Grazing rate prediction data processing method based on plant morphological characteristics

Abstract

The invention relates to a grazing rate prediction data processing method based on plant morphological characteristics, and belongs to the technical field of electric digital data processing and machine learning application. The method solves the problems that the grazing response signal is weak due to the dependence on the overall characteristics of the vegetation community, and the reproducibility is poor due to the lack of a standardized process. The method comprises the following steps: in a model training stage, receiving a plant individual form measurement value containing a plant height and a crown diameter and a grazing rate reference value, calculating an individual form index based on the plant height and the crown diameter, dividing a space unit and calculating spatial statistical characteristics of the form index in the unit, and training to obtain a grazing rate estimation model; in the model application stage, the same index calculation and space statistics steps are executed on the new plant morphology measurement value, the result is input into the trained model, and the estimated grazing rate is output. According to the method, through building individual index calculation, space unit statistics, model training and application, effective prediction of the grazing rate based on plant morphological characteristics is realized.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of digital data processing and machine learning application, and particularly relates to a grazing rate prediction data processing method based on plant morphological characteristics. BACKGROUND

[0002] In the machine learning application scenarios related to ecological environment monitoring and analysis, accurately and efficiently extracting effective information from plant data collected by remote sensing or in the field to assess grazing pressure is a long-standing and challenging technical problem.

[0003] Firstly, when using machine vision and data analysis technology to evaluate grazing activities, the core difficulty lies in how to extract quantitative features that are sensitive to grazing intensity and stable from the complex changes in plant morphology. Traditional methods rely heavily on remote sensing vegetation indices (such as NDVI, EVI) or average features at the community level (such as average height, total biomass). However, these features essentially reflect the overall spectrum or macro physiological state of the vegetation community, which is extremely sensitive to environmental background noise such as climate and soil moisture, and responds slowly and lacks specificity to the subtle changes in individual plant morphological structure caused by selective grazing behavior. Grazing activities often result in non-uniform and nonlinear responses in plant individual morphological dimensions such as height and crown width, for example, the grazed plants become shorter but may grow laterally. Using community average features will smooth out these key spatial heterogeneity information, causing the model to fail to effectively distinguish between morphological changes caused by grazing and similar changes caused by other environmental factors, thus severely restricting the accuracy and robustness of the prediction model. The fundamental reason is that the feature engineering level has failed to deeply explore the internal correlation of individual-scale morphological parameters and their spatial distribution patterns.

[0004] Secondly, in the data processing process of building prediction models, another prominent problem is the lack of systematic and automated feature construction and optimization mechanism. Existing technologies usually preset a single morphological indicator (such as only using plant height) or a fixed spatial statistical scale based on prior knowledge or simple experience. However, the combination of morphological features that can most effectively correlate grazing pressure (such as what mathematical relationship between plant height and crown width) and the best spatial aggregation scale will vary with vegetation type, geographic region, and seasonal dynamics. Manually trying and empirically selecting these key parameters not only is tedious and inefficient, but also makes it difficult to ensure that the global or approximately optimal solution is found, resulting in the performance of the final model being highly dependent on the operator's professional experience and being unstable. This reflects the lack of a technical framework that can adaptively and data-drivenly optimize feature extraction functions and spatial analysis units at the data processing method level.

[0005] Furthermore, in the model building stage, there is often a complex nonlinear relationship between grazing rate and plant morphological characteristics, and the data is often affected by various confounding factors. Simple linear regression models struggle to capture this complex mapping, while directly applying complex nonlinear models (such as deep neural networks) risks overfitting in small sample scenarios and suffers from poor model interpretability. Simultaneously, models trained on data from a single region often experience significant performance degradation when applied to new regions with different ecological backgrounds due to variations in data distribution, indicating insufficient spatiotemporal generalization ability. This necessitates that the model architecture itself possess both strong nonlinear fitting capabilities and good generalization ability, but traditional single modeling strategies often fail to simultaneously meet both objectives.

[0006] Furthermore, obtaining the label data—the "grazing rate"—which is crucial for training supervised learning models, is also a technical bottleneck. Grazing rate data obtained through manual recording or surveys often suffers from inaccuracies, discontinuities, or high costs. How to automatically generate quantifiable grazing pressure proxy labels that precisely match plant morphology observations in time and space using readily available objective data sources (such as livestock GPS tracks) through deterministic data processing algorithms is key to building a reliable data loop and is also a weak link in the current technical process.

[0007] In summary, existing technologies, when applied to the specific task of grazing rate prediction, have limitations in areas such as the effectiveness and automation of feature engineering, the model's ability to characterize complex nonlinear relationships and its generalization capabilities, and the automated generation of high-quality supervision labels. These issues collectively restrict the accuracy, automation level, and practical application scope of grazing rate prediction methods based on plant morphology data, necessitating a systematic solution from the perspective of data processing methodologies. Summary of the Invention

[0008] One object of the present invention is to solve at least the above-mentioned problems and / or defects, and to provide at least the advantages described below.

[0009] One objective of this invention is to address the problems of incomplete data processing and insufficient standardization in existing technologies for predicting grazing rates based on plant morphological characteristics. Traditional methods often separate data collection, feature extraction, and model application, lacking a unified and reproducible data processing framework from raw individual morphological measurements to final regional estimates, resulting in poor method versatility and weak comparability of results.

[0010] One objective of this invention is to address the difficulty and high cost of acquiring high-quality labeled data in supervised learning models. Traditional grazing rate data relies on manual recording, which suffers from subjective errors, spatiotemporal mismatches, and high acquisition costs, thus hindering the large-scale application of data-driven models.

[0011] An object of the present application is to solve the problem of relying on artificial experience to select key morphological indexes in feature engineering, and lacking objective optimization basis. The prior art usually uses a single morphological index (such as plant height) based on prior knowledge, ignoring the possibility of more effective combination relationship between plant height and crown width and other multi-morphological parameters.

[0012] An object of the present application is to solve the problem of difficulty in determining the optimal spatial aggregation scale in spatial data analysis. When individual plant data in point form is aggregated into spatial statistical features in surface form, scale selection has a decisive influence on feature effectiveness, but traditional methods often fix the scale by experience.

[0013] An object of the present application is to solve the problem of insufficient fitting ability or weak generalization ability of a single model in modeling complex nonlinear relationships. The relationship between grazing rate and morphological features is complex and affected by many interference factors, and simple models are difficult to fit, and complex models are prone to overfitting.

[0014] An object of the present application is to solve the problem of large grid data noise and poor continuity caused by uneven spatial distribution of original trajectory data. The distribution of GPS trajectory points is affected by the behavior of livestock and may be uneven, and the spatio-temporal occupancy grid generated directly has data holes or sharp noise, affecting the accuracy of subsequent label calculation.

[0015] An object of the present application is to solve the problem that simple linear accumulation cannot accurately reflect the group superposition effect when converting the number of livestock and residence time into grazing pressure indicators. The ecological pressure of multiple livestock existing at the same time is not the simple arithmetic sum of individual effects.

[0016] An object of the present application is to solve the problem of randomness and overfitting risk of evaluation results due to single data division in the process of automated feature selection. Evaluating candidate functions on only a single data set may result in selecting functions that are overfitted to that particular data and have poor generalization ability.

[0017] An object of the present application is to solve the problem that a single spatial scale feature may not be able to fully capture multi-scale spatial patterns. The expression of grazing pressure in space may be present at different scales, and fixing a single scale will lose information.

[0018] An object of the present application is to solve the problem that base learners in ensemble learning may have too high prediction correlation and insufficient diversity due to the same input features. This limits the effect of ensemble learning in improving model performance.

[0019] An object of the present application is to provide a grazing rate prediction data processing method based on plant morphological features, including a model training phase and a model application phase: The model training phase includes: receiving a training dataset, the training dataset comprising morphological measurements of a plurality of plant individuals collected in a first geographical region, the morphological measurements comprising at least plant height and crown diameter; receiving a grazing rate benchmark value corresponding to the first geographical region; for each plant individual, calculating a morphological index based on its plant height and crown diameter; dividing the first geographical region into a plurality of first spatial units according to a preset spatial division rule, and calculating spatial statistical features of each first spatial unit based on the morphological indices of all plant individuals falling within the first spatial unit; training a regression model with the spatial statistical features of the first spatial units as input features and the grazing rate benchmark value as a supervised label, to obtain a grazing rate estimation model; the model application stage comprises: receiving a to-be-evaluated region, the to-be-evaluated region comprising a to-be-tested dataset of a plurality of plant individuals; for each plant individual in the to-be-tested dataset, calculating a morphological index based on its plant height and crown diameter; dividing the to-be-evaluated region into a plurality of second spatial units according to a preset spatial division rule, and calculating spatial statistical features of each second spatial unit based on the morphological indices of all plant individuals falling within the second spatial unit; inputting the spatial statistical features of the second spatial units into the grazing rate estimation model to output corresponding estimated grazing rates.

[0020] Preferably, in the grazing rate prediction data processing method based on plant morphological features, the grazing rate benchmark value corresponding to the first geographical region is received, specifically comprising: obtaining livestock trajectory data of the first geographical region in a preset historical time period, the livestock trajectory data comprising livestock position information with timestamps; calculating a livestock spatiotemporal occupation grid of the first geographical region in the preset historical time period according to the livestock trajectory data, wherein the value of each pixel in the grid is the cumulative residence time of all livestock in the geographical range of the pixel in the preset historical time period; obtaining a plant morphological measurement collection time point; defining a time backtracking window with the collection time point as the endpoint; extracting all pixel values corresponding to the range of the first geographical region from the livestock spatiotemporal occupation grid, and calculating the cumulative sum of these pixel values in the time backtracking window, and taking the cumulative sum as the grazing rate benchmark value.

