A method for estimating surface biomass of typical temperate grasslands

Abstract

The present invention relates to the fields of unmanned aerial vehicle (UAV) multispectral remote sensing image analysis and the construction of inversion models based on machine learning algorithms. Specifically, it relates to a method for estimating the surface biomass of typical temperate grasslands. Using a multi-rotor UAV equipped with a multispectral sensor, multi-band spectral reflectance information of pastures is acquired from near-low altitudes. A physical spectral reflectance model is constructed to calculate true and accurate reflectance data for the same spectrum band. Vegetation coverage, soil properties, and forage canopy information are recorded in the sampling area, and surface biomass is measured by cutting samples. Multiple vegetation indices and 3D spatial models are combined with a multimodal input structure. Data enhancement and small sample learning strategies are employed to construct a biomass inversion model within the MANBA machine learning framework. This method improves the accuracy of grassland biomass estimation and provides effective technical support for grassland ecological monitoring and management.

Description

Technical Field

[0001] The present invention relates to the fields of unmanned aerial vehicle (UAV) multispectral remote sensing image analysis and inversion model construction based on machine learning algorithms, and in particular to a method for estimating the surface biomass of typical temperate grasslands. Background Art

[0002] Typical grasslands play a key role in climate regulation, water conservation, and maintaining ecological balance. They also offer advantages in producing high-quality, pollution-free, and low-cost herbivorous livestock and poultry products. However, grassland degradation, desertification, and soil erosion threaten this ecological balance. Therefore, it is crucial to rapidly survey grassland resources, ensure their rational allocation, and ultimately enhance grassland management and ecological restoration capabilities.

[0003] The inventors discovered that while satellite remote sensing technology is widely used in grassland resource monitoring, its resolution limits its ability to provide detailed surface information, hindering accurate assessments of grassland conditions. Furthermore, traditional manual sampling methods, while accurate, are time-consuming and labor-intensive, making it difficult to cover large areas of grassland and resulting in incomplete data. Furthermore, models fitted using a single spectral band often exhibit inaccurate reflections of grassland conditions and fail to capture complex ecological changes. These limitations restrict resource surveys and management of grassland ecosystems. Summary of the Invention

[0004] In order to address the shortcomings of the existing technology, this paper proposes a multi-parameter fitting model based on unmanned aerial vehicle multispectral quantitative remote sensing. It combines high-resolution multispectral data from unmanned aerial vehicles with advanced machine learning algorithms and a small sample learning strategy to improve the accuracy and efficiency of grassland biomass estimation, providing a new technical means for the rational allocation of grassland resources.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] The first aspect of the present invention provides a method for estimating the surface biomass of a typical temperate grassland:

[0007] A multi-rotor drone equipped with a multispectral sensor was used for aerial mapping. Single-band spectral reflectance information for blue, green, red, red edge, and near-infrared (NIR) was obtained from the pasture at low altitude within the study area. The multi-band spectral imagery was processed and analyzed using QGIS Desktop and ENVI 5.6 professional remote sensing software. Remote sensing information of the pasture grassland was extracted, and 11 common vegetation index data were constructed. A weighted fusion method was used to obtain a highly accurate leaf area index (LAI).

[0008] Manual sampling areas were selected within the pasture, and areas with balanced grass growth and rich species were selected as sampling areas as much as possible. A 1m*1m sample box was used to select the optimal sampling range in this area, and the RTK coordinates of the center point of the sample box were recorded. The total percentage of grassland coverage within the sample box, the individual percentage of various vegetation types, the height of the vegetation canopy, the chlorophyll content of the canopy leaves, the distribution of leaf inclination, the regional soil firmness, and the soil moisture content were recorded.

[0009] All the grass in the sample box was cut evenly with scissors, and the fresh grass was placed in a kraft paper bag for fresh weight weighing. Then, it was dried in a DGF electric blast drying oven at 75°C for 48 hours until constant weight was obtained, and then the dry weight was weighed. The difference between the two weights was recorded as the aboveground biomass (AGB).

