ICESat-2 adaptive photon denoising method and system based on vegetation layering

Abstract

The invention discloses an ICESat-2 adaptive photon denoising method and an ICESat-2 adaptive photon denoising system based on vegetation layering. According to the method, a vegetation layering strategy is introduced, the photon data is divided into different gradient grades according to the vegetation coverage, and the problem that a traditional global denoising method is insufficient in adaptability in a complex vegetation environment is effectively solved. And the optimal denoising parameter of each vegetation level is adaptively determined in combination with the Bayesian optimization algorithm, so that the photon denoising precision and efficiency of the DBSCAN clustering in the vegetation region with uneven density are remarkably improved. According to the method, the advantages of multi-source remote sensing data are fully utilized, efficient screening and separation of signal photons and noise under different vegetation density conditions are achieved, and a higher-quality data basis is provided for subsequent forest structure parameter inversion and ground feature classification.

Description

Technical Field

[0001] This invention relates to the field of remote sensing information processing and application technology, specifically to an ICESat-2 adaptive photonic denoising method and system based on vegetation stratification. Background Technology

[0002] Remote sensing technology, as a core means of Earth observation, has demonstrated irreplaceable value in many fields such as resource surveys, environmental monitoring, and ecological assessment due to its advantages of wide coverage, short observation cycles, and efficient data acquisition. Among them, spaceborne lidar (LiDAR) systems, represented by the Ice, Cloud and Land Elevation Satellite-2 (ICESat-2), can accurately capture three-dimensional information about the Earth's surface structure and are widely used in scenarios such as forest resource surveys, glacier monitoring, and topographic mapping.

[0003] However, due to factors such as the satellite operating environment, atmospheric interference, and complex surface reflections, the signal-to-noise ratio of ICESat-2 satellite-borne lidar data is low, with significant noise interference. If this noise is not effectively removed, it will severely interfere with the extraction of key vegetation parameters such as vegetation height and cover, leading to a substantial decrease in parameter extraction accuracy and consequently affecting the reliability of subsequent ecological assessments and resource management. Traditional denoising methods typically perform global denoising, failing to consider local noise variations. Therefore, developing an ICESat-2 adaptive photonic denoising method based on vegetation stratification has significant application value and practical significance. Summary of the Invention

[0004] Therefore, it is necessary to provide an ICESat-2 adaptive photonic denoising method based on vegetation layering that can achieve high-precision photonic denoising, addressing the aforementioned technical problems.

[0005] To achieve the above objectives, this invention provides an ICESat-2 adaptive photon denoising method based on vegetation stratification, comprising the following steps: S1. Acquire discrete footprint points of the ICESat-2 spaceborne lidar, Sentinel-2 optical data, and airborne LiDAR data; S2. Preprocess the discrete footprint point data, Sentinel-2 data and airborne LiDAR data to obtain the geographical location and elevation information, Sentinel-2 surface reflectance and airborne LiDAR point cloud dataset corresponding to each spaceborne LiDAR footprint point. S3. Based on the Sentinel-2 surface reflectance, the spatial distribution data of NDVI is calculated; S4. Based on the NDVI spatial distribution data and the geographical location information of the spaceborne lidar footprint points, extract the FVC information of each photon location from the FVC spatial distribution; S5. Based on the FVC information of each extracted photon location, the photon data is divided into several vegetation gradient levels, and the photon data of each vegetation gradient level is denoised using DBSCAN clustering. S6. Based on the elevation information of the spaceborne lidar footprint points and the airborne LiDAR point cloud dataset, evaluate the accuracy of the denoised photon data.

[0006] Preferably, the preprocessing step in step 2 includes: Perform data quality checks on discrete footprint point data, remove outliers, and extract footprint points from the spaceborne lidar. Radiometric calibration, atmospheric correction, and topographic correction were performed on the Sentinel-2 optical data, and the Sentinel-2 surface reflectance was calculated. Geographic registration, point cloud denoising, and point cloud classification were performed on airborne LiDAR data to calculate continuous digital elevation model (DEM) and digital surface model (DSM). The coordinate system of all data is unified, and the spaceborne lidar footprint points are spatially matched with airborne LiDAR data and Sentinel-2 optical data to extract the latitude, longitude and elevation information corresponding to each spaceborne lidar footprint point.

[0007] Preferably, the method for calculating the NDVI spatial distribution data in step 3 includes: calculating the continuous NDVI spatial distribution data within the study area based on the red band reflectance data and near-infrared band reflectance data of Sentinel-2 imagery and according to the standard NDVI calculation model.

