Multi-source DEM accuracy evaluation based on ICESat-2 in Qinghai-Tibet Plateau, China

ABSTRACT

Digital Elevation Models (DEMs) are critical datasets in the field of Earth sciences, essential for accurate measurement and analysis of the Earth's surface. We used ICESat-2 to quantitatively evaluate 5 DEMs (ALOS PALSAR, ASTER GDEM V3, COPERNICUS, NASADEM and TanDEM-X) on the Tibetan Plateau. The research findings indicate that the ALOS exhibits the highest level of accuracy, as evidenced by its root mean square error (RMSE) value of 5.05m. It is closely followed by NASA and COP, which have RMSE values of 6.23m and 8.10m, respectively. In contrast, the ASTER and TDX90 demonstrate comparatively lower levels of accuracy, as indicated by their respective RMSE values of 11.47m and 12.32m. It is worth mentioning that the accuracy of DEMs is significantly influenced by land cover, especially during the transition from areas with low vegetation to those with high vegetation. This transition often results in a decrease in accuracy. Furthermore, the accuracy of DEMs tends to decrease with increasing slope values. Aspect exhibits a notable spatial distribution pattern characterized by a "low in the Southwest direction, high in all other directions" phenomenon. This study offers significant contributions to the evaluation of DEMs accuracy and its suitability in the Qinghai-Tibet Plateau.

KEYWORDS:

• Digital elevation models
• ICESat-2 ATL08
• vertical accuracy assessment
• Qinghai-Tibet Plateau

1. Introduction

Digital Elevation Models (DEMs) are pivotal tools for the digital representation of Earth's surface morphology, offering immense potential for a wide array of geoscientific studies. Elevation-related research in various fields benefits from the diverse applications of DEMs. These applications include glacier mass balance assessments (Berthier et al. 2007), terrain classification (H. Li et al. 2012), surface process modeling (Tarolli 2014), and flood inundation mapping (Saksena and Merwade 2015). The accuracy of DEMs is crucial in guaranteeing the dependability and practical usefulness of these applications.

Since 2003, researchers have had access to multiple open-source DEM datasets, including Shuttle Radar Topography Mission (SRTM) DEM (Farr et al. 2007), the Advance Land Observing Satellite (ALOS) world 3D-30 m (AW3D30) DEM (Tadono et al. 2016), the Advanced Land Observing Satellite Phased Array Type L-band Synthetic Aperture Radar (ALOS PALSAR) (Nitheshnirmal et al. 2019), the Advanced Spaceborne Thermal Emission and Reflection radiometer (ASTER) DEM (Uuemaa et al. 2020), COPERNICUS (Fahrland et al. 2022), the National Aeronautics and Space Administration’s DEM (NASADEM) (W. Chen et al. 2022), and TerraSAR-X add-on for Digital Elevation Measurement (TanDEM-X) DEM (Zink et al. 2014), among others.

Among these datasets, AW3D30 and ASTER GDEM utilize optical photogrammetric techniques to generate stereo data products, while SRTM DEM, ALOS PALSAR, COPERNICUS, NASADEM, and TanDEM-X rely on Interferometric Synthetic Aperture Radar (InSAR) technology for acquisition. However, the vertical accuracy of DEMs is influenced by various factors, including horizontal errors, terrain features, and land cover. Additionally, inherent errors due to variations in imaging configurations and data collection technologies, as well as other systematic biases resulting from the original data collection and production methods, come into play. Hence, evaluating the vertical accuracy of these open-source DEMs is of paramount importance (Hirt, Filmer, and Featherstone 2010; Mouratidis and Ampatzidis 2019; Tran et al. 2023).

To assess the vertical accuracy of DEMs, the reliance on accurate and reliable reference datasets is imperative. Previous research has primarily relied on Ground Control Points (GCP) acquired through Global Positioning System (GPS) measurements as validation data (Briole et al. 2021; Florinsky, Skrypitsyna, and Luschikova 2018; Leitão et al. 2016; P. Li et al. 2012). Although GCPs provide a high level of precision, collecting elevation data for ground points is a time-consuming task and can be challenging in remote mountainous regions and other inaccessible areas. Additionally, GCPs are often situated in relatively flat regions with limited land cover. This can result in an insufficient representation of the spatial distribution of errors, which in turn can introduce significant biases in the assessment results (Guth and Geoffroy 2021). Several researchers have utilized pre-existing high-precision DEMs for the purpose of evaluation (Hu et al. 2017; Mukherjee et al. 2013; Shortridge and Messina 2011). However, this approach is only applicable to specific regions, thereby limiting the assessment of accuracy for large-scale DEMs (Berthier et al. 2006; Leitão et al. 2016). As technology has evolved, point cloud data generated by airborne Light Detection and Ranging (LiDAR) and height measurements obtained from satellite laser altimeters like Ice, Cloud, and Land Elevation Satellite (ICESat) have become crucial tools for assessing elevation accuracy. Particularly, precise LiDAR height data help overcome some of the limitations of ground control points and have become the most effective means for evaluating the accuracy of large-scale DEMs. Although the ICESat/GLAS satellite launched by NASA in 2003 provided accurate laser altimetry data globally and found extensive use in DEMs evaluation (Bhang, Schwartz, and Braun 2007; C. Chen, Yang, and Li 2020), it is important to note that the previous ICESat/GLAS data, characterized by larger footprints, were prone to being influenced by terrain slope, resulting in an accuracy of approximately 2 m (Wessel et al. 2018). This implies that validating the accuracy of DEMs accuracy in rugged areas was not always applicable (Z. Liu et al. 2020). With the launch of the Ice, Cloud, and Land Elevation Satellite-2 (ICESat-2) in 2018, equipped with the Advanced Topographic Laser Altimeter System (ATLAS), there is now a greater availability of densely and precisely surveyed elevation information. Research has shown that using ATL08 terrain height data generated by ATLAS, accuracy can reach 0.70 m (Neuenschwander et al. 2020). Furthermore, some scholars have demonstrated that by utilizing ICESat-2 ATL08 data, it is possible to automatically extract high-density and high-precision GCP data can be achieved by selecting elevation data and attribute parameters (Wang et al. 2021). This approach is more suitable for large-scale assessments of vertical accuracy in GDEM (Hui Li et al. 2022; Neuenschwander and Pitts 2019).

