Testing a computed tomography imaging spectrometer for Earth observations on the HEIMDAL stratospheric balloon mission

1 Introduction

Stratospheric high-altitude balloons (HABs) offer a reusable and adaptable platform for remote sensing that can accelerate the availability of high-resolution Earth observation (EO) data (Floberghagen, 2024). At an altitude of 20–35 km, they operate at a distance of approximately 20 km from the surface of the Earth compared to low Earth orbit satellites. This allows HABs to acquire higher-resolution optical imagery than satellites with comparable optics while still covering significantly larger areas and staying afloat longer than conventional drones. Therefore, HABs have significant potential as EO platforms (Pankine et al., 2009) both by themselves and in co-operation with satellites to deliver very-large-area coverage and very high spatial and temporal resolution EOs (Floberghagen, 2024; Drinkwater, 2019).

There are two other important imaging parameters for EO imaging, namely, spectral resolution and spectral range. Therefore, HABs should also be able to carry sufficient payloads for spectral imaging, which is a hybrid modality that combines imaging and spectroscopy with spectral range and resolution beyond red–green–blue (RGB) wavelengths. Hyperspectral imaging (HSI) (Goetz et al., 1985) is a spectral imaging technique that combines imaging and spectroscopy by recording the light intensities in many narrow contiguous spectral bands at each spatial point ( x , y ). Since different features on the Earth’s surface, such as trees or water bodies, reflect (absorb, scatter, and emit) light differently at specific wavelengths, this creates a characteristic spectral fingerprint that can be identified with HSI.

The main acquisition technique in HSI is push-broom (spatial) scanning (Boldrini et al., 2012), in which the spectral information of a single spatial-dimension slice is projected onto the sensor. However, push-broom scanning requires that the velocities of the imaged objects relative to the camera be known with high precision. Uncertainties in the relative motions can result in significant motion artifacts and image distortions. The push-broom technique is therefore well-suited to large satellites with little motion relative to the orbit but less suited to HABs, where the gondola generally moves more relative to the flight trajectory.

Herein, we study a snapshot (non-scanning) HSI technique in which the entire scene is recorded in a single sensor integration time. This approach reduces sensitivity to motion artifacts from the HAB platform. In particular, the snapshot HSI system that we developed and tested is a computed tomography imaging spectrometer (CTIS) (Descour and Dereniak, 1995; Bulygin et al., 1992; Okamoto and Yamaguchi, 1991) developed by Newtec Engineering A/S. This imaging system was developed, tested, and operated as part of the HEIMDAL mission. The test flight was launched from Kiruna in October 2024 as part of the REXUS/BEXUS 34/35 2024 campaign organized by DLR-SNSA.

The prototype camera used in the system has a spectral range from 600 to 850 nm and captures both spatial and spectral information simultaneously using a 2D diffraction grating, which disperses the light into a 3 × 3 diffraction pattern with a central zeroth-order and eight surrounding first-order components. From this diffraction image, a tomographic reconstruction is required to obtain the hyperspectral image (Descour and Dereniak, 1995; Bulygin et al., 1992; Okamoto and Yamaguchi, 1991). The image processing used in the system is described elsewhere (Huang et al., 2022; Ahlebæk et al., 2023), while some of the first proximal sensing applications have been documented in earlier reports (Peters et al., 2022; Peters et al., 2025).

HSI-based remote sensing has many applications in monitoring the climate, the environment, and nature. For example, satellite- and drone-based HSI has been used extensively for land-cover and biomass classifications, such as tree-species classification (Mäyrä et al., 2021; Ballanti et al., 2016; Masaitis and Mozgeris, 2012), and for studying vegetation dynamics, including in the Arctic region (Nelson et al., 2022). As a simple proof-of-principle application of the acquired CTIS data, we analyzed the flight data to distinguish landmass from water bodies.

The outline of this paper is as follows. In Section 2, we present the CTIS HAB system and image reconstruction algorithms. In Section 3, we discuss the integration of the imaging system with the BEXUS HAB gondola, mission control software, and laboratory tests. In Section 4, we discuss the actual BEXUS HAB flight launched from Kiruna; in Section 5, we present an analysis of the acquired data. Section 6 provides the conclusion.

