Analysis of GCP Chip Resolution for Automatic Accurate Sensor Modeling of Very High-Resolution Satellite Images

Abstract

Establishing accurate sensor models is an essential process for ensuring positional accuracy in remote sensing applications. This process needs to be developed further, in particular, for Very High-Resolution (VHR) satellite images due to their high level of accuracy requirements. This study analyzed the effects of the resolution of Ground Control Point(GCP) chips on automated, accurate sensor modeling for VHR satellite images. Using WorldView-3 images with a resolution of 31.6 cm, accurate sensor models were established by matching the images against GCP chips constructed from aerial images with a resolution of 25 cm and from Unmanned Aerial Vehicle (UAV) images with a resolution of 1.9-3 cm, respectively. The accuracy of sensor models from the two types of GCP chips was evaluated. Results showed that GCP chips from UAV images yielded comparable or better performance than those from aerial images. On average, accurate sensor models were automatically established with an accuracy of 1.08 pixels at the rural site and 1.57 pixels at the urban site using UAV-based GCP chips, with a statistically significant improvement over aerial chips at the rural site. Results also showed that upsampling of raw satellite images improved the accuracy of sensor models. Future research is required on the selection of an optimal upsampling scale, the use of orthorectified UAV images as GCP chips, and the feasibility of achieving sub-pixel-level accuracy. Nevertheless, this study provides an important basis for advancing automated geometric correction in VHR satellite images.

1. Introduction

With rapid advancements in satellite imaging technology and the expansion of application domains, demand for Very High-Resolution (VHR) satellite images has been increasing rapidly. Accordingly, satellite images with sub-meter level Ground Sampling Distance (GSD) are being provided. As the resolution of satellite images becomes finer, higher geometric accuracy is required. Sensor modeling is the process of establishing the geometric relationship between image coordinates and their corresponding ground coordinates. This relationship can be expressed through the Rational Function Model (RFM), derived from Rational Polynomial Coefficients (RPCs) provided by satellite vendors (Grodecki & Dial, 2003). However, since position and attitude sensors onboard satellites inherently contain geometric distortions, the initial sensor model established by the RPCs provided includes errors. These distortions mainly arise from uncertainties in orbit and attitude measurements, sensor misalignment, and other physical factors that cause discrepancies between predicted and actual image coordinates. For the effective utilization of satellite images, it is essential to perform geometric correction to remove the geometric distortions contained in raw satellite images and to establish an accurate correspondence between image coordinates and ground coordinates.

According to the internationally established CEOS Analysis Ready Data for Land (CARD4L) standards designed to improve the usability and reliability of satellite images, the residual error after geometric correction for medium-resolution satellite images must satisfy a requirement of ≤0.5 pixels radial Root Mean Square Error (rRMSE) (CEOS, 2023). However, for VHR satellite images, there are inherent technical limitations in enhancing positional accuracy to the sub-pixel level. The positional accuracy achievable for VHR satellite images with current technologies has been estimated to be within approximately 1–1.5 pixels (Choi et al., 2023). Currently, there are no explicit standard requirements for the geometric correction of VHR satellite images. Nevertheless, continuous efforts are required to achieve higher positional accuracy in VHR satellite images. In addition, automated geometric correction technologies are required to reduce cost for achieving higher positional accuracy.

In general, when establishing an accurate sensor model, Ground Control Points (GCPs) are used to refine the correspondence between image coordinates and ground coordinates and to reduce errors in the initial sensor model. To automate this process, GCP chips are defined as small image patches centered on the ground-surveyed positions of GCPs (Lee et al., 1998). GCPs are matched against raw satellite images, extracted, and an accurate geometric correction is then applied automatically. Research on the automated, accurate geometric correction using GCP chips has been actively conducted. To enhance matching performance, it was reported that constructing image pyramids and applying iterative procedures of accurate geometric correction could improve the sensor model performance (Son et al., 2021). Others proposed to improve the spatial resolution of raw satellite images using pan-sharpening techniques (Shin et al., 2018).

This paper aims to achieve fully automated, accurate geometric correction of VHR satellite images. For this, we attempt to analyze the effect of the resolution of GCP chips for matching against VHR satellite images and to check the level of accuracy achievable. The performance of GCP chip matching may be influenced by various factors. Among them, the effects of resolution difference between GCP chips and raw satellite images have been studied for medium and high-resolution satellite images (Choi & Kim, 2022; Lee & Kim, 2021; Shin et al., 2018). These studies recommended GCP chips with higher resolution than raw satellite images and demonstrated that matching performance could be improved by adjusting the resolution of GCP chips and satellite images to an optimal scale (Choi et al., 2024; Choi & Kim, 2022; Lee & Kim, 2021). Scale adjustment was also applied to match satellite images with different resolutions (Kim et al., 2022). In this paper, we extend previous studies to VHR satellite images with a Ground Sampling Distance (GSD) below 0.3 m and GCP chips with a GSD of a few centimeters.

