Recent advancements in satellite image utilization technology have led to growing demands for high-resolution satellite imagery. In response, Korea has developed and launched a Korean Multi-Purpose Satellite (KOMPSAT) with a resolution of 0.55 m. Furthermore, Compact Advanced Satellite 500 (CAS500) with a resolution of 0.5 m has been developed (Lee et al., 2017; Kim, 2020). Korea is currently participating in the development of comprehensive land management technology using satellite image information big data to improve land management technology (Kim, 2023).

The correction of geometric distortions caused by errors in satellite global positioning system (GPS) receivers and attitude control sensors is necessary when using satellite imagery. This correction involves aligning image coordinates with ground coordinates using ground control points (GCPs). The establishment of a precise sensor orientation through this process is crucial in improving the accuracy of positioning in high-resolution satellite imagery. In this context, even minor errors can have a significant impact, limiting the achievable positioning accuracy to a range of 1–1.5 pixels (Choi et al., 2023).

Traditionally, establishing precise sensor orientation requires manually acquiring GCPs, which incurs significant time and economic costs. To address this challenge, previous studies have introduced GCP chips - small image fragments with precise ground coordinates. They have developed a technology that can automatically match GCP chips and satellite images to establish an automatic precision sensor orientation (Park et al., 2020; Yoon, 2019).

The use of GCP chips for precise sensor orientation establishment automation provided a cost-effective means of producing and distributing calibrated satellite images. As a result, researchers have actively pursued studies aimed at achieving higher positioning accuracy through the establishment of automatic precise sensor orientation. Oh et al. (2022) investigated the feasibility of automating the establishment of high-resolution satellite image precision sensor orientation using GCP chips.

Meanwhile, Shin et al. (2018) conducted an experiment in GCP chip matching using pan-sharpened images to establish an automatic precision sensor orientation tailored for high-resolution satellite imagery. The findings indicated a significant improvement in matching performance when using pan-sharpened images compared to original multispectral images. Lee and Kim (2021) investigated the effectiveness of using high-resolution GCP chips to establish a precise sensor orientation for medium-resolution satellite imagery. They used satellite image upsampling instead of pan-sharpening to improve the matching performance of medium-resolution satellite images without pan bands.

The study’s results indicated that upsampling to match at the medium resolution was more effective in terms of geometric accuracy than matching at the original resolution. Choi and Kim (2022a) conducted a follow-up study in which they adjusted the number of chips used to establish the precision sensor orientation during satellite image upsampling. This was done to reduce the time required to establish a medium-resolution automatic precision sensor orientation and improve GCP chip matching performance. Previous studies have shown that the performance of GCP chip matching is affected by the spatial resolution of both the GCP chip and satellite imagery.

The objective of this study is to develop an improved method for establishing automatic precision sensor orientation in high-resolution satellite images. To achieve this, we conducted chip matching by upsampling both the satellite and GCP chip images. We then performed a comprehensive comparative analysis of the results using the established precision sensor orientation. The study concluded that upsampling both the high-resolution satellite images and the GCP chip images can improve the performance of the sensor orientation when matching them.

2.1. Materials

This study utilized the pre-existing spatial information data of the Korean Peninsula, as described in a previous study (Park et al., 2020). The GCP chips for the South Korean region were created using precise aerial orthoimages with a spatial resolution of 0.25 m. It includes three different types of ground coordinate information (Yoon et al., 2018). These three types of ground coordinate information include the unified control point, triangulation point, and image control point. The Korea National Geographic Information Institute generates and manages the unified control point and triangulation point. The image control point is used in the production of the national basic map and aerial orthoimage.

In the case of North Korea, it is difficult to acquire and continuously manage GCP because it is inaccessible. Therefore, the GCP chips for the North Korean region were created using satellite orthoimages with a spatial resolution of 0.5 m. It includes orthoimagery plane coordinates and digital elevation model (DEM) ellipsoid height. Therefore, The GCP chips in the South Korean region were created from color images extracted from aerial orthoimagery. In contrast, the GCP chips in the North Korean region were created from single-band images extracted from satellite orthoimagery generated from panchromatic images. Table 1 provides detailed specifications of the GCP chips used. An example of the GCP chips is shown in Fig. 1. Fig. 1(a) shows GCP chips in South Korea, while Fig. 1(b) shows GCP chips in North Korea.

Figure 1. Example of GCP chip.

Table 1 . Specifications for GCP chip used.