[0021] Preferably, in the grazing rate prediction data processing method based on plant morphological features, the calculation relationship of the morphological index is determined through the following function selection steps: predefining a plurality of candidate functions, each candidate function being used to calculate a numerical value based on plant height and crown diameter; Before the model training stage, a function selection procedure is performed, which includes: The first geographical area is divided into a plurality of evaluation units using the same preset spatial division rule as in the model training stage; For each candidate function, the following evaluation operations are performed: Using the candidate function, the plant height and crown diameter of all plant individuals in the training data set are calculated to obtain the candidate morphological index of each plant individual; For each evaluation unit, the candidate statistical characteristics of the unit are calculated based on the candidate morphological index of all plant individuals falling into the unit; The correlation measure value between the candidate statistical characteristics of all evaluation units and the corresponding grazing rate reference value is calculated; From all candidate functions, the candidate function with the highest correlation measure value is selected as the final morphological index calculation relationship for the model training stage and the application stage.

[0022] Preferably, in the plant morphological feature-based grazing rate prediction data processing method, the size of the spatial unit in the preset spatial division rule is determined through the following scale optimization steps: A set of candidate spatial scales is predefined; A baseline calculation relationship for calculating the morphological index is preselected; A scale evaluation procedure is performed, which includes: The first geographical area corresponding to the training data set is divided into a candidate spatial unit grid of corresponding size according to each candidate spatial scale; For each candidate spatial scale, the following operations are performed: The morphological index of each plant individual in the training data set is calculated according to the baseline calculation relationship; For each candidate spatial unit under this scale division, the candidate statistical characteristics of the unit are calculated based on the morphological index of all plant individuals falling into it; The candidate statistical characteristics of all candidate spatial units obtained under this scale are divided into a training subset and a validation subset together with the corresponding grazing rate reference value; A scale evaluation regression model is trained using the training subset, and the prediction accuracy index of the scale evaluation regression model is calculated using the validation subset; The prediction accuracy indexes corresponding to all candidate spatial scales are compared, and the candidate spatial scale with the optimal prediction accuracy index is determined as the spatial unit size used in the preset spatial division rule.

[0023] Preferably, in the plant morphological feature-based grazing rate prediction data processing method, the training regression model to obtain the grazing rate estimation model is implemented through a stacked ensemble training method, which includes the following steps: The dataset consisting of spatial statistical features and corresponding grazing rate benchmarks is divided into a first training set and a second training set. Base learner training and meta-feature generation steps: Using the first training set, support vector regression model, gradient boosting tree regression model and multilayer perceptron neural network model are trained respectively; then, the spatial statistical features of the second training set are used as inputs and input into the three trained models respectively to obtain the predicted values ​​of their respective outputs. The three predicted values ​​corresponding to the same second training set sample are combined into a three-dimensional meta-feature vector. Meta-learner training steps: Using the three-dimensional meta-feature vector as input features and the grazing rate benchmark value corresponding to the second training set as output labels, train a linear regression model as the final ensemble grazing rate estimation model.

[0024] Preferably, in the grazing rate prediction data processing method based on plant morphological characteristics, after generating the livestock spatiotemporal occupancy raster, the method further includes: Spatial smoothing filtering is applied to the livestock spatiotemporal occupancy raster to obtain smoothed raster data; Smoothed raster data is used to replace the original livestock spatiotemporal occupancy raster for subsequent extraction of cumulative sums.

[0025] Preferably, in the grazing rate prediction data processing method based on plant morphological characteristics, the step of calculating the livestock spatiotemporal occupancy raster includes calculating the cumulative value of each pixel within each unit time interval, which includes: Count the number of different livestock individuals located within the geographic range of this pixel within this unit of time interval; Using the number of different livestock individuals within the geographic range of the pixel in the unit time interval as input, a pressure weight value is calculated through a nonlinear transformation function, wherein the nonlinear transformation function is configured to increase the output pressure weight value in a superlinear manner as the number of inputs increases; The product of the pressure weight value and the unit time length is taken as the effective grazing pressure contribution value of the pixel within the unit time interval. The effective grazing pressure contribution values ​​calculated within all unit time intervals are summed to obtain the final value of the corresponding cell in the livestock spatiotemporal occupation raster.

[0026] Preferably, in the grazing rate prediction data processing method based on plant morphological characteristics, the step of calculating the correlation metric value specifically includes: The training dataset is randomly divided into a training subset and a validation subset; The evaluation operation is performed using a subset of the training set to obtain the first correlation measure between the candidate statistical features and the corresponding grazing rate benchmark. The evaluation operation is performed using a validation subset to obtain a second correlation measure between candidate statistical features and the corresponding grazing rate benchmark. The weighted average of the first and second correlation measures is used as the final correlation measure of the candidate function.

[0027] Preferably, in the grazing rate prediction data processing method based on plant morphological characteristics, the regression model is a multi-scale feature fusion regression model; its training method is as follows: For the first geographic region, multiple different candidate spatial scales are used to divide it, resulting in multiple sets of spatial units; For each spatial unit, the spatial statistical characteristics at that scale are calculated based on the morphological index of the individual plants that fall into it. Spatial statistical features calculated at different scales for the same geographical location are spliced ​​together to form a multi-scale fused feature vector. Using multi-scale fused feature vectors as input features and grazing rate benchmark values ​​as supervision labels, a multi-scale feature fusion regression model is trained, which serves as the final grazing rate estimation model.

[0028] Preferably, in the grazing rate prediction data processing method based on plant morphological characteristics, in the base learner training and meta-feature generation steps: For the spatial statistical features of the first training set, a set of random noise features is added to each set to form three different enhanced feature sets; The three enhanced feature sets were used as input features for training the support vector regression model, gradient boosting tree regression model, and multilayer perceptron neural network model, respectively. When generating meta-feature vectors using the second training set, the same random noise features as before are added to the spatial statistical features of the second training set, and then input into the corresponding base learner to obtain the predicted value.

[0029] The present invention has at least the following beneficial effects: This invention achieves effective prediction of regional grazing rates based on individual plant morphological characteristics by constructing a complete technical process encompassing individual index calculation, spatial unit statistics, model training, and application. Starting with individual-scale morphological data, this scheme systematically addresses core issues such as insufficient signal specificity due to reliance on community average characteristics, information loss due to ignoring spatial heterogeneity, and poor reproducibility due to a lack of standardized procedures through spatial structured aggregation and machine learning modeling. The entire process provides a computable path from raw data to prediction results.

[0030] This invention provides an objective and automatically obtainable source of supervisory signals for supervised learning model training by designing an algorithmic label generation method based on spatiotemporal trajectory data. This method transforms grazing rate data, which originally relied on manual recording, into a quantifiable indicator that can be directly calculated, reducing data acquisition costs and subjective errors. The generated grazing rate benchmark value has a clear correspondence with plant morphology observation data in the spatiotemporal dimensions, improving the internal consistency and reliability of the training dataset and laying the foundation for building a robust prediction model.

[0031] This invention introduces a data-driven automatic selection mechanism for morphological index functions, transforming feature construction from an experience-based process to an optimization process based on quantitative evaluation. This method automatically selects the morphological index calculation method with the highest statistical correlation to the target variable from a set of candidate functions, thereby improving the objectivity and adaptability of the feature engineering process. It enables feature representations to better adapt to specific datasets and task objectives, helping to uncover morphological feature combinations that are more sensitive to grazing pressure.

[0032] This invention proposes a spatial scale optimization method based on model performance feedback, providing a quantitative decision-making basis for determining the optimal size of spatial statistical units. This method avoids the blindness of selecting scales based on experience, and automatically selects the spatial aggregation granularity most beneficial to model performance by evaluating the performance of features constructed at different candidate scales in the prediction task. This helps to extract regional statistical features with stronger spatial representativeness and richer information, thereby improving the quality of input data for subsequent models.

[0033] This invention integrates the advantages of multiple heterogeneous base learners by employing a stacked ensemble learning model construction strategy. This method utilizes base learners to capture different patterns in the data, and then effectively fuses them through meta-learners, forming a composite prediction model capable of handling complex nonlinear relationships. This structure enhances the fitting ability to complex relationships in the data without excessively increasing the complexity of individual models, and helps improve the overall predictive stability and generalization potential of the model.

[0034] This invention improves the data quality of the spatiotemporal occupancy raster generated from the original trajectory data by introducing a spatial smoothing filtering process. This process effectively reduces data noise and spatial discontinuities caused by uneven distribution of trajectory points, generating a smoother and more consistent raster surface. The processed raster data provides a more stable and reliable input for subsequent calculations, helping to reduce errors introduced by this step, thereby improving the consistency and accuracy of the final grazing rate benchmark value.

[0035] This invention improves the calculation method for grazing stress contribution by introducing a nonlinear transformation function with superlinear growth characteristics. When quantifying the effects of herd grazing, this method considers the potential nonlinear superposition effects of multiple livestock simultaneously, making the calculated stress values ​​more ecologically reasonable. This allows the generated grazing stress baseline values ​​to more precisely characterize the stress differences under different livestock numbers, providing more discriminative and informative supervision labels for model learning.

[0036] This invention enhances the robustness of the morphological index selection process by introducing a training-validation set partitioning and weighted average evaluation strategy into the function selection process. This method simultaneously considers the fitting ability of candidate functions on the training set and their generalization performance on the validation set, and combines these two factors in a weighted manner, reducing reliance on the results of a single data partitioning. This helps to select morphological index functions with stronger generalization ability and less prone to overfitting, thereby making the results of the entire feature construction process more reliable.

[0037] This invention constructs a multi-scale feature fusion regression model, enabling the model to simultaneously utilize statistical feature information at different spatial scales. This method concatenates spatial pattern representations from multiple scales into a fused feature vector, providing the model with more comprehensive and richer spatial contextual information. This allows the model to adaptively weigh and utilize effective signals at different scales during learning and prediction, thereby enhancing its overall ability to characterize complex spatial distribution patterns.

