installed to collect groundwater depth data every 10 minutes.

[0072] The Granier heat diffusion probe consists of two identical induction probes, each containing an ultrafine copper-nickel alloy thermocouple. The upper probe is heated with a constant current, while the lower probe is not heated. The temperature difference between the two probes is measured by a data acquisition unit. The movement of sap within the tree trunk carries away the heat from the probes, thus reducing the temperature difference between the two probes.

[0073] like Figure 2 As shown, when installing the probe at a height of 1.3 meters above the target tree, a hole with a diameter of 2.0 mm and a depth of 22 mm was first drilled using an electric drill. The probe's sensing part was then evenly coated with insulating and thermally conductive silicone grease before being carefully inserted. The area around the probe insertion point was then sealed with sealant to prevent heat loss and rainwater infiltration. The probe was connected to the CR1000X data acquisition unit, recording and storing data every 10 minutes. Figure 3 This paper presents an example of raw temperature difference data of sap flow density in tree trunks collected by the Granier thermal diffusion probe (recorded every 10 minutes), with the temperature difference values ​​at 24 consecutive time points used for subsequent calculation of sap flow density.

[0074] At the same time, such as Figure 2 As shown, a high-precision growth meter (such as the Milang TR4-10MM spring self-resetting displacement sensor) was installed 10-20cm above the height of the sample tree at breast height, and the micron-level changes in the trunk diameter were continuously recorded at 10-minute intervals using a CR1000X data acquisition device.

[0075] Figure 4 The paper shows the cumulative changes in daily stem radius, daily tree water deficit, and monthly radial growth for two typical tree species during 2020-2021. Figure 4 In this model, trunk radial variation (SRV) is decomposed into irreversible radial growth (GRO, histogram) and reversible trunk water deficit (TWD, line graph). The zero-growth model works by defining radial growth as a point in time when the trunk diameter exceeds that of any previous time period, and trunk water deficit as a point in time when the trunk diameter falls below that of any previous time period. This method can accurately distinguish between reversible radial fluctuations caused by changes in tree water content and actual xylem growth.

[0076] In addition, during four key hydrological periods—dry season (March), rising water season (June), high water season (August), and receding water season (October)—water samples were collected from lakes, precipitation, groundwater, stratified soil water, and xylem and leaf samples of various tree species. Water was extracted using low-temperature vacuum distillation, and hydrogen and oxygen isotopes (δ¹²H₂O) were determined using a liquid water stable isotope analyzer (Picarro L2130-i). 2 H, δ 18 O), used to calculate the proportion of water source.

[0077] The spatial distribution of canopy temperature, normalized difference vegetation index (NDVI), and photochemical reflectance index (PRI) were obtained by using a drone equipped with a hyperspectral camera.

[0078] The clustering results automatically classify trees in different water level gradient zones into different water use strategy types, for example: High water level sensitive type (corresponding to seasonally flooded lakeside forest G1): Trees experience suppressed transpiration and stunted growth during flooding in the high water season, but recover rapidly after the water recedes; Water level buffer adaptation type (corresponding to occasional flooding wetland transition forest G2): Trees maintain a relatively stable growth rhythm under the combined effects of precipitation and water level; Water level insensitive type (corresponding to G3 terrestrial forest around lakes without water inundation): Trees are mainly driven by precipitation and temperature, and their growth curves show a single-peak pattern.

[0079] This invention provides an embodiment that takes the typical evergreen broad-leaved forest of the water-land transition zone of Poyang Lake, China's largest freshwater lake wetland, as the research object, and specifically describes the implementation process of this invention.

[0080] like Figure 5 As shown, a 1-hectare long-term monitoring plot was established near Leigongling in Lushan City (29°50′N, 116°08′E). Based on the degree of flooding during the high-water season, the plot was divided into three water level gradient zones: seasonally flooded lakeside forest (G1), intermittently flooded wetland transitional forest (G2), and non-flooded lakeside terrestrial forest (G3). The interval between the three gradient zones was ≥40 meters. Five representative tree species (including Camphora officinarum, Castanopsis sclerophylla, Quercus faberi, Triadicasebifera, and Loropetalum chinense) were selected for each gradient zone, with four individuals of each species, totaling 60 trees for long-term monitoring.

