Multi-source data fusion machine learning grassland grass yield inversion method

Abstract

The invention provides a grassland grass yield inversion method based on multi-source data fusion and a machine learning algorithm, and aims to improve the estimation precision and efficiency of grassland biomass. According to the method, satellite remote sensing data, unmanned aerial vehicle high-resolution images, climate data acquired by a meteorological station and a digital elevation model are integrated, key features are extracted by using a feature selection technology, and model training and optimization are performed by using machine learning algorithms such as random forest, a support vector machine and extreme gradient lifting. Through model optimization strategies such as cross validation and grid search, the generalization ability and stability of the model are ensured. According to the method, real-time inversion prediction of the grassland grass yield can be realized, a scientific basis is provided for grassland ecological protection, grass and livestock balance management and animal husbandry production, and the method has important practical application value and wide market prospect.

Description

Technical Field

[0001] The present invention relates to the field of grassland ecological monitoring and biomass estimation, and in particular to a grassland grass yield inversion method combining multi-source data fusion technology with a machine learning algorithm. Background Art

[0002] With the global population growth and ecological changes, grasslands, as important terrestrial ecosystems, have attracted extensive attention for their productivity and ecological functions. Accurate estimation of grassland grass yield is of great significance for grassland ecological protection, animal husbandry development and biodiversity maintenance. Traditional methods for estimating grass yield mainly rely on ground measurements, which are not only time-consuming and labor-intensive, but also difficult to achieve large-scale continuous monitoring.

[0003] In recent years, the development of remote sensing technology has provided new technical means for monitoring grassland grass yield. Satellite remote sensing data such as MODIS are widely used in remote sensing estimation of grassland grass yield due to their high temporal resolution and wide coverage. By analyzing the correlation between vegetation index and grass yield, a remote sensing estimation model can be established and spatial inversion of grass yield can be performed. However, a single data source is often difficult to meet the needs of high-precision inversion. Therefore, integrating multi-source data, such as meteorological data, elevation models, and drone images, can improve the accuracy and robustness of the inversion model.

[0004] The development of machine learning algorithms has provided powerful tools for the analysis of remote sensing data. Algorithms such as random forests, support vector machines, and extreme gradient boosting have been used to extract grass yield information from remote sensing data. These algorithms are able to handle a large number of variables and identify complex relationships in the data, thereby improving the accuracy of grass yield estimation. Although some progress has been made in grassland grass yield inversion, some challenges still exist. For example, how to fully integrate multi-source data to improve inversion accuracy, how to optimize machine learning models to adapt to different grassland types and environmental conditions, and how to achieve real-time monitoring and prediction of grass yield. In addition, the estimation of grassland grass yield needs to consider not only biomass, but also the ecological function and sustainability of the grassland.

[0005] This paper aims to propose a new grassland grass yield inversion method by integrating multi-source data and applying advanced machine learning algorithms. This method can improve the accuracy and efficiency of grass yield estimation and provide a scientific basis for the rational use of grassland resources and ecological protection. Summary of the invention

[0006] The present invention proposes a grassland grass yield inversion method based on multi-source data fusion and machine learning algorithm, aiming to improve the estimation accuracy and efficiency of grassland biomass. The method is implemented by the following steps:

[0007] S1: Collect multi-source data of the grassland area, including MODIS satellite data, high-resolution drone image data, climate data collected by at least three meteorological stations, and digital elevation model data. The temporal resolution of MODIS satellite data is 16 days and the spatial resolution is 250 meters. The spatial resolution of drone image data is not less than 0.1 meters. Climate data includes temperature, precipitation and humidity. The spatial resolution of digital elevation model data is 30 meters.

[0008] S2: Preprocess the collected data, including using the threshold filtering method to clean the data, normalize the data to the [0, 1] interval, and extract key features, including the normalized vegetation index (NDVI), enhanced vegetation index (EVI), land surface temperature (LST) and altitude. The calculation formulas of NDVI and EVI are NDVI = (ρ NIR -ρ Red ) / (ρ NIR +ρ Red ) and EVI = 2.5 × (ρ NIR / ρ Red ), where ρ NIR and ρ Red represent the reflectance in the near-infrared and red bands, respectively;

[0009] S3: using a feature selection algorithm, such as a recursive feature elimination (RFE) algorithm based on information gain, to select the features that have the greatest impact on grass yield inversion, ensuring that the correlation coefficient of the selected features is greater than 0.5, and the number of selected features does not exceed 80% of the total number of features;

[0010] S4: Use one or more of the random forest, support vector machine (SVM) or extreme gradient boosting (XGBoost) algorithms, combined with the feature selection results, to train the grass yield inversion model. The model training is validated using at least two independent data sets, one of which is used for model training and the other for model validation. The number of samples in the validation set is not less than 20% of the total number of samples.