This study draws attention to the lack of empirical analyses on geometric correction for VHR satellite images. Previous GCP chip matching methods developed for medium- to high-resolution images may not be suitable for VHR images (Han et al., 2015). This study also draws attention to the effect of GSDs of GCP chips on automated geometric correction of VHR images. Previous studies analyzed matching performance according to scale adjustment between GCP chips and raw satellite images. The impact of GSDs of GCP chips on matching performance has not been considered.

In addition to studies on geometric correction, a broader set of deep learning-based image-matching and fusion approaches has leveraged Unmanned Aerial Vehicle (UAV) images as high-resolution references to align satellite images. One line of work improved matching accuracy by adjusting both the geometric orientation and spatial resolution of UAV or aerial images to mitigate viewpoint and resolution differences with satellite images (Cheng & Xu, 2025). Furthermore, fusion studies combining UAV and Sentinel-2 images used upsampling and downsampling to alleviate resolution inconsistencies between the two sensors (Zhang et al., 2025). These studies commonly recognize the importance of resolution differences in UAV–satellite matching and apply resolution adjustments as a practical remedy. However, they do not treat resolution differences as an explicit experimental variable, nor do they quantitatively analyze how controlled changes in reference-image resolution influence matching performance.

Lastly, this study also draws attention to the use of images from UAVs as GCP chips. Previous studies have considered GCP chips generated from aerial and satellite imagery for medium- to high-resolution imagery (Choi et al., 2024; Choi & Kim, 2022; Lee & Kim, 2021; Shin et al., 2018). Since VHR images contain finer spatial information, the GSD of GCP chips needs to be finer. Therefore, we investigate whether high-resolution GCP chips from UAV imagery could effectively improve the automated geometric correction of VHR satellite imagery.

We evaluated the impact of GCP chips constructed from two types of images, one from aerial images at a GSD of 25 cm and the other from UAV images at a GSD of 1.9–3 cm, on automated geometric correction of WorldView-3 images at a GSD of 31.6 cm. The two types of GCP chips were generated at identical ground locations. The scales of the GCP chips and the WorldView-3 images were adjusted to an identical grid spacing. All GCP chips and the WorldView-3 images were resampled to the same target ground sampling distance before GCP chip matching. GCP chip matching was applied, and accurate sensor models were constructed. The chips from UAV images generally yielded better performance, while properly adjusted chips from aerial images achieved comparable accuracy. The absolute positional accuracy of the resulting orthorectified images validated the practicality of the proposed geometric correction strategy. Overall, this study demonstrates an effective method for automated, accurate geometric correction of VHR satellite images.

This paper is organized as follows. Section 2 describes the satellite image, the datasets used for constructing GCP chips, and the automated geometric correction procedure. Section 3 presents the experimental results obtained from the two study areas, and Section 4 discusses the key findings and their implications.

2. Materials and Methods

2.1. Data and Preprocessing

2.1.1. Satellite Images

In this study, for the purpose of analyzing matching performance under different environmental conditions, two areas with different land surface characteristics were selected as experiment sites. The locations of the study areas are shown in Fig. 1. Geumsan (Fig. 1a) is a rural, mountainous area, whereas Incheon (Fig. 1b) is an urban area. Geumsan represents a relatively homogeneous rural landscape with wide areas of farmland and forested slopes, where artificial structures are sparse and land-cover changes occur slowly. In contrast, Incheon is a highly built-up urban environment with dense artificial structures, frequent changes, and more complex texture patterns, requiring more frequent updates of GCP chip information. These two sites were selected specifically to represent contrasting land surface characteristics for analyzing matching performance, and were also chosen as areas where UAV flight operations are relatively unrestricted.

Fig. 1. Location map of the study areas in the Republic of Korea: (a) Geumsan and (b) Incheon.

Panchromatic bands of WorldView-3 images with a GSD of 0.316 m were used for experiments. One image per study area was prepared. The images used corresponded to Level 1B Basic Product, which were raw images with radiometric correction only and without geometric correction. The list of satellite images used in this study is provided in Table 1 and in Fig. 2.

Table 1. List of satellite images used in this study

Fig. 2. WorldView-3 images covering the study areas: (a) Geumsan and (b) Incheon.

2.1.2. Aerial Orthoimages

To construct GCP chips from aerial images, aerial orthoimages were utilized. The aerial orthoimages used in this study were provided by the National Geographic Information Institute (NGII) at a GSD of 25 cm. These aerial images had already been orthorectified and color-corrected to produce nationwide orthoimages. Image patches were extracted directly without additional preprocessing to construct GCP chips. The image coordinates of the center pixels of the extracted patches were assigned ground coordinates provided by NGII, and these assigned coordinates were used as the GCPs. The ground coordinates assigned to the GCP chips were obtained from NGII’s Unified Control Points (UCPs), which constitute a nationwide high-accuracy geodetic reference network. According to NGII specifications, the absolute positional accuracy of these control points is within 30 mm in planimetry and within 2.5 mm × √S in height, where S denotes the one-way observation distance in kilometers. Examples of the aerial orthoimages are shown in Fig. 3.

