An Open Source DEM Accuracy Assessment Method Based on ICESat-2

Abstract

This invention discloses an open-source DEM accuracy assessment method based on ICESat-2, relating to surveying and mapping science and technology, and remote sensing data processing technology. The method includes: Step S1, acquiring laser point data and DEM data to be evaluated; Step S2, using high-confidence elevation control points as elevation benchmarks, calculating the root mean square error (RMSE) of the DEM to be evaluated; Step S3, constructing a comprehensive screening mechanism based on multi-dimensional attribute parameters and setting attribute indices; Step S4, designing a control point extraction strategy based on a multi-attribute joint isolated forest algorithm, including: basic screening based on reference ground elevation, preliminary screening of attribute indices, advanced screening of attribute indices, and isolated forest algorithm screening; using the elevation accuracy values ​​of the laser points after advanced screening as input features, constructing an isolated forest model for anomaly detection, removing laser points identified as abnormal by the algorithm, and obtaining high-confidence elevation control points.

Description

Technical Field

[0001] This invention relates to the fields of surveying and mapping science and technology, and remote sensing data processing technology, and in particular to an open-source DEM accuracy assessment method based on ICESat-2. Background Technology

[0002] Digital Elevation Models (DEMs), as fundamental spatial data in Earth sciences, play a crucial role in numerous fields, including geomorphology, topography, surveying, engineering construction, biodiversity, and geological hazards. Higher elevation accuracy of a DEM reflects more realistic topographic details, providing reliable data support for applications such as engineering design, geological hazard analysis, and vegetation erosion monitoring. Therefore, a systematic evaluation of the elevation accuracy of existing DEM products is of significant theoretical and practical importance for the scientific selection and reconstruction correction under different application contexts. There are indeed many existing DEM accuracy verifications. For example, Nahed Osama et al. verified the accuracy of open-source DEMs in the American highlands using filtered ATL08 product data, demonstrating that the strong beams of the ATL08 product can serve as an elevation reference for vegetated mountainous areas. Liu et al. used GPS to verify the elevation accuracy of four open-source DEMs in parts of China, analyzing the impact of longitude, latitude, and slope on DEM elevation accuracy. Liu et al. also used the ATL08 product to evaluate global DEMs at different resolutions. However, existing DEM elevation accuracy verification mainly focuses on small-scale, small-area applications, lacking research on DEM accuracy verification for large-scale, large-area applications, applications spanning latitude and longitude, and complex terrain conditions. Accuracy verification of global wide-area DEMs requires high-precision, reliable elevation control points, but obtaining a large number of high-precision control points presents significant challenges and high costs. With the rapid development of spaceborne laser altimetry technology, utilizing spaceborne laser data such as ICESat-2 provides a feasible path for DEM accuracy verification. ICESat-2 spaceborne laser altimetry data has advantages such as wide coverage, high accuracy, repeatable observations, and publicly available data, making it possible to use ICESat-2 data for large-area DEM elevation accuracy verification.

[0003] Because the photon-counting lidar on ICESat-2 is susceptible to interference from environmental factors such as atmospheric scattering, solar radiation, and instrument thermal noise during target detection, the acquired raw data contains a mixture of signal and noise photons, resulting in a high background noise rate. Therefore, it is necessary to perform refined denoising processing to improve the accuracy of ICESat-2 elevation products. Existing research has explored the accuracy of ICESat-2 elevation data and its influencing factors in depth. For example, Wang et al.'s research revealed the impact of signal-to-noise ratio, slope, vegetation elevation, and vegetation cover on the elevation accuracy of ICESat-2 (Ice, Cloud, and Land Elevation Satellite-2). Wang et al. proposed a photon screening process based on the attribute parameters of ICESat-2's ATL08 product and a reference DEM to evaluate the accuracy of ICESat-2 data. Zheng et al. proposed a processing flow that integrates quality inspection, reference DEM screening, and reference attribute screening to improve the applicability of ATL08 products under different terrain coverage conditions and complexity. Shang et al. proposed a control point extraction strategy based on the attribute parameters of ATL08 and the accuracy and location requirements of control points. Wang M et al. established a global laser point control database containing 560 million points using ATL08-level data and verified that the elevation accuracy of 90% of the laser points was better than 0.7m. However, while the above method has made progress in improving laser point accuracy, its filtering effect still has limitations in areas with complex terrain and diverse land cover types. Furthermore, the reliability of the filtered ICESat-2 data for evaluating ASTER GDEM, RTCDEM, SRTM, and TanDEM-X has not yet been confirmed. Summary of the Invention

[0004] This invention provides an open-source DEM accuracy assessment method based on ICESat-2 to solve the problem of difficulty in obtaining high-precision control points in large-scale DEM accuracy verification. It can robustly and accurately extract high-reliability elevation control points from ICESat-2 data.

[0005] To solve the above-mentioned technical problems, the technical solution proposed by this invention is as follows:

[0006] An open-source DEM accuracy assessment method based on ICESat-2 includes the following steps:

[0007] Step S1, Data Acquisition: Acquire ICESat-2 laser point data and the digital elevation model (DEM) data to be evaluated;

[0008] Step S2, set DEM evaluation value and accuracy index: use high-confidence elevation control points as elevation benchmarks, calculate the root mean square error (RMSE) of the DEM to be evaluated, and evaluate the elevation accuracy of the DEM.

[0009] Step S3: Based on multi-dimensional attribute parameters, set attribute indicators to construct a comprehensive screening mechanism;

[0010] Step S4: Design a control point extraction strategy based on the multi-attribute joint isolated forest algorithm, including: basic screening based on reference ground elevation, preliminary screening of attribute indicators, advanced screening of attribute indicators, and screening using the isolated forest algorithm.

[0011] Using the elevation accuracy values ​​of the laser points after advanced screening as input features, an isolated forest model is constructed for anomaly detection. Laser points identified as abnormal by the algorithm are removed, and finally, highly reliable elevation control points are obtained.

[0012] A further improvement to the above technical solution is as follows:

[0013] Preferably, in step S1, at least 10 regions spanning latitude and longitude are selected, with significant differences in regional geoscientific environment requirements, and including at least four major landforms: plateau, hills, basins, and plains.

[0014] Preferably, in step S2, the evaluation index G_RMSE is introduced to quantify the degree of agreement between the DEM accuracy evaluation based on laser points and GPS measured points. The calculation formula is as follows:

[0015]

[0016] in, and These represent the RMSE (Resolution of Emptiness) measured with respect to the DEM data when using ICESat-2 laser points and GPS control points as elevation references, respectively.

[0017] Preferably, the attribute indicators in step S3 include segmented elevation information, uncertainty parameters, ground parameter markers, cloud cover markers, topographic indicators, photon quantity parameters, photon distribution quality parameters, and signal-to-noise ratio, wherein:

[0018] The segmented elevation information includes the maximum ground elevation, minimum ground elevation, average ground elevation, median ground elevation, interpolated ground surface elevation, best-fit ground elevation, and reference ground elevation.