Error analysis of global open-source DEMs data has been widely researched both domestically and internationally. Some studies have shown that the actual vertical accuracy of DEMs is generally higher than the officially reported vertical accuracy (González-Moradas and Viveen 2020). However, DEMs error assessments for the Qinghai-Tibet Plateau (TP) remain relatively scarce. In earlier research, some scholars used ICESat/GLAS data to compare SRTM and map-based DEMs, providing a basis for SRTM as a potential dataset to monitor glacier volume changes since 2000 (Huang et al. 2011). Additionally, some researchers in high-altitude mountainous regions in Asia utilized ICESat/GLAS data and Global Positioning System (GPS) measurements to assess the quality of seven different open-source DEMs datasets, with results indicated that AW3D30 performed the best in high mountain areas of Asia (K. Liu et al. 2019). Other scholars, by analyzing the elevation differences between SRTM and TanDEM-X data, outlined the spatial patterns of glacier mass balance in the Tibetan Plateau. This research provides essential evidence for comprehending the influence of global climate change on the cryosphere (Ke et al. 2020). Recent studies have examined the performance of six DEM datasets in the Tibetan Plateau, and the findings suggest that NASADEM demonstrates the highest vertical accuracy. This research also tested the impact of these six DEM datasets on four ice thickness inversion models, revealing that NASADEM is the best choice for estimating the entire TP ice sheet thickness (W. Chen et al. 2022).

While there have been numerous validation studies on global DEMs data, there still exist some gaps in understanding the error distribution characteristics of DEMs data in the Qinghai-Tibet Plateau. If the accuracy of these new digital elevation model products, based on the latest remote sensing data and more advanced processing methods, is verified, they will be of great value as they can capture natural or anthropogenic changes in surface terrain. The objective of this study is to comprehensively assess the Tibetan Plateau using ICESat-2 ATL08 laser point cloud data as a reference. Five recently released global free open-source DEMs products will be selected for evaluation: ALOS PALSAR (referred to as ALOS, 12.5 m), ASTER GDEM V3 (referred to as ASTER, 30 m), COPERNICUS (referred to as COP, 30 m), NASADEM (referred to as NASA, 30 m), and TanDEM-X (referred to as TDX90, 90 m). These products will be collectively referred to as GDEM5 in this text. We will investigate the vertical accuracy of GDEM5 and its correlations with terrain features such as elevation, land cover, slope, and aspect in order to understand the random and systematic effects among each DEM. By understanding the error distribution characteristics of GDEM products with varying spatial resolutions and acquisition techniques in the Tibetan Plateau, we can offer guidance for selecting future DEMs data in this region. This will not only ensure the reliability and accuracy of GDEM applications, but it will also provide valuable insights for improving DEMs data in the area.

2. Study area and data sources

2.1. Study area

The Qinghai-Tibet Plateau, located in western China, extends from approximately 26°00’ to 39°47’ North latitude and 73°19’ to 104°47’ East longitude. It is the highest plateau in the world and the largest in Asia. Its topography is characterized by a wide variety of features, including vast highland plains, steep mountain slopes and ranges, deep valleys, and canyons.

This diversity places exceptionally high demands on the accuracy and precision of DEMs data, particularly in areas with rapidly changing terrain. Moreover, the TP is home to numerous glaciers and high-altitude lakes, which require even higher accuracy in DEMs to precisely depict the distribution and shape of glaciers, as well as the water surface outlines of high-altitude lakes. Given the complexity of the TP topography and its significant impact on global climate change through its influence on weather patterns, it is crucial to analyze the accuracy errors in different DEMs data for this region (Figure 1).

Figure 1. NASA DEM-based study area profile map, the blue line is the screened ICESat-2 ALT08 sample sites.

2.2. Datasets

The data used in our study mainly include ALOS PALSAR, ASTER GDEM V3, COPERNICUS, NASADEM, TanDEM-X, and the reference data are ICESat-2 ATL08 (Table 1).

Table 1. Overview of the global digital elevation model evaluated in this study.

Datasets	Acquired	Producer	Grid	Coverage	Datum plain/Vertical	Method source	Vertical accuracy
ALOS PALSAR	2018	ASF	12.5 m	60°N∼59°S	WGS84/WGS84	L band radar	∼5 m
ASTER GDEM V3	2019	NASA METI	30 m	83°N∼83°S	WGS84/EGM96	Stereo NIR imagery	∼20 m
COPERNICUS	2020	ESA	30 m	Global	WGS84/EGM2008	X band radar	4 m
NASADEM	2020	LP DAAC	30 m	60°N∼56°S	WGS84/EGM96	C band radar	Not reported
TanDEM-X	2016	DLR	90 m	85°N∼90°S	WGS84-G1150/WGS84-G1150	X band radar	<10 m
ICESat-2 ATL08	2018	NASA	70 cm	90°N∼90°S	WGS84/WGS84	Photon-counting	0.1 m

2.2.1. ALOS PALSAR

ALOS PALSAR data is a remote sensing dataset developed by the Japan Aerospace Exploration Agency (JAXA) and derived from the Advanced Land Observing Satellite. It offers a spatial resolution of 12.5 m and utilizes L-band frequencies, providing high-resolution, all-weather, and continuous surface observation capabilities. In comparison to optical remote sensing, SAR has the advantage of penetrating cloud cover and conducting observations at night.

2.2.2. ASTER GDEM

ASTER GDEM is a digital elevation model generated by the ASTER instrument aboard the Terra satellite. The initial version of this product was known as ASTER GDEM V1, which covered the region from 83°N to 83°S and had an approximate spatial resolution of 30 m. However, V1 suffered from artifacts and outliers. In an effort to enhance data quality, ASTER GDEM V2 was released in 2011. This version incorporated improved processing algorithms to derive high-quality DEMs from a greater number of ASTER stereo pairs. Despite these improvements, V2 still exhibited anomalies and artifacts. The release of ASTER GDEM V3 in 2019 addressed these issues by fixing elevation void areas and enhancing data availability. This version represents an improvement over its previous versions, making it a valuable resource for geospatial applications.

2.2.3. COPERNICUS

COPERNICUS is an Earth observation program established through a collaboration between the European Union (EU) and the European Space Agency (ESA). Its primary mission is to create high-precision DEM. COPERNICUS provides three distinct DEM products with varying resolutions: 0.4″, 1″, and 3″. These products undergo various processing steps, including terrain correction and hydrological editing, to eliminate peaks and voids, identify and fill gaps, edit coastlines, and correct unreliable terrain structures and random biases (Hui Li et al. 2022). This rigorous processing ensures the accuracy and reliability of the resulting DEM data.

2.2.4. NASADEM

NASADEM is a DEM initially released by NASA's Jet Propulsion Laboratory (JPL) in 2019, with an updated version released in 2020. It is a global, nearly void-free, 1 arc-second (30 m) DEM. It serves as a successor to SRTM V3 and is developed by reprocessing the original SRTM radar data using improved algorithms and reference data from the Ice, Cloud, and Land Elevation Satellite (ICESat). This dataset is a compilation of various sources aimed at improving the accuracy and coverage of SRTM data. Some researchers have demonstrated that NASADEM exhibits higher vertical accuracy compared to SRTM DEM (Uuemaa et al. 2020).