Fig. 3. Examples of aerial orthoimages.

2.1.3. UAV Images

This study required UAV images to construct GCP chips with a resolution much finer than that of VHR satellite images. UAV images were acquired over the Geumsan and Incheon study areas using a DJI Phantom 4 Real Time Kinematic (RTK) with an FC6310R optical camera. Flights were conducted at each location where GCP chips were constructed from the aerial orthoimages.

The UAV ascended to a predefined altitude and captured nadir-view images. For Geumsan, the flight altitude was set to 110 m, and UAV images were acquired at a GSD of 3 cm. For Incheon, the flight altitude was set to 69 m, and UAV images were acquired at a GSD of 1.9 cm. These differences in flight altitude were due to UAV flight regulations.

In addition to NGII’s UCPs, several ground points were surveyed in the field using a Global Navigation Satellite System (GNSS) receiver (RTAP2U, specifications in Table 3). Real-Time Correction Messages (RTCM) were provided via the NGII’s Virtual Reference Station (VRS) service, enabling network-based RTK positioning (Kim et al., 2021). To verify the reliability of the GNSS measurements, selected UCPs were independently resurveyed and compared with the official NGII coordinates. The GNSS-surveyed points differed by less than 3.6 mm in horizontal position and matched the UCP elevation values up to the second decimal place, confirming equivalent geodetic accuracy.

UAV flight information is presented in Table 2, and the specifications of the GNSS receiver are summarized in Table 3. Examples of UAV images from Geumsan and Incheon are shown in Fig. 4. A comparative view of the same region using satellite, aerial, and UAV images is provided in Fig. 5.

Table 2. Specifications of the UAV used for the survey (Phantom 4 RTK)

Table 3. Specifications of the GNSS receiver (RTAP2U)

Fig. 4. Examples of UAV images.

Fig. 5. Comparison of satellite, aerial, and UAV images.

2.1.4. UAV Image Preprocessing for GCP Chip Generation

UAV images contain exterior orientation parameters (roll, pitch, and yaw) derived from the onboard Inertial Measurement Unit (IMU), as well as geographic coordinates (latitude and longitude) measured by the RTK-GNSS sensor. In this study, we utilized the metadata to align UAV images to the Transverse Mercator (TM) coordinate system. Based on this alignment, base images were established to generate GCP chips. The UAV imagery was TM-aligned but not orthorectified, since the study areas were selected to be sufficiently flat that relief displacement is negligible within the 45×45 m chip footprint. To achieve this, we introduced a rotation transformation R_img^obj, which defines the mapping between the image coordinate system and the object coordinate system. This transformation represents the camera sensor model that governs how each pixel in the image is projected onto the ground. Specifically, it is constructed as a sequential product of rotation matrices that describe the relationships among the image, body, navigation, Earth-Centered and Earth-fixed (ECEF), and TM reference frames.

\(\begin{align}R_{i m g}^{o b j}=R_{N 0}^{o b j} \cdot R_{E C E F}^{N 0} \cdot R_{N a v i}^{E C E F} \cdot R_{b o d y}^{\text {Navi }} \cdot R_{i m g}^{b o d y}\end{align}\)       (1)

In Eq. (1), each rotation matrix describes the coordinate transformation between successive reference frames. R_img^body converts image coordinate system to the UAV body frame, while R_body^Navi incorporates roll, pitch, and yaw angles to express the platform’s orientation in the navigation frame. R_Navi^ECEF represents the mapping between the ECEF system and the local navigation system defined at the UAV position. R_ECEF^N0 provides the transformation from the ECEF system to the navigation frame defined at the map projection center. Finally, R_N0^obj connects this reference frame to the object space, defined by the TM projection coordinates.

Once R_img^obj is established, the transformation is applied to the four corner pixels of the UAV image. These corner pixels are then used to calculate their projected ground positions in the TM coordinate system. These projected corner points define the image footprint. Based on the footprint, a regular grid aligned to the true north direction is constructed with the target GSD. UAV image is then resampled onto this grid, thereby producing a north-aligned, TM-rectified image. Through the above procedure, north-aligned and resampled UAV images were prepared. Subsequently, ground coordinates obtained from the GNSS receiver (RTAP2U) were assigned to the grid points identified as the chip center. GCP chips were cropped from these resampled images. Fig. 6 illustrates the overall process of generating chip images from the original UAV image.

Fig. 6. UAV image preprocessing steps for GCP chip generation: (a) original UAV image before yaw correction, (b)yaw-adjusted UAV image aligned to true north, and (c) extracted GCP chip from the yaw-corrected image.