[0019] The uncertainty parameters include pointing angle uncertainty, radial trajectory uncertainty, local slope uncertainty, timing uncertainty, geographic positioning uncertainty, and scattering uncertainty;

[0020] The ground parameter markings include ground slope and ground elevation skewness;

[0021] The photon quantity parameters include ground photon count, canopy photon count, and top canopy photon count.

[0022] Preferably, the basic screening based on the reference ground elevation in step S4 is as follows: a threshold is preset, the interpolated ground surface elevation of the ICESat-2 laser point is compared with the reference ground elevation, and laser points whose absolute difference exceeds the preset threshold are removed.

[0023] Preferably, the preliminary screening of attribute indicators in step S4 specifically includes:

[0024] The interpolated ground surface elevation is compared with the ground median elevation. A deviation threshold is set, and laser points that exceed the deviation threshold are removed.

[0025] The difference between the median ground elevation and the reference ground elevation is used for filtering. A threshold is set to remove laser points whose absolute value of the difference between the median ground elevation and the reference ground elevation exceeds the threshold range.

[0026] By comparing the reference ground elevation with the best-fit ground elevation, laser points with an absolute difference exceeding three times the absolute elevation accuracy are eliminated.

[0027] Geometric constraints are applied using ground slope indices, and slope thresholds are set to eliminate laser points whose ground slope indices exceed the slope thresholds.

[0028] Data is filtered using the number of surface photons, and a threshold is set to retain laser points whose surface photon count exceeds the threshold.

[0029] Data is filtered using the surface photon rate, and a threshold is set to remove laser points with a surface photon rate lower than the threshold.

[0030] Preferably, the advanced filtering of attribute indicators in step S4 specifically includes:

[0031] Data filtering was performed using uncertainty parameters and ground elevation skewness. A threshold was set to remove laser points whose ground elevation skewness was greater than three times the RMSE value.

[0032] Data is filtered using photon distribution quality parameters, retaining laser points whose photon distribution quality parameters are greater than -1 and whose three middle flag bits are equal to 1.

[0033] Cloud cover markers are introduced for data filtering, and laser points with cloud cover markers greater than 3 are removed;

[0034] Based on land cover data, laser points located in water bodies are removed;

[0035] Remove laser points with a signal-to-noise ratio less than 1;

[0036] Remove laser points whose ground index is greater than twice the RMSE;

[0037] The standard deviation of the laser point elevation accuracy is calculated using the Laida criterion, with 3 times the standard deviation as the final threshold, and laser points exceeding the final threshold are removed.

[0038] Preferably, the method of constructing an isolated forest model for anomaly detection includes: randomly selecting subsamples from the input elevation accuracy training data to construct multiple isolated binary trees; during the construction of tree nodes, randomly selecting a cutting point to recursively divide the data space until the termination condition is met; calculating the path length of each laser point in all isolated trees and obtaining an anomaly score accordingly; and identifying laser points with anomaly scores exceeding a preset threshold as anomalies and removing them.

[0039] The open-source DEM accuracy assessment method based on ICESat-2 provided by this invention has the following advantages compared with existing technologies:

[0040] (1) The open-source DEM accuracy assessment method based on ICESat-2 of the present invention utilizes ICESat-2 satellite laser data with wide coverage and open data, and replaces the traditional high cost and difficult-to-obtain GPS measurement points with a systematic and refined screening strategy, providing a feasible technical solution for large-scale DEM accuracy verification.

[0041] (2) The open-source DEM accuracy assessment method based on ICESat-2 of the present invention introduces key parameters such as ground standard deviation into the screening process for the first time. Through the progressive steps of multi-attribute joint constraints and isolated forest anomaly detection, low-quality photon points affected by factors such as noise, cloud layer, ground slope, and vegetation are eliminated to the greatest extent, so that the accuracy of the finally extracted elevation control points reaches the sub-meter level. Attached Figure Description

[0042] Figure 1 This is a spatial distribution map of GPS control points in the area used for experimental verification of this invention, where (a) is Zhangye; (b) is Weinan; (c) is Fangshan; (d) is Beichen; (e) is Chengdu; (f) is Yantai 1; (g) is Yantai 2; (h) is Zhengzhou; (i) is Wuhan; (j) is Hefei; (k) is Nanjing; (l) is Pudong; (m) is Zhanjiang; (n) is Wenchang; (o) is Ganzhou; and (p) is Ningbo.

[0043] Figure 2 This is a spatial distribution map of laser points within the experimental verification area of ​​this invention, where (a) is Zhangye; (b) is Weinan; (c) is Fangshan; (d) is Beichen; (e) is Chengdu; (f) is Yantai 1; (g) is Yantai 2; (h) is Zhengzhou; (i) is Wuhan; (j) is Hefei; (k) is Nanjing; (l) is Pudong; (m) is Zhanjiang; (n) is Wenchang; (o) is Ganzhou; and (p) is Ningbo.

[0044] Figure 3This is a flowchart illustrating the overall process of the elevation control point extraction strategy of this invention.

[0045] Figure 4 This is a statistical chart of the uncertainty experimental results in the experimental verification of this invention, where (a) is Zhangye; (b) is Weinan; (c) is Fangshan; (d) is Beichen; (e) is Chengdu; (f) is Yantai 1; (g) is Yantai 2; (h) is Zhengzhou; (i) is Wuhan; (j) is Hefei; (k) is Nanjing; (l) is Pudong; (m) is Zhanjiang; (n) is Wenchang; (o) is Ganzhou; and (p) is Ningbo.

[0046] Figure 5 The following are the evolution trends of the average G_RMSE of different DEM products during the screening process in the experimental verification of this invention, where (a) is the ASTER result; (b) is the RTC DEM result; (c) is the SRTM result; and (d) is the TanDEM-X result.

[0047] Figure 6 The figures show the interval distribution statistics of the results of four sets of digital elevation models in sixteen research areas after screening in the experimental verification of this invention. Among them, (a) is the interval distribution statistics of ASTER, (b) is the interval distribution statistics of RTC DEM, (c) is the interval distribution statistics of SRTM, and (d) is the interval distribution statistics of TanDEM-X.

[0048] Figure 7 The high-precision, high-spatial-resolution ground verification results for this invention are shown in the following figures: (a) DSM (Radio Signal Management System) of the Baiyangping UAV; (b) DSM of the Xinweijiang UAV; (c) DSM of the Bajiaolong UAV; (d) Spatial distribution of GPS control points (red triangles) in the Shaoyang experimental area; (e) GPS measurement scene, corresponding to the position of the red pentagram in figure (d); (f) Geographical location of the UAV DSM and GPS experimental area in Hunan Province; (g) Statistical histogram of UAV DSM accuracy verification; (h) Statistical histogram of GPS accuracy verification.