2 HEIMDAL HAB imaging system

The technical objective of the mission presented here was to successfully acquire images with the CTIS system under emulated stratospheric conditions and during flight.

Despite the potential of HABs to provide high-resolution EO imaging—both spatially and spectrally—there is scant literature on HSI systems for HABs (Delauré et al., 2017; Navarro et al., 2022). To the best of our knowledge, this study is the first effort at demonstrating a CTIS system on a HAB and one of the first to demonstrate snapshot HSI from a HAB.

2.1 Imaging system

The complete imaging system integrated into the HAB gondola is shown in Figure 1.

The four modules are (from right to left) the CTIS camera (long lens with red cap), a companion RGB camera, the on-board computer (OBC), and the electronics box.

The camera system (OBC, RGB camera, and CTIS camera) was co-developed by Newtec Engineering A/S and Qtechnology A/S and is based on their QT-5222 platform. This is a dual-head camera system, meaning that both cameras are controlled and triggered and that all acquired images are handled by the OBC. It is an upgraded version of the single-head system described in Peters et al. (2022); Peters et al. (2025), where the OBC and CTIS camera were combined into a single module. Despite the change in system architecture, the CTIS optics and specifications remain the same: it houses a monochrome 4 MP GSENSE2020 CMOS sensor, a custom diffractive optical element (DOE), two 35-mm Vis-NIR VS-technology lenses (VS-H3520-IRC), and an outer 50-mm Vis-NIR VS-technology lens (VS-H1620-IRC) that effectively determine the focal length of the camera. The spectral range of the camera was 600–850 nm, and this window was defined using a 600-nm long-pass filter (FELH0600) and an 850-nm short-pass filter (FESH0850) from Thorlabs that was situated at the front of the system. While the RGB camera is not central to the scope of this study, it is included in the dual-head setup and features a Sony IMX420 sensor with a 25-mm lens.

2.2 Spatial and spectral resolution datacubes

The system outputs 3D hyperspectral datacubes I ( x , y , λ ), where the intensity I is a function of the spatial coordinates ( x , y ) and wavelength λ, with spatial dimensions of 312 × 312 p i x e l s and either 145 or 236 spectral channels depending on whether the reconstruction algorithm used is based on convolutional neural network (CNN) or expectation maximization (EM), as presented in the following subsection.

At the expected float altitude of 26 km, the system has a ground sampling distance (GSD) of approximately 3.4 m per pixel.

The desired focal length and spatial resolution were motivated by the need to allow tree-type classifications in future flight campaigns (e.g., the crown spread of adult spruce specimens typically ranges from 7 to 9 m 2 (The Morton Arboretum, 2025)), although tree clustering is anticipated to increase the area over which the specimens vary. Similarly, the spectral resolution for tree-type classification would need to be ≤10 nm to ensure that the relevant absorption features are distinguishable (e.g., chlorophyll a and b peaks are observed at 662 nm and 642 nm, respectively (Parizek et al., 2019)).

2.3 CTIS image reconstruction

Some portions of this section have been adapted from Peters et al. (2025).

The CTIS captures the spatial and spectral information from the datacube simultaneously using a 2D diffraction grating, thereby dispersing the light into a 3 × 3 diffraction pattern comprising a central zeroth-order and eight surrounding first-order components, as shown in Figure 2.

For this diffraction image, tomographic reconstruction is required to obtain the hyperspectral datacube (Descour and Dereniak, 1995; Bulygin et al., 1992; Okamoto and Yamaguchi, 1991). The relation between the CTIS diffraction image and datacube is described by the linear imaging equation (Equation 1) that maps the datacube f in vectorized form onto the CTIS diffraction image g in vectorized form using the system matrix H (Descour and Dereniak, 1995):

g = H f + n , (1)