Area	South Korea	North Korea
Ground coordinates reference	Unified control point (UCP), Image control point (ICP), Triangulation point (TP)	Plane coordinates of the orthogonal image, DEM ellipsoidal height
Source image	Aerial orthoimage	Satellite orthoimage
Chip size	1,027 x 1,027 pixel	513 x 513 pixel
GSD	0.25 m	0.50 m
Band	Red, Green, Blue	Gray

GSD: ground sample distance..

The CAS500-1 image, a high-resolution satellite image with a spatial resolution of 0.5 meters, was utilized in this study. The specifications of the satellite images are shown in Table 2. The study area includes Sejong, Andong, and Pyongyang. Table 3 presents a list of the satellite images used in this study. Additionally, the study area’s geographic distribution is displayed in Figs. 2 to 4.

Figure 2. Distribution of study area (Sejong).

Figure 3. Distribution of study area (Andong).

Figure 4. Distribution of study area (Pyongyang).

Table 2 . Specifications for satellite images used.

Satellite image	CAS500-1
Product Level	Level 1R
Spectral resolution	Panchromatic: 450–900 nm Multispectral: 450–900 nm (Blue, Green, Red, NIR)
GSD	Panchromatic: 0.5 m Multispectral: 2.0 m
Orbit	Circular sun-synchronous (500 km)
Swath	≥ 12km
Radiometric resolution	12 bits

Table 3 . List of used satellite images.

Data	Acquisition date	Upper left coordinate (Lat, Lon)	Bottom low coordinate (Lat, Lon)	Quantity of GCP chip
Sejong 1	2021.12.07	36.509042465, 127.266395653	36.426378705, 127.427498911	56
Sejong 2	2021.12.07	36.612209951, 127.238683259	36.529558877, 127.400014661	40
Sejong 3	2022.02.27	36.625661023, 127.201747702	36.553286018, 127.406668861	68
Sejong 4	2022.03.03	36.533186804, 127.227716611	36.450030787, 127.390117310	55
Sejong 5	2023.03.16	36.506973164, 127.244234044	36.427195422, 127.414961162	58
Sejong 6	2023.03.16	36.609898079, 127.214833673	36.530137740, 127.385835309	40
Andong 1	2021.11.11	36.724915722, 128.875059965	36.642481614, 129.038039928	39
Andong 2	2022.06.12	36.619879282, 128.688020424	36.537777343, 128.850648011	36
Andong 3	2022.12.02	36.449640721, 128.893645792	36.368206234, 129.058636036	43
Pyongyang 1	2021.10.19	39.156498918, 125.645494344	39.073884439, 125.839485291	92
Pyongyang 2	2023.01.10	39.004960711, 125.710489853	38.922023330, 125.888358084	74

2.2. Methods

Fig. 5 presents a schematic of the procedure for establishing automatic precision sensor orientation using GCP chip matching, as applied in this paper.

Figure 5. Flowchart in this study.

Based on the findings of Shin et al. (2018), we conducted image fusion, also known as pan-sharpening, to enhance the matching accuracy between high-resolution satellite images and GCP chips. Pan-sharpening is a technology that combines high-resolution panchromatic band images with low-resolution multi-spectral band images to reconstruct high-resolution color images. This study utilized an algorithm based on the component-substitution (CS) fusion technique, as described in previous works (Park et al., 2020; Vivine et al., 2015).

This paper proposed establishing the initial sensor orientation first. The sensor orientation was established using the rational function model (RFM) for this study. The RFM is a mathematical sensor model mainly used for satellite imagery. It is a model expression that mathematically expresses the relationship between image coordinates and ground coordinates, without physical sensor information. It can be established by applying the rational polynomial coefficients (RPCs) provided with the original satellite image to the formula. Eq. (1) and (2) delineate the formulas of the RFM corresponding to the columns and rows of the image.

The formula provided uses c_n and r_n to represent the normalized image coordinates, and X_n, Y_n, and Z_n to denote the normalized ground coordinates. In the given equation, a_ijk, b_ijk, c_ijk, and d_ijk are the coefficients of the rational polynomials P₁, P₂, P₃, and P₄, respectively. These coefficients constitute the rational function expression defining the relationship between the image’s column, row, and ground coordinates. We set the values of the RPCs for i + j + k ≥ 4 to zero to formulate the RFM model expression using only the coefficients of the third term or lower in the ground coordinates. Additionally, we set b₀₀₀ and d₀₀₀ to 1 to fix the proportionality constant of the rational function expression. The remaining 78 coefficients for ground coordinate terms with i + j + k from 1 to 3 were provided as RPC files with satellite images.

