Grass yield multi-space-time simulation method combining field sampling and integrated machine learning

Abstract

The invention relates to the field of grassland ecological index calculation, in particular to a grass yield multi-space-time simulation method combining field sampling and integrated machine learning, which comprises the following steps of: collecting sampling data of field investigation, key remote sensing vegetation indexes influencing the grass yield of a natural grass land, and covariants related to surrounding environment, terrain, climate and the like; integrating a plurality of machine learning algorithm models for training, and evaluating the performance of each algorithm model by adopting cross validation; determining the weight of each algorithm model through an optimization algorithm; establishing an integrated model based on the weight of each algorithm model; and carrying out simulation calculation by utilizing an integrated machine learning model, and outputting a multi-period grass yield spatial distribution diagram. According to the method, an integrated algorithm is adopted to construct a precise spatial distribution model and time sequence change simulation, geography, climate, soil and multi-source remote sensing data are effectively integrated, and an integrated machine learning method is utilized to perform high-resolution, high-precision and multi-time-sequence spatial simulation on the grass yield of the natural grass land; the method has great reference value for grass and livestock balance monitoring and sustainable management of grassland.

Description

Technical Field

[0001] The present invention relates to the field of grassland ecological index calculation, and specifically to a multi-temporal and spatial simulation method of grass yield combining field sampling and integrated machine learning. Background Art

[0002] The grass yield of natural grassland refers to the total amount of material accumulated in the aboveground part of grassland vegetation that can be economically utilized within a production cycle. Grass yield is not only an important indicator for quantitatively evaluating the rational use of grassland resources and the balance between grass and livestock, but also a sensitive indicator for accurately diagnosing the health status of grassland. It represents the production status and production potential of grassland and is the basis for the development of animal husbandry. Accurate estimation of grass yield and its changes help to reveal the dynamic characteristics of natural grasslands and accurately evaluate the restoration and degradation of natural grasslands. How to accurately estimate the grass yield of natural grasslands in multiple time series on a large regional scale and avoid damage to grasslands has become a key issue that needs to be urgently addressed in current grassland management and ecological protection.

[0003] The traditional method of estimating grass yield is mainly to directly obtain the measured data of grassland biomass in the survey area by sampling at field points and using direct harvesting method (mowing method). Grassland plants are mowed during the peak growth season of grassland vegetation, then bagged and dried at a constant temperature and weighed to obtain the dry weight of grassland yield. This method has high measurement accuracy and is suitable for precise measurement of yield on a small scale, but it is destructive to grassland, time-consuming and labor-intensive, and can only obtain the grass yield of the sampling area in the current year. It relies on ground-measured data and is constrained by temporal and spatial limitations, making it difficult to expand to larger scales and productive grasslands in history and future years.

[0004] With the continuous development of remote sensing technology, multispectral or hyperspectral optical remote sensing has a good sensitivity to green vegetation. Due to its advantages such as high resolution and mature technology, remote sensing empirical models have gradually been applied to grass yield estimation and have formed technical specifications for business operations. Remote sensing data has the advantage of strong spatiotemporal continuity and can be used to monitor the continuous changes in grass yield on a large scale in a region. For example, vegetation indices based on remote sensing inversion (such as normalized vegetation index, NDVI) are used to estimate the grass yield of natural grasslands by building empirical models between them and grass yield. However, remote sensing technology has a significant effect in the inversion of grass yield of natural grasslands with a single grassland type and flat terrain over a large area. However, the accuracy of the grass yield empirical model based on remote sensing technology is limited by the quality and quantity of field sampling data, and the spectral characteristics of grassland vegetation are easily interfered with, and it is impossible to effectively and comprehensively capture the complex relationship between the grass yield of natural grasslands with multiple grassland types and surrounding environmental factors under complex ecological environments. In addition, remote sensing data are easily affected by meteorological conditions such as weather and clouds, resulting in discontinuous data acquisition or reduced quality.