[0010] According to the RTK position information recorded in the sampling area, the corresponding mowing sampling area is found from the remote sensing image, and the leaf area index (LAI) corresponding to the 1m*1m range in the image is extracted;

[0011] The relationship between independent variables and dependent variables was set for the extracted parameter information and the measured surface biomass data, and then imported into the machine learning framework of the MANBA (Multi-Agent Nonlinear Bayesian Algorithm) model. At the same time, using the multimodal input structure of the model, the RGB image information of the sample box and the 3D spatial model data of the forage were also imported, and the small sample learning method was integrated to construct a typical grassland surface biomass inversion model.

[0012] As a further limitation of the first aspect of the present invention, when acquiring and processing remote sensing images, the following additional details are included:

[0013] First, Metashape is used to align and align aerial images for mapping, correct spectral parameters, and use ground control points to correct distortion in the stitched images and improve image processing accuracy.

[0014] Using QGIS Desktop and ENVI 5.6, we calculated and extracted image data for 11 vegetation indices across all areas of the pasture, including the Normalized Difference Vegetation Index (NDVI), Optimized Soil Adjusted Vegetation Index (OSAVI), Enhanced Vegetation Index (EVI), Red Edge Chlorophyll Index (RECI), Difference Vegetation Index (DVI), Modified Soil Adjusted Vegetation Index (MSAVI), Green Normalized Difference Vegetation Index (GNDVI), Green Chlorophyll Index (GCI), Normalized Flavonoid Index (NFBDI), Ratio Vegetation Index (RVI), and Restricted Difference Vegetation Index (RDVI). We then fused and analyzed the 11 vegetation indices, using a multi-weighted approach to obtain a more accurate leaf area index (LAI), while minimizing the effects of soil background and atmospheric spectral reflectance.

[0015] As a further limitation of the first aspect of the present invention, when collecting grassland ground data, the following additional details are included:

[0016] When selecting sampling areas, areas with balanced grass growth and rich species should be used as the main sampling areas. At the same time, areas with sparse and dense grass should also be sampled so that the overall data presents a normal distribution.

[0017] Typical temperate grasslands have a wide variety of vegetation. When recording the individual proportions of various vegetation species within the sample box, only the proportions of the following 19 species were recorded: Stipa, Artemisia frigida, Caragana microphylla, Leymus chinensis, Artemisia frigida-longflowered, Potentilla bisecta, Glechoma longituba, Chenopogon chinensis, Chenopogon chinensis (Green Chenopogon), Saposhnikovia divaricata, Saussurea davidiana, Cleistogenes scabra (Cleistogenes scabra), Artemisia annua, Altai dogwa flower, Dawuli cyperus, High two-lobed, Leontopodium lanceolata, and Yellow-flowered lanceolate.

[0018] When recording regional soil firmness and soil moisture information, data should be collected at the four corners and the middle of the sample box, and the probes of the corresponding equipment should be inserted about 20 cm underground to obtain data;

[0019] Take stereoscopic photos of various vegetation canopies in the sampling area, and then calculate the vegetation canopy height using stereo vision technology;

[0020] Broken biological samples were collected from 19 representative vegetation species in the study area, placed in sealed bags, and transported to the laboratory in a low-temperature environment. The relative chlorophyll content was extracted using a SPAD instrument and a spectrometer to obtain the chlorophyll content parameters of the canopy leaves of the 19 typical vegetation species.

[0021] The LAI-2000 optical sensor was used to measure the diffuse light intensity passing through the vegetation canopy in the sample box, and this was used to further calibrate the leaf area index (LAI). The same instrument was also used to measure the inclination angle distribution of leaves in the vegetation canopy.

[0022] As a further limitation of the first aspect of the present invention, when weighing the cut and dried forage, the following details are included:

[0023] When mowing the grass flush, the grass should be cut strictly within the scope of the sample box and the stubble height on the ground should not exceed 2cm.