[0008] Preferably, Sentinel-2 red band reflectance and near-infrared band reflectance are used as input variables, and pixel-by-pixel calculations are performed according to the standard calculation model to generate spatially continuous NDVI spatial distribution data. The standard calculation model for NDVI is as follows: , In the formula, NIR is the near-infrared reflectance of Sentinel-2, and Red is the red reflectance of Sentinel-2.

[0009] Preferably, the method for extracting FVC information for each photon location in step 4 includes: calculating the spatial distribution of surface vegetation cover (FVC) based on the NDVI spatial distribution data and combining it with a pixel binary model; and extracting FVC information for each photon location from the FVC spatial distribution based on the geographical location information of the spaceborne lidar footprint points.

[0010] Preferably, the method for denoising photon data at each vegetation gradient level using DBSCAN clustering in step 5 includes: The DBSCAN algorithm is used to denoise photons under different vegetation gradient levels. The DBSCAN algorithm defines sample points as core points, boundary points and noise points. For dataset D, if sample p e -If a neighborhood contains MinPts samples, then sample p is called a core point; For a sample b that is not a core point, if b lies at any core point p... e - Within the neighborhood, sample b is called a boundary point; For a sample n that is not a core point, if n is not in any core point... e - Within the neighborhood, sample n is called a noise point.

[0011] Preferably, in step S6, the method for evaluating the accuracy of the denoised photon data includes: The denoised photon elevation information and the elevation information obtained from the airborne LiDAR point cloud dataset were used to evaluate accuracy. Accuracy evaluation was achieved through recall (R), precision (P), and F scores, calculated using the following formula: In the formula, TP represents the number of signal photons that are correctly identified, FP represents the number of noise photons that are mistakenly identified as signal photons, and FN represents the number of signal photons that are not correctly identified.

[0012] The present invention also provides an ICESat-2 adaptive photon denoising system based on vegetation stratification. The system is used to implement the above method and includes: an acquisition module, a processing module, a calculation module, an extraction module, a denoising module, and an evaluation module.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention introduces a vegetation stratification strategy, dividing photon data into different gradient levels based on vegetation cover, effectively addressing the challenge of insufficient adaptability of traditional global denoising methods in complex vegetation environments. By combining this with a Bayesian optimization algorithm to adaptively determine the optimal denoising parameters for each vegetation level, the accuracy and efficiency of DBSCAN clustering for photon denoising in areas with uneven vegetation density are significantly improved. This method fully leverages the advantages of multi-source remote sensing data, achieving efficient identification and separation of signal photons and noise under different vegetation density conditions, providing a higher-quality data foundation for subsequent forest structure parameter inversion and land cover classification. Attached Figure Description

[0014] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention; Figure 2 The figure shows a comparison of the results of the vegetation stratification algorithm based on the present invention with other classic denoising algorithms; where (a) is the Jixi natural forest research area; and (b) is the Wangyedian artificial forest research area. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] First, the technical terms used in this invention will be explained.

[0019] The ε-neighborhood is a fundamental geometric concept in the DBSCAN algorithm used to define local density and determine point types. It transforms the abstract problem of determining data point density into a computable spatial point count problem using a simple distance radius ε.

[0020] Example 1 This embodiment provides an ICESat-2 adaptive photonic denoising method based on vegetation stratification, including the following steps: S1. Acquire discrete footprint points of the ICESat-2 spaceborne lidar, Sentinel-2 optical data, and airborne LiDAR data.

[0021] Acquire discrete footprint points of the ICESat-2 spaceborne lidar, Sentinel-2 optical data, and airborne LiDAR data; preprocess the discrete footprint point data, Sentinel-2 data, and airborne LiDAR data to obtain the geographical location and elevation information, Sentinel-2 surface reflectivity, and airborne LiDAR point cloud dataset corresponding to each spaceborne lidar footprint point.

[0022] The ICESat-2 spaceborne lidar, with its direct measurement advantage, can acquire key structural parameters such as forest canopy height. However, its discrete footprint data is often affected by noise interference from atmospheric scattering and surface clutter, making it difficult to directly meet the requirements of high-precision applications. Sentinel-2 optical data, on the other hand, has a wide coverage range and can provide horizontal distribution characteristics of surface vegetation (such as vegetation type and cover density). These horizontal characteristics can serve as important criteria for distinguishing signal photons from noise photons. Combining this data with ICESat-2 data for denoising is expected to reduce noise interference through multi-dimensional information complementarity, thereby improving the accuracy of lidar data. Meanwhile, high-precision airborne LiDAR data serves as a field reference, acquiring high-precision and reliable three-dimensional information on the surface and vegetation, which can be used to verify the accuracy of preprocessed data.