2.2.5. TanDEM-X

TanDEM-X is a SAR interferometric measurement system formed by two satellites, the German TerraSAR-X and TanDEM-X. During their flights, they remotely sense the Earth's surface. This configuration reduces or eliminates the extensive time traditionally required for InSAR and the impact of atmospheric interference on the collection of high-precision cross-track interferograms (Zink et al. 2014). In 2016, a 0.4 arc-second (12 m) spatial resolution TanDEM-X DEM was released. However, access to it required payment. In 2018, the German Aerospace Center (DLR) made a 90 m resolution TanDEM-X DEM freely available to the scientific community. The current version is an unedited release, which means that there may be unreliable elevation values in areas with complex terrain, such as water bodies, steep slopes, and dense vegetation. Additionally, no gap-filling or interpolation processing has been performed, which may result in data gaps or invalid values (Yu et al. 2021).

2.2.6. ICESat-2

In September 2018, the United States’ NASA launched its latest spaceborne laser altimetry mission, ICESat-2, which carries the ATLAS. This system utilizes the time of flight of laser pulses to measure the elevation of the Earth's surface. In the spring of 2019, ICESat-2 provided its measurements to the scientific community. Compared to the ICESat/GLAS data released in 2003, ICESat-2 achieves more accurate Earth surface elevation data by providing high-density laser point clouds. It can more precisely measure the elevations of mountain glaciers and ice caps, land and canopy heights, and monitor the Earth's carbon storage. ICESat-2 offers detailed topographic information and enables more precise analysis of elevation changes. Researchers have made ATLAS data available and validated the accuracy of its ground elevations, confirming the geolocation precision of ICESat-2 laser points and the accuracy of its data products (Neuenschwander and Pitts 2019).

3. Accuracy evaluation method

3.1. Data pre-processing

To accurately assess the vertical accuracy of GDEM5, data preprocessing is of paramount importance. We downloaded ICESat-2 ATL08 data and used Python to extract eight parameters from the original HDF5 files (Table 2) (Zheng et al. 2022). All the laser point cloud data within the Tibetan Plateau region was extracted, resulting in a total of 421,110 elevation points. The WGS84 coordinate system was selected as a reference in this study.

Table 2. Description of ATL08 data extraction parameters.

Parameters	Describe
latitude	Latitude of the central photon within each statistical cell
longitude	Longitude of the central photon within each statistical cell
dem_h	The best available DEM elevation for this geographic location, with a reference ellipsoid of WGS84
h_te_best_fit	The best-fit topographic elevation at the center within each statistical cell
h_te_interp	Interpolated land elevation marked as ground photons
terrain_slope	Slope of the ground in each statistical cell in the along-track direction, calculated by linear fitting
night_flag	Night marker, 0 = day, 1 = night
cloud_flag_atm	Cloud amount marker, the value range is [0,10], above 0 indicates the presence of clouds or aerosols, [0,2] indicates that the cloud amount is less than or equal to 20%

3.1.1. Coarse filtering of reference data

Before undertaking the validation of DEM elevation accuracy, it is imperative to carry out an initial filtering process on the ICESat-2 ATL08 elevation data in order to ensure the reliability of the data. These data are subject to the influence of multiple factors, such as slope, cloud cover, signal-to-noise ratio, acquisition time, and land cover (Zheng et al. 2022; Zhu et al. 2018).

In the preliminary filtering, special attention should be given to the influence of slope on elevation accuracy. Research has shown that slope significantly affects elevation accuracy, with steeper slopes leading to lower accuracy. As a result, laser footprint points with a slope parameter greater than 0.01 were excluded to ensure the removal of elevation outliers. Cloud cover is another crucial factor to consider. Lower cloud cover typically has a positive effect on elevation accuracy, so data with lower cloud cover is preferred as it often provides higher elevation accuracy. Acquisition time is another influencing factor. Data collected at night generally exhibits higher accuracy because it is less affected by atmospheric scattering and solar background noise. Therefore, data collected at night was chosen to obtain more reliable elevation information.

The filtering criteria for ICESat-2 ATL08 attribute parameters are as follows (Wang et al. 2021):

1. $\mathrm{terrain}\mathrm{\_}\mathrm{slope}<{\mathrm{T}}_{\mathrm{slope}}$

2. $\mathrm{cloud}\mathrm{\_}\mathrm{flag}\mathrm{\_}\mathrm{atm}<{\mathrm{T}}_{\mathrm{cloud}}$

3. $\mathrm{night}\mathrm{\_}\mathrm{flag}<{\mathrm{T}}_{\mathrm{night}}$

In condition (1), T_slope represents the slope threshold. Research results indicate that when T_slope is set to 0.01, the laser points can achieve elevation accuracy below 0.1 m. In condition (2), T_cloud represents the cloud cover threshold, set to 3, indicating that the cloud_flag_atm value range is [0, 2], representing the extraction of data with cloud cover not exceeding 20%. In condition (3), T_night represents the time range, with 0 indicating daytime and 1 indicating nighttime. Setting the time range to 1 means extracting only nighttime observation data.

After the initial filtering, there are 306,186 ICESat-2 elevation point samples available for further analysis.

3.1.2. Fine filtering of reference data

After the initial filtering of ICESat-2 data, a more rigorous filtering process was conducted to ensure the accuracy of the elevation data. Based on the assumption that ICESat-2 ATL08 data serves as a reference truth, low-quality GCPs were excluded through a comparison with the GDEM datasets. For each GCP, the height difference (Dh) between its values in the ATL08 and GDEM datasets was calculated. Points with a Dh outside the range of ±100 were removed as outliers in order to achieve the objective of precise filtering. This method has been widely applied in previous research (Ding et al. 2022; Huang et al. 2011). While some scholars propose using the 3-sigma rule to obtain more accurate GCP data (Hawker, Neal, and Bates 2019; Hui Li et al. 2022), it is important to note that many data points that did not meet the criteria were already excluded in the initial filtering stage. Additionally, considering the requirement for an adequate number of GCP samples in the Tibetan Plateau, the 3-sigma rule was not utilized in this study.

In the end, 248,754 high-quality GCPs were obtained. These GCPs will be utilized to assess the accuracy of the GDEM, ensuring the dependability and scientific integrity of the research.

3.2. Elevation datum conversion

Before conducting DEM accuracy assessment, it's essential to address the elevation discrepancies arising from different vertical datums between ATL08 and GDEM data to ensure consistency and comparability (Hui Li et al. 2022). This is a crucial step in reducing sources of error and ensuring the accuracy of the final elevation analysis.

First, we utilized the vertical datum transformation tool VDatum, developed by the National Oceanic and Atmospheric Administration (NOAA), to convert the vertical datums of EGM2008 and G1150 to EGM96 (Hui Li et al. 2022). This conversion helps align the vertical datums of both datasets, reducing sources of error. Subsequently, we converted EGM96 geoid heights to the common WGS84 ellipsoid height, which helps eliminate errors caused by differences in vertical datums (Cai et al. 2022; Das, Agrawal, and Mohan 2014; Z. Liu et al. 2020). The formula for vertical datum transformation is as follows: (1) ${\mathrm{H}}_{\mathrm{elip}}={\mathrm{H}}_{\mathrm{ortho}}+\mathrm{N}$(1) In the given formula, H_elip denotes the ellipsoid height of the WGS84 ellipsoid, H_ortho represents the orthometric height on the EGM96 geoid, and N signifies the disparity between the ellipsoid height and orthometric height. In the processing, the value of N can be derived by utilizing the WGS84.img file, which is made available through the ArcGIS software (Cai et al. 2022).

