On-Orbit Performance and Hyperspectral Data Processing of the TIRSAT CubeSat Mission

Abstract

A miniaturized hyperspectral camera, developed by integrating a linear variable band-pass filter (LVBPF) with an image sensor, was installed on the TIRSAT 3U CubeSat, launched on 17 February 2024 by Japan’s H3 launch vehicle. The satellite and its onboard hyperspectral camera conducted on-orbit experiments and successfully acquired hyperspectral data from multiple locations. The required attitude control for the hyperspectral mission was also achieved. CubeSat-based hyperspectral missions often face challenges in image alignment due to factors such as parallax, distortion, and limited attitude stability. This study presents solutions to these issues, supported by actual observational hyperspectral data. To verify the consistency of the hyperspectral data acquired by TIRSAT and processed using the proposed method, a validation analysis was conducted.

1. Introduction

Hyperspectral cameras are valuable tools for Earth observation, supporting a wide range of applications such as agricultural, forestry, and ocean remote sensing. These cameras capture both images and spectral data simultaneously, enabling the detection and classification of target conditions [1,2,3,4]. Although several operational systems and development plans exist for hyperspectral cameras on large satellites and the International Space Station [5,6,7,8,9], these remain limited in number, and their effectiveness has not been fully demonstrated. The ideal attributes for hyperspectral remote sensing include full Earth coverage and high observation frequency. This is particularly critical in agricultural monitoring, where frequent and regular observations are essential for effective crop management. However, when relying on a single satellite in Low Earth Orbit, the observation frequency remains low (approximately once every two weeks). A promising solution is the deployment of a constellation of CubeSats equipped with hyperspectral cameras.

Traditionally, spaceborne hyperspectral cameras have required large telescopes to gather sufficient light for spectroscopic observations. Recent advances in image sensor technology have significantly improved sensitivity, making hyperspectral imaging feasible even with smaller telescopes. Consequently, new initiatives are underway to deploy microsatellites and CubeSats with hyperspectral capabilities. In recent years, numerous start-up companies have entered the Earth observation sector, focusing on the mass production of small satellites, including microsatellites, and offering services through satellite constellations. A satellite constellation refers to a system in which multiple microsatellites are interconnected and operated in a coordinated manner. Start-up companies are increasingly constructing such constellations to create a global Earth observation network [10,11,12,13,14]. Integrating hyperspectral cameras into CubeSats and microsatellites can significantly enhance the temporal resolution of Earth observation data collection.

Examples of compact hyperspectral cameras include HYPSO-1, developed by the Norwegian University of Science and Technology [15], and HyperScout-1, developed by Cosine Remote Sensing B.V. [16]. Dragonette, developed by Canada’s Wyvern, has a wavelength resolution of 20 nm, resulting in coarser spectral data than some alternatives. However, it achieves a ground sampling distance of 5.3 m, allowing for higher spatial resolution imaging [17]. Kuva Space, a Finnish hyperspectral satellite start-up, aims to provide global images two to three times daily using a constellation of 100 6U CubeSats called Hyperfield, which will cover the spectral range from visible near-infrared to short-wavelength infrared [18]. Planet’s Tanager satellite is equipped with a hyperspectral sensor that spans wavelengths from 400 nm to 2500 nm. While many small satellites are limited to visible and near-infrared observations, Tanager’s ability to extend into the short-wavelength infrared range (up to 2500 nm) is a key advantage [19]. In this range, methane absorbs between 2150 and 2450 nm and CO₂ absorbs between 1980 and 2100 nm. This capability enables the distinct detection of CO₂ and methane, contributing to efforts toward carbon neutrality.

In this context, the University of Fukui and SEIREN Co., Ltd. developed and installed a hyperspectral camera on a CubeSat named TIRSAT and conducted an in-orbit demonstration. TIRSAT is a 3U CubeSat developed primarily by SEIREN Co., Ltd., managed by Japan Space Systems as a commissioned project for Japan’s Ministry of Economy, Trade and Industry (METI). TIRSAT’s primary payload is a bolometer-type camera intended for thermal infrared measurements of Earth’s surface temperature, including heat sources such as factories to estimate operational status. The hyperspectral camera developed by the University of Fukui was installed as an additional payload for in-orbit demonstration. TIRSAT was successfully launched into a sun-synchronous sub-recurrent orbit at an altitude of approximately 680 km on 17 February 2024 by Japan’s H3 launch vehicle. After orbital insertion, the satellite’s basic functions were verified, and mission operations began. The hyperspectral camera on TIRSAT incorporates a linear variable band-pass filter (LVBPF), enabling significant miniaturization suitable for CubeSat integration [20]. This advancement allows for the convenient spectral measurement of ground features and expands the potential of hyperspectral imaging applications. The camera can be easily integrated into CubeSats carrying multiple instruments and is particularly effective for hyperspectral CubeSat constellations.

This paper presents the detailed performance specifications of the hyperspectral camera, along with the satellite’s specifications and attitude-control results related to TIRSAT observations. It also provides an analysis of the observational data and the in-orbit data processing methods. In particular, LVBPF-based hyperspectral data processing requires precise image alignment, which is critical for data accuracy. This paper proposes a geometric transformation-based alignment method to correct image distortion and discusses the validity of the acquired hyperspectral data.