[0049] Figure 8 The following is an experimental verification of the land cover type and spatial distribution of laser points in the Wuhan and Weinan study areas before and after screening: (a) Wuhan before screening; (b) Wuhan after screening; (c) Weinan before screening; (d) Weinan after screening. Detailed Implementation

[0050] The following provides a detailed description of specific embodiments of the present invention. It should be understood that the specific embodiments described herein are for illustrative and explanatory purposes only and are not intended to limit the scope of the invention.

[0051] The open-source DEM accuracy assessment method based on ICESat-2 of this invention specifically includes the following steps:

[0052] Step S1, data acquisition.

[0053] At least 10 regions spanning a wide range of latitude and longitude were selected, with significant differences in regional geoscientific environmental requirements, including at least four major landforms: plateau, hills, basins, and plains.

[0054] This invention uses high-precision GPS control points as reference data to evaluate the elevation accuracy of open-source DEMs. By utilizing real-time network motion technology, a continuously operating GPS reference station is created. The elevation and horizontal accuracy of the control points obtained from the GPS reference station are better than 0.1m, which can provide reliable ground truth values ​​for large-scale DEM accuracy evaluation.

[0055] Step S2: Set the DEM evaluation value and accuracy index.

[0056] This invention uses both GPS field measurement points and laser point data from the ICESat-2 product as elevation benchmarks to comprehensively evaluate the accuracy of an open-source DEM. Let the elevation value of a certain reference point be... , The extracted value of this point in the DEM is then defined as the DEM accuracy of that point. The formula is:

[0057]

[0058] in, The closer a value is to 0, the smaller the deviation between the DEM elevation value and the benchmark value, and the higher the elevation accuracy.

[0059] This invention uses evaluation indicators As a measure of accuracy, let's assume a region contains N verification points, and its root mean square error is... Standard deviation )and The calculation formulas are as follows:

[0060]

[0061]

[0062]

[0063]

[0064] in, and These represent the RMSE (Relative Emergent Sequence) measured with respect to DEM data when using ICESat-2 laser points and GPS control points as elevation benchmarks, respectively. This invention introduces the following indicators: To measure the difference between the two, it is defined as follows: and difference, The smaller the absolute value, the better the evaluation results based on laser points match the results based on GPS measured points, thus demonstrating the reliability and accuracy of using laser points as an open-source DEM elevation evaluation benchmark.

[0065] Step S3: By combining the constraints of multi-dimensional attribute parameters, a comprehensive screening mechanism is constructed to extract highly reliable elevation control points.

[0066] The key attribute metrics selected are as follows:

[0067] (1) Segmented elevation information: Segmented elevation information mainly includes reference ground elevation and surface elevation values. Specifically, it includes: maximum ground elevation (h_te_max), minimum ground elevation (h_te_min), average ground elevation (h_te_mean), median ground elevation (h_te_median), interpolated ground surface elevation (h_te_interp), best-fit ground elevation (h_te_best_fit), and reference ground elevation (dem_h).

[0068] The reference ground elevation is a built-in attribute field of ICESat-2; this parameter is obtained by extracting the best DEM corresponding to the geographical location of each data segment, and its elevation value is taken from various global and regional elevation datasets, including: MERIT, GIMP, GMTED and MSS.

[0069] (2) Uncertainty parameter (h_te_uncertainty): h_te_uncertainty represents the uncertainty of the measured elevation of a point on the Earth's surface, including pointing angle uncertainty, radial trajectory uncertainty, local slope uncertainty, time synchronization uncertainty, geographic positioning uncertainty and scattering uncertainty.

[0070] (3) Ground parameter labels: attribute parameters used to describe the terrain and landform, mainly including ground slope (terrain_slope) and ground elevation skew (h_te_skew).

[0071] (4) Cloud_flag_atm: When the ATLAS laser beam passes through clouds or atmospheric aerosols, it causes a scattering effect, resulting in a reduction in the number of photons echoed from the ground and causing ranging errors. The cloud_flag_atm index records the number of cloud or aerosol layers identified in each 25Hz atmospheric profile (the value ranges from 0 to 10). A cloud_flag_atm greater than 0 indicates that clouds or aerosols may be present.

[0072] (5) Topographic index (h_te_std): h_te_std represents the standard deviation of the topographic points around the ground interpolated along the track segment, and is a key indicator of surface roughness. By setting a reasonable threshold for the topographic index, outliers caused by drastic topographic fluctuations can be effectively eliminated, thereby improving the accuracy of topographic interpolation.

[0073] (6) Photon quantity parameters: including ground photon count (n_te_photons), canopy photon count (n_ca_photons), and top canopy photon count (n_toc_photons). Ground photon count describes the number of laser points reaching the ground; the higher the ground photon count, the more accurate the ground elevation. Surface photon ratio (ratio_te_photons) is defined as the ratio of ground photon count to the total number of signal photons; the larger this ratio, the better the laser penetration and the higher the elevation accuracy. The formula for calculating surface photon ratio is as follows:

[0074]

[0075] (7) Photon distribution quality parameters: Due to factors such as cloud cover or signal attenuation, the distribution of photons in ground echoes may be uneven. The subset_te_flag is used to characterize the spatial quality distribution of topographic photons identified within each 100-meter road segment. This indicator reflects the uniformity of surface photons and is an important basis for screening high-quality laser measurement areas.

[0076] (8) Signal-to-noise ratio (SNR): The SNR value is defined as the ratio of the number of signal photons identified by the DRAGANN (Density-Removing Adaptive Gaussian Nearest Neighbor) algorithm to the number of background noise photons within the segment processed by the ICESat-2 algorithm. This indicator directly reflects the degree of interference to the signal. By setting a reasonable SNR threshold, abnormal data with a large proportion of noise can be effectively removed, thereby improving the extraction ratio of useful signals.

[0077] Step S4: This invention designs a control point extraction strategy based on a multi-attribute joint isolated forest algorithm. This strategy consists of the following four key steps: basic screening based on reference ground elevation, preliminary screening of attribute indicators, advanced screening of attribute indicators, and isolated forest algorithm screening.

[0078] S4-1, based on reference ground elevation for basic screening.

[0079] The interpolated ground surface elevation (h_te_interp) of the ICESat-2 laser points is compared with the reference ground elevation. If the absolute value of the difference, |h_te_interp-dem_h|, exceeds a set threshold, it is identified as a gross error and discarded. This invention selects laser points with |h_te_interp-dem_h| values ​​greater than the threshold of 30m as error points for rejection.

[0080] S4-2, Preliminary screening of attribute indicators.

[0081] After basic screening using reference ground elevation, this invention further introduces multi-dimensional attribute indicators for refined screening of laser points, specifically:

[0082] S4-2-1 compares the interpolated ground surface elevation with the ground median elevation, sets the deviation threshold to 2 m, and judges all laser points exceeding the deviation threshold as gross errors and removes them.

