Accuracy Assessment of a Digital Elevation Model Constructed Using the KOMPSAT-5 Dataset

Abstract

The Interferometric Synthetic Aperture Radar (InSAR) has significantly advanced in its usage for analyzing surface information such as displacement or elevation. In this study, we evaluated a digital elevation model (DEM) constructed using X-band KOMPSAT-5 interferometric datasets provided by the Korea Aerospace Research Institute (KARI). The 28-day revisit cycle of KOMPSAT-5 poses challenges in maintaining interferometric correlation. To address this, four KOMPSAT-5 images were employed in a multi-baseline interferometric approach to mitigate temporal decorrelation effects. Despite the slightly longer temporal baselines, the analysis revealed sufficient coherence (>0.8) in three interferograms. The height of ambiguity ranged from 59 to 74 m, which is a moderate height of sensitivity to extract topography over the study area of San Francisco in the USA. Unfortunately, only ascending acquisition mode datasets were available for this study. The derived DEM was validated against three reference datasets: Copernicus GLO-30 DEM, ICESat-2, and GEDI altimetry. A high coefficient of determination (R² > 0.9) demonstrates the feasibility of the interferometric application of KOMPSAT-5.

1. Introduction

The digital elevation model (DEM), a quantitative representation of surface elevation, can be utilized in various research fields, such as urban development management, geological or hydrological studies, and natural disaster monitoring [1]. DEMs are crucial resources for monitoring dynamic environments, such as coastal areas or intertidal flats, which are considerably affected by global climate change. Although terrestrial leveling field surveys [2], light detection, and ranging measurements using unmanned aerial vehicles or drones [3] have successfully constructed very high spatial resolution and accurate DEMs, they have limitations because of their narrow spatial coverage and high labor costs.

A synthetic aperture radar (SAR) transmits microwave signals from the payload and receives backscattered signals from surface targets. Compared with optical sensors, SAR, which utilizes its own signal with longer wavelengths, can be less affected by weather and day-or-night conditions [4]. One advantage of SAR observations is that the differences between two-phase observables are calculated, enabling the extraction of accurate topographic elevations and precise measurements of surface displacement using interferometric SAR (InSAR) techniques [5]. However, temporal decorrelation owing to surface changes and atmospheric conditions in repeat-pass InSAR observations often causes a loss of interferometric coherence, preventing the construction of an accurate DEM [6].

To eliminate temporal decorrelation, the Shuttle Radar Topography Mission (SRTM), designed to create the first near-global DEM using single-pass interferometry, flowed from 11–22 February 2000 [7]. A bistatic SAR flight formation with an approximate temporal baseline of TerraSAR-X and TanDEM-X belonging to the German Aerospace Center has been the only operational SAR system to generate a global DEM with very high spatial resolution [8].

We evaluated the feasibility of DEM derivation from multi-baseline repeat-pass interferometry with the KOMPSAT-5 dataset. Although a few InSAR studies using KOMPSAT-5 have been reported [9,10], research leveraging its interferometric phase remains relatively limited. Moreover, applications involving multi-baseline interferometric pairs to enhance the accuracy of KOMPSAT-5-based DEMs are even rarer. Therefore, this study focuses on the performance evaluation of InSAR DEMs derived from KOMPSAT-5 data, with the primary objective of assessing the feasibility and quantitative accuracy of KOMPSAT-5 InSAR products. There are two obstacles to creating a coherent interferogram using KOMPSAT-5: (1) a slightly long revisit cycle of 28 days and (2) large orbital tubes due to intermittent orbit control. The orbit control of the KOMPSAT-5 has been tested with a repeat ground track within ±1 and ±2 km to assess the InSAR capability [11]. To mitigate temporal decorrelation from large temporal baselines, a multi-baseline InSAR approach can suppress randomly distributed phase components [12,13].