[0023] S2. Preprocess the discrete footprint point data, Sentinel-2 data and airborne LiDAR data to obtain the geographical location and elevation information, Sentinel-2 surface reflectance and airborne LiDAR point cloud dataset corresponding to each spaceborne LiDAR footprint point.

[0024] Data quality checks were performed on discrete footprint data to remove outliers and extract the spaceborne lidar footprint points. Radiometric calibration, atmospheric correction, and topographic correction were performed on Sentinel-2 optical data, and Sentinel-2 surface reflectance was calculated. Georegistration, point cloud denoising, and point cloud classification were performed on airborne LiDAR data, and continuous digital elevation models (DEM) and digital surface models (DSM) were calculated. The coordinate systems of all data were unified, and the spaceborne lidar footprint points were spatially matched with the airborne LiDAR data and Sentinel-2 optical data to extract the latitude, longitude, and elevation information corresponding to each spaceborne lidar footprint point.

[0025] S3. Based on Sentinel-2 surface reflectance, the spatial distribution data of NDVI is calculated.

[0026] First, the spatial distribution data of NDVI is calculated based on Sentinel-2 image data. Then, the spatial distribution of surface vegetation cover (FVC) is calculated based on the spatial distribution data of NDVI using a pixel binary model. Finally, the correlation between the spatial distribution of FVC and photon coordinates is established by combining photon geographic location information, and the FVC information of the corresponding photon location is extracted. While the Normalized Difference Vegetation Index (NDVI) effectively characterizes the growth status and green cover of surface vegetation, making it a commonly used indicator for reflecting vegetation information, its ability to distinguish areas with high vegetation cover is relatively weak. Once vegetation cover reaches a certain level, the NDVI value tends to saturate, making it difficult to accurately reflect subtle differences in vegetation density within high-coverage areas. It is also susceptible to interference from factors such as soil background reflection and atmospheric aerosols, leading to biases in vegetation information representation. In contrast, Surface Vegetation Cover (FVC), calculated using a pixel-based binary model, decomposes pixels into vegetated and non-vegetated portions, directly quantifying the proportion of surface vegetation cover. This provides a more intuitive and accurate reflection of differences in vegetation distribution across different areas, especially in complex vegetation environments, where it more clearly depicts vegetation density gradients. Based on the accurate representation of vegetation information by FVC, subsequent vegetation stratification denoising based on FVC gradients allows for more targeted denoising of photon data from areas with different coverage levels, making high-precision denoising possible.

[0027] Based on red and near-infrared reflectance data from Sentinel-2 imagery, and according to the standard NDVI calculation model, continuous NDVI spatial distribution data within the study area were calculated. The expression is as follows: , In the formula, NIR is the near-infrared reflectance of Sentinel-2, and Red is the red reflectance of Sentinel-2.

[0028] S4. Based on the spatial distribution data of NDVI and the geographical location information of the spaceborne lidar footprint points, extract the FVC information of each photon location from the FVC spatial distribution.

[0029] By combining a pixel-based binary model, the spatial distribution of surface vegetation cover (FVC) is calculated, clarifying the vegetation cover degree at each spatial location within the study area. Photon geographic location information acquired by spaceborne lidar is introduced, and a spatial coordinate matching method is used to map the FVC spatial distribution map to photon coordinates, thereby accurately extracting the vegetation cover information for each photon's corresponding spatial location from the FVC spatial distribution map. The specific process includes: based on continuous NDVI spatial distribution data, combined with a pixel-based binary model, the core logic of which is: assuming each image pixel consists of a vegetation-covered portion and a non-vegetation-covered portion, the proportion of each portion in the pixel is calculated using the NDVI's full vegetation cover threshold and no vegetation cover threshold, thus obtaining the pixel's vegetation cover degree. Pixel-by-pixel calculations are performed, and by calculating the coverage of all pixels in the image area, an FVC spatial distribution map is generated, thereby clarifying the vegetation cover degree at each spatial location. The resulting FVC calculation results are as follows: Among them, the pixel value of pure vegetation partial coverage is from This indicates that the pixel value of the partially covered bare soil is determined by... This indicates that the pixel value at the 5% confidence level is taken as... The pixel value at the 95% confidence level is taken as... ,Will and The vegetation cover of the study area was calculated by substituting the values ​​into the formula.