2. Development of a Linear Variable Band-Pass Filter-Based Hyperspectral Camera

2.1. Observation Method of LVBPF-Based Hyperspectral Camera

Conventional hyperspectral cameras typically consist of a telescope and a slit-based spectrometer, which includes a collimator, a dispersion element (usually a prism or grating), and an objective lens. Due to the numerous optical components required, such cameras are difficult to miniaturize, particularly for space applications where suitable optical materials are limited. Recently, a miniaturization method employing a linear variable band-pass filter (LVBPF)—a band-pass filter whose wavelength transmittance changes linearly with position—has been proposed [21,22,23,24]. This approach enables the acquisition of both spatial and spectral information using a single compact filter.

The configuration and observation mechanism of the LVBPF-based hyperspectral camera are shown in Figure 1. While traditional grating-based hyperspectral cameras offer high spectral resolution, they reduce the amount of light reaching the image sensor due to the dispersion of incident light into multiple orders. In contrast, the sensitivity of an LVBPF-based camera primarily depends on the transmittance of the filter, resulting in higher sensitivity. However, as with conventional types, LVBPF-based hyperspectral cameras also require push-broom scanning. The filter’s wavelength transmittance varies in the along-track direction, necessitating image acquisition by scanning the scene spatially to capture data at the same wavelength. These spectral images are then assembled into three-dimensional hyperspectral data.

2.2. Specification

In this study, a miniaturized hyperspectral camera utilizing an LVBPF was developed and installed as an additional payload on the 3U-CubeSat TIRSAT. Figure 2 shows the development of the LVBPF-based camera. The unit, including the lens, filter, and image sensor, was approximately 3 cm long and weighed only 35 g. This compact design allows integration not only into 3U-CubeSats but also into 1U-CubeSats.

Table 1. Specifications of the developed hyperspectral camera.

Item	Specification
Size	3.6 cm × 3.6 cm × 2.4 cm
Weight	35 g
Ground sampling distance	450 m/pixel
Swath	460 km
Available wavelength range	400–770 nm
Spectral sampling distance	5 nm
Spectral resolution	18.1 nm
Number of bands	75 bands
Focal length of telescope lens	8 mm
F-number of telescope lens	F/2.5
Detector format	1280 along track × 1024 cross track
Pixel size	5.3 μm
Dynamic range	8-bit

lists the camera specifications. Since the hyperspectral camera was designed as an auxiliary instrument aboard a multi-mission CubeSat, a small lens with coarse resolution was selected to ensure a wide ground coverage of up to 450 km.

Table 2. Specifications of the LVBPF.

Item	Specification
Size	10.1 mm × 8 mm
Thickness	0.5 mm
Spectral detection range	380–850 nm
Dispersion	67.7 nm/mm
Peak transmission	65%
Spectral blocking property	<1%
Half bandwidth	15 nm at 430 nm 20.6 nm at 780 nm

presents the LVBPF specifications. The filter, which was tailored to match the size of the image sensor, enables coverage of visible to near-infrared wavelengths. As shown in Figure 2, the filter has not been deposited directly onto the sensor but mechanically fixed. To achieve minimal spacing, the image sensor’s cover glass has been removed, and the filter is positioned as close as possible to the sensor.

This design offers flexibility, allowing the easy modification of spectral resolution and wavelength range by changing the filter, sensor, or telescope lens. The image sensor used was the UI-1242LE-NIR from IDS Imaging Development Systems GmbH. It features a global shutter, making it suitable for space applications. The system utilizes a USB 2.0 interface, and imaging control was executed using a Raspberry Pi Compute Module 3. Additionally, onboard image processing using Python 3.11 and OpenCV 4.10 was supported, resulting in a highly flexible camera system based on consumer-grade technology.

2.3. Pre-Flight Test on Ground

2.3.1. Imaging Performance

The imaging performance of the hyperspectral camera flight model was evaluated in ground tests. Point spread functions (PSFs) were measured using a collimator with a pinhole chart illuminated by green light. The environmental tests included random vibration, thermal cycling, and thermal vacuum testing. The thermal tests simulated expected space conditions, ranging from −10 °C to +50 °C. The random vibration test was conducted with the camera mounted on the TIRSAT flight model, following the launch-vehicle specifications.

Figure 3 shows the PSFs measured before and after the environmental testing. Prior to the testing, the PSFs exhibited a standard deviation (σ) of 1.4 pixels, indicating good imaging quality. After the tests, the vertical-axis PSF increased slightly, to 1.6 pixels, while the horizontal axis remained at 1.4 pixels. The overall change in this PSF was 0.2 pixels. Despite this minor deviation, the camera maintained sufficient imaging performance, demonstrating its robustness under launch and space conditions.