[0011] S5: Optimize model parameters through cross-validation and grid search to improve the generalization ability and inversion accuracy of the model. The optimization process includes adjusting parameters such as tree depth, minimum sample split, and learning rate. The depth of the random forest tree is set in the range of 10 to 20, the minimum sample split number is set in the range of 2 to 20, and the learning rate of XGBoost is set in the range of 0.01 to 0.2.

[0012] S6: Use the trained model and the latest multi-source data to perform inverse prediction of grassland grass yield. The inverse prediction includes but is not limited to monthly and seasonal grass yield estimates. The prediction results are expressed in dry matter weight. The dry matter conversion factor is determined based on historical data. The conversion factor ranges from 0.8 to 0.95.

[0013] S7: Analyze the inversion results, evaluate the spatial distribution and temporal changes of grass yield, generate spatial distribution maps and time series maps of grassland grass yield, and visualize them through GIS software. The resolution of the spatial distribution map should not be less than 250 meters, and the time interval of the time series map should not exceed 1 month.

[0014] The data collection and preprocessing (S1) involves collecting multi-source data from the grassland area, which is the basis of the grassland grass yield inversion method. The specific operation is as follows: First, MODIS satellite data is collected, and its temporal resolution is set to 16 days to ensure that the periodic changes in vegetation growth can be captured, and the spatial resolution is 250 meters, so as to cover a larger grassland area and maintain the spatial continuity of the data. At the same time, a high-resolution camera or multispectral sensor carried by an unmanned aerial vehicle is used to obtain high-resolution image data with a spatial resolution of not less than 0.1 meters to capture the fine structure and details of grassland vegetation. In addition, data on climate parameters including temperature, precipitation, humidity and solar radiation are collected from at least three meteorological stations, which are essential for understanding the climatic conditions for grassland growth. Digital elevation model (DEM) data is obtained through GPS and other remote sensing technologies to provide topographic information of the grassland area. In the data preprocessing (s2) stage, the threshold filtering method is used to denoise and remove outliers from satellite and unmanned aerial vehicle image data to ensure data quality. All data will be normalized and converted to the [0,1] interval to eliminate the dimensional effects between different data sources. Finally, key features were extracted, including the Normalized Difference Vegetation Index (NDVI) and Enhanced Vegetation Index (EVI) calculated by analyzing the reflectance of the red and near-infrared bands, the Land Surface Temperature (LST) calculated by the radiance of the thermal infrared band, and the altitude directly extracted from the DEM data. The extraction of these characteristic parameters provides the necessary input variables for the subsequent grassland grass yield inversion model.

[0015] In the feature selection step (S3), we use a recursive feature elimination (RFE) algorithm based on information gain to identify and select the features that have the greatest impact on the grassland grass yield inversion model. The specific process is as follows: First, we calculate the information gain between each feature (such as NDVI, EVI, LST and altitude) and grass yield. Information gain is an indicator to measure the correlation between a feature and a label (grass yield), which reflects the degree of reduction in label uncertainty under the condition of known feature information. The larger the information gain, the stronger the correlation between the feature and the label. According to the calculated information gain value, we sort all features from high to low according to the gain value. This ensures that we prioritize those features that are most helpful for grass yield prediction. The RFE algorithm builds a model by recursively eliminating the least important features. In each round of recursion, the algorithm trains a base model (such as random forest, SVM or XGBoost) and determines the importance of the feature based on the coefficients of the model or the feature importance. Then, the least important feature (i.e., the feature with the lowest information gain) is removed and the model is retrained on the remaining feature set. In order to ensure that the selected features have a strong linear relationship with grass yield, we further screened for features with a correlation coefficient greater than 0.5. The correlation coefficient is a statistical indicator that measures the strength of the linear relationship between two variables, and its value ranges from -1 to 1. A correlation coefficient greater than 0.5 indicates a strong positive correlation between the feature and grass yield. On the basis of ensuring the correlation between the selected features and grass yield, we also control the number of selected features to ensure that it does not exceed 80% of the total number of features. This step helps prevent model overfitting and improves the generalization ability of the model.