[0083] S4-2-2, the difference between the median ground elevation and the reference ground elevation is h_dif_ref. Using the h_dif_ref parameter, a larger |h_dif_ref| value indicates a greater deviation between the median ground elevation and the reference ground elevation, signifying lower observation accuracy of the original ATL08 laser point. When |h_dif_ref| exceeds 25m, the RMSE of the laser point increases sharply. Therefore, this invention considers laser points within the range of |h_dif_ref|>25m as abnormal errors and discards them.

[0084] S4-2-3, by comparing the reference ground elevation with the best-fit ground elevation, the best-fit ground elevation (h_te_best_fit) has extremely high elevation accuracy and is suitable as a benchmark reference for segmented ground. The filtering effect of the difference between the reference ground elevation and the best-fit ground elevation on ICESat-2 increases as the threshold decreases. Therefore, this invention selects and removes laser points whose absolute elevation accuracy exceeds three times that of |reference ground elevation - best-fit ground elevation|.

[0085] S4-2-4 utilizes terrain slope for geometric constraints. The elevation accuracy of the laser points decreases significantly with increasing slope. In flat areas (slope < 2°), ICESat-2 exhibits good elevation accuracy. The slope is calculated based on trigonometric transformations, and the corresponding terrain slope index is approximately 0.03. In complex areas, this can be relaxed to 0.05. Considering the complex topographic features of the area, to ensure robustness, this invention uniformly sets the slope threshold to 0.05. Laser points with a terrain slope index > 0.05 are discarded as error points.

[0086] S4-2-5 uses the number of ground photons (n_te_photons) for data filtering. The number of ground photons is positively correlated with the elevation accuracy. The elevation accuracy slowly increases as the number of ground photons increases. Laser points with n_te_photons>50 are retained.

[0087] S4-2-6. Data filtering is performed using surface photon rate (ratio_te_photons). When the surface photon rate is greater than 0.6, the RMSE (Recovery Mean Squared Error) decreases slowly as the surface photon rate increases. Laser points with a surface photon rate < 0.6 are discarded as error points.

[0088] S4-3, Advanced filtering of attribute metrics.

[0089] After completing the initial screening, a more refined screening strategy will be implemented.

[0090] S4-3-1 uses the uncertainty parameter (h_te_uncertainty) and ground elevation skewness (h_te_skew) for data filtering. First, twice the RMSE of the uncertainty parameter is used as the threshold. If the uncertainty parameter is greater than twice the RMSE, the corresponding laser point is removed as an error point. Then, three times the RMSE of the ground elevation skewness is calculated. If the value of the ground elevation skewness is greater than three times the RMSE value, the coordinates of the corresponding laser point are removed as an error point.

[0091] S4-3-2, Data filtering is performed using the photon distribution quality parameter (subset_te_flag), which reflects whether photons are evenly distributed on the Earth's surface. Therefore, this flag can be used to filter areas where laser beam measurements are more accurate. This invention selects data from the five flag bits of the photon distribution quality parameter that satisfy a value greater than -1 and have the middle three flag bits equal to 1.

[0092] S4-3-3, considering that clouds and aerosols may cause backscattering effects that affect the quantity and quality of photon echoes, this invention introduces cloud amount flags (cloud_flag_atm) for data filtering. This invention limits the cloud amount flag threshold to below 3, that is, only retains data with cloud amount less than or equal to 20%, so as to minimize the impact of the atmospheric environment on measurement accuracy.

[0093] S4-3-4, considering the influence of different land cover types on photon echo signals, the introduced error is usually greater than the system differences between different DEMs. This invention introduces 10-resolution global land cover data based on the European Space Agency (ESA) Sentinel-2 satellite, and uses its high-precision 9-category land cover classification map to constrain the study area, eliminating laser points located in water bodies.

[0094] In section S4-3-5, to suppress the negative impact of background noise on signal photon extraction, the signal-to-noise ratio (SNR) is used as an evaluation index. When SNR < 1, it indicates that the intensity of the background noise has exceeded the intensity of the reflected signal, and it is highly likely to be a "false point". Therefore, this invention can effectively improve the reliability of laser points by eliminating laser points with SNR < 1.

[0095] S4-3-6. Considering the degree of deviation of interpolated elevation under complex terrain, the terrain index (h_te_std) is selected for filtering. By calculating the RMSE of the remaining data, twice the RMSE is set as the threshold. If the terrain index is greater than the threshold, it is regarded as an outlier with a large interpolation error and is removed.

[0096] S4-3-7 To further enhance data consistency, based on the aforementioned screening steps, the standard deviation (STD) of the laser point elevation accuracy is calculated. Using the Raida criterion, 3 times the STD is set as the final threshold to remove outliers that exceed this range, ensuring the statistical significance of the data.

[0097] S4-4, Deep optimization based on the Isolation Forest algorithm.

[0098] After completing the screening based on physical and attribute indicators, in order to further eliminate residual nonlinear outliers, this invention introduces the isolated forest algorithm to perform final filtering on the elevation accuracy values, aiming to achieve a deep improvement in data accuracy through an anomaly detection mechanism.

[0099] To balance the accuracy of laser point elevation with the sample retention rate, this invention sets an anomaly score threshold. Below this threshold, the algorithm can selectively remove laser points with significantly shorter path lengths (i.e., abnormal elevation distributions), thereby reducing G_RMSE while maximizing the retention of more laser points. The isolated forest is composed of multiple isolated binary trees, and the construction process of a single isolated binary tree is as follows:

[0100] (a) Subsample initialization: 256 laser points are randomly selected from the elevation accuracy training dataset as subsamples and placed at the root node of the isolated tree.

[0101] (b) Generation of the dividing hyperplane: Randomly generate a cutting point p between the maximum and minimum values ​​of the elevation value dimension (specified dimension) of the current node data.

[0102] (c) Recursive partitioning of space: Use the cutting point to generate a hyperplane to divide the data space of the current node into two subspaces: sample points less than p are assigned to the left branch, and sample points greater than or equal to p are assigned to the right branch.

[0103] (d) Iteration and Termination: Recursively execute steps (b) and (c) on the left and right branch nodes to continuously build new child nodes until the termination condition is met (the leaf node contains only a single data, or the height of the tree reaches the preset value).

[0104] By screening key aspects, the impact of factors such as clouds and aerosols, ground slope, and water reflection on the evaluation results can be eliminated to the greatest extent, ensuring the reliability of laser points in terms of elevation accuracy and spatial distribution.

[0105] Experimental verification:

[0106] To ensure the generalizability of the research findings, 16 regions spanning a wide range of latitudes and longitudes were selected as experimental areas. These regions exhibit significant differences in their geological environments, encompassing four main landform types: plateaus, hills, basins, and plains. Specifically: Chengdu and Ganzhou represent basin topography; Hefei, Yantai 1, Yantai 2, and Zhengzhou represent hilly topography; Zhangye is located in a typical plateau region; and the plains areas selected include Ningbo, Nanjing, Pudong, Weinan, Zhanjiang, Wenchang, Fangshan, Beichen, and Wuhan. Detailed topographical classification characteristics and the number of laser points in each corresponding region are shown in Table 1.