The study area and dataset used in this study are described in Section 2. The data processing used to create the DEM is presented in Section 3. Section 4 evaluates the constructed DEM using three datasets: Copernicus GLO-30 DEM, Ice, Cloud, and Land Elevation Satellite-2 (ICESat-2), and a Global Ecosystem Dynamics Investigation (GEDI). Finally, concluding remarks and discussion are presented in Section 5.

2. Materials

2.1. Study Area

This study aimed to demonstrate the interferometric capability of KOMPSAT-5 and assess the accuracy of its DEMs using multiple datasets. The test area was San Francisco on the northern coast of California, USA (Figure 1). This region is an international testbed site for evaluating SAR data processing and performance [14]. Figure 1a illustrates the area of interest (AoI) and the tracks of the datasets used in this study. Figure 1b depicts the topographic features of the AoI, with 50 m contour intervals overlaid on the 2022 land cover map provided freely by Esri Inc. The southwestern part of San Francisco is characterized by mountainous and vegetated terrain, whereas the northern part is relatively flat.

2.2. Datasets

KOMPSAT-5, developed by the Korea Aerospace Research Institute (KARI), is the first Korean SAR satellite equipped with an X-band sensor operating at a radar carrier frequency of 9.66 GHz (a wavelength of 3.1 cm). The GOLDEN mission was designed for Geographical Information Systems (GISs), ocean monitoring, land management, disaster monitoring, and environmental monitoring. The Corea SAR Instrument (COSI) payload supports the primary mission by providing three high-resolution mode SAR images: (1) a 1 m high-resolution spotlight mode with a 5 km swath, (2) a 3 m resolution standard strip map mode with a 30 km swath, and (3) a 20 m resolution-wide ScanSAR mode with a 100 km swath [15,16]. All acquisition modes of KOMPSAT-5 use the selection of single polarization among the full-polarization options. The revisit cycle of KOMPSAT-5 is 28 days, which is a relatively long time span for radar interferometry, similar to that of other optical KOMPSAT satellite series.

We collected four KOMPSAT-5 Stripmap images from November 6, 2021, to February 26, 2022. The temporal baselines of each interferometric pair were 56, 84, and 112 d, respectively, which may be quite long for InSAR applications. The difference in Doppler frequency was estimated to range from 2.1 to 16.4 Hz. The characteristics of the datasets are listed in

Table 1. Characteristics of the dataset.

Acquisition Parameter
Acquisition mode	Stripmap
Polarization	HH
Carrier frequency	X-band (9.66 GHz)
Incidence angle	30.27°
Pulse repetition frequency	3426.23 Hz
ADC sampling rate	1543.75 MHz
Azimuth pixel spacing	2.05 m
Range pixel spacing	0.97 m

.

To process the interferometric data and estimate the accuracy of the constructed DEM, we collected four datasets: SRTM DEM; Copernicus GLO-30 DEM; Ice, Cloud, Land Elevation Satellite-2 (ICESat-2); and Global Ecosystem Dynamics Investigation (GEDI). The SRTM DEM covers an area from 60°N to 60°S, with a vertical accuracy ranging from 4.56 to 7.10 m. The SRTM was created using single-pass interferometric data with X- and C-band frequencies acquired in February 2000 and post-processed to avoid data anomalies through void filling with a water body mask. We utilized the SRTM DEM to subtract the topographic phase and reduce the phase unwrapping errors.

The Copernicus DEM is a high-precision digital surface model that includes information on buildings and vegetation based on WorldDEM data acquired from the TanDEM-X mission [17]. The TanDEM-X mission utilized simultaneous bistatic SAR observations collected between 2011 and 2015, supported by the German Aerospace Center and Airbus Defense and Space. This study used globally distributed GLO-30 data from the EEA-10m, GLO-30, and GLO-90 datasets. The Copernicus DEM provides high accuracy in both vertical and horizontal dimensions. The absolute vertical and horizontal accuracies of the Copernicus DEM were approximately 4 and 6 m, respectively. Additionally, the relative vertical accuracy varied depending on the terrain slope: it was better than 2 m for slopes ≤ 20% (≈11.31°) and better than 4 m for slopes > 20% within a 1° × 1° area. These accuracy specifications ensured the dataset’s reliability across diverse landscapes and elevation conditions [18]. The Copernicus DEM was primarily used to assess how well the derived DEM represents continuous topographic trends and to evaluate the spatial distribution of elevation concerning complex terrain and land cover types.