[0016] In the model training step (S4), we use random forest, support vector machine (SVM) or extreme gradient boosting (XGBoost) algorithm to train the grassland grass yield inversion model through the following detailed process: First, we divide the feature selected data set into two independent parts: training set and validation set. Usually, we use 80% of the data as the training set and the remaining 20% ​​of the data as the validation set. This division ratio is designed to ensure that the model has enough data to learn the relationship between grass yield and features during the training phase, and enough data to evaluate the generalization ability of the model during the validation phase. Random forest is an ensemble learning method that improves the accuracy of prediction by building multiple decision trees and outputting average results. During the training process, we set the number of decision trees in the forest to 100, and each tree randomly selects a feature subset from the original feature set for splitting during training to increase the diversity of the model and reduce the risk of overfitting. SVM is a powerful classification technique that distinguishes different categories by finding the optimal hyperplane in a high-dimensional space. In this step, we choose radial basis function (RBF) as the kernel function, which is a common choice because it can handle nonlinear relationships. We use grid search and cross-validation to optimize two key parameters of SVM: the regularization parameter C and the kernel function parameter γ. XGBoost is an efficient gradient boosting framework that improves the performance of the model by gradually adding trees to correct the errors of the previous tree. We set the depth of XGBoost to 6, the learning rate to 0.1, and the default tree structure parameters. The optimization process of XGBoost involves splitting the features to find the best feature combination and splitting point. After the model training is completed, we use an independent validation set to evaluate the performance of the model. This includes calculating the root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R 2 ). These metrics provide a comprehensive picture of the model’s prediction accuracy and fit. Based on the performance metrics on the validation set, we may need to go back and adjust the model’s parameters to further improve the model’s accuracy. This process may involve adjusting the number of decision trees in a random forest, optimizing the C and γ parameters of an SVM, or adjusting the learning rate and tree depth of XGBoost.

[0017] In the model optimization step (S5), we use methods such as cross-validation and grid search to carefully optimize the model parameters to improve the generalization ability and inversion accuracy of the model. The specific process includes the adjustment of the following key parameters: Use grid search (GridSearchCV) to optimize the `n_estimators` parameter of the random forest model, test different numbers of decision trees, such as 100, 200, 300, etc., to determine the optimal number of trees. Adjust the `max_depth` parameter to control the maximum depth of the decision tree, and test values ​​such as 10, 20, 30, etc. to avoid overfitting. Evaluate the impact of different `min_samples_split` parameter values ​​on model performance through cross-validation, such as 2, 5, 10, etc., which is the minimum number of samples required for decision tree splitting. Use grid search to optimize the regularization parameter `C` of SVM, and test values ​​ranging from 0.1 to 100 to find the best penalty coefficient. Adjust the kernel function parameter `gamma`, and test values ​​such as 0.01 to 10. This parameter controls the influence range of a single training sample in the RBF kernel. Optimize the `max_depth` parameter and set the maximum depth of the tree, such as 4, 5, and 6 layers, to control the model complexity. Adjust the learning rate `eta` and test values ​​such as 0.01, 0.05, and 0.1 to determine the step size when improving the tree. Optimize the `min_child_weight` parameter and test values ​​such as 1, 2, and 3, which is the minimum weight sum required when dividing child nodes. During the optimization process, we use 5-fold cross validation (cv=5) to evaluate the performance of each set of parameters and select the parameter combination that performs best on the validation set. In this way, we can ensure that the model not only performs well on the training set, but also maintains high generalization ability on unseen data. Ultimately, we will select the parameter settings that perform best in cross validation to obtain the most optimized grass yield inversion model.

[0018] In the inversion prediction step (S6), we use the optimal model trained and verified in the model optimization step (S4) and combine it with the latest multi-source data to perform inversion prediction of grassland grass yield. The specific process is as follows: First, collect the latest multi-source data, including MODIS satellite data, drone high-resolution image data, climate data collected by weather stations, and digital elevation model data. These data must undergo the same preprocessing and feature extraction process as in step (S1) to ensure data consistency and availability. Load the optimal model trained in step (S4). This model has been parameter optimized and has demonstrated good prediction performance on the validation set. The latest multi-source data after preprocessing is input into the trained model to perform prediction. The model will output estimates of grassland grass yield, which are usually expressed in wet weight. Since the commercial and scientific value of grass yield is often related to its dry matter content, it is necessary to convert the wet weight output by the model into dry matter weight. This conversion is based on the dry matter conversion coefficient determined by historical data. For example, if historical data show that the dry matter content of grass is 20% of the wet weight on average, the wet weight output by the model is multiplied by 0.20 to obtain the dry matter weight. Generate inverse prediction results for grassland grass yield, including spatial distribution maps and time series maps of grassland grass yield. These results are expressed in dry matter weight and can provide quantitative data support for grassland management. Use the most recent period of ground-measured grass yield data to verify the model's prediction results and calculate prediction accuracy indicators such as the coefficient of determination (R 2 ), root mean square error (RMSE) and mean absolute error (MAE) to ensure the reliability of the prediction results.

