Method for obtaining consumption yield and related device

Abstract

The application discloses an acquisition method of consumed grass yield and related equipment, relates to the technical field of agricultural remote sensing, precise livestock raising and ecological monitoring, and comprises the following steps: collecting multi-band image data of a preset area by a multispectral sensor carried on a UAV and performing pretreatment; obtaining the surface reflectivity of each pixel in the pretreated multi-band image data, then calculating at least one vegetation index of each pixel, and inversely calculating at least one environmental factor of each pixel based on the surface reflectivity of each pixel; inputting the vegetation index and the environmental factor of each pixel into a trained machine learning model to obtain the residual grass yield of each pixel; inputting the environmental factor of each pixel and pretreated meteorological data into a light energy utilization rate model to obtain the theoretical grass yield of each pixel, and then calculating the consumed grass yield of the preset area. The application provides a scientific and quantitative basis for fine grassland management and ecological protection.

Description

Technical Field

[0001] This invention relates to the fields of agricultural remote sensing, precision animal husbandry and ecological monitoring, and in particular to a method and related equipment for obtaining the amount of grass consumed. Background Technology

[0002] In the practice of refined grazing management and grassland ecological protection, accurately quantifying the actual forage production consumed by herbivores within a certain period, i.e., the amount of forage consumed, is a key prerequisite for assessing grassland carrying capacity and achieving dynamic regulation of the grass-livestock balance. However, due to the spatial heterogeneity and complexity of animal grazing activities, how to quickly, accurately, and non-destructively obtain data on the amount of forage consumed over a large area of ​​grassland has always been a technical challenge.

[0003] Currently, the methods for obtaining grassland consumption and forage yield mainly rely on two types of approaches. The first type is an estimation method based on statistical data and ground surveys. Specifically, this involves obtaining livestock inventory data from county-level livestock statistical yearbooks, estimating total demand by combining this with theoretical daily feed intake, obtaining average forage yield through sampling plot harvesting, and then extrapolating the forage-livestock balance. The second type is an indirect assessment using remote sensing technology. For example, this involves using satellite or drone platforms to obtain vegetation indices, establishing empirical relationship models between vegetation indices and aboveground biomass, and inferring consumption by comparing differences in biomass before and after grazing or in different areas.

[0004] However, the aforementioned existing methods have significant limitations. Statistical data-based estimation methods heavily rely on the accuracy of livestock data, which is typically only at the county level and cannot reflect the spatial distribution differences of livestock within pastures. Furthermore, they completely ignore the consumption by wild herbivores and insects, leading to an underestimation of the actual total consumption. In addition, these methods cannot achieve rapid, dynamic diagnosis. On the other hand, while traditional remote sensing assessment methods provide spatial information, they are often limited to using simple vegetation indices to invert total biomass, making it difficult to effectively distinguish between theoretical grass yield under no grazing pressure and actual residual grass yield after grazing. Therefore, they cannot directly and accurately calculate the true consumption. Simultaneously, existing remote sensing methods fail to fully integrate key environmental factors affecting grassland growth and do not address the bottlenecks in multispectral data acquisition quality and processing efficiency.

[0005] Therefore, existing technologies are insufficient to meet the urgent need for high-precision, high-efficiency, and spatially clear measurement of grassland consumption and yield. There is a pressing need for a new technical solution that can integrate multi-source data, accurately separate theoretical yield from actual consumption, and enable rapid processing and analysis. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to address the shortcomings of the prior art, and specifically provides a method and related equipment for obtaining the amount of consumed forage, as detailed below: 1) In a first aspect, the present invention provides a method for obtaining the amount of consumed forage, the specific technical solution of which is as follows: Multi-band image data of a preset area is acquired using a multispectral sensor mounted on a drone. The multi-band image data is preprocessed. The surface reflectance of each pixel in the preprocessed multi-band image data is obtained. Based on the surface reflectance of each pixel, at least one vegetation index is calculated for each pixel, and at least one environmental factor for each pixel is inverted based on the surface reflectance of each pixel. The vegetation index and environmental factor of each pixel are input into a trained machine learning model to obtain the residual grass yield of each pixel. The environmental factor of each pixel and the preprocessed meteorological data are input into a light energy utilization model to obtain the theoretical grass yield of each pixel. The difference between the theoretical grass yield and the residual grass yield of each pixel is taken as the consumed grass yield of the corresponding pixel. The sum of the consumed grass yields of each pixel is taken as the consumed grass yield of the preset area.

[0007] The beneficial effects of the method for obtaining the amount of consumed grass provided by this invention are as follows: By acquiring the theoretical and residual grass yields for each pixel and calculating the difference, the pixel-level grass consumption is directly obtained, overcoming the shortcomings of traditional methods that rely on county-level statistical yearbook data and cannot characterize the spatial heterogeneity within pastures. The calculation results reflect the actual spatial distribution of grass consumption, rather than a regional average. Furthermore, since the theoretical grass yield simulates the potential yield under no grazing pressure, and the residual grass yield represents the actual remaining amount after grazing, the difference objectively quantifies the actual total consumption by multiple consumers, including livestock, wild herbivores, and herbivorous insects, correcting the problem of underestimating carrying capacity pressure due to neglecting wildlife consumption in traditional methods. In addition, the entire process, based on UAV multispectral remote sensing and automated model calculations, enables rapid and non-destructive monitoring and assessment of grass consumption in a pre-defined area, providing direct and scientific quantitative evidence for refined grassland management and ecological protection.

[0008] 2) In a second aspect, the present invention also provides a system for obtaining the amount of grass produced, the specific technical solution of which is as follows: The system includes a data acquisition module, an image data preprocessing module, a surface reflectance acquisition module, an index factor calculation module, a grass yield calculation module, a grass consumption calculation module, and a total grass consumption calculation module. The data acquisition module is used to acquire multi-band image data of a preset area using a multispectral sensor mounted on a drone. The image data preprocessing module is used to preprocess the multi-band image data. The surface reflectance acquisition module is used to acquire the surface reflectance of each pixel in the preprocessed multi-band image data. The index factor calculation module is used to calculate at least one vegetation index for each pixel based on its surface reflectance. Based on the surface reflectance of each pixel, at least one environmental factor for each pixel is inverted; the vegetation yield calculation module is used to: input the vegetation index and environmental factor of each pixel into the trained machine learning model to obtain the residual vegetation yield of each pixel; input the environmental factor of each pixel and the preprocessed meteorological data into the light energy utilization model to obtain the theoretical vegetation yield of each pixel; the vegetation consumption calculation module is used to: take the difference between the theoretical vegetation yield and the residual vegetation yield of each pixel as the vegetation consumption of the corresponding pixel; the total vegetation consumption calculation module is used to: take the sum of the vegetation consumption of each pixel as the vegetation consumption of the preset area.

[0009] 3) In a third aspect, the present invention also provides an electronic device, the electronic device including a processor coupled to a memory, the memory storing at least one computer program, the at least one computer program being loaded and executed by the processor, so that the electronic device implements any of the above-mentioned methods for obtaining the amount of grass consumed.

[0010] 4) In a fourth aspect, the present invention also provides a computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements any of the above-mentioned methods for obtaining the amount of grass consumed.

[0011] It should be noted that the beneficial effects of the technical solutions of the second to fourth aspects of the present invention and their corresponding possible implementations can be found in the above description of the technical effects of the first aspect and its corresponding possible implementations, and will not be repeated here. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments of the present invention will be briefly introduced below: Figure 1 This is a flowchart illustrating a method for obtaining the amount of consumed grass according to an embodiment of the present invention; Figure 2 This is a schematic diagram of a system for obtaining the amount of grass produced according to an embodiment of the present invention. Detailed Implementation

[0013] The principles and features of the present invention are described below. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0014] The technical solution of the present invention and how the technical solution of the present invention solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of the present invention will now be described with reference to the accompanying drawings.

[0015] like Figure 1 As shown, an embodiment of the present invention provides a method for obtaining the amount of consumed grass, which includes the following steps: S1. Multi-band image data of a preset area is collected using a multispectral sensor mounted on a drone. The specific implementation process is as follows: S10. Determine and input the spatial range of the preset area. The operator uses geographic information system software or professional UAV flight mission planning software to clearly delineate the grassland area where multi-band image data needs to be collected, i.e., the preset area, by drawing a polygon on the map interface or directly importing boundary coordinate files. This step ensures that the subsequent flight path can completely cover the target area and may be optimized according to the shape of the area to avoid invalid flights.

[0016] S11. Conduct UAV flight mission planning. In the flight planning software, use the boundaries of the preset area as a basis to set specific flight parameters. Key flight parameters include flight altitude, forward overlap, and lateral overlap. Flight altitude is typically set between 80 and 120 meters. This range allows for obtaining high-resolution ground images with a resolution of approximately 3cm to 5cm while ensuring safety. Forward overlap is set to no less than 80%, and lateral overlap to no less than 75%. The high overlap setting is for subsequent generation of high-precision orthophotos and 3D models using photogrammetry. Furthermore, the mission should be carried out during periods of clear weather, stable lighting, and low wind speeds; the optimal time window is usually around noon when the local solar altitude angle is high.

[0017] S12. Perform automated data acquisition flight. Following the route planned in step two, the UAV autonomously takes off, flies, and completes data acquisition. During flight, the UAV's onboard multispectral sensor operates using a global shutter, ensuring strict synchronization of images across all spectral bands at the exposure time, eliminating pixel misalignment issues between different bands that may occur due to time-sharing exposures. The multispectral sensor continuously acquires raw multispectral image data and stores it in common lossless or low-compression formats. Simultaneously, the UAV's built-in light intensity sensor synchronously records the total solar irradiance data in each spectral band at each exposure time; this data is a crucial input for subsequent accurate radiometric calibration.

[0018] S13. During data acquisition, the intelligent real-time dimming algorithm runs continuously. This algorithm analyzes the brightness histogram distribution of each band in the current frame image captured by the multispectral sensor in real time. When the system detects that the maximum digital quantization value of a pixel in one or more bands of the image is close to the sensor's saturation limit, for example, reaching 85% of the saturation threshold, the algorithm automatically and slightly adjusts the sensor's exposure time or sensitivity parameters. The adjustment strategy prioritizes ensuring that the red-edge band and near-infrared band, which are most sensitive to vegetation analysis, are not overexposed, thereby preserving detailed information of highly reflective vegetation areas. All adjustments to exposure parameters are accurately recorded and compensated for in subsequent radiometric calibration processing to ensure the physical accuracy of the final reflectance data. This process effectively copes with drastic changes in ground lighting conditions caused by passing clouds during flight, improving data availability under complex weather conditions.

[0019] S14. Preliminary Data Processing and Storage. After completing the flight path within the preset area, the UAV will land, and the operator will export the multi-band image data, corresponding geographic location information, and synchronously recorded solar irradiance data stored in the multispectral sensor and UAV flight controller. This data will be organized according to preset naming rules to prepare for the next step of data preprocessing. The entire acquisition process achieves rapid, synchronous, and high-quality acquisition of multi-band image data for the preset area.