[0005] Field sample surveys are the most important means of obtaining accurate values ​​of natural grassland grass production. Remote sensing inversion is the key technology for spatial simulation of natural grassland grass production. Machine learning combined with large language models and open cloud platforms that store massive data (such as the GEE platform) are effective means for large-scale and long-term monitoring of natural grasslands. Machine learning is not limited to the number and type of samples, nor does it require specific assumptions about the model structure. It can flexibly, efficiently, and robustly fit the relationship between dependent and independent variables, and can continuously improve simulation accuracy through feedback errors. A single machine learning model has distinct advantages, but it also has its own limitations. For example, the random forest algorithm has high simulation accuracy and model robustness, but it is prone to poor performance in high-dimensional data, and the support vector machine has a high computational cost when processing large-scale data. In addition, a single algorithm is usually difficult to fully capture the complex interactions between grass production and multiple environmental factors. Ensemble learning combines models by constructing machine learning models such as Random Forests (RF), Artificial Neural Networks (ANN), Support Vector Machines (SVM) and Multivariate Adaptive Regression Spline Models. Based on field sample surveys and integrating multi-source remote sensing data, ensemble machine learning can not only accurately estimate the grass production of natural grasslands through deep learning of massive historical data, classification, zoning and segmentation, and conduct macro-monitoring and evolution analysis of natural grassland grass production, but also effectively predict the changing trends of grass production in different scenarios in the future, providing technical support for the scientific formulation of grassland management strategies.

[0006] Therefore, there is an urgent need for a multi-temporal and spatial simulation method of grass production that combines field sampling and integrated machine learning. Summary of the invention

[0007] In order to overcome the problems existing in the prior art, the purpose of the present invention is to provide a multi-temporal and spatial simulation method of grass yield that combines field sampling and integrated machine learning, thereby improving the flexibility and reliability of data processing by integrating multi-source and multi-factor data.

[0008] To achieve the above object, the present invention provides the following technical solution: a multi-temporal and spatial simulation method of grass yield combining field sampling and integrated machine learning, comprising the following steps:

[0009] S1: Collect and organize sampling data from field surveys of natural grassland field plots;

[0010] S2: Data preprocessing of survey data, and selection of covariates related to grass production based on remote sensing data;

[0011] S3: Establish multiple algorithm models for training, and use cross-validation to evaluate the performance of each algorithm model during the training process;

[0012] S4: Determine the weight of each algorithm model through the optimization algorithm;

[0013] S5: Establish an integrated model based on the weights of each algorithm model;

[0014] S6: Use the integrated model to simulate and calculate, obtain the simulated results of grass production in the field, and output the spatial distribution map of grass production.

[0015] The present invention is further configured as follows: the survey data in step S1 includes longitude and latitude data, coverage and grass yield of herbaceous plants and small shrubs.

[0016] The present invention is further configured as follows: Step S1 specifically comprises taking the center of the field sample plot as the starting point, setting three 40m long sample lines with an angle of 120° in the north, southeast and southwest directions respectively, selecting a typical area representing the sample plot within 5m on any side of each sample line to set three 1m×1m sample plots, and extracting the survey data in each sample plot area.

[0017] The present invention is further configured as follows: Step S2 specifically comprises taking grass yield as the dependent variable, using R language to extract remote sensing data as covariates, and calculating the correlation coefficient matrix between each covariate and grass yield, and selecting covariates with correlation coefficient r>0.3 for subsequent steps.

[0018] The R language may be R-4.3.2 and RStudio, and an integrated learning method is used to simulate the dry weight of grass yield with a resolution of 100m.

[0019] The present invention is further configured as follows: the remote sensing data includes normalized vegetation index, annual growing season average precipitation, annual growing season average temperature, standardized precipitation index, soil organic matter content, soil moisture, altitude, slope, longitude and latitude.

[0020] The present invention is further configured as follows: in step S3, the algorithm model includes a random forest model (RandomForests, RF), an artificial neural network model (Aritificial Neural Networks, ANN), a support vector machine model (Support Vector Machines, SVM) and a multivariate adaptive regression spline model.

[0021] The present invention is further configured to use k-fold cross validation (k=10) to evaluate the performance of each model. In each iteration, the data set is randomly divided into 10 parts, 9 for training and 1 for validation, and this cycle is repeated to ensure that the model has good generalization ability.