where n is the linear noise. In our case, f is a vector of 312 × 312 × ( 145 o r 236 ) voxels, and g is a vector of 1910 × 1910 pixels. The system matrix H incorporates the diffraction sensitivity (lens transmission, sensor response, and diffraction efficiency of the diffractive optical element), illumination, and vignetting. For non-trivial system dimensions, the CTIS system is an underdetermined linear equation without an exact solution. Therefore, iterative reconstruction algorithms such as EM (Descour and Dereniak, 1995; White et al., 2020) or CNNs (Huang et al., 2022; Ahlebæk et al., 2023) are used to approximate the inverse of H and reconstruct the datacubes from the CTIS images. Equation 2 presents our description and implementation of the EM algorithm followed that reported by Huang et al. (2022) and consisted of an expectation step, where we computed g ̂ = H f ̂ ( k ), and a maximization step, where we updated f ̂ ( k ):Here, k is the iteration index, f ̂ ( k ) is the current estimate, ∑ i = 1 q 2 H i j is the sum of the rows in H, H T is the transpose of H, and ⊙ denotes element-wise multiplication. This equation combines the expectation and maximization steps; we initialized the equation with f ̂ ( 0 ) = H T g and performed 20 iterations, as typically 10–30 iterations are needed (White et al., 2020). Both H construction and f reconstruction were implemented in MATLAB using sparse matrix manipulations. As detailed in Section 2, the optical system of the CTIS cameras was the same as that outlined in Ahlebæk et al. (2023), with the sole exception of the outermost lens, which also applies to the construction of the matrix H. A series of experiments was carried out to precisely replicate the inclusion of spatial and spectral correction terms (such as point-spread function, illumination, and quantum efficiency), as elaborated in the supplementary material of Ahlebæk et al. (2023).

f ̂ k + 1 = f ̂ k ∑ i = 1 q 2 H i j ⊙ H T g H f ̂ k . (2)

CNNs have been used previously for direct classification of CTIS images (Douarre et al., 2021) and reconstruction of datacubes from CTIS images (Huang et al., 2022; Ahlebæk et al., 2023; Mel et al., 2022; Wu and Cai, 2023; Zimmermann et al., 2022). As an extension, an auto-decoder network architecture was proposed in Peters et al. (2025). The CNN reconstruction model used in the present work was adopted directly from Peters et al. (2025) without modification. A detailed description of the network architecture, training data, and training procedure is provided in the original publication. This physics-guided convolutional auto-decoder consists of the UNet presented in Huang et al. (2022) as the decoder, followed by an encoder that uses physical knowledge of the imaging system encoded in the system matrix H. It utilized the linear imaging equation (Equations 2, 3) to obtain an estimated CTIS image g ̂ = H ̂ f (since the exact system matrix H is unknown), and an additional UNet structure was employed to refine this estimated image.

3 Imaging system integration

In this section, we detail the mechanical and thermal requirements and designs for the imaging system to function on the BEXUS HAB platform under stratospheric conditions.

The trajectory of the HAB flight is shown in Figure 3 and passes over both landmass and water bodies. The flight was planned around solar noon to ensure uniform lighting conditions for the test flight. During flight, the balloon reached an altitude of 28 km, as shown in Figure 4. At this height, temperatures of < −50 C° and pressures as low as 0.05 atm are reached, during which time the imaging system must remain operational.

3.1 Mechanical

The robustness and stability of the mechanical gondola and payload structure are crucial for HAB flights. A structural failure could result in experiment failure or even loss of equipment mid-flight. Although most of the flight is expected to be smooth, three critical events can subject the structure to forces of up to 20g: balloon cutdown, parachute engagement, and gondola touchdown.

An overview of the BEXUS gondola carrying the imaging system for the HEIMDAL mission is shown in Figure 5. The four green modules are (from left to right) the CTIS camera, RGB camera, OBC, and electronics box.

The gondola was built using 45 × 45 mm Rexroth aluminum profiles and assembled with various angle gussets at the corner interfaces for an effective volume of 2 × 0.6 × 0.6 (L × W × H) m 3.

The mechanical structure for the imaging system was designed using six Rexroth 20× 20 mm profiles (Figure 6): four vertical 470-mm legs to elevate the four experiment modules from the gondola frame and two horizontal 600-mm legs connecting the four vertical legs. These six profiles were secured with corner gussets and additional custom braces that connected the leg frames to enhance stability.

The custom braces were initially designed and manufactured from 2-mm aluminum bent sheet metal; however, they were deemed too soft following a shock test, so a new set was manufactured using steel.

Rubber buffers were used for shock absorption and electrical insulation when mounting the experiment modules on the gondola.