The National Aeronautics and Space Administration (NASA) launched ICESat-2 on 16 September 2018 to monitor changes in cryosphere ice sheets and glaciers, providing altitude information over oceans, land, and vegetation with centimeter-level precision [19]. ICESat-2 had a Ku-band (13.575 GHz) Advanced Topographic Laser Altimeter System (ATLAS) that emitted six laser beams per pulse. The beams of a single pulse were divided into strong and weak beams, with the strong beam delivering four times the energy of the weak beam [20]. We collected ATL08 strong beam data acquired on 19 July 2022 and 17 January 2023 to validate the results (yellow lines in Figure 1). A total of 1400 elevation points were used to calculate the correlation between the ICESat-2 measurements and the derived DEM, enabling a statistical accuracy assessment.

The GEDI mission, launched by NASA on 5 December 2018, provides high-resolution laser-ranging observations of the Earth’s forests and topography. Its primary objective was to characterize the structure and dynamics of ecosystems. We incorporated the GEDI Level 2A elevation and height metric data global footprint level V002, which offers precise ground elevation measurements beneath the forest canopy and other height metrics with a vertical accuracy of 10 cm [21]. For validation, 9700 samples were collected from 10 tracks between September 2021 and January 2023 (black dotted line in Figure 1). The GEDI data were evenly distributed across the study area, minimizing potential bias in the validation process. In total, 9600 points were randomly sampled to ensure a robust evaluation and statistical correlation analysis was performed to validate the results. Additionally, the spatial distribution of height differences was statistically analyzed using this dataset to better understand the variability of residual errors across the study area.

Although the Global Navigation Satellite System (GNSS) provides highly accurate elevation data, its sparse distribution in the study area limits its applicability to comprehensively validating the KOMPSAT-5 DEM. Unlike GNSS, which includes elevation data at discrete locations, the Copernicus DEM, ICESat-2, and GEDI datasets offer broader spatial coverage, making them more suitable for evaluating large-scale topographic variations. Combining these datasets enabled a more robust assessment of the KOMPSAT-5 DEM.

3. Methodology

Our data-processing scheme for constructing a DEM using multi-baseline interferometric pairs assumed the absence of significant surface displacement, meaning that the constructed interferograms reflect the topographic height. We processed four KOMPSAT-5 SAR acquisitions to obtain interferograms using a commercial GAMMA software package [22]. Figure 2 illustrates a block diagram of data processing. To ensure precise DEM generation, we employed a method that compensated for the differential phase calculated from multi-baseline interferometric pairs and an external DEM to simulate the topographic phase of the external DEM. This approach is a standard technique for deriving an improved interferometric DEM. However, the key contribution of this study lies in utilizing the interferometric phase of KOMPSAT-5 to generate and evaluate the DEM.

Before InSAR processing, we used the SRTM 30 m DEM as an external DEM for topographic phase simulation and initial elevation correction. Since SRTM data are referenced to orthometric height, we converted them to WGS84 ellipsoidal height by adding geoid heights. The external DEM was then cropped to match the KOMPSAT-5 single-look complex (SLC) data coverage. Using SAR acquisition parameters, we simulated the topographic phase. To ensure consistency with the interferogram resolution, we oversampled the data to achieve an approximate spatial resolution of 6 m in both the latitude and longitude directions. This process facilitated the generation of an initial lookup table for conversion into range-Doppler geometry.