[0020] A multispectral sensor is an imaging device capable of simultaneously capturing the reflected light information of ground objects in multiple discrete, narrow spectral bands. Unlike ordinary cameras that only record red, green, and blue light, multispectral sensors typically include additional red-edge and near-infrared bands, which are particularly crucial for vegetation monitoring. Each band corresponds to a specific wavelength range; for example, the center wavelength of the blue band is approximately 450 nm, the green band is approximately 560 nm, the red band is approximately 650 nm, the red-edge band is approximately 730 nm, and the near-infrared band is approximately 840 nm. By acquiring information from these specific bands, various vegetation indices can be calculated, thereby more accurately retrieving the physiological and biochemical parameters of vegetation.

[0021] The pre-defined area refers to the target grassland region for which grazing carrying capacity assessment is required. This area is clearly defined before data collection, typically based on management boundaries, ecological functional zoning, or research plot design. In the field, the boundary coordinates of the pre-defined area are input into flight planning software as the spatial basis for the drone's automatic flight. Clearly defining the pre-defined area is a prerequisite for ensuring complete data collection coverage and spatially specific analysis results.

[0022] Multiband image data refers to a set of spatially aligned images generated synchronously by a multispectral sensor during a single imaging process. Each image records the reflectance intensity information of the same ground feature scene in a specific spectral band. For example, a single acquisition may simultaneously obtain a blue band image, a green band image, a red band image, a red-edge band image, and a near-infrared band image. Each pixel in these images contains a digitally quantified value of its corresponding geographical location. This information can be converted into reflectance after subsequent processing, serving as the foundational data for quantitative remote sensing analysis.

[0023] Specifically, when using a multispectral sensor to acquire multi-band image data of a preset area, the exposure parameters of the multispectral sensor are dynamically adjusted through an intelligent real-time dimming algorithm. The specific implementation process is as follows: 1) Before the drone takes off and begins its data collection mission, the intelligent real-time dimming algorithm initializes. It loads a set of preset baseline exposure parameters based on sensor characteristics and the typical reflectance spectrum of grass, including initial exposure time and ISO. Simultaneously, the algorithm sets key control thresholds, the most important being a saturation safety threshold, such as 85% of the sensor's maximum recordable digital quantization value. This threshold serves as a warning line to trigger exposure parameter adjustments. The algorithm also defines adjustment step sizes and priority rules, such as specifying small increments in exposure time for each adjustment, and prioritizing the exposure suitability of the red-edge and near-infrared bands when multiple bands require adjustment.

[0024] 2) As the UAV flies along the planned route and continuously collects multi-band image data, the intelligent real-time dimming algorithm starts working simultaneously. For each newly captured frame of multi-band image data, the algorithm immediately reads the raw digital quantization value matrix of each band image from the sensor. Subsequently, the algorithm performs real-time statistics on each band image in parallel, calculating the histogram distribution of the digital quantization values ​​of all pixels in that band image. By analyzing the histogram, the algorithm quickly extracts key statistics such as the maximum and average digital quantization values ​​of the current frame for each band image. This monitoring process is continuous, providing a data foundation for real-time judgment of changes in lighting conditions.

[0025] 3) The maximum digital quantization value of each acquired band image is compared with the preset saturation safety threshold from the first step. The judgment logic is: if the maximum digital quantization value of any band, especially the red-edge band or near-infrared band, reaches the saturation safety threshold, it is determined that there is a risk of overexposure under the current exposure parameters, and the data quality of that frame image may be damaged. Once the judgment condition is triggered, the algorithm immediately enters the adjustment decision stage. The decision logic is based on preset priorities: a new, slightly reduced exposure time or ISO value is calculated, and the adjustment range is made according to a preset step size. The decision process prioritizes reducing the overexposure risk of the red-edge and near-infrared bands, because these two bands are most sensitive to vegetation biomass inversion. At the same time, the algorithm evaluates the impact of the adjustment on other bands, striving to ensure that the digital quantization values ​​of all bands are generally within a reasonable range under the new parameters.

[0026] 4) After the adjustment decision is generated, the algorithm sends the new exposure time or sensitivity parameter values ​​to the sensor in real time through the control interface between the UAV flight control system and the multispectral sensor. Upon receiving the instruction, the sensor immediately applies the adjusted parameters when acquiring the next frame of imagery. This process is completed in milliseconds, ensuring a very fast response speed from detecting changes in illumination to completing parameter adjustments, effectively tracking rapid changes in surface illumination such as cloud shadows. Simultaneously, the algorithm generates a detailed adjustment log record, including at least a timestamp, the exposure parameters before adjustment, the exposure parameters after adjustment, the band that triggered the adjustment, and the corresponding maximum digital quantization value. These records are completely saved.

[0027] 5) The intelligent real-time dimming algorithm not only dynamically adjusts parameters but also provides crucial compensation information for subsequent radiometric calibration. According to photographic principles, reflectance calculation relies on accurate exposure parameters. The algorithm automatically calculates the corresponding radiometric compensation coefficient using the exposure parameter adjustment logs recorded in step four. For example, if the exposure time is from... Adjusted to Therefore, for images acquired within this time period, a compensation coefficient proportional to the change in exposure time needs to be introduced into the radiometric calibration formula. .here, Indicates the exposure time before the adjustment. This indicates the adjusted exposure time. This represents the exposure change compensation coefficient. The algorithm closely associates and stores these calculated compensation coefficients with the corresponding multi-band image data files, ensuring that in the subsequent radiometric calibration process, the correct parameters can be used to restore the digital quantization value to the accurate apparent reflectance, thereby guaranteeing the physical accuracy of the final reflectance data from the source.

[0028] In multispectral sensors, exposure parameters refer to the set of adjustable variables that control the amount of light entering the photosensitive element of each independent spectral band during imaging. These parameters mainly include exposure time and photosensitivity. Exposure time determines the duration of light exposure to the photosensitive element, typically measured in milliseconds; photosensitivity represents the sensitivity of the photosensitive element to light signals, commonly expressed as ISO values. By coordinating the adjustment of these parameters, the level of the raw digital quantization value output by the sensor can be controlled, preventing information loss due to excessively strong light causing the digital quantization value to reach the sensor's limit, or excessively weak light causing the digital quantization value to be submerged in noise. Real-time optimization of these parameters under dynamic lighting conditions is crucial for acquiring high-quality, usable multispectral image data.

[0029] S2. Preprocess the multi-band image data, specifically performing spatiotemporal alignment, radiometric calibration, and atmospheric correction. The specific implementation process is as follows: S20. The goal of spatiotemporal alignment is to ensure that multiple single-band images (such as blue, green, red, red-edge, and near-infrared images) acquired in a single acquisition by a multispectral sensor are perfectly matched in spatial pixel location and unified under a consistent geographic coordinate system. This process relies on a combination of hardware and algorithms. Because this invention uses an integrated sensor, all bands share the same lens and imaging plane, thus ensuring strict synchronization of images at the moment of exposure at the hardware level, eliminating time and viewing angle differences that may result from time-division acquisition. In post-processing, high-precision POS data recorded by the UAV and real-time dynamic differential positioning system information, including latitude, longitude, altitude, and attitude angle at the time of each image capture, are used. This data is then processed using a motion reconstruction algorithm in photogrammetry software. The motion reconstruction algorithm automatically matches corresponding feature points in all band images and calculates precise camera intrinsic and extrinsic parameters, thereby generating a precise 3D point cloud and digital surface model shared by all bands. Based on this unified spatial model, the original images for each band are geometrically corrected and resampled to generate orthophotos with accurate geographic coordinates. The final output is an independent orthophoto for each band with strictly aligned pixel positions. The geographic extent of all images is completely consistent with the coordinate reference, providing a guarantee for subsequent pixel-level exponential calculations and model inversion.

[0030] S21. Radiometric calibration is the process of converting the raw digital quantization value of each pixel in a spatiotemporally aligned orthophoto image of each band into physically meaningful apparent reflectance. Apparent reflectance refers to the ratio of the energy of radiation reflected from ground objects received by the sensor at the top of the atmosphere to the energy of incident solar radiation. The calibration process requires two key inputs: one is the sensor's inherent radiometric calibration coefficient, and the other is the actual solar irradiance data at the time of acquisition. The calibration formula is: ,in, The apparent reflectivity of band λ is a dimensionless quantity, typically ranging from 0 to 1. This represents the original digital quantization value of a pixel in a λ-band image; it is an integer value recorded by the sensor. The sensor calibration coefficient for band λ is a conversion factor that converts a digital quantization value into the radiance at the sensor's entrance pupil. It is usually provided by the sensor manufacturer at the factory or obtained through laboratory calibration or on-site calculation using data from a ground calibration board. This represents the solar irradiance in band λ at the moment of image acquisition, measured in watts per square meter per micrometer. This data is measured and recorded in real-time by the light intensity sensor built into the drone during flight.

[0031] When applying this formula, for each spatiotemporally aligned single-band orthophoto, the data for each pixel is read. Value, and call the corresponding And the time recorded when the image was taken The value is obtained by calculation for that pixel. This step eliminates the influence of the sensor's own response characteristics and changes in sunlight intensity at the time, making data collected at different times and from different devices comparable on a radiometric scale. In particular, if an intelligent real-time dimming algorithm is used and exposure parameters are adjusted during the acquisition process, then the formula... value or The coefficients need to be corrected accordingly based on the compensation coefficients recorded by the algorithm to ensure the physical accuracy of the calculation.

[0032] S22. The goal of atmospheric correction is to eliminate the scattering and absorption of light by atmospheric molecules and aerosols, and to reduce the apparent reflectivity at the top of the atmosphere. Converted to the true surface reflectance of the Earth's surface Surface reflectivity is an inherent property of ground features, unaffected by atmospheric conditions, and serves as a direct input for quantitative analysis and model inversion. This invention employs a simplified method based on the principle of radiative transfer, such as dark target subtraction, to achieve atmospheric correction. The correction formula is: ,in, The reflectance of the Earth's surface in band λ is the final result after atmospheric correction. This represents the apparent reflectivity of band λ, i.e., the output result after radiometric calibration. This represents the path radiance value for band λ, also known as the haze value. It indicates a portion of the radiation that the sensor can receive even if the ground is a perfect blackbody due to atmospheric influence. Operationally, it is usually estimated by selecting a known dark target area in the image (such as clear deep water or dense shadows) and taking its minimum apparent reflectance. This represents the atmospheric transmittance of band λ, a value between 0 and 1. It indicates the proportion of sunlight that passes through the atmosphere, reaches the ground, and is reflected back to the sensor without being scattered or absorbed. This value can be estimated using atmospheric model lookup tables based on meteorological observation data (such as visibility) at the time and location of acquisition, or by using empirical values. The process involves, for radiometrically calibrated images of each band, first automatically or manually identifying dark target areas, and then statistically calculating the transmittance of each band. Then, based on the meteorological conditions during the flight operation, the frequency bands are determined or estimated. Finally, each pixel's Value, corresponding and Substituting into the above formula, calculations are performed pixel by pixel to obtain the surface reflectance image for each band. After atmospheric correction, the multi-band image data more accurately reflects the physicochemical properties of ground vegetation and soil, laying a reliable physical foundation for subsequent high-precision feature extraction and model calculation.