[0019] In the result analysis step (S7), we conduct a detailed analysis of the inversion prediction results to evaluate the spatial distribution and temporal variation of grassland grass yield. First, the trained model is used in combination with the latest multi-source data, including MODIS satellite data, drone high-resolution image data, climate data collected by meteorological stations, and digital elevation model data to perform inversion prediction. These data are preprocessed, including normalization and key feature extraction, such as NDVI, EVI, LST, etc. The prediction results are expressed in dry matter weight, and the wet weight is converted to dry matter weight by multiplying by a dry matter conversion coefficient determined based on historical data (e.g., 0.20). Next, the spatial distribution map and time series graph of grassland grass yield are generated, and these graphs are visualized through GIS software to intuitively present the spatial variability and temporal trend of grass yield. In order to ensure the accuracy of the prediction results, we use statistical indicators to evaluate the performance of the model, including calculating the root mean square error (RMSE), mean absolute error (MAE), and determination coefficient (R 2 ). For example, the calculation formula for RMSE is where y i is the actual grass yield at the i-th observation point, is the predicted value, and n is the total number of observation points. These indicators can fully reflect the prediction accuracy and fit of the model. In addition, we will generate a histogram of feature importance to identify the features that contribute most to grass yield prediction, which helps to understand the prediction ability of the model and guide future data collection. Through these detailed analysis steps, this invention can provide a scientific basis for grassland management, help decision makers better understand the spatiotemporal changes in grassland productivity, and formulate reasonable management strategies accordingly. "

[0020] The accuracy of the model was evaluated by comparing the measured grass yield data with the inverse prediction data. We calculated the coefficient of determination (R 2 ), root mean square error (RMSE), and mean absolute error (MAE). For example, RMSE is calculated using the following formula: where y i is the actual grass yield at the ith observation point, is the model prediction value, and n is the total number of observation points. These indicators can fully reflect the prediction accuracy and fitting degree of the model, ensuring the reliability and effectiveness of the model in practical applications.

[0021] Using the inversion results, the present invention provides decision support for grassland ecological protection, grass-livestock balance management and animal husbandry production. Carrying capacity assessment considers at least three factors including grass yield, grassland area and livestock number. For example, by analyzing grass yield data, we can determine the number of livestock that each hectare of grassland can carry, thereby avoiding overgrazing and protecting the ecological balance of the grassland. In addition, the inversion results can also help formulate reasonable grassland utilization strategies, such as adjusting grazing cycles and intensity, and optimizing forage reserves to support sustainable livestock production. The present invention also includes establishing a grassland grass yield database for storing and analyzing historical and real-time grass yield data. The database covers classification information of grassland vegetation types and records of grassland utilization status, such as grass yield per unit area, interannual changes and spatial distribution differences of different grassland types. These data not only provide a scientific basis for grassland management, but can also be used to monitor grassland ecological changes, evaluate the effectiveness of protection and restoration measures, and provide data support for future grassland protection policies.

[0022] The technical solution of the present invention can bring the following significant technical effects:

[0023] 1. Improve the accuracy of grass yield estimation: By integrating MODIS satellite data, high-resolution image data from drones, climate data collected by weather stations, and digital elevation model data, combined with feature selection and machine learning algorithms, the present invention can significantly improve the estimation accuracy of grassland grass yield. For example, in the study of Songnen grassland, the remote sensing estimation model established by statistical analysis methods using MODIS data combined with ground-measured data has a simulation accuracy of 78%, which can effectively simulate grassland grass yield and conduct scientific research.

[0024] 2. Enhance the spatiotemporal resolution and robustness of data: The multi-source data fusion technology adopted by the present invention can provide more continuous and detailed data support in space and time, and enhance the spatiotemporal resolution of the data. In addition, through the optimization of the machine learning algorithm, the model can adapt to different grassland types and environmental conditions, and improve the robustness of the inversion model. Multi-source information fusion technology can make full use of multiple information sources, and combine redundant or complementary information of multiple information sources in space or time according to specific standards to obtain a consistent interpretation or description of the object being measured, so that the information system has better performance than the system composed of the various subsets it contains.