[0030] S5. Based on the FVC information of each extracted photon location, the photon data is divided into several vegetation gradient levels, and the photon data of each vegetation gradient level is denoised using DBSCAN clustering.

[0031] Based on the obtained vegetation cover (FVC), the photon data is divided into different vegetation gradient levels. DBSCAN clustering is then performed on the photon data of each stratified interval. To improve the efficiency and accuracy of denoising, the optimal parameter combination in the DBSCAN clustering denoising process is found using the Bayesian optimization algorithm. Based on the optimal parameter combination, the clustering denoising of the photon data in each interval is completed, and finally, accurate denoising of the photon data is achieved.

[0032] DBSCAN, a classic density clustering denoising algorithm, is widely used in photon data denoising due to its ability to identify noise clusters of arbitrary shapes without requiring a preset number of clusters. It is suitable for distinguishing between discrete outliers caused by noise and real signal photons in ICESat-2 data. However, the denoising effect of this algorithm is highly dependent on the settings of parameters such as the neighborhood radius and the minimum number of neighboring points. Improper parameter values ​​can easily lead to incomplete noise removal or the accidental deletion of effective photons. Furthermore, traditional parameter tuning often relies on manual trial and error, which is not only inefficient but also makes it difficult to find the globally optimal parameter combination.

[0033] Bayesian optimization algorithms, by constructing a probabilistic model of the objective function, can dynamically adjust the search direction based on historical parameter tuning results, quickly locate the optimal parameter combination of DBSCAN, avoid the blindness of manual trial and error, and significantly shorten the parameter optimization time. This effectively solves the problem of balancing efficiency and accuracy in traditional denoising, ensuring efficient and accurate photon denoising in all vegetation gradient intervals, and meeting the requirements for denoising efficiency and accuracy in complex vegetation environments.

[0034] The DBSCAN algorithm is used to denoise photons under different vegetation gradient levels. The sample points of the DBSCAN algorithm are composed of core points, boundary points, and noise points. The definitions of these three types of sample points are as follows: For a given dataset D, if sample p -The neighborhood contains at least If there are n samples (including sample p), then sample p is called the core point, that is: .

[0035] Let p be the core point, where - Neighborhood The expression is: in, q This represents another sample point in dataset D, used to determine whether it is a sample point. p Within the ε-neighborhood; D This represents the entire dataset, containing all the sample points to be analyzed. distance This represents the distance metric function.

[0036] For a sample b that is not a core point, if b lies at any core point p... - Within the neighborhood, sample b becomes the boundary point, that is: .

[0037] For a sample n that is not a core point, if n is not in any core point... - Within the neighborhood, sample n is called a noise point, that is: .

[0038] The optimal parameter combination in the DBSCAN clustering denoising process is found using a Bayesian optimization algorithm. Based on this optimal parameter combination, clustering denoising of photon data in each interval is completed, ultimately achieving accurate denoising of the photon data, including: Bayesian optimization utilizes Bayesian statistical principles to guide the search process, establishing a Gaussian process model to predict the impact of hyperparameter combinations on model performance and automatically selecting the next set of hyperparameters to be evaluated. This method uses a Gaussian process (GP) as the objective function, which is a stochastic process defined on a continuous input space; the joint distribution of any finite set of points is a multivariate Gaussian distribution. A GP is defined by a mean function... Sum of covariance functions Complete definition, that is: in, x It's about configuring hyperparameters. That is the corresponding objective function. In the experiment, It is usually assumed to be zero. Covariance function Define any two points x , The covariance between them reflects their output values and The correlation. Commonly used covariance functions include the squared exponential function, i.e.: in, It is the magnitude of the difference parameter control function. The length ratio parameter controls the relevant distance of the point in the input space.

[0039] When new observation data is introduced, the distribution of the objective function can be updated based on Bayes' theorem and combined with Gaussian processes. At any point... The posterior distribution predicted by the model is a Gaussian distribution, and its mean and variance are calculated using the following formulas: in, It's a new point. Its own covariance value, It is its covariance vector with all observed points. It is the covariance matrix of the observed points. It is the output vector of the observed points; T Indicates matrix transpose; I Represents the identity matrix; m Indicates Gaussian process in x i The predicted mean at the location; s 2 Variance represents the uncertainty in a prediction.

[0040] S6. Based on the elevation information of the spaceborne lidar footprint points and the airborne LiDAR point cloud dataset, the accuracy of the denoised photon data is evaluated.