2.1.5. GCP Chips

As explained in the previous section, two types of GCP chips were prepared for each study area using the aerial orthoimages and the UAV images. For each study area, the locations of GCP chips were selected to allocate GCP chips evenly distributed within the extent of the WorldView-3 images. In the Geumsan area, 20 locations were selected, and in the Incheon area, 16 locations. Fig. 7 shows the location of GCP chips selected for the WorldView-3 images over the Geumsan and Incheon areas. There were no GCP chip locations in the upper left region of the Geumsan image. This was because the region contained very steep mountains, and UAV flight could not be attempted. Note that in both study areas, GCP chips were generated from aerial orthoimages and UAV images at the same locations. This ensured consistency in the performance comparison between the two types of GCP chips, with resolution as the only variable.

Fig. 7. Distribution of GCP chips: (a) Geumsan and (b) Incheon.

GCP chips were constructed by extracting image patches of a predefined size, whose centers were at the selected GCP chip locations. From the aerial orthoimages, GCP chips of 180 × 180 pixels with 25 cm GSD were constructed. Each chip covered an area of 45 × 45 m on the ground. From the UAV images, the size of GCP chips was adjusted to cover the same ground extent as the corresponding GCP chips from the aerial orthoimages. For Geumsan, the GSD of the UAV images was 3 cm. The size of each GCP chip was defined as 1,500 × 1,500 pixels to cover the same ground area of 45 × 45 m. For Incheon, the GSD of UAV images was 1.9 cm. The size of each GCP chip was defined as 2,380 × 2,380 pixels, which corresponded to the same ground area. The specifications of the chips are summarized in Tables 4 and 5. Fig. 8 shows examples of GCP chips from the aerial orthoimage and the UAV image at the same GCP location for Geumsan and Incheon.

Table 4. Specifications of GCP chips in Geumsan

Table 5. Specifications of GCP Chips in Incheon

Fig. 8. Example of GCP chips: (a) Geumsan from aerial orthoimages, (b) Geumsan from UAV images, (c) Incheon from aerial orthoimages, and (d) Incheon from UAV images.

2.2. Automated Accurate Sensor Modeling

This paper proposes an automated, accurate sensor model establishment for VHR satellite images based on GCP chip matching. The workflow for the establishment is illustrated in Fig. 9. We adjust the GSDs of the satellite images and GCP chips to make their GSDs a designed value. After this, each GCP chip is matched against the satellite images. Once matching is complete for all chips, an accurate sensor model is established from the results. We repeat this procedure with different designed values of GSDs using the GCP chips from aerial images and the chips from UAV images. We compare the accuracy of sensor models generated from the two types of GCP chips.

Fig. 9. Workflow of accurate sensor modeling for satellite images based on GCP chip matching.

2.2.1. Establishment of the Initial Sensor Model

A sensor model mathematically represents the geometric relationship between the image coordinates and the ground coordinates at the time of imaging. It is used to estimate the ground position corresponding to a specific pixel in the image. In this study, an initial sensor model is established based on the coefficients for a Rational Function Model (RFM) provided with satellite images. The RFM used consists of two rational polynomial expressions.

\(\begin{align}c_{n}=\frac{P_{1}\left(X_{n}, Y_{n}, Z_{n}\right)}{P_{2}\left(X_{n}, Y_{n}, Z_{n}\right)}=\frac{\sum_{i=0}^{3} \sum_{j=0}^{3} \sum_{k=0}^{3} a_{i j k} X_{n}^{i} Y_{n}^{j} Z_{n}^{k}}{\sum_{i=0}^{3} \sum_{j=0}^{3} \sum_{k=0}^{3} b_{i j k} X_{n}^{i} Y_{n}^{j} Z_{n}^{k}}\end{align}\)       (2)

\(\begin{align}r_{n}=\frac{P_{3}\left(X_{n}, Y_{n}, Z_{n}\right)}{P_{4}\left(X_{n}, Y_{n}, Z_{n}\right)}=\frac{\sum_{i=0}^{3} \sum_{j=0}^{3} \sum_{k=0}^{3} c_{i j k} X_{n}^{i} Y_{n}^{j} Z_{n}^{k}}{\sum_{i=0}^{3} \sum_{j=0}^{3} \sum_{k=0}^{3} d_{i j k} X_{n}^{i} Y_{n}^{j} Z_{n}^{k}}\end{align}\)       (3)

In Eqs. (2) and (3), X_n, Y_n, Z_n denote normalized ground coordinates and c_n, r_n denote normalized image coordinates. Each polynomial function P_i is expressed as a cubic polynomial and contains 20 coefficients in the numerator and denominator. Since four such polynomials are used, the complete RFM is defined by a total of 80 coefficients. This initial RFM-based sensor model is to be subsequently refined using the results of GCP chip matching, and updated into an accurate sensor model.