[0022] The present invention is further configured as follows: in step S4, the optimization algorithm adopts L-BFGS-B (Limited-memory Broyden-Fletcher-Goldfarb-Shanno); the residual sum of squares RSS and the determination coefficient R of each algorithm model are calculated. 2 and predicted values, and uses L-BFGS-B to find the optimal weight combination that minimizes the fitting error of each algorithm model.

[0023] The present invention is further configured as follows: the integrated model f established in step S5 e (x) is expressed as:

[0024] f e (x) = k1 × f NN (x)+k2×f RF (x)+k3×s SVM (x)+k4×(x)

[0025] k1+k2+k3+k4=100

[0026] Among them, f NN (x) is the fitting result of the artificial neural network model, and its weight is k1; f RF (x) is the fitting result of the random forest model, and its weight is k2; f SVM (x) is the fitting result of the support vector machine model, and its weight is k3; f MARS (x) is the fitting result of the multivariate adaptive regression spline model, and its weight is k4.

[0027] In summary, the beneficial effects of the above technical solution of the present invention are as follows:

[0028] The present invention adopts an integrated algorithm to construct an accurate spatial distribution model, effectively combines geographical, climate, soil and remote sensing data using integrated learning methods, and conducts high-resolution and high-precision spatial simulation of grass production at the provincial scale, providing a reliable reference for grassland management in large areas and complex environments.

[0029] The present invention also improves the data input and output methods, and improves the flexibility and reliability of data processing by integrating multi-source data. The input data includes remote sensing images, ground monitoring data and meteorological data, and the output is a refined grassland yield distribution map, which greatly improves the accuracy and operability of the simulated grass yield spatial distribution. It ensures the diversity and representativeness of the input data, and also improves the practicality and visualization level of the output results. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for describing the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0031] Figure 1 Technical flow chart of the multi-temporal and spatial simulation method of grass yield combining field sampling and integrated machine learning.

[0032] Figure 2 Schematic diagram of the correlation coefficient between grass production in a certain province and various factors in the embodiment.

[0033] Figure 3 This is a comparative analysis chart of the simulated values ​​and measured values ​​of the multi-temporal and spatial simulation method of grass yield combining field sampling and integrated machine learning in the embodiment.

[0034] Figure 4 For the example, output the spatial distribution map of natural grassland production in a certain province in the current year (2022).

[0035] Figure 5 For the embodiment, a spatial distribution map of natural grassland production in a province in the historical period (2016) is output. DETAILED DESCRIPTION

[0036] In order to enable those skilled in the art to better understand the technical solution of the present invention, the technical solution of the present invention is clearly and completely described below in conjunction with the accompanying drawings of the present invention. Based on the embodiments of the present invention, other similar embodiments obtained by ordinary technicians in the field without making any creative work should all fall within the scope of protection of the present invention.

[0037] The present invention will be further described below in conjunction with the accompanying drawings and preferred embodiments.

[0038] Example:

[0039] like Figure 1 As shown in the figure, a preferred embodiment of the present invention is a multi-temporal and spatial simulation method of grass yield combining field sampling and integrated machine learning, comprising the following steps:

[0040] S1: Collect and organize sampling data from field surveys of natural grassland field plots;

[0041] Starting from the center of the field plot, three 40m long sample lines with an angle of 120° were set in the north, southeast and southwest directions respectively. Three 1m×1m sample plots were set in a typical area representing the plot within 5m on either side of each sample line, and the survey data in each sample plot area were extracted. The survey data included latitude and longitude data, coverage and grass yield of herbaceous plants and small shrubs.

[0042] S2: Data preprocessing of survey data, and selection of covariates related to grass production based on remote sensing data;

[0043] Remote sensing data were extracted as covariates using R language, and the correlation coefficient matrix between each covariate and grass yield was calculated. Covariates with correlation coefficient r>0.3 were selected for subsequent steps.