3.3. Elevation accuracy evaluation index

To evaluate the dependability and suitability of GDEM5 data across various resolutions, the errors of GDEM5 were measured by employing GCPs. This study aimed to quantitatively characterize the distribution pattern of GDEM5 elevation errors under various terrain features by selecting five evaluation indexes. The evaluation indexes are presented as follows:

3.3.1. Height disparity

The elevation error represents the discrepancy between the GDEM5 elevation (${\mathrm{h}}_{\mathrm{dem}}$) and the ICESat-2 elevation (${\mathrm{h}}_{\mathrm{ICESat}-2}$). It is calculated using the following formula: (2) $d={\mathrm{h}}_{\mathrm{dem}}-{\mathrm{h}}_{\mathrm{ICESat}-2}$(2)

3.3.2. Mean error

The mean error represents the average difference between the GDEM5 and ICESat-2 elevations. It is calculated by summing up all elevation differences and dividing by the total number of ICESat-2 ATL08 sample points: (3) $\mathrm{ME}={\displaystyle \frac{\sum d}{n}}$(3)

3.3.3. Standard deviation

The standard deviation indicates the dispersion of the elevation errors around the mean error. It is calculated using the following formula: (4) $\mathrm{SD}=\sqrt{{\displaystyle \frac{\sum {(d-\mathrm{ME})}^2}{n}}}$(4)

3.3.4. Mean absolute error

The mean absolute error represents the average absolute difference between the GDEM5 and ICESat-2 elevations. It is calculated by summing up the absolute values of the elevation differences and dividing by the total number of ICESat-2 ATL08 sample points: (5) $MAE={\displaystyle \frac{\sum |d|}{n}}$(5)

3.3.5. Root mean square error

The root mean square error provides a measure of the overall accuracy of the GDEM5 data. It is calculated by taking the square root of the average of the squared elevation differences: (6) $RMSE=\sqrt{{\displaystyle \frac{\sum d^2}{n}}}$(6) These evaluation indices allow for a quantitative assessment of the distribution pattern of GDEM5 elevation errors across different terrain features (Jiang et al. 2020; W. Li, Wang, and Zhu 2020). The analysis provides valuable insights into the accuracy and reliability of the GDEM5 data compared to the reference ICESat-2 data.

4. Results

4.1. Overall DEM error assessment

We conducted a comprehensive assessment of the elevation errors in GDEM5. The results are presented in the statistical analysis histograms and scatter plots shown in Figure 2.

Figure 2. Histogram of error distribution of GDEM5: (a) ALOS, (b) ASTER (c) COP (d) NASA (e) TDX90.

From visual analysis, although there is some variation in the shape of the error curves in GDEM5, the error curves for all DEMs seem to roughly follow a normal distribution, with the centers of the distributions being close to zero. It is worth noting that, with the exception of TDX90, all DEMs show a small positive bias and positive skewness values. Additionally, the error distribution plots for COP and NASA DEMs show strong peaks. The concentration of the Gaussian curve distribution indicates that smaller data errors correspond to a more tightly concentrated error distribution, thus indicating higher accuracy.

Table 3 summarizes the descriptive statistics for all DEMs, representing their performance in terms of ME,SD,MAE, and RMSE.

Table 3. Basic statistics of height difference of GDEM5.

Data sets	Number of points	ME	SD	MAE	RMSE
ALOS PALSAR	248754	0.22	5.04	2.83	5.05
ASTER GDEM V3	248754	3.41	10.94	7.88	11.47
COPERNICUS	248754	2.96	7.53	4.55	8.10
NASADEM	248754	2.57	5.67	3.33	6.23
TanDEM-X	248754	−0.72	12.03	6.92	12.32

The ME values for GDEM5 range from −0.72 m to 3.41 m. Specifically, ALOS, ASTER, COP, and NASA have positive ME values, indicating that they generally underestimated the elevation distribution of the Tibetan Plateau. In contrast, TDX90 has a negative ME value, suggesting that it underestimated the terrain height. The positive ME values can be partially attributed to delays in laser echoes caused by ground cover, vegetation, buildings, or other objects. The negative ME values may be partly due to the influence of dense vegetation canopies, which can prevent shorter-wavelength SAR or optical satellite sensors from ‘seeing through’ the canopy and thus affect elevation accuracy. The MAE values range from 2.83 m to 7.88 m, with a smaller MAE indicating greater accuracy. Interestingly, the MAE of ASTER at 30 m resolution (7.88 m) is larger than the MAE of TDX90 at 90 m resolution (6.92 m), which is an unexpected result.

The SD and RMSE values range from 5.04 m to 12.03 m, while the RMSE values range from 5.05 m to 12.32 m. SD is used to describe the dispersion of data, while RMSE is also used to evaluate the accuracy of a model or measurement. From the perspective of GDEM5, the ALOS data exhibits the smallest dispersion, with SD and RMSE of 5.04 m and 5.05 m, respectively, indicating the highest level of accuracy. NASA and COP follow, with relatively concentrated error distributions, having SD and RMSE values of (5.67 m, 7.63 m) and (6.23 m, 8.10 m), respectively. The ASTER and TDX90 datasets have the most dispersed error distributions, indicating lower accuracy. The standard deviation and root mean square error (RMSE) values for ASTER are 10.94 m and 12.03 m, respectively. For TDX90, the SD and RMSE values are 11.47 m and 12.32 m, respectively. These statistics provide detailed information about the performance and accuracy of various DEMs.

The results presented in Figure 2 and Table 3 provide a comprehensive evaluation of the accuracy and reliability of the GDEM products. Research has shown that the resolution of DEMs significantly impacts terrain analysis. Changes in DEM resolution, from fine to coarse, also affect the statistical characteristics of elevation values(Sharma and Bagri 2023). We observe that ALOS exhibits the highest quality due to the significant advantage of high-resolution DEMs in capturing terrain details. As a reprocessed product of SRTM, NASA outperforms COP in various metrics (Hui Li et al. 2022; M. Li et al. 2023). Studies suggest that the use of COP performs relatively poorly in rugged and densely vegetated areas (W. Chen et al. 2022). Scholars have also confirmed that NASA offers superior quality in terms of the entire terrain and ice thickness modeling for the TP.

While ASTER GDEM V3 includes an additional 360,000 optical stereo pairs compared to the V2 version, its actual accuracy improvement is limited. The main focus of this version is to reduce data gaps and water-related anomalies. ASTER GDEM V3's relatively lower accuracy at 30 m resolution may be attributed to its dependence on traditional optical stereo photogrammetry, which has limited penetration capabilities. Furthermore, researchers (Carrera-Hernández 2021; M. Li et al. 2023) have already demonstrated that the data accuracy of ASTER GDEM V3 is higher in flat areas but generally lower in mountainous regions, dense vegetation, and river areas.