Using the established initial sensor orientation, we conducted a comprehensive search for all GCP chips within the entire coverage of the satellite image. The image corner coordinates, derived from the initial sensor orientation, were used to calculate the minimum rectangular area encompassing the image region. To account for potential errors in the initial sensor orientation, a margin was incorporated into the image area. The margin dimensions were set at 250 meters in both the x and y directions of the satellite image. High-resolution satellites have a limited field of view and are affected by the area’s topography. Therefore, the margin size was determined based on the search for stable GCP chips and the potential for overestimation to establish accurate sensor orientation. The search for GCP chips was confined to the area within which the margin was applied, referencing the GCP database.

Subsequently, we defined the region surrounding each GCP chip and executed upsampling of the satellite image to the desired scale. Previous research has shown that bilinear interpolation is advantageous for upsampling satellite images in terms of both performance and time efficiency (Choi and Kim, 2022).

Therefore, we adopted bilinear interpolation for the upsampling process. In this study, we applied upsampling factors ranging from 1 to 3 times for each set of experimental data, resulting in spatial resolutions of 0.5 m, 0.25 m, and 0.167 m. The corresponding upsampling factors are shown in Fig. 6. Fig. 6 illustrates the original satellite image transformed according to the upsampling factor. It was evident that a larger upsampling factor leads to a smoother appearance of the image. To save processing speed and memory, the upsampling was only applied to the navigation area relative to the GCP, rather than the entire original satellite image.

Figure 6. Different satellite image patches and a GCP chip according to upsampling factors.

Afterward, we upsampled and remapped the GCP chip image to match the geometry of the satellite image and its newly achieved spatial resolution. To ensure automatic matching using an area-based image-matching algorithm, we deformed the GCP chip to synchronize its resolution and geometry with that of the satellite image. Fig. 7 presents a depiction of the resampling process for the GCP chip, considering the geometry and resolution of the satellite image. This procedure was executed for all identified GCP chips.

Figure 7. Concept of resampling GCP chip.

Next, we created an image pyramid to improve processing efficiency for image matching. Four levels of image pyramids were constructed for each satellite image patch. The matching process began at the top layer of the pyramid and used the Census algorithm in the last layer and the zero-mean normalized cross-correlation (ZNCC) algorithm in the remaining layers, except for the last one. The previous study (Yoon, 2019) has shown that the ZNCC algorithm is robust to image size variations, but its accuracy in extracting matching points is somewhat lower. In contrast, the Census algorithm has relatively good accuracy in extracting matching points but is sensitive to variations in image size.

Therefore, we used the Census algorithm for the original image and the ZNCC algorithm for the other reduced images. Despite our efforts, there were still some mismatch points among the automatically matched GCPs. To address this issue, we utilized the random sample consensus (RANSAC) technique (Fischler and Bolles, 1981) to automatically remove them. The detailed process of GCP chip matching and mismatch point removal is omitted here for brevity and to avoid redundancy, as it has been covered in previous research papers (Yoon et al., 2018; Son et al., 2021).

As the final step, we established a precise sensor orientation based on the auto-matching results with removed mismatches. We estimated the coefficients of the following error correction equation to rectify any errors in the initial sensor orientation.

In the equation above, c₀, r₀, c, and r represent the image coordinates before and after correction, respectively. The coefficients of the error correction formula are represented by a₁₁ to a₂₃. After estimating the coefficients of the error correction equation, the RPC coefficients were recalculated using this equation to establish precise sensor orientation. The precision sensor orientation was established for each upsampling factor and then converted back to the original resolution for accuracy comparison (Lee and Kim, 2021). The accuracy of the precision sensor orientation was compared for each upsampling factor.

In this study, model error and check error were employed as performance evaluation metrics. Model error represents the relative root mean square error (rRMSE) measured for the GCPs used to refine the initial sensor model. Check error, on the other hand, is expressed as the rRMSE measured for checkpoints manually extracted from reference points not included in the model points. Eq. (5) details the calculation of rRMSE, where RMSE_col represents the RMSE in the column direction of the image, and RMSE_row represents the RMSE in the row direction of the image. Accuracy analyses were conducted based on the original image with a spatial resolution of 0.5 m.

Table 4 shows the number of GCP chips within the acquisition area and the manually acquired number of GCP chips for verification across the 11 CAS500-1 satellite images used in the experiment.

Table 4 . Quantity of GCP chips per experimental case.