[0033] By sequentially performing three steps—spatial-temporal alignment, radiometric calibration, and atmospheric correction—the original multi-band image data is transformed into a spatially precisely aligned data cube with real physical surface reflectance values, thus meeting the stringent data quality requirements of precision agriculture remote sensing quantitative analysis.

[0034] S3. Obtain the surface reflectance of each pixel in the preprocessed multi-band image data. The specific implementation process is as follows: S30. After atmospheric correction, the data processing system outputs a separate digital image file for each spectral band. These files are the final output of the preprocessing workflow. It searches for and confirms all necessary surface reflectance image files from a predefined project directory or database table, following the band naming rules. Typical bands include blue, green, red, red-edge, and near-infrared bands. Using a geospatial data read / write library, these GeoTIFF files are opened one by one. During loading, not only is the two-dimensional numerical matrix containing surface reflectance values ​​read, but also the embedded geographic metadata in the file is extracted and verified, including the top-left corner coordinates, cell size, number of rows, number of columns, and coordinate system definition. This spatial metadata of all band images is compared to ensure complete consistency, a prerequisite for subsequent cell-level data fusion. The surface reflectance matrices of all bands are loaded into a specific data buffer in the computer's memory for high-speed random access.

[0035] S31. Since all imagery bands share a strictly consistent spatial reference, a unified, two-way coordinate transformation index can be established. For the entire geographic area covered by the imagery, a regular grid is generated based on the ground resolution of the pixels. Each grid cell corresponds to one pixel. The key to achieving this mapping is using affine transformation parameters, which are typically obtained from the image metadata. The affine transformation formula is as follows: in, and Geographic coordinates (e.g., longitude and latitude, or projected coordinates) representing the center point of a pixel. and This indicates the column and row number of the pixel in the image matrix; the index usually starts from 0. This indicates the X-coordinate of the center point of the top-left pixel in the image. This indicates the pixel size in the X direction (ground resolution). This represents the rotation component of the row, which is typically 0 in orthophotos. This represents the Y-coordinate of the center point of the top-left pixel in the image. This represents the rotation component of the column, which is typically 0 in orthophotos. This represents the cell size in the Y direction, usually a negative value indicating that the Y coordinate increases downwards. This transformation relationship is established and stored in memory. When processing cells at a specific geographic location, the inverse transformation of this formula is used to obtain the pixel size from the geographic coordinates. Quickly calculate its corresponding row and column numbers This allows the same spatial unit to be located in the matrix across all bands.

[0036] S32. Access each valid cell in the image using a sequential traversal or block traversal strategy. For the currently being processed cell located at the... Line number For each cell in a column, a synchronous read operation is performed. This operation accesses the two-dimensional data array in memory for each loaded band and extracts the cells located in the array. The element value at a given location. The extracted values ​​constitute the spectral reflectance vector of that pixel. Mathematically, for a given set of elements containing... For a system with multiple bands, this vector is represented as: in, Indicates the first Line number Multiband surface reflectivity column vector of column pixels. This indicates that the cell is in the th order. bands (e.g.) (For the blue band) Surface reflectance values. Each They are all based on an atmospheric correction formula. The calculated floating-point number. (Superscript) This represents the transpose of a vector. For example, for a five-band system, the vector of a vegetation cover cell might be... This clearly demonstrates the typical characteristics of vegetation: low reflectivity in the visible light band and high reflectivity in the near-infrared and red-edge bands.

[0037] S33. Construct and manage the pixel spectral database. To improve the efficiency of subsequent feature extraction, the massive extracted pixel spectral vectors need to be effectively organized. One efficient memory organization method is to construct a three-dimensional data cube. The dimensions of this cube are... Through this data structure, the value of any pixel in any band can be obtained... One approach is to provide instantaneous access. Another object-oriented or database-driven approach is to create a structure or record for each cell, containing its unique identifier, geographic coordinates, and spectral vector. The organized data is then stored in memory or a cache, and a fast query mechanism is established. For large areas exceeding memory capacity, a chunking strategy is employed, dividing the large image into regular tiles, performing the extraction and organization operations for each tile, and writing the results to a temporary file or database.

[0038] S34. After extracting and organizing the surface reflectance of all pixels, an automated quality check process is implemented. Verification includes several aspects. It involves calculating the global minimum, maximum, mean, and standard deviation of surface reflectance for all pixels in each band, checking for any abnormal deviations from reasonable physical ranges. It also involves judging the reasonableness of spectral curves, for example, by randomly sampling a batch of pixels, plotting their spectral curves, and procedurally checking whether their shapes conform to the spectral characteristics of typical land features such as vegetation, soil, and water bodies. Furthermore, it checks for missing or invalid values ​​to ensure the integrity of the data matrix. If an abnormal value is found in a certain pixel in a certain band, it will be marked or interpolated according to a predetermined strategy. This verification step is a crucial step in ensuring that data quality is not degraded throughout the pipeline from preprocessing to feature extraction, providing a high-quality data source for generating reliable normalized difference vegetation indices and retrieving environmental factors.

[0039] In multi-band image data, pixels are the most basic building blocks and carriers of spatial information. Each preprocessed orthophoto consists of a two-dimensional matrix, and each smallest square in the matrix is ​​called a pixel. Each pixel represents a region of a specific size and shape on the ground, the size of which is determined by the ground resolution of the image. For a specific pixel location, it corresponds to a value in each image of a different spectral band; for example, it has one numerical value in the blue band image and another numerical value in the near-infrared band image. Therefore, each pixel in multi-band image data is essentially a vector containing information in multiple spectral dimensions, serving as a key node connecting spatial location and spectral attributes. All subsequent feature calculations and model inversions will be performed independently on a pixel-by-pixel basis.

[0040] Surface reflectance of a pixel is a key physical quantity describing the ability of the ground unit it represents to reflect solar radiation. It is a dimensionless value, typically ranging from 0 to 1. The specific value of surface reflectance represents the ratio of the radiant energy reflected by a ground object to the total solar radiant energy incident on that object in a specific spectral band. This value has undergone radiometric calibration and atmospheric correction, removing the influence of sensor characteristics and atmospheric scattering and absorption, thus accurately reflecting the physicochemical properties of the ground object, such as chlorophyll content, moisture status, and cell structure. At a specific pixel, surface reflectance is stored as a floating-point number and uniquely corresponds to its represented ground location through precise geocoding. It is the absolute fundamental data input for calculating all vegetation indices and for quantitative remote sensing model inversion.

[0041] S4. Calculate at least one vegetation index for each pixel based on the surface reflectance of each pixel, and invert at least one environmental factor for each pixel based on the surface reflectance of each pixel. Among them, at least one vegetation index includes at least one of the Normalized Differential Vegetation Index (NDVI), the Normalized Differential Red Edge Index (NDRE), and the Soil-Regulated Vegetation Index (SAVI), and at least one environmental factor includes at least one of air temperature, relative humidity, and soil texture.

[0042] The specific implementation process for calculating at least one vegetation index for each pixel is as follows: S40. Before starting the calculation, ensure that the multi-band surface reflectance data for each pixel has been correctly acquired and organized. This data is typically stored in memory as a three-dimensional data cube, where the first two dimensions represent the rows and columns of the image, and the third dimension represents the spectral bands. The specific band sequence corresponding to each dimension must be clearly defined. For example, confirm that layer 0 of the data cube represents the blue band surface reflectance. The first layer is the surface reflectivity in the green band. The second layer is the surface reflectivity in the red band. The third layer is the surface reflectivity in the red-edge band. The fourth layer is the surface reflectivity in the near-infrared band. For each vegetation index that needs to be calculated, a specific band index or name on which its calculation depends will be predefined, and the data for these bands will be verified to be valid and complete within the specified spatial range.

[0043] S41. The calculation of the Normalized Difference Vegetation Index (NDVI) is performed independently for each pixel. For row and column coordinates... From the pixels, two key physical quantities are extracted from their surface reflectance data: near-infrared surface reflectance. and red band surface reflectivity The formula for calculating the Normalized Difference Vegetation Index is: in, Indicates the first Line number Normalized difference vegetation index values ​​of the column pixels. This indicates the surface reflectance of the pixel in the near-infrared band. This indicates the surface reflectance of the red band of that pixel.

[0044] The process iterates through each pixel in the image, performing the aforementioned subtraction, addition, and division operations sequentially. In program implementation, this is typically represented by parallel element-wise calculations of the reflectance matrices for the entire red and near-infrared bands. Calculation results... It is a real number between -1 and 1. A higher value usually indicates that the vegetation at that pixel location is denser and has a larger biomass. The calculation results for each pixel are stored in a new two-dimensional matrix with the same number of rows and columns as the original image, thus generating the normalized difference vegetation index distribution map.

[0045] S42. The calculation process for the Normalized Differential Red Edge Index is similar to that of the Normalized Differential Vegetation Index, but the band combination used is different. For pixels... Extract its near-infrared surface reflectance and red-edge band surface reflectivity The formula for calculating the normalized difference red-edge index is: in, Indicates the first Line number Normalized difference red-edge index value of column pixels. This indicates the surface reflectance of the pixel in the near-infrared band. This indicates the surface reflectivity of the red-edge band of that pixel.

[0046] Similarly, calculations are performed by traversing all pixels. The red-edge band is more sensitive to the chlorophyll content and nitrogen status of vegetation, especially in areas with high vegetation cover. The normalized difference red-edge index is more effective than the normalized difference vegetation index in avoiding signal saturation. Calculation results It is also stored in an independent two-dimensional matrix to form a normalized difference red edge index distribution map.

[0047] S43. The calculation of the soil-modified vegetation index aims to reduce the interference of bare soil background on the vegetation index signal. Calculation of this index requires three inputs: near-infrared surface reflectance. Red band surface reflectivity And a preset soil adjustment coefficient The formula for calculating the soil-regulated vegetation index is: in, Indicates the first Line number Soil-regulated vegetation index values ​​for each pixel. This indicates the surface reflectance of the pixel in the near-infrared band. This indicates the surface reflectance of the red band of that pixel. This represents the soil conditioning coefficient, an empirical constant typically ranging from 0 to 1. For grasslands with moderate vegetation cover, The value is usually set to 0.5; however, different values ​​can be configured depending on the specific conditions of the study area. Value. During calculation, the molecule is calculated first. Then calculate the denominator. Finally, a division operation is performed. After traversing all pixels, a distribution map of the soil-regulated vegetation index is generated.