[0024] Freshly cut grass should be weighed within 1 hour to avoid evaporation of moisture in the kraft paper bag, which may lead to a decrease in weighing accuracy. Similarly, dry grass taken out of the drying box should be weighed within 30 minutes to avoid prolonged exposure to the air, which may cause it to absorb moisture again and result in a higher dry weight measurement result.

[0025] As a further limitation of the first aspect of the present invention, when deeply analyzing remote sensing images and aligning them with ground data, the following additional details are included:

[0026] When searching for the corresponding mowing area in the remote sensing image, all coordinate point information was first normalized and imported into QGIS Desktop software using the Add Delimited Text Layer function to determine the sample location. The software's shape-based digital tools were then used to select the mowing sample area, thereby extracting information related to the 11 vegetation indices in the region and generating the leaf area index (LAI).

[0027] As a further limitation of the first aspect of the present invention, when constructing a biomass inversion model based on machine learning, the following details are supplemented:

[0028] The fusion of multiple vegetation indices, including leaf area index (LAI), leaf chlorophyll content, canopy height, and leaf inclination, was used as input parameters to construct a radiation transmission model using a physical model approach. This model simulates the path and energy loss of light after it enters the canopy and undergoes multiple scattering and reflections between leaves, stems, and the ground, thereby calculating the reflection status of light of different wavelengths.

[0029] The PROSAIL model is used to analyze the reflection information of different wavelengths of light and the geometric structure of the canopy (SAIL), ultimately obtaining true and accurate reflectance data for different spectral bands based on physical principles.

[0030] The model input parameters include accurate reflectance data of different spectral bands, total grassland coverage within the sample frame, the proportion of 19 plant species and their canopy leaf chlorophyll content parameters, soil firmness and soil moisture information data of the sampling area;

[0031] When using the MANBA framework to construct a surface biomass inversion model, the model input data is first preprocessed, including noise removal, standardization, and normalization, to ensure data consistency and reliability and improve the model's prediction accuracy;

[0032] All pre-processed input parameters were integrated into a multimodal dataset, and the input data were analyzed using a tree-based model and LASSO regression feature selection technique to identify the features most correlated with surface biomass.

[0033] When building a model, if the sample size is small, data augmentation techniques such as rotation, translation, and affine transformation can be used to derive relevant image training samples. Using a small sample learning framework, Siamese Networks can be used to calculate the similarity between samples for classification. Generative Adversarial Networks (GANs) are then used to generate new samples to expand the dataset. By fine-tuning existing key parameters, the model can quickly adapt to new tasks, allowing for a small number of iterative updates on small samples to improve performance on new samples. This method effectively enables rapid model training iterations when samples are scarce, improving model performance.

[0034] Under the MANBA model framework, the RGB image information of the sample plot will be used to extract edge and texture information through multi-layer convolution and pooling operations. The text information will be processed into embedding vectors through natural language processing to represent semantic information. Then, a multi-head attention mechanism will be used to fuse the image and text features into a unified feature space. The strategy of the multi-agent system is used to simulate the impact of different input features on biomass estimation. Each agent is responsible for analyzing a portion of the feature vectors in the feature space and optimizing the overall model through information sharing and collaboration.

[0035] During the training process, Monte Carlo cross-validation technology is used to ensure the stability and generalization ability of the model under different data partitions. Mean square error (MSE) and cross entropy are selected as loss functions for model training, and the RMSprop adaptive optimization algorithm is used to dynamically adjust the learning rate and optimization process to accelerate convergence.