[0038] This invention increases the diversity among base learners in ensemble learning at a lower computational cost by adding differentiated fixed random noise features to different base learners. This method reduces the correlation of their prediction errors by causing each base learner to learn from slightly different data perspectives through slight and controlled perturbations to the input features. This helps improve the overall performance of the stacked ensemble, enabling the ensemble model to achieve better generalization results than a single base learner or a homogeneous ensemble.

[0039] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Attached Figure Description

[0040] Figure 1 This is a flowchart of a data processing method for predicting grazing rates based on plant morphological characteristics. Detailed Implementation

[0041] The present invention will now be described in further detail so that those skilled in the art can implement it based on the description.

[0042] This invention provides a data processing method for predicting grazing rates based on plant morphological characteristics, including a model training stage and a model application stage. The model training stage includes: receiving a training dataset, which includes morphological measurements of multiple plant individuals collected within a first geographical region, the morphological measurements including at least plant height and crown diameter; receiving a grazing rate benchmark value corresponding to the first geographical region; calculating a morphological index for each plant individual based on its plant height and crown diameter; dividing the first geographical region into multiple first spatial units according to a preset spatial division rule, and calculating the spatial statistical characteristics of each unit based on the morphological indices of all plant individuals falling within each first spatial unit; and applying the morphological index of each plant individual within each first spatial unit to the next grazing rate. The spatial statistical features of a spatial unit are used as input features, and the grazing rate benchmark is used as the supervision label to train a regression model to obtain a grazing rate estimation model. The model application stage includes: receiving the test dataset of the region to be evaluated, which includes morphological measurements of multiple plant individuals; calculating the morphological index for each plant individual in the test dataset based on its plant height and crown diameter; dividing the region to be evaluated into multiple second spatial units according to a preset spatial division rule, and calculating the spatial statistical features of each unit based on the morphological indices of all plant individuals falling into each second spatial unit; inputting the spatial statistical features of each second spatial unit into the grazing rate estimation model, and outputting the corresponding estimated grazing rate.

[0043] For ease of understanding, the following explanations are provided for some key terms used in this specification: The first geographic region refers to the specific geographic area used to collect training data (plant morphology measurements and corresponding grazing rate baselines) during the model training phase.

[0044] The second spatial unit refers to the basic analytical unit obtained after spatially dividing the region to be evaluated during the model application phase. Its division rules are the same as the "preset spatial division rules" used during the model training phase to divide the "first spatial unit".

[0045] The area to be evaluated refers to a new geographical area where the grazing rate needs to be estimated during the model application phase.

[0046] A morphological index is a scalar value that combines two basic morphological parameters of an individual plant—plant height and crown diameter—through a mathematical formula, and is used to characterize the overall morphological state of the plant.

[0047] Spatial statistical characteristics refer to the characteristic values ​​obtained after statistical analysis of the morphological indices of all plant individuals falling within a spatial unit, such as, but not limited to, the mean, standard deviation, median, and coefficient of variation of the morphological indices within that unit.

[0048] The grazing rate baseline refers to a quantitative grazing pressure label calculated during the model training phase using objective data sources (such as livestock trajectory data), which is spatiotemporally matched with plant morphology observation data and used to supervise model training.

[0049] One specific implementation of the present invention is as follows: First, within a target grazing grassland, a high-precision visible light and lidar sensor mounted on a drone is used to conduct aerial photography, acquiring high-resolution images and 3D point cloud data covering the area. Through point cloud segmentation and image recognition algorithms, the plant height and crown diameter of each major herbaceous plant are automatically extracted, forming a dataset of individual plant morphological measurements. Simultaneously, GPS collars deployed in the area record trajectory data over the same time period.

[0050] The data processing system receives the two sets of data mentioned above. For morphological data, the system calculates a comprehensive morphological index for each plant, which is obtained by using a preset ratio function between plant height and crown diameter. Subsequently, the system automatically divides the grassland into multiple square grids with sides of ten meters as spatial units on a digital map. For each grid, the system calculates the arithmetic mean and standard deviation of the comprehensive morphological index of all individual plants falling within it, using these two values ​​as the spatial statistical characteristics of that grid.

[0051] For trajectory data, the system processes it into a spatiotemporal occupancy grid, calculates the cumulative number of hours all livestock stay in each grid cell, and uses the cumulative value of thirty days prior to the time of plant data collection as the benchmark value for the grazing rate of the area.

[0052] Next, the system uses the spatial statistical features of all grids as input features and the corresponding grazing rate benchmark values ​​as labels, and trains the model using the gradient boosting tree algorithm to generate a grazing rate estimation model. When a new grassland needs to be evaluated, the system repeats the above steps of data collection, individual index calculation, grid division, and feature extraction, and inputs the obtained spatial statistical features into the trained model to output the estimated grazing rate value of the grassland.

[0053] In embodiments of the present invention, the "preset spatial division rule" can be used to divide the space using a fixed scale based on domain experience, such as a square grid with a side length of 10 meters. The "preset spatial division rule" can employ a regular grid division method. Specifically, the geographical region (whether it is the first geographical region or the region to be evaluated) is divided into multiple square grid units of the same size on a two-dimensional plane coordinate system (such as the UTM coordinate system), with the grid side length denoted as L. The value of L ranges from 1 meter to 100 meters, preferably 10 meters. During division, a regular grid covering the entire region is generated using the bounding rectangle of the region as the scope; grid units exceeding the region boundary are not included in the statistics. Each grid unit serves as an independent spatial analysis unit. The system records its center point coordinates or boundary coordinates and determines which grid unit each plant individual's coordinates fall into based on a spatial index (such as an R-tree or a grid index), thereby determining the attribution relationship between the plant individual and the spatial unit. The "morphological index" can be directly calculated using a simple mathematical relationship (such as a ratio or product) between plant height and crown diameter. The "regression model" can be trained using machine learning regression algorithms commonly used in the field, such as Gradient Boosting Decision Tree (GBDT) or Random Forest.

[0054] The grazing rate estimation model is a regression model based on machine learning, and its mathematical form can be expressed as: in, The output is the estimated grazing rate (scalar), x is the input feature vector, and Θ is the model parameters. The input feature vector x consists of the spatial statistical features of each spatial unit, specifically including but not limited to: the mean of the morphological indices of all individual plants within that unit. Standard deviation Coefficient of variation For example, if the mean and standard deviation are chosen as features, then... The model can be any regression algorithm or an ensemble thereof, such as linear regression, decision tree, random forest, gradient boosting tree, or neural network. During the model training phase, the predicted value is minimized. The optimal model parameters Θ are obtained by solving the loss function (such as mean squared error) between the model and the true grazing rate benchmark value y.

[0055] During the model application phase, for each second spatial unit of the region to be evaluated, the following steps are performed: (1) Determine the scope of the unit according to the preset spatial division rules; (2) Extract the morphological indices of all plant individuals falling into the unit; (3) Calculate the spatial statistical characteristics of the unit (such as the mean of the morphological index). and standard deviation ,in j Indicates the first j (units) (4) Calculate the eigenvectors Input the pre-trained grazing rate estimation model f ; (5) The model outputs the estimated grazing rate of this unit. .

[0056] Ultimately, the overall grazing rate of the area to be evaluated can be obtained by summing the estimated values ​​of each unit (such as taking the mean or weighted sum), or by directly outputting the grazing rate distribution map of each unit.

[0057] Existing technologies typically perform calculations directly on the spectral information of satellite image pixels to obtain indices reflecting vegetation greenness or productivity, and then establish statistical relationships with grazing activities at a regional scale. This method acquires overall, averaged-state information about the vegetation community, and its signals are easily affected by strong interference from weather, phenological stages, and soil background. Most importantly, it cannot detect changes in individual plant morphology caused by selective foraging, such as dwarfing or altered crown width—microscopic morphological adjustments that are precisely the most direct characteristic responses of grazing activities. Therefore, existing technologies for indicating grazing stress suffer from low sensitivity, poor specificity, and ambiguous interpretation.

[0058] In contrast, the implementation of this invention fundamentally changes the data source and analysis path. It directly extracts precise morphological parameters of individual plants from high-resolution remote sensing or field measurements, and uses these as the basis for subsequent calculations. By constructing a data processing chain of "individual morphological index - spatial unit statistical characteristics," this invention can directly quantify the specific morphological changes that grazing may induce. Spatial unit statistical characteristics further capture the spatial heterogeneity of these changes. Finally, a machine learning model is used to establish a mapping relationship between these physically meaningful morphological spatial features and grazing activities. This technical approach allows the analysis signal to originate directly from the grazing process itself, significantly improving the specificity, accuracy, and mechanistic interpretability of grazing stress monitoring, achieving a technological leap from monitoring "vegetation status" to assessing "grazing impacts."

[0059] The data processing flow of this invention can be summarized as follows: collecting individual plant morphological data and calculating morphological indices; dividing the region into spatial units according to preset spatial division rules (such as fixed-size grids) and calculating spatial statistical characteristics; establishing a mapping relationship between spatial statistical characteristics and grazing rate benchmark values ​​using a regression model, thereby estimating the grazing rate of the region. This core flow is the foundation of the method of this invention, and its specific implementation can be based on empirical parameters, or its performance and adaptability can be further improved through the optimization steps described below.

[0060] In a preferred embodiment, the method of receiving a grazing rate benchmark value corresponding to a first geographic region specifically includes: acquiring livestock trajectory data of the first geographic region within a preset historical time period, the livestock trajectory data containing livestock location information with timestamps; calculating a livestock spatiotemporal occupancy grid of the first geographic region within the preset historical time period based on the livestock trajectory data, wherein the value of each pixel in the grid is the cumulative dwell time of all livestock within the geographic range of that pixel during the preset historical time period; acquiring the collection time point of plant morphology measurements; defining a time backtracking window backward from the collection time point as the endpoint; extracting all pixel values ​​corresponding to the range of the first geographic region from the livestock spatiotemporal occupancy grid, and calculating the cumulative sum of these pixel values ​​within the time backtracking window, using the cumulative sum as the grazing rate benchmark value.

