High and cold meadow aboveground biomass monitoring method based on PROSAIL-BP

Abstract

The invention discloses an alpine meadow aboveground biomass monitoring method based on PROSAIL-BP, and belongs to the field of ecological remote sensing information processing. In order to overcome the defect that samples in the alpine region are insufficient and have high precision and high reliability, the method comprises the steps that an alpine meadow mask, remote sensing images and field biomass data in a target region are obtained, the remote sensing images are spliced, the wave band reflectivity is normalized, and the grassland region reflectivity is reserved in combination with mask cutting; predefining a PROSAIL model matched with the region features; performing Sobol global sensitivity analysis to obtain a first-order sensitivity index and a total-order sensitivity index of the parameter; screening a key wave band and four high-sensitivity parameters; uniformly sampling to generate a parameter group, simulating a hyperspectrum through PROSAIL, extracting the reflectivity of a key wave band, and constructing a data set by multiplying a leaf area index by a dry matter content as a target variable; and training a three-layer BP neural network, and processing the reflectivity output pixel biomass of the remote sensing key wave band. The method is applied to a remote sensing information processing system and has high precision and high reliability.

Description

TECHNICAL FIELD

[0001] The application relates to the field of ecological remote sensing information processing, and particularly relates to a high-cold meadow aboveground biomass monitoring method based on PROSAIL-BP. BACKGROUND

[0002] The high-cold meadow is a core component unit of the Qinghai-Tibet Plateau ecosystem, and the integrity of its ecological function has a profound impact on regional and even global ecological balance. From the perspective of regional climate regulation, the high-cold meadow affects local water cycle and energy exchange through processes such as vegetation transpiration and surface albedo regulation, which is crucial for maintaining the climate stability of the "Asian Water Tower" of the Qinghai-Tibet Plateau; in terms of water conservation, the complex structure formed by the vegetation layer and root system of the meadow can enhance the soil water holding capacity and reduce surface runoff, which is an important guarantee for the water source of the upper reaches of the Yangtze River and the Yellow River; and in terms of carbon sink function, the high-cold meadow, as an important carbon pool of the terrestrial ecosystem, fixes atmospheric carbon through photosynthesis, and the dynamic change of its carbon storage is directly related to the global carbon cycle balance.

[0003] In recent years, under the double disturbance of climate change (such as rising temperature and changing precipitation pattern) and human activities (such as overgrazing and infrastructure construction), the high-cold meadow has shown degradation phenomena such as decreased vegetation cover, soil desertification and reduced biodiversity, leading to significant weakening of its ecological function. Under this background, large-scale and dynamic ecological monitoring and evaluation become the premise of curbing degradation and achieving ecological restoration, and the precise monitoring of aboveground biomass (AGB) is particularly critical. AGB not only directly reflects grassland productivity (such as vegetation growth rate and community structure stability) and degradation degree (such as biomass reduction often accompanied by intensified degradation), but also is the basic data for estimating the carbon storage of grassland ecosystem - accurate AGB data can provide scientific basis for carbon source / sink evaluation under the "double carbon" target and support the formulation of regional carbon neutralization path.

[0004] However, traditional ground survey methods (such as sampling method and harvesting method) have obvious limitations: on the one hand, due to the limitation of manpower and material resources, the number of survey sample points is usually limited and unevenly distributed, which is difficult to represent the real situation of large-area heterogeneous grassland and lacks spatial representativeness; on the other hand, the complex terrain of plateau (such as alpine, hypoxia, rugged terrain) leads to high cost and low efficiency of field operation, which cannot meet the needs of long-term continuous monitoring. Compared with traditional methods, remote sensing technology has the advantages of wide coverage (can quickly obtain regional scale data), temporal and spatial consistency (multi-temporal observation under unified standard), which provides a breakthrough path for AGB inversion (i.e. monitoring biomass through remote sensing data), especially in areas such as alpine meadow where human activities are rare, its applicability is more prominent. In recent years, with the progress of remote sensing sensor technology, the data resolution (spatial, spectral, temporal) and multi-spectral observation ability continue to improve (such as hyperspectral data can capture the subtle spectral response differences of vegetation), and the AGB inversion method based on remote sensing image has gradually become the mainstream technical means of ecological monitoring.

[0005] In the development of AGB inversion method, traditional statistical methods have dominated, including linear regression, exponential regression, polynomial regression, etc. This kind of method realizes inversion by establishing the mapping relationship between remote sensing index and measured AGB, which is simple to operate and efficient, but also has obvious limitations: the performance of the method is highly dependent on environmental background (such as soil brightness, atmospheric interference) and species difference (different spectral characteristics of different grasses), which leads to poor transferability of the method in different regions or different vegetation types - for example, the method established in Kobresia meadow cannot be directly applied to the meadow of miscellaneous grass, which ultimately leads to insufficient universality and stability of the method.

[0006] In recent years, with the rapid development of artificial intelligence and computer technology, machine learning methods (such as random forest, gradient boosting tree, support vector machine, etc.) have been widely used in remote sensing inversion field. This kind of method can effectively handle the complex correlation between remote sensing data and AGB by constructing a complex nonlinear mapping relationship, and the inversion accuracy is significantly higher than that of traditional statistical methods. However, machine learning methods still have obvious shortcomings: first, they are highly dependent on a large number of high-quality training samples, and the field sampling of alpine meadow is difficult and costly, which is prone to problems such as sample scarcity or uneven distribution, directly leading to decreased generalization ability of the method; second, the method is mostly a "black box" structure, lacking the description of the physical mechanism between vegetation spectral response and biomass accumulation (such as the correlation between leaf chlorophyll content and photosynthesis intensity), which makes the method lack of interpretability and difficult to deal with uncertainty in complex environments.

[0007] Therefore, there is an urgent need to develop a new AGB inversion scheme: it can break through the limitation of insufficient field survey data, reduce the dependence on sample size, and integrate vegetation spectral physical processes and advanced machine learning methods (such as deep learning), which can improve the inversion accuracy while enhancing the interpretability and generalization ability of the scheme, that is, it has high precision and high reliability. SUMMARY

[0008] In order to alleviate or partially alleviate the above technical problems, the solution of the present application is as follows:

[0009] A high-cold meadow aboveground biomass monitoring method based on PROSAIL-BP, comprising:

[0010] Obtain high-cold meadow mask data, remote sensing image data and field aboveground biomass data about the target area, and perform geometric splicing on the remote sensing image data to obtain continuous regional image data set, and perform normalization processing on each band reflectivity and cutting combined with high-cold meadow mask data, only keep the remote sensing reflectivity data of the grassland region;

[0011] Predefine a PROSAIL model matched with the characteristics of the high-cold meadow of the target area;

[0012] Based on Sobol global sensitivity analysis of variance decomposition, Sobol first-order sensitivity index and Sobol total-order sensitivity index of the input parameters of the PROSAIL model are obtained;

[0013] According to the sensitivity index, a key band set in the remote sensing image is screened, and the four high-sensitivity model input parameters most significantly affecting the output result are screened according to the sensitivity index sorting, wherein the key band set is composed of the following center wavelengths: 442.9nm, 492.4nm, 832.8nm, 864.7nm, 945.1nm and 2202.4nm, and the four high-sensitivity model input parameters are leaf area index, dry matter content, leaf water content and soil factor;

[0014] A number of sets of scheme input parameter combinations covering different vegetation growth states and soil background variations of high-cold meadow are generated by using uniform sampling strategy, each set of parameters is input into the PROSAIL model to generate high spectral canopy reflectivity in the range of 400nm-2500nm by forward simulation, then the corresponding reflectivity is extracted according to the center wavelengths of the key bands to match the physical response of the remote sensing image, and finally the leaf area index x dry matter content is taken as the target variable, and a number of structured data sets containing four input feature fields and one target variable field are constructed;

[0015] The three-layer fully connected BP neural network is trained by using the structured data set, and the trained BP neural network is used to process the reflectivity corresponding to the key band set extracted from the remote sensing image data, and the ground biomass corresponding to each pixel is output.