[0048] S44. After completing the calculation of all specified vegetation indices, the generated Normalized Difference Vegetation Index (NDVE) matrix, Normalized Difference Red Edge Index (NDRE) matrix, and Soil-Regulated Vegetation Index (SRE) matrix are integrated and managed. Integration can be achieved by stacking them as multiple "layers" into a new multidimensional data product. Subsequently, automated verification checks are performed. Verification includes checking whether the numerical ranges of each index matrix conform to theoretical expectations; for example, checking whether the majority of the Normalized Difference Vegetation Index (NDVE) values ​​fall within the reasonable vegetation range of 0.2 to 0.8. Spatial consistency checks are also performed, comparing whether the vegetation spatial patterns reflected by different indices in the same area have a reasonable correlation. Abnormal pixels with a division by zero that may occur during the calculation process are marked and filled with null values ​​or specific codes. Finally, these calculated and verified vegetation index matrices will serve as important input features for subsequent calculations of theoretical grass yield models and light energy utilization models.

[0049] The specific implementation process for obtaining at least one environmental factor for each pixel through inversion includes: S45. Before conducting the specific inversion, two aspects of data preparation are required. The first aspect is preparing the measured dataset for training and validation. For air temperature and relative humidity, it is necessary to collect measured air temperature and relative humidity values ​​recorded by multiple meteorological stations within and around the study area during the synchronous or quasi-synchronous flight of the UAV, ensuring that these stations have accurate geographical coordinates. For soil texture, it is necessary to collect soil samples from bare surfaces or areas with extremely low vegetation cover, obtain the percentage content of sand, silt, and clay particles through laboratory analysis, and record the location coordinates of the sampling points. The second aspect is constructing the spectral features used for inversion. Read the multi-band surface reflectance of each pixel: , , , , , It represents the surface reflectivity in the blue band, with a center wavelength of approximately 450 nm. This value quantifies the reflectivity of the ground unit represented by the pixel in the blue light band. This indicates the surface reflectivity in the green band, with a center wavelength of approximately 560 nm. This value quantifies the reflectivity of the ground unit represented by the pixel in the green band. This indicates the surface reflectivity in the red band, with a center wavelength of approximately 650 nm. This value quantifies the reflectivity of the ground unit represented by the pixel in the red band. The red-edge band surface reflectance, with a center wavelength of approximately 730 nm, quantifies the reflectance characteristics of the ground unit represented by the pixel in the red-edge band, i.e., the transition region at the edge of chlorophyll absorption, and is particularly sensitive to the physiological state of vegetation. This represents the surface reflectance in the near-infrared band, with a center wavelength of approximately 840 nm. This value quantifies the reflectance characteristics of the ground unit represented by the pixel in the near-infrared band and is highly correlated with the cellular structure and biomass of vegetation. A series of derived spectral indices are calculated, such as the Normalized Difference Vegetation Index (NDVEI) and the Normalized Difference Red Edge Index (NDRI). These original band reflectance values ​​and derived indices will serve as input features for machine learning models or spectral matching algorithms. The coordinates of meteorological stations or soil sampling points are matched with multispectral orthophotos to extract all spectral features of the corresponding pixel locations, forming a "feature-label" paired sample dataset.

[0050] S46. The inversion of air temperature and relative humidity is accomplished using a supervised machine learning algorithm. Its core is to establish a nonlinear mapping function from spectral features to environmental factors. Taking the random forest regression algorithm as an example, the inversion process is divided into two stages: model training and spatial prediction.

[0051] During the model training phase, spectral features extracted from meteorological station locations are used as input variables, i.e., feature vectors. The vector contains Features, such as Use the measured temperature values ​​from the corresponding stations. Or measure the relative humidity of the air. As target label The training process aims to find a function. This ensures that the predicted values ​​are as close as possible to the actual values. The model accomplishes this by constructing multiple decision trees and performing ensemble learning. The training process can be represented conceptually by the following: in, This represents the predicted temperature or relative humidity value from the model. The dataset is divided into training and validation sets. The training set is used to fit the model parameters, and the validation set is used to evaluate the model performance. The model is optimized by adjusting hyperparameters such as the number of decision trees and the maximum depth, until the coefficient of determination of the predicted results is reached. Achieve acceptable level and root mean square error Small enough.

[0052] In the spatial prediction stage, the trained random forest model It applies to every cell across the entire study area. For any cell... Obtain its spectral feature vector And input it into the model: in, and These are the temperature and relative humidity values ​​retrieved from the pixel, respectively. By iterating through all pixels and applying the corresponding temperature and relative humidity models, spatially continuous temperature and relative humidity raster maps are generated.

[0053] S47. Soil texture inversion mainly relies on matching the spectral characteristics of exposed soil pixels with known soil spectral libraries. This will be illustrated using the spectral angle mapping method as an example.

[0054] 1) Establish a reference soil spectral library. Correlate the soil sample data analyzed in the laboratory with the multispectral surface reflectance data of the corresponding sampling points. Each soil sample constitutes a reference spectral curve: The sample is accompanied by soil texture data, such as clay content. Multiple such records constitute a localized soil spectral reference library.

[0055] 2) Identify bare soil pixels to be inverted. Using a vegetation index threshold (e.g., setting the normalized difference vegetation index < 0.2), pixels with low vegetation cover and dominant soil information are automatically identified from the image. These pixels will be used for soil texture inversion.

[0056] 3) Calculate and match the spectral angles. For each soil pixel to be inverted... Its spectral vector is denoted as Calculate the spectrum of this pixel and compare it with each reference spectrum in the spectral library. spectral angles between The formula for calculating the spectral angle is: in, This represents the dot product operation of vectors. The Euclidean norm (modulus) of a vector. The smaller the value, the more similar the shapes of the two spectral curves.

[0057] 4) Assign soil texture. Find the reference spectrum with the smallest spectral angle to the pixel to be inverted. ,Right now Then, the soil texture data (such as clay content) corresponding to this reference spectrum is... The soil texture (such as clay content) is directly assigned to the cell. After traversing all exposed soil pixels, a spatial distribution map of soil texture (such as clay content) can be obtained. For sand and silt content, the same process is used, matching and assigning values ​​using the corresponding reference attributes respectively.

[0058] S48. After completing the spatial inversion of each environmental factor, integrate all result layers. For temperature and relative humidity, the inversion results cover all pixels. For soil texture, the inversion results mainly cover bare soil areas; for vegetation-covered areas, they may be marked or appropriately filled using spatial interpolation methods. Perform basic plausibility checks on all inversion results, such as checking whether the temperature value is within the possible range of the regional climate, whether the relative humidity is between 0 and 100%, and whether the sum of soil texture percentages is close to 100%. The validated temperature, relative humidity, and soil texture raster data, along with the previously calculated vegetation indices, will serve as key input variables for subsequent theoretical and residual grass yield models.

[0059] S5. Input the vegetation index and environmental factors of each pixel into the trained machine learning model to obtain the residual grass yield of each pixel; input the environmental factors of each pixel and the preprocessed meteorological data into the light energy utilization model to obtain the theoretical grass yield of each pixel. The process of obtaining the residual grass yield for each pixel includes: S50. The machine learning model used in this invention is Random Forest Regression. Random Forest Regression is an ensemble learning algorithm that improves the accuracy and stability of predictions by constructing and combining a large number of independent decision trees. Its basic principle is to use bootstrap sampling to extract multiple subsets of samples from the original training dataset, and train a decision tree for each subset. When splitting nodes in the growth of each decision tree, the optimal splitting feature is not selected from all input features, but from a randomly selected subset of features, which further enhances the overall diversity and generalization ability of the model. For regression problems, the final prediction output of the random forest is the arithmetic mean of the predictions from all decision trees. This model can effectively handle high-dimensional features, automatically assess feature importance, and is relatively insensitive to overfitting, making it very suitable for complex regression problems such as retrieving residual grass yield from multispectral and multi-environmental factors, which may involve multicollinearity.

[0060] S51. Training machine learning models requires high-quality, well-matched feature-label datasets. Feature data comes from preprocessing and inversion results, while label data comes from field measurements.

[0061] For each training sample point located in the grazing quadrat area outside the fence, extract all preset features of the corresponding pixel and construct a feature vector. Feature vectors typically contain two main categories: one is environmental factors, such as the retrieved temperature. relative humidity of air Soil sand content Soil clay content and altitude obtained from digital elevation models and slope Secondly, vegetation indices, such as the normalized difference vegetation index (NDVI). and Normalized Difference Red Edge Index Therefore, the feature vector of a sample can be represented as: , where subscript Indicates the first Training samples. Label data. This refers to the residual forage yield measured directly using the manual standard quadrat harvesting method in the grazing quadrats outside the aforementioned fenced areas. The unit is typically grams of dry matter per square meter. Ensure that each feature vector... All are associated with a precise, concurrently measured forage yield label. Strictly correspondence.

[0062] Before inputting the data into the model, the feature data is standardized or normalized to eliminate differences in the units and ranges of different features. For example, the Z-score standardization method is used to make the mean of each feature 0 and the standard deviation 1. For label data, logarithmic transformation is sometimes performed to improve its normal distribution characteristics.

[0063] S52. The goal of model training is to find a set of optimal model parameters so that the model can adapt to the input features. Predict grass yield as accurately as possible The training process follows these steps: All prepared sample data The dataset is randomly divided into training and validation sets. A common ratio is 70% for training and 30% for validation. The training set is used to directly fit the model, while the validation set is used to evaluate model performance and tune hyperparameters. A random forest regression model is initialized by setting initial hyperparameters, such as the number of decision trees. The maximum number of features considered when splitting each tree. Maximum depth of decision tree Then, the model is fitted using the training set data. During the fitting process, the algorithm constructs... Each decision tree is a decision tree. The growth of each tree minimizes the mean squared error within each node by recursively selecting features and split points. For nodes The dataset it contains is Finding the optimal splitting feature and dividing point The goal is to minimize the sum of the errors of the left and right child nodes after the split: ,in, and It is the dataset of the left and right child nodes generated after the split. Indicates the number of samples in the dataset. The calculation formula is: , It is a dataset The mean of the sample labels. For a new input feature vector. Each decision tree A predicted value will be given. The final prediction of a random forest is the average of the predictions from all the trees. .

[0064] S53. Evaluate the trained initial model using the reserved validation set. The main evaluation metrics include the coefficient of determination. and root mean square error Coefficient of determination The calculation formula is: ,in, It is the number of samples in the validation set. It is the first The actual residual yield of each sample It is the residual grass yield predicted by the model. It is the average value of the actual residual grass yield in the validation set. The closer the value is to 1, the stronger the model's ability to explain variation. Root mean square error (RMSE) The calculation formula is: , It reflects the level of absolute error in the model's predictions; the smaller the value, the better.