[0036] After model training is complete, the validation set is used to evaluate model performance. The root mean square error (RMSE) is calculated to evaluate the gap between the model's predicted value and the actual value. The coefficient of determination (R²) reflects the model's ability to explain the variation in real data. Precision, recall, and F1-score are used to evaluate the model's performance in different categories for classification tasks. The model is visualized and analyzed, with scatter plots and residual plots of the predicted and actual values ​​drawn to intuitively understand the model's prediction effect. After confirming that the model performance meets expectations, it is applied to the actual grassland biomass estimation task. Multispectral remote sensing data collected by drones is used for real-time monitoring to obtain timely information on biomass changes.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] 1. The present invention combines spectral information of multiple bands, including blue, green, red, red-edge, and near-infrared, with ground vegetation coverage information and constructs a physical model to simulate light scattering and reflection conditions to generate true and accurate reflectance data of spectra in different bands. These data are then integrated into the construction of a surface biomass estimation model for typical temperate grasslands, which can significantly improve the ability to accurately measure grassland biomass and analyze its spatial distribution.

[0039] 2. Using professional software such as QGIS Desktop and ENVI 5.6, we extracted multiple vegetation indices from remote sensing images. This was combined with detailed data from the ground sampling areas, including soil firmness, moisture content, vegetation type, canopy height, canopy leaf chlorophyll content, and leaf inclination distribution, to form a multi-dimensional, multi-level data integration and analysis system.

[0040] 3. The MANBA (Multi-Agent Nonlinear Bayesian Algorithm) model was introduced, leveraging a multimodal input structure and integrating multi-source data to construct an accurate biomass inversion model. Through small-sample learning and a multi-agent system strategy, the model's adaptability and predictive accuracy were significantly improved. In the case of scarce samples, data augmentation techniques and a small-sample learning framework were employed to expand the dataset through methods such as rotation, translation, and generative adversarial networks. Monte Carlo cross-validation and adaptive optimization algorithms were used to ensure the model's stability and generalization capabilities.

[0041] 4. This invention not only focuses on the accurate estimation of biomass, but also forms a complete process system through refined data collection, in-depth analysis, and model construction. The comprehensive management method ensures the accurate and rapid investigation of grassland ecosystem resources.

[0042] Advantages of additional aspects of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] The accompanying drawings, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0044] Figure 1 The present invention provides a technical roadmap for the method of estimating the surface biomass of typical temperate grasslands based on UAV multispectral quantitative remote sensing. DETAILED DESCRIPTION

[0045] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0046] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present invention belongs.

[0047] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0048] In the absence of conflict, the embodiments of the present invention and the features thereof may be combined with each other.

[0049] The technical solutions of the embodiments of the present application will be described below in conjunction with the drawings in the embodiments of the present application.

[0050] A method for estimating surface biomass of a typical temperate grassland comprises the following steps:

[0051] S10. Use a multi-rotor drone equipped with a multispectral sensor for aerial mapping to obtain single-band spectral reflectance information for blue, green, red, red edge, and near-infrared (NIR) near-low altitude within the study pasture. Use QGIS Desktop and ENVI 5.6 professional remote sensing software to process and analyze the multi-band spectral imagery, extract remote sensing information of the pasture grassland, construct data for 11 common vegetation indices, and use a weighted fusion method to obtain a highly accurate leaf area index (LAI).

[0052] S20. Manually select sampling areas within the pasture, choosing areas with balanced grass growth and a rich variety of grass species as sampling areas. Use a 1m*1m sample box to select the optimal sampling range within this area, and record the RTK coordinates of the sample box center point. Record the total percentage of grassland cover within the sample box, the individual percentage of each type of vegetation, vegetation canopy height, canopy leaf chlorophyll content, leaf inclination distribution, regional soil firmness, and soil moisture content;

[0053] S30. Use scissors to cut all the grass in the sample box evenly, put the fresh grass into a kraft paper bag and weigh the fresh weight. Then use a DGF electric blast drying oven to dry at 75°C for 48 hours until the weight is constant, then weigh the dry weight. Record the difference between the two weights as the aboveground biomass (AGB).

[0054] S40, finding the corresponding mowing sampling area from the remote sensing image according to the RTK position information recorded in the sampling area, and extracting the leaf area index (LAI) corresponding to the 1m*1m range in the image;

[0055] S50. Set the relationship between independent variables and dependent variables for the extracted parameter information and the measured surface biomass data, and import them into the machine learning framework of the MANBA (Multi-Agent Nonlinear Bayesian Algorithm) model. At the same time, use the multimodal input structure of the model to import the sample box RGB image information and the forage 3D spatial model data together, and jointly integrate the small sample learning method to construct a typical grassland surface biomass inversion model.