[0025] 3. Realize real-time monitoring and prediction of grassland grass yield: The method of the present invention can realize real-time inversion prediction of grassland grass yield, including monthly and seasonal grass yield estimation. The confidence interval of the inversion result is estimated by Monte Carlo simulation or bootstrap, the number of simulations is not less than 1000 times, and the confidence level of the confidence interval is 95%, thereby providing a scientific basis for decision-making support for grassland ecological protection, grass-livestock balance management and animal husbandry production. In addition, the study of using vegetation index to predict grassland grass yield in the next 16 days found that the prediction accuracy of the power exponential model based on NDVI was 74%, indicating that it is feasible to use vegetation index to predict grass yield in the next 16 days. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 The step diagram of the grassland grass yield inversion method of the present invention is DETAILED DESCRIPTION

[0027] The following will be combined with the attached embodiment of the present invention Figure 1 , the technical solutions in the embodiments of the present invention are clearly and completely described. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0028] The present invention proposes a grassland grass yield inversion method based on multi-source data fusion and machine learning algorithm, aiming to improve the estimation accuracy and efficiency of grassland biomass. The method is implemented by the following steps:

[0029] S1: Collect multi-source data of the grassland area, including MODIS satellite data, high-resolution drone image data, climate data collected by at least three meteorological stations, and digital elevation model data. The temporal resolution of MODIS satellite data is 16 days and the spatial resolution is 250 meters. The spatial resolution of drone image data is not less than 0.1 meters. Climate data includes temperature, precipitation and humidity. The spatial resolution of digital elevation model data is 30 meters.

[0030] S2: Preprocess the collected data, including using the threshold filtering method to clean the data, normalize the data to the [0, 1] interval, and extract key features, including the normalized vegetation index (NDVI), enhanced vegetation index (EVI), land surface temperature (LST) and altitude. The calculation formulas of NDVI and EVI are NDVI = (ρ NIR -ρ Red ) / (ρ NIR +ρ Red ) and EVI = 2.5 × (ρ NIR / ρ Red ), where ρ NIR and ρ Red represent the reflectance in the near-infrared and red bands, respectively;

[0031] S3: using a feature selection algorithm, such as a recursive feature elimination (RFE) algorithm based on information gain, to select the features that have the greatest impact on grass yield inversion, ensuring that the correlation coefficient of the selected features is greater than 0.5, and the number of selected features does not exceed 80% of the total number of features;

[0032] S4: Use one or more of the random forest, support vector machine (SVM) or extreme gradient boosting (XGBoost) algorithms, combined with the feature selection results, to train the grass yield inversion model. The model training is validated using at least two independent data sets, one of which is used for model training and the other for model validation. The number of samples in the validation set is not less than 20% of the total number of samples.

[0033] S5: Optimize model parameters through cross-validation and grid search to improve the generalization ability and inversion accuracy of the model. The optimization process includes adjusting parameters such as tree depth, minimum sample split, and learning rate. The depth of the random forest tree is set in the range of 10 to 20, the minimum sample split number is set in the range of 2 to 20, and the learning rate of XGBoost is set in the range of 0.01 to 0.2.

[0034] S6: Use the trained model and the latest multi-source data to perform inverse prediction of grassland grass yield. The inverse prediction includes but is not limited to monthly and seasonal grass yield estimates. The prediction results are expressed in dry matter weight. The dry matter conversion factor is determined based on historical data. The conversion factor ranges from 0.8 to 0.95.

[0035] S7: Analyze the inversion results, evaluate the spatial distribution and temporal changes of grass yield, generate spatial distribution maps and time series maps of grassland grass yield, and visualize them through GIS software. The resolution of the spatial distribution map should not be less than 250 meters, and the time interval of the time series map should not exceed 1 month.

[0036] In the data collection and preprocessing step, MODIS satellite data are first collected, with a temporal resolution of 16 days to ensure that the periodic changes in vegetation growth can be captured, and a spatial resolution of 250 meters to cover a larger grassland area and maintain the spatial continuity of the data. At the same time, high-resolution image data with a spatial resolution of not less than 0.1 meters are obtained using high-resolution cameras or multispectral sensors carried by drones to capture the fine structure and details of grassland vegetation. In addition, data on climate parameters such as temperature, precipitation, humidity, and solar radiation are collected from at least three meteorological stations, which are essential for understanding the climatic conditions for grassland growth. Digital elevation model (DEM) data are obtained through GPS and other remote sensing technologies to provide topographic information of the grassland area. In the data preprocessing stage, the threshold filtering method is used to denoise and remove outliers from satellite and drone image data to ensure data quality. All data will be normalized and converted to the [0,1] interval to eliminate the dimensional effects between different data sources. Finally, key features were extracted, including the Normalized Difference Vegetation Index (NDVI) and Enhanced Vegetation Index (EVI) calculated by analyzing the reflectance of the red and near-infrared bands, the Land Surface Temperature (LST) calculated by the radiance of the thermal infrared band, and the altitude directly extracted from the DEM data. The extraction of these characteristic parameters provides the necessary input variables for the subsequent grassland grass yield inversion model.