[0016] In some embodiments, the remote sensing image data is Sentinel-2 L2A image data.

[0017] In some embodiments, the PROSAIL model input parameters are: dry matter content, chlorophyll content, leaf structure parameters, leaf water content, carotenoid content, brown pigment component, leaf area index, hotspot size, soil factor, solar zenith angle, observation azimuth angle, relative azimuth angle, average leaf inclination angle, and leaf inclination angle distribution shape.

[0018] In some embodiments, the four high-sensitivity model input parameters are set to variable ranges, and the remaining PROSAIL model input parameters are set to the mean values of predefined parameter value intervals.

[0019] In some embodiments, the input layer of the BP neural network is configured with 6 nodes, the hidden layer uses ReLU activation function and introduces Dropout regularization to suppress overfitting, and the output layer is set to 1 node.

[0020] In some embodiments, the neuron dropout rate in the Dropout regularization is 0.5.

[0021] In some embodiments, during the training of the three-layer fully connected BP neural network, the mean square error is used as the loss function, the Adam optimizer is used to optimize the loss function, and the learning rate is set to 0.001.

[0022] In some embodiments, the alpine meadow ground biomass monitoring method based on PROSAIL-BP is applied to a grassland monitoring system, an agricultural and pastoral management platform, or an ecological evaluation system.

[0023] In some embodiments, the key band set reflectivity is extracted from the pre-processed Sentinel-2 L2A image according to the sample plot center point coordinates in the field, the BP neural network model is input to obtain the ground biomass prediction value, and the prediction value is linearly regressed with the measured value to verify the model generalization ability.

[0024] In some embodiments, the ground biomass spatialization estimation of the target area is carried out: the data corresponding to the key band set of the Sentinel-2 L2A image after geometric correction, radiation normalization and grassland mask cropping are loaded, and the key band set reflectivity of the corresponding pixels in the grassland region is extracted one by one to input the BP neural network to obtain the ground biomass prediction value.

[0025] The technical scheme of the present application has one or more of the following beneficial technical effects:

[0026] (1) The present application overcomes the defect of insufficient field samples in the prior art, solves the problem of sample scarcity or difficulty in obtaining samples, and is particularly suitable for high-cold meadow, desert grassland and other areas where field measurement is difficult.

[0027] (2) The present application innovatively and organically combines the core advantages of physical mechanism and data-driven methods for the first time, forming a scheme with high reliability and high precision. The present application not only simulates the reflectivity under the structure and spectral characteristics of high-cold meadow, so that the scheme of the present application has high interpretability and generalization, but also learns complex nonlinear relationships from large sample data by referring to machine learning methods, so that the scheme has high precision.

[0028] (3) The present application can be easily extended to different regions or different years of scenes, and can process remote sensing images in batches and automatically to predict / monitor aboveground biomass, with high spatiotemporal scale adaptability. In addition, the scheme of the present application can be connected to grassland monitoring systems, agricultural and pastoral management platforms, ecological evaluation systems, etc., to automatically monitor or / and process remote sensing ecological image information, and has strong practical value.

[0029] In addition, the present application has other beneficial effects which will be mentioned in the specific embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 is a flow chart of the aboveground biomass monitoring method of the high-cold meadow based on PROSAIL-BP of the present application;

[0031] Figure 2 is a schematic diagram of field collection of aboveground vegetation according to an exemplary embodiment of the present application;

[0032] Figure 3 is a schematic diagram of the PROSAIL model according to an exemplary embodiment of the present application;

[0033] Figure 4 is a schematic diagram of Sobol global sensitivity analysis according to an exemplary embodiment of the present application;

[0034] Figure 5 is a distribution diagram of the first-order sensitivity and total-order sensitivity of the parameter in the wavelength of 400-2500 nm according to an exemplary embodiment of the present application;

[0035] Figure 6 is a distribution diagram of the first-order sensitivity and total-order sensitivity of the parameter in each wavelength band according to an exemplary embodiment of the present application;

[0036] Figure 7It is a schematic diagram of generating a training sample data set based on a PROSAIL model according to an exemplary embodiment of the present application.

[0037] Figure 8 It is a schematic diagram of a BP neural network scheme according to an exemplary embodiment of the present application.

[0038] Figure 9 It is a schematic diagram of a BP neural network scheme stability test result according to an exemplary embodiment of the present application.

[0039] Figure 10 It is a schematic diagram of a BP neural network scheme precision test result according to an exemplary embodiment of the present application.

[0040] Figure 11 It is a comparison diagram of a predicted AGB and a measured AGB according to an exemplary embodiment of the present application.

[0041] Figure 12 It is a target area alpine meadow AGB spatial prediction diagram according to an exemplary embodiment of the present application. DETAILED DESCRIPTION

[0042] In order to make the objects, technical solutions and advantages of the present application clearer, the technical solutions in the present application will be described clearly and completely below with reference to the drawings in the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0043] In order to clearly describe the technical solutions of the embodiments of the present application, in the embodiments of the present application, the same items or similar items with basically the same functions and effects are distinguished by using "first", "second", etc. The skilled in the art can understand that "first", "second", etc. do not limit the quantity and execution order.

[0044] Figure 1 It is a flowchart of an alpine meadow aboveground biomass monitoring method based on PROSAIL-BP of the present application.

[0045] The present application will be described below according to the flowchart shown in Figure 1 The specific embodiments of each step will be described carefully.

[0046] Step S1: Obtain alpine meadow mask data, remote sensing image data and field aboveground biomass data about a target area, perform geometric splicing on the remote sensing image data to obtain continuous regional image data set, perform normalization processing on the reflectivity of each band and cutting in combination with the alpine meadow mask data, and only retain the remote sensing reflectivity data of the grassland region.

[0047] In one embodiment of the present application, the data preparation process for aboveground biomass inversion is as follows: first, the geographical scope of the target area is determined, and multi-source data related to aboveground biomass inversion is systematically obtained, including alpine meadow (grassland) mask data, Sentinel-2 L2A level remote sensing image and field measured aboveground biomass sample data. Then the above data is integrated and preprocessed. The acquisition and preprocessing process of each data is described as follows:

[0048] In order to obtain alpine meadow mask data, in one example, grassland distribution information can be obtained based on Google Earth Engine (GEE) cloud computing platform, and ESA World Cover 2021 v200, a global land cover data product released by European Space Agency (ESA), is selected as the main data source. This product has a spatial resolution of 10m and good global applicability. The "Grassland" category in its land cover classification system is used as the basis for grassland identification. Through reclassification processing, the "Grassland" category in the original land cover map is assigned as the target category, and the remaining categories are set as the background. Then the grassland distribution range of the target study area is extracted, and a spatially consistent grassland mask layer is generated for subsequent remote sensing image cropping and feature extraction.

[0049] To obtain remote sensing image data, the multispectral remote sensing data used in the present application is derived from the EP Sentinel-2 satellite and provided by the United States Geological Survey (USGS), with a data level of surface reflectance product (Level-2A). In one example, 6 images taken on August 21, 2024 are selected to ensure coverage of the entire study area, with consideration of cloud cover control and temporal and spatial consistency.

[0050] During data preprocessing, geometric stitching is first performed on multiple images to construct a continuous and seamless regional image dataset. Then, normalization processing technology is used to standardize the reflectance values of each band to eliminate the radiation inconsistency caused by multiple stitching. Finally, combined with the aforementioned grassland mask layer, image mask cropping processing is performed to retain only the remote sensing reflectance data of the grassland area, in order to improve the relevance and accuracy of subsequent modeling analysis. In Table 1, the details of the Sentinel-2 L2A image band settings are shown.