[0061] One specific implementation of the present invention is as follows: During the model training phase, the system needs to acquire a grazing rate benchmark value that matches the plant morphology measurement data. To achieve this, the system first accesses the trajectory data stream continuously uploaded by the GPS positioning collars deployed on the target sheep flock. This data stream includes each collar number, timestamp, and latitude and longitude coordinates.

[0062] The system sets a historical time period, such as the past six months, and reads all trajectory data within this period. Next, on a geographic information system platform, the system digitizes the geographic extent of the target grassland into a regular grid, with each grid cell measuring 10 meters by 10 meters. For each grid cell, the system iterates through all trajectory points, filters out points falling within its geographic boundary, and calculates the cumulative dwell time of all livestock represented by these points within that grid. For example, if a sheep's trajectory point shows that it spent a total of five hours within a certain grid, these five hours are added to the value of that grid. The dwell time of all livestock is accumulated according to this rule, ultimately generating a digital map called the "Livestock Spatiotemporal Occupation Raster," where the value of each pixel represents the cumulative number of livestock dwell hours within the corresponding geographic grid during the historical time period.

[0063] Subsequently, the system acquires the precise acquisition time point of the plant morphology measurement data. Using this time point as the endpoint, the system defines a backward time window, for example, thirty days. From the aforementioned spatiotemporal occupancy grid, the system extracts the cumulative dwell time value of all grid cells within the target grassland area within the backward time window. Finally, the system sums these time values, and the resulting total cumulative hours are defined as the grazing rate benchmark value for this area under this implementation method, used for training the subsequent supervised learning model. Through the above algorithmic process based on livestock trajectory data, the system can automatically generate a grazing rate benchmark value that precisely matches the plant morphology observation data in terms of time and space. This provides the supervised learning model with an objective, automatically acquireable, high-quality supervisory signal source, effectively solving the core problems of high cost of label data acquisition, large subjective errors, and spatiotemporal mismatch caused by relying on manual recording.

[0064] Current technologies primarily rely on manual recording by herders, fixed-point camera observations, or estimations based on theoretical formulas for carrying capacity to obtain grazing rate data. These technologies typically suffer from problems such as high data subjectivity, low spatiotemporal resolution, high acquisition costs, or a disconnect between theoretical assumptions and actual field conditions. Manual recording is prone to omissions and errors; fixed-point observations have limited fields of view and are difficult to represent the entire area; theoretical formula estimations often ignore the actual spatial distribution and activity intensity differences of livestock. The "grazing rate" data obtained by these methods may be multi-day or monthly averages in terms of time, and a general total in terms of space, making it difficult to accurately match with high spatiotemporal resolution remote sensing observation data of plant morphology, thus introducing significant label noise into model training.

[0065] In contrast, the embodiments of this invention provide an objective, automatic, and high spatiotemporal resolution grazing pressure quantification and matching scheme. It utilizes continuous trajectory data automatically collected by IoT devices as raw input, and through deterministic rasterization calculations and time window extraction algorithms, generates a grazing pressure quantification index that is precisely located in time and spatially decomposed to a grid level of ten meters. This index directly reflects the actual spatial utilization intensity and time investment of livestock, with clear physical meaning. More importantly, through the "time backtracking window" algorithm parameter, it can flexibly explore and match with the response lag of plant morphology, thereby providing high-quality, accurately aligned supervisory labels for machine learning models. This method transforms the acquisition of grazing rate data from relying on manual labor or simple theoretical models into a programmable, repeatable, and high-precision data computation task, significantly improving the reliability of training data and the training effect of the model.

[0066] In this invention, determining the morphological index calculation formula and optimizing the spatial unit size are enhancement and optimization steps of the core process, aiming to improve the model's predictive performance. In practical applications, the execution can be selected based on data conditions and accuracy requirements: empirically preset morphological indices (such as the ratio of plant height to crown width) and spatial scales (such as a 10-meter grid) can be directly used to enter the core training and application process. If better performance is desired, a function selection process can be executed first to determine the optimal morphological index, then a scale optimization process can be executed based on this index to determine the optimal spatial unit size, and finally, the optimized parameter combination can be used to execute the core training and application process. The implementation order of the function selection and scale optimization processes, or whether to simultaneously consider multi-scale feature fusion, can be flexibly configured according to the actual situation without affecting the implementation of the core method of this invention.

[0067] In a preferred embodiment, the method involves determining the morphological index calculation formula through the following function selection steps: predefining multiple candidate functions, each used to calculate a value based on plant height and crown diameter; before the model training phase, executing a function selection process, which includes: dividing the first geographical region into multiple evaluation units using the same preset spatial partitioning rules as in the model training phase; for each candidate function, performing the following evaluation operations: using the candidate function, calculating the plant height and crown diameter of all plant individuals in the training dataset to obtain the candidate morphological index for each plant individual; for each evaluation unit, calculating the candidate statistical features of the unit based on the candidate morphological indices of all plant individuals falling into the unit; calculating the correlation measure between the candidate statistical features of all evaluation units and the corresponding grazing rate benchmark value; and selecting the candidate function with the highest corresponding correlation measure value from all candidate functions as the final morphological index calculation formula used in the model training and application phases.

[0068] One specific implementation of the present invention is as follows: Before formally training the grazing rate estimation model, the system first executes an automated morphological index function selection process. This process predefines a set of candidate mathematical relationships, such as several simple functions including the ratio of plant height to crown diameter, the product of the two, and the approximate volume calculated based on the two.

[0069] After the process begins, the system uses the same preset rules as the subsequent main training process to divide the target grassland into fixed spatial evaluation units. Then, for each candidate function, the system uses the plant height and crown diameter data of all plant individuals in the entire training dataset to calculate the candidate morphological index for each plant individual based on the function.

[0070] Subsequently, for each spatial assessment unit, the system statistically analyzes the candidate morphological indices of all plant individuals falling within it, and calculates the statistical characteristics of that unit, such as the mean and standard deviation of these indices. After calculation, the system obtains a sequence of spatial statistical characteristics generated by the candidate function.

[0071] Finally, the system calculates the statistical correlation between this set of spatial statistical feature sequences and the known grazing rate benchmark sequence, such as the Pearson correlation coefficient. The system repeats the entire process for all candidate functions and compares their respective correlation coefficients. Ultimately, the candidate function that produces the highest correlation coefficient is selected and formally adopted as the morphological index calculation formula for all subsequent model training and application stages. For example, if the feature derived from the ratio function has the highest correlation with grazing rate, then all subsequent steps uniformly use plant height divided by crown diameter as the morphological index. This data-driven automatic function selection process transforms the construction of morphological indices from relying on expert prior experience to an optimization process based on quantitative evaluation, ensuring that the selected combination of morphological features can best represent grazing pressure and overcoming the potential lack of adaptability of manually preset single indicators.

[0072] The formula for calculating morphological indices is determined through the following function selection steps: A set of candidate functions is predefined, each used to calculate a value based on plant height H and crown diameter D; before the model training phase, a function selection process is executed, which includes: dividing the first geographical region into multiple evaluation units using the same preset spatial partitioning rules as in the model training phase; for each candidate function, the following evaluation operations are performed: using the candidate function, the plant height and crown diameter of all plant individuals in the training dataset are calculated to obtain the candidate morphological index for each plant individual; for each evaluation unit, based on the candidate morphological indices of all plant individuals falling into that unit, the candidate statistical features (e.g., the mean of the morphological indices) of that unit are calculated; the correlation measure between the candidate statistical features of all evaluation units and the corresponding grazing rate benchmark is calculated; from all candidate functions, the candidate function with the highest corresponding correlation measure is selected as the final formula for calculating morphological indices used in the model training and application phases.

[0073] For example, the candidate function set may include, but is not limited to, the following forms: Ratio function: f(H,D)=H / D, where H is the plant height and D is the crown diameter.

[0074] Product function: f(H,D) =H×D.

[0075] Area approximation function: f(H,D) =π×(D / 2) 2 (Approximate crown projection area).

[0076] Volume approximation function: f(H,D)=(π×(D / 2)) 2 ) ×H×k, where k is the morphological coefficient (generally taken as 0.5~1.0), used to approximate plant volume.

[0077] Composite function: f(H,D)=(H a ) / (D b ), where a and b are adjustable positive real number parameters, typically ranging from 0.5 to 2.0.

[0078] The system will automatically select the function that is most correlated with the grazing rate benchmark from a predefined set of such functions through the evaluation process.

[0079] The "correlation metric" is calculated using the Pearson correlation coefficient. For each candidate function, the spatial statistical characteristics (such as the mean morphological index) of all evaluation units are first calculated, and then the Pearson correlation coefficient between the characteristic sequence and the corresponding grazing rate benchmark sequence is calculated. r The value range is [−1, 1]. Choose | r The function with the largest correlation (i.e., the strongest correlation) is used as the optimal morphological index for calculating the relationship. If multiple functions have similar correlation coefficients, further screening can be performed using statistical significance tests (e.g., p-value < 0.05).

[0080] Before calculating the morphological index, plant height H and crown diameter D must be standardized to the same unit of length, such as centimeters (cm) or meters (m), to ensure physical consistency of the calculation results. For plants with non-circular crowns, crown diameter D can be defined as the diameter of the equivalent circle of the crown's projected area, or the arithmetic mean of the crown lengths in two perpendicular directions can be used as a substitute. This approach is based on practical measurement feasibility and balances geometric simplification of plant morphology with computational robustness.

[0081] Grazing stress often leads to a decrease in plant height and an increase or change in canopy width. Therefore, the ratio function H / D more sensitively reflects this morphological response trend of "dwarfing and expansion": the higher the grazing intensity, the lower the H / D value is generally. The product function H×D, on the other hand, tends to characterize the overall morphological scale of the plant. By setting a set of candidate functions, the system can automatically compare the sensitivity of different mathematical combinations to grazing signals without relying on a single biological hypothesis beforehand.