[0044] In this embodiment, the remote sensing data uses the 30m resolution annual maximum normalized vegetation index (NDVI) data of a province between 2012 and 2022, and the environmental data includes terrain factors (such as altitude, slope, etc.) extracted from the digital elevation model (DEM) with a resolution of 90m. The Mean Annual Growing Season Precipitation (MAGSP), the Mean Annual Growing Season Temperature (MAGST), and the Standardized Precipitation Index (SPI) of a province from 2010 to 2022 are also collected with a resolution of 1km. The soil data includes the soil organic matter (SOM) and soil moisture (SH) data of China with a resolution of 1km; the human activity data includes the population density data of 2010, 2015, and 2020 with a resolution of 1km.

[0045] Based on R-4.3.2, covariates were extracted and the correlation coefficient matrix with grass production was calculated. The results are as follows: Figure 2 As shown in the figure, the covariates were selected into the model when the correlation coefficient between them and grass yield reached r>0.3. Among the environmental factors, NDVI had the highest correlation with grass yield (r=0.48), followed by average precipitation in the growing season (r=0.34), and soil moisture also had a certain influence (r=0.30). Therefore, NDVI, average precipitation in the growing season, slope and soil moisture were selected as covariates to participate in the simulation of grass yield.

[0046] S3: Establish multiple algorithm models for training, and use cross-validation to evaluate the performance of each algorithm model during the training process;

[0047] In this embodiment, four algorithms, namely, random forest model (RF), artificial neural network model (ANN), support vector machine model (SVM) and multivariate adaptive regression spline model, are selected for integrated learning.

[0048] Among them, RF is an integrated learning method for classification and regression, which is achieved by building a large number of decision trees when training the model. The simulation result will return the average value of each tree; ANN is a computing system inspired by biological neural networks, which imitates the structure and function of the animal brain. Since artificial neural networks can reproduce and simulate nonlinear processes, they are currently also widely used in ecological modeling; MARS is an algorithm for complex nonlinear regression problems. It can automatically model nonlinearity and interactions between variables. During the training process, k-fold cross validation (k=10) is used to evaluate the performance of each model. In each iteration, the data set is randomly divided into 10 parts, 9 for training and 1 for validation, and this cycle is repeated to ensure that the model has good generalization ability.

[0049] S4: Determine the weight of each algorithm model through the optimization algorithm;

[0050] L-BFGS-B (Limited-memory Broyden-Fletcher-Goldfarb-Shanno) is used to calculate the residual sum of squares RSS and the coefficient of determination R of each algorithm model. 2 and predicted values, and uses L-BFGS-B to find the optimal weight combination that minimizes the fitting error of each algorithm model.

[0051] In this embodiment, the MARS weight is 14.6, the RF weight is 63.3, the SVM weight is 22.2, and the ANN weight is 0.

[0052] S5: Establish an integrated model based on the weights of each algorithm model e (x);

[0053] f e (x) = k1 × f NN (x)+k2×f RF (x)+k3×f SVM (x)+k4×(x)

[0054] k1+k2+k3+k4=100

[0055] Among them, f NN (x) is the fitting result of the artificial neural network model, and its weight is k1; f RF (x) is the fitting result of the random forest model, and its weight is k2; f SVM (x) is the fitting result of the support vector machine model, and its weight is k3; f MARS (x) is the fitting result of the multivariate adaptive regression spline model, and its weight is k4.

[0056] The integrated model includes four alternative algorithms, and the final integrated model mainly includes three algorithms: random forest (RF), multivariate adaptive regression spline (MARS) and support vector machine (SVM).

[0057] S6: Use the integrated model to simulate and calculate, obtain the simulated results of grass yield in the field, and output the spatial distribution map of grass yield in the current year and historical period, such as Figure 4-Figure 5 As shown in the table, the average grass yield in the province in 2022 is 88.04 g / m 2 , the minimum value is 26.26g / m 2 , the maximum value is 222.46g / m 2 , grass yield is 100~150g / m 2 The area of ​​the range accounts for the largest proportion. Figure 4-Figure 5 The results of the simulation of the spatial distribution of grass production in 2016 (historical period) and 2022 (current year) are shown. Based on the accurate assessment of the grassland productivity in the current year, the present invention can trace the spatial pattern of grass production in the historical period, thereby providing an important scientific basis for the study of grassland ecosystem changes. Based on field sampling data and multi-source remote sensing information, by integrating machine learning algorithms, not only high-precision spatial simulation of grass production is achieved, but also time series reconstruction can be performed, which provides an effective technical means for studying the response of grassland ecosystems to global changes, and has important practical significance for understanding the evolution of grassland ecosystems and formulating grassland protection policies.