[0037] In the feature selection step (S3), we use the recursive feature elimination (RFE) algorithm based on information gain to select features from the preprocessed data to determine the features that contribute most to the grassland grass yield inversion model. Specifically, we first calculate the information gain between each feature (including NDVI, EVI, LST, and altitude) and grass yield, then sort the features according to the gain value, and recursively eliminate the least important features using the RFE algorithm. In each round of recursion, we train a random forest model, which determines which features contribute the least to the model's predictive ability by evaluating the importance of the features. We set the correlation coefficient threshold to 0.5, and only when the correlation coefficient between the feature and grass yield is greater than this value, the feature will be retained. This ensures that the selected features have a strong linear relationship with grass yield. Finally, we control the number of selected features to not exceed 80% of the total number of features to avoid model overfitting and improve the generalization ability of the model. Through this series of precisely controlled parameters and steps, we are able to screen out the most predictive features from the original feature set, providing a solid foundation for building an accurate and efficient grassland grass yield inversion model.

[0038] In the model training step (S4), we carefully designed the training process to ensure that the grassland grass yield inversion model trained using the random forest, support vector machine (SVM) or extreme gradient boosting (XGBoost) algorithm is both accurate and efficient. First, we divided the dataset into a training set and a validation set, usually in a ratio of 80% training set and 20% validation set, to ensure that the model has enough data when training, while leaving enough samples to evaluate the generalization ability of the model. For the random forest model, we set the number of decision trees to 100, which can improve the stability and accuracy of the model by integrating the prediction results of multiple decision trees. When training the SVM model, we choose the radial basis function (RBF) as the kernel function, and optimize the regularization parameter C (usually in the range of [0.1,100]) and the kernel function parameter γ (usually in the range of [0.001,1]) through grid search and cross-validation to find the best classification hyperplane. For the XGBoost model, we set the tree depth to 6 and the learning rate to 0.1, and balance the bias and variance of the model by adjusting these parameters. After the model training is completed, we use the validation set to evaluate the performance of the model and calculate the root mean square error (RMSE), mean absolute error (MAE) and coefficient of determination (R 2), which will comprehensively reflect the prediction accuracy and fit of the model. Based on the performance indicators on the validation set, we may need to adjust the model parameters, such as increasing or decreasing the number of decision trees in the random forest, adjusting the C and γ parameters of the SVM, or adjusting the learning rate and tree depth of XGBoost to further improve the accuracy of the model. Through this series of finely controlled parameters and steps, we can ensure that the trained model has high accuracy and good generalization ability in predicting grassland grass yield.

[0039] In the model optimization step (S5), we used cross-validation and grid search techniques to fine-tune the parameters of the random forest, support vector machine (SVM) and extreme gradient boosting (XGBoost) models to optimize the generalization ability and inversion accuracy of the grassland grass yield inversion model. Specifically, for the random forest model, we tested the number of decision trees `n_estimators` (ranging from 100 to 300, with a step of 100), the maximum depth of the tree `max_depth` (ranging from 10 to 30, with a step of 10), and the minimum number of samples required for node splitting `min_samples_split` (with values ​​of 2, 5, and 10) through grid search based on 5-fold cross-validation. For the SVM model, we adjusted the regularization parameter `C` (ranging from 0.1 to 100, with a step of 0.1), and the parameter `gamma` of the RBF kernel function (ranging from 0.01 to 10, with a step of 0.01). The optimization of the XGBoost model focuses on the maximum depth `max_depth` (ranging from 4 to 6), the learning rate `eta` (ranging from 0.01 to 0.1, with a step size of 0.01), and the minimum weight required for child node division `min_child_weight` (with values ​​of 1, 2, and 3). Through the systematic adjustment and optimization of these parameters, we can ensure that the selected model fits the training set well and has strong generalization ability on the validation set, thus providing strong model support for the accurate inversion of grassland grass yield.

[0040] In the inversion prediction step (S6), we first collect the latest MODIS satellite data with a temporal resolution of 16 days and a spatial resolution of 250 meters, as well as high-resolution drone image data at least once a month with a spatial resolution of no less than 0.1 meters. At the same time, we collect climate data including temperature, precipitation, and humidity from at least three weather stations, and obtain terrain information with a resolution of 30 meters from the digital elevation model. These data undergo the same preprocessing process as the data collection and preprocessing step (S1), including threshold filtering method for denoising, normalization to the [0,1] interval, and extraction of key features such as NDVI, EVI, LST, etc. Subsequently, we load the optimal model obtained in the model training step (S4), for example, a random forest model containing 100 decision trees, the maximum tree depth is 10, and the minimum sample split is set to 2. Using this model, we use the preprocessed data as input, perform inversion prediction, and obtain the estimated value of grassland grass yield. The wet weight estimates output by the model are then converted to dry matter weight by multiplying by a dry matter conversion factor determined based on historical data, such as 0.20. Finally, we generate spatial distribution maps and time series maps of grassland grass yield expressed in dry matter weight and visualize them through GIS software to provide a scientific basis for grassland management. If necessary, we also use recent ground-truth grass yield data to verify the accuracy of the model prediction and calculate R 2 , RMSE, MAE and other statistical indicators to ensure the reliability of the prediction results.