2.3.2. Spectral Calibration

The spectral performance was evaluated using an integrating sphere equipped with a 10 nm full-width at half-maximum (FWHM) band-pass filter and illuminated by a halogen lamp. The band-pass filter allowed specific wavelength ranges to pass, enabling the assessment of the camera’s spectral resolving ability. A halogen lamp provided stable, broad-spectrum illumination for consistent and accurate spectral measurements.

Spectral images were acquired by the hyperspectral camera flight model at discrete wavelengths corresponding to band-pass filters centered at 400 nm, 450 nm, 500 nm, 550 nm, 600 nm, 650 nm, and 700 nm, as shown in Figure 4. These images were analyzed to calibrate the spectral data and evaluate the resolution at each wavelength. The raw images exhibited noticeable distortion, characterized by a pronounced smile effect. This distortion, commonly caused by optical components such as lenses, was uncorrected in the acquired images. Additionally, the positional variation in the wavelength transmission of the LVBPF contributed to the overall distortion.

Figure 5 quantitatively illustrates the smile distortion. The centroid of the signal was calculated for each image column, and the center positions were plotted against the row number. The centroid of the signal was calculated using Equation (1), where i is the column number (horizontal direction), j is the row number (vertical direction), $\lambda$ is the measured spectral band, and C is the centroid of the signal at each row and each spectral band. The smile distortion shown in Figure 5 is defined as the difference between the centroid of the signal at row j = 512 pixels and that at each row.

$$C(j,\lambda )={\displaystyle \frac{\sum _{i=0}i·I(i,j,\lambda )}{\sum _{i=0}I(i,j,\lambda )}}$$

(1)

$$Smile distortion(j,\lambda )=C\left(512,\lambda \right)-C(j,\lambda )$$

(2)

The resulting data revealed a shift of approximately 20–30 pixels from the center to the image edge. Using a calibration coefficient of 0.336 nm/pixel, calculated at the central row (512 pixels), this displacement corresponds to a wavelength shift of approximately 6.72–10.08 nm. Although relatively small, this shift significantly affects the spectral accuracy and cannot be ignored.

To address this, a wavelength calibration formula was calculated for each row based on the spectral data in Figure 4. The wavelength calibration formula for each row was derived by linear interpolation between the plotted points corresponding to the seven spectral bands shown in Figure 4. The results were used to generate a calibrated wavelength map, presented in Figure 6. This map allows for accurate wavelength assignment across the entire image and compensates for distortion-related errors. The implications of this calibration will be further discussed in the section on observed image analysis.

2.3.3. Spectral Performance Measurement

Figure 7 shows the spectral resolution at the image center (row 512) for each band. The FWHM values, representing the spectral resolution, are as follows: 49.5 pixels (16.6 nm) at 400 nm; 56.5 pixels (18.9 nm) at 450 nm; 61.1 pixels (20.6 nm) at 500 nm; 68.3 pixels (23.0 nm) at 550 nm; 87.1 pixels (29.3 nm) at 600 nm; 110.7 pixels (37.2 nm) at 650 nm; and 113.0 pixels (38.0 nm) at 700 nm. The spectral resolution has been interpreted as the width of the light beam passing through the filter, calculated as the product of the reciprocal F-number, dispersion, and distance between the filter and the focal plane. With an F-number of F/2.5, a filter-to-focal-plane distance of 0.67 mm, and a dispersion of 67.7 nm/mm, the theoretical spectral resolution is approximately 18.1 nm.

The measured and theoretical values show good agreement at the shorter wavelengths; however, the resolution becomes coarser at the longer wavelengths. This trend likely resulted from the wavelength-dependent half-bandwidth of the LVBPF, which is broader on the long-wavelength side, increasing the allowable incident angle and reducing resolution. These factors increase spectral cross-talk among adjacent pixels, degrading spectral purity and reducing the effective number of bands.

To meet the mission’s requirements, the camera was equipped with a fast F/2.5 lens to capture high-quality Earth images. However, this lens increased the beam width through the filter, resulting in coarser resolution. The dependence of the spectral resolution on the lens F-number was confirmed and should be considered in future optimization of LVBPF-based hyperspectral camera design.

3. Description of Satellite Bus for TIRSAT

3.1. Introduction of TIRSAT

This section outlines the specifications of the TIRSAT satellite. An external view of the satellite flight model is shown in Figure 8. The satellite measured 117 mm × 117 mm × 381 mm before the deployment of its solar array panels and had a mass of 4.97 kg. It was equipped with a 1U-sized (approximately 100 mm × 100 mm × 100 mm) bolometer-type camera that operated in the thermal infrared wavelength range as its primary mission payload. Accordingly, the satellite bus occupied approximately 2U of volume. The main specifications of the TIRSAT satellite bus are summarized in

Table 3. Specifications of the TIRSAT satellite bus.