[0058] The spatial distribution map of grass production in the historical period output by the present invention, combined with the current year data, can be used to: (1) quantitatively evaluate the interannual fluctuations and spatial heterogeneity of grassland productivity in the study area; (2) identify key areas of grassland degradation or improvement, and provide spatial positioning guidance for grassland ecological restoration; (3) provide a scientific basis for grassland carrying capacity assessment and grazing management.

[0059] Grassland production is closely related to climatic conditions. The province has good water and heat conditions, and the overall grassland production performance is at a high level. From the perspective of spatial distribution, the yield of temperate grassland, temperate meadow grassland and mountain meadow distribution areas in the east and south is relatively high, while the yield of temperate desert grassland and temperate desert distribution areas in the central part is relatively low, showing an overall pattern of high in the east and low in the west. The grass yield of each grassland type in 2022 basically conforms to the distribution law of each grassland type, among which: mountain meadow is the type with the highest grass yield, which is 153.64g / m 2 ; Temperate steppe desert has the lowest grass yield, which is 87.48g / m 2 ; The grass yield of temperate meadow grassland is 130.46g / m 2 ; The grass yield of temperate desert is 100.22g / m 2 ; The grass yield of temperate grassland is 100.32g / m 2 ; The grass yield of temperate desert steppe is 81.79g / m 2 .

[0060] The simulation results of this embodiment are compared with the measured values. Figure 3 As shown in Table 1, the fitting line of the integrated model is close to the 1:1 line, indicating that the simulated value is highly consistent with the measured value. The model can simulate grass production well and has strong generalization ability.

[0061] Table 1. Estimation effect of grass yield by integrated model

[0062]

[0063] It is compared with the model accuracy involved in similar studies, as shown in Table 2:

[0064] Table 2 Estimation of grass yield by similar research models

[0065]

[0066]

[0067] In Table 2, the documents corresponding to the document numbers are as follows:

[0068] [1] Yang S, Feng Q, Liang T, et al. Modeling grassland above-ground biomass based on artificial neural network and remote sensing in the Three-River Headwaters Region[J]. Remote Sensing of Environment, 2018, 204: 448-455.

[0069] [2] Li Shuzhen, Xu Dawei, Fan Kaikai, et al. Research on grassland aboveground biomass inversion based on UAV and satellite remote sensing[J]. Remote Sensing Technology and Applications, 2022, 7(1): 272-278.

[0070] [3] Su Wensen, Sun Bin, Hong Liang, et al. Spatiotemporal remote sensing estimation technology of grass yield in Hulunbuir grassland based on GF-6WFV data[J]. Forestry Science and Technology Newsletter, 2022(10):14-19.

[0071] [4] Zeng Na, Ren Xiaoli, He Honglin, et al. Spatiotemporal dynamics of aboveground biomass in grasslands of Sanjiangyuan National Park and its impact on climate [J]. Acta Ecologica Sinica, 2023, 43(3): 1175-1184.

[0072] [5] Yang Yongqiang, Zhang Qingqing, Xu Hailiang, et al. Spatial and temporal distribution patterns of aboveground biomass of major grassland types in the northern foothills of the western Tianshan Mountains [J]. Journal of Ecology, 2024, 43(7): 2075-2082.

[0073] [6]Li H, LiF, XiaoJ, et al.A machine learning scheme for estimatingfine-resolution grassland aboveground biomass over China with Sentinel-1 / 2satellite images[J].Remote Sensing of Environment,2024,311:114317.

[0074] [7]Yao Y,RenH.Estimation of grassland aboveground biomass in northernChina based on topography-climate-remote sensing data[J].EcologicalIndicators,2024,165:112230.

[0075] It can be seen that the model accuracy RMSE index (16.35g / m 2 ), the regional natural grassland yield is better than previous similar studies in terms of spatial simulation accuracy in terms of correlation coefficient and key error parameters.