[0041] In the result analysis step (S7), we conduct a detailed analysis of the inversion prediction results to evaluate the spatial distribution and temporal variation of grassland grass yield. First, the inversion prediction is performed using the trained model combined with the latest multi-source data, including MODIS satellite data, drone high-resolution image data, climate data collected by meteorological stations, and digital elevation model data. These data are preprocessed, including normalization and key feature extraction, such as NDVI, EVI, LST, etc. The prediction results are expressed in dry matter weight, and the wet weight is converted to dry matter weight by multiplying by the dry matter conversion factor determined based on historical data (e.g., 0.20). Next, the spatial distribution map and time series graph of grassland grass yield are generated, and these graphs are visualized using GIS software to intuitively present the spatial variability and temporal trend of grass yield. In order to ensure the accuracy of the prediction results, we use statistical indicators to evaluate the model performance, including calculating the root mean square error (RMSE), mean absolute error (MAE), and determination coefficient (R 2 ). For example, the calculation formula for RMSE is where y i is the actual grass yield at the i-th observation point, is the predicted value, and n is the total number of observation points. These indicators can fully reflect the prediction accuracy and fit of the model. In addition, we will generate a histogram of feature importance to identify the features that contribute most to grass yield prediction, which helps to understand the prediction ability of the model and guide future data collection. Through these detailed analysis steps, this invention can provide a scientific basis for grassland management, help decision makers better understand the spatiotemporal changes in grassland productivity, and formulate reasonable management strategies accordingly. "

[0042] The accuracy of the model was evaluated by comparing the measured grass yield data with the inverse prediction data. We calculated the coefficient of determination (R 2 ), root mean square error (RMSE), and mean absolute error (MAE). For example, RMSE is calculated using the following formula: where y i is the actual grass yield at the ith observation point, is the model prediction value, and n is the total number of observation points. These indicators can fully reflect the prediction accuracy and fitting degree of the model, ensuring the reliability and effectiveness of the model in practical applications.

[0043] Using the inversion results, the present invention provides decision support for grassland ecological protection, grass-livestock balance management and animal husbandry production. Carrying capacity assessment considers at least three factors including grass yield, grassland area and livestock number. For example, by analyzing grass yield data, we can determine the number of livestock that each hectare of grassland can carry, thereby avoiding overgrazing and protecting the ecological balance of the grassland. In addition, the inversion results can also help formulate reasonable grassland utilization strategies, such as adjusting grazing cycles and intensity, and optimizing forage reserves to support sustainable livestock production. The present invention also includes establishing a grassland grass yield database for storing and analyzing historical and real-time grass yield data. The database covers classification information of grassland vegetation types and records of grassland utilization status, such as grass yield per unit area, interannual changes and spatial distribution differences of different grassland types. These data not only provide a scientific basis for grassland management, but can also be used to monitor grassland ecological changes, evaluate the effectiveness of protection and restoration measures, and provide data support for future grassland protection policies.

[0044] The grassland grass yield inversion method of the present invention improves the accuracy and efficiency of grass yield estimation by fusing multi-source data and applying advanced machine learning algorithms, providing a scientific basis for the rational utilization of grassland resources and ecological protection.