[0041] The denoising algorithm accuracy was evaluated by comparing the denoised photon elevation information with the elevation information obtained from the airborne LiDAR point cloud dataset using recall R, precision P, and F scores. The evaluation results were used to guide the optimization of the ICESat-2 photon denoising algorithm. In the formula, TP represents the number of correctly identified signal photons, FP represents the number of noise photons misidentified as signal photons, and FN represents the number of signal photons that were not correctly identified. The process of this embodiment is as follows: Figure 1 As shown.

[0042] This embodiment introduces a Bayesian optimization algorithm to intelligently search for the optimal parameter combination of DBSCAN at each level, achieving the theoretically optimal denoising effect for each subset. This method constructs a well-defined and structurally clear search space for Bayesian optimization by using vegetation stratification, while Bayesian optimization ensures that the denoising potential of each level is fully utilized. This organic coupling not only effectively solves the problem of adapting global parameters in complex vegetation environments but also achieves a synergistic denoising efficiency of "1+1>2," significantly improving the processing accuracy and efficiency of ICESat-2 photon data under multi-layered vegetation conditions.

[0043] Example 2 This embodiment also provides an ICESat-2 adaptive photon denoising system based on vegetation stratification, including: an acquisition module, a processing module, a calculation module, an extraction module, a denoising module, and an evaluation module.

[0044] The following will describe in detail, with reference to this embodiment, how the present invention solves the technical problems in practical work.

[0045] First, the discrete footprint points of the ICESat-2 spaceborne lidar, Sentinel-2 optical data, and airborne LiDAR data are acquired using the acquisition module.

[0046] Acquire discrete footprint points of the ICESat-2 spaceborne lidar, Sentinel-2 optical data, and airborne LiDAR data; preprocess the discrete footprint point data, Sentinel-2 data, and airborne LiDAR data to obtain the geographical location and elevation information, Sentinel-2 surface reflectivity, and airborne LiDAR point cloud dataset corresponding to each spaceborne lidar footprint point.

[0047] The ICESat-2 spaceborne lidar, with its direct measurement advantage, can acquire key structural parameters such as forest canopy height. However, its discrete footprint data is often affected by noise interference from atmospheric scattering and surface clutter, making it difficult to directly meet the requirements of high-precision applications. Sentinel-2 optical data, on the other hand, has a wide coverage range and can provide horizontal distribution characteristics of surface vegetation (such as vegetation type and cover density). These horizontal characteristics can serve as important criteria for distinguishing signal photons from noise photons. Combining this data with ICESat-2 data for denoising is expected to reduce noise interference through multi-dimensional information complementarity, thereby improving the accuracy of lidar data. Meanwhile, high-precision airborne LiDAR data serves as a field reference, acquiring high-precision and reliable three-dimensional information on the surface and vegetation, which can be used to verify the accuracy of preprocessed data.

[0048] The processing module preprocesses the discrete footprint point data, Sentinel-2 data, and airborne LiDAR data to obtain the geographical location and elevation information, Sentinel-2 surface reflectance, and airborne LiDAR point cloud dataset corresponding to each spaceborne LiDAR footprint point.

[0049] Data quality checks were performed on discrete footprint data to remove outliers and extract the spaceborne lidar footprint points. Radiometric calibration, atmospheric correction, and topographic correction were performed on Sentinel-2 optical data, and Sentinel-2 surface reflectance was calculated. Georegistration, point cloud denoising, and point cloud classification were performed on airborne LiDAR data, and continuous digital elevation models (DEM) and digital surface models (DSM) were calculated. The coordinate systems of all data were unified, and the spaceborne lidar footprint points were spatially matched with the airborne LiDAR data and Sentinel-2 optical data to extract the latitude, longitude, and elevation information corresponding to each spaceborne lidar footprint point.

[0050] The calculation module calculates the spatial distribution data of NDVI based on the Sentinel-2 surface reflectance.

[0051] First, the spatial distribution data of NDVI is calculated based on Sentinel-2 image data. Then, the spatial distribution of surface vegetation cover (FVC) is calculated based on the spatial distribution data of NDVI using a pixel binary model. Finally, the correlation between the spatial distribution of FVC and photon coordinates is established by combining photon geographic location information, and the FVC information of the corresponding photon location is extracted. While the Normalized Difference Vegetation Index (NDVI) effectively characterizes the growth status and green cover of surface vegetation, making it a commonly used indicator for reflecting vegetation information, its ability to distinguish areas with high vegetation cover is relatively weak. Once vegetation cover reaches a certain level, the NDVI value tends to saturate, making it difficult to accurately reflect subtle differences in vegetation density within high-coverage areas. It is also susceptible to interference from factors such as soil background reflection and atmospheric aerosols, leading to biases in vegetation information representation. In contrast, Surface Vegetation Cover (FVC), calculated using a pixel-based binary model, decomposes pixels into vegetated and non-vegetated portions, directly quantifying the proportion of surface vegetation cover. This provides a more intuitive and accurate reflection of differences in vegetation distribution across different areas, especially in complex vegetation environments, where it more clearly depicts vegetation density gradients. Based on the accurate representation of vegetation information by FVC, subsequent vegetation stratification denoising based on FVC gradients allows for more targeted denoising of photon data from areas with different coverage levels, making high-precision denoising possible.