[0081] A pressure-recording water level gauge (CTD-Diver) was installed near the G1 gradient zone to automatically record the water level elevation and flooding duration every 10 minutes. A 2-meter-deep groundwater observation pipe (PVC filter pipe) was installed in each of the G1, G2, and G3 gradient zones to continuously record the seasonal dynamics of groundwater depth. Multi-layer soil moisture sensors (10cm / 30cm / 50cm) were vertically deployed in each gradient zone to monitor the temporal dynamics of soil moisture content in the root zone. A small weather station (VantagePro2) was installed in the center of the study plot to provide meteorological data with a 10-minute resolution.

[0082] Following the method described in the first step, the 60 sample trees were... Figure 2The Granier thermal diffusion probe and high-precision growth instrument were installed. Water samples were collected from various sources during four key hydrological periods (dry season in March, rising season in June, wet season in August, and receding season in October), and stable hydrogen and oxygen isotopes were measured. The proportion of water sources was quantitatively analyzed using a Bayesian mixture model (MixSIAR).

[0083] Follow steps two through six in sequence. The results will show: During the dry and receding water periods, trees in lakeside forests (G1) mainly rely on shallow soil water and groundwater, resulting in vigorous sap flow and low internal water use efficiency. During the wet season, due to root zone hypoxia, the water source shifts to deep soil water, transpiration is limited, and water use efficiency increases.

[0084] Through clustering, G1 trees were classified as high water level sensitive, G2 as water level buffer adaptive, and G3 as water level insensitive.

[0085] Furthermore, interpretable analysis yielded the following conclusions: For the high-water-level sensitive type (G1), lake water level elevation and groundwater depth are the dominant factors influencing transpiration rate, with their contribution weights significantly increasing during the wet season, indicating that flooding stress is a key factor limiting transpiration water consumption in this type of plant; for the water-level buffer-adapted type (G2), the contribution weights of precipitation and soil volumetric water content remained at a high level and fluctuated little in all periods, reflecting the dual buffer-adaptive capacity of this type of plant to water level and precipitation; for the water-level insensitive type (G3), air temperature and saturated vapor pressure difference had the most prominent contribution weights, indicating that atmospheric evaporation demand is the main environmental factor driving transpiration water consumption in this type of plant.

[0086] Furthermore, to verify the effectiveness and advancement of the collaborative coupling model constructed in this invention, this embodiment uses the measured data of 60 sample trees in the Poyang Lake wetland forest as a benchmark, and compares the performance of the model of this invention with the following baseline models: (1) the first prediction model alone, namely the LSTM-Transformer hybrid model, without coupling the second prediction model and without collaborative training; (2) the traditional mechanism model, namely the transpiration estimation model based on the Penman-Monteith formula. The evaluation indicators include: root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²) of plant transpiration rate prediction. 2 The study also examined the robustness of the model's predictive performance under different levels of data missing. The missing data condition was set to randomly remove 30% of the measured values ​​of soil volumetric water content and matrix potential from the environmental time series data to simulate the actual situation in field monitoring where these two parameters are easily missing due to sensor failure or environmental interference. The results are shown in Table 1.

[0087] The key hyperparameters and structural parameters of the model were determined on the validation set by combining grid search with early stopping.

[0088] In some specific implementations, in the multimodal fusion network, the first encoder (hyperspectral) adopts a ConvNeXtV2-Tiny backbone network with an output dimension of 512; the second encoder (temporal) adopts a two-layer LSTM with 128 hidden units; and the gated fusion unit is a single-layer fully connected network that outputs fusion weights α.

[0089] In some specific implementations, in the first prediction model (LSTM-Transformer hybrid model), the number of LSTM layers is 2, with 128 hidden units per layer; the number of Transformer encoder layers is 4, the number of multi-head attention heads is 8, the feedforward network dimension is 512, and the dropout rate is 0.1.