It should be noted that despite the lower resolution of the TDX90 (90 m), its accuracy is comparable to that of the 30 m-resolution ASTER. The RMSE for TDX90 is 12.32 m, while for ASTER it is 11.47 m. This is a surprising result. This may reflect the excellent quality of TanDEM-X data and superior data processing techniques. TDX90, as the latest product, utilizes InSAR interferometry, which makes the mapping process more efficient and accurate. However, this result has been previously confirmed in studies such as Liu (Z. Liu et al. 2020) which evaluated various DEMs including SRTM-1 DEM, SRTM-3 DEM, ASTER GDEM2, AW3D30 DEM, and TanDEM-X 90-m DEM, as well as a 30 m resampled TanDEM-X DEM. The results indicate that the quality of ASTER GDEM2, AW3D30 DEM, and TDX30 DEM is even lower than that of SRTM3 and TDX90 DEM, especially for mountainous terrain. Similar conclusions have been drawn by other researchers (Briole et al. 2021; Grohmann 2018; X. Liu et al. 2022). Therefore, the fact that TDX90's RMSE is approaching that of ASTER suggests the inherent quality advantage of TDX90 data. This advantage will be further analyzed in relation to terrain factors, land cover, and other relevant factors.

4.2. Influence of different elevations on the accuracy of DEM

Due to the combined effects of terrain relief and variations in relative elevation, terrain features display different characteristics in different elevation zones.

In high-altitude areas, as the terrain relief increases, the elevation shows a logarithmic increase followed by fluctuating changes, while the relative elevation displays an exponential increase followed by fluctuations. Studies have shown (Z. Feng et al. 2020) that there is a positive correlation between terrain relief, relative elevation, and mean altitude. Therefore, we categorized elevation into different ranges: ≤ 1200 m, 1200 m–2000 m, 2000 m–4000 m, 4000 m–5000 m, and ≥5000 m. Figure 3 illustrates the distribution characteristics of elevation errors within different elevation ranges.

Figure 3. Differences between GDEM5 and ICESat-2 ATL08, shown as violin plots according to different altitudes: (a) ALOS, (b) ASTER (c) COP (d) NASA (e) TDX90.

Overall, the datasets provided by ALOS, COP, and NASA demonstrate a consistent distribution across all elevation intervals, suggesting a state of relative stability. However, the TDX90 exhibits significant deviations at elevations lower than 1200 m and higher than 5000 m. Additionally, the error distribution of ASTER data is found to be highly unstable, potentially due to a significant influence from elevation. Therefore, the GDEM5 data demonstrates diverse terrain representation capabilities across distinct elevation zones.

To further investigate the impact of different elevations on DEM errors, we utilize heatmaps (Figure 4) to visually depict the distribution of GDEM5 assessment values within each elevation range. Concerning the ME values, ALOS, COP, NASA, and TDX90 show minimal variations across different elevation ranges, with ME values approaching zero. However, ASTER's ME shows significant differences, ranging from −1.46 m to 5.37 m. This suggests that there is a highly uneven distribution of errors in ASTER data. As for SD, MAE, and RMSE, ALOS, COP, and NASA values exhibit slight upward trends from low to high elevations, while showing minimal variations across different elevation zones.

Figure 4. Differences between GDEM5 and ICESat-2 ATL08, ME, SD, MAE and RMSE calculated according to different altitudes shown as heat maps: (a) ALOS, (b) ASTER (c) COP (d) NASA (e) TDX90.

It's worth noting that ASTER still demonstrates substantial error variations in various elevation zones, indicating significant data quality instability. SD ranges from 6.46 m to 12.11 m, MAE varies between 4.33 m and 8.58 m, and RMSE fluctuates from 6.53 m to 13.25 m. Furthermore, TDX90's SD, MAE, and RMSE values show significant increases from low to high elevation zones. The lowest values are observed in the ≤1200 m range, with values of 6.78 m, 3.10 m, and 6.82 m for SD, MAE, and RMSE, respectively. Conversely, the highest values are found in the ≥5000 m range, with values of 13.93 m, 8.71 m, and 13.96 m for SD, MAE, and RMSE, respectively.

These results reaffirm that the errors of GDEM5 are influenced by different elevation zones, particularly in 90 m resolution DEMs. The variations in errors become more pronounced as the elevation increases. This also suggests that high-resolution DEMs are better at capturing fine details of mountainous terrain, have a higher tolerance for elevation changes, and exhibit smaller errors compared to low-resolution DEMs. This aspect can be further validated through the analysis of slope, aspect, and land cover.

4.3. Effect of different land covers on the accuracy of DEM

To gain a deeper understanding of the correlation between DEM accuracy and land cover attributes, we selected the GlobaLand30 V2020 dataset, which was created by China's National Geomatics Center (NGCC) and possesses a resolution of 30 m. This dataset comprises 10 land cover categories and attains an overall classification accuracy of 83.51%. Within the Qinghai-Tibet Plateau, a total of nine distinct land cover types have been identified. These include cropland (CL), forest (FR), grassland (GL), shrubland (SL), wetland (WL), water bodies (WB), artificial surfaces (AS), bare land (BL), and glaciers and permanent snow (SN).

Figure 5 illustrates the distribution of errors in GDEM5 across various land cover types within the Qinghai-Tibet Plateau. We have observed that the accuracy of the DEM is significantly influenced by various land cover types. The error distribution of GDEM5 is found to be most dispersed in forests, glaciers, and permanent snow. On the other hand, ASTER exhibits the most unstable distribution across different land cover types, suggesting its heightened sensitivity to variations in land cover categories. Additionally, the utilization of box plots and normal distribution curves provides further support for these observations, indicating that ALOS demonstrates the highest level of accuracy across a range of land cover types. This is followed by NASA, COP, ASTER, and TDX90. The minimum MAE values for ALOS, NASA, and ASTER are observed in water bodies, whereas the minimum MAE values for COP and TDX90 are found in wetlands. Additionally, it is observed that GDEM5 demonstrates the highest MAE in forested areas, potentially due to the influence of vegetation on elevation measurements.

Figure 5. Differences between GDEM5 and ICESat-2 ATL08, based on box line plots and normal distribution curves for different land cover types (red diamonds in the figure indicate MAE): (a) ALOS, (b) ASTER (c) COP (d) NASA (e) TDX90; (CL) cropland, (FR) forest, (GL) grassland, (SL) shrubland, (WL) wetland, (WB) water body, (AS) artificial surface, (BL) bare land, (SN) glacier and permanent snow.