The four modules were fastened to the rails with T-nuts through the rubber buffers and onto their respective component brackets. Additionally, Loctite was used to ensure that none of the components came loose during the launch. The electronics box was wrapped in multilayer insulation (MLI) to protect the solid-state devices (SSDs) housed within, as deemed necessary by the first thermal vacuum chamber test in the Mars Laboratory at Aarhus University. The final integrated structure is as shown in Figure 1, which also shows that the frame beneath the electronics box is wrapped in brown insulation tape. This precaution was necessary because the MLI is electrically conductive, and direct contact with the frame could have caused a short circuit.

The final total payload weight was 7.5 kg, with dimensions of 0.60 × 0.25 × 0.47 m 3.

3.2 Software, camera control, and image storage

The flight software for the mission consisted of two separate systems: the OBC and the software for the ground-station control room. Furthermore, the OBC software was split into two parts: image storage and reconstruction, and camera control. The communication between the OBC and ground station was enabled by an Ethernet interface called E-link that was provided by the launch provider Swedish Space Corporation (SSC).

Images were captured and stored every second on three separate SSD drives during flight, where each image was assigned an incremental suffix, such as image 12255, as shown in Figure 2. To validate on-board image reconstruction while limiting the power consumption, we reconstructed the datacubes from the CTIS images every 10th image, yielding approximately 2,000 cubes. A fraction of the images was also downlinked during flight through E-link at a rate of one image per 90 s.

A graphical user interface (GUI) was developed to monitor and control the imaging system during flight. A snapshot of the GUI is shown in Figure 7. The GUI allowed the operations team to switch between manual and automatic exposure, control the bandwidth at which the images were transmitted to the ground station, and monitor the image quality of the downlinked images visually.

3.3 Power system

The power consumption of the camera system, including the OBC, was estimated to be 180 Wh, with 1 h in idle mode prior to launch and 5 h of active time during flight. The system was powered using two parallel-connected battery packs, each with 8x SAFT LSH20 (3.6 V) batteries configured in series, delivering a nominal 28.8 V and a capacity of 13000 mAh corresponding to 750 Wh at 20 °C. The expected nominal temperature inside the battery compartment during flight is − 20 °C, which reduces the effective capacity of one battery pack to roughly 173 Wh. Hence, we used two battery packs to ensure sufficient power for the entire flight.

Moreover, the battery pack delivered an unstable 28.8 V, while the OBC demands a stable voltage in the range 20–28.8 V, with tests done at 24 V. Voltage regulation was achieved with a power distribution unit (PDU) consisting of a TDK-Lambda i7C buck-boost converter, upon which the PDU was designed in KiCad and produced by ICAPE Denmark A/S to lower the operating voltage to 24 V.

The complete electronic design scheme is shown in Figure 8.

3.4 Pre-flight tests

The most significant concern regarding the system performance was the thermal design surrounding the OBC and on-board computing units of the cameras, particularly whether the experiment would overheat in the low-pressure environment of the stratosphere due to limited convective cooling. The high power consumption of the prototype camera combined with its low heat dissipation raised concerns about the potential thermal bottlenecks.

To address these concerns, the experiment was tested in the Mars Simulation Laboratory at Aarhus University. This state-of-the-art vacuum chamber was designed to replicate Martian environmental conditions and is equipped with a cooling plate that circulates liquid nitrogen to achieve extremely low temperatures. Figure 9(a) shows the exterior of the Mars Laboratory vacuum chamber, while Figure 9(b) shows the experiment as the thermal test was in progress.

The test protocol involved subjecting the experiment to a 5-mbar pressure environment while varying the environmental temperature between −20 °C and −80 °C for over 5 h. These test conditions were designed to simulate conservative worst-case scenarios for pressure, temperature, and duration. The first test revealed that the SSDs would not withstand the harsh environment, prompting us to wrap the electronics box in MLI. This insulation effectively reduces the radiated heat and helps maintain the operating temperature of the electronics for longer. A follow-up test was performed after the insulation was added, which confirmed that the experiment would remain fully operational throughout the thermal vacuum test.