[0045] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. The above embodiments and descriptions are only preferred examples of the present invention and are not intended to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A grassland grass yield inversion method based on multi-source data fusion and machine learning algorithm, characterized in that: The following steps are involved: S1: Collect multi-source data of the grassland area, including MODIS satellite data, high-resolution drone image data, climate data collected by at least three meteorological stations, and digital elevation model data. The temporal resolution of MODIS satellite data is 16 days and the spatial resolution is 250 meters. The spatial resolution of drone image data is not less than 0.1 meters. Climate data includes temperature, precipitation and humidity. The spatial resolution of digital elevation model data is 30 meters. S2: Preprocess the collected data, including using the threshold filtering method to clean the data, normalize the data to the [0, 1[ interval, and extract key features, including the normalized vegetation index (NDVI), enhanced vegetation index (EVI), land surface temperature (LST) and altitude. The calculation formulas of NDVI and EVI are NDVI = (ρ NIR -ρ Red ) / (ρ NIR +ρ Red ) and EVI = 2.5 × (ρ NIR / ρ Red ), where ρ NIR and ρ Red represent the reflectance in the near-infrared and red bands, respectively; S3: using a feature selection algorithm, such as a recursive feature elimination (RFE) algorithm based on information gain, to select the features that have the greatest impact on grass yield inversion, ensuring that the correlation coefficient of the selected features is greater than 0.5, and the number of selected features does not exceed 80% of the total number of features; S4: Use one or more of the random forest, support vector machine (SVM) or extreme gradient boosting (XGBoost) algorithms, combined with the feature selection results, to train the grass yield inversion model. The model training is validated using at least two independent data sets, one of which is used for model training and the other for model validation. The number of samples in the validation set is not less than 20% of the total number of samples. S5: Optimize model parameters through cross-validation and grid search to improve the generalization ability and inversion accuracy of the model. The optimization process includes adjusting parameters such as tree depth, minimum sample split, and learning rate. The depth of the random forest tree is set in the range of 10 to 20, the minimum sample split number is set in the range of 2 to 20, and the learning rate of XGBoost is set in the range of 0.01 to 0.

2. S6: Use the trained model and the latest multi-source data to perform inverse prediction of grassland grass yield. The inverse prediction includes but is not limited to monthly and seasonal grass yield estimates. The prediction results are expressed in dry matter weight. The dry matter conversion factor is determined based on historical data. The conversion factor ranges from 0.8 to 0.

95. S7: Analyze the inversion results, evaluate the spatial distribution and temporal changes of grass yield, generate spatial distribution maps and time series maps of grassland grass yield, and visualize them through GIS software. The resolution of the spatial distribution map should not be less than 250 meters, and the time interval of the time series map should not exceed 1 month.

2. The method according to claim 1, characterized in that The collection period of MODIS satellite data in the multi-source data is from April to October of each growing season, and the collection period of UAV high-resolution image data is at least once a month.

3. The method according to claim 1 or 2, characterized in that: The feature selection algorithm also includes a maximum information coefficient (MIC) algorithm based on mutual information, and a model-based feature selection method, such as a feature importance evaluation based on random forests, where the feature importance threshold is set to 1.5 times the average importance value.

4. The method according to claim 1, characterized in that: The number of trees in the random forest in the machine learning algorithm is set to 100, the radial basis function (RBF) is selected as the kernel function of the support vector machine (SVM), the depth of the extreme gradient boosting (XGBoost) is set to 6 layers, and the subsample ratio of XGBoost is set to 0.

8.

5. The method according to claim 1, characterized in that The model optimization strategy also includes using Akaike Information Criterion (AIC) or Bayesian Information Criterion (BIC) for model selection, and evaluating the stability of the clustering algorithm by using Silhouette Coefficient, and the threshold of Silhouette Coefficient is set to 0.

5.

6. The method according to claim 1, characterized in that The inversion prediction further includes uncertainty assessment of grass yield, and the confidence interval of the inversion result is estimated by Monte Carlo simulation or bootstrap, with the number of simulations being no less than 1000 times and the confidence level of the confidence interval being 95%.

7. The method according to claim 1, characterized in that The verification step includes calculating the determination coefficient (R) between the inverted grass yield and the measured grass yield. 2 ), root mean square error (RMSE) and mean absolute error (MAE), the number of samples in the validation data set is not less than 30, R 2 The threshold of was set to 0.7, the threshold of RMSE was set to 15% of grass yield, and the threshold of MAE was set to 10% of grass yield.

8. The method according to claim 1, characterized in that The decision support further includes grassland carrying capacity assessment and grazing management strategy optimization based on the inversion results. The carrying capacity assessment considers at least three factors including grass yield, grassland area and livestock number. The grazing management strategy optimization includes the adjustment of grazing cycle, grazing intensity and grazing area.

9. The method according to claim 1, characterized in that: The database also includes classification information of grassland vegetation types and records of grassland utilization status. The vegetation types include at least three main grassland types, such as tallgrass prairie, shortgrass prairie and mixed grass prairie. The grassland utilization status records include grazing frequency and grazing intensity.

10. A computer program product, characterized in that The method comprises an instruction sequence for executing the method described in any one of claims 1 to 9, and a software module for storing and processing data, wherein the software module comprises a data preprocessing module, a feature selection module, a model training module, an inversion prediction module and a result analysis module, and each module comprises a user interface for inputting parameters, displaying processing progress and outputting results.

Images