Geometric accuracy evaluation and analysis of ZY-1 02E IRS thermal infrared image data using GCP extraction based on phase correlation matching method

The Ziyuan-1 (ZY-1) 02E launched on December 26, 2021 is equipped with a thermal infrared sensor (IRS), which has a ground resolution of better than 16m and a width priority of 115km, balancing the advantages of high resolution and large wide observation. The geometric performance of image data is the premise of remote sensing application, and the difficulty in evaluating the geometric performance of thermal infrared image data lies in the extraction of well-distributed, reliable and accurate GCPs. To extract GCP from high-precision reference images, it is necessary to overcome the feature differences between images caused by different spectral responses. This paper adopts a phase correlation matching method based on frequency domain to realize the fine registration of the data obtained by the emission thermal spectral band with the data from the reflectance spectral band, which can not only solve the GCP extraction of conventional thermal infrared images collected during the day, but also obtain satisfactory GCP data from thermal infrared data acquired at night. In order to test the GCP method proposed in this paper, three typical areas are selected as the experimental areas, including Yiyang City in Hunan, Nagqu City in Xizang and Hami City in Xinjiang, and the internal geometric accuracy and absolute geolocation accuracy of the thermal infrared data spanning one year are evaluated and analyzed by using the reference data composed of the DOM with an accuracy of 2m and the DEM with an accuracy of 10m. The research results indicate that the internal geometric accuracy of ZY-1 02E IRS satellite image data is better than 1.0 pixels, and the performance is satisfactory. However, its absolute geolocation accuracy needs to be continuously improved, especially there are systematic errors in the ascending data at night that require further research. Overall, it meets the design accuracy indicators of satellites and can meet the application requirements of thermal infrared remote sensing.


Introduction
The Ziyuan-1 (ZY-1) 02E launched on December 26, 2021 is equipped with three main payloads, including a 9-band visible near-infrared camera (VNIC), a 166-band hyper-spectral imager, and a single-band thermal infrared sensor (IRS).It is also known as the 5m Optical 02 Satellite, which is the successor satellite of ZY-1 02D, and they are all important components of China's space-based infrastructure to continue to carry out the moderate resolution Earth observation missions.It is mainly devoted to providing panchromatic, multispectral, hyperspectral and thermal infrared data for the operation of natural resource monitoring and management, disaster prevention and control, environmental protection, etc.The ZY-1 02E has added the IRS compared to its predecessor, which has a ground resolution of better than 16m and a width priority of 115km, taking into account the advantages of high resolution and large wide observation.The IRS mainly collect thermal radiation information about the target and can work all day.
In order to ensure the high quality of radiometric and geometrically corrected image data, geometric accuracy and radiometric quality need to be evaluated and verified.Among them, absolute geolocation accuracy, internal geometric accuracy and band registration accuracy are the most important geometric performance indicators, which are the main contents of geometric calibration and verification activities, and these performance indicators need to be continuously verified, monitored and improved.
Geometric accuracy analysis is a challenge for thermal infrared image data, and the main problem is that it is difficult to obtain satisfactory tie-points between the reference data of the reflective bands and the long-wave thermal infrared emissive bands.
For the Landsat 8 and Landsat 9 the Thermal Infrared Sensor (TIRS) geometric characterization and calibration, the Operational Land Imager (OLI) short-wave infrared (SWIR) band image acquired simultaneously with TIRS is used as a reference, and the normalized gray scale correlation with a surface fitting polynomial algorithm is utilized to extract the tiepoints (Storey J. et al., 2014;Choate, M.J. et al., 2023;Choate, M.J. et al., 2024).In other words, the geometric performance analysis of Landsat 8 and Landsat 9 TIRS is premised on the rigid constraints between SWIR and TIRS, so a well-calibrated OLI is a prerequisite for on-orbit geometric analysis of TIRS.
For the ZY-1 02E IRS thermal infrared image data, this paper adopts the phase-correlated matching method based on frequency domain to achieve the accurate registration between the data obtained by the emission spectral band and the data of the reflectance spectral band, and further analyzes the geometric performance of the thermal infrared image, especially for the nighttime imaging data, and carries out a reliable geometric accuracy evaluation and analysis for the first time.