The first step in KOMPSAT-5 InSAR processing is precise co-registration, a critical procedure in InSAR data processing. Although the geometric baselines of the selected interferometric pairs were relatively small, we performed a coarse-to-fine co-registration approach to account for local terrain variations. The SLC data acquired on 6 November 2021 were selected as the reference image, and intensity cross-correlation was employed to achieve precise sub-pixel alignment. A multi-looking process with a 4 × 4 window was applied to estimate the correlation between the intensity data and radar-coded DEM. A fourth-order Lanczos interpolator was used to resample secondary data for a selected reference scene. After using a multi-looking tool to suppress interferometric noise, the azimuth and range spacings of the created interferograms were 8 m and 4 m, respectively. Spectral adaptive phase-shift filtering in both the range and azimuth directions was used to enhance the interferograms [23], and coherence was computed using a 5 × 5 window to evaluate the quality of the interferograms.

During this process, ghost signals were detected around the top of the scene caused by aliasing artifacts of the Doppler phase history in the target, as indicated by the yellow polygon in Figure 3a. These artifacts arise from repeatedly acquiring strong backscattered signals from a moving platform. In the dataset used in this study, ghost signals were predominantly observed over ocean areas. Given that the primary objective of this study is to extract topographic height values for inland regions accurately, we determined that signals from the ocean could potentially introduce phase unwrapping errors in the inland areas. Therefore, to mitigate these artifacts, we utilized water body data from the SRTM 1-arc second DEM to mask the ghost signals in ocean areas. Phase unwrapping errors often occur on steep slopes in mountainous regions, preventing accurate topographic height information extraction. Although the height of ambiguity (HoA) of the interferometric pairs was not too small, the topographic phase from the SRTM 1-arc second DEM was subtracted to minimize these unwrapping errors. The residual phases from inaccurate orbit information were removed by estimating a polynomial phase model (a0 + a1∗y + a2∗x: x and y are image coordinates in the range and azimuth direction). The minimum cost flow (MCF) algorithm based on Delaunay triangulation was utilized, incorporating both the coherence threshold (>0.6) and the water body mask. During the phase-unwrapping process, interpolation with a 25-pixel radius window was conducted to fill the void pixels masked by the coherence threshold.

We generated six initial interferograms with various perpendicular baselines, ranging from −3 to −86 m (

Table 2. Baseline and HoA of the interferograms.

Interferometric Pairs	Temp. Baseline (days)	Perp. Baseline (m)	HoA (m)
6 November 2021–1 January 2022	56	−68.6	−74.7
6 November 2021–29 January 2022	84	−86.8	−59.0
6 November 2021–26 February 2022	112	−72.1	−71.1
1 January 2022–29 January 2022	28	−18.2	−272.7
1 January 2022–26 February 2022	56	−3.6	−1392.2
29 January 2022–26 February 2022	28	14.7	339.5

). The perpendicular baselines of the interferograms used in this study ranged from −68.6 to −86.8 m, providing a moderate HoA of 59.0 to 74.7 m. HoA was inversely proportional to the perpendicular baseline, which is a factor that determines the height sensitivity according to the slope degree. However, interferometric pairs with very short baselines (below 20 m) can cause a significantly large HoA and are affected by severe atmospheric artifacts. Therefore, interferometric pairs were excluded from the DEM generation process. In the height conversion step, only three interferograms were used. The height maps derived from individual interferograms were averaged to mitigate randomly distributed errors, as height estimation errors may vary depending on different observation geometries.

In the last step, precise geocoding was applied using a refined lookup table, considering the updated height map [24]. Since the magnitude of height error correction was minimal, the initial lookup table generated using the SRTM DEM was deemed sufficient. However, an additional refinement step was conducted to minimize further coordinate transformation errors. Finally, the KOMPSAT-5 DEM was geocoded in geographic coordinates on a WGS84 ellipsoid.

4. Results

Our dataset showed high coherence in the inland area, ranging from 0.79 to 0.85 of the three selected KOMPSAT-5 interferometric pairs (Figure 3b–d). A slight degradation in overall coherence was detected as the temporal baseline increased from day 56 to 112. Additionally, severe volume decorrelation was observed in mountainous areas with dense tree cover, varying based on land cover type (

Table 3. Coherence by land cover type.