[0051] To obtain field aboveground biomass data, in one example, ground measurement data was obtained from August to September 2024. Standardized sampling plots were laid out using repeated measurement method to obtain aboveground biomass measurement values.

[0052] Figure 2 is a schematic diagram of field collection of aboveground vegetation according to an exemplary embodiment of the present application. In this example, a circular plot with a radius of 10 meters is set up in each sample plot, and 3 standard quadrats of 1 m x 1 m are evenly arranged on the circumference of the circular plot at an angle of 120°. Independent sampling is performed in each quadrat to ensure sample representativeness. The collected aboveground plant samples are numbered on site and then taken back to the laboratory, and are placed in a constant-temperature drying oven at 65°C to dry to a constant weight. The obtained dry weight is used as the aboveground biomass index of the plant. The dry weights measured for the 3 quadrats in each sample plot are arithmetically averaged, and the average value is taken as the aboveground biomass reference value corresponding to the center point of the sample plot, to provide accurate ground verification data for remote sensing inversion modeling.

[0053] Table 1: Details of band settings of Sentinel-2 L2A images

[0054]

[0055] Step S2: Predefining a PROSAIL model matching the characteristics of the target alpine meadow region.

[0056] To achieve accurate remote sensing inversion of aboveground biomass of the target alpine meadow region, the core in the process of remote sensing inversion of aboveground biomass is to construct a radiation transfer scheme that can accurately describe the spectral response mechanism of vegetation, so as to provide a reliable physical basis for the inversion scheme. Based on this, the present application selects a vegetation canopy radiation transfer scheme, PROSAIL model, which can combine the physiological and ecological characteristics of the target region (alpine meadow) vegetation (such as leaf thickness, canopy structure), the published remote sensing inversion literature of alpine meadow, the experience value of the parameters of the scheme of the same type of region, and the remote sensing database (such as leaf spectrum database LAI product library), and can determine the reasonable value range of each parameter of the scheme through the "literature verification + data matching" method, to provide accurate input basis for subsequent scheme simulation and inversion scheme.

[0057] Figure 3 is a PROSAIL model schematic diagram according to an exemplary embodiment of the present application. This diagram clearly shows the radiation transfer simulation logic of the PROSAIL model known in the art from the leaf level to the canopy level, which specifically includes the coupling relationship and signal transmission path of the PROSPECT (Leaf Optical Properties Spectra) sub-scheme and the SAIL (Scattering by Arbitrarily Inclined Leaves) sub-scheme. First, the PROSPECT sub-scheme is used to calculate the leaf scale spectral characteristics, and then the SAIL sub-scheme is used to realize the canopy scale radiation transfer simulation.

[0058] PROSAIL model is a classical and widely used vegetation canopy bidirectional reflectance simulation scheme in the field of vegetation remote sensing. Its core advantage is to break the scale limit and realize the continuous simulation of radiation transmission from leaf scale to canopy scale by coupling PROSPECT (leaf level) and SAIL (canopy level) two sub-schemes. This scheme has dual application value: on the one hand, it can be used as a forward scheme of remote sensing image to simulate the canopy reflectance spectrum under different vegetation conditions; on the other hand, it can be used for inverse inversion of key physiological parameters of plants such as chlorophyll a+b content, leaf area index (LAI), leaf water content, leaf dry matter content, etc., to provide technical support for vegetation physiological and ecological monitoring.

[0059] PROSPECT sub-scheme (Leaf Optical Properties Spectra, leaf optical properties spectrum scheme) is mainly used to simulate the spectral reflectance and transmittance of single leaf in the wavelength range of 400~2500 nm. Its modeling core is based on the physical and chemical reaction mechanism between leaf spectral response and its internal biochemical composition (including chlorophyll a+b, carotenoids, leaf water, leaf dry matter). The difference in absorption and scattering of different biochemical components to light of specific wavelength directly affects the spectral performance of leaf. This scheme regards leaf as a stack of multiple equal-thickness optical layers (industry refers to it as "stack scheme"), uses radiation transmission theory and Beer-Lambert law (describes the relationship between light absorption and substance concentration, light path), and quantitatively simulates the propagation path, absorption loss and transmission law of light in leaf tissue, ensuring the physical authenticity of leaf spectral simulation results.

[0060] SAIL sub-scheme (Scattering by Arbitrarily Inclined Leaves, scattering by arbitrarily inclined leaves) focuses on the canopy level and simulates the scattering and reflection process of infinite and homogeneous vegetation canopy to incident solar radiation. It is especially suitable for modeling the bidirectional reflectance distribution function (BRDF) characteristics of canopy level, i.e. the law of canopy reflectance changing with observation angle and solar angle. This scheme is based on one-dimensional radiation transfer equation and assumes that the leaf is a plane mirror, the leaf inclination is arbitrarily distributed, and the vegetation background is homogeneous soil. Its input parameter system covers canopy structure parameters (such as leaf area index, average leaf inclination), observation geometry parameters (such as solar zenith angle, observation zenith angle) and soil background parameters (such as soil reflectance factor). Through the combination calculation of the above parameters, the reflectance data of canopy under specific observation conditions are output. The standard expression of this scheme belongs to public technology and is not described in detail here.

[0061] In the present application, the PROSPECT sub-scheme and the SAIL sub-scheme can be described by the following formula:

[0062]

[0063]

[0064] In the formula, N represents a leaf structure parameter; Cab represents chlorophyll a+b; Cw represents water content; Cm represents dry matter content; Car represents carotenoid content; and Cbp represents brown pigment component. represents reflectivity; represents transmittance, and LAI represents leaf area index; Hspot represents a hotspot effect factor; Psoil represents a soil factor; LIDFa represents an average leaf inclination angle; LIDFb represents a leaf inclination angle distribution shape; tts represents a solar zenith angle; tto represents an observation zenith angle; and psi represents a relative azimuth angle.

[0065] From the perspective of scheme applicability, the PROSAIL model has three core advantages: first, the physical driving characteristics are significant, and the scheme output depends on the physical mechanism rather than a large number of training samples; second, the parameter system has high physiological interpretability, each input parameter in the scheme corresponds to a specific leaf biochemical property (such as chlorophyll content) or canopy structure property (such as leaf area index), rather than an abstract mathematical variable, which facilitates subsequent interpretation of the scientific significance of the inversion results combined with ecological and physiological knowledge; and third, the function is compatible, and can be used for forward simulation, that is, calculating the spectral reflectance of the vegetation canopy under different wavebands (such as the blue, green, red, and near-infrared wavebands of the Sentinel-2 image) and observation conditions under given parameter conditions, and can also be used for reverse calculation, that is, based on the measured or remotely sensed reflectance to retrieve key biophysical parameters, and is widely applicable to various vegetation remote sensing inversion tasks.

[0066] In some preferred embodiments, in order to ensure that the input parameters of the PROSAIL model have good adaptability and representativeness under the environmental conditions (such as low temperature and strong radiation) and vegetation ecological characteristics (such as dwarfing and dense growth) of the target area (alpine meadow), in the parameter setting process, the strategy of “multi-source data fusion + regional adaptability adjustment” is adopted: on the one hand, the previous research results in the field of alpine meadow remote sensing inversion, the small-scale field measurement experience data obtained by the present application in the early stage, and the authoritative remote sensing product information are fully referred to; on the other hand, combined with the vegetation physiological characteristics (such as thicker leaves and lower water content) of the alpine meadow ecosystem in the target area, the reasonable value range of each scheme parameter is systematically defined, so as to avoid simulation deviation caused by mismatch between parameters and regional characteristics.

[0067] In the selection of parameter data sources, the application follows the principle of "priority matching regional characteristics": under the support of available experimental data, priority is given to literature and experimental databases that are highly consistent with the target area alpine meadow vegetation type (dominated by Kobresia, Stipa) in terms of ecological attributes (such as growth cycle), physiological structure (such as leaf thickness, biochemical component proportion) and geographical distribution (such as high altitude, low temperature environment) for the predefinition of scheme input parameters, to ensure the pertinence of parameter basic data.