Item	Specification
Size	117 mm × 117 mm × 381 mm
Weight	4.97 kg
Attitude Determination and Control Subsystem	Three-axis stabilization control using geomagnetic sensor, MEMS gyroscope, three-sun sensor, GPS receiver, magnetic torque, and reaction wheel
Electrical Power Subsystem	Solar array panel: four deployable panels, four body-mounted panels Maximum power generation: 20 W Typical power consumption: 10 W Battery: 5.8 Ah, nominal 8 V (lithium ion battery)
Communication Subsystem	Telemetry/command: S-band Command uplink: 4 kbps, Telemetry downlink: 4 kbps–64 kbps Mission data downlink: X-band (5 Mbps, 10 Mbps)
Orbit	Sun-synchronous sub-recurrent orbit Altitude: 680 km (approximately); inclination: 98 degrees

, and its mechanical configuration is illustrated in Figure 9.

The satellite bus, named SRN-3U, was developed by SEIREN Co., Ltd. in collaboration with the University of Fukui and the University of Tokyo. It was based on the flight heritage of the TRICOM-2 satellite bus [25,26], with enhancements including high-speed communication capability, high-precision attitude control, and deployable solar panels. The satellite was designed to be compatible with the ISI Space Quad Pack Type III CubeSat deployer, which matches the release mechanism dimensions of the H3 launch vehicle.

The ground station, equipped with S-band and X-band antennas, was located at ArkEdge Space in Japan. The satellite operations, including pass scheduling, mission planning, and data analysis, were conducted at SEIREN’s satellite operation center in Fukui, Japan. The hyperspectral data acquired by TIRSAT was analyzed at the University of Fukui.

TIRSAT was equipped with an X-band transmitter capable of downlink speeds of up to 10 Mbps. It used two types of X-band antennas: a broad-directional low-gain antenna (LGA) and a highly directional medium-gain antenna (MGA). For telemetry and command, an S-band transponder was employed, with antennas mounted on the +Y and −Y panels. These antennas were connected via an RF combiner and splitter, enabling communication in nearly all directions by covering half the space in the +Y or −Y direction.

The satellite also featured four deployable solar panels that could generate up to 20 W of effective power. In addition, body-mounted solar cells were installed on four sides of the satellite. This redundant configuration supported stable power generation during the initial operational phase and ensured the power budget was maintained even if sun-pointing control was lost. The satellite’s battery consisted of lithium ion cells arranged in a two-series, two-parallel configuration.

The attitude determination and control subsystem (ADCS) was based on a compact 1U-sized integrated module [27], which was customized for this mission by excluding a star tracker. The onboard computer (OBC) was derived from the TRICOM-2 project and featured a fault-tolerant design that included automatic rebooting approximately every four hours via a reset counter, regardless of operational mode. This served as a countermeasure against single-event faults. The status data was recorded in the high-speed, non-volatile memory and retrieved immediately after rebooting to ensure continuous operation.

Both the ADCS and OBC employed command-centric architecture (C2A) for software implementation [28], providing high flexibility and ease of in-orbit reconfiguration. This architecture defined all satellite actions through commands, enabling functional changes without memory rewriting and allowing software updates to be implemented efficiently in orbit.

3.2. Attitude Determination and Control Subsystem of TIRSAT

The block diagram of the ADCS is presented in Figure 10, and its specifications are summarized in

Table 4. Specifications of the ADCS onboard TIRSAT.

Item	Specification
Reaction wheel	Three-axis mounted Max angular momentum: 3 mNms (nominal)               5 mNms (peak)
Magnetic torquer	Three-axis mounted Magnetic moment: 0.35 Am2
Sun sensor	Three-surface mounted Accuracy: ≤0.5°(3σ)
Fine geomagnetic sensor	Resolution: 13 nT
Fine gyroscope	Random noise: 4.36 × 10−5 rad/s (1σ)
Microcontroller	Clock: 80 MHz; ROM: 512 kiB; RAM: 128 KiB

. TIRSAT employed a three-axis attitude-control system composed of two microelectromechanical systems (MEMS) gyroscopes, three sun sensors, two geomagnetic sensors, three-axis reaction wheels, and three-axis magnetic torquers, all integrated into a compact ADCS module. Redundant units were included for both the gyroscopes and geomagnetic sensors, with one unit offering fine accuracy and the other coarse accuracy.

Although the ADCS module included a Global Navigation Satellite System Receiver (GNSSR), the onboard position determination was primarily performed using Simplified General Perturbations Satellite Orbit Model 4 (SGP4) and two-line elements (TLEs) uploaded from the ground station. During the orbital operations, SGP4 with the TLEs was predominantly used for position estimation.

The ADCS software supported multiple control modes, including detumbling using only magnetic torquers, three-axis nadir pointing, and three-axis sun pointing. Each pointing mode was achieved using a combination of three-axis reaction wheels and magnetic torquers or with magnetic torquers alone. Thus, even in the event of a reaction-wheel failure, attitude control could continue with reduced accuracy. Additionally, an offset angle could be applied to each of the three rotational axes—roll, pitch, and yaw—in any pointing mode. As the ADCS was not equipped with a star tracker, the attitude determination relied mainly on the sun sensors, which offered an accuracy of approximately 0.5°.