[0090] In some specific implementations, the second prediction model has 5 fully connected hidden layers with 128 neurons per layer and tanh activation function; the output layer has 2 neurons, corresponding to soil volumetric water content θ and matrix potential ψ, respectively. In collaborative training, the alternating optimization cycle is 50 steps for the second prediction model followed by 50 steps for the first prediction model per round; the backpropagation coefficient for prediction error is set to 0.3; the maximum number of iterations is 500, and the early stopping round is 20. In the loss function, the ecological constraint term weight λ_phy = 0.1. In the clustering module, the K-means algorithm is used, with K = 3 clusters, corresponding to the three water use strategy types. The optimizer is Adam, with a learning rate of 1e-3 and a batch size of 64.

[0091] Traditional mechanistic models estimate vegetation transpiration rates based on energy balance and aerodynamic principles, without using any neural networks or data-driven methods. The specific formula is as follows: ET is the transpiration rate (mm·h) -1 After conversion, it is consistent with the unit of this invention (kg·m). -2 ·h -1 Alignment (1mm·h) -1 ≈1kg·m -2 ·h -1 Net radiation Albedo was calculated from measured photosynthetically active radiation at a meteorological station, with a value of 0.23 (typical forest canopy). Soil heat flux G: Neglected (assuming G≈0, time step ≥10 minutes). Air density. Calculated in real time based on air temperature and pressure. Specific heat at constant pressure of air. 1013 J·kg -1 ·K -1 Saturated vapor pressure Calculated using the Tetens formula, with air temperature as the input. Actual vapor pressure. : Determined by relative humidity and Calculation. Wet / dry surface constant γ: 0.065 kPa·K -1 (Under standard atmospheric pressure). Aerodynamic drag Using Monteith's formula, assuming a canopy roughness length of 0.1m, a zero-plane displacement height of 0.7 × tree height, and wind speed as measured at a meteorological station at a height of 2m, the canopy surface resistance is... The reciprocal of the canopy conductance (G) derived from the measured trunk sap flow density. c -1 The model uses daily average values. The time resolution is 10 minutes, consistent with the data acquisition device.

[0092] Table 1 Comparison of prediction performance of different models under different data integrity conditions As shown in Table 1, under the condition of 100% data integrity, the cooperative coupling model of this invention achieves the best predictive performance, with a determination coefficient R0. 2 The accuracy reached 0.90, with a root mean square error (RMSE) of only 0.038 and a mean absolute error (MAE) of 0.025. In contrast, the first prediction model alone... 2 The R value was 0.85, RMSE was 0.057, and MAE was 0.041; the traditional mechanistic model R... 2 The R² value was only 0.71, the RMSE was 0.092, and the MAE was 0.078. Compared to the first prediction model alone, the model of this invention reduced the RMSE by 33.3%, and the R² value was significantly lower. 2 It improved by 5.9%; compared to the traditional mechanistic model, RMSE decreased by 58.7%, and R... 2 The accuracy was improved by 26.8%. The above results show that, under sufficient data conditions, this invention, through the synergistic coupling of the second prediction model and the first prediction model, enables the two models to promote each other during training, and can more fully explore the temporal dependencies and ecological patterns in multimodal data, thereby achieving more accurate prediction of plant transpiration rate.

[0093] When 30% of the measured values ​​of soil volumetric water content and matrix potential were randomly removed from the environmental time-series data, the predictive performance of all models decreased to varying degrees. However, the synergistic coupling model of this invention exhibited the best robustness. Specifically, under the condition of 30% data missing, the R... 2 The R² decreased from 0.90 to 0.87, a drop of only 3.3%; the RMSE increased from 0.038 to 0.049, an increase of 28.9%. However, the first prediction model alone, under the same missing data conditions, showed a lower R². 2The performance index (PMI) plummeted from 0.85 to 0.61, a decrease of 28.2%; the RMSE rose from 0.057 to 0.118, an increase of 107.0%, indicating severe performance degradation and significant overfitting. Traditional mechanistic models rely on complete input data, and the key variable, canopy surface resistance (r), is problematic. s Since calculations were not possible when data was missing, evaluations were performed only under conditions of complete data.