[0068] For the predefinition of leaf biochemical and structural parameters, the determination of the value range of key leaf biochemical and structural parameters such as leaf structure parameter (N, representing the number of internal optical layers of the leaf), chlorophyll a+b content (Cab), leaf water content (Cw) and dry matter content (Cm) in the application takes the international classic leaf spectrum database LOPEX'93 (Leaf Optical Properties EXperiment, LOPEX) as the core data source: from the database, a sample group similar in leaf biochemical composition to the target area alpine meadow vegetation is selected (the specific selection criteria are: leaf dry matter content 0.0019~0.0165g / cm², chlorophyll a+b content 1~80μg / cm², matching the leaf characteristics of dominant species in alpine meadow), the statistical distribution characteristics (including mean, standard deviation, minimum value, maximum value) of the above-mentioned parameters in the sample group are extracted, and the chlorophyll content range inverted by MODIS vegetation index products (such as EVI index) is further compared and analyzed with the 50 leaf sample data measured in the field in the previous stage of the application, on the basis of verifying the consistency of multiple data sources (such as data deviation ≤15%), the final value range is determined, for example: leaf structure parameter N is set to 1~3, chlorophyll a+b content Cab is set to 1~80μg / cm², to enhance the physiological reality of the scheme input.

[0069] For carotenoid content (Car) and brown pigment content (Cbrown), due to the limitations of high altitude and poor transportation in the target area, there is still a lack of targeted field observation and experimental data, in order to avoid the underestimation or overestimation of the scheme due to the subjective setting of narrow range parameters, the parameter range is set to the theoretical maximum value interval supported by the PROSAIL model (Car is 0.1~20 μg / cm², Cbrown is 0.1~1) in the modeling stage, which can maximize the parameter fluctuation space without violating the physical stability constraints of the scheme (the parameter value does not exceed the allowed range of the scheme algorithm), covering the changes of carotenoid and brown pigment content of alpine meadow vegetation in different growth states (such as the green-up period, the yellowing period).

[0070] Meanwhile, in the canopy scale parameters, the setting of the leaf area index (LAI, representing the degree of leafiness of the canopy) parameter relies on the MODIS product MOD15A2H dataset (full name: MODIS / Terra + Aqua Leaf Area Index / FPAR 8-Day L4 Global 500m) released by NASA, which has the advantages of high temporal resolution (8 days) and wide spatial coverage, and has been globally verified for accuracy (LAI inversion accuracy ≥ 85%). The specific operation steps can be exemplarily: first, extract the MOD15A2H dataset corresponding to the same period (August 2024) as the Sentinel-2 image from the NASA Earth data platform; second, perform quality control on the dataset, and remove low-quality pixels caused by cloud coverage, aerosol interference, and sensor failure (retain high-quality pixels with QA value of "00") through the QA band; then, use the bilinear interpolation method to resample the 500m resolution LAI data to 10m resolution consistent with the Sentinel-2 L2A image, ensuring spatial scale matching; finally, statistics the LAI value distribution range (such as minimum value 0.8, maximum value 3.2, median value 1.8) of the resampled dataset, and accordingly set the representative interval of the LAI parameter in the PROSAIL model as 0.1~6.

[0071] For the important geometric and structural variables affecting the bidirectional reflectance characteristics of vegetation, the hotspot effect parameter (Hspot, representing the reflection enhancement effect when the sun's direct direction and the observation direction are consistent) and the soil background reflection factor (Psoil, representing the contribution of soil to canopy reflectance), considering the complex terrain (many gentle slopes and depressions) of the study area, the lack of detailed canopy-soil structure field measurement data (such as the influence of soil texture and organic matter content on reflectivity), the maximum parameter value range allowed by the PROSAIL model is adopted (Hspot is 0.01~1, Psoil is 0.01~1), which can cover the radiation response differences caused by different soil light and dark degrees (such as black calcareous soil and chestnut calcareous soil) and canopy structure combinations (such as sparse meadow and dense meadow) in the target area as much as possible, and reduce the simulation error caused by the lack of soil background information.

[0072] In addition, regarding the observation geometric parameters, i.e. the solar zenith angle (tts, the angle between the solar ray and the ground normal), the observation zenith angle (tto, the angle between the satellite sensor and the ground normal), and the solar-observation relative azimuth angle (psi, the horizontal angle between the solar incident direction and the satellite observation direction), in the preferred embodiment of the present application, the corresponding XML format metadata file of the Sentinel-2 L2A image product provided directly by the USGS is selected to extract: the solar zenith angle tts is obtained by locating the value of "SolarZenithAngle" under the "Solar_Angles" field in the metadata file, the observation zenith angle tto is obtained from the value of "ViewZenithAngle" under the "Viewing_Angles" field, and the relative azimuth angle psi is obtained from the value of "RelativeAzimuthAngle". The observation geometric parameters of each Sentinel-2 image are extracted separately and correspond to the sub-regions covered by the image, ensuring that the PROSAIL model simulation of different sub-regions uses angle parameters that completely match the actual remote sensing observation, significantly improving the matching degree between the simulated spectrum and the observed spectrum.

[0073] Finally, for the average leaf inclination angle (LIDFa) and the leaf inclination distribution function shape parameter (LIDFb), since the target area is wide (such as covering more than 1000 km²), there is a lack of detailed field measurement data for each region (such as measuring leaf inclination by a canopy analyzer), the present application sets their values to 0, i.e. the default crown leaf inclination is a spherically symmetric distribution. This processing method can not only simplify the modeling complexity to a certain extent (reduce 2 parameters that need to be calibrated), but also cover the crown leaf distribution state of most meadow communities (such as Kobresia meadow and Stipa steppe) in the target area through the universal characteristics of spherically symmetric distribution, taking into account the requirements of scheme simplification and simulation accuracy. Referring to Table 2, it shows the details of the pre-defined parameter range of the PROSAIL model of the present application.

[0074] Table 2: Details of the pre-defined parameter range of the PROSAIL model

[0075]

[0076] Step S3: Sobol global sensitivity analysis based on variance decomposition, obtaining Sobol first-order sensitivity index and Sobol total-order sensitivity index of the input parameters of the PROSAIL model.

[0077] Based on the pre-defined results of the PROSAIL model parameters in the receiving step S2 (including leaf biochemical parameters, crown structure parameters, observation geometry parameters and the like), the core S3 of the present step is to quantize the influence degree of each input parameter on the simulation results of the target band reflectivity of the Sentinel-2 L2A image through global sensitivity analysis based on variance decomposition, and finally realize the optimization of the input parameter system - that is, to eliminate low-sensitive parameters to reduce the parameter dimension of the subsequent inversion algorithm and improve the calculation efficiency, and to retain key parameters to ensure the accuracy of the above-ground biomass inversion.

[0078] In order to accurately identify the contribution difference of each input parameter (such as leaf structure parameter N, chlorophyll a+b content Cab, leaf area index LAI and the like) of the PROSAIL model to the remote sensing reflectivity output of the 13 bands (Band1-Band12, including one panchromatic band) of the Sentinel-2 L2A image, the present application does not use the traditional local sensitivity analysis (which can only evaluate the influence of the change of the parameter in the local interval, cannot capture the nonlinear interaction effect between parameters, and is not suitable), but selects the Sobol global sensitivity analysis method based on variance decomposition. The method can simultaneously quantize the influence of the independent action of a single parameter and the interaction of multiple parameters on the reflectivity output through the decomposition of the scheme output variance in the full space range of the parameters pre-defined in step S2, can perfectly adapt to the physical nonlinear characteristics of the PROSAIL model (such as the coupling radiation response between the leaf biochemical parameters and the crown structure parameters), and can provide reliable quantitative basis for parameter optimization.