[0065] If the verification results are unsatisfactory, the hyperparameter optimization process will be initiated. This will be achieved through methods such as grid search or random search within a predefined hyperparameter space (e.g., adjusting...). , , The range of parameters was explored, and cross-validation was used to evaluate the performance of each parameter group. Finally, the parameter set was selected based on the validation set. highest and The smallest set of hyperparameters is used to retrain the final residual grass yield machine learning model.

[0066] S54. Obtain a validated random forest regression model that meets performance requirements. Then, it can be applied to the entire study area to calculate the residual grass yield of each pixel. For each pixel within the study area... Following the exact same feature extraction and preprocessing procedure as the training data, the feature vector of that pixel is generated. This vector contains the same feature types and order as during the training phase. It represents the feature vector of a pixel. Input into the trained random forest regression model In this model, all decision trees are computed in parallel, each providing a prediction. These predictions are then averaged to obtain the final predicted residual grass yield for that pixel. , Repeat the above input construction and model inference process for each pixel in the image. Then, calculate the prediction results for all pixels. Arranged according to their original spatial location, a two-dimensional matrix with the same number of rows and columns as the input image is generated, which is the spatial distribution map of residual grass yield. This raster layer, with grams of dry matter per square meter as the unit, intuitively shows the true spatial distribution pattern of grassland biomass in the entire study area after grazing, providing a direct basis for subsequent calculation of grass yield consumption and assessment of carrying capacity pressure.

[0067] The process of obtaining the theoretical grass yield for each pixel includes: S55. The light energy utilization efficiency model is a remote sensing estimation model based on the principles of plant physiological ecology. It is used to simulate the efficiency with which vegetation converts absorbed photosynthetically active radiation into organic dry matter, thereby estimating the net primary productivity or biomass of vegetation. In this invention, this model is used to calculate the potential maximum biomass of grassland under conditions of no grazing pressure, i.e., the theoretical grass yield. The construction and application of the model is a parameterization and spatialization process, the core of which lies in accurately quantifying the various key limiting factors in the light energy conversion process. The input to the model is the environmental factors for each pixel and preprocessed meteorological data. The light energy utilization efficiency model is built on the Monteith model framework, and its basic assumption is that vegetation dry matter yield is proportional to the photosynthetically active radiation absorbed by the vegetation canopy and the light energy conversion efficiency. The model formula used to calculate the theoretical grass yield is expressed as follows: in, The theoretical grass yield of a pixel is the potential dry matter mass of the aboveground portion of a unit area of ​​grassland under no grazing pressure, usually expressed in grams per square meter. Photosynthetically active radiation refers to the radiant energy in the solar radiation that can be utilized by plant photosynthesis within a wavelength range (usually 400-700nm), and the unit is megajoules per square meter. The proportion of photosynthetically active radiation absorbed refers to the ratio of photosynthetically active radiation absorbed by the vegetation canopy to the incident photosynthetically active radiation. It is a dimensionless quantity between 0 and 1. It represents the maximum light energy utilization rate under ideal conditions, that is, the theoretical maximum efficiency by which vegetation converts absorbed photosynthetically active radiation into dry matter when there are no environmental stresses (such as water or temperature stresses). The unit is grams of dry matter per megajoule. The optimal environmental coefficient is a dimensionless quantity between 0 and 1, used to quantify the promoting effect of ideal or optimal environmental conditions (such as optimal temperature and no moisture stress) on light energy utilization efficiency. It is usually set to 1 or close to 1 in the simulation of ideal conditions. The optimal climate yield coefficient is an empirical correction coefficient between 0 and 1, used to comprehensively adjust model systematic errors to better reflect actual observations under ideal local growth conditions. The physical meaning of this formula is: of the photosynthetically active radiation reaching the ground, a portion is absorbed by the vegetation canopy. This absorbed radiation energy has a theoretical maximum conversion efficiency under ideal environmental conditions, thus yielding an estimated potential maximum biomass.

[0068] S56. The application of the solar energy utilization rate model relies on the accurate acquisition or estimation of the five key parameters in the formula. The construction process involves the fusion of remote sensing data, preprocessed meteorological data, inverted environmental factors, and ground-based measured data. Specifically: 1) Preprocessed meteorological data refers to raw meteorological observation data obtained from meteorological stations in and around the study area, synchronous or quasi-synchronous with the UAV flight time. This data, after preprocessing steps such as temporal consistency checks, missing value imputation, and spatial interpolation, generates meteorological element raster data with spatial resolution and extent consistent with the multispectral imagery. This data must include at least total solar radiation. Temperature and relative humidity of air .

[0069] 2) Photosynthetically active radiation The total solar radiation data was estimated from preprocessed meteorological data. The unit is megajoules per square meter. Photosynthetically active radiation accounts for approximately 45% of total radiation, therefore it can be estimated using an empirical conversion formula: Using the spatially interpolated total solar radiation raster layer, a photosynthetically active radiation raster layer matching the spatial resolution of the remote sensing image is calculated.

[0070] 3) Photosynthetically active radiation absorption ratio It is highly correlated with vegetation cover and leaf area index. Using UAV remote sensing data, the vegetation index is estimated through normalized difference indices. A linear empirical relationship is established: .in, and These are empirical coefficients. These coefficients can be obtained through synchronous ground measurements: simultaneously measuring the normalized difference vegetation index and the actual photosynthetically active radiation absorption ratio (RAVA) in ungrazing quadrats within fenced areas (using equipment such as canopy analyzers), and then determining the coefficients through linear regression analysis. and The value was used to simulate the light absorption capacity of the canopy under undisturbed grazing conditions.

[0071] 4) Maximum light energy utilization rate This is a vegetation type-dependent parameter. For specific grassland types such as temperate grasslands or alpine meadows, it needs to be determined through localization experiments. Typically, during the growing season, completely ungrazed, optimally grown, enclosed plots are selected, and their biomass increase and absorbed photosynthetically active radiation are measured simultaneously to calculate the localization parameter. value.

[0072] 5) Optimal environmental coefficient It is a parameter used to quantify the promoting effect of environmental conditions on light energy utilization. In simulating ideal growth conditions without grazing pressure, The value is determined by the environmental factors of each pixel. Specifically, the temperature of each pixel obtained through inversion is... and relative humidity of air Compare with the preset optimal range for grassland growth. If the environmental factor value of a pixel is within the optimal range, then assign... A value of 1 indicates no environmental stress; if the value deviates from the optimal range, a coefficient less than 1 is calculated based on a preset attenuation function to reflect the slight limitation of the suboptimal environment on potential light energy utilization. Thus, It then established a connection with environmental factors.

[0073] 6) Optimal climate harvest coefficient It is a comprehensive correction parameter. Its determination depends on measured data from ungrazing quadrats within the fence. (Using unprocessed...) The corrected model (i.e., the model is set up) Calculate the theoretical forage yield of these ungrazing plots and compare the calculated results with the measured theoretical forage yield. Adjust the yield using optimization algorithms such as least squares. The ideal value is determined so that the coefficient of determination between the calculated and measured values ​​is maximized, and the root mean square error is minimized. Finally, a value suitable for the current region and grassland type under ideal conditions is determined. value.

[0074] S57. After completing the construction, inversion, and calibration of all parameters, the model is applied to every pixel in the entire study area to perform spatial calculations of the theoretical grass yield. Specifically: 1) It is necessary to prepare input data and parameter spatial layers corresponding to each pixel. This includes: a photosynthetically active radiation layer calculated from preprocessed meteorological data of total solar radiation; and a layer of normalized difference vegetation index and calibration coefficients. , The calculated photosynthetically active radiation absorption ratio layer; the maximum light energy utilization constant assigned to each pixel based on the grassland type. Layers; environmental factors (temperature) retrieved based on each pixel. and relative humidity of air The optimal environmental coefficient obtained by calculation Layers; and the calibrated optimal climate harvest coefficient constant. .

[0075] The photosynthetically active radiation (PADR) layer refers to a raster image composed of the PADR values ​​for each pixel in space. The generation of this layer relies on pre-processed total solar radiation data from meteorological data. This total solar radiation data originates from observations at regional meteorological stations and undergoes preprocessing steps such as time synchronization and spatial interpolation to form a total solar radiation raster with spatial resolution consistent with the UAV imagery and covering the entire study area. Based on empirical proportions in plant physiology, PADR accounts for approximately 45% of total solar radiation. Therefore, the PADR value for each pixel can be calculated by multiplying the total solar radiation value of each pixel by a coefficient of 0.45. Performing this calculation pixel-by-pixel generates the PADR layer, reflecting the spatial distribution of PADR within the study area. This layer serves as the energy input basis for the light energy utilization efficiency model to estimate theoretical grass yield.

[0076] The photosynthetically active radiation (PADR) absorption ratio layer refers to a raster image composed of the PDR absorption ratio of the vegetation canopy at each pixel location in space. The calculation of this layer requires two inputs: a Normalized Difference Vegetation Index (NDVI) layer calculated from multispectral imagery, and a set of empirical coefficients calibrated using ground-based measurement data. These empirical coefficients are obtained through simultaneous measurements in ungrazing quadrats within fenced areas; that is, by simultaneously measuring the NDVI and the actual PDR within the quadrats, and then determining the slope and intercept of the linear relationship through linear regression analysis. Substituting the NDVI value of each pixel into this linear relationship formula allows the calculation of the PDR for that pixel. This layer quantifies the potential capacity of the vegetation canopy to capture solar energy for photosynthesis under ideal conditions without grazing disturbance, with a value ranging from 0 to 1.

[0077] The maximum light energy utilization constant layer refers to a raster image composed of the maximum light energy conversion efficiency values ​​of the grassland vegetation type represented by each pixel location under ideal conditions. Maximum light energy utilization is a physiological parameter specific to a vegetation type, representing the theoretically highest efficiency with which vegetation converts absorbed photosynthetically active radiation into dry matter when there is no environmental stress. Different grassland types, such as temperate grasslands and alpine meadows, have different maximum light energy utilization values. This layer is constructed based on the grassland type distribution map of the study area. Each grassland type is assigned a specific maximum light energy utilization constant value obtained through inversion calculations using localized control experiments, and this constant value is assigned to all pixels within the corresponding grassland type area. This layer provides a theoretical upper limit parameter for calculating theoretical grass yield.

[0078] The optimal environmental coefficient layer is a raster image composed of the numerical values ​​of the degree to which environmental conditions at each pixel location promote or maintain light energy utilization. This coefficient is used to fine-tune light energy utilization to reflect the quality of background environmental conditions under ideal conditions without grazing pressure. Its calculation relies on environmental factors retrieved from each pixel, mainly air temperature and relative humidity. The air temperature and relative humidity values ​​of each pixel are compared with the predefined optimal temperature and humidity ranges for the growth of that grassland type. If the environmental factor values ​​of a pixel are within the optimal range, the environmental conditions are considered optimal, and the pixel is assigned an optimal environmental coefficient value of 1; if they deviate from the optimal range, a coefficient value between 0 and 1 is calculated according to a preset attenuation function. This layer transforms the spatial distribution information of environmental factors into a fine-tuning coefficient for theoretical light energy utilization efficiency, thereby making the estimation of theoretical grass yield more consistent with the actual environmental background of the region.