[0094] The model in this invention maintains high prediction accuracy even under data loss conditions, primarily due to the role of the second prediction model in the collaborative coupling mechanism. When soil volumetric water content and matrix potential are missing from environmental time-series data, the second prediction model generates ecologically consistent complete values ​​at the missing spatiotemporal points based on the soil moisture transport law described by the Richards equation. This effectively restores the integrity of environmental information in the multimodal fusion features, allowing the first prediction model to still predict evaporation rates based on complete environmental input. Simultaneously, the alternating optimization and error backpropagation mechanism during collaborative training creates a virtuous cycle where the two models mutually reinforce each other: the complete data generated by the second prediction model is more beneficial to the evaporation prediction of the first prediction model, and the prediction error of the first prediction model in turn guides the second prediction model to optimize its completeness strategy. In contrast, the standalone first prediction model lacks ecological consistency guidance and the completeness support of the second prediction model. When environmental data is missing, it can only rely on limited samples for purely data-driven fitting, resulting in a significant decrease in model generalization ability.

[0095] In summary, the experimental results in Table 1 fully verify the effectiveness and advancement of the synergistic coupling method of the first and second prediction models proposed in this invention. In particular, in practical applications where soil volumetric water content and matrix potential data are often missing, the synergistic coupling model of this invention can improve prediction accuracy and model robustness.

[0096] like Figure 6 As shown, a plant water use strategy analysis system includes: The data acquisition and fusion module is used to acquire multimodal data of the target plant, perform feature-level fusion on the multimodal data, and obtain multimodal fusion features to characterize the interaction between the plant and the environment. The multimodal data includes environmental time series data, plant physiological time series data, and hyperspectral image time series data. The model building module is used to build and train a first prediction model based on an attention mechanism. The first prediction model is used to predict the transpiration rate of the target plant at future times based on multimodal fusion features. The module also builds and trains a second prediction model based on ecological consistency constraints. The second prediction model is used to predict the soil volumetric water content and matrix potential of the target plant in the spatiotemporal space based on multimodal fusion features and the spatial coordinates of the target plant. This is to complete the missing parts of soil volumetric water content and matrix potential in environmental time series data. The module also performs synergistic coupling training of the first and second prediction models to enhance the robustness of the first prediction model in predicting plant transpiration rate under the condition of missing environmental time series data. The water use strategy classification module is used to predict the transpiration rate of all target plants in the future time by the first prediction model after training. It uses a clustering algorithm to cluster all target plants based on the transpiration rate in the future time and classifies each target plant into different water use strategy types. The contribution quantification module is used to apply interpretability analysis algorithms to extract the weights of the attention mechanism in the first prediction model, obtain the contribution weights of various environmental factors included in the multimodal fusion features at different sampling times to the plant transpiration rate, calculate the contribution value of each environmental factor to the output of the first prediction model based on the contribution weights, and quantify the degree of influence of each environmental factor on the plant transpiration water consumption behavior in different water use strategy types.

[0097] A computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement a method for analyzing plant water use strategies.

[0098] 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 analyzing plant water use strategies, characterized in that, Includes the following steps: Step S1: Collect multimodal data of the target plant, perform feature-level fusion on the multimodal data to obtain multimodal fusion features for characterizing the interaction state between the plant and the environment. The multimodal data includes environmental time series data, plant physiological time series data and hyperspectral image time series data. Step S2: Construct and train a first prediction model based on an attention mechanism. The first prediction model is used to predict the transpiration rate of the target plant at future times based on the multimodal fusion features. Step S3: Construct and train a second prediction model based on ecological consistency constraints. The second prediction model is used to predict the soil volumetric water content and matrix potential of the target plant in time and space according to the multimodal fusion features and the spatial coordinates of the target plant, so as to complete the missing parts of soil volumetric water content and matrix potential in the environmental time series data. Step S4: Co-couple the first prediction model and the second prediction model for training to enhance the robustness of the first prediction model in predicting plant transpiration rate under conditions of missing environmental time series data. Step S5: The trained first prediction model predicts the transpiration rate of all target plants at future times. A clustering algorithm is used to cluster all target plants based on the plant transpiration rate at the future times, and each target plant is divided into different water use strategy types. Step S6: Apply the interpretability analysis algorithm to extract the weights of the attention mechanism in the first prediction model, obtain the contribution weights of each environmental factor included in the multimodal fusion feature to the plant transpiration rate at different sampling times, calculate the contribution value of each environmental factor to the output of the first prediction model based on the contribution weights, and quantify the degree of influence of each environmental factor on the plant transpiration water consumption behavior in different water use strategy types.