A software-in-the-loop simulation (SILS) was conducted to validate the nadir-pointing control mode. The initial angular velocity for all axes was set to 0.1°/s, and the simulation commenced under daylight conditions. The satellite was assumed to be in a sun-synchronous orbit at an altitude of 680 km. The initial attitude conditions were defined such that in Case A, the −X surface was oriented toward the sun, while in Cases B to D, offset angles were applied to each axis. As shown in Figure 11, the simulation results indicated that the attitude convergence time was within 400 s and the pointing error was maintained within 2°. Furthermore, after 3000 s, the attitude remained stable, with a pointing error of less than 1°.

The target values for the attitude control in the hyperspectral imaging mission are presented in

Table 5. Required attitude-control accuracy and stability for the ADCS.

Item	Attitude-Control Accuracy (Roll, Pitch)	Attitude-Control Accuracy (Yaw)	Attitude Stability (Roll)
Requirements	7.3°	5.8°	0.048°/s

. The target pointing-control angle was set to 7.3°, corresponding to a positional difference of 200 pixels between the target location and the captured image. Because the hyperspectral camera utilized a linear variable band-pass filter (LVBPF), the spectral direction pixels had to align with the along-track direction. As a result, high precision was required for the yaw-axis attitude control and roll-axis attitude stability. A conceptual diagram illustrating these requirements is shown in Figure 12.

The required yaw-axis attitude accuracy was derived using Equation (3), where α denotes the acceptable overlap ratio between the first and last images and the IFOV (instantaneous field of view) represents the angle per pixel, which was 0.038° for the hyperspectral camera. Assuming an allowable overlap ratio of 70%, the required yaw-axis pointing accuracy was calculated to be 5.8°. Moreover, the roll-axis angle during the imaging needed to be maintained within 5.8° of its initial value.

The required image-capture duration was calculated using Equation (4). At an orbital altitude of 680 km, TIRSAT’s orbital velocity was approximately 7.5 km/s, and its ground sampling distance (GSD) was 450 m/pixel, as shown in Table 1. Given an along-track pixel count of 1024, the time required to capture an image over the full spatial field of view and the full spectral range was estimated to be approximately 120 s. During this period, the roll-axis angle had to remain within 5.8°, yielding a required attitude stability of 0.048°/s. Based on the SILS simulation results, the ADCS of TIRSAT was predicted to have sufficient capability to meet these pointing and stability requirements.

$$∆\phi =IFOV\times {\displaystyle \frac{(1-\alpha )\times AcrossPixels}{2}}$$

(3)

$$CaptureTime={\displaystyle \frac{2\times AlongTrackPixels\times GSD}{Orbit velocity}}$$

(4)

3.3. Mission Data Handling and Communication Subsystem

The control of the mission payload, including the hyperspectral camera, was managed by the mission data handling unit (MDH), which was developed using a field-programmable gate array (FPGA). Mission data transmission to the ground station was carried out using the X-band transmitter. The block diagram of the mission-processing and communication subsystem is shown in Figure 13. As previously mentioned, the hyperspectral camera was controlled by a Raspberry Pi. The captured images were transferred to the MDH unit via Serial Peripheral Interface (SPI) communication. The MDH unit stored the received raw images in the embedded multimedia card (eMMC) memory, which served as the data storage. The eMMC memory had a total capacity of 8 GB, with 1 GB allocated for the hyperspectral camera, which could store up to 10 hyperspectral datasets in orbit. The data read from the eMMC memory was transmitted to the ground station through the X-band transmitter, using a data format compliant with the Consultative Committee for Space Data Systems (CCSDS) standard.

To perform map projection, satellite attitude and position data corresponding to the image timestamp were required. These ancillary data were transmitted from the ADCS module and transferred via the onboard computer (OBC) and MDH unit to the hyperspectral camera, which included the Raspberry Pi. The ancillary data were appended to the header of the raw image and also stored in the eMMC memory. The MDH unit and eMMC memory utilized commercial off-the-shelf (COTS) components, which were newly developed. It was confirmed that these components had a radiation tolerance exceeding 20 krad. Proton-irradiation experiments also evaluated their single-event resilience, and the results confirmed that these components were sufficiently capable of functioning in orbit.

The X-band transmitter supported downlink communication speeds of 5 Mbps or 10 Mbps using Offset Quadrature Phase-Shift Keying (OQPSK) modulation. The output power of the transmitter could be set to 1 W or 2 W. TIRSAT was equipped with a single patch-array antenna as the low-gain antenna (LGA) and a 2 × 2 patch-array antenna as the medium-gain antenna (MGA). These antennas could be switched using an RF switch. The maximum gain of the LGA was 4 dBi, while the MGA had a maximum gain of 10 dBi. The LGA was mounted on the +Y surface, and the MGA was mounted on the −Y surface. The half-power beamwidth was ±35° for the LGA and ±20° for the MGA. While attitude control was required during the X-band communication, coarse attitude control sufficed when using the LGA.