[0107] Table 1. Topographic types and number of ICESat-2 laser points in the sixteen study areas.

[0108]

[0109] Four global digital elevation model (DEM) data were used: ASTER GDEM, ALOS PALSAR RTCDEM, SRTM DEM, and TanDEM-X. Their basic parameters and information are as follows: (1) ASTER GDEM (Advanced Spaceborne Thermal Emission and Reflection Radiometer Global DEM): This product is based on optical stereo image pairs acquired by the ASTER sensor and is produced using digital photogrammetry technology, covering the global land area between 83°N and 83°S. The experiment used ASTER GDEM v3 version, with a spatial resolution of 30 meters, a horizontal datum of WGS84 coordinate system, and a vertical datum of EGM96 geoid.

[0110] (2) ALOS PALSAR RTC DEM (ALOS PALSAR Radiometric Terrain Corrected DEM): This data was generated by the Alaska Satellite Facility (ASF) using PALSAR sensor data from the ALOS-1 satellite, and refined through synthetic aperture radar (SAR) imagery with a spatial resolution of 12.5 meters. Its coverage area is the land between 60°N and 57°S (excluding Antarctica and high-latitude regions of Eurasia), and both the horizontal and vertical references use the WGS84 coordinate system.

[0111] (3) SRTM DEM (Shuttle Radar Topography Mission DEM): This data comes from the Space Shuttle Radar Topography Mission and is acquired using airborne interferometric synthetic aperture radar (InSAR) technology, covering approximately 80% of the global landmass (60°N~56°S). The experiment used the SRTM1 version with a spatial resolution of 30 meters, a horizontal reference of the WGS84 coordinate system, and a vertical reference of the EGM96 geoid.

[0112] (4) TanDEM-X (TerraSAR-X for Digital Elevation Measurement): This data was generated by the German Aerospace Center (DLR) using two-satellite interferometric synthetic aperture radar (InSAR) technology, creating a global elevation model. It covers a range from 90°N to 85°S. The TanDEM-30 dataset was selected, with a spatial resolution of 30 meters. Both the horizontal and vertical references use the WGS84 coordinate system (the vertical reference is the WGS84 ellipsoidal height).

[0113] High-precision GPS control points were used as reference data to verify the elevation accuracy of four open-source DEMs. The experiment utilized real-time network motion technology to create continuously operating GPS reference stations. The elevation and horizontal accuracy of the control points obtained from these GPS reference stations were better than 0.1m, providing reliable ground truth values ​​for large-scale DEM accuracy assessment.

[0114] The number of GPS control points and their geographical latitude and longitude ranges in each experimental area are detailed in Table 2, and their spatial distribution within the study area is as follows: Figure 1 As shown.

[0115] Table 2. Number and latitude / longitude range of GPS control points within the experimental area.

[0116]

[0117] The Ice, Cloud, and Land Elevation Satellite-2 (ICESat-2), currently in orbit, operates in a near-polar orbit at an altitude of approximately 500 km and an inclination of 92°, with an orbital period of 91 days. Thanks to its onboard photon counting system, ICESat-2 can conduct high-frequency continuous data acquisition along its orbital direction. Its measurement range is extensive, covering the global landmass between 88°N and 88°S. The core sensor onboard ICESat-2, ATLAS (Advanced Terrain Laser Altimeter), emits six laser beams, divided into three groups. Each group contains one strong beam and one weak beam (the strong and weak beams are spaced 90m apart), with an energy ratio of 4:1. The three groups of laser beams are approximately 3.3 km apart across the orbital direction. ICESat-2 digital products are categorized into levels 0 to 3 (ATL00 to ATL21) based on their application. The ATL03 product records the latitude, longitude, and longitude of all photons detected by the ATLAS sensor, as well as the WGS84 reference ellipsoidal altitude. By classifying and processing the raw photons of ATL03, the generated ATL08 product provides topographic elevation, canopy elevation, and other parameters describing the elevation of land and vegetation in segments every 100 m along the orbital direction.

[0118] Although both ATL03 and ATL08 products possess precise spatial location information and can theoretically serve as the foundational data for ICESat-2 high-precision elevation control point extraction, they differ significantly in data structure and information density. ATL03 contains a large number of raw photons (including signal photons and solar background noise), resulting in a significantly larger data volume compared to ATL08 for the same region. Since this experiment requires large-scale data extraction across 16 experimental areas with significant topographical differences, directly using ATL03 data would lead to problems such as read latency and low processing efficiency due to its excessive data volume. In contrast, ATL08 not only has a more refined data structure but also provides rich geoscientific descriptive parameters such as surface roughness, slope, canopy elevation, and canopy cover, providing crucial attribute support for the control point extraction strategy. Balancing processing efficiency with the need to utilize multi-dimensional features, this experiment ultimately selected ATL08 data as the core foundational data.

[0119] The distribution of laser points in each experimental area is as follows: Figure 2 As shown.

[0120] Given that ASTER and SRTM use the EGM96 geoid as their vertical reference, while RTC DEM, TanDEM-X, and ICESat-2 ATL08 products use the WGS84 ellipsoidal height as their vertical reference, and WGS84 is used as the reference, the elevations of ASTER and SRTM are converted from the EGM96 reference to the WGS84 ellipsoidal height using the geoid model, thus enabling the comparison of all elevation models and laser points under the same spatial reference.

[0121] Data from 16 experimental areas were subjected to basic screening based on reference ground elevation (dem_h), preliminary screening based on attribute indicators, advanced screening based on attribute indicators, and screening using the Isolation Forest algorithm.

[0122] (a) Verify the correctness of the selected attribute indicators.

[0123] (1) This invention utilizes topographic indices (h_te_std) to optimize and filter ICESat-2 laser point data. To evaluate its effectiveness, experiments were conducted in 16 study areas. The effect of topographic indices on the filtering of laser point data was analyzed by comparing the G_RMSE before filtering with the SG_RMSE after filtering.

[0124] The specific experimental procedure is as follows: First, invalid values ​​in the terrain indicators (such as 3.4028235E+38) were identified as outliers and removed. Such data usually reflects insufficient terrain photon number density. Then, twice the RMSE of the terrain indicators was selected as the discrimination threshold. If the terrain indicator value of a laser point exceeded this threshold, it was determined that the point had a large interpolation elevation difference and was considered an error point, thus being removed. Finally, using high-precision GPS measured points as a benchmark, the accuracy of the filtered laser point data was checked, and the final SG_RMSE was calculated. The experimental results are shown in Table 3.