Further statistical analysis reveals significant differences among various land cover types (Figure 6). For instance, the forest has the highest RMSE, which can be attributed to vegetation obstructing signal reflection and penetration for optical photogrammetry and short-wavelength SAR. The laser beam is obstructed by the vegetation canopy, which affects the accuracy of elevation measurements of the Earth's surface. Additionally, with the exception of TDX90, artificial surface types exhibit the lowest RMSE. This is due to the DEM's excellent performance in accurately measuring the height of built structures. TDX90 exhibits the lowest RMSE in wetland types, while showing larger RMSE in other types. This may be attributed to the low spatial resolution and characteristics of SAR interferometric measurements. Low coherence or low backscattering can introduce noise over water surfaces (Rizzoli et al. 2017). Furthermore, water bodies, glaciers, and permanent snow exhibit higher RMSE values due to variations in terrain ruggedness and the challenges associated with seasonal changes in water bodies, as well as complex electromagnetic wave reflection and scattering. These findings are consistent with the research of scholars such as (Hui Li et al. 2022; M. Li et al. 2023; Uuemaa et al. 2020).

Figure 6. Differences between GDEM5 and ICESat-2 ATL08, ME, SD, MAE and RMSE calculated according to different altitudes are shown as histograms for (a) ALOS, (b) ASTER (c) COP (d) NASA (e) TDX90; (CL) cropland, (FR) forest, (GL) grassland, (SL) shrubland, (WL) wetland, (WB) water body, (AS) artificial surface, (BL) bare land, (SN) glacier and permanent snow.

As mentioned above, different land cover types significantly impact the accuracy of DEM. It is worth noting that we observed a significant increase in GDEM5 RMSE when transitioning from areas with low vegetation cover to areas with high vegetation cover. This indicates a significant drop in DEM accuracy, with different land cover types ranked as follows: forest > glaciers and permanent snow > bare land > water bodies > grassland > shrubland > wetland > cropland > artificial surfaces. This is due to variations in the phase center of reflected electromagnetic waves caused by different vegetation covers, which results in significant differences in DEM errors. This finding is consistent with the conclusions of other studies (Gdulová, Marešová, and Moudrý 2020; Shetty et al. 2021). Moreover, InSAR-based DEMs exhibit higher accuracy across different land cover types and significantly outperform optical products. This implies that DEMs generated using InSAR technology perform better under various land cover types, which is of great practical significance for terrain analysis and precise measurements.

4.4. Effect of slope on DEM accuracy

The slope is a topographical characteristic that quantifies the extent of terrain inclination, indicating the magnitude of changes in elevation across horizontal distances (Varga and Bašić 2015). Geometric distortions caused by optical photogrammetry and InSAR-related imaging processing have been identified as obstacles to obtaining accurate terrain information (Z. Liu et al. 2020; Ludwig and Schneider 2006; Sun et al. 2003). These distortions, coupled with variations in slope, further contribute to the degradation of DEM accuracy.

The findings depicted in Figure 7 demonstrate a distinct correlation between accuracy and the slope of the terrain. By conducting an analysis of ME, SD, MAE, and RMSE across various slope ranges, a notable escalation in errors is observed with an increase in slope. The aforementioned statement highlights the significance of slope as a crucial determinant that directly impacts the accuracy and reliability of DEM products acquired using optical and InSAR technologies.

Figure 7. Differences between GDEM5 and ICESat-2 ATL08, ME, SD, MAE and RMSE calculated according to different slopes are shown as dotted line plots: (a) ALOS, (b) ASTER (c) COP (d) NASA (e) TDX90.

To gain a more detailed understanding of how GDEM5 responds to terrain slope and its impact on terrain representation, Figure 8 presents box plots and hexagonal scatter plots of DEM errors based on slope. The expanded range of box plots and the increase in MAE indicate the substantial influence of slope on the vertical accuracy of DEMs. The scatter plots in Figure 8 provide a clearer visualization of error dispersion, with ASTER and TDX90 exhibiting larger error spreads. Conversely, ALOS, NASA, and COP demonstrate greater stability across various slope conditions.

Figure 8. Differences between GDEM5 and ICESat-2 ATL08 according to box line plots (at 5° intervals) and hexagonal scatter plots (red diamonds in the figure indicate MAE) for different slopes: (a) ALOS, (b) ASTER (c) COP (d) NASA (e) TDX90.

Specifically, TDX90 exhibits the highest SD of 46.33 m, indicating the greatest level of uncertainty within this dataset. Furthermore, all DEMs exhibit a significant increase in MAE values. Among them, ALOS demonstrates the lowest MAE values across all slope ranges, with an average of 5.14 m. This indicates excellent accuracy in various slope conditions, which can be attributed to the favorable slope invariance of the high-resolution data. NASA and COP exhibit similar accuracy across various slope ranges, with average MAEs of 7.02 m and 8.79 m, respectively. Overall, ASTER outperforms TDX90 in terms of accuracy, with average MAEs of 11.17 m and 17.14 m, respectively. It is worth noting that TDX90 demonstrates higher accuracy in small slope ranges (<15°), even surpassing ASTER.

However, with an increasing slope (>15°), its quality rapidly declines. Hence, in areas with gentle slopes, apart from ASTER, other DEM datasets can be used as alternatives for high-quality DEMs. Moreover, the TanDEM-X product with a resolution of 90 m accurately represents mountainous terrain features. This finding underscores that, when compared to traditional optical photogrammetry products, InSAR products demonstrate higher accuracy and reliability in slope-changing conditions. These results align with findings from other researchers (M. Li et al. 2023; Uuemaa et al. 2020).

4.5. Influence of slope direction on DEM accuracy

Aspect, a crucial element of terrain characteristics, signifies the direction and steepness of the Earth's surface.

Figure 9 illustrates a rose diagram that showcases the fluctuation of elevation errors in GDEM5 in relation to aspect. The findings from the visual analysis indicate that the impact of aspect on the accuracy of DEMs is relatively insignificant. Additionally, hexagonal scatter plots is employed to visually depict the variability of GDEM5 accuracy in relation to aspect, thereby offering an intuitive portrayal of error dispersion. The range of MAE values in the scatter plot reflects the consistency of DEM accuracy across different aspects, with a larger range indicating lower consistency.

Figure 9. Differences between GDEM5 and ICESat-2 ATL08 according to radial and hexagonal scatter plots for different slope directions (red diamonds in the figure indicate MAE): (a) ALOS, (b) ASTER (c) COP (d) NASA (e) TDX90.

From the findings depicted in Figure 9, it can be observed that DEM errors are not significantly affected by aspect and remain relatively stable. ALOS, COP, and NASA exhibit consistent error distributions in all directions. For ALOS, the maximum and minimum MAE values are found in the northwest direction (MAE = 3.16 m) and the due south direction (MAE = 2.04 m), respectively. COP's maximum and minimum values are observed in the due west direction (MAE = 5.32 m) and due south direction (MAE = 2.74 m), respectively. NASA's maximum and minimum values occur in the northwest direction (MAE = 3.74 m) and due south direction (MAE = 1.93 m), respectively. Next is the TDX90 dataset, which has the highest MAE in the Northwest direction (MAE = 8.00 m) and the lowest MAE in the Southwest direction (MAE = 3.78 m). Compared to other DEMs, ASTER displays the most dispersed error distribution and exhibits the largest MAE in all eight directions, with an average MAE of 7.62 m. This indicates the strongest influence of aspect, lowest stability, and poorest quality.