[0082] In real-world datasets (such as typical grasslands in northern China), H / D values ​​were sampled and calculated along known grazing intensity gradients (e.g., light, moderate, and heavy grazing areas). Statistical tests (e.g., one-way ANOVA) showed that there were significant differences in H / D values ​​under different grazing intensities (p<0.05), and that the values ​​showed a regular decreasing trend with increasing grazing intensity. This result provides experimental support for the ratio function as one of the candidate functions. The function selection process is based on such datasets, automatically selecting the optimal morphological index form through quantitative evaluation.

[0083] Current technologies rely on the knowledge of domain experts to determine which morphological characteristics might be sensitive to grazing responses. The limitations of this approach are that expert judgments may be based on experience in specific species or environments and are not universally applicable. Furthermore, when faced with multiple potentially relevant parameters such as plant height and canopy width, it is difficult to determine the optimal mathematical combination of these parameters based on experience alone; often, only limited, subjective attempts are possible, lacking systematic and objective comparative standards. This results in the selected characteristics not being the most discriminative form in the current data environment, thus limiting the potential performance of the model from the outset.

[0084] In contrast, the implementation of this invention transforms the morphology index from an "experience-based preset" to a "data-driven selection." It constructs a candidate function set containing multiple possible relationships and designs a standardized evaluation process to quantitatively evaluate the correlation strength between the features generated by each function and the target variable. This process essentially transforms a key hyperparameter selection problem in feature engineering into an automatically executable optimization search problem. It does not rely on specific prior assumptions but allows the data itself to reveal which mathematical relationship between plant height and crown width most effectively represents grazing pressure, thereby improving the objectivity, scientific rigor, and adaptability to specific datasets in the feature construction process, providing a more reliable foundation for obtaining better model performance.

[0085] In a preferred embodiment, in the method, the size of the spatial unit in the preset spatial partitioning rule is determined by the following scale optimization steps: predefining a set of candidate spatial scales; preselecting a benchmark calculation formula for calculating morphological indices; executing a scale evaluation process, which includes: dividing the first geographic region corresponding to the training dataset into candidate spatial unit grids of corresponding sizes according to each candidate spatial scale; for each candidate spatial scale, performing the following operations: calculating the morphological index of each plant individual in the training dataset according to the benchmark calculation formula; for each candidate spatial unit divided at this scale, calculating the candidate statistical features of the unit based on the morphological indices of all plant individuals falling within it; dividing the candidate statistical features of all candidate spatial units obtained at this scale and the corresponding grazing rate benchmark values ​​into a training subset and a validation subset; using the training subset to train a scale evaluation regression model, and using the validation subset to calculate the prediction accuracy index of the scale evaluation regression model; comparing the prediction accuracy indices corresponding to all candidate spatial scales, and determining the candidate spatial scale with the best prediction accuracy index as the spatial unit size adopted by the preset spatial partitioning rule.

[0086] One specific embodiment of the present invention is as follows: After determining the morphological index calculation formula, the system further executes a spatial scale optimization process. First, the system predefines a set of candidate spatial unit side lengths, such as five meters, ten meters, twenty meters, and fifty meters. Simultaneously, the selected morphological index calculation formula is used consistently.

[0087] Subsequently, for each candidate side length, the system divides the entire target grassland area into square grid cells of the corresponding size according to that dimension. For each division scheme, the system counts all index values ​​falling within each grid cell based on the pre-calculated plant individual morphology index, and calculates the statistical characteristics of each cell, such as the average value of the index within the cell.

[0088] Next, the system randomly divides the statistical feature data of all grid cells generated at this scale and their corresponding grazing rate benchmark values ​​into two parts: a large portion of the data serves as the training subset, and a small portion serves as the validation subset. Using the training subset, the system quickly trains a lightweight regression model, such as a linear regression model. Then, this model is used to predict the validation subset, and the root mean square error between the predicted values ​​and the actual grazing rate benchmark values ​​is calculated as the prediction accuracy index for this candidate scale.

[0089] The system repeats the entire process of partitioning, modeling, and validating for each candidate side length. Finally, by comparing the prediction accuracy metrics corresponding to all candidate side lengths, the side length that minimizes the root mean square error of the validation set is determined as the preset spatial partitioning scale for subsequent main model training. This scale optimization method, guided by model prediction performance, provides an objective and quantitative basis for determining the optimal size of spatial statistical units, avoiding the risks of information loss or feature instability caused by fixing the scale based on experience or convention in traditional methods.

[0090] In ecological or remote sensing spatial analysis, a fixed spatial analysis scale is often artificially set based on the size of the study area, data resolution, or domain experience and conventions (e.g., directly using the pixel size of remote sensing images or a fixed hectare grid). When choosing a spatial scale, there is often a lack of direct, quantitative correlation with the final analytical objective. The chosen scale may simply be to maintain consistency with other studies or be limited by the inherent resolution of the raw data, but this scale may not be the optimal granularity for extracting the spatial features most relevant to grazing rates. Inappropriate scale selection can lead to information loss: an excessively large scale smooths out local heterogeneity, while an excessively small scale makes statistical characteristics unstable and increases noise. Due to the lack of objective evaluation criteria, the rationality and optimality of this scale selection method, which relies on experience or convention, are difficult to guarantee, becoming a potential but unverified uncertainty factor affecting model performance.

[0091] In contrast, the implementation of this invention transforms the spatial scale from a fixed parameter that needs to be pre-defined into an optimization variable oriented towards model prediction performance. It systematically constructs spatial features at different candidate scales and utilizes a standardized, validation set-based evaluation process to quantify the effectiveness of features at each scale. This approach directly links scale selection to the final modeling objective, making the determination of the optimal scale quantifiable, repeatable, and goal-oriented. It adaptively seeks the feature aggregation granularity that best reveals the spatial pattern of grazing pressure based on specific data distributions and target variables, thereby enhancing the relevance and scientific rigor of spatial feature engineering and laying a more solid foundation for building high-performance prediction models.

[0092] In a preferred embodiment, the method involves training a regression model to obtain a grazing rate estimation model, specifically implemented through a stacked ensemble training approach, including the following steps: dividing a dataset consisting of spatial statistical features and corresponding grazing rate benchmark values ​​into a first training set and a second training set; base learner training and meta-feature generation steps: using the first training set, a support vector regression model, a gradient boosting tree regression model, and a multilayer perceptron neural network model are trained respectively; subsequently, using the spatial statistical features of the second training set as input, these are input into the three trained models respectively to obtain their respective output prediction values, and the three prediction values ​​corresponding to the same sample in the second training set are combined into a three-dimensional meta-feature vector; meta-learner training steps: using the three-dimensional meta-feature vector as input features and the grazing rate benchmark value corresponding to the second training set as the output label, a linear regression model is trained as the final ensemble grazing rate estimation model.

[0093] The method described above trains a regression model to obtain a grazing rate estimation model, specifically through a stacked ensemble training approach, including the following steps: The dataset partitioning strategy is as follows: to prevent data leakage caused by temporal or spatial correlations, a simple random partitioning is not used; instead, a "spatiotemporal grouping cross-validation" strategy is employed. Specifically, the dataset is sorted according to the collection time and divided into several consecutive groups in chronological order. The first 70% of the groups are used as the first training set, and the last 30% as the second training set. If the data exhibits significant spatial clustering characteristics, it is clustered according to spatial location to ensure that the training and validation sets do not overlap spatially. This partitioning method effectively simulates the model's generalization scenarios when facing new time periods or regions in real-world applications, avoiding performance overestimation due to excessive data correlation.

[0094] The base learners are specifically configured as follows: The base learners include three different types of regression models with pre-tuned parameters, and the specific configurations are as follows: 1. Support Vector Regression (SVR): Radial Basis Function (RBF) is used as the kernel function, the regularization parameter C is set to 1.0, and the kernel coefficient γ is set to 'scale'; 2. Gradient Boosting Tree Regression (GBDT): The maximum tree depth is set to 5, the learning rate is set to 0.1, and the number of trees is set to 100. 3. Multilayer Perceptron Neural Network (MLP): The hidden layer structure is [64, 32], the activation function is ReLU, the optimizer is Adam, and the learning rate is set to 0.001.

[0095] Each base learner is trained independently using the spatial statistical features of the first training set and the corresponding grazing rate benchmark.

[0096] The selection criteria for the meta-learner are as follows: linear regression was chosen as the meta-learner, mainly based on the following three considerations: 1. Prevent overfitting: The input of the meta-learner is the prediction result of the base learner. It has a low feature dimension (three base learners correspond to three-dimensional features), few linear regression parameters, and a simple structure, making it less prone to overfitting on the second training set. 2. Interpretability: The weight coefficients of linear regression can intuitively reflect the contribution of the prediction results of each base learner to the final decision; 3. Computational efficiency: Linear regression has a fast training speed and is suitable for quickly fusing base learner outputs in the second layer of stacked ensembles.

[0097] The coupling method between the stacked ensemble and the aforementioned multi-layer framework is as follows: the entire stacked ensemble model is coupled end-to-end with the aforementioned multi-layer framework of "individual morphological index → ​​spatial statistical features" through feature flow and training process. Spatial statistical features are used as input features for base learners; The supervision signal used during the training phase of the base learner is exactly the same as the grazing rate benchmark value in the aforementioned framework; In the model application phase, the new data processing flow is as follows: plant individual morphological measurement values ​​→ morphological index calculation → spatial statistical feature extraction → base learner prediction → meta learner fusion → final grazing rate output; During training, the parameters of both the base learner and the meta learner are learned through the same training set (divided into the first and second training sets) to ensure that the multi-layer structure remains consistent in its optimization objective.

[0098] The aforementioned stacked ensemble structure captures the nonlinear relationships of different patterns in the data through the diversity of heterogeneous base learners, and then achieves robust fusion through linear meta-learners. This not only enhances the model's ability to fit complex mapping relationships, but also avoids overfitting through a structured training process, thereby improving the generalization performance and stability of the final grazing rate estimation model.