Data	Quantity of GCP chips	Quantity of checkpoint
Sejong 1	56	6
Sejong 2	40	6
Sejong 3	68	6
Sejong 4	55	6
Sejong 5	58	7
Sejong 6	40	6
Andong 1	39	5
Andong 2	36	4
Andong 3	43	5
Pyongyang 1	92	10
Pyongyang 2	74	8

The experimental results are presented in Table 5, and the performance evaluation indicator graphs for each study area are shown in Figs. 8 to 18, illustrating the observed trends in the experimental outcomes. In each figure, (a) and (b) represent the model error and check error, respectively, as functions of the number of GCPs and the upsampling scale.

Figure 8. Result of Sejong1.

Figure 9. Result of Sejong2.

Figure 10. Result of Sejong3.

Figure 11. Result of Sejong4.

Figure 12. Result of Sejong5.

Figure 13. Result of Sejong6.

Figure 14. Result of Andong1.

Figure 15. Result of Andong2.

Figure 16. Result of Andong3.

Figure 17. Result of Pyongyang1.

Figure 18. Result of Pyongyang2.

Table 5 . Matching result.

Sejong 1					Sejong 2
GSD (m)	Error (pixel)		Matching rate		GSD (m)	Error (pixel)		Matching rate
	Model	Check				Model	Check
0.500	1.285	1.546	45%		0.500	1.198	1.125	50%
0.250	0.626	1.036	36%		0.250	0.624	0.963	35%
0.167	0.403	0.738	34%		0.167	0.423	1.043	38%
0.125	0.277	1.807	36%		0.125	0.346	1.081	45%
Sejong 3					Sejong 4
GSD (m)	Error (pixel)		Matching rate		GSD (m)	Error (pixel)		Matching rate
	Model	Check				Model	Check
0.500	1.367	2.039	34%		0.500	1.230	0.936	45%
0.250	0.739	1.419	25%		0.250	0.605	0.632	36%
0.167	0.502	4.363	28%		0.167	0.477	2.651	40%
0.125	0.352	3.590	22%		0.125	0.324	0.881	33%
Sejong 5					Sejong 6
GSD (m)	Error (pixel)		Matching rate		GSD (m)	Error (pixel)		Matching rate
	Model	Check				Model	Check
0.500	1.433	1.573	41%		0.500	1.312	2.225	55%
0.250	0.661	0.714	34%		0.250	0.559	1.665	35%
0.167	0.465	2.233	33%		0.167	0.386	2.387	43%
0.125	0.294	1.809	34%		0.125	0.335	2.639	40%
Andong 1					Andong 2
GSD (m)	Error (pixel)		Matching rate		GSD (m)	Error (pixel)		Matching rate
	Model	Check				Model	Check
0.500	1.174	1.866	41%		0.500	1.034	1.396	47%
0.250	0.666	1.476	36%		0.250	0.600	1.211	39%
0.167	0.446	2.054	31%		0.167	0.435	1.002	33%
0.125	0.328	1.773	44%		0.125	0.325	1.549	31%
Andong 3					Pyongyang 1
GSD (m)	Error (pixel)		Matching rate		GSD (m)	Error (pixel)		Matching rate
	Model	Check				Model	Check
0.500	1.292	1.453	42%		0.500	1.346	2.189	41%
0.250	0.654	1.340	37%		0.250	0.705	1.860	27%
0.167	0.418	1.315	42%		0.167	0.435	4.302	24%
0.125	0.304	2.349	28%		0.125	0.347	3.141	21%
Pyongyang 2
GSD (m)	Error (pixel)		Matching rate
	Model	Check
0.500	1.477	1.932	35%
0.250	0.561	1.001	29%
0.167	0.438	1.979	26%
0.125	0.341	3.217	26%

Part (a) of each experiment shows that the model error consistently decreased as the upsampling factor increased and the ground sample distance (GSD) became smaller. These results indicated that coefficient estimation was successfully applied in the error correction equation for the RPC update. It can be interpreted that the higher the upsampling factor was applied, the closer the GCPs were to the estimated error correction, the closer they were selected as model points, and consequently, the smaller the GSD compared to the original resolution, the smaller the model estimation error.

Part (b) of each experiment demonstrates a decrease in check error when upsampling by 2x from the original resolution for all experiments. However, when upsampling by 3x, the error rate generally increased again. The 4x results also showed inconsistent results for each experiment. This observation is consistent with previous studies, such as Lee and Kim (2021), which have shown that excessive upsampling scaling relative to the original data reduces model estimation error but does not necessarily result in improved accuracy.