[0079] 2) Traverse every cell of the entire study area. For the current cell, extract the value of that cell's position from the data and parameter layers mentioned above: , , , , Then, these values ​​are substituted into the light energy utilization model formula for calculation: This calculation process is completed in parallel through array operations, efficiently generating the theoretical grass yield for each pixel. .

[0080] 3) Calculation results for all pixels Arranged according to the original spatial order, a complete spatial distribution map of theoretical grass yield is generated. To verify the model's accuracy, measured data from reserved, ungrazing independent quadrats that were not involved in parameter calibration are used. The model estimates for each quadrat location are compared with the measured values, and validation indices such as the coefficient of determination and root mean square error are calculated. If the accuracy meets the requirements, this theoretical grass yield distribution map is the final result; if not, the parameter inversion and calibration process needs to be reviewed and optimized, especially the methods for determining the photosynthetically active radiation absorption ratio and the optimal climate harvest coefficient. This final result visually reflects the spatial distribution of grassland potential productivity under no-grazing pressure and serves as the foundational data layer for calculating animal consumption and assessing carrying capacity.

[0081] S6. The difference between the theoretical grass production and the residual grass production of each pixel is taken as the grass production consumed by the corresponding pixel.

[0082] S7. The sum of the grass production consumed by each pixel is used as the grass production consumed in the preset area.

[0083] The grass yield consumed in the pre-defined area refers to the total dry matter mass of forage consumed by herbivores (including grazing livestock, wild herbivores, and herbivorous insects) across all pixels within a specific time frame and a clearly defined grassland space (i.e., the pre-defined area). This indicator is obtained by subtracting the residual grass yield from the theoretical grass yield of each pixel within the area, and then summing the differences across all pixels. Its application is significant, providing a quantitative and spatial core decision-making basis for the scientific management of grassland ecosystems, the sustainable development of animal husbandry, and biodiversity conservation. Specifically: 1) Achieving precise diagnosis and zoned management of grass-livestock balance. Traditional grass-livestock balance assessments rely heavily on livestock inventory data from county statistical yearbooks, failing to reflect the actual spatial distribution and grazing intensity of livestock within pastures. Pre-defined zones provide a clear heat map of total forage consumption and distribution within the pasture. Managers can compare this total consumption with the theoretical demand per standard sheep unit calculated based on the actual number and type of livestock, accurately determining whether the entire area is overgrazing. More importantly, by analyzing the spatial distribution map of forage consumption, overgrazing areas (consumption hotspots) and underutilized areas can be visually identified. This provides a direct and scientific blueprint for implementing zoned rotational grazing, setting differentiated grazing intensities, and planning the layout of fencing and watering points, achieving a leap from "average pasture management" to "precise plot management."

[0084] 2) Quantifying the ecological carrying capacity of wild herbivores to support human-wildlife harmony and ecological protection. Traditional carrying capacity models often overlook the consumption of grasslands by wild herbivores (such as rabbits, rodents, and ungulates) and insects, leading to an underestimation of the total carrying capacity pressure on grasslands. The grass production from the consumption in a predefined area can be further separated from the total consumption by methods such as diet analysis and population surveys to determine the contribution ratio of livestock consumption to wildlife consumption. This provides crucial data for assessing the actual carrying capacity of grassland ecosystems for wild populations. In the field of ecological protection, this data helps to scientifically formulate wildlife population management plans, find a balance between livestock development and wildlife protection, and provide a quantitative basis for establishing ecological compensation mechanisms and mitigating human-wildlife conflict.

[0085] 3) Serving grassland resource asset accounting and ecological product value realization. The amount of forage consumed is a direct quantitative reflection of the supply services (forage production) provided by grasslands. Accurately measuring this amount is the foundation for grassland resource asset accounting and assessing the annual physical quantity and value of forage products. In the context of promoting the realization of ecological product value, this data can serve as one of the assessment indicators for the effectiveness of grassland ecological protection, or provide important baseline data and monitoring basis for grassland carbon sequestration projects (increasing soil carbon fixation through rational grazing), transforming the ecological value of grasslands into economic value.

[0086] 4) As a dynamic monitoring indicator for evaluating grassland health and grazing systems. The amount of forage consumed in a pre-defined area is a dynamic variable. By monitoring at different time points (such as different seasons and different years), a spatiotemporal sequence of forage consumption can be obtained. Analyzing this change can assess the effect of grazing system adjustments (such as changes in rotational grazing cycles) and determine the trend of grassland recovery or degradation. If the amount of forage consumed in a certain area is consistently close to or even exceeds the theoretical forage yield, it indicates continuous over-exploitation and warns of grassland degradation risks. Conversely, it may indicate under-exploitation, with problems such as forage waste or vegetation aging. Therefore, it is a core early warning indicator for evaluating whether grazing activities are within the ecological security threshold.

[0087] 5) Providing crucial validation data for model optimization and scientific research. Methodologically, the estimated grass production from pre-defined areas can be cross-validated with consumption indirectly extrapolated from traditional livestock census and wildlife survey data, thereby assessing and improving the accuracy of the UAV remote sensing inversion model. Simultaneously, this indicator provides unprecedented, large-scale, and high-precision empirical data support for fundamental ecological questions such as food competition among multiple consumers (livestock and wild animals) and the impact of grazing activities on the material cycle and energy flow of grassland ecosystems.

[0088] The technical solution of the present invention will be further described through another embodiment. In this embodiment, an integrated multispectral sensor is used on a UAV. This sensor includes blue, green, red, red-edge, and near-infrared bands, and a global shutter is used to achieve strict synchronous acquisition of images in all bands, obtaining high-resolution, multi-band ground image data of a preset area. During the acquisition process, the exposure parameters of the multispectral sensor are dynamically adjusted through an intelligent real-time dimming algorithm to cope with complex lighting conditions. Through an edge-cloud collaborative processing architecture, high-quality automated preprocessing is performed on the acquired multi-band image data in the cloud. The preprocessing process specifically includes spatiotemporal alignment, radiometric calibration, and atmospheric correction of the multi-band images to obtain the surface reflectance of each pixel in the multispectral band. Subsequently, from the preprocessed data, at least one vegetation index for each pixel is calculated based on its surface reflectance, such as the Normalized Difference Vegetation Index (NDVI), the Normalized Difference Red Edge Index (NDRI), and the Soil-Regulated Vegetation Index (SRI). Based on the surface reflectance of each pixel, at least one environmental factor for each pixel is obtained through machine learning inversion or spectral matching methods, such as air temperature, relative humidity, and soil texture. Then, the vegetation index and environmental factor of each pixel are used as input features and fed into a pre-trained machine learning model, such as a random forest regression model. This model calculates the residual grass yield for each pixel by fusing these multispectral derived vegetation indices and environmental factors, denoted as _____. Residual grass yield represents the actual remaining grass yield of grassland after grazing. Simultaneously, a light energy utilization efficiency model is used to calculate the theoretical grass yield based on UAV data and preprocessed meteorological data; specifically, environmental factors for each pixel and preprocessed meteorological data are input into the light energy utilization efficiency model, which uses a formula... Perform the calculations. Among them, This represents the theoretical grass yield per pixel. Indicates photosynthetically active radiation; It represents the proportion of photosynthetically active radiation absorbed, estimated by the normalized difference vegetation index, etc. This represents the maximum light energy utilization rate under ideal conditions; Indicates the optimal environmental coefficient; This represents the optimal climate-based harvest coefficient calibrated using measured data. The model calculates the theoretical grass yield per pixel, denoted as . Theoretical grass yield characterizes the maximum potential biomass of grassland under no grazing pressure. Finally, based on the validated model results, the theoretical grass yield for each pixel is determined. and residual grass yield The difference is calculated to determine the amount of grass produced by various herbivores in each pixel, denoted as . ,Right now This grass production consumption includes the combined consumption by livestock, wild herbivores, and herbivorous insects. It calculates the grass production consumption of all pixels within the preset area. The total grass production consumed in the area is obtained by summing the results. Then, using the standard sheep unit annual forage intake (e.g., 1.548 kg hay / day, calculated over 365 days) as a benchmark, the equivalent number of sheep units corresponding to this total grass production consumption is calculated. This yields spatially clear assessments of carrying capacity, grassland utilization, and carrying capacity pressure, enabling rapid, accurate, and non-destructive monitoring of grassland productivity and livestock stress, and providing a more scientific and effective methodology for grassland management and ecological protection.

[0089] The data acquisition layer completes flight planning based on the preset area, then controls the UAV to perform automatic cruise missions. During flight, it utilizes the integrated multispectral sensor onboard the UAV to acquire high-resolution ground image data in the blue, green, red, red-edge, and near-infrared bands through multispectral synchronous acquisition technology. During acquisition, intelligent real-time dimming technology dynamically adjusts sensor exposure parameters to ensure data quality. The acquired raw data is transmitted to the cloud via an edge-cloud collaborative data upload link. After entering the cloud processing layer, an automated preprocessing process is initiated based on the edge-cloud collaborative architecture. This process sequentially performs radiometric correction on the uploaded multi-band image data. After geometric correction and stitching, a high-precision digital surface model and orthophotos of reflectance in each band are generated. Based on these orthophotos, feature and factor extraction is performed, specifically including calculating a series of vegetation indices such as the Normalized Difference Vegetation Index (NDVI) and the Normalized Difference Red Edge Index (NDRI), and obtaining environmental factors such as air temperature, relative humidity, and soil texture through remote sensing inversion methods. Subsequently, the processing flow is divided into two parallel model computation paths. One path inputs the extracted vegetation indices and environmental factors into a trained machine learning model, such as a random forest or XGBoost algorithm model. This model calculates the residual grass yield for each pixel by fusing these features, denoted as . The residual grass yield model requires parameter optimization based on validation data from outside the fenced quadrats during the training and validation phases to ensure its accuracy. Another approach involves inputting environmental factors and pre-processed meteorological data into the light energy utilization model, which is based on the formula... Calculate the theoretical grass yield per pixel, denoted as . ,in Photosynthetically active radiation, The proportion of photosynthetically active radiation absorbed. To maximize light energy utilization, For the optimal environmental coefficient, To determine the optimal climate harvest coefficient, the theoretical forage yield model also needs to be calibrated and validated using quadrat data within the enclosure. After both forage yield models have been validated, animal consumption is calculated, with the core formula being: That is, the theoretical grass yield per pixel. Subtract residual grass yield The amount of grass produced by the animal that obtains this pixel Finally, the process enters the application layer, where the above calculation results are visualized and applied to generate spatial distribution maps of theoretical and residual grass yield, heat maps of animal consumption, and comprehensive grass-livestock balance decision reports, thereby providing clear spatial decision-making basis for grassland management.