In general, the error distribution of GDEM5 on the Qinghai-Tibet Plateau exhibits a trend of ‘low in the Southwest direction, high in all other directions’. The relationship between aspect and DEM accuracy may be influenced by the topography and geomorphology of the Qinghai-Tibet Plateau. It could also be influenced by factors such as data acquisition time, satellite sensor heading in ascending and descending orbits, and the radar's angle of incidence to the ground in the InSAR sensor (Jiang et al. 2020). While the impact of aspect may be relatively small, it is still necessary to consider aspect factors in specific geographical contexts to ensure the accuracy of DEMs.

5. Discussion and conclusion

5.1. Discussion

5.1.1. Overall accuracy of ICESat-2 and GDEM5

ICESat-2 data offers a unique advantage in verifying DEM accuracy, a claim supported by numerous researchers. For instance, (Neuenschwander and Pitts 2019) provided ATLAS data and verified the accuracy of ground elevation, demonstrating that ICESat-2 ATL08 data shows elevation accuracy within approximately 0.70 m. Similarly, other researchers, such as (Carabajal and Boy 2020), have come to similar conclusions, asserting that ICESat-2 ATL08 data possesses sufficient accuracy for assessing DEM accuracy. Additionally, this study removed GCP outliers and selected data with higher accuracy to ensure the accuracy of the assessment. In the end, 248,754 GCP points were selected, which helped to overcome the limitations of a small sample size and reduce the impact of sparse sample points within the study area.

In the Qinghai-Tibet Plateau region, ALOS data provides the most robust and accurate DEM with an RMSE of 5.05 m. NASA and COP follow, with RMSE values of 6.23 m and 8.10 m, respectively. NASA's overall accuracy slightly surpasses that of COP. These findings may differ from those of some researchers, such as (Guth and Geoffroy 2021; Marešová et al. 2021), and (Ghannadi et al. 2023). However, there are also supporting conclusions from other researchers. For instance, studies by (Carrera-Hernández 2021) and (Adiri et al. 2022) suggest that NASA's DEM represents an improvement over SRTM V3, particularly in flat terrains. It even offers better vertical accuracy than AW3D30 DSM. Furthermore, in steep mountainous areas, NASA outperforms COP in depicting terrain details, in line with the conclusions of (Hui Li et al. 2022). Similarly, (W. Chen et al. 2022) and other researchers have validated the performance of glacier topography and ice thickness inversion models for the entire Tibetan Plateau. Their findings demonstrate that NASA data achieves the highest accuracy and provides a more accurate representation of the terrain features of the Qinghai-Tibet Plateau. In the Nepalese region, researchers (Maharjan, Bhattarai, and Bhattarai 2022) used a hydro-meteorological model and found that NASA performed better than COP, with RMSE values of 14.1 m and 16.4 m, respectively. (Tran et al. 2023) arrived at the same conclusion, considering NASA as the most accurate DEM. Therefore, these differences in critical performance might largely be attributed to the heterogeneous features of the study area, such as steep mountainous terrain, humid forested areas, and variations in land use. Additionally, variations in the timing of reference data collection and the introduction of noise or ‘speckle’ in the results could also contribute to these differences. Further research is required to determine whether NASA or COP data offers higher accuracy and performance in accurately representing the various terrain features of the high-altitude mountainous regions of the Qinghai-Tibet Plateau.

Regarding ASTER and TDX90 data, their RMSE values are very close, at 11.47 m and 12.32 m, respectively. However, ASTER has slightly higher overall accuracy. This result can be attributed to the influence of the two datasets having different spatial resolutions. However, in specific areas with certain land cover types and slopes below 15°, TDX90 shows higher vertical accuracy than ASTER in multiple instances.

5.1.2. Effect of elevation and land cover on accuracy

The relationship between DEM errors and elevation is closely linked because the DEM surface has limitations in accurately representing the rugged terrain of high-altitude mountains. With increasing elevation, the RMSE gradually increases, indicating that the accuracy of measurements is hindered at higher elevations. The high-resolution ALOS data exhibits an average RMSE of 5.08 m across all elevation ranges, making it widely applicable. For 30 m resolution DEMs, NASA presents an average RMSE of 6.10 m and should be considered the preferred choice. Following NASA, COP and ASTER have average RMSE values of 8.35 m and 10.09 m, respectively. It's worth noting that TDX90 has an average RMSE of 9.96 m, indicating superior vertical accuracy compared to ASTER in areas below 4000 m in elevation.

Land cover affects the accuracy of GDEM, which is influenced by the production technology of the product. The evaluation results reveal the average RMSE rankings as follows: ALOS (5.93 m) < NASA (6.82 m) < COP (7.83 m), TDX90 (11.28 m) < ASTER (11.40 m). ALOS maintains a leading position in this domain due to its advantage in high-resolution capabilities. In the 30 m resolution dataset, NASA is more suitable for the study area than COP. ASTER accuracy is significantly influenced by land cover types, while TDX90 exhibits better accuracy than ASTER in specific land cover types such as cropland, shrubland, wetland, artificial surfaces, and bare land.

Scholars such as (Varga and Bašić 2015) have observed that areas with forest cover significantly impact the accuracy of all DEMs, especially those generated through optical techniques. This is because optical techniques are limited in their ability to penetrate vegetative canopies to obtain true surface elevation, particularly in rugged terrain and glacier accumulation areas. Therefore, in future work, accurately correcting DEM biases caused by vegetative canopies and exploring data obtained by different technologies are crucial for achieving precise elevation measurements.

5.1.3. Influence of slope and slope direction on accuracy

Studies have indicated that slope is a significant factor that affects the accuracy of DEMs. Vertical accuracy decreases significantly as the slope angle increases. In this regard, high-resolution ALOS is the least affected by slope and aspect. In the 30 m resolution DEM, NASA is the preferred choice, followed by COP. It's worth noting that TDX90 reaffirms the superior accuracy of InSAR-derived DEMs compared to optical photogrammetry. For slopes less than 15°, TDX90 exhibits higher accuracy than ASTER. However, as the slope increases, the RMSE also increases linearly, leading to a significant decrease in accuracy. Therefore, slope significantly impacts both optical photogrammetry and InSAR-derived DEM errors. When considering resolution, InSAR-derived DEM products (COP, NASA, and TDX90) are less affected by slope compared to optical products (ALOS and ASTER). Among these five DEM products, ASTER is the most affected by aspect, displaying a pronounced error dispersion. However, research indicates that aspect does not have a significant impact on DEM. In the study of the Qinghai-Tibet Plateau, the error distribution with respect to aspect exhibits a pattern of ‘low in the Southwest direction, high in all other directions’.

5.1.4. Effect of spatial resolution

Spatial resolution plays a critical role in capturing variations in terrain, as lower spatial resolutions lead to corresponding alterations in pixel values. Therefore, a higher spatial resolution is beneficial for accurately depicting terrain features (Sharma and Bagri 2023; Uuemaa et al. 2020). Our research findings are consistent with prior studies, suggesting that increased spatial resolution plays a crucial role in mitigating the impact of elevation, land cover, slope, and aspect on the accuracy of DEMs. This advantage allows for a more comprehensive representation of terrain features.