2.2.2. Resolution Adjustment

Since the GSDs of satellite images and GCP chips are not identical, their GSDs must be unified before matching. Therefore, both datasets were resampled to a common designed GSD. This helped to minimize the differences in spatial resolution and pixel coverage. In turn, it enables the comparison of matching performance under various resolution conditions. The designed GSD values were set to 30 cm, 15 cm, 10 cm, 7.5 cm, and 6 cm. Although the actual GSD of the WorldView-3 panchromatic image used in this study is 31.6 cm, a nominal GSD of 30 cm was adopted to define the scale factors used throughout the experiments. These values correspond to upsampling ratios of ×1, ×2, ×3, ×4, and ×5 relative to the 30 cm nominal GSD.

Resolution adjustment was performed differently depending on whether the target GSD was coarser or finer than the original GSD. When the target GSD was coarser than the original GSD, downsampling was performed using area-based resampling. In this approach, each output pixel is computed as the weighted average of the input pixels contributing to its area, which mitigates aliasing during decimation. Conversely, when the target GSD was finer than the original GSD, upsampling was performed using bicubic interpolation. These resolution adjustment procedures were applied to align the pixel spacing of both the satellite image and the GCP chips to a common grid.

Fig. 10 illustrates the visual changes in the image according to the upsampling scale. It compares the satellite image patches with GSDs of 30 cm, 15 cm, and 10 cm. It shows smoother visual transitions as the pixel spacing becomes finer.

Fig. 10. WorldView-3 satellite image patches at different GSDs according to upsampling scale: (a) 30 cm, (b) 15 cm, and (c) 10 cm.

2.2.3. GCP Chip Remapping

Area-based matching methods adopted in this paper are highly sensitive to geometric inconsistencies (Ye & Shan, 2014). Therefore, it is essential to realign GCP chips to conform to the grid structure of satellite images. In this study, as illustrated in Fig. 11, GCP chips are remapped to conform to the spatial resolution and geometric extent of the satellite images. This process serves as a preprocessing step for automated image matching between satellite images and GCP chips and directly affects the accuracy of area-based matching algorithms (Son et al., 2021). Based on the initial sensor model of a raw satellite image, the center and four corners of each GCP chip are projected onto the satellite image frame. Based on the projected image locations, the GCP chips are remapped to conform to the satellite image frame. Bilinear interpolation was applied to resample the GCP chips. Fig. 11 demonstrates how GCP chips are aligned with the image frame of satellite images.

Fig. 11. Resampling of GCP chips to match the satellite image.

2.2.4. Matching between Satellite Images and GCP Chips

To ensure both processing efficiency and accuracy in matching between satellite images and GCP chips, this study applies an iterative accurate geometric correction method based on an image pyramid (Son et al., 2021). For each satellite image patch, a four-level image pyramid is constructed, as shown in Fig. 12. Figs. 12(a–c) show images reduced by factors of 8, 4, and 2, respectively, while Fig. 12(d) represents the original-resolution image. Area-based matching is performed from the top layer to lower layers. The results at higher layers are propagated to lower layers as the initial positions. Given that the satellite image and the GCP chip had already been normalized to the same GSD in Section 2.2.2, both were subsequently downsampled to match each pyramid layer using an averaging approach. This hierarchical approach progressively refines point locations by correcting mismatches at upper layers before passing them downward, thereby improving overall matching accuracy.

Fig. 12. Image pyramid configuration examples: (a) pyramid layer 4 (8xzoom), (b) pyramid layer 3 (4xzoom), (c) pyramid layer 2 (2xzoom), and (d) pyramid layer 1 (original image).

The upper three layers are matched using the Zero-means Normalized Cross Correlation (ZNCC) algorithm (Di Stefano et al., 2005), while the final level employs the Census Transform (CT) algorithm (Zabih & Woodfill, 1994). ZNCC is a representative similarity metric that computes cross-correlation between two image patches, normalized by subtracting the mean of each patch. CT is a non-parametric method that encodes local structure by comparing each neighboring pixel to the center pixel, assigning a binary value of 1 when the neighbor is greater, and 0 otherwise. At the upper levels of the pyramid, ZNCC is used because the reduced resolution images require less computation, allowing this relatively slow but accurate algorithm to be applied efficiently. At the final original resolution level, the faster CT algorithm is used to reduce processing time, ensuring overall computational efficiency while maintaining matching accuracy (Son et al., 2021). For the ZNCC matching stages, a 39 × 39 pixel search window and a correlation threshold of 0.3 were used, and RANSAC outlier rejection was applied with an inlier threshold of 2 pixels.