[0052] Based on red and near-infrared reflectance data from Sentinel-2 imagery, and according to the standard NDVI calculation model, continuous NDVI spatial distribution data within the study area were calculated. The expression is as follows: , In the formula, NIR is the near-infrared reflectance of Sentinel-2, and Red is the red reflectance of Sentinel-2.

[0053] Based on the NDVI spatial distribution data and the geographical location information of the spaceborne lidar footprint points, the extraction module extracts the FVC information of each photon location from the FVC spatial distribution.

[0054] By combining a pixel-based binary model, the spatial distribution of surface vegetation cover (FVC) is calculated, clarifying the vegetation cover degree at each spatial location within the study area. Photon geographic location information acquired by spaceborne lidar is introduced, and a spatial coordinate matching method is used to map the FVC spatial distribution map to photon coordinates, thereby accurately extracting the vegetation cover information for each photon's corresponding spatial location from the FVC spatial distribution map. The specific process includes: based on continuous NDVI spatial distribution data, combined with a pixel-based binary model, the core logic of which is: assuming each image pixel consists of a vegetation-covered portion and a non-vegetation-covered portion, the proportion of each portion in the pixel is calculated using the NDVI's full vegetation cover threshold and no vegetation cover threshold, thus obtaining the pixel's vegetation cover degree. Pixel-by-pixel calculations are performed, and by calculating the coverage of all pixels in the image area, an FVC spatial distribution map is generated, thereby clarifying the vegetation cover degree at each spatial location. The resulting FVC calculation results are as follows: Among them, the pixel value of pure vegetation partial coverage is from This indicates that the pixel value of the partially covered bare soil is determined by... This indicates that the pixel value at the 5% confidence level is taken as... The pixel value at the 95% confidence level is taken as... ,Will and The vegetation cover of the study area was calculated by substituting the values ​​into the formula.

[0055] The denoising module divides the photon data into several vegetation gradient levels based on the FVC information of each extracted photon location, and uses DBSCAN clustering to denoise the photon data of each vegetation gradient level.

[0056] Based on the obtained vegetation cover (FVC), the photon data is divided into different vegetation gradient levels. DBSCAN clustering is then performed on the photon data of each stratified interval. To improve the efficiency and accuracy of denoising, the optimal parameter combination in the DBSCAN clustering denoising process is found using the Bayesian optimization algorithm. Based on the optimal parameter combination, the clustering denoising of the photon data in each interval is completed, and finally, accurate denoising of the photon data is achieved.

[0057] DBSCAN, a classic density clustering denoising algorithm, is widely used in photon data denoising due to its ability to identify noise clusters of arbitrary shapes without requiring a preset number of clusters. It is suitable for distinguishing between discrete outliers caused by noise and real signal photons in ICESat-2 data. However, the denoising effect of this algorithm is highly dependent on the settings of parameters such as the neighborhood radius and the minimum number of neighboring points. Improper parameter values ​​can easily lead to incomplete noise removal or the accidental deletion of effective photons. Furthermore, traditional parameter tuning often relies on manual trial and error, which is not only inefficient but also makes it difficult to find the globally optimal parameter combination.

[0058] Bayesian optimization algorithms, by constructing a probabilistic model of the objective function, can dynamically adjust the search direction based on historical parameter tuning results, quickly locate the optimal parameter combination of DBSCAN, avoid the blindness of manual trial and error, and significantly shorten the parameter optimization time. This effectively solves the problem of balancing efficiency and accuracy in traditional denoising, ensuring efficient and accurate photon denoising in all vegetation gradient intervals, and meeting the requirements for denoising efficiency and accuracy in complex vegetation environments.

[0059] The DBSCAN algorithm is used to denoise photons under different vegetation gradient levels. The sample points of the DBSCAN algorithm are composed of core points, boundary points, and noise points. The definitions of these three types of sample points are as follows: For a given dataset D, if sample p -The neighborhood contains at least If there are n samples (including sample p), then sample p is called the core point, that is: .