Among the DEM data we selected, the ALOS data with a spatial resolution of 12.5 m effectively overcomes the impact of terrain features and land cover, making it a top choice. In the 30 m DEM category, NASA is the preferred choice, followed by COP and ASTER. Although ASTER has a spatial resolution of 30 m, the evaluation results suggest that its elevation information provided in the Qinghai-Tibet Plateau region is relatively low quality. Especially in low-elevation areas and land cover types such as cropland, shrubland, wetland, artificial surfaces, and bare land, the elevation accuracy of ASTER is even lower than that of TDX90. This is consistent with the findings of researchers such as (DeWitt, Warner, and Conley 2015) and (Uuemaa et al. 2020). Furthermore, ASTER may also suffer from artifacts caused by cloud cover, mismatches between different scenes, and processing techniques, which are its major drawbacks (González-Moradas and Viveen 2020; Uuemaa et al. 2020). Overall, while ASTER's performance is slightly superior to TDX90, the accuracy of TDX90 is comparable to that of ASTER. This can be explained by the low spatial resolution of TDX90, which fails to capture changes within larger pixels, resulting in significant errors in elevation measurements. This conclusion aligns with the findings of (L. Feng and Muller 2016; Z. Liu et al. 2020; Uuemaa et al. 2020). Therefore, considering that TDX90 adopts novel measurement techniques, it holds promise as a new standard for global DEMs in terms of geometric resolution, accuracy, and the ability to depict complex terrains (Rizzoli et al. 2017; Zink et al. 2014). Future research can further examine the accuracy and applicability of TDX90 products, taking into account data calibration in specific regions such as the Qinghai-Tibet Plateau.

5.1.5. Limitations and recommendations

While this study has conducted accuracy assessments on five GDEM datasets, it primarily focuses on vertical accuracy, with limited consideration of the potential impact of horizontal displacements. Research indicates that there is a significant sine-cosine relationship between DEM errors and terrain (slope) when there is no data registration due to horizontal offsets between different datasets (Grohmann 2018; Nuth and Kääb 2011). To address this issue, data registration would be an effective and practical method, which has been validated in related research (C. Chen, Yang, and Li 2020; Hui Li, Deng, and Wang 2017; Ravanbakhsh and Fraser 2013).

In this study, the five selected GDEM datasets differ in their production timing from the reference data ICEsat-2. For instance, the NASADEM data were collected at a very short interval, whereas COPERNICUS took several years to complete its data collection, and the TanDEM-X data were collected in 2015. Therefore, there is a temporal disparity between these datasets, which could potentially result in variations in elevation in specific regions, particularly in areas with high dynamics like ice and snow, water bodies, wetlands, and urban areas. This could impact the assessment results (Hui Li et al. 2022).

The five GDEMs differ in spatial resolution. Therefore, when selecting an appropriate GDEM, it is important to carefully consider the size of the study area and the research purpose. The high-resolution ALOS PALSAR dataset is more suitable for applications in fine-scale terrain or hydrological features. In 30 m resolution DEMs, NASADEM is the preferred option, followed by COPERNICUS. The ASTER GDEM V3 is not recommended as the top choice data. TanDEM-X exhibits similar overall accuracy to ASTER GDEM V3 and shows promise as a significant alternative for future elevation data research.

Additionally, we did not conduct detailed zone-specific discussions about the specific terrain features of the Qinghai-Tibet Plateau, such as glaciers and permanent snow areas. In these specific areas, the ICEsat-2's ATL06 data product may be considered, as it provides precise elevation information for land glaciers. Research has shown that the ATL06 product achieves elevation accuracy of up to 5 cm in Antarctica, with surface measurement precision reaching 20 cm (Brunt et al. 2021; Zhang et al. 2021). Therefore, future research may consider utilizing this data to verify the accuracy of glacier and permanent snow area elevations in the Qinghai-Tibet Plateau. Additionally, this data can be used for applications related to ice thickness inversion.

5.2. Conclusion

Through a comprehensive evaluation of five DEM products in the Qinghai-Tibet Plateau, this study analyzed and compared their accuracy by considering multiple indicators (ME, SD, MAE, RMSE) and multiple factors (elevation, land cover, slope, and aspect).

The research results indicate that different DEMs with varying spatial resolutions offer distinct advantages in providing detailed information about terrain. The high-resolution ALOS PALSAR product achieved the lowest RMSE of 5.05 m, making it the most robust and accurate DEM for all surface features. This is because finer spatial scales make it easier to capture terrain details and are more robust against various terrain factors. In the case of 30-meter resolution DEMs, the accuracy ranking is as follows: NASADEM, COPERNICUS, and ASTER GDEM V3, with RMSE values of 6.23 m, 8.10 m, and 11.47 m, respectively. The 90 m resolution TanDEM-X product exhibited an RMSE of 12.32 m. Elevation and land cover have a significant impact on the accuracy of DEMs. Among these factors, ALOS PALSAR performed the best, followed by NASADEM and COPERNICUS. TanDEM-X demonstrated higher accuracy than ASTER GDEM V3 in land cover types such as cropland, shrubland, wetland, and artificial surfaces. Overall, the accuracy of all DEMs decreases as the land cover transitions from open terrain to rugged mountainous terrain and from low vegetation to high vegetation. Artificial surfaces exhibited the lowest RMSE (1.74 m), whereas forested areas had the highest RMSE (25.45 m).

Additionally, water bodies, glaciers, or permanent snow caused an overestimation of elevations in all DEMs. The dependence of DEM accuracy on slope is pronounced. As slope increases, the RMSE of all DEM products significantly rises, resulting in decreased accuracy. This provides a reference for future correction of systematic errors caused by terrain slope, especially when the slope is greater than 15°, where the 90 m TanDEM-X DEM exhibits the lowest RMSE. However, when the slope is less than 15°, the 30 m ASTER GDEM V3 displayed lower accuracy than the 90 m TanDEM-X DEM. The variation in aspect did not significantly affect DEM errors, and in the Qinghai-Tibet Plateau, the overall error distribution trend exhibited ‘low in the Southwest direction, high in all other directions’.

This study further emphasizes the effectiveness of ICEsat-2 data as a robust instrument for assessing GDEMs, providing valuable references for selecting more accurate and suitable DEMs in the Qinghai-Tibet Plateau and for future corrections of DEMs under different terrains.

Acknowledgements

The authors would like to thank three anonymous reviewers for their constructive comments and suggestions, which greatly improved the quality of the manuscript.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

All the data used in this study are for free and publicly available.

Additional information

Funding

This work was supported by the Second Qinghai-Tibet Plateau Scientific Research Program under Grant [grant number 2019QZKK0307] and the National Key Research and Development Program of China under Grant [Project No. 2019YFE0115200].