[0099] In predictive modeling, directly selecting a single type of machine learning model based on problem characteristics or experimental comparisons—such as using only random forests or only neural networks—relies on choosing a single model architecture deemed best suited for the current task. The limitation of this approach is that any single model has its inherent inductive bias and capability boundaries. For example, support vector machines excel at handling high-dimensional nonlinear problems with small to medium sample sizes but may be inefficient on large-scale data; gradient boosting trees are friendly to heterogeneous data but may be sensitive to noise; and neural networks have strong fitting capabilities but require large amounts of data and have weak interpretability. When facing potentially unpredictable and complex mapping relationships between grazing rates and complex morphological spatial features, relying on a single model carries the risk of underfitting or poor generalization. The model selection process itself becomes an uncertain process requiring extensive trial and error.

[0100] In contrast, the implementation of this invention abandons the approach of relying on a single optimal model and instead adopts an integrated and hierarchical modeling strategy. It uses multiple heterogeneous base learners in the first layer to model the data from different algorithmic perspectives, aiming to capture diverse latent patterns within the data. The meta-learner in the second layer specifically learns how to optimally fuse these diverse first-layer predictions. This structure is equivalent to constructing a decision-making system composed of multiple experts and a coordinator. It does not attempt to find a universal model, but rather, through architectural design, systematically integrates the advantages of different models and allows the data itself to learn the optimal fusion method, thereby achieving more robust and powerful predictive performance overall than any single model and reducing sensitivity to the initial model selection.

[0101] In a preferred embodiment, after generating the livestock spatiotemporal occupancy raster, the method further includes: performing spatial smoothing filtering on the livestock spatiotemporal occupancy raster to obtain smoothed raster data; and using the smoothed raster data to replace the original livestock spatiotemporal occupancy raster for subsequent extraction of cumulative sums.

[0102] One specific embodiment of the present invention is as follows: After generating the original livestock spatiotemporal occupancy raster, the system automatically performs a data post-processing step to improve the raster quality.

[0103] The system first loads the generated raw raster data into memory, where the value of each pixel in the raster is the initially calculated cumulative dwell time. Then, the system calls a preset spatial smoothing filtering algorithm to process the raster. For example, the system can use a 3x3 Gaussian filter kernel to perform a convolution operation on the raster. During convolution, the value of each central pixel is weighted by a Gaussian average of its own value and the values ​​of its eight neighboring pixels, generating a new, smoothed pixel value. This convolution kernel traverses every pixel of the entire raster in a sliding window manner, ultimately generating a new raster image.

[0104] After this smoothing process, any small speckled noise or sharp numerical abrupt changes that may exist in the original raster due to accidental clustering or missing trajectory points will be smoothed and mitigated to a certain extent, making the numerical change trend on the raster surface appear more continuous and gradual. Finally, the system replaces the original, unprocessed raster with this smoothed raster data for subsequent calculations of the cumulative sum within the time backtracking window. This spatial smoothing filtering effectively suppresses high-frequency noise and data gaps introduced by the uneven spatial distribution of the original GPS trajectory data, generating a more continuous and stable rasterized grazing pressure surface, thereby improving the consistency and reliability of the subsequently extracted grazing rate benchmark value as a model training label.

[0105] For example, a Gaussian low-pass filter, a mean filter, or a median filter with a 3×3 or 5×5 window can be used for convolution operations. Parameters such as the standard deviation of the Gaussian filter and the window size of the mean filter can be preset according to the spatial resolution and noise level of the trajectory data.

[0106] Existing technologies for generating density or intensity distribution maps using trajectory point data typically use raw count or accumulated value rasters directly without any specialized post-processing aimed at improving spatial continuity and reducing noise. The quality of the raster data generated by existing technologies is highly dependent on the spatial uniformity of the original trajectory points. Due to the randomness of livestock movement, GPS positioning errors, and limitations in device sampling frequency, trajectory points are often spatially unevenly distributed, resulting in scattered point clusters or blank areas in the generated rasters, and potentially drastic jumps in values ​​between spatial neighborhoods. This noise not only affects the visual appeal of the raster data, but more importantly, as data-level noise, it is directly incorporated into all subsequent calculations based on this raster, introducing undue random error into the final generated grazing rate baseline label, thus reducing the data quality and reliability of the supervision label.

[0107] In contrast, the embodiments of this invention identify and specifically address the spatial noise problem introduced by defects in the original data acquisition. It introduces mature and robust spatial filtering techniques from digital image processing into the post-processing workflow of spatiotemporal raster data. This step does not alter the macroscopic spatial pattern and total volume of the data, but rather purposefully suppresses high-frequency noise and anomalous mutations that are spatially unrepresentative due to data sparsity and randomness through local weighted averaging, thereby enhancing the spatial continuity of the raster data. This processing is equivalent to denoising and smoothing the original observation data, improving the consistency and stability of the derived grazing rate benchmark value, and providing a cleaner and more reliable supervisory signal foundation for the training of downstream machine learning models.

[0108] In a preferred embodiment, the method, in the step of calculating the livestock spatiotemporal occupancy grid, includes the following steps for calculating the cumulative value of each pixel within each unit time interval: counting the number of different livestock individuals located within the geographic range of the pixel within the unit time interval; using the number of different livestock individuals located within the geographic range of the pixel within the unit time interval as input, calculating a pressure weight value through a nonlinear transformation function, wherein the nonlinear transformation function is configured such that its output pressure weight value increases in a superlinear manner as the number of inputs increases; multiplying the pressure weight value by the unit time length as the effective grazing pressure contribution value of the pixel within the unit time interval; and summing the effective grazing pressure contribution values ​​calculated in all unit time intervals to obtain the final value of the corresponding pixel in the livestock spatiotemporal occupancy grid.

[0109] One specific embodiment of the present invention is as follows: When calculating the spatiotemporal occupancy grid based on livestock trajectory data, the method for calculating the contribution of each grid cell within a specific time unit is improved. The system first uses one hour as a basic unit time interval.

[0110] For each unit time interval, the system counts all livestock trajectory points within the geographic range of the target pixel within that hour, and performs deduplication based on the individual identifiers they carry to obtain the number of different livestock individuals within that pixel within that hour.

[0111] The system then uses this quantity as input and applies a preset non-linear transformation rule. The core characteristic of this rule is that the output pressure weight value increases at a rate greater than linear with the increase in the input quantity; for example, the input quantity is squared to calculate the weight. Using this rule, the system calculates the grazing pressure weight value for that pixel within the current hour.

[0112] Next, the system multiplies this pressure weight value by the length of a unit of time. Since the unit of time is one hour, the product is numerically equal to the weight value, but its physical meaning is the effective grazing pressure contribution value borne by the pixel in the current hour.

[0113] Finally, the system sums up the effective grazing pressure contribution values ​​of each pixel calculated within all unit time intervals. This sum is the final value for that pixel. Once all pixels have been calculated according to this rule, they collectively form a new and improved livestock spatiotemporal occupancy grid. By introducing a nonlinear transformation function to superlinearly weight the number of livestock, this improved method more accurately simulates the group superposition effects such as competition and trampling that may occur when multiple livestock are present simultaneously. This makes the calculated grazing pressure index more reasonable in an ecological sense and provides more discriminative and information-rich supervisory labels for model learning.

[0114] For example, the nonlinear transformation function can be in the form of a power function: W(n) = n α Where n is the number of livestock and α is an exponent greater than 1 (e.g., α = 1.5, 2.0) to ensure that the output value grows superlinearly with the input. Piecewise functions or saturated growth functions (e.g., combinations of logarithmic functions) can also be used to simulate more complex group effects.

[0115] The method of simply summing the contributions of different individuals at the same time and place when calculating animal spatiotemporal occupancy or density is essentially linear; it directly sums the dwell times of different individuals. This method implicitly assumes that the ecological impact of multiple livestock in the same location equals the simple sum of their individual impacts. However, ecological research shows that when animal groups move within a limited space, the combined ecological pressure generated by the interactions of foraging competition, trampling, and social behavior often exceeds the simple arithmetic sum of individual impacts, exhibiting a non-linear "superadditive" effect. The linear summation method fails to characterize this true group effect, leading to a potential systematic underestimation of the actual ecological pressure caused by intensive grazing.

[0116] In contrast, the embodiments of this invention translate ecological knowledge into concrete algorithmic rules. By introducing a well-designed nonlinear transformation step into the computational chain, it enables the computational model to simulate the herd effect where grazing pressure accelerates with increasing livestock numbers. This improvement allows the generated spatiotemporal occupancy grid to not only quantify livestock activity in the temporal and spatial dimensions but also to more reasonably characterize the impact of livestock allocation in the intensity dimension. The resulting grazing rate benchmark, used as a model label, contains richer and more ecologically consistent information about grazing intensity, helping machine learning models learn more realistic stress-response relationships, thereby improving the ecological rationality and accuracy of predictions.

[0117] In a preferred embodiment, the step of calculating the correlation metric in the method specifically includes: randomly dividing the training dataset into a training subset and a validation subset; performing an evaluation operation using the training subset to obtain a first correlation metric between the candidate statistical feature and the corresponding grazing rate benchmark value; performing an evaluation operation using the validation subset to obtain a second correlation metric between the candidate statistical feature and the corresponding grazing rate benchmark value; and using the weighted average of the first correlation metric and the second correlation metric as the final correlation metric of the candidate function.

[0118] One specific implementation of the present invention is as follows: In the function selection process, when it is necessary to calculate the final relevance metric of each candidate function, the system adopts a robust evaluation strategy that combines the training set and the validation set.

[0119] Specifically, the system first randomly divides the complete training dataset used for evaluation into two parts: one part as the training subset and the other part as the validation subset, for example, in a 7:3 ratio.