[0056] The step S10 further includes:

[0057] S11. First, Metashape is used to align and align aerial images for mapping, correct spectral parameters, and use ground control points to correct the distortion of the stitched images and improve image processing accuracy.

[0058] S12. Use QGIS Desktop and ENVI 5.6 to calculate and extract image data of 11 vegetation indices in all areas of the pasture, including the normalized difference vegetation index (NDVI), optimized soil-adjusted vegetation index (OSAVI), enhanced vegetation index (EVI), red edge chlorophyll index (RECI), difference vegetation index (DVI), modified soil-adjusted vegetation index (MSAVI), green normalized difference vegetation index (GNDVI), green chlorophyll index (GCI), normalized flavonoid index (NFBDI), ratio vegetation index (RVI), and restricted difference vegetation index (RDVI);

[0059] S13. The 11 vegetation indices obtained were fused and analyzed, and a leaf area index (LAI) with high accuracy was obtained through a multi-weighted approach, while reducing the impact of soil background and atmosphere on spectral reflectance.

[0060] The step S20 further includes:

[0061] S21. When selecting sampling areas, areas with balanced grass growth and rich grass species should be the main sampling areas. Areas with sparse and dense grass should also be sampled to ensure that the overall data presents a normal distribution.

[0062] S22. The vegetation of typical temperate grasslands is diverse. When recording the individual percentages of each vegetation type within the sample box, only the percentages of the following 19 species were recorded: Stipa grassifolia, Artemisia frigida, Caragana microphylla, Leymus chinensis, Stipa grassifolia-long flower, Potentilla bisecta, Glechoma longituba, Chenopogon chinensis, Chenopogon chinensis (green quinoa), Saposhnikovia divaricata, Saussurea davidiana, Cleistogenes scabra (cleistogenes), Artemisia annua, Altai dogwas, Dawuli cyperus, Tall bifid, Leontopodium lanceolata, and lanceolate yellow flower.

[0063] S23. When recording regional soil firmness and soil moisture information, data should be collected at the four corners and the center of the sample box, and the probes of the corresponding equipment should be inserted about 20 cm underground to obtain data;

[0064] S24. Take stereoscopic photos of various vegetation canopies in the sampling area, and then calculate the vegetation canopy height using stereo vision technology;

[0065] S25. Collect damaged biological samples from 19 representative plant species in the study area, place them in sealed bags, and transport them to the laboratory in a low-temperature environment. Use a SPAD instrument and a spectrometer to extract the relative chlorophyll content and obtain the chlorophyll content parameters of the canopy leaves of the 19 typical plant species.

[0066] S26. Use the LAI-2000 optical sensor to measure the diffuse light intensity passing through the vegetation canopy in the sample frame, and use this to further calibrate the leaf area index (LAI). Also use this instrument to measure the inclination angle distribution information of the leaves in the vegetation canopy.

[0067] The step S30 further includes:

[0068] S31. When mowing the grass flush, the grass should be cut strictly within the scope of the sample box.

[0069] And the stubble height on the ground should not exceed 2cm;

[0070] S32. Newly cut grass should be weighed within 1 hour to avoid evaporation of moisture in the kraft paper bag, which may lead to a decrease in weighing accuracy. Similarly, dry grass taken out of the drying box should be weighed within 30 minutes to avoid prolonged exposure to air, which may cause it to absorb moisture again and result in a higher dry weight measurement result.

[0071] The step S40 further includes:

[0072] S41. When searching for the corresponding mowing area in the remote sensing image, all coordinate point information was first normalized and imported into QGIS Desktop software by adding a delimited text layer function to determine the sample location. The shape-based digital tool of the software was then used to select the range of the mowing sample, thereby extracting information related to the 11 vegetation indices of the region and generating the leaf area index (LAI).