[0079] Figure 4 It is the Sobol global sensitivity analysis schematic diagram according to the exemplary embodiments of the present application. The core process is "parameter sampling→ scheme forward→ variance decomposition→ sensitivity calculation": first, a sample combination covering the full parameter space is generated through random sampling, and the reflectivity output is obtained by inputting the PROSAIL model; then, based on the variance decomposition theory, the total variance of the reflectivity output is decomposed into the independent contribution of each parameter and the interaction contribution between parameters, and finally the proportion of each part is quantified through the sensitivity index. The Sobol global sensitivity analysis method was proposed by RU mathematician Ilya M. Sobol in 1993, and belongs to the category of global sensitivity analysis. Its core advantage is that it can directly process the nonlinear radiation transfer scheme of the PROSAIL model and the like which is physically driven without assuming linearity or monotonicity, and can simultaneously identify the independent influence of a single parameter and the high-order interaction influence of multiple parameters, avoiding the omission of key interaction effects by local analysis.

[0080] The application adopts two types of core sensitivity indexes to evaluate the influence of parameters: one is Sobol first-order sensitivity index (Sobol first-order sensitivity index, denoted as S1), which represents the contribution ratio of the independent action of a single parameter to the total variance of the reflectivity output of the PROSAIL model. The greater the S1 value, the more significant the influence of the parameter change on the reflectivity simulation result. The other is Sobol total sensitivity index (Sobol Total sensitivity index, denoted as ST), which represents the contribution ratio of the independent action of a single parameter and the interaction of the parameter with all other input parameters (such as Cab and LAI, N and Cw) to the total variance of the reflectivity output. The greater the difference between the ST value and the S1 value, the stronger the interaction effect of the parameter with other parameters.

[0081] Suppose a model is: where each parameter x i has a certain range of variation and is uniformly distributed within the range, and f(x) is integrable, then the total variance of the scheme output result can be represented as:

[0082] ;

[0083] In the formula, Var(Y) is the total variance of the scheme result, V i is the direct variance of the input parameter x i , V i,j and V 1,2,…,n are the coupling variances of the interaction of each parameter. The first-order sensitivity index of the parameter x i can be represented as:

[0084] ;

[0085] The second-order and higher-order sensitivity indexes of the input parameter x i can be represented as:

[0086] , , .

[0087] The total-order sensitivity index of the input parameter x i can be represented as:

[0088] .

[0089] Based on the Sobol global sensitivity analysis principle, combined with the input parameter range of the PROSAIL model defined in step S2, the sample module in the open source Python sensitivity analysis library SALib (Sensitivity Analysis Library) is used to call the Saltelli sampling method to sample the input variables. The Saltelli sampling method is a numerical approximation estimation method based on the Monte Carlo method, which can support the sensitivity calculation of first-order effect and high-order interaction, has the advantages of high sampling efficiency and small estimation variance, and can ensure that the generated samples are representative and can uniformly cover the whole parameter space. For example, in the specific operation, first, the value range of each parameter determined in step S2 (such as N: 1-3, Cab: 1-80 μg / cm², LAI: 0.1-6) is input into the parameter configuration dictionary of SALib, and the parameter distribution type is set to “uniform distribution” (which meets the assumption that the parameter has no significant bias in the predefined interval); then, combined with the calculation efficiency and accuracy requirement, the basic sampling number M is set to 1000, according to the sample size formula of Saltelli sampling algorithm (sample size N = 2xDxM, D is the number of parameters), finally 24000 groups of parameter combination matrix are generated (PROSAIL model has 12 core input parameters for sampling, and the observation geometry parameters are fixed as the image metadata extraction value, and are not involved in the analysis); considering that there are differences in the dimensions of different parameters (such as Cab unit μg / cm², LAI is dimensionless), “min-max standardization” is used to map all parameter samples to the [0, 1] interval (standardization formula is ), wherein x is the original sample value, x min , x max is the parameter extreme value predefined in step S2), so as to avoid the influence of dimension difference on sampling uniformity.

[0090] For example, after the above 24000 groups of standardized parameter samples are inversely standardized to the original dimension one by one, they are input into the coupled PROSAIL model, and the observation geometry parameters (sun zenith angle tts, observation zenith angle tto, relative azimuth angle psi) extracted from the Sentinel-2L2A image metadata in step S2 are substituted, the forward simulation of the canopy reflectivity is carried out on the 13 bands (Band1: 443 nm to Band12: 2202 nm) of the Sentinel-2L2A image, and finally the simulation reflectivity output results of the 13 bands corresponding to each group of input parameters are obtained (see Table 1), forming a complete reflectivity simulation data set.

[0091] Based on the above PROSAIL radiation transmission scheme, the canopy bidirectional reflectance simulation results of Sentinel-2 L2A image band are further quantitatively sensitive to the evaluation of the input parameters of the scheme: calling the analyze module in the Python sensitivity analysis toolkit SALib, using the Sobol global sensitivity analysis function (sobol.analyze) in it, calculating the first order sensitivity index S1 and the total order sensitivity index ST of each input parameter for each band in each observation area, setting the confidence level to 95% during the calculation process, and verifying the stability of the index through 1000 times bootstrap sampling to ensure that the result error is less than or equal to 5%.

[0092] The sensitivity analysis step of the present application can not only accurately identify the dominant influence factors of each input parameter on the reflectivity simulation of each spectral band (such as LAI, Cab, Cm, which usually show high sensitivity parameters), but also clearly reflect the nonlinear interaction mechanism between variables (such as the influence of the interaction effect of Cab and LAI on the reflectivity of near-infrared band), thereby providing solid theoretical support and quantitative basis for subsequent parameter optimization (eliminating low sensitive parameters, fixing medium sensitive parameters), scheme simplification (reducing parameter dimension), feature band selection (screening bands that respond significantly to sensitive parameters) and inversion strategy construction (focusing on key parameter inversion), laying a foundation for improving the efficiency and accuracy of alpine meadow aboveground biomass remote sensing inversion.

[0093] Step S4: According to the sensitivity index, the key band set in the remote sensing image is screened, and the four high sensitive model input parameters that have the most significant influence on the output result are screened according to the sensitivity index sorting, wherein the key band set is composed of the following center wavelengths: 442.9nm, 492.4nm, 832.8nm, 864.7nm, 945.1nm and 2202.4nm, and the four high sensitive model input parameters are leaf area index, dry matter content, leaf water content and soil factor.

[0094] Based on the Sobol global sensitivity analysis results of the PROSAIL model input parameters in step S3, the core of this step is to simultaneously carry out Sentinel-2 L2A image key band screening and scheme parameter setting optimization based on the quantitative indicators of Sobol first order sensitivity index (S1) and Sobol total order sensitivity index (ST) - by identifying the bands that respond significantly to high sensitive parameters and simplifying the setting of low sensitive parameters, further reducing the calculation complexity of the subsequent inversion method, while ensuring the simulation accuracy and parameter interpretation ability required for aboveground biomass inversion, providing high-quality input basis for subsequent modeling.

[0095] Figure 5This is a distribution diagram of the first-order sensitivity and total-order sensitivity of the parameters in the wavelength range of 400~2500nm according to an exemplary embodiment of the present invention. Figure 6 This is a distribution diagram of the first-order sensitivity and total-order sensitivity of the parameters in each band according to an exemplary embodiment of the present invention.