Spacecraft and IRS Overview
The ZY-1 02E satellite orbits the Earth in a sun-synchronous, at an altitude of 778 km, inclined at 98.5 degrees.The satellite has a 55-day repeat cycle with an equatorial crossing time: 10:30 a.m.Table 1 summarize the main characteristics of the ZY-1 02E.
The IRS adopts long-line array push-broom imaging, which has better sensitivity than traditional scanning imaging and can obtain ground object temperature information more accurately.Moreover, it is a push-broom sensor with no moving parts, and there is no influence of mechanism swing in the imaging process, which helps to improve the geometric performance of the image.The IRS optical system adopts a catadioptric optical system, including the main optical system and the relay optical system, the optical system has a pupil diameter of 435mm, a focal length of 1038mm, and a field of view of 8.6° in the pushsweeping direction.With high-precision thermal control technology, the stability design index of the internal azimuth element is ensured, and the absolute distortion is less than 1μm.In addition, the infrared payload arranges a double-line detector in the focal plane, which can obtain two staggered images at the same time, and has the ability of super-resolution imaging.As shown in Figure 1, the thermal infrared camera detector is composed of 8 single-module detectors spliced together according to the zigzag structure, and the total number of pixels in the linear array direction reaches 8192 pixels, and the pixel size is 20 μm × 20 μm (Tong et al., 2023).The main specifications for the design of the IRS instrument are given in Table 2.

Geometric Model
The ZY-1 02E satellite calculates orbital coordinates through the GPS receiver system, and relies on the combination of highprecision gyroscope and star sensor to solve the attitude.After completing the internal geometry calibration and external geometry calibration of the IRS, the physical imaging model of the camera can be built using the orbital position, attitude data, and time system.
In order to facilitate the application of image data, the rational polynomial coefficients (RPC) models are often used instead of strict physical models.In this paper, the affine transformation model in image space based on the RPC is used to study the geometric accuracy of the IRS.

Datasets
In order to study the absolute geolocation accuracy and internal geometric accuracy of the IRS thermal infrared data, three typical areas are selected as experimental areas, including Hunan, Xizang and Xinjiang.The reference data for the area is composed of a 2m accurate DOM and a 10m accurate DEM, and the IRS imagery data spans the four seasons of the year.

Study Area
Referring to the historical weather conditions and regional data acquisition in China, and considering the characteristics of terrain undulation and land cover type, several experimental areas such as Yiyang City in Hunan Province, Nagqu City in Xizang and Hami City in Xinjiang are preferentially selected.
The vegetation in the Hunan experimental area is abundant, and the surface characteristics are clear.The Xizang experimental area has a high altitude and simple land cover.The air in the Xinjiang experimental area is dry, and the land surface cover is mainly Gobi and the vegetation is scarce.

Reference Data
As shown in Table 3, the reference datasets for Hunan, Xizang, and Xinjiang are part of the datasets covering the entire range of China, with planimatric and elevation accuracy of 2.0m and 10m, respectively.The accuracy of the dataset is reliable, and it can meet the accuracy research of IRS data with 16m GSD.

IRS Thermal Infrared Image
In the experimental regions, a total of 80 scenes of thermal infrared data with good quality are used for accuracy analysis from 2022 to 2024, including 40 scenes of images obtained on descending passes mode during the day and 40 scenes of data obtained on ascending passes mode at night. Figure 2 shows an example of a typical IRS images on descending and ascending passes mode in Yiyang City, Hunan Province.(c) The IRS images on descending and ascending passes mode overlaid by geographic coordinates Figure 2.An example of a typical IRS images on descending and ascending passes mode in Yiyang City, Hunan Province.

GCP Extraction
Extracting GCP between emission thermal infrared image data and reflection spectral data using the matching method based on the spatial domain is difficult to obtain satisfactory tiepoints in most cases, while the phase correlation methods based on the frequency domain can generally achieve better results and can be used for studying the geometric accuracy of IRS data.