[0073] The step S50 further includes:

[0074] S51. Using the leaf area index (LAI), leaf chlorophyll content, canopy height, and leaf inclination angle information, which are integrated into multiple vegetation indices, as input parameters, a radiation transmission model is constructed using a physical model method. This model simulates the path and energy loss of light after entering the vegetation canopy and undergoing multiple scattering and reflection between leaves, stems, and the ground surface, thereby calculating the reflection status information of light of different wavelengths.

[0075] S52. Use the PROSAIL model to analyze the reflectance information of different wavelengths of light and the geometric structure of the canopy (SAIL), ultimately obtaining true and accurate reflectance data for different spectral bands based on physical principles.

[0076] S53. The model input parameters include the real and accurate reflectance data of different spectral bands, the total grassland coverage within the sample frame, the proportion of 19 plant species and their canopy leaf chlorophyll content parameters, the soil firmness and soil moisture content information data of the sampling area;

[0077] S54. When using the MANBA framework to construct a surface biomass inversion model, the model input data should first be preprocessed, including noise removal, standardization, and normalization, to ensure data consistency and reliability and improve the model's prediction accuracy;

[0078] S55. All pre-processed input parameters were integrated into a multimodal dataset. The input data were analyzed using a tree-based model and LASSO regression feature selection technique to identify the features most correlated with surface biomass.

[0079] S56. When building a model, if the sample size is small, data augmentation techniques such as rotation, translation, and affine transformation can be used to derive relevant image training samples. Using a small sample learning framework, Siamese networks can be used to calculate the similarity between samples for classification. Generative adversarial networks (GANs) can then be used to generate new samples to expand the dataset. By fine-tuning existing key parameters, the model can be quickly adapted to new tasks, with a small number of iterative updates on small samples to improve performance on new samples. This method can effectively perform rapid model training iterations when samples are scarce, thereby improving model performance.

[0080] S57. Under the MANBA model framework, the RGB image information of the sample plot will be used to extract edge and texture information through multi-layer convolution and pooling operations. The text information will be processed into embedding vectors through natural language processing to represent semantic information. Then, the image and text features will be integrated into a unified feature space through a multi-head attention mechanism. The strategy of the multi-agent system is used to simulate the impact of different input features on biomass estimation. Each agent is responsible for analyzing a portion of the feature vectors in the feature space and optimizing the overall model through information sharing and collaboration.

[0081] S58. During the training process, Monte Carlo cross-validation technology is used to ensure the stability and generalization ability of the model under different data partitions. Mean square error (MSE) and cross entropy are selected as loss functions for model training, and the RMSprop adaptive optimization algorithm is used to dynamically adjust the learning rate and optimization process to accelerate convergence.

[0082] S59. After model training is completed, the validation set is used to evaluate the model performance. The root mean square error (RMSE) is used to evaluate the gap between the model's predicted value and the actual value. The coefficient of determination (R²) reflects the model's ability to explain the variation in real data. Precision, recall, and F1-score are used to evaluate the model's performance in different categories for classification tasks. The model is visualized and scatter plots and residual plots are drawn between the predicted and actual values ​​to intuitively understand the model's prediction effect. After confirming that the model performance meets expectations, it is applied to the actual grassland biomass estimation task. Multispectral remote sensing data collected by drones is used for real-time monitoring to obtain timely information on biomass changes.