[0076] In summary, the present invention improves the flexibility and reliability of data processing by integrating multi-source data. The input data includes remote sensing images, ground monitoring data and meteorological data, and the output is a refined grassland yield distribution map, which greatly improves the accuracy and operability of the simulated grass yield spatial distribution. The diversity and representativeness of the input data are guaranteed, and the practicality and visualization level of the output results are also improved.

[0077] Finally, it should be noted that the above content is only used to illustrate the technical solution of the present invention, rather than to limit the scope of protection of the present invention. Simple modifications or equivalent substitutions of the technical solution of the present invention by ordinary technicians in this field do not deviate from the essence and scope of the technical solution of the present invention.

Claims

1. A multi-temporal and spatial simulation method of grass production combining field sampling and integrated machine learning, characterized in that: The following steps are involved: S1: Collect and organize sampling data from field surveys of natural grassland field plots; S2: Data preprocessing of survey data, and selection of covariates related to grass production based on remote sensing data; S3: Establish multiple algorithm models for training, and use cross-validation to evaluate the performance of each algorithm model during the training process; S4: Determine the weight of each algorithm model through the optimization algorithm; S5: Establish an integrated model based on the weights of each algorithm model; S6: Use the integrated model to simulate and calculate, obtain the simulated results of grass production in the field, and output the spatial distribution map of grass production.

2. The multi-temporal and spatial simulation method of grass production combining field sampling and integrated machine learning according to claim 1, characterized in that: The survey data in step S1 include latitude and longitude data, coverage and grass yield of herbaceous plants and small shrubs.

3. The multi-temporal and spatial simulation method of grass production combining field sampling and integrated machine learning according to claim 2 is characterized in that: Step S1 specifically includes taking the center of the field sample plot as the starting point, setting three 40m long sample lines with an angle of 120° in the north, southeast and southwest directions respectively, selecting three 1m×1m sample plots within 5m on any side of each sample line in a typical area representing the sample plot, and extracting the survey data in each sample plot area.

4. The multi-temporal and spatial simulation method of grass production combining field sampling and integrated machine learning according to claim 1, characterized in that: Specifically, step S2 takes grass yield as the dependent variable, uses R language to extract remote sensing data as covariates, calculates the correlation coefficient matrix between each covariate and grass yield, and selects covariates with correlation coefficient r>0.3 for subsequent steps.

5. The multi-temporal and spatial simulation method of grass yield combining field sampling and integrated machine learning according to claim 4, characterized in that: The remote sensing data include normalized difference vegetation index, annual growing season average precipitation, annual growing season average temperature, standardized precipitation index, soil organic matter content, soil moisture, altitude, slope, longitude and latitude.

6. The multi-temporal and spatial simulation method of grass production combining field sampling and integrated machine learning according to claim 1, characterized in that: In step S3, the algorithm model includes a random forest model, an artificial neural network model, a support vector machine model and a multivariate adaptive regression spline model.

7. The multi-temporal and spatial simulation method of grass yield combining field sampling and integrated machine learning according to claim 6, characterized in that: K-fold cross validation was used to evaluate the performance of each model.

8. The multi-temporal and spatial simulation method of grass yield combining field sampling and integrated machine learning according to claim 7, characterized in that: In step S4, the optimization algorithm adopts L-BFGS-B to calculate the residual sum of squares RSS and the determination coefficient R of each algorithm model. 2 and predicted values, and find the optimal weight combination that minimizes the fitting error of each algorithm model.

9. The multi-temporal and spatial simulation method of grass production combining field sampling and integrated machine learning according to claim 8, characterized in that: The integrated model f established in step S5 e (x) is expressed as: f e (x)=k1×f NN (x)+k2×f RF (x)+k3×f SVM (x)+k4×(x) k1+k2+k3+k4=100 Among them, f NN (x) is the fitting result of the artificial neural network model, and its weight is k1; f RF (x) is the fitting result of the random forest model, and its weight is k2; f SVM (x) is the fitting result of the support vector machine model, and its weight is k3; f MARS (x) is the fitting result of the multivariate adaptive regression spline model, and its weight is k4.

Images