[0096] After completing step S3, which involves the Sobol global sensitivity analysis of the PROSAIL model input variables, this invention uses the S1 and ST values ​​obtained from this analysis as the core basis, combined with... Figure 5 and Figure 6 The sensitivity curves shown (the curve trends intuitively reflect the changes in the response intensity of different bands to the target parameters) and the exponential distribution (clearly presenting the numerical differences in sensitivity of each band) were used to selectively screen 13 multispectral bands (Band1~Band12, including 1 panchromatic band) of Sentinel-2 L2A imagery, aiming to identify band combinations that have a significant response to specific key biophysical parameters. Specifically, the two shortwave bands, Band 1 (442.9 nm) and Band 2 (492.4 nm), showed high sensitivity to leaf area index (LAI) and soil background reflectance factor (Psoil); the three near-infrared bands, Band 8 (832.8 nm), Band 8A (864.7 nm), and Band 9 (945.1 nm), showed outstanding sensitivity due to their strong responses to LAI, leaf dry matter content (Cm), and Psoil; and Band 12 (2202.4 nm), as a shortwave infrared band, showed significant sensitivity to LAI, leaf water content (Cw), and Psoil. Statistical analysis shows that the S1 and ST values ​​corresponding to the above-mentioned bands are all greater than 0.1, significantly higher than other bands (where S1 is mostly below 0.05 and ST is mostly below 0.08). This indicates that these bands make a substantial contribution to the total variance of the reflectance simulation output of the PROSAIL model, possessing not only good parameter identification capabilities but also effectively avoiding interference from low signal-to-noise ratio bands on the inversion results. Based on this, this invention selects the above six bands (Band1, Band2, Band8, Band8A, Band9, Band12) to construct a key band set for subsequent remote sensing feature extraction and biophysical parameter inversion modeling processes.

[0097] At the same time, according to the sensitivity index ranking result of each parameter in step S3, the present application further identifies the four high-sensitive model input parameters (referred to as high-sensitive parameters) that have the most significant impact on the output result of the scheme, i.e. LAI, Cm, Cw and Psoil; the remaining scheme input variables (such as low-sensitive model input parameters, referred to as low-sensitive parameters, such as carotenoid content Car, brown pigment content Cbrown, hot spot effect parameter Hspot, etc.) have significantly lower contribution to the output band reflectivity, and their sensitivity index is either close to zero or has a very narrow fluctuation range (such as the average S1 of Car is only 0.02), and the influence on the inversion result can be ignored, so in the subsequent parameter setting, low-sensitive variables can be simplified and processed.

[0098] After identifying the high-sensitive parameters and the key bands, the present application re-calibrates the input parameters of the PROSAIL model according to the optimization principle of "low-sensitive parameter fixed assignment, high-sensitive parameter reasonable range setting" to achieve the balance between scheme structure simplification and inversion efficiency improvement. Specifically, for low-sensitive parameters such as Car, Cbrown, Hspot, etc., the average value of the pre-defined parameter value interval in step S2 (such as Car 10 μg / cm², Cbrown 0.1, Hspot 0.1) is fixed and assigned - this assignment method not only conforms to the physiological and ecological characteristics of alpine meadow vegetation, but also reduces the number of adjustable parameters and the dimension of the scheme, significantly improving the efficiency of subsequent inversion calculation; and for the four high-sensitive parameters LAI, Cm, Cw and Psoil, a wider and more scientific variable range (such as LAI: 0.1-6, Cm: 0.0019-0.0165 g / cm², Cw: 0.004-0.01 g / cm², Psoil: 0.14-0.51) is set according to ecological rationality (such as LAI needs to conform to the actual distribution range of 0.1-6 of alpine meadow) and sensitivity intensity (such as Cm has a significant impact on near-infrared band reflectivity, and needs to be appropriately widened to cover vegetation in different growth states), which not only covers the parameter variation interval of alpine meadow vegetation in the target area, but also provides sufficient parameter space for subsequent lookup table construction and training sample generation. This optimization strategy focuses on high-sensitive parameters and key bands, effectively improving the sensitivity (more significant response to target parameter changes) and specificity (reducing the interference of irrelevant variables) of the PROSAIL model inversion, while ensuring the accuracy of the reflectivity simulation (the R² of the simulated value and the measured reflectivity of the Sentinel-2 L2A image is maintained above 0.85) and the physiological interpretability of the parameters, and lays a solid data foundation and scheme foundation for subsequent aboveground biomass remote sensing inversion (such as lookup table-based inversion algorithm). Table 3 is an example of PROSAIL model parameter range calibration details.

[0099] Table 3: PROSAIL model parameter range calibration details

[0100]

[0101] Step S5: Generate several sets of scheme input parameter combinations covering different vegetation growth states and soil background variations in alpine meadow in a uniform sampling strategy, input each set of parameters into the PROSAIL model to generate hyperspectral canopy reflectance in the range of 400 nm to 2500 nm through forward simulation, then extract the corresponding reflectance according to the key band center wavelength to match the physical response of the remote sensing image, and finally take leaf area index x dry matter content as the target variable to construct a structured data set containing 4 input feature fields and 1 target variable field for several samples.

[0102] The core of this step is to generate a large-scale training sample data set through PROSAIL model forward simulation--based on uniform sampling of parameter space, link the Sentinel-2L2A image key band reflectance, and construct a structured mapping relationship of "multispectral features-target variables", which provides high-quality data support under the physical mechanism constraint for subsequent machine learning-based aboveground biomass inversion scheme. Figure 7 is a schematic diagram of generating a training sample data set based on the PROSAIL model according to an exemplary embodiment of the present application, which clearly shows the complete process of "parameter sampling-forward simulation-band extraction-target variable correlation", and the specific construction process is as follows:

[0103] Based on the completed parameter re-calibration of the PROSAIL radiation transfer scheme, the application first carries out sample parameter space construction and forward simulation: based on the high sensitivity parameter value range (LAI, Cm, Cw, Psoil, other low sensitivity parameters Car, Cbrown, Hspot, etc. have been fixed as the average value) of the PROSAIL model calibrated in step S4, the sample module of SALib library under Python platform is used to generate a high-dimensional parameter combination space through the uniform sampling (Uniform Sampling) strategy. In order to ensure the coverage of different vegetation growth states (such as the green-up period, the full-bloom period, and the dry-yellow period) and soil background variations (such as different organic matter content soil) of the target alpine meadow, combined with the parameter dimension (4 high sensitivity parameters) and the simulation accuracy requirement, the sampling number is set to 50000 groups, each parameter is strictly within the calibration range and uniformly distributed in the parameter space (the interval of each parameter sample quantile is ≤2%), avoiding sample bias. The 50000 groups of parameters are input into the PROSAIL model one by one to perform forward simulation, and the hyperspectral canopy reflectance data in the range of 400-2500nm corresponding to each group of parameters are generated - each simulated spectrum curve corresponds to the canopy bidirectional reflectance characteristics under specific vegetation physiological state (such as high LAI representing dense meadow, high Cm representing sufficient leaf dry matter accumulation) and observation geometry condition (the observation parameters of Sentinel-2 L2A image extracted in step S2 are followed), which completely covers the potential variation range of alpine meadow canopy spectrum.

[0104] In order to realize the consistency of the simulation samples and the actual remote sensing observation data in the spectral dimension, the multispectral features need to be extracted from the above hyperspectral reflectance and aligned with the Sentinel-2 L2A image: for the 400-2500nm continuous hyperspectral data generated by the PROSAIL model, the reflectance values at the center wavelengths of each waveband are extracted according to the 6 key waveband center wavelengths screened in step S4 - specifically by finding the wavelength point closest to each waveband center wavelength in the hyperspectral data (wavelength deviation ≤1nm), the reflectance of this point is taken as the simulated reflectance of the corresponding waveband, which ensures that the simulated spectral features are completely matched with the sensor physical response of Sentinel-2 L2A image, avoids the error of subsequent inversion scheme caused by spectral mismatch, and provides input feature data with comparability and physical consistency for the construction of inversion scheme.