Interferometric Pairs	Coherence by Land Cover Type
	Grassland	Trees	Urban
6 November 2021–1 January 2022	0.65	0.53	0.94
6 November 2021–29 January 2022	0.51	0.46	0.92
6 November 2021–26 February 2022	0.49	0.40	0.91

). The performance of the interferometric application was better than expected, given that the relatively large temporal baselines due to the revisit cycle of 28 days frequently inhibited the maintenance of good coherence.

Figure 4a presents the DEM constructed using multi-baseline interferometric pairs. We noticed that the topographic features of moderately sloped and mountainous regions were well preserved. However, detailed topographic height information could not be retrieved because of the intermediate range of the HoAs. To assess the performance of the constructed DEM, we calculated a difference map using the Copernicus GLO-30 DEM (Figure 4b). The difference map indicates that the constructed DEM aligns well with the Copernicus GLO-30 DEM in low-relief regions. However, we noted relatively large elevation deviations near the hilly and mountainous areas with lower coherence (the yellow circles in Figure 4b). The discrepancy in the vegetated mountainous areas might result from low coherence due to the temporal decorrelation caused by the long revisit cycle. In addition, our results relied only on data acquired in an ascending orbit. Thus, our dataset could not compensate for topographic distortions, such as the shadow effect. To compare the topographic changes in the DEMs, a profile along the A-A′ traverse line is presented in Figure 4c. The A-A′ and B-B′ transects were selected as they represent the longest cross-sections of the study area, encompassing diverse terrain slopes and land cover types. By analyzing the elevation distribution along these transects, we assessed the ability of the derived DEM to maintain a continuous topographic representation. We examined how elevation errors vary with slope and land cover characteristics. The A-A′ profile extends from mountainous regions in the southwest to moderate- and low-slope urban areas in the northeast, while the B-B′ profile traverses from the northwest to the southeast, passing through moderate- to steep-slope vegetation areas. The black and red lines represent the topographic variations in the constructed DEM and Copernicus GLO-30, respectively, showing a similar overall trend. The blue line depicts the moving average of elevation differences, revealing the fact that larger discrepancies occur primarily in high-elevation areas. A high coefficient of determination (R² = 0.99) was estimated, and the root-mean-square error (RMSE) was approximately 9.9 m and 7.84 m. Figure 4e presents a detailed analysis of the relationship between height differences, slope, and land cover types. The terrain slope along the lines was calculated and categorized into 5-degree intervals. The average height differences for each interval were analyzed based on the following land cover types: grassland, trees, and urban areas. The results indicate that in low-slope areas (<10°), height differences remain relatively small across all land cover types. However, height differences increase in moderate slope ranges (10–20°), particularly in vegetated areas with trees and grassland. In steep slopes (>20°), height differences become more pronounced, with trees exhibiting the most significant discrepancies. In contrast, urban areas do not show substantial elevation differences in steep slope regions, as such slopes are uncommon in urban environments. Slopes exceeding 25° were relatively rare along this line, resulting in fewer measurements in these ranges. By averaging all residual height values presented in Figure 4e according to land cover types, we found that urban areas had an average height difference of 5.08 m. In contrast, vegetated areas, specifically those with trees and grasslands, exhibited more significant discrepancies with mean height differences of 9.39 m and 8.52 m, respectively. These findings suggest that, compared to vegetated areas, the constructed DEM demonstrates higher accuracy in representing high-resolution topography in urban areas.