[0060] Let p be the core point, where - Neighborhood The expression is: in, q Represents the dataset D Another sample point in the dataset is used to determine whether it is in the sample point. p Within the ε-neighborhood; D This represents the entire dataset, containing all the sample points to be analyzed. distance This represents the distance metric function.

[0061] For a sample b that is not a core point, if b lies at any core point p... - Within the neighborhood, sample b becomes the boundary point, that is: .

[0062] For a sample n that is not a core point, if n is not in any core point... - Within the neighborhood, sample n is called a noise point, that is: .

[0063] The optimal parameter combination in the DBSCAN clustering denoising process is found using a Bayesian optimization algorithm. Based on this optimal parameter combination, clustering denoising of photon data in each interval is completed, ultimately achieving accurate denoising of the photon data, including: Bayesian optimization utilizes Bayesian statistical principles to guide the search process, establishing a Gaussian process model to predict the impact of hyperparameter combinations on model performance and automatically selecting the next set of hyperparameters to be evaluated. This method uses a Gaussian process (GP) as the objective function, which is a stochastic process defined on a continuous input space; the joint distribution of any finite set of points is a multivariate Gaussian distribution. A GP is defined by a mean function... Sum of covariance functions Complete definition, that is: in, x It's about configuring hyperparameters. That is the corresponding objective function. In the experiment, It is usually assumed to be zero. Covariance function Define any two points x , The covariance between them reflects their output values and The correlation. Commonly used covariance functions include the squared exponential function, i.e.: in, It is the magnitude of the difference parameter control function. The length ratio parameter controls the relevant distance of the point in the input space.

[0064] When new observation data is introduced, the distribution of the objective function can be updated based on Bayes' theorem and combined with Gaussian processes. At any point... The posterior distribution predicted by the model is a Gaussian distribution, and its mean and variance are calculated using the following formulas: in, It's a new point. Its own covariance value, It is its covariance vector with all observed points. It is the covariance matrix of the observed points. It is the output vector of the observed points; T Indicates matrix transpose; I Represents the identity matrix; m Indicates Gaussian process in x i The predicted mean at the location; s 2 Variance represents the uncertainty in a prediction.

[0065] The evaluation module assesses the accuracy of the denoised photon data based on the elevation information of the spaceborne lidar footprint points and the airborne LiDAR point cloud dataset.

[0066] The denoising algorithm accuracy was evaluated by comparing the denoised photon elevation information with the elevation information obtained from the airborne LiDAR point cloud dataset using recall R, precision P, and F scores. The evaluation results were used to guide the optimization of the ICESat-2 photon denoising algorithm. In the formula, TP represents the number of signal photons that are correctly identified, FP represents the number of noise photons that are mistakenly identified as signal photons, and FN represents the number of signal photons that are not correctly identified.

[0067] Example 3 Figure 2 Comparison of ICESat-2 photon denoising accuracy under different methods. From the perspective of beam type, strong beams, due to their more accurate representation of vegetation information, show a greater advantage in denoising accuracy. In the Jixi Natural Forest Research Area (… Figure 2(a) After standard deviation denoising, DBSCAN denoising, OPTICS denoising, and optimized vegetation layer denoising, the F-scores of the strong beam, weak beam, and total beam in Jixi all showed an upward trend. Among them, the vegetation layer denoising algorithm achieved the best denoising accuracy, with the F-value of the strong beam in Jixi reaching 0.9616, the weak beam 0.9425, and the total beam 0.9578. The strong beam showed higher denoising accuracy, which proves the effectiveness of the method under natural forest conditions; in the Wangyedian artificial forest research area ( Figure 2 (b) The F-scores of different beams also showed an increasing trend with the denoising method. Under the vegetation stratification denoising algorithm, the F-score of the strong beam in Wangyedian was 0.9628, the weak beam was 0.8702, and the total beam was 0.9385. The strong beam also performed better, which further illustrates the effectiveness of the method under artificial forest conditions. Overall, the two study areas showed different forest type heterogeneity, but both demonstrated the advantage of the vegetation stratification optimization denoising method in improving denoising accuracy. Moreover, the denoising results of this invention are in good agreement with actual airborne LiDAR data, verifying the effectiveness of this invention in ICESat-2 photon denoising applications under different forest types. The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A vegetation-layer-based ICESat-2 adaptive photonic denoising method, characterized in that, Includes the following steps: S1. Acquire discrete footprint points of the ICESat-2 spaceborne lidar, Sentinel-2 optical data, and airborne LiDAR data; S2. Preprocess the discrete footprint point data, Sentinel-2 data and airborne LiDAR data to obtain the geographical location and elevation information, Sentinel-2 surface reflectance and airborne LiDAR point cloud dataset corresponding to each spaceborne LiDAR footprint point. S3. Based on the Sentinel-2 surface reflectance, the spatial distribution data of NDVI is calculated; S4. Based on the NDVI spatial distribution data and the geographical location information of the spaceborne lidar footprint points, extract the FVC information of each photon location from the FVC spatial distribution; S5. Based on the FVC information of each extracted photon location, the photon data is divided into several vegetation gradient levels, and the photon data of each vegetation gradient level is denoised using DBSCAN clustering. S6. Based on the elevation information of the spaceborne lidar footprint points and the airborne LiDAR point cloud dataset, evaluate the accuracy of the denoised photon data.