2. The method for analyzing plant water use strategies according to claim 1, characterized in that: The environmental time series data are environmental factor data collected continuously over time, including lake water level elevation, groundwater depth, soil volumetric water content, matrix potential, air temperature, relative humidity, photosynthetically active radiation, precipitation, and saturated vapor pressure difference. The plant physiological time series data are plant physiological response data collected continuously over time. The plant physiological response data includes trunk sap flow density continuously monitored by the Granier thermal diffusion probe, trunk water deficit extracted from trunk radial changes continuously monitored by a high-precision growth instrument, leaf water potential at dawn and noon measured according to the hydrological cycle, and leaf stable carbon isotopes. The hyperspectral image time series data is multi-temporal hyperspectral remote sensing data repeatedly collected at preset time intervals. The multi-temporal hyperspectral remote sensing data includes the spatial distribution of canopy temperature, normalized vegetation index, and photochemical reflectance index.

3. The method for analyzing plant water use strategies according to claim 2, characterized in that: The fusion method for the multimodal fusion features includes: A feature fusion network is constructed for feature-level fusion of multimodal data. The feature fusion network consists of a dual encoder-decoder structure with a gated fusion mechanism, wherein the first encoder is used to extract the spatial-spectral features of the hyperspectral image time series data. The second encoder is used to extract the temporal features of the environmental time series data and the plant physiological time series data. ; The spatial-spectral features are dynamically calculated using a gating fusion mechanism. and the aforementioned time-series features The fusion weights α are used to generate multimodal fusion features F, where: ； ； In the formula, σ is the sigmoid activation function. and Here are the learnable weights and bias parameters in the feature fusion network, and [:] represents the feature concatenation operation.

4. The method for analyzing plant water use strategies according to claim 3, characterized in that: The first prediction model is an LSTM-Transformer hybrid model; The multimodal fusion features are input into the LSTM layer of the LSTM-Transformer hybrid model. The LSTM layer performs temporal encoding on the multimodal fusion features, extracts the short-term temporal dependencies contained therein, and outputs latent features with short-term temporal memory. The latent features are input into the Transformer encoder in the LSTM-Transformer hybrid model. The multi-head self-attention mechanism in the Transformer encoder is used to capture the long-term dependencies and global interaction features between variables in the multimodal fusion features. The output of the Transformer encoder is mapped through a fully connected layer to obtain the predicted value of the plant transpiration rate at future times for the target plant; Specifically, the attention weights of the multi-head self-attention mechanism in the Transformer encoder are aggregated and normalized to obtain the contribution weights of various environmental factors to the transpiration rate of the plant at different sampling times.

5. The method for analyzing plant water use strategies according to claim 4, characterized in that: The second prediction model uses a fully connected neural network structure; The spatial coordinates of the target plant and the multimodal fusion features are input into the input layer of a fully connected neural network structure. The input layer concatenates the spatial coordinates of the target plant and the multimodal fusion features to obtain a fusion vector. The fused vector is input into the hidden layer of a fully connected neural network structure, and the hidden layer performs a nonlinear transformation on the fused vector to extract the spatiotemporal correlation features contained therein. The spatiotemporal correlation features are mapped through the output layer to obtain the soil volumetric water content and matrix potential of the target plant in the spatiotemporal space.

6. The method for analyzing plant water use strategies according to claim 5, characterized in that: The training loss of the second prediction model consists of a weighted sum of two parts: data fitting loss and ecological consistency constraint loss. The formula for calculating the training loss is as follows: In the formula, For data fitting loss, Loss due to ecological consistency constraints. For hyperparameters; The data fitting loss is the mean square error between the predicted soil volumetric water content and matrix potential output by the second prediction model and the corresponding measured values. The formula for calculating the data fitting loss is as follows: In the formula, N is the total number of measured samples. For the first The measured volumetric water content of the soil in each sample For the first Predict soil volumetric water content for each sample For the first Measured matrix potential of each sample For the first Predicting matrix potential for each sample The ecological consistency constraint loss is the residual norm of the water flow equation in the soil-plant-atmosphere continuum, and the formula for calculating the ecological consistency constraint loss is as follows: In the formula For soil hydraulic conductivity, The rate of change of soil volumetric water content over time. The gradient of the matrix potential along the vertical depth is given by the soil hydraulic conductivity, which is a function of the soil volumetric water content.