[0125] Table 3. Statistics of filtering results based on interpolated ground topography standard deviation

[0126]

[0127] The following conclusions can be drawn from the data analysis in Table 3:

[0128] ASTER accuracy improved: In most regions, the G_RMSE of ASTER data decreased, proving that the h_te_std-based filtering strategy effectively eliminated laser points with significant errors. For example, the G_RMSE in Zhanjiang decreased from 2.93m to 0.99m, and in Nanjing from 2.08m to 1.54m, showing significant improvement. However, the G_RMSE in areas such as Wenchang actually increased, indicating that this method still has certain limitations for ASTER products.

[0129] Applicability of Multi-Source DEMs: For TanDEM-X, RTC DEM, and SRTM products, the filtered G_RMSE showed a significant reduction in most areas. This indicates that using topographic indicators as filtering parameters has high reliability and wide applicability in improving the laser point elevation accuracy of ATL08 products.

[0130] (2) In order to evaluate the screening effectiveness of uncertainty parameters on laser points, a series of sampling thresholds were set in the range of 0.25 to 2 with a step size of 0.25 to optimize the screening of ATL08 laser point data.

[0131] The specific experimental steps are as follows: First, identify and remove invalid values ​​from the uncertainty parameters (such as 3.4×). These were considered abnormal error points and removed. Then, twice the RMSE of the uncertainty parameter was used as the judgment threshold. If the uncertainty parameter of a laser point exceeded this threshold, the point was deemed not to meet the accuracy requirements and was removed as an error point. Finally, high-precision GPS measured points were used as a benchmark to verify the accuracy of the filtered laser point data. The G_RMSE of the four DEM products was finally calculated, and the statistical results are as follows: Figure 4 As shown.

[0132] according to Figure 4 Statistical analysis shows that, after optimization and screening of uncertainty parameters, the accuracy performance of each DEM product is as follows:

[0133] Multi-source DEM accuracy performance: For ASTER products, G_RMSE values ​​decreased in 50% of areas. For RTC DEM and SRTM products, G_RMSE showed a decreasing trend in most areas. Although G_RMSE did not decrease in Chengdu and Weinan, the increased G_RMSE remained within 0.38 m / 0.2 m (RTC DEM) and 0.36 m / 0.23 m (SRTM), respectively, indicating that high vertical accuracy was maintained in these areas. In contrast, TanDEM-X products showed a decrease in G_RMSE in all experimental areas.

[0134] Screening effectiveness evaluation: Experimental results show that the uncertainty parameter can significantly eliminate laser points with obvious deviations in the laser point data of the ATL08 product, demonstrating good universality. However, this indicator has relatively limited effectiveness in screening ASTER products and has certain limitations.

[0135] Threshold sensitivity analysis: When the threshold is set to 0.25, although G_RMSE reaches its minimum in some areas, the overly strict screening conditions lead to a sharp decrease in the number of laser points, with the rejection rate exceeding 95% in some areas. The severe shortage of samples induces drastic fluctuations in G_RMSE; for example, the G_RMSE of ASTER products in Weinan and Ganzhou surged to 14.08m and 3.98m, respectively. When the threshold is less than 1, G_RMSE exhibits significant numerical fluctuations due to the limited size of the remaining laser point samples. When the threshold is in the range [1,2], G_RMSE shows a slow decreasing trend as the threshold decreases, but the number of laser points screened out is relatively large as the threshold decreases. Considering both the number of laser points retained and the degree of G_RMSE optimization, a threshold of 2 is selected as the optimal threshold. This threshold effectively reduces G_RMSE.

[0136] (ii) Verify the screening effect of the Isolation Forest algorithm.

[0137] The abnormal score threshold was set to 0.7. A comparative analysis was conducted on the G_RMSE before screening and the SG_RMSE after screening (G_RMSE after screening), and the experimental statistical results are shown in Table 4.

[0138] Table 4 Statistical data of filtering results based on the Isolation Forest algorithm

[0139]

[0140] According to the statistical results in Table 4, after optimization and screening using the Isolation Forest algorithm, the accuracy performance of each DEM product is as follows:

[0141] For ASTER products, the G_RMSE value decreased in nearly 50% of areas, demonstrating that the screening strategy effectively eliminated some laser points with significant elevation deviations. For example, the G_RMSE in Nanjing was optimized from 2.08m to 0.79m. However, the G_RMSE in areas such as Wenchang actually increased after screening, reflecting the limitations of this method in verifying the accuracy of ASTER products due to the data quality and terrain complexity of ASTER itself. In contrast, the G_RMSE of TanDEM-X, RTC DEM, and SRTM products showed a significant decrease in most areas. The G_RMSE of TanDEM-X products decreased after algorithm screening. Experimental results confirm that using the isolated forest algorithm for optimization screening has high reliability and wide applicability in improving the elevation accuracy of laser points in ATL08 products.

[0142] (iii) Verify the overall vertical accuracy.

[0143] 1) Verification of vertical accuracy of ATL08 data

[0144] This study evaluates the elevation accuracy of four open-source digital surface models (DEMs) using laser point elevation data extracted from the ICESat-2 ATL08 product. To ensure the robustness of the evaluation results, the inherent elevation accuracy of the ATL08 data is first verified. Given that the 16 study areas span a wide range of latitudes and longitudes and have complex terrain, and considering the limitations of field surveying costs and timelines, this invention adopts a strategy of key area field measurements plus multi-source auxiliary verification. Specifically, Shaoyang area in Hunan Province is selected as the core verification area, and a large number of GNSS control points are deployed for direct verification. Simultaneously, high-precision digital surface models (DSMs) for parts of Bajiaolong, Xinweijiang, and Baiyangping are introduced as auxiliary references to further enhance the reliability of the verification.

[0145] The measured GNSS points were obtained using the WGS84 coordinate system and through manual stereoscopic surveying. All four verification areas are located within Hunan Province, encompassing typical landforms ranging from river valleys and plains to mid-to-high mountains, and are therefore highly representative. Figure 7 The distribution of laser points and measured GNSS points is shown, with blue dots representing extracted ICESat-2 laser points. A field operation scene and a histogram of elevation deviation distribution are also included. Results show that the RMSE between the laser points of the ATL08 product and the measured GNSS and DSM references is less than 2m, indicating that the ATL08 product has high vertical accuracy in complex terrain.

[0146] The geographic coordinates of the validation area, the number of samples, and the corresponding RMSE evaluation results are detailed in Table 5.

[0147] Table 5 Validation Area Results

[0148]

[0149] 2) Experiments on topographic index parameters

[0150] The laser point data from 16 experimental areas, including Hefei, Fangshan, and Zhangye, were screened. The comparison results of G_RMSE and SG_RMSE (G_RMSE obtained after screening through the extraction strategy) before and after screening are shown in Table 6.