Laser measurements from ICESat-2 and GEDI were used for further validation. Along the tracks of each dataset shown in Figure 1a, approximately 1400 points from ICESat-2 and 9600 points from GEDI were used to validate the DEM across the entire study area. Figure 5 presents scatter plots comparing the constructed DEM with the two reference datasets. To examine the performance of the multi-baseline interferometric approach, we compared the DEMs from individual interferograms (Figure 5a–c) and the averaged interferogram (Figure 5d). The comparison confirmed that the multi-baseline interferometric approach provides the highest coefficient of determination (R² = 0.98) and the lowest RMSEs (6.0 m for the GEDI dataset and 8.4 m for the ICESat-2 dataset). Interestingly, topographic trends were similar at low elevations (<150 m). In comparison, an elevation offset between the two fitting results was observed, owing to the slightly different slopes at high elevations (>250 m). These results imply that KOMPSAT-5 might be weak in constructing DEMs over heavily vegetated areas owing to temporal decorrelation. Figure 5b shows the best DEM performance when using an individual interferometric pair. The temporal baseline of the data was 84 days, and the coherence was lower than that of the first interferometric pair. We suspect that determining the level of the lowest HoA may be helpful in creating a more accurate DEM. The worst correlation and the highest RMSE were detected for the first interferometric pair with the largest HoA.

Figure 6 presents the variogram and histogram analysis of the residual height differences between the GEDI dataset and the K5 DEM. Figure 6a visualizes the spatial correlation of residuals using a variogram, revealing a strong spatial correlation within a range of approximately 1500 m. This suggests that residuals between closely located points exhibit similar patterns, indicating that the DEM effectively captures local terrain characteristics. Beyond this range, the variogram flattens at approximately 80 reaching its sill, and implying that residuals become spatially uncorrelated and independent of local terrain features. Additionally, the spherical model fitted to the experimental variogram closely aligns with the observed data, confirming the reliability of the spatial correlation analysis. Meanwhile, Figure 6b illustrates the histogram of residuals between the GEDI dataset and the K5 DEM, showing a near-normal distribution centered around zero. This suggests that the DEM exhibits minimal bias and performs reliably overall. Most residuals fall within the range of −20 m to 20 m, indicating strong agreement between the two datasets. However, a few outliers beyond −40 m and 40 m have been observed, likely corresponding to regions with steep slopes or densely vegetated areas where the DEM struggles to accurately capture terrain features. This combined analysis highlights the fact that the DEM maintains high accuracy within specific distance ranges and terrain conditions, demonstrating a strong spatial correlation up to 1500 m and minimal residual bias overall.

5. Conclusions

In this study, we successfully developed a high-resolution digital elevation model (DEM) using multi-baseline interferometric pairs derived from KOMPSAT-5 datasets in San Francisco, CA, USA. The analysis utilized four KOMPSAT-5 SLC datasets, all acquired in the ascending direction. Despite the challenges posed by intermittent orbit control, the dataset’s perpendicular baseline ranged from −68.6 to −86.8 m, corresponding to a moderate height of ambiguity (HoA) between 59.0 and 74.7 m. Although the temporal baselines were relatively large due to KOMPSAT-5’s 28-day revisit cycle, high interferometric coherence (>0.79) was consistently observed across most of the study area, except in hilly or mountainous regions.

Reference datasets from ICESat-2 and GEDI laser altimetry were employed for accuracy validation. The multi-baseline interferometric approach demonstrated the highest accuracy, achieving an RMSE of 6.0 m with ICESat-2 and 8.4 m with GEDI. Notably, the individual interferometric pair with the lowest HoA yielded the best performance in DEM generation. These findings suggest that incorporating multi-baseline InSAR techniques with varying HoAs enhances the precision and reliability of the DEM.

While the results demonstrate the feasibility of KOMPSAT-5 interferometry and the high accuracy of derived DEMs, this study also indicates that coherence loss in vegetated areas can influence accuracy. The shorter wavelength of the X-band contributes to temporal decorrelation, emphasizing the need for frequent observations to maintain coherence. The development of KOMPSAT-6, the successor to KOMPSAT-5, is expected to address these limitations. KOMPSAT-6 features a significantly shorter revisit cycle of 11 days, which is likely to improve interferometric coherence and overall application performance. The anticipated advancements of KOMPSAT-6 highlight its potential to further enhance the reliability and efficiency of SAR-based DEM construction in future studies.