Phase correlation matching method
The principle of the phase correlation method (Kuglin and Hines, 1975) is based on the Fourier shift property, that is, when there is only a shift between the two image blocks, a linear phase difference will be reflected in the Fourier domain.The phase correlation image matching method uses the Fourier transform to convert the image blocks to be matched into the frequency domain for cross-correlation, which only uses the phase information in the mutual power spectrum of the frequency domain of the image blocks, which reduces the influence of the image grayscale value, has good antiinterference.It pays more attention to the texture information of the image, which can not only overcome the difficulty of matching caused by the change of ground coverage of multitemporal images, but also eliminate the adverse effects brought by different image resolutions.It has excellent performance in the sub-pixel translation estimation of images with different properties and different spectral bands (Foroosh et al., 2002), and has better adaptability to thermal infrared image registration.

Accuracy Analysis
After completing the IRS accuracy evaluation using the data from three experimental areas, further trend analysis is conducted on the internal geometric accuracy and absolute geolocation accuracy.

Absolute Geolocation Analysis
According to the satellite roll angle and the date of data acquisition, Figures 12 and 13   (b) The error distribution of the data obtained on ascending passes mode at night. Figure 13.The absolute geolocation error distribution based on the acquisition date (mean error, in meters).

Conclusion
Aiming at the problem of analyzing the geometric performance of thermal infrared data of ZY-1 02E TIRS instrument, this paper uses the phase correlation matching algorithm based on frequency domain to solve the problem of reliable matching between the data obtained from the emission spectral band and the data of the reflectance spectral band, extracts satisfactory GCP data, and then carries out the evaluation and analysis of the geometric accuracy of thermal infrared image data using a total of 80 scenes data obtained in one year in experimental areas such as Hunan, Xizang and Xinjiang.For the first time, a relatively more accurate and reliable exploration of geometric characteristics has been carried out, especially for nighttime imaging thermal infrared remote sensing image data obtained on ascending passes mode.
The research results indicate that the internal geometric accuracy of IRS image data is good, which meets the specification indicators and can meet application requirements.In terms of absolute geolocation accuracy without GCP support, further improvement is needed, especially there are systematic errors in the ascending data at night, which requires research in the future.Although the absolute geolocation errors are more prominent compared with the widely used Landsat and Sentinel-2 data, the accuracy can be improved through the GCP, and the adverse effects on the remote sensing application of IRS thermal infrared data can be overcome.
(a) The image obtained on descending passes mode during the day.(b)The image obtained on ascending passes mode at night.

Figure 3 ,
Figure 3, 4 and 5 show the Schematic diagram of GCP distribution of Hunan, Xizang and Xinjiang, respectively.Due to the rich texture of Hunan data, the GCP distribution is uniform and the extraction effect is very good, but for the data of Xizang and Xinjiang, the results of GCP extraction are acceptable and can be used for geometric accuracy evaluation, which will not affect the evaluation results, but it is still necessary to continue to explore and continuously improve the GCP extraction strategy in follow-up research.

Figure 7 .
Figure 7.The absolute geolocation error distribution of each scene of Hunan, Xizang and Xinjiang (mean error, in meters).

Figure 8 .
Figure 8.The absolute geolocation error distribution of each scene on descending passes mode of Hunan, Xizang and Xinjiang (mean error, in meters).

Figure 9 .
Figure 9.The absolute geolocation error distribution of each scene on ascending passes mode of Hunan, Xizang and Xinjiang (mean error, in meters).

Figures
Figures 10 and 11 show the internal geometric error distribution of IRS data in the sample and line directions, respectively, based on the satellite roll angle and data acquisition date.Analysis of these graphs shows that there are no systematic errors related to the roll angle and the date of acquisition in both the descending and ascending orbit data.
depict the absolute geolocation error distribution of IRS data in the easting and northing directions, respectively.By observing these distribution, no systematic errors related to the roll angle and the date are found in the data on descending and ascending passes modes.(a) The error distribution of the data obtained on descending passes mode during the day.(b) The error distribution of the data obtained on ascending passes mode at night. Figure 12.The absolute geolocation error distribution based on the roll angle (mean error, in meters).
(a) The error distribution of the data obtained on descending passes mode during the day.

Table 2 .
Main technical specifications of IRS thermal infrared image

Table 3 .
Basic information of the experimental area.