[0090] The drone platform used is either the DJI Phantom 4 Multispectral Edition, the DJI Mavic 3M, or an equivalent integrated multispectral drone platform. These drone platforms have built-in light intensity sensors that automatically record solar irradiance in various bands during flight, providing crucial data for subsequent radiometric correction. The multispectral sensor includes at least five core bands: blue (center wavelength approximately 450nm), green (center wavelength approximately 560nm), red (center wavelength approximately 650nm), red-edge (center wavelength approximately 730nm), and near-infrared (center wavelength approximately 840nm). The positioning system employs a real-time dynamic differential positioning system to ensure that each acquired image has high-precision geographic location information with centimeter-level accuracy. This is fundamental for achieving high-precision geometric correction and spatial matching with ground survey sample plot data. Ground calibration boards, using grayscale boards and pure white boards with known reflectivity, are laid out in the survey area before aerial photography to assist in absolute radiometric calibration.

[0091] The mission planning requires the use of professional flight mission planning software to set flight parameters for the preset area. A flight altitude of 80-120 meters is recommended to obtain a ground resolution of 3-5 cm. The forward overlap should be no less than 80%, and the lateral overlap no less than 75% to ensure the accuracy of subsequent photogrammetric 3D modeling. The operation should be conducted during local times with clear weather, low wind speed, and a high solar altitude angle to minimize shadows and ensure stable lighting conditions. During data acquisition, the UAV automatically flies along a preset route, simultaneously acquiring image data and solar irradiance data for all bands, and capturing images of the ground calibration board. This process employs multispectral synchronous acquisition technology. Through an integrated sensor and global shutter, strict spatiotemporal alignment of all band images is achieved, fundamentally eliminating the registration errors present in traditional multi-camera systems and laying a data foundation for subsequent accurate quantitative inversion. Simultaneously, an intelligent real-time dimming algorithm is integrated during the acquisition process. This algorithm dynamically adjusts the exposure parameters of the multispectral sensor to cope with changes in lighting, ensuring the quality of the raw data.

[0092] Data processing is automated in the cloud within an edge-cloud collaborative processing architecture. The first step is radiometric calibration, which converts the raw digital quantization values ​​recorded by the camera into physically meaningful apparent reflectance. The formula is: ,in, It is the apparent reflectance of the λ band, which is a dimensionless quantity; It is the original digital quantization value of the pixel in the λ-band image, which is the integer value recorded by the sensor; These are the camera calibration coefficients for band λ, used to convert digital quantization values ​​into radiance; This represents the solar irradiance in band λ received by the sensor during flight, measured in watts per square meter per micrometer. Next, atmospheric correction is performed to eliminate the effects of atmospheric scattering and absorption, obtaining the true surface reflectance of ground features. A simplified method based on a radiative transfer model is used, with the following formula: ,in, It is the surface reflectance of the λ band, which is the final physical quantity after atmospheric correction; It is the apparent reflectance of band λ; It is the path radiation value of band λ, representing the additional radiation caused by atmospheric scattering; λ is the atmospheric transmittance of band λ, representing the attenuation ratio of light propagating through the atmosphere. Finally, geometric correction and stitching are performed. Using the positioning and attitude determination system data onboard the UAV and the information from the real-time dynamic differential positioning system, a high-precision digital surface model and orthophoto images of the surface reflectance for each band are generated through motion-reconstruction photogrammetry algorithms. The resulting images for all bands must be strictly aligned at the pixel level.

[0093] Features for subsequent model calculations are extracted from the processed surface reflectance orthophoto image. Vegetation index calculations include: Normalized Difference Vegetation Index (NDV), Normalized Difference Red Edge Index (NDRI), and Soil-Regulated Vegetation Index (SRI). The formula for calculating the NPV is as follows: Where NIR represents the surface reflectance in the near-infrared band, and Red represents the surface reflectance in the red band; the formula for calculating the normalized difference red-edge index is: Where RedEdge represents the surface reflectance in the red-edge band; the formula for calculating the soil-modified vegetation index is: Where L is the soil conditioning coefficient, an empirical constant. Environmental factor inversion includes: air temperature and relative humidity can be inverted using machine learning methods, i.e., by establishing a correlation model between measured data from regional meteorological stations and the multispectral characteristics of UAVs for prediction; soil texture can be inverted from a pre-established soil spectral library using spectral matching methods such as spectral angle mapping, utilizing the spectral information of exposed soil areas; photosynthetically active radiation can be estimated by multiplying the total solar radiation data obtained from meteorological stations by an empirical coefficient. These extracted vegetation indices and inverted environmental factors will serve as common inputs for calculating residual and theoretical grass yields.

[0094] Feature extraction and factor inversion are performed based on preprocessed multispectral orthophotos. This step specifically includes calculating a series of vegetation indices, such as the Normalized Difference Vegetation Index (NDVI), the Normalized Difference Red Edge Index (NDRI), and the Soil-Regulated Vegetation Index (SRI). The formula for calculating the NDVI is as follows: Here, NIR represents the surface reflectance in the near-infrared band, Red represents the surface reflectance in the red band, and the formula for calculating the normalized differential red-edge index is: Here, RedEdge represents the surface reflectance in the red-edge band. Simultaneously, environmental factors are inverted from the imagery, including soil texture inversion based on spectral analysis of exposed soil areas, and temperature and relative humidity inversion using machine learning methods. These extracted and inverted data are then split to support two parallel computational models. One is a residual grass yield model, which uses machine learning algorithms such as random forest or XGBoost regression. Its training sample input comes from grazing quadrat data outside fences, including a feature vector X composed of vegetation indices and environmental factors, such as normalized difference vegetation index, normalized difference red-edge index, temperature, relative humidity, and soil texture, as well as the measured grass yield of the quadrat outside the fence as a label Y. A predictive model is generated through training. During training, the model output is compared with the measured quantity of quadrat outside the fence for feedback optimization and error analysis to optimize model parameters, forming a model validation and optimization closed loop. The trained and optimized model is then applied to the entire study area to generate a residual grass yield distribution map. The other is a theoretical grass yield model, based on a light energy utilization rate model, whose core calculation formula is... , Theoretical grass yield is given by PAR, which represents photosynthetically active radiation, usually derived from weather station data or estimates. fPAR represents the percentage of photosynthetically active radiation absorbed, which can be calculated as follows: Where a and b are empirical coefficients, This represents the maximum light energy utilization constant after localization calibration. This represents the optimal environmental coefficient. This indicates the need to calibrate the optimal climate harvest coefficient using measured data from quadrats within the fenced area. After determining the model parameters, this coefficient is applied to the entire study area to generate a theoretical forage yield distribution map. Finally, the theoretical forage yield distribution map and the residual forage yield distribution map are spatially overlaid for calculation, based on the formula... Generate a spatial distribution map of animal consumption, thus completing the entire technical process from multispectral imagery to quantification results of grassland productivity consumption.

[0095] The goal of the residual forage yield model is to simulate the actual remaining forage yield of grassland after grazing. The model's input features are extracted from an out-of-fence grazing quadrat representing the actual grazing state. Core input features include air temperature, relative humidity, soil texture (such as sand and clay percentages) retrieved from UAV multispectral data, and environmental factors such as altitude and slope obtained from a digital elevation model, collectively forming the feature vector X. The model's training labels Y are derived from the measured forage yield of the corresponding out-of-fence grazing quadrat, measured manually using a quadrat harvesting method, typically in grams per square meter of dry weight. The algorithms used to construct the model are random forest regression or XGBoost regression, which effectively handle the complex nonlinear relationship between features and the target variable and are insensitive to overfitting. The conceptual formula of the model can be expressed as: ,in, Represents the residual grass yield, and RF represents the random forest regression function. Indicates temperature, Indicates relative humidity of the air. This indicates the percentage of sand content in the soil. This represents the percentage of soil clay content. The model validation process uses reserved, unused quadrat data outside the fence to evaluate the performance of the trained model, calculates validation metrics such as the coefficient of determination and root mean square error, and optimizes and adjusts the model parameters based on the evaluation results to improve prediction accuracy.

[0096] The goal of the theoretical forage yield model is to calculate the maximum potential biomass of grassland under no grazing pressure. The model is based on the principle of light energy use efficiency, and its calculation formula is as follows: In this formula, Indicates the theoretical grass yield; This represents photosynthetically active radiation, measured in megajoules per square meter, and is estimated from the total solar radiation in preprocessed meteorological data. The proportion of photosynthetically active radiation absorbed is highly correlated with vegetation cover and is linearly estimated using the normalized difference vegetation index (NDVI). The relationship is as follows: ,in and These are empirical coefficients obtained by fitting measured data; It represents the maximum light energy utilization rate under ideal conditions, and the unit is grams of dry matter per megajoule. It is a constant given according to the grassland type and needs to be determined through localized experiments. The optimal environment coefficient is a dimensionless quantity between 0 and 1, whose value is determined by whether the environmental factors (such as temperature and relative humidity) of each pixel are within the optimal range for grassland growth. The optimal climate yield coefficient, representing an empirical correction coefficient between 0 and 1, is calibrated using measured data from ungrazing quadrats within fenced areas. The model validation process uses measured data from ungrazing quadrats within fenced areas to validate key parameters in the light energy utilization model, particularly... and Estimating coefficients in a relation , Perform calibration and verification.

[0097] Animal-generated forage consumption is calculated as the difference between theoretical forage production and residual forage production, using the following formula: ,in, Indicates the amount of grass consumed by animals. Indicates the theoretical grass yield. This represents the remaining grass yield. The calculation process is performed independently at each pixel level, thus generating a spatial distribution map of animal consumption covering the entire pasture area.

[0098] Multi-consumer carrying capacity assessments can be conducted based on forage production generated by animal consumption. The calculation formula is: ,in, This represents the annual hay intake per consumer, measured in sheep units. A standard sheep unit consumes 1.548 kg of hay per day, calculated annually over 365 days. It can also be used to calculate grazing rates. The formula reflects the proportion of grassland being utilized. .

[0099] In scenarios with drastically changing lighting conditions, traditional multispectral cameras, due to their fixed exposure times, are prone to overexposure or underexposure of data. This invention integrates an intelligent real-time dimming algorithm on the UAV. The algorithm flow is as follows: First, the multispectral sensor analyzes the brightness histogram distribution of each band in the current acquisition frame in real time. Second, if the system detects that the maximum digital quantization value of a pixel in any band of the image exceeds 85% of a preset saturation threshold, it automatically fine-tunes the sensor's exposure time or sensitivity parameters, prioritizing the avoidance of overexposure in the red-edge band and near-infrared band, which are crucial for vegetation analysis. Third, the algorithm synchronously records all exposure parameter adjustment information and performs precise compensation in the subsequent radiometric calibration calculation formula, thereby ensuring the physical accuracy of the final calculated apparent reflectance. This technology effectively improves the availability and quality of data collected by UAVs under uneven lighting or cloudy weather conditions.