7. The method for analyzing plant water use strategies according to claim 6, characterized in that: The collaborative coupling training of the first prediction model and the second prediction model includes: Step S401: Using an alternating optimization strategy, in each training round, first fix the parameters of the first prediction model, and use the second prediction model to generate predicted values ​​at the missing time points of soil volumetric water content and matrix potential in the environmental time series data to complete the environmental time series data; Step S402: Fix the parameters of the second prediction model, reconstruct new multimodal data using the completed environmental time series data, plant physiological time series data and hyperspectral image time series data, extract new multimodal fusion features from the new multimodal data using a feature fusion network, train the first prediction model based on the new multimodal fusion features, and calculate the prediction error of the first prediction model. Step S403: The prediction error of the first prediction model is used as an additional loss term and backpropagated to the second prediction model to update the parameters of the second prediction model so that the soil volume water content and matrix potential data of the missing segment generated by the second prediction model are more conducive to improving the prediction accuracy of the first prediction model. Step S404: Repeat steps S401 to S403 until the loss functions of the first prediction model and the second prediction model converge, thus obtaining the trained first prediction model and the second prediction model.

8. The method for analyzing plant water use strategies according to claim 7, characterized in that: Methods for clustering all target plants based on the plant transpiration rate at the future time using clustering algorithms include: The transpiration rates of all target plants output by the first prediction model after training are used to form the time series feature vector of the transpiration rate of each target plant at future times. Set the number of clusters K, and randomly select K transpiration rate time series feature vectors of target plants as the initial cluster centers; Calculate the Euclidean distance between the transpiration rate time series feature vector of each target plant and each cluster center, and assign each target plant to the category of the nearest cluster center; Update each cluster center based on the allocation results, and repeat the allocation and update steps until the cluster centers no longer change or the preset number of iterations is reached, so that each target plant is classified into different water use strategy types.

9. A system for analyzing plant water use strategies, characterized in that, The system, which is applied to the plant water use strategy analysis method according to any one of claims 1-8, comprises: The data acquisition and fusion module is used to acquire multimodal data of the target plant, perform feature-level fusion on the multimodal data, and obtain multimodal fusion features to characterize the interaction state between the plant and the environment. The multimodal data includes environmental time series data, plant physiological time series data, and hyperspectral image time series data. The model building module is used to build and train a first prediction model based on an attention mechanism. The first prediction model is used to predict the transpiration rate of the target plant at future times based on the multimodal fusion features. The module also builds and trains a second prediction model based on ecological consistency constraints. The second prediction model is used to predict the soil volumetric water content and matrix potential of the target plant in the space-time of the target plant based on the multimodal fusion features and the spatial coordinates of the target plant, so as to complete the missing parts of soil volumetric water content and matrix potential in the environmental time series data. The module also performs collaborative coupling training of the first prediction model and the second prediction model to enhance the robustness of the first prediction model in predicting the transpiration rate of the plant under the condition of missing environmental time series data. The water use strategy classification module is used to predict the transpiration rate of all target plants at future times by the first prediction model that has been trained. The clustering algorithm is used to cluster all target plants based on the plant transpiration rate at the future time, and classify each target plant into different water use strategy types. The contribution quantification module is used to apply an interpretability analysis algorithm to extract the weights of the attention mechanism in the first prediction model, obtain the contribution weights of each environmental factor included in the multimodal fusion feature to the plant transpiration rate at different sampling times, calculate the contribution value of each environmental factor to the output of the first prediction model based on the contribution weights, and quantify the degree of influence of each environmental factor on the plant transpiration water consumption behavior in different water use strategy types.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, implement the method as described in any one of claims 1-8.

Images