[0120] For each candidate function to be evaluated, the system first performs a complete evaluation operation using a training subset. That is, it uses the candidate function to calculate the morphological index of plant individuals in the training subset, thereby generating spatial unit statistical features, and calculates the correlation between these features and the corresponding grazing rate benchmark value of the training subset to obtain the first correlation metric.

[0121] Next, the system independently performs the same complete evaluation operation again using the validation subset. That is, it uses the same candidate function to calculate the morphological index of plant individuals in the validation subset, generates the corresponding spatial unit statistical features, and calculates the correlation between these features and the corresponding grazing rate benchmark value of the validation subset to obtain the second correlation metric.

[0122] Finally, the system performs a weighted average of the two correlation metrics, assigning 70% weight to the first metric and 30% weight to the second, to calculate a comprehensive final correlation metric. This value represents the overall performance of the candidate function and is used for subsequent comparison and selection with other candidate functions. This weighted evaluation strategy, combining the training set and independent validation set, effectively reduces the risk of overfitting caused by the randomness of a single data partition by simultaneously considering the fitting potential and generalization performance of the candidate function. This makes the selection process of the morphological exponential function more robust, and the selected function is more likely to maintain good performance on new data.

[0123] When performing feature selection or hyperparameter tuning, evaluation metrics are often calculated and directly compared on a single dataset. For example, the correlation coefficients between different features and labels are calculated on a single dataset, and the highest one is selected. This method carries a significant risk: the selected features may be "overfitted" to the random noise or specific distribution patterns of that particular dataset. This is because the evaluation process does not distinguish between the feature's performance in "learning" the intrinsic relationships in the data and "memorizing" the randomness of the data. Directly relying on evaluation results on a single dataset can easily lead to overfitting, causing the selected features or parameters to significantly deteriorate on new data, thus impairing the model's generalization ability.

[0124] In contrast, the implementation of this invention addresses this problem by introducing a structured "training-validation" evaluation framework. It divides the dataset into training and validation subsets and calculates the performance of candidate functions on each subset separately. Evaluation using the training subset reflects the function's fitting potential, while evaluation using the independent validation subset verifies its generalization ability. Finally, a weighted average is used to synthesize the information from both subsets. This method essentially adds a "generalization ability test" to the feature selection process. By simulating new data scenarios the model might encounter in the future, it provides a more robust and comprehensive basis for decision-making, thus helping to select feature calculation methods that not only perform well on current data but are also more likely to maintain stable performance on future data. This improves the reliability of the entire feature engineering process and the generalization robustness of the selected features.

[0125] In a preferred embodiment, the regression model in the method is a multi-scale feature fusion regression model. Its training method is as follows: for a first geographical region, multiple different candidate spatial scales are used to divide it into multiple sets of spatial units; for each set of spatial units, spatial statistical features at that scale are calculated based on the morphological index of the individual plants falling within it; the spatial statistical features calculated at different scales for the same geographical location are concatenated to form a multi-scale fusion feature vector; using the multi-scale fusion feature vector as input features and the grazing rate benchmark value as a supervision label, a multi-scale feature fusion regression model is trained, serving as the final grazing rate estimation model.

[0126] One specific implementation of the present invention is as follows: During the model training phase, the system adopts a multi-scale feature fusion strategy to construct the final regression model, rather than using only a single-scale spatial statistical feature.

[0127] First, the system predefines a set of candidate spatial scales, such as side lengths of five meters, ten meters, twenty meters, and fifty meters. For the target geographic region, the system no longer divides the area using only a single scale, but instead uses all of the above candidate scales to divide it independently, thus obtaining four sets of corresponding spatial unit grids.

[0128] Next, for each grid at a specific scale, the system, based on pre-calculated plant individual morphology indices, counts all index values ​​falling within each cell and calculates the spatial statistical characteristics of that cell, such as the average value. Thus, four sets of spatial statistical characteristic data at different scales are generated for the same geographical area.

[0129] Subsequently, the system performs feature fusion. For each geographic location, the system extracts its spatial statistical feature values ​​calculated at four scales: five meters, ten meters, twenty meters, and fifty meters, and concatenates these four values ​​in sequence to form a multi-scale fused feature vector.

[0130] Finally, the system uses this multi-scale fused feature vector as input features and the corresponding grazing rate benchmark value as supervision label, and employs a regression algorithm, such as a gradient boosting tree or neural network, to train the model. The resulting model is the multi-scale feature fusion regression model, which can simultaneously accept and process feature information from different scales. By constructing a multi-scale feature fusion model, this method enables the model to simultaneously capture multi-scale spatial patterns, such as local details and overall layout, avoiding the loss of key information that may occur with a fixed single scale, thereby enhancing the overall ability to characterize complex spatial distribution patterns.

[0131] This involves stitching together spatial statistical features calculated at different scales for the same geographical area (or the same set of plant individual data) to form a multi-scale fused feature vector. Specifically, spatial indexing relationships can be established between grids of different scales, and feature values ​​of multiple scale units that overlap or contain each other in space can be associated and stitched together.

[0132] In spatial data analysis, a fixed and singular spatial scale is typically chosen to generate features and build predictive models, based on prior knowledge or resolution limitations. The limitation of this approach is that spatial patterns and ecological processes often manifest and interact simultaneously at different spatial scales. For example, subtle livestock grazing traces may be clearly captured at a smaller scale, while changes in vegetation gradient across an entire slope aspect require a larger scale to summarize. Artificially selecting a single scale is tantamount to forcibly filtering out complementary information that may be present at other scales before modeling even begins. Regardless of the chosen scale, there is an inherent risk of information loss, and such selection is often based on subjective experience, making its rationality difficult to verify. Consequently, the model loses its ability to utilize multi-scale information for collaborative decision-making.

[0133] In contrast, the implementation of this invention changes the paradigm of feature construction. It no longer treats scale as a pre-selected "single-choice parameter," but rather as a series of "multi-perspective information sources" that can be provided in parallel. By simultaneously calculating and fusing spatial features at multiple scales, the model gains the ability to represent spatial heterogeneity across scales. The model itself can automatically learn how to balance and integrate signals at different scales within this fused feature space through the training process. This approach is equivalent to equipping the model with multiple sets of observation lenses at different magnifications, enabling it to simultaneously perceive local details and overall patterns. This promises to capture complex spatial patterns related to grazing pressure more comprehensively and robustly, avoiding the loss of crucial information due to inappropriate selection of a single scale.

[0134] In a preferred embodiment, in the method, during the base learner training and meta-feature generation steps: a set of random noise features is added to the spatial statistical features of the first training set to form three different enhanced feature sets; the three enhanced feature sets are used as input features for training the support vector regression model, the gradient boosting tree regression model, and the multilayer perceptron neural network model, respectively; when generating meta-feature vectors using the second training set, the same random noise features as before are added to the spatial statistical features of the second training set, and then input to the corresponding base learner to obtain the predicted value.

[0135] One specific embodiment of the present invention is as follows: In the stacked ensemble training process, before performing base learner training, the system performs differential enhancement processing on the input features.

[0136] Specifically, after obtaining the spatial statistical feature matrix of the first training set, the system generates corresponding augmented feature sets for the three base learners to be trained. The system first generates three independent random number matrices, each with the same dimension as the original feature set, where each value is randomly sampled from a standard normal distribution with a mean of zero and a small variance. For example, the standard deviation of the noise can be set to 5% to 20% of the statistical value of the standard deviation of the corresponding feature in the original feature set. Subsequently, the system performs element-wise addition operations on the original spatial statistical feature matrix with these three random number matrices, thereby forming three augmented feature sets that are slightly different in value but share the same basic information.

[0137] Next, the system dedicates the first augmented feature set to training the support vector regression model, the second augmented feature set to training the gradient boosting tree regression model, and the third augmented feature set to training the multilayer perceptron neural network model. Each base learner completes independent training using its own dedicated augmented feature set.

[0138] When it is necessary to generate meta-feature vectors using the second training set, the system employs the exact same processing method: the spatial statistical feature matrix of the second training set is added to the three previously generated and saved identical random number matrices to obtain three corresponding enhanced feature sets. These are then input into the corresponding pre-trained base learners to obtain predicted values, which are combined into a three-dimensional meta-feature vector for subsequent meta-learner training. By adding differentiated but fixed random noise features to different base learners, this technique proactively increases the input diversity among base learners at a low cost, prompting them to learn from slightly different data perspectives, thereby reducing the correlation of their prediction errors and ultimately helping to improve the overall generalization performance of the stacked ensemble model.

[0139] When building stacked ensemble models, the exact same training data can be simultaneously input into all base learners for training. While this method aggregates different types of base learners, because their input features are completely identical, these learners tend to learn similar patterns in the data and associate similar errors. This can lead to high correlations between the predicted outputs of the base learners, weakening the core advantage of ensemble learning: reducing variance and improving generalization ability by combining diverse weak learners. If the predictions of the base learners are highly consistent, the differences in information that the meta-learners can learn and fuse are very limited, and the overall performance of the ensemble model may only slightly outperform the single best-performing base learner, making it difficult to achieve significant performance improvements.

[0140] In contrast, the embodiments of the present invention provide a low-cost technique for actively intervening in the diversity of base learner inputs. By adding fixed but different random noise to the original features, the system creates a slightly differentiated, slightly perturbed data perspective for each base learner. This perturbation forces each base learner to avoid simply memorizing completely consistent global patterns during training, and instead to capture stable, essential mapping relationships between features and the target even when faced with slightly different inputs. This process effectively reduces the correlation of prediction errors among different base learners and increases the diversity of their prediction results. This allows subsequent meta-learners to access richer and more complementary prediction information, enabling more effective fusion decisions and ultimately making the entire stacked ensemble model more likely to achieve more robust prediction performance than any single base learner.