[0100] Model validation and evaluation employed hold-out validation or k-fold cross-validation. For the residual forage yield model validation, data from outside fenced quadrats were divided into training and validation sets in approximately a 7:3 ratio. A random forest regression model was trained using the training set data, and validation metrics, including the coefficient of determination and root mean square error, were calculated using the validation set data. If the root mean square error was too high, the model parameters were adjusted, or additional input features were considered. For the theoretical forage yield model validation, measured data from ungrazing quadrats within fenced areas were used. Key parameters in the light energy utilization model were optimized using fitting methods such as least squares, with the goal of maximizing the calculated values. The model exhibits the largest coefficient of determination and the smallest root mean square error compared to the measured values. Furthermore, spatial visualization validation is performed by applying the validated model to the entire study area to generate a spatial distribution map of grass yield. The rationality of the spatial distribution pattern is then examined through visual interpretation and other methods.

[0101] This invention proposes a multispectral synchronous acquisition technology, employing an integrated sensor where all spectral bands share a single lens and exposure is controlled by a global shutter. This fundamentally solves the spatiotemporal inconsistency problem caused by differences in acquisition time and viewing angle among multi-band image data, laying a solid data foundation for subsequent high-precision quantitative inversion of biophysical parameters. Furthermore, this invention proposes an intelligent real-time dimming algorithm. This algorithm dynamically optimizes the exposure parameters of the multispectral sensor in real time at the UAV data acquisition end, effectively coping with the complex and variable lighting environment in the field. It overcomes the technical bottleneck of poor data quality in traditional UAV remote sensing systems under unstable weather conditions such as cloudy skies, significantly expanding the system's effective operating time window. This invention also proposes an edge-cloud collaborative processing architecture. In this architecture, the edge device is responsible for data acquisition, primary compression and encryption, and intelligent dimming control; after receiving the data, the cloud service platform automatically triggers a complete pipeline operation including data preprocessing, feature extraction, model calculation, result visualization, and distribution. This innovative architecture enables cloud-based, real-time online calculation of the entire grassland carrying capacity assessment process. After users upload data, the system can automatically complete the analysis and quickly generate reports and maps without complicated manual operations, greatly improving work efficiency and lowering the barrier to entry for using professional technical software.

[0102] The drone is equipped with a multispectral sensor to perform flight missions, collect raw spectral data and record flight logs. At the same time, the ground calibration board and the real-time dynamic differential positioning system base station provide the reference data required for radiometric calibration and geometric correction. Smart mobile terminals, such as remote controllers or tablets, are used to send flight control commands and upload encrypted raw data and processing task requests via wireless network. These data and requests are transmitted via 4G, 5G or WiFi wireless links in the network transmission layer, and after identity verification and data encryption are completed by a secure access gateway, they arrive at the cloud platform. In the cloud platform, the uploaded raw data is first stored in the object storage service. Upon receiving the data, the task queues and triggers in the access and triggering layers automatically initiate subsequent processing flows. Subsequently, the parallel processing engine coordinates a series of automated processing services, including radiometric and geometric correction services, feature extraction and inversion services, residual forage yield model calculation services, and theoretical forage yield model calculation services. These services rely on a distributed computing cluster to achieve efficient parallel computation. The optimized algorithm models used in the processing are stored in a model library and continuously updated, while the generated output data, such as surface reflectance orthophotos, residual forage yield distribution maps, and theoretical forage yield distribution maps, are stored in the output database. The API and application service layers provide functions such as data visualization and analysis report downloads based on these outputs. On the user side, managers log in to the system through a web browser or mobile application, send processing requests, and interact with the platform. The cloud platform returns various results, including grassland productivity reports and animal consumption heatmaps, through application programming interfaces (APIs), for managers to view and make grazing management decisions.

[0103] In the above embodiments, although the steps are numbered S1, S2, etc., they are only specific embodiments given by the present invention. Those skilled in the art can adjust the execution order of S1, S2, etc. according to the actual situation. The scheme after adjusting the order is also within the protection scope of the present invention. It can be understood that in some embodiments, some or all of the above embodiments may be included.

[0104] like Figure 2 As shown, an embodiment of the present invention provides a system for obtaining the amount of grass consumed, including a data acquisition module, an image data preprocessing module, a surface reflectance acquisition module, an exponential factor calculation module, a grass production calculation module, a consumed grass production calculation module, and a total consumed grass production calculation module. The data acquisition module is used to: acquire multi-band image data of a preset area using a multispectral sensor mounted on a drone; the image data preprocessing module is used to: preprocess the multi-band image data; the surface reflectance acquisition module is used to: acquire the surface reflectance of each pixel in the preprocessed multi-band image data; and the exponential factor calculation module is used to: calculate the surface reflectance of each pixel based on the surface reflectance of each pixel. The system calculates at least one vegetation index for each pixel and, based on the surface reflectance of each pixel, inverts at least one environmental factor for each pixel; the grass yield calculation module is used to: input the vegetation index and environmental factor of each pixel into the trained machine learning model to obtain the residual grass yield of each pixel; input the environmental factor of each pixel and the preprocessed meteorological data into the light energy utilization model to obtain the theoretical grass yield of each pixel; the grass consumption calculation module is used to: take the difference between the theoretical grass yield and the residual grass yield of each pixel as the grass consumption of the corresponding pixel; the total grass consumption calculation module is used to: take the sum of the grass consumption of each pixel as the grass consumption of the preset area.

[0105] Optionally, the above technical solution also includes an exposure parameter adjustment module, which is used to dynamically adjust the exposure parameters of the multispectral sensor through an intelligent real-time dimming algorithm when multi-band image data of a preset area is acquired using a multispectral sensor.

[0106] Optionally, in the above technical solution, the image data preprocessing module is specifically used for: performing spatiotemporal alignment, radiometric calibration, and atmospheric correction on multi-band image data.

[0107] Optionally, in the above technical solution, at least one vegetation index includes at least one of the Normalized Differential Vegetation Index (NDVI), the Normalized Differential Red Edge Index (NDRE), and the Soil-Adjusted Vegetation Index (SAVI), and at least one environmental factor includes at least one of air temperature, relative humidity, and soil texture.

[0108] It should be noted that the beneficial effects of the system for obtaining consumed forage yield provided in the above embodiments are the same as those of the method for obtaining consumed forage yield described above, and will not be repeated here. Furthermore, the system provided in the above embodiments is only illustrated by the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the system can be divided into different functional modules according to the actual situation to complete all or part of the functions described above. In addition, the system and method embodiments provided in the above embodiments belong to the same concept, and their specific implementation process is detailed in the method embodiments, and will not be repeated here.

[0109] An electronic device according to an embodiment of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements any of the above-mentioned methods for obtaining the amount of grass consumed.

[0110] An embodiment of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements any of the above-mentioned methods for obtaining the amount of grass consumed.

[0111] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for obtaining the amount of consumed grass, characterized in that, include: Multi-band image data of a preset area is collected using a multispectral sensor mounted on a drone; The multi-band image data is preprocessed; Obtain the surface reflectance of each pixel in the preprocessed multi-band image data; Based on the surface reflectance of each pixel, at least one vegetation index for each pixel is calculated, and based on the surface reflectance of each pixel, at least one environmental factor for each pixel is retrieved. The vegetation index and environmental factors of each pixel are input into the trained machine learning model to obtain the residual grass yield of each pixel; the environmental factors of each pixel and the preprocessed meteorological data are input into the light energy utilization model to obtain the theoretical grass yield of each pixel. The difference between the theoretical grass production and the residual grass production of each pixel is taken as the grass production consumed by the corresponding pixel. The sum of the grass production consumed by each pixel is used as the grass production consumed in the preset area.

2. The method for obtaining the amount of consumed forage according to claim 1, characterized in that, Also includes: When the multispectral sensor is used to acquire multi-band image data of a preset area, the exposure parameters of the multispectral sensor are dynamically adjusted through an intelligent real-time dimming algorithm.

3. The method for obtaining the amount of consumed forage according to claim 1 or 2, characterized in that, The preprocessing of the multi-band image data includes: spatiotemporal alignment, radiometric calibration, and atmospheric correction of the multi-band image data.

4. A method for obtaining the amount of consumed forage according to claim 1 or 2, characterized in that, The at least one vegetation index includes at least one of the Normalized Differential Vegetation Index (NDVI), the Normalized Differential Red Edge Index (NDRE), and the Soil-Adjusted Vegetation Index (SAVI), and the at least one environmental factor includes at least one of air temperature, relative humidity, and soil texture.

5. A system for obtaining the amount of grass produced, characterized in that, It includes a data acquisition module, an image data preprocessing module, a surface reflectance acquisition module, an index factor calculation module, a grass yield calculation module, a grass consumption calculation module, and a total grass consumption calculation module; The data acquisition module is used to: acquire multi-band image data of a preset area using a multispectral sensor mounted on a drone; The image data preprocessing module is used to: preprocess the multi-band image data; The surface reflectance acquisition module is used to: acquire the surface reflectance of each pixel in the preprocessed multi-band image data; The index factor calculation module is used to: calculate at least one vegetation index for each pixel based on the surface reflectance of each pixel, and invert at least one environmental factor for each pixel based on the surface reflectance of each pixel. The grass yield calculation module is used to: input the vegetation index and environmental factors of each pixel into the trained machine learning model to obtain the residual grass yield of each pixel; input the environmental factors of each pixel and the preprocessed meteorological data into the light energy utilization model to obtain the theoretical grass yield of each pixel. The grass production consumption calculation module is used to: take the difference between the theoretical grass production and the residual grass production of each pixel as the grass production consumption of the corresponding pixel. The total grass production calculation module is used to: use the sum of the grass production of each pixel as the grass production of the preset area.

6. The system for obtaining the amount of consumed forage according to claim 5, characterized in that, It also includes an exposure parameter adjustment module, which is used to dynamically adjust the exposure parameters of the multispectral sensor through an intelligent real-time dimming algorithm when the multispectral sensor is used to collect multi-band image data of a preset area.

7. The system for obtaining the amount of consumed forage according to claim 5 or 6, characterized in that, The image data preprocessing module is specifically used to perform spatiotemporal alignment, radiometric calibration, and atmospheric correction on the multi-band image data.

8. The system for obtaining the amount of consumed forage according to claim 5 or 6, characterized in that, The at least one vegetation index includes at least one of the Normalized Differential Vegetation Index (NDVI), the Normalized Differential Red Edge Index (NDRE), and the Soil-Adjusted Vegetation Index (SAVI), and the at least one environmental factor includes at least one of air temperature, relative humidity, and soil texture.

9. An electronic device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement a method for obtaining consumed grass production as described in any one of claims 1 to 4.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements a method for obtaining consumed grass production as described in any one of claims 1 to 4.

Images