[0105] Finally, the target variable is set to complete the training data set construction. In view of the clear physiological correlation between the aboveground biomass of alpine meadow and the leaf area index (LAI) and the leaf dry matter content (Cm) of the vegetation, LAI represents the total number of leaves per unit area, Cm represents the dry matter weight of the leaves per unit area, and the product of the two can indirectly reflect the core contribution factor of the dry matter of the aboveground part of the vegetation per unit area, which has clear ecological significance and can reduce the deviation of a single parameter in the biomass prediction. Therefore, the application selects "LAI x Cm" as the target variable of remote sensing inversion, so as to enhance the ecological rationality and prediction accuracy of the modeling process. The 6 key band reflectances (input features) and "LAI x Cm" values (target variable) corresponding to each group of simulation samples are corresponded one by one, and a structured training sample data set is constructed, which contains 50000 sample records, each record covers 4 input feature fields and 1 target variable field, which not only ensures the scientificity of the data through the physical mechanism constraint of the PROSAIL model, but also establishes a clear input-output mapping relationship between the "multispectral reflectance-biomass correlation factor", so as to provide sufficient data and clear physical meaning for the subsequent construction of random forest, support vector machine and other machine learning methods based on machine learning method of aboveground biomass inversion scheme.

[0106] Step S6: training a three-layer fully connected BP neural network using the structured data set, and processing the reflectance corresponding to the key band set extracted from the remote sensing image data using the trained BP neural network to output the aboveground biomass corresponding to each pixel.

[0107] Preferably, the input layer of the BP neural network is configured as 6 nodes, the hidden layer adopts ReLU activation function and introduces Dropout regularization to suppress overfitting, and the output layer is set as 1 node.

[0108] Immediately after the construction of the "multispectral feature-target variable" structured training sample data set (containing 50000 samples, the input features are Sentinel-2 L2A key band reflectances, and the target variable is LAI x Cm) in step S5, the core of step S6 is to construct a back propagation neural network (Back Propagation Neural Network, BP-NN) scheme, which is verified by system stability test and precision verification to ensure that the scheme can realize effective fitting of the training samples and high-precision prediction of the aboveground biomass of alpine meadow, and has good robustness and generalization ability, thereby providing reliable scheme support for the final remote sensing inversion of the aboveground biomass.

[0109] Figure 8is the BP neural network scheme principle diagram according to the exemplary embodiment of the application. The three-layer network structure of "input layer-hidden layer-output layer" and the error back propagation path are clearly shown. BP-NN is a typical feedforward neural network, whose core mechanism is based on the error back propagation algorithm, and belongs to the supervised learning scheme of minimizing the loss function by gradient descent method - in the forward propagation process, the input features are calculated to obtain the predicted value through each layer of neurons; in the back propagation process, the error between the predicted value and the true value is transmitted in the reverse direction along the network level, and the connection weights and bias between the neurons in each layer are adjusted synchronously until the loss function converges to the minimum value. This scheme is widely used in the fields of nonlinear modeling, function fitting, pattern recognition and remote sensing inversion because it can approximate any continuous function. In the specific network structure, the input layer is used to receive the multi-spectral feature variables (Band1, Band2, Band8, Band8A, Band9, Band12 reflectance of Sentinel-2L2A) determined in step S5, the number of input nodes is set to 6 (consistent with the feature dimension), and only responsible for data transmission without performing calculation; the hidden layer is composed of multiple neurons, which first performs linear transformation (sum of weight and input feature product plus bias) on the input signal, and then applies a nonlinear activation function to extract deep nonlinear features of the data. In this embodiment, the activation function is selected as the rectified linear unit (Rectified Linear Unit, ReLU), which can effectively alleviate the gradient vanishing problem and improve the fitting efficiency of the scheme; the output layer generates the predicted value of the scheme according to the processing result of the hidden layer, and the number of output nodes is set to 1 (corresponding to the target variable LAIx Cm).

[0110] In one example, the application is based on Python3.9 programming environment and open source machine learning framework TensorFlow (version 2.10.0) to design and build the above-mentioned three-layer fully connected BP-NN scheme. In order to suppress the overfitting problem caused by sample redundancy in the training process of the scheme and enhance the generalization performance of the scheme to new data, the Dropout regularization mechanism is introduced in the hidden layer, and the neuron dropout rate (Dropout rate) is set to 0.5 - that is, 50% of the hidden layer neurons are randomly discarded during training to avoid excessive dependence between neurons; the mean squared error (Mean Squared Error, MSE) is used as the loss function (Loss) in the training process of the scheme, and the MSE calculates the square mean of the difference between the predicted value and the true value to accurately measure the deviation between them; at the same time, the Adam optimizer (learning rate is set to 0.001) is used to optimize the loss function, which can speed up the convergence speed of the scheme and improve the stability of weight update.

[0111] After the scheme structure is determined, in order to test the performance stability of the BP-NN under different data division conditions, the Bootstrap resampling method is used to process the training sample data set of step S5: 50000 samples are extracted from the original data set in each time (consistent with the size of the original data set) in a sampling manner with replacement, to form a resampling training set, and then the resampling training set is randomly divided into a sub-training set (used for scheme training) and a sub-test set (used for intermediate verification) in a ratio of 7:3, based on which a modeling and testing process is completed; the above resampling, division, modeling and testing process is repeated 200 times to fully simulate the influence of different data sampling scenarios on the performance of the scheme. The whole stability testing process takes the coefficient of determination (Coefficient of Determination, R²) and the root mean square error (Root Mean Square Error, RMSE) as the core precision evaluation indexes: R² is used to measure the linear fitting degree of the predicted value and the true value of the scheme (the value range is 0-1, and the closer to 1, the better the fitting effect); and RMSE is used to quantify the average level of prediction error (the unit is consistent with the target variable, and the smaller the value, the higher the precision).

[0112] Figure 9 FIG. 1 is a BP neural network scheme stability test result display diagram according to an exemplary embodiment of the present application. The 200 times of repeated training and testing results of the scheme show that the R² distribution interval of the scheme output under different sampling conditions is 0.85-0.87 (statistical test p<0.05, indicating that the results have statistical significance), and the RMSE is between 76.90~83.53gm⁻² (also p<0.05), and the standard deviation of the 200 results is small (R² standard deviation 0.005, RMSE standard deviation 1.21gm⁻²), which shows that the scheme maintains high fitting precision and prediction consistency under different data division scenarios, and there is no significant performance fluctuation, proving that it has good stability and anti-interference ability.

[0113] After the stability verification is passed, the trained BP-NN scheme is further independently verified to evaluate the actual generalization ability of the scheme: an independent test set (15000 samples, accounting for 30% of the original data set, and no data overlap with the training process) in the data set of step S5 which does not participate in Bootstrap resampling is selected, the reflectivity of the 6 key wavebands of the test set is input into the scheme, the corresponding target variable predicted value is obtained, and then the true target variable value of the test set is compared and analyzed.

[0114] Figure 10is a BP neural network scheme precision test result display figure according to an exemplary embodiment of the present application. The scheme has excellent fitting performance on the independent test set: the determination coefficient R² reaches 0.86, indicating that the linear correlation degree of the predicted value and the true value is high; the root mean square error RMSE is 74.15 gm⁻², the prediction error is at a low level, and the error distribution is uniform (no obvious systematic deviation). The verification result fully proves that the BP-NN scheme constructed in this embodiment not only can effectively fit the training data, but also can accurately predict new data that has not been seen before, and has practical application potential and practical value in the task of remote sensing driven alpine meadow biomass inversion.

[0115] Finally, in order to further evaluate the reliability and precision performance of the BP neural network model constructed in the actual remote sensing inversion application, the present application carries out independent verification of the model in the actual application scene by field measured aboveground biomass data. After confirming the reliability of the model under real ground conditions, further spatialization estimation of aboveground biomass in the target area of alpine meadow is carried out, which provides quantitative data support for grassland resource monitoring and ecological management.