[0141] To verify the effectiveness of this invention, a comparative experiment was conducted using a measured dataset from a typical grassland in northern China during the growing season. The experimental area was divided into three gradients: light, moderate, and heavy grazing. A total of 1200 plant individual morphological data (plant height, crown width) samples were collected, and livestock GPS trajectory data for the corresponding areas were acquired simultaneously. The dataset was divided into training and test sets in chronological order to ensure temporal independence. The baseline models for comparison included: (1) a linear regression model based on the average plant height of the community; and (2) a random forest regression model based on the Sentinel-2 satellite NDVI vegetation index. All models were evaluated using the same training and test sets, and the evaluation metrics were root mean square error (RMSE) and coefficient of determination (R²).

[0142] Experimental results show that the model constructed by the method of this invention has an RMSE of 0.14 and an R² of 0.91 on the test set. In comparison, the baseline model 1 (average plant height) has an RMSE of 0.29 and an R² of 0.60; the baseline model 2 (NDVI vegetation index) has an RMSE of 0.33 and an R² of 0.55. Compared with baseline model 1, the method of this invention reduces the RMSE by approximately 51.7% and increases the R² by approximately 51.7%; compared with baseline model 2, the RMSE reduces by approximately 57.6% and increases the R² by approximately 65.5%.

[0143] To further verify the model's robustness, the experiment was repeated after adding 10% Gaussian noise to the training data. The results show that the RMSE of the method described in this invention increased by only 8.3% on noisy data (from 0.14 to 0.152), while the RMSE of baseline model 1 and baseline model 2 increased by 22.1% and 31.4%, respectively. This indicates that the method described in this invention has stronger anti-interference ability and stability in terms of feature extraction and model structure. The above experimental results fully demonstrate that the method described in this invention has significant beneficial effects in improving the accuracy of grazing rate prediction and enhancing model robustness.

[0144] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. It can be applied to various fields suitable for the present invention. Other modifications can be readily implemented by those skilled in the art. Therefore, the present invention is not limited to the specific details without departing from the general concept defined by the claims and their equivalents.

Claims

1. A data processing method for predicting grazing rates based on plant morphological characteristics, characterized in that, This includes the model training phase and the model application phase: The model training phase includes: Receive a training dataset, which includes morphological measurements of multiple plant individuals collected within a first geographic region. The morphological measurements include at least plant height and crown diameter. Receive a grazing rate baseline corresponding to the first geographic region. For each plant individual, calculate a morphological index based on its plant height and crown diameter. According to the preset spatial division rules, the first geographical region is divided into multiple first spatial units, and the spatial statistical characteristics of each unit are calculated based on the morphological index of all plant individuals falling into each first spatial unit. The spatial statistical characteristics of each first spatial unit are used as input features, and the grazing rate benchmark value is used as the supervision label to train a regression model and obtain a grazing rate estimation model. The model application phase includes: Receive the test dataset of the region to be evaluated, which includes morphological measurements of multiple plant individuals; for each plant individual in the test dataset, calculate the morphological index based on its plant height and crown diameter; According to the preset spatial division rules, the area to be evaluated is divided into multiple second spatial units, and the spatial statistical characteristics of the unit are calculated based on the morphological index of all plant individuals falling into each second spatial unit. The spatial statistical characteristics of each second spatial unit are input into the grazing rate estimation model, and the corresponding estimated grazing rate is output.

2. The grazing rate prediction data processing method based on plant morphological characteristics according to claim 1, characterized in that, Receive the grazing rate baseline value corresponding to the first geographic region, specifically including: Obtain livestock trajectory data for the first geographic region within a preset historical time period. The livestock trajectory data includes livestock location information with timestamps. Based on livestock trajectory data, the spatiotemporal occupancy grid of livestock in the first geographic area within a preset historical time period is calculated. The value of each cell in the grid is the cumulative dwell time of all livestock within the geographic range of that cell during the preset historical time period. Time points for obtaining plant morphological measurement values; A time backtracking window is defined backward from the point of data collection; Extract all cell values ​​corresponding to the first geographic region from the livestock spatiotemporal occupancy raster, and calculate the cumulative sum of these cell values ​​within the time backtracking window. Use the cumulative sum as the grazing rate benchmark value.

3. The grazing rate prediction data processing method based on plant morphological characteristics according to claim 1, characterized in that, The formula for calculating the morphological index is determined through the following function selection steps: Multiple candidate functions are predefined, each of which is used to calculate a value based on plant height and crown diameter; Prior to the model training phase, a function selection process is executed, which includes: Using the same pre-defined spatial partitioning rules as in the model training phase, the first geographic region is divided into multiple evaluation units; For each candidate function, perform the following evaluation operation: Using this candidate function, the plant height and crown diameter of all plant individuals in the training dataset are calculated to obtain the candidate morphological index for each plant individual; For each evaluation unit, candidate statistical characteristics of the unit are calculated based on the candidate morphological indices of all plant individuals falling into that unit. Calculate the correlation measure between the candidate statistical characteristics of all evaluation units and the corresponding grazing rate benchmark; From all candidate functions, select the candidate function with the highest corresponding correlation metric value as the final morphological index calculation formula used in the model training and application stages.

4. The grazing rate prediction data processing method based on plant morphological characteristics according to claim 1, characterized in that, In the preset spatial partitioning rules, the size of the spatial unit is determined through the following scale optimization steps: A predefined set of candidate spatial scales; A benchmark calculation formula for calculating the morphology index is pre-selected; The implementation of a metric assessment process includes: The first geographic region corresponding to the training dataset is divided into candidate spatial unit grids of corresponding sizes according to each candidate spatial scale. For each candidate spatial scale, perform the following operations: Based on the baseline calculation formula, calculate the morphological index of each plant individual in the training dataset; For each candidate spatial unit at this scale, the candidate statistical characteristics of the unit are calculated based on the morphological indices of all plant individuals falling within it. The candidate statistical features of all candidate spatial units obtained according to this scale and their corresponding grazing rate benchmark values ​​are divided into training subsets and validation subsets. A scale-evaluation regression model is trained using a training subset, and a prediction accuracy metric for the scale-evaluation regression model is computed using a validation subset. Compare the prediction accuracy indices corresponding to all candidate spatial scales, and determine the candidate spatial scale with the best prediction accuracy index as the spatial unit size adopted by the preset spatial division rule.

5. The grazing rate prediction data processing method based on plant morphological characteristics according to claim 1, characterized in that, The regression model is trained to obtain the grazing rate estimation model, which is specifically achieved through a stacked ensemble training method, including the following steps: The dataset consisting of spatial statistical features and corresponding grazing rate benchmarks is divided into a first training set and a second training set. Base learner training and meta-feature generation steps: Using the first training set, support vector regression model, gradient boosting tree regression model and multilayer perceptron neural network model are trained respectively; then, the spatial statistical features of the second training set are used as inputs and input into the three trained models respectively to obtain the predicted values ​​of their respective outputs. The three predicted values ​​corresponding to the same second training set sample are combined into a three-dimensional meta-feature vector. Meta-learner training steps: Using the three-dimensional meta-feature vector as input features and the grazing rate benchmark value corresponding to the second training set as output labels, train a linear regression model as the final ensemble grazing rate estimation model.

6. The grazing rate prediction data processing method based on plant morphological characteristics according to claim 2, characterized in that, After generating the livestock spatiotemporal occupancy raster, the following is also included: Spatial smoothing filtering is applied to the livestock spatiotemporal occupancy raster to obtain smoothed raster data; Smoothed raster data is used to replace the original livestock spatiotemporal occupancy raster for subsequent extraction of cumulative sums.

7. The grazing rate prediction data processing method based on plant morphological characteristics according to claim 2, characterized in that, In the step of calculating the spatiotemporal occupancy raster of livestock, the cumulative value of each pixel within each unit time interval is calculated, including: Count the number of different livestock individuals located within the geographic range of this pixel within this unit of time interval; Using the number of different livestock individuals within the geographic range of the pixel in the unit time interval as input, a pressure weight value is calculated through a nonlinear transformation function, wherein the nonlinear transformation function is configured to increase the output pressure weight value in a superlinear manner as the number of inputs increases; The product of the pressure weight value and the unit time length is taken as the effective grazing pressure contribution value of the pixel within the unit time interval. The effective grazing pressure contribution values ​​calculated within all unit time intervals are summed to obtain the final value of the corresponding cell in the livestock spatiotemporal occupation raster.

8. The grazing rate prediction data processing method based on plant morphological characteristics according to claim 3, characterized in that, The specific steps for calculating the correlation metric are as follows: The training dataset is randomly divided into a training subset and a validation subset; The evaluation operation is performed using a subset of the training set to obtain the first correlation measure between the candidate statistical features and the corresponding grazing rate benchmark. The evaluation operation is performed using a validation subset to obtain a second correlation measure between candidate statistical features and the corresponding grazing rate benchmark. The weighted average of the first and second correlation measures is used as the final correlation measure of the candidate function.

9. The grazing rate prediction data processing method based on plant morphological characteristics according to claim 4, characterized in that, The regression model is a multi-scale feature fusion regression model; The training method is as follows: For the first geographic region, multiple different candidate spatial scales are used to divide it, resulting in multiple sets of spatial units; For each spatial unit, the spatial statistical characteristics at that scale are calculated based on the morphological index of the individual plants that fall into it. Spatial statistical features calculated at different scales for the same geographical location are spliced ​​together to form a multi-scale fused feature vector. Using multi-scale fused feature vectors as input features and grazing rate benchmark values ​​as supervision labels, a multi-scale feature fusion regression model is trained, which serves as the final grazing rate estimation model.

10. The grazing rate prediction data processing method based on plant morphological characteristics according to claim 5, characterized in that, In the base learner training and meta-feature generation steps: For the spatial statistical features of the first training set, a set of random noise features is added to each set to form three different enhanced feature sets; The three enhanced feature sets were used as input features for training the support vector regression model, gradient boosting tree regression model, and multilayer perceptron neural network model, respectively. When generating meta-feature vectors using the second training set, the same random noise features as before are added to the spatial statistical features of the second training set, and then input into the corresponding base learner to obtain the predicted value.

Images