2.2.5. Establishment of the Accurate Sensor Model

After establishing the initial sensor model, this study refines the RFM by applying bias compensation with GCPs obtained from the matching process (Fraser & Hanley, 2005; Oh et al., 2022; Yoon et al., 2018). Bias compensation is performed by adding correction terms (Δc, Δr) obtained from a correction model to the image coordinates predicted by the initial RFM. An affine-based first-order polynomial model is used as the correction model, a widely recognized effective method for establishing accurate sensor models (Jeong et al., 2014; Teo, 2011; Yoon et al., 2018). The correction terms are estimated using an iterative least squares adjustment, enabling optimal parameter estimation (Grodecki & Dial, 2003). The bias-compensated RFM expresses the image coordinates as follows:

c = c_RFM(X, Y, Z) + Δc + є_c       (4)

r = r_RFM(X, Y, Z) + Δr + є_r       (5)

where c and r are the actual image coordinates measured at the GCPs, c_RFM (X, Y, Z) and r_RFM (X, Y, Z) are the image coordinates computed from the initial RFM, ∆c and ∆r are the correction terms estimated, є_c and є_r are the residuals in the column and row directions, respectively. The correction model used in this study is formulated as follows:

c = a₁₁c_o + a₁₂r_o + a₁₃       (6)

r = a₂₁c_o + a₂₂r_o + a₂₃       (7)

where c_o and r_o are the image coordinates derived from the initial RFM, c and r are the corrected image coordinates, and a₁₁~a₂₃ are the coefficients to be estimated through iterative least squares adjustment. As the automatically extracted GCP correspondences from the preceding step may include outliers, the Random Sample Consensus (RANSAC) algorithm is applied to remove them (Fischler & Bolles, 1981).

3. Results

The GCP chips of the two types were matched against the raw WorldView-3 images over Geumsan and Incheon areas and accurate sensor models were established. In this study, automated accurate sensor modeling of VHR satellite images was evaluated in terms of matching performance and sensor model accuracy. Table 6 shows the number of successfully matched GCP chips for each experiment case. Successfully matched points refer to the inliers retained after RANSAC filtering and used to establish the sensor model. More GCP chips were matched on average in Geumsan than in Incheon. This is because the Geumsan area had a larger number of available GCP chips than Incheon, and their centers appeared more distinctly in the images, which led to more successful matches. The number of matches decreased slightly as the adjusted GSD values decreased. However, the decrease was not severe. The matching algorithm seemed to work well for adjusted GSDs designed for VHR satellite images.

Table 6. Number of successfully matched GCP chips for each experiment case​​​​​​​

Sensor model accuracy was evaluated in terms of model point errors and check point errors. The model point error was defined as the RMSE between the observed image coordinates of the GCP chips used to establish the accurate sensor model and the image coordinates predicted by the model. It indicates how accurately the model reproduces the training data used for its establishment. The check point error was obtained by computing RMSE between the predicted and observed image coordinates of GCPs that were excluded from model establishment. Seven check points were prepared in each study area. This error reflects the model performance at independent check points that were not used for sensor model establishment, and it serves as an indicator of the model’s validation quality as well as its ability to generalize to unseen locations.

Table 7 presents the sensor model accuracy for Geumsan, while Table 8 for Incheon. For the two areas, model errors consistently decreased as the adjusted GSD values decreased. For the aerial GCP chips over Geumsan, the error was approximately 1.32 pixels at the GSD of 30 cm. The error decreased on the 15 cm and 10 cm scales and finally reached about 0.30 pixels at the 6 cm scale. The UAV-based GCP chips over Geumsan showed a similar trend, and the final errors for both chips converged to around 0.26 pixels. The Incheon area exhibited similar patterns. For the aerial chips, the error was 1.29 pixels at a GSD of 30 cm, which improved significantly at a 10 cm scale to 0.25 pixels at 6 cm. The UAV chips started at 1.35 pixels and reduced by more than half at the 15 cm and 10 cm scales, reaching 0.21 pixels at 6 cm.

Table 7. Overall experiment results for the Geumsan study area

Table 8. Overall experiment results for the Incheon study area​​​​​​​

To establish a reference baseline for the subsequent check point error analysis, the initial and corrected sensor model errors are decomposed into column and row Mean Errors (ME) and Standard Deviations (SD). As shown in Table 9, the initial sensor model RMSEs were 7.4 pixels for Geumsan and 10.5 pixels for Incheon, mainly due to large mean errors in the column direction (Col ME), indicating strong systematic biases in both areas. Such error patterns are commonly observed for vendor-provided RPC models, which often show systematic biases along the row or column direction. Our initial RPC-based sensor model exhibited this typical behavior, as summarized in Table 9. In Geumsan, the RMSE was reduced from 7.4 to 1.0 pixels (UAV 30 cm), and in Incheon from 10.5 to 1.7 pixels, corresponding to an improvement of more than 80%. This demonstrates that the accurate sensor modeling effectively removed most of the systematic translational errors dominated by the column-direction mean error. The improvement was most evident in the column direction, which is consistent with geometric shifts in raw satellite images typically occurring along the column axis. After correction, the SD values in both directions remained within 0.2–1.7 pixels, indicating that the remaining discrepancies are primarily attributed to random errors rather than systematic biases.

Fig. 13. Model point error for the Geumsan study area.

Fig. 14. Model point error for the Incheon study area.