Claims

1. A method for estimating surface biomass of typical temperate grasslands, characterized in that: The following steps are involved: S10. Use a multi-rotor drone equipped with a multispectral sensor to conduct aerial mapping and obtain single-band spectral reflectance information in blue, green, red, red-edge, and near-infrared near-low altitude within the pasture to be studied. Use QGIS Desktop and ENVI 5.6 professional remote sensing software to process and analyze the multi-band spectral imagery, extract remote sensing information of the pasture grassland, construct data for 11 common vegetation indices, and use a weighted fusion method to obtain the leaf area index. S20. Manually select sampling areas within the pasture, choosing areas with balanced grass growth and a rich variety of grasses as sampling areas. Use a 1m*1m sample box to select the optimal sampling range within the sampling area, and record the RTK coordinates of the sample box center point. Record the total percentage of grassland cover within the sample box, the individual percentage of each type of vegetation, vegetation canopy height, canopy leaf chlorophyll content, leaf inclination distribution, regional soil firmness, and soil moisture content; S30. Use scissors to cut all the grass in the sample box evenly, put the fresh grass into a kraft paper bag and weigh the fresh weight, then use a DGF electric hot air drying oven to dry at 75°C for 48 hours until constant weight is obtained, then weigh the dry weight, and record the difference between the two weights as the surface biomass; S40, finding the corresponding mowing sampling area from the remote sensing image according to the RTK position information recorded in the sampling area, and extracting the leaf area index corresponding to the 1m*1m range in the image; S50. Set the relationship between independent variables and dependent variables for the extracted parameter information and the measured surface biomass data, import them into the machine learning framework of the MANBA model, and use the multimodal input structure of the model to import the RGB image information of the sample box and the 3D spatial model data of the forage grass together, and jointly integrate the small sample learning method to construct a typical grassland surface biomass inversion model.

2. The biomass estimation method according to claim 1, wherein: The step S10 further includes: S11. First, Metashape is used to align and calibrate the spectral parameters of the aerial images for mapping. Ground control points are used to correct the distortion of the stitched images and improve the image processing accuracy. S12. Use QGIS Desktop and ENVI 5.6 to calculate and extract image data for 11 vegetation indices in all areas of the pasture, including the normalized difference vegetation index, optimized soil-adjusted vegetation index, enhanced vegetation index, red edge chlorophyll index, difference vegetation index improved soil-adjusted vegetation index, green normalized difference vegetation index, green chlorophyll index, normalized flavonoid index, ratio vegetation index, and restricted difference vegetation index; S13. The 11 vegetation indices obtained are fused and analyzed, and a leaf area index with high accuracy is obtained through a multi-weighted approach, while reducing the impact of soil background and atmosphere on spectral reflectance.

3. The biomass estimation method according to claim 2, characterized in that: The step S20 further includes: S21. When selecting sampling areas, areas with balanced grass growth and rich grass species should be the main sampling areas. Areas with sparse and dense grass should also be sampled to ensure that the overall data presents a normal distribution. S22. The vegetation of typical temperate grasslands is diverse. When recording the individual percentages of each vegetation type within the plot, only the percentages of the following 19 species were recorded: Stipa grassifolia, Artemisia frigida, Caragana microphylla, Leymus chinensis, Stipa grassifolia-longiflora, Potentilla bisecta, Crassula serrata, Chenopodium truncatum, Saposhnikovia divaricata, Cleistogenes scabra, Artemisia annua, Altai dogwas, Dawuli cyperus, High bifid, Leontopodium lanceolata, and lanceolate yellow flower. S23. When recording regional soil firmness and soil moisture information, data should be collected at the four corners and the center of the sample box, and the probes of the corresponding equipment should be inserted about 20 cm underground to obtain data; S24. Take stereoscopic photos of various vegetation canopies in the sampling area, and then calculate the vegetation canopy height using stereo vision technology; S25. Collect damaged biological samples from 19 representative plant species in the study area, place them in sealed bags, and transport them to the laboratory in a low-temperature environment. Use a SPAD instrument and a spectrometer to extract the relative chlorophyll content and obtain the chlorophyll content parameters of the canopy leaves of the 19 typical plant species. S26. Use the LAI-2000 optical sensor to measure the diffuse light intensity passing through the vegetation canopy in the sample frame, and use it to further calibrate the leaf area index. At the same time, measure the inclination angle distribution information of the leaves in the vegetation canopy.