4. Data-Construction Method for LVBPF-Based Hyperspectral Data

Images acquired using an LVBPF capture different transmission wavelengths in the along-track direction. Therefore, a single image represents a mixture of spatial and spectral information. Spectral images are captured at various positions using push-broom observations to capture the same wavelength, with the overlapping parts of each image synthesized to generate a spectral image, forming hyperspectral data. In general, push-broom imaging sequentially synthesizes images at regular intervals. However, factors such as lens distortion, parallax at the capture position, and attitude instability can affect image alignment. As a result, when overlapping regions are synthesized using only temporal capture information, the resulting hyperspectral data exhibit distortions, and the individual spectral images do not align properly. To address this, we estimated overlap coordinates using image feature-point matching and homography transformation, synthesizing the images to generate corrected hyperspectral data.

Figure 14 illustrates the data-construction method implemented on the ground segment. This process is divided into two stages. The first step involves feature-point matching to automatically detect the overlapping regions of each image. Displacement between the images is then estimated, and alignment is achieved using affine transformation. The images used in this step are raw data captured by the LVBPF-based hyperspectral camera, referred to as RAW images. Each RAW image is accompanied by a wavelength map that indicates the transmission wavelength for each pixel, as shown in Figure 6. These RAW images contain distortions due to lens and parallax effects and represent unprocessed data. Feature-point matching is performed between the first image (I_n) and the following image (I_n+1). To facilitate the feature-point matching, the RAW images undergo image enhancement processes such as contrast enhancement and sharpening. Gamma correction is applied for the contrast enhancement, and unsharp masking is used for the sharpening. Accelerated Keypoint and Descriptor Extraction (AKAZE) is employed for the feature-point detection.

After applying these enhancements, the top three feature-point pairs identified are selected and affine transformation is performed. The wavelength map is transformed using the same matrix applied to the images. The wavelength map represents physical wavelengths, but it needs to be mapped as discrete band numbers. In this study, mapping was performed with a 5 nm sampling interval. Because the wavelength interval per pixel was significantly smaller than 5 nm, no discrete band numbers were skipped during this quantization process. This process is repeated for each image captured by the hyperspectral camera, yielding affine-transformed RAW images and the corresponding wavelength maps, which are stored in a three-dimensional structure to form the hyperspectral data.

At this stage, the RAW image and wavelength map are affine-transformed, meaning only operations such as translation or rotation are applied. While this step results in roughly aligned hyperspectral data, distortions due to parallax and other factors are not removed, and the images do not align perfectly. The next step involves performing a homography transformation based on feature-point matching between the generated hyperspectral data to correct position and distortion simultaneously. The result is distortion-free and properly aligned hyperspectral data.

The spectral images for each band are read from the hyperspectral data, and the positions of the feature points are extracted. Feature-point detection and extraction are performed using the same method as in the first step. The complete set of extracted feature-point pairs is then used with Random Sample Consensus (RANSAC) to compute the homography transformation matrix. RANSAC is an iterative algorithm used to estimate a model from data containing outliers. All spectral images, except for the reference image, are transformed using this matrix and stored in the final hyperspectral data.

5. On-Orbit Results

5.1. Observation Result and Validation of Data-Construction Method

TIRSAT successfully pointed to the nadir and scanned and acquired ground-surface images using the LVBPF-based hyperspectral camera. The first of the hyperspectral data was acquired on 29 March 2024, with data acquisition continuing thereafter, incorporating parameter adjustments for precise attitude control. Several hyperspectral datasets acquired by TIRSAT are presented. These datasets were constructed using the method outlined in Figure 14. Figure 15 shows a composite color image created using the red, green, and blue spectral bands, acquired near Baja California, Mexico. The image reveals the spectral data for various regions: the ocean (e.g., location A), urban areas (e.g., location B), and mountainous areas (e.g., locations C and D). Next, Figure 16 displays data captured over Romania and Bulgaria, including the spectral data for vegetation areas (e.g., locations A, B, and D) and bare land (e.g., location C). Figure 17 shows a false-color image generated using data collected near Nevada, USA, including spectral information from vegetated areas (e.g., locations A, B, C, and D).

To demonstrate the effectiveness of the proposed hyperspectral data-construction method, several methods were compared: (1) the sequentially synthesized method at regular time intervals; (2) affine transformation; (3) homography transformation; and (4) the proposed method, which combines both affine and homography transformations.

Table 6. Parameters for each hyperspectral data-construction method.

Item	(1) Sequentially Synthesized	(2) Affine	(3) Homography	(4) Proposed Method
Image enhancement	-	Gamma correction (gamma = 2.0)
		Unsharp masking
Feature detection		AKAZE
Transformation		Affine	Homography and RANSAC	(1) + (2)
Library		OpenCV, pillow

provides the parameters for each method. Figure 18 illustrates examples of positional shifts caused by the data-construction methods. The base data, taken near Kazakhstan, show comparisons of overlap for the red, green, and blue bands at three locations using the four methods. The results reveal that at the central location (b), all methods provided relatively accurate alignment. However, significant positional shifts were observed at the other locations (a and c), where the bands did not overlap correctly, except for the proposed method.