Table 9. Initial sensor model check error measured by validation points in this study

Table 10. Overall experiment results for the Geumsan study area (ME, SD, RMSE)​​​​​​​

Table 11. Overall experiment results for the Incheon study area (ME, SD, RMSE)​​​​​​​

Figs. 15 and 16 show the check point errors over Geumsan and Incheon. For the aerial chips over Geumsan, the check error of the initial sensor model (7.4 pixels) was reduced to 2.21 pixels at the adjusted GSD of 30 cm and to 1.61 pixels at 15 cm. The errors were further decreased to 1.27 pixels and 1.18 pixels at 10 cm and 7.5 cm, respectively, but increased again to 2.05 pixels at 6 cm. For the UAV chips, the check point errors were consistently lower, ranging from 0.96 to 1.28 pixels across all adjusted GSDs. These results indicate a clear performance advantage when using high-resolution GCP chips from UAV images.

Fig. 15. Check point error for the Geumsan study area.

Fig. 16. Check point error for the Incheon study area.​​​​​​​

For the aerial chips in Incheon, the check error of the initial sensor model (10.5 pixels) was reduced to 1.99 pixels at 30 cm GSD and to 1.31 pixels at 10 cm. At 7.5 cm and 6 cm, the errors converged to within the 1.33-pixel range. For the UAV chips, the error was reduced to 1.66 pixels at 30 cm. However, the error slightly increased to 2.01 pixels at 15 cm and gradually decreased to the minimum value of 1.31 pixels. In the Incheon area, the mean check point errors of the aerial and UAV chips were both 1.6 pixels; however, at 30 cm GSD, the check point error for the UAV chips (1.66 pixels) was smaller than that for the aerial chips (1.99 pixels). As the adjusted GSD decreased, the errors for both platforms improved and, at 10 cm and finer, tended to converge to a similar level of about 1.3–1.4 pixels.

Overall, the check point errors showed different patterns from the number of successful matches. While the two chip types yielded similar match counts, UAV chips achieved comparable or slightly better accuracy across the two study areas. To assess whether the accuracy difference between the two chip types was statistically meaningful, the five GSD-specific errors within each check point were averaged to yield one paired observation per check point (n = 7 per region), thereby accounting for repeated measurements across GSD conditions. A paired Wilcoxon signed-rank test was performed, and effect sizes and bootstrap 95% confidence intervals for the mean paired difference (Aerial − UAV) were also reported. For Geumsan, UAV chips yielded lower error at every check point (p = 0.016; r_rb = 1.00; mean paired difference (Aerial − UAV) = 0.59 pixels, 95% CI [0.44, 0.73]). For Incheon, no statistically significant difference was detected (p = 0.688; r_rb = −0.21; mean paired difference (Aerial − UAV) = −0.01 pixels, 95% CI [−0.14, 0.11]).

In both areas, the check point errors decreased as the adjusted GSD values decreased. In Geumsan, the lowest error was 0.96 pixels for the UAV chips at the 7.5 cm scale, while in Incheon it was 1.31 pixels for the UAV chips at the 6 cm scale. However, beyond a certain scale, no meaningful error reduction was observed at finer resolutions. Investigation into the optimal adjusted GSD value is required as a topic for future research. Orthorectified images were generated using the initial sensor models and the best accurate sensor models generated from UAV chips. As shown in Fig. 17, the orthorectified images from the proposed GCP chip matching process exhibited higher positional consistency when overlaid with NGII reference maps.

Fig. 17. Visual comparison of orthoimages overlaid on the reference map (NGII digital topographic map): (a) Geumsan and (b) Incheon.​​​​​​​

4. Discussion

This study evaluated the impact of GCP chip resolution on the accuracy of automated accurate sensor modeling using VHR WorldView-3 images. The evaluation results raised several points of interest, which are discussed below.

First, the use of high-resolution GCP chips improved sensor model accuracy. Experimental results showed that check point errors obtained from UAV chips were generally lower than those from aerial chips. This improvement can be attributed to the richer image textures of high-resolution chips, which facilitate more accurate area-based matching. The results suggest the potential for achieving sub-meter (i.e., absolute positional accuracy <1 m) accuracy using VHR satellite images. The statistical test in Section 3 further supported the superiority of UAV chips at Geumsan, whereas no statistically significant difference between the two chip types was detected at Incheon.

Second, performing GCP chip matching at a higher resolution than the raw satellite images tended to improve sensor model accuracy. As the resolution increased, check point errors generally exhibited a decreasing trend. Although upsampling does not create new spatial information, it provides a denser sampling grid that enables finer sub-pixel localization during matching. However, excessive upsampling did not yield further accuracy improvements, as the underlying information content is bounded by the original GSD. This trend aligns with the findings of previous studies (Choi et al., 2024; Choi & Kim, 2022; Lee & Kim, 2021). Ultimately, the results of this study demonstrate both the utility and the limitations of upsampling. For future work, further analysis is required to quantitatively determine the optimal target GSD, taking into account the efficiency of the geometric correction process.