4. The biomass estimation method according to claim 3, characterized in that: The step S30 further includes: S31. When mowing the grass flush against the ground, cut strictly within the sample box and ensure that the stubble height on the ground does not exceed 2 cm; S32. Newly cut grass should be weighed within 1 hour to avoid evaporation of moisture in the kraft paper bag, which may lead to a decrease in weighing accuracy. Similarly, dry grass taken out of the drying box should be weighed within 30 minutes to avoid prolonged exposure to air, which may cause it to absorb moisture again and result in a higher dry weight measurement result.

5. The biomass estimation method according to claim 4, characterized in that: The step S40 further includes: S41. When searching for the corresponding mowing area in the remote sensing image, all coordinate point information was first normalized and imported into QGIS Desktop software by adding a delimited text layer function to determine the sample location. The shape-based digital tool of the software was then used to select the range of the mowing sample, thereby extracting the relevant information of the 11 vegetation indices in the region and generating the leaf area index.

6. The biomass estimation method according to claim 5, characterized in that: The step S50 further includes: S51. Using the leaf area index, leaf chlorophyll content, vegetation canopy height, and leaf inclination information fused from multiple vegetation indices as input parameters, a physical model method is used to construct a radiation transmission model. This model simulates the paths and energy losses of light that enters the vegetation canopy and undergoes multiple scattering and reflections between leaves, stems, and the ground surface, thereby calculating the reflection status information of light of different wavelengths. S52. Use the PROSAIL model to analyze the reflection information of different wavelengths of light and the geometric structure of the canopy, and ultimately obtain true and accurate reflectance data of different spectral bands based on physical principles; S53. The model input parameters include the real and accurate reflectance data of different spectral bands, the total grassland coverage within the sample frame, the proportion of 19 plant species and their canopy leaf chlorophyll content parameters, the soil firmness and soil moisture content information data of the sampling area; S52. When constructing a surface biomass inversion model using the MANBA framework, the input data of the model should first be preprocessed, including noise removal, standardization, and normalization, to ensure the consistency and reliability of the data and improve the prediction accuracy of the model; S53. All pre-processed input parameters were integrated into a multimodal dataset. The input data were analyzed using a tree-based model and LASSO regression feature selection technique to identify the features most correlated with surface biomass. S54. When building a model, if the sample size is small, data augmentation techniques such as rotation, translation, and affine transformation can be used to derive relevant image training samples. Using a small sample learning framework, a twin network can be used to calculate the similarity between samples for classification. Then, a generative adversarial network can be used to generate new samples to expand the data set. By fine-tuning existing key parameters, the model can quickly adapt to new tasks, allowing a small number of iterative updates on small samples to improve performance on new samples. This method can effectively perform rapid model training iterations when samples are scarce, thereby improving model performance. S55. Under the MANBA model framework, the RGB image information of the sample plot will be used to extract edge and texture information through multi-layer convolution and pooling operations. The text information will be processed into embedding vectors through natural language processing to represent semantic information. Then, the image and text features will be integrated into a unified feature space through a multi-head attention mechanism. The strategy of the multi-agent system is used to simulate the impact of different input features on biomass estimation. Each agent is responsible for analyzing a portion of the feature vectors in the feature space and optimizing the overall model through information sharing and collaboration. S57. During the training process, Monte Carlo cross-validation technology is used to ensure the stability and generalization ability of the model under different data partitions. Mean square error and cross entropy are selected as loss functions for model training, and the RMSprop adaptive optimization algorithm is used to dynamically adjust the learning rate and optimization process to accelerate convergence. S58. After model training is completed, use the validation set to evaluate model performance and calculate the root mean square error (RMS): evaluates the gap between the model's predicted value and the actual value; the coefficient of determination: reflects the model's ability to explain the variation in real data; precision, recall, and F1-score: for classification tasks, evaluate the model's performance in different categories; perform a visual analysis of the model, draw scatter plots and residual plots of the predicted and actual values, and intuitively understand the model's prediction effect. After confirming that the model performance meets expectations, apply it to the actual grassland biomass estimation task, use multispectral remote sensing data collected by drones for real-time monitoring, and obtain timely information on biomass changes.

Images