[0116] In order to fully evaluate the precision and applicability of the BP neural network model in the actual remote sensing inversion task, the present application introduces field measured aboveground biomass independent samples which have no overlap with the training samples to carry out verification - the measured samples are derived from the standardized sampling sample plots (a total of N sample plots, each sample plot takes the average dry weight of 3 1m x 1m quadrats as the sample AGB true value, unit: gm⁻²) laid in the target area from August to September 2024 in the example land 1. In the verification process, first, the 6 key bands (Band1, Band2, Band8, Band8A, Band9, Band12) reflectance of the corresponding pixels are extracted from the preprocessed Sentinel-2L2A image through the latitude and longitude coordinates of the sample center point; then the reflectance data is input into the trained BP neural network model to obtain the AGB prediction value of the sample center point pixel; finally, the linear regression analysis is carried out between the measured true value and the model prediction value of each sample plot to test the fitting effect and generalization ability of the model under the condition of non-simulation data.

[0117] Figure 11is a comparison chart of predicted AGB and measured AGB according to an exemplary embodiment of the present application. The linear correlation between the two is clearly shown. The statistical results show that there is a significant linear correlation between the model prediction value and the field measured data: the determination coefficient (R²) reaches 0.82, indicating that the model can explain 82% of the variation of the measured AGB; the root mean square error (RMSE) is 89.07 gm⁻², and the prediction error is within the reasonable range of the actual variation of the AGB of the target region (the average AGB of the target region is about 350 gm⁻², and the error ratio is ≤25%); and the above R² and RMSE results pass the statistical significance test at the 0.05 level (p<0.05), proving that the model can maintain high fitting accuracy and prediction stability after being separated from the simulation training data, and has good generalization ability and practical application value.

[0118] After confirming that the model passes the field measurement verification, the present application is applied to the spatialization estimation of aboveground biomass in the target region: first, load the key band reflectivity data of the Sentinel-2 L2A image of the target region which has been preprocessed in step S1 (geometric correction, radiation normalization and grassland mask cutting have been completed, and only the grassland area pixels are retained); then for each 10m resolution pixel within the grassland mask range, extract its Band1, Band2, Band8, Band8A, Band9, Band12 reflectivity as input features one by one, input the BP neural network model for AGB prediction, and obtain the AGB prediction value (unit: gm⁻²) corresponding to each pixel; finally, based on the spatial analysis function of ArcGIS software, integrate the AGB prediction values of all pixels into a spatial distribution map of the aboveground biomass of the target region, realizing the spatialization expression of AGB from the pixel level to the regional level.

[0119] Figure 12 is a spatial prediction map of AGB of the target region alpine meadow according to an exemplary embodiment of the present application. The map clearly presents the spatial heterogeneity characteristics of the AGB of the study area: for example, the AGB value in the valley area is generally higher (mostly in 400~500 gm⁻²) due to sufficient water conditions, while the AGB value in the slope top area is relatively lower (mostly in 200~300 gm⁻²) due to thin soil and water shortage, which is highly consistent with the actual ecological environment characteristics of the target region. The spatial distribution results can provide accurate and quantitative basic data support for subsequent dynamic monitoring of alpine meadow grassland resources, evaluation of ecosystem service functions and sustainable management decision of grassland.

[0120] In order to better illustrate the present application, numerous specific details are given in the foregoing specific embodiments. Those skilled in the art should understand that the present application can also be implemented without some specific details. In some examples, methods, means, elements and circuits familiar to those skilled in the art are not described in detail in order to highlight the main idea of the present application.

[0121] The above merely provides the specific implementation of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of the changes or replacements within the technical range disclosed by the present application, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for monitoring aboveground biomass in alpine meadows based on PROSAIL-BP, characterized in that, include: Acquire alpine meadow mask data, remote sensing image data, and field biomass data for the target area. Geometrically stitch the remote sensing image data to obtain a continuous regional image dataset. Normalize the reflectance of each band and crop the data in conjunction with the alpine meadow mask data, retaining only the remote sensing reflectance data of the grassland area. A predefined PROSAIL model that matches the alpine meadow characteristics of the target area; Based on variance decomposition, Sobol global sensitivity analysis is used to obtain the Sobol first-order sensitivity index and Sobol total-order sensitivity index of the PROSAIL model input parameters. The key band set in the remote sensing image was selected based on the sensitivity index, and the four most sensitive model input parameters with the most significant impact on the output results were selected according to the sensitivity index. The key band set consists of the following center wavelengths: 442.9nm, 492.4nm, 832.8nm, 864.7nm, 945.1nm and 2202.4nm. The four highly sensitive model input parameters are leaf area index, dry matter content, leaf water content and soil factors. Using a uniform sampling strategy, several sets of input parameter combinations for different vegetation growth states and soil background variations covering alpine meadows were generated. Each set of parameters was input into the PROSAIL model to generate hyperspectral canopy reflectance in the range of 400nm~2500nm in a forward simulation. Then, the corresponding reflectance was extracted based on the center wavelength of the key band to match the physical response of the remote sensing image. Finally, with leaf area index × dry matter content as the target variable, a structured dataset containing several samples with 4 input feature fields and 1 target variable field was constructed. A three-layer fully connected BP neural network is trained using the structured dataset, and the trained BP neural network is used to process the reflectance corresponding to the key band set extracted from the remote sensing image data, outputting the aboveground biomass corresponding to each pixel.

2. The method for monitoring aboveground biomass in alpine meadows based on PROSAIL-BP according to claim 1, characterized in that: The remote sensing image data is Sentinel-2 L2A image data.

3. The method for monitoring aboveground biomass in alpine meadows based on PROSAIL-BP according to claim 2, characterized in that: The input parameters for the PROSAIL model are: dry matter content, chlorophyll content, leaf structure parameters, leaf water content, carotenoid content, brown pigment composition, leaf area index, hotspot size, soil factors, solar zenith angle, observation azimuth angle, relative azimuth angle, average leaf tilt angle, and leaf tilt angle distribution shape.

4. The method for monitoring aboveground biomass in alpine meadows based on PROSAIL-BP according to claim 3, characterized in that: The four highly sensitive model input parameters are set to a variable range; For the remaining PROSAIL model input parameters, they are set to the average of the predefined parameter value range.

5. The method for monitoring aboveground biomass in alpine meadows based on PROSAIL-BP according to claim 4, characterized in that: The input layer of the BP neural network is configured with 6 nodes, the hidden layer uses the ReLU activation function and introduces Dropout regularization to suppress overfitting, and the output layer is set to 1 node.

6. The method for monitoring aboveground biomass in alpine meadows based on PROSAIL-BP according to claim 5, characterized in that: The neuron dropout rate in the Dropout regularization is 0.

5.

7. The method for monitoring aboveground biomass in alpine meadows based on PROSAIL-BP according to claim 6, characterized in that: During the training of the three-layer fully connected BP neural network, the mean squared error was used as the loss function, and the Adam optimizer was used to optimize the loss function. The learning rate was set to 0.

001.

8. The method for monitoring aboveground biomass in alpine meadows based on PROSAIL-BP according to claim 7, characterized in that: The PROSAIL-BP-based method for monitoring aboveground biomass in alpine meadows can be applied to grassland monitoring systems, agricultural and pastoral management platforms, or ecological assessment systems.

9. The method for monitoring aboveground biomass in alpine meadows based on PROSAIL-BP according to claim 8, characterized in that: Based on the coordinates of the center point of the sample plot in the field, the reflectance of the key band set was extracted from the preprocessed Sentinel-2 L2A image and input into the BP neural network model to obtain the predicted value of aboveground biomass. The predicted value and the measured value were subjected to linear regression analysis to verify the generalization ability of the model.

10. The method for monitoring aboveground biomass in alpine meadows based on PROSAIL-BP according to claim 9, characterized in that: Spatial estimation of aboveground biomass in the target area: Load the data corresponding to the key band set of Sentinel-2 L2A image after geometric correction, radiometric normalization and grassland masking, extract the reflectance of the key band set for each corresponding pixel in the grassland area and input it into the BP neural network to obtain the predicted value of aboveground biomass.

Images