2. The ICESat-2 adaptive photon denoising method based on vegetation stratification according to claim 1, characterized in that, The preprocessing steps in step 2 include: Perform data quality checks on discrete footprint point data, remove outliers, and extract footprint points from the spaceborne lidar. Radiometric calibration, atmospheric correction, and topographic correction were performed on the Sentinel-2 optical data, and the Sentinel-2 surface reflectance was calculated. Geographic registration, point cloud denoising, and point cloud classification were performed on airborne LiDAR data to calculate continuous digital elevation model (DEM) and digital surface model (DSM). The coordinate system of all data is unified, and the spaceborne lidar footprint points are spatially matched with airborne LiDAR data and Sentinel-2 optical data to extract the latitude, longitude and elevation information corresponding to each spaceborne lidar footprint point.

3. The ICESat-2 adaptive photon denoising method based on vegetation stratification according to claim 1, characterized in that, The method for calculating the spatial distribution data of NDVI in step 3 includes: based on the red band reflectance data and near-infrared band reflectance data of Sentinel-2 imagery, and according to the standard calculation model of NDVI, the continuous spatial distribution data of NDVI in the study area is calculated.

4. The ICESat-2 adaptive photonic denoising method based on vegetation stratification according to claim 3, characterized in that, Using Sentinel-2 red band reflectance and near-infrared band reflectance as input variables, pixel-by-pixel calculations are performed according to the standard calculation model to generate spatially continuous NDVI spatial distribution data. The standard calculation model for NDVI is as follows: ， In the formula, NIR is the near-infrared reflectance of Sentinel-2, and Red is the red reflectance of Sentinel-2.

5. The ICESat-2 adaptive photonic denoising method based on vegetation stratification according to claim 1, characterized in that, The method for extracting FVC information for each photon location in step 4 includes: calculating the spatial distribution of surface vegetation cover (FVC) based on the NDVI spatial distribution data and combining it with a pixel bisection model; and extracting the FVC information for each photon location from the FVC spatial distribution based on the geographical location information of the spaceborne lidar footprint points.

6. The ICESat-2 adaptive photon denoising method based on vegetation stratification according to claim 1, characterized in that, Step 5 involves using DBSCAN clustering to denoise the photon data for each vegetation gradient level, including: The DBSCAN algorithm is used to denoise photons under different vegetation gradient levels. The DBSCAN algorithm defines sample points as core points, boundary points and noise points. For dataset D, if sample p ε -If a neighborhood contains MinPts samples, then sample p is called a core point; For a sample b that is not a core point, if b lies at any core point p... ε - Within the neighborhood, sample b is called a boundary point; For a sample n that is not a core point, if n is not in any core point... ε - Within the neighborhood, sample n is called a noise point.

7. The ICESat-2 adaptive photon denoising method based on vegetation stratification according to claim 1, characterized in that, In step S6, the method for evaluating the accuracy of the denoised photon data includes: The denoised photon elevation information and the elevation information obtained from the airborne LiDAR point cloud dataset were used to evaluate accuracy. Accuracy evaluation was achieved through recall (R), precision (P), and F scores, calculated using the following formula: In the formula, TP represents the number of signal photons that are correctly identified, FP represents the number of noise photons that are mistakenly identified as signal photons, and FN represents the number of signal photons that are not correctly identified.

8. A vegetation-stratified ICESat-2 adaptive photonic denoising system, the system being used to implement the method according to any one of claims 1-7, characterized in that, include: The module includes an acquisition module, a processing module, a calculation module, an extraction module, a noise reduction module, and an evaluation module.

Images