[0151] Table 6 Comparison of G_RMSE and SG_RMSE after all screening processes are completed

[0152]

[0153] The specific performance of each DEM product is as follows:

[0154] The SRTM product exhibited the highest and most robust accuracy among the four DEM products. The G_RMSE values ​​after screening were all below 1m, with the mean decreasing from 1.63m to 0.39m. Zhengzhou had the largest error (0.75m), while Wenchang and Beichen had the highest accuracy, with an error of only 0.01m. Both the RTC DEM and Tan DEM-X products demonstrated excellent elevation accuracy. For the RTC DEM product, the G_RMSE values ​​for all experimental areas after screening were less than 1m, with the mean decreasing from 1.6m to 0.47m. The maximum value was 0.94m in Nanjing, and the minimum was 0.09m in Ningbo. For the TanDEM-X product, the G_RMSE values ​​for all 16 areas decreased, with the mean decreasing from 2.66m to 0.55m. The maximum value was 1.23m in Zhengzhou, and the minimum was 0.11m in Yantai. ASTER performed worse than the three DEMs mentioned above. The average G_RMSE value after screening decreased from 1.41m to 1.31m. Among them, the error in Zhangye was the largest at 2.97m, while the accuracy in Nanjing was the highest at only 0.02m.

[0155] After screening using the method of this invention, the mean G_RMSE of the four DEM products all decreased, indicating that the deviation between the elevation values ​​of the laser points and the GPS measured elevation values ​​was significantly reduced, achieving sub-meter accuracy. This strongly demonstrates that the laser point cloud optimized using the extraction strategy proposed in this invention has extremely high reliability, wide geographical applicability, and technical feasibility in replacing traditional GPS measured points for multi-source DEM product accuracy verification. This provides an effective alternative for large-scale, high-efficiency global DEM vertical accuracy assessment.

[0156] Combination Figure 5 As the 16 screening steps were implemented progressively, the evolution trends of the mean G_RMSE of the four DEM products showed significant differences. The curve of the ASTER product exhibited a clear non-linear fluctuation trend. Throughout the screening process, the mean G_RMSE failed to show a monotonically decreasing trend, reflecting that its vertical accuracy and stability were relatively sluggish in response to the extraction method. Ultimately, its mean G_RMSE remained at 1.31m, exhibiting the worst accuracy and highest volatility among the four product types. For RTC DEM, SRTM, and TanDEM-X, the mean G_RMSE curves showed a high degree of consistency, exhibiting a significant step-like decreasing trend overall. As the screening progressed, the mean G_RMSE of the 16 study areas continued to decrease, eventually stabilizing at 0.47m, 0.39m, and 0.55m, respectively. The mean G_RMSE of all radar source DEMs (RTC DEM, SRTM, TanDEM-X) was below 1m. Experimental results demonstrate that, after a full-process screening, RTC DEM, SRTM, and TanDEM-X significantly outperform ASTER in both accuracy and system robustness.

[0157] Figure 5 The evolution trajectory clearly verifies the significant effectiveness of the extraction strategy proposed in this invention in improving the accuracy of ATL08 laser elevation points. This strategy can efficiently and robustly identify and eliminate elevation anomalies, enabling the vertical accuracy of laser points to reach sub-meter levels.

[0158] pass Figure 6 Comparative analysis shows that at a resolution of 30 meters, SRTM and TanDEM-X, based on the InSAR system, exhibit superior elevation accuracy. After screening using the extraction strategy proposed in this invention, SRTM and TanDEM-X achieved G_RMSE as low as 0.28m and 0.32m in 50% of the studied areas, respectively. This is mainly attributed to the high penetration of active microwave remote sensing in complex atmospheric environments and the high sensitivity of phase measurements to terrain undulations. In stark contrast, ASTER products based on optical passive imaging showed the most inadequate accuracy. The upper limit of G_RMSE for ASRER in 50% of the area (1 m) even exceeded the maximum G_RMSE of SRTM in all areas (0.75m). This phenomenon strongly demonstrates that, under the same conditions, the elevation products generated by optical stereo photogrammetry have significantly lower consistency with laser point clouds and GPS benchmarks than InSAR products. With a high spatial resolution of 12.5m, RTC DEM showed good stability after screening, with G_RMSE below 1m in all areas. Its G_RMSE is below 0.46m in 50% of the region, and its accuracy performance is significantly better than ASTER.

[0159] Experimental data show that using the selected laser points as a truth verification source reveals the accuracy differences of DEMs from different systems. DEMs derived from InSAR have an overwhelming advantage in vertical alignment with high-precision laser point clouds. The screening strategy proposed in this invention provides a highly reliable data foundation for this cross-platform data verification.

[0160] (iv) The impact of land cover type on ASTER measurement accuracy.

[0161] After completing the entire screening process, ASTER products from different study regions exhibited significant differences. To explore the physical mechanisms behind this, this invention selected typical regions for comparative analysis of topographic and land cover classification. Figure 8 The study demonstrates the spatial distribution relationship between land cover types and laser points in the Wuhan and Weinan regions before and after screening.

[0162] Wuhan was one of the regions with the most significant screening results. Analysis showed that the retained experimental data points after screening were mainly distributed in building areas, with the proportion increasing from 81.72% to 90.36%, while the proportion in water areas decreased from 15.68% to 0%. Correspondingly, the G_RMSE significantly decreased from 2.12m to 0.31m. This indicates that the screening strategy successfully eliminated outliers caused by water interference and retained building photons with high coherence characteristics. Weinan and Zhengzhou are areas with poor accuracy. In these areas, G_RMSE increased instead of decreased after screening. In Weinan, the land cover types after screening were mainly crops and buildings. The proportion of crops increased from 41.01% to 49.55%, while the proportion of buildings decreased from 54.6% to 41.96%, causing G_RMSE to soar from 0.13m to 2.69m. In Zhengzhou, the land cover types were mainly crops and buildings. The proportion of crops increased from 41.9% to 56.48%, while the proportion of buildings decreased to 31.61%, and G_RMSE increased from 0.95m to 1.51m.

[0163] Statistical analysis of land cover types in most study areas revealed that areas with poor ASTER performance generally exhibited a significant increase in the proportion of crops. Research by Wang et al. showed that when vegetation cover is extensive, photons have difficulty penetrating the vegetation to reach the surface, resulting in insufficient surface photon counts and consequently, significant elevation errors. This invention suggests that the underlying cause lies in the fact that crops and low-lying weeds are rich in moisture, while ASTER's stereo matching performance is weak in areas with high water content and at the edges of water bodies. Simultaneously, the low height of vegetation makes it difficult for the photon classification algorithm to accurately distinguish between vegetation canopy photons and actual surface photons, thus introducing systematic elevation errors.

[0164] Experimental results confirm that areas with a significantly reduced proportion of crops after screening exhibited better ASTER validation accuracy than other areas. Conversely, an increased proportion of crops led to a significant decrease in ASTER validation accuracy and stability. This finding underscores the crucial role of land cover type in optical DEM accuracy assessment.