If orbital velocity were the only factor, alignment could be achieved using the sequentially synthesized method or affine transformation, but distortion would still occur. Specifically, the hyperspectral camera installed on TIRSAT is a wide-angle camera, and due to the Earth’s curvature and parallax, the height difference between the center and edges of the image exceeds 40 km. As a result, affine transformation alone cannot remove distortion. Homography transformation can correct geometric distortions in the image, but it cannot fully account for distortions caused by Earth’s curvature and parallax, nor can it completely convert the image into a flat representation. Therefore, as proposed in this research, it is best to first generate hyperspectral data with coarse alignment using affine transformation to account for parallel movement, followed by homography transformation to correct distortions such as parallax. These results confirm the effectiveness of the proposed hyperspectral data-construction method.

5.2. On-Orbit Analysis Results of the Satellite Bus Performances

5.2.1. Evaluation of Attitude Precision and Stability

The attitude-control accuracy and stability of TIRSAT were analyzed based on the imaging data captured by the hyperspectral camera and the attitude determination data from the ADCS. The attitude-control accuracy was calculated based on the geometric correction accuracy of the images. Geometric correction was first applied to the raw images using the attitude determination data, which included quaternions and satellite position, followed by map projection. Landsat-9 OLI-2 served as the reference data for verifying the accuracy of the geometric correction. Mexico, selected for its clear coastal features, was chosen as the sample location. Figure 19 shows the map projection results for the raw image and the geo-referenced image captured by Landsat-9. Figure 19 illustrates the map projection results of the raw image captured by the TIRSAT hyperspectral camera compared with the geo-referenced image from Landsat-9. Six ground-control points (GCPs) were manually selected, and the distances between the pairs were calculated and used to determine the geometric correction errors. The results are summarized in

Table 7. Geo-referenced coordinates of satellite images and geometric correction error distances.

Location	TIRSAT		Landsat-9		Distance [km]
	Latitude [°]	Longitude [°]	Latitude [°]	Longitude [°]
A	31.7117	−113.8237	31.3447	−113.6420	44.297
B	30.9371	−114.7187	31.4722	−114.9723	64.209
C	31.4933	−114.0276	31.9123	−114.1741	48.612
D	33.0178	−114.7806	32.4926	−114.8430	58.692
E	32.1983	−114.7021	31.6918	−114.5939	57.241

. The average geometric correction error was found to be 54.6 km, which, when converted to a pointing error, corresponds to approximately 4.6°. This pointing error includes both attitude-control errors and errors from the camera’s shutter timing and alignment. However, the results satisfy the required values presented in Table 5.

The attitude stability was evaluated using the time history of the angular velocity from the gyroscope onboard the ADCS. Figure 20 shows the time histories of the angular velocities for all three axes, with an expanded graph for the roll axis. The time axis is referenced from the start of the imaging, set as the zero-second point. TIRSAT’s hyperspectral camera specifies a stability requirement for the roll axis of approximately 0.048°/s. The results show that, from the start to the end of imaging over a period of approximately 120 s, the angular velocity remained below 0.04°/s, thus meeting the stability requirement. Based on these analysis results, it can be concluded that the attitude-control system of TIRSAT has sufficient performance to perform imaging with the hyperspectral camera.

5.2.2. Thermal Analysis

This section presents the thermal environment of the hyperspectral camera, based on the thermal analysis results. While resistance temperature detectors (RTDs) have been installed on each panel and certain components, the hyperspectral camera onboard TIRSAT does not have its own dedicated RTD. However, the thermal simulations confirmed that the aluminum panel located adjacent to the camera exhibits similar temperature characteristics, allowing it to serve as a proxy for monitoring the camera’s temperature. Thus, the temperature of the aluminum panel near the camera was monitored and evaluated in orbit as a proxy for the camera’s temperature.

Figure 21 shows a comparison between the thermal simulation results and on-orbit measured data when the satellite was at an altitude of 680 km and in a tumbling state. The satellite can maintain its power budget even without sun-pointing, as its attitude-control mode allows it to remain in a tumbling state with no attitude control during normal operations. This graph includes the simulated temperature of the hyperspectral camera and the panel near the camera, along with the on-orbit measured temperature of the panel. The simulation results generally match the on-orbit measurements, with the simulated camera temperature ranging between +6 °C and −5 °C. However, around the 65,000 s mark, deviations are observed, with temperatures at approximately +10 °C higher and −5 °C lower than the simulated temperatures at certain points. Further investigation is required to determine the cause of these discrepancies. Despite this, the hyperspectral camera was successfully operated in ground tests within the temperature range of −10 °C to 50 °C, indicating sufficient thermal margins for its operation.

5.2.3. Power Control Subsystem

TIRSAT is equipped with four deployable solar panels, which were successfully deployed in orbit, confirming a maximum effective power output of 16 W. Figure 22 illustrates the power generation status after the deployment of the solar array panels. Each deployable solar panel generated 4 W, resulting in a total power output of 16 W from all four panels. Power generation from the body-mounted solar array panels, previously obscured by the deployable panels, was also confirmed.