When establishing the orientation of a precision sensor at the original resolution, the check error exceeded 2 pixels in many cases, with an average check error of 1.7 pixels. When establishing the orientation of a precision sensor using an upsampling factor of 2, the check error was less than 2 pixels in all experiments, with an average check error of 1.2 pixels. We achieved a 50% improvement over the original resolution. In several experiments, where the original resolution resulted in check errors of over 1 pixel, performance considerably improved when the 2x upsampling factor was applied, resulting in less than 1 pixel check error.

To validate the experimental results objectively, we compared the matching results for each resolution. Fig. 19 shows the results of the matching GCP chip. The matching results of the GCP chip at the original resolution of 0.5 m were roughly aligned to the resolution grid. However, the matching results of the GCP chip at 0.25 m upsampled were more precisely aligned. However, when upsampled to 0.167 meters, the GCP chip over-interpolates, resulting in a smoothed image that makes it difficult to find the exact match point.

Figure 19. Matching results for each GSD. (a) Original satellite image, (b) matching result of original GSD (0.5 m), (c) matching result of 0.25 m GSD, and (d) matching result of 0.167 m GSD.

In this chapter, we delve into the distinctions between our findings and those of preceding studies utilizing medium-resolution satellite imagery (Lee and Kim, 2021; Choi and Kim, 2022a). Previous research utilizing medium-resolution satellite imagery, with a resolution of 5 m and high-resolution GCP chips featuring a resolution of 0.25 m, identified an optimal upsampling ratio of 3 times, corresponding to a GSD of 1.67 m. In our present study, employing high-resolution satellite imagery at a resolution of 0.5 m along with a GCP chip sharing the same specifications, the optimal upsampling ratio was determined to be 2 times, equating to a GSD of 0.25 m.

In the case of North Korea, both the satellite image and the GCP chip had a resolution of 0.5 m. However, the results were still improved by upsampling to 0.25 m, which may be due to a limitation of the algorithm used for chip matching. The image-matching algorithm used in this study was a pixel-by-pixel similarity search method. In summary, when upsampled to 0.25 m, the similarity search interval was applied to 0.5 pixels of the original satellite image. This improvement was believed to be a result of this adjustment.

During the study, a decrease in accuracy was observed when the upsampling ratio was tripled, which deviates from the findings in low- and medium-resolution images. This phenomenon may be attributed to disparities in the sharpness of the original satellite image or the act of upsampling to a GSD higher than the chip resolution. Future research should investigate whether the clarity of the original image or the chip resolution has a greater impact on matching results. This is crucial for advancing our understanding and achieving higher accuracy in the context of high-resolution satellite images.

In this study, we applied upsampling to both satellite images and pre-existing high-resolution GCP chips to establish a highresolution satellite image precision sensor orientation. We subsequently conducted a comparative analysis of the matching performance. The experimental outcomes confirmed a notable enhancement in matching performance when upsampling was applied to both satellite images and GCP chips.

When establishing a precision sensor orientation with a 2x upsampling factor, an improvement of up to 50% was observed compared to the original resolution. In certain experimental outcomes, the check error approached or fell below 1 pixel. Due to technical constraints, achieving sub-pixel positioning accuracy when establishing precision sensor orientation on high-resolution satellite imagery at its original resolution was challenging. However, our results demonstrated that subpixel-level positioning accuracy can be achieved through meticulous upsampling of both satellite images and GCP chips.

Moreover, despite employing different specifications for GCP chips in the South and North Korean regions, the experimental results exhibited comparable trends. Through this, we confirmed that if it is possible to build GCP chips in inaccessible areas such as North Korea as well as South Korea, this method is sufficiently applicable, and it is possible to establish an improved precision sensor orientation.

In this study, we tested a GCP chip with a GSD similar to that of satellite images. However, if a more sophisticated GCP chip is built and utilized, the position error could be further reduced through the method proposed in this study. Therefore, further research using high-resolution and high-accuracy GCP chips is needed for precise sensor modeling of high-resolution satellite images. It is also necessary to consider the use of ultra-high-resolution data such as drones. We hope that this research will help in the processing and utilization of high-resolution satellite images.

This work was carried out with the support of “Cooperative Research Program for Agriculture Science and Technology Development (Project No. PJ 016233)” Rural Development Administration, Republic of Korea. This work is supported by the Korea Agency for Infrastructure Technology Advancement grant funded by the Ministry of Land, Infrastructure and Transport (Grant No. RS-2022-00155763).

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