[0165] In summary, through experimental analysis of 16 study areas in China, the effectiveness of the method described in this invention and its performance in accuracy assessment using four open-source DEMs are demonstrated. The main conclusions are as follows:

[0166] (1) This method is the first to introduce a high-precision elevation control point extraction strategy for ATL08 products using terrain index parameters. Experimental results show that the parameters have a good screening effect on laser points and effectively ensure the elevation accuracy of the extracted laser points. In terms of algorithm performance, the isolated forest algorithm has the best screening effect on Tan DEM-X and RTC DEM data, followed by SRTM, while the screening effect of ASTER is relatively limited.

[0167] (2) The accuracy of the four DEM products varied across the 16 study areas. ASTER had a mean G_RMSE of 1.31m; while RTC DEM, SRTM, and Tan DEM-X had mean G_RMSEs of 0.47m, 0.39m, and 0.55m, respectively, all below 1m. Among the four open-source DEMs, SRTM had the highest overall accuracy, TanDEM-X and RTC DEM performed robustly, and ASTER had the worst performance. Experiments demonstrate that the method of this invention improves the overall elevation accuracy of the laser points in the ATL08 product.

[0168] (3) The method of this invention has good universality and robustness, and has been successfully verified in 16 experimental areas with distinct geographical features in China. Its high-precision screening results provide an effective solution to the problem of obtaining high-precision elevation control points over a large area. The laser point cloud optimized by the method of this invention has extremely high reliability, wide geographical applicability, and technical feasibility in replacing traditional GPS measured points for multi-source DEM product accuracy verification. It provides an effective alternative for large-scale, high-efficiency global DEM vertical accuracy assessment.

[0169] The above embodiments are merely preferred examples of the present invention and are not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Therefore, any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention should fall within the protection scope of the present invention.

Claims

1. An open-source DEM accuracy assessment method based on ICESat-2, characterized in that, Includes the following steps: Step S1, Data Acquisition: Acquire ICESat-2 laser point data and the digital elevation model (DEM) data to be evaluated; Step S2, set DEM evaluation value and accuracy index: use high-confidence elevation control points as elevation benchmarks, calculate the root mean square error (RMSE) of the DEM to be evaluated, and evaluate the elevation accuracy of the DEM. Step S3: Based on multi-dimensional attribute parameters, set attribute indicators to construct a comprehensive screening mechanism; Step S4: Design a control point extraction strategy based on the multi-attribute joint isolated forest algorithm, including: basic screening based on reference ground elevation values, preliminary screening of attribute indicators, advanced screening of attribute indicators, and screening using the isolated forest algorithm. Using the elevation accuracy values ​​of the laser points after advanced screening as input features, an isolated forest model is constructed for anomaly detection. Laser points identified as abnormal by the algorithm are removed, and finally, highly reliable elevation control points are obtained.

2. The open-source DEM accuracy assessment method based on ICESat-2 according to claim 1, characterized in that, In step S1, at least 10 regions spanning latitude and longitude are selected, with significant differences in regional geoscientific environment requirements, including at least four major landforms: plateau, hills, basins, and plains.

3. The open-source DEM accuracy assessment method based on ICESat-2 according to claim 1, characterized in that, In step S2, the evaluation index G_RMSE is introduced to quantify the degree of agreement between the DEM accuracy evaluation based on laser points and GPS measured points. Its calculation formula is as follows: ； in, and These represent the RMSE (Resolution of Emptiness) measured with respect to the DEM data when using ICESat-2 laser points and GPS control points as elevation references, respectively.

4. The open-source DEM accuracy assessment method based on ICESat-2 according to claim 3, characterized in that, The attribute indicators in step S3 include segmented elevation information, uncertainty parameters, ground parameter markers, cloud cover markers, topographic indicators, photon quantity parameters, photon distribution quality parameters, and signal-to-noise ratio, wherein: The segmented elevation information includes the maximum ground elevation, minimum ground elevation, average ground elevation, median ground elevation, interpolated ground surface elevation, best-fit ground elevation, and reference ground elevation. The uncertainty parameters include pointing angle uncertainty, radial trajectory uncertainty, local slope uncertainty, timing uncertainty, geographic positioning uncertainty, and scattering uncertainty; The ground parameter markings include ground slope and ground elevation skewness; The photon quantity parameters include ground photon count, canopy photon count, and top canopy photon count.

5. The open-source DEM accuracy assessment method based on ICESat-2 according to claim 4, characterized in that, The basic screening based on the reference ground elevation in step S4 is as follows: a threshold is preset, the interpolated ground surface elevation of the ICESat-2 laser point is compared with the reference ground elevation, and laser points whose absolute difference exceeds the set threshold are removed.

6. The open-source DEM accuracy assessment method based on ICESat-2 according to claim 4, characterized in that, The preliminary screening of attribute indicators in step S4 is specifically as follows: The interpolated ground surface elevation is compared with the ground median elevation. A deviation threshold is set, and laser points that exceed the deviation threshold are removed. The difference between the median ground elevation and the reference ground elevation is used for filtering. A threshold is set to remove laser points whose absolute value of the difference between the median ground elevation and the reference ground elevation exceeds the threshold range. By comparing the reference ground elevation with the best-fit ground elevation, laser points with an absolute difference exceeding three times the absolute elevation accuracy are eliminated. Geometric constraints are applied using ground slope indices, and slope thresholds are set to eliminate laser points whose ground slope indices exceed the slope thresholds. Data is filtered using the number of surface photons, and a threshold is set to retain laser points whose surface photon count exceeds the threshold. Data is filtered using the surface photon rate, and a threshold is set to remove laser points with a surface photon rate lower than the threshold.

7. The open-source DEM accuracy assessment method based on ICESat-2 according to claim 6, characterized in that, The advanced filtering of attribute indicators in step S4 is specifically as follows: Data filtering was performed using uncertainty parameters and ground elevation skewness. A threshold was set to remove laser points whose ground elevation skewness was greater than three times the RMSE value. Data is filtered using photon distribution quality parameters, retaining laser points whose photon distribution quality parameters are greater than -1 and whose three middle flag bits are equal to 1. Cloud cover markers are introduced for data filtering, and laser points with cloud cover markers greater than 3 are removed; Based on land cover data, laser points located in water bodies are removed; Remove laser points with a signal-to-noise ratio less than 1; Remove laser points whose ground index is greater than twice the RMSE; The standard deviation of the laser point elevation accuracy is calculated using the Laida criterion, with 3 times the standard deviation as the final threshold, and laser points exceeding the final threshold are removed.

8. The open-source DEM accuracy assessment method based on ICESat-2 according to claim 7, characterized in that, The method for constructing an isolated forest model for anomaly detection includes: randomly selecting subsamples from the input elevation accuracy training data to construct multiple isolated binary trees; during the construction of tree nodes, randomly selecting a cutting point to recursively divide the data space until the termination condition is met; calculating the path length of each laser point in all isolated trees and obtaining an anomaly score accordingly; and identifying laser points with anomaly scores exceeding a preset threshold as anomalies and removing them.

Images