The power consumption increased during the hyperspectral camera observations and X-band data transmission. To evaluate the power budget, Figure 23 and Figure 24 show the battery voltage profiles during the observation and transmission periods, respectively. The onboard battery has a maximum voltage of approximately 8.0 V and maintains a nearly fully charged state under steady sunlight conditions. Figure 23 shows the battery voltage and power consumption of the hyperspectral camera during the observations. The camera consumes approximately 2 W, and the battery voltage decreases by only about 0.5 V. This minor drop does not pose any issues, as the voltage returns to its original level once the observation is completed. Figure 24 shows the battery voltage and power consumption of the X-band transmitter during the data transmission. The transmitter consumes over 10 W, and the battery voltage drops by approximately 0.9 V during transmission, indicating a significant decrease. The battery’s state of charge (SOC) is estimated to decrease from approximately 80% to 72% during this period. Nevertheless, the data transmission is sustained for approximately 7 min. Given that the satellite’s visible operation window is around 10 min, a 7 min transmission duration is considered reasonable. These results confirm that TIRSAT’s power budget is sufficient to support both hyperspectral observations and X-band data transmission.

5.3. Comparison of Hyperspectral Data Across Two Time Periods

A comparison of the hyperspectral data from two time periods was conducted using the datasets from the coastal area of Mexico, as shown in the previous section, projected onto a map. The acquisition times were nearly identical, with only a three-month difference between the datasets (November 2024 and January 2025). The spectral data were extracted from multiple locations: the vegetation areas of the mountainous region (locations A and B), the intermountain region (locations C and D), and the coastal region (locations E and F). The comparison results are shown in Figure 25 and Figure 26. Since the hyperspectral data were not radiometrically corrected, the comparison was based solely on Digital Numbers (DNs).

Although the data were captured during the winter season, no significant absorption features are observed in the vegetation areas. However, it is noticeable that the red edge between 700 nm and 750 nm has shifted to the left. In the intermountain region, the values in the red to near-infrared range (600 nm to 750 nm) are found to be higher. For the coastal region, minimal differences have been observed between the two time periods.

To compare the differences in the vegetation areas, vegetation analysis maps for the two periods were created. Typically, comparisons are made in the near-infrared and red absorption bands for vegetation analysis. However, since there was little red absorption during the winter season, the comparison was performed using a quantity inspired by the Green Normalized Difference Vegetation Index (GNDVI), which involves the near-infrared and green visible light bands. Because the TIRSAT hyperspectral data were not radiometrically calculated and did not represent reflectance values, a new quantity, termed the uncalibrated GNDVI (uGNDVI), was defined using the raw, uncalibrated digital signals, as shown in Equation (5). Here, D_G and D_IR represent the DNs at 547.8 nm for the green (D_G) band and 737.8 nm for the red-infrared (D_IR) band, respectively. The analysis results are shown in Figure 27. The results show a decrease in the uGNDVI across the entire analysis area from November to January. However, the uGNDVI in the vegetation areas of the mountainous region have increased. While a detailed discussion of the analysis results is not included in this paper, these findings confirm that the comparison analysis between the two time periods was successfully performed, with no issues regarding geospatial accuracy or the stability of the spectral data acquisition needed for comparison.

$$\mathrm{u}\mathrm{G}\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}={\displaystyle \frac{D_{IR}-D_G}{D_{IR}+D_G}}$$

(5)

6. Conclusions

This paper presents the results of observations made using the LVBPF-based hyperspectral camera installed on TIRSAT. The hyperspectral camera covers the visible to near-infrared spectral range and has successfully acquired hyperspectral data over a broad ground surface area. Regular imaging experiments are ongoing, and this paper outlines and validates a proposed method for constructing LVBPF-based hyperspectral data. This method involves removing coarse alignment using affine transformation to account for parallel movement, followed by the application of homography transformation to correct distortions such as parallax. The effectiveness of this data-construction method was validated by the on-orbit results, as demonstrated by comparative experiments showing that only the proposed combination of affine and homography transformations achieved consistent band alignment across the entire image.

Despite TIRSAT being a small 3U-CubeSat, its satellite bus has demonstrated the ability to fully meet the observation requirements of the camera. This confirms that high-precision attitude control can be achieved at a low cost, as demonstrated by the stable pointing during multiple operations and the acquisition of the hyperspectral data. Additionally, other subsystems, such as the X-band communication system and mission data handling, have proven sufficient to operate the hyperspectral camera, as demonstrated by the on-orbit results.

To validate the consistency of the hyperspectral data acquired by TIRSAT, a successful comparison analysis between two time periods was conducted. No issues were found with the geospatial accuracy or spectral data stability. The results also revealed temporal variations in the spectral characteristics across multiple land cover types, including vegetation areas, further confirming the robustness of the data for time-series analysis.

In conclusion, both the LVBPF-based hyperspectral camera and the TIRSAT satellite bus have demonstrated their capabilities in orbit. Future plans include further verifications for practical use, radiometric calibrations, and obtaining spectral reflectance data. The LVBPF-based hyperspectral camera also offers flexibility by allowing filter replacements to change spectral characteristics and the convenience of replacing the telescope lens. Additionally, the high optical transmission of the filters enhances spatial resolution. We plan to develop a hyperspectral camera with higher spatial resolution suitable for CubeSat applications.