In addition to resolution effects, several other factors may have contributed to matching failures. These include changes in the real-world scene, difficulty in visually identifying GCPs within the images, and various environmental conditions. In urban areas, relief displacement, building shadows, and complex man-made structures often made GCP identification difficult, whereas in mountainous regions, limited UAV accessibility affected the usability of GCPs. The Incheon area experienced more frequent temporal scene changes than the Geumsan area between the satellite acquisition and the reference data, and such changes may influence the matching success rate. These observations suggest that practical geometric correction applications require not only resolution optimization but also matching strategies that are robust to such environmental factors.

It should also be noted that the resolution adjustment process in this study was designed as a preprocessing step for image matching, focusing on aligning pixel spacing and preserving spatial features. Accordingly, explicit modeling of the sensor-specific Point Spread Function (PSF) was not included within the scope of this study, and a standard resampling method was applied to mitigate aliasing. For example, Wolfe and Nishihama (2011) used the PSF of a coarse-resolution sensor as a weighting function when downsampling a finer-resolution sensor image for automated ground control matching. In future work, if accurate PSF or Modulation Transfer Function (MTF) information becomes available, or validated approximations of the WV3 PSF/MTF are utilized, the proposed approach could be extended to incorporate resolution adjustment methods that more precisely reflect the physical optical characteristics of the sensor.

This study has several limitations that suggest directions for future research. First, the proposed UAV-based GCP chip approach does not claim practical superiority over conventional workflows in terms of cost, efficiency, or scalability, since UAV acquisition itself also requires field operations. The contribution of this study lies in demonstrating that, given the same UCP coordinates, higher-resolution UAV chips can improve automated sensor modeling accuracy over conventional aerial orthoimage chips. Additionally, the accuracy assessment relied on only seven check points per region, and temporal inconsistencies between satellite and reference image acquisition may have influenced matching performance. Both factors limit the strength of the conclusions and motivate the need for larger check point sets and temporally aligned data acquisition in future studies.

From an operational perspective, UAV operation is restricted in mountainous, island, and no-fly zones. The upper-left region of the Geumsan scene was excluded for this reason. In such areas, UCPs are typically not established, and triangulation points, which are less accurate than UCPs, are commonly used instead. Future work will analyze how the inclusion of such less accurate reference points affects sensor modeling accuracy. Performance under varying elevation conditions should also be investigated together.

Additionally, flat terrain was intentionally selected in this study to minimize the influence of orthorectification when evaluating the matching algorithm. Further investigation is needed on how orthorectified UAV images affect geometric accuracy in more complex terrains. Another aspect warranting attention is that as the spatial resolution of satellite images increases, the coverage area of individual scenes tends to decrease. Additional studies are needed to develop efficient geometric correction strategies that utilize a limited number of GCPs, and further analysis is required to examine how high-resolution GCP chips can contribute to optimizing the number of GCPs and achieving cost-efficient geometric correction while maintaining accuracy.

Finally, this study provides only a preliminary validation of the proposed approach under the tested conditions, as it was evaluated using a limited number of WorldView-3 scenes from two contrasting sites. Further validation across a larger image set, additional VHR satellite platforms, and diverse terrain types is required to confirm the general applicability of the proposed method.

5. Conclusions

This study investigated automated geometric correction for VHR satellite images using GCP chips generated from UAV images. In particular, we analyzed the impact of GCP chip resolution on the correction process. Experiments were conducted using GCP chips generated from aerial images with a GSD of 25 cm and from UAV images with GSDs of 1.9–3 cm. While the achieved accuracy did not reach the CARD4L requirement, the proposed method achieved sub-meter (i.e., absolute positional accuracy <1 m) accuracy and produced reliable orthorectified images from VHR satellite images. Overall, the results indicate that finer-resolution GCP chips and upsampling of raw satellite images can improve the geometric accuracy of VHR satellite images, with the magnitude of improvement depending on land-surface characteristics. An additional contribution of this study is the direct generation of GCP chips from UAV images and their application to automated geometric correction. This approach highlights the practical value of UAV-based reference data and suggests that geometric correction accuracy could be further improved when such data are utilized more efficiently and systematically in future applications. As the present validation was conducted with WorldView-3 imagery over two study areas, broader validation across additional VHR satellite platforms and diverse terrain types is required to confirm the general applicability. The findings of this study may contribute to improved positional accuracy in VHR satellite images.

Author Contributions

Conceptualization: Kim TJ; Data curation: Kim HA, Choi HG, Kim HJ, Kim SH; Methodology, Software: Kim HA, Choi HG; Formal analysis, Investigation, Validation, Visualization, Writing–original draft: Kim HA; Writing–review & editing: Kim TJ.

Conflicts of Interest

No potential conflict of interest relevant to this article was reported.

Funding

This work is supported by the Korea Agency for Infrastructure Technology Advancement (KAIA) grant funded by the Ministry of Land, Infrastructure and Transport(GrantRS-2022-00155763).

Data Availability Statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.