4 Flight and operations

The system was powered up several hours prior to launch to facilitate communications and systems checks, resulting in the capture of approximately 9,500 images while the gondola remained on the ground. During this phase, external power was supplied via cable to conserve battery resources. Consequently, the images presented in the results are numbered from 10,000 to 12,000, corresponding to the initial 500–2,500 s of flight. Note that the image numbering effectively elucidates the timeline of the balloon flight.

The balloon was launched at 07:52 on October 2 from the Balloon Platform at Esrange. The predicted (brown) and actual (blue) flight paths are shown in Figure 3.

Initially, the paths were aligned closely but began diverging after approximately 1.5 h. The flight, which commenced from Esrange in Northern Sweden, concluded in a forested area of Finnish Lapland. Although the planned duration was approximately 3 h, the balloon’s deviation from the predicted trajectory caused it to enter a “no-landing” zone, resulting in a total flight time of approximately 5.5 h.

The altitude profile for the flight is shown in Figure 4 and presents a nearly linear ascent of roughly 4.8 m/s, as expected; however, the balloon reached a higher maximum height of approximately 28 km than expected. This could be why the predicted trajectory deviated from the actual one.

5 Results

Both cameras captured images during the flight, and all images were safely stored on the SSDs that survived the gondola’s descent and landing. However, several issues were noted during the flight.

It is noted that all images presented in the Results section and Supplementary Appendix A were acquired during the ascent phase of the balloon flight. Consequently, the GSD varies slightly between images owing to altitude differences. From the images shown, the highest spatial resolution was achieved for image 11367 as an estimated GSD of 1.5 m/pixel, while the lowest resolution was obtained for image 12277 at a GSD of 2.1 m/pixel. These GSD values were theoretically derived from the sensor’s pixel size, focal length of the optical system, and corresponding acquisition altitude. The image suffix indices follow the second-by-second acquisition sequence.

5.1 RGB camera

The auto-exposure function did not operate correctly for the RGB camera, resulting in consistently underexposed images, as shown in the example RGB image in Figure 10. Although the camera captured images at 1 frame per second (FPS), the 500 kbps transmission limit caused an image delay of approximately 90 s at the ground station, significantly hindering real-time assessment and manual exposure adjustments during flight. A second issue during the flight test was that some of the RGB images were completely black, likely owing to trigger-signal synchronization between the RGB and CTIS cameras. The cameras were configured in a master-slave setup, and a possible misalignment in their frame rates may have resulted in missed or improperly captured exposures. Lastly, toward the latter part of the flight, condensation was visible in the captured images, which caused a hazy appearance and introduced a distinct artifact in the top-right corners of the frames.

It should be noted that these reported issues have limited impact on the analyses presented in the following sections. The RGB imagery was used to primarily provide visual context and spatial awareness for the hyperspectral measurements and was not used in any quantitative analyses. Furthermore, the synchronization inconsistencies resulted in the loss of only a small number of RGB frames, while sufficient overlap remained between consecutive acquisitions. Condensation was observed primarily during the later stages of the flight, whereas the data used in the subsequent analyses were acquired earlier, before condensation became a significant issue. Consequently, these operational challenges do not affect the presented results.

5.2 CTIS camera

The CTIS camera operated smoothly throughout the flight without major technical issues. However, as with the RGB camera, condensation was apparent during the later stages of the flight, significantly impacting image quality. The CTIS data used for the analyses presented in the following sections were acquired before the condensation became significant and are therefore not materially affected by this issue. An example CTIS image from the early stage of the flight is shown in Figure 2.

Both the RGB and CTIS cameras captured images at 1 FPS with exposure times between 4 and 10 ms. These short exposure times effectively mitigate any potential motion blur caused by gondola rotation or swinging.

In Figure 11, the zeroth-order components of CTIS images 11770–11800 are overlaid on a corresponding RGB image captured by the wider field of view (FOV) RGB camera. This overlay highlights the motion between each frame caused by the gondola’s rotation and pendular oscillations during flight, demonstrating that the snapshot HSI technique is the most suitable method for hyperspectral image acquisition under these conditions. This overlaying was done using an autocorrelation algorithm through a feature-based registration approach called Oriented FAST and Rotated BRIEF (ORB) (Rublee et al., 2011) keypoints. The ORB algorithm consists of four primary steps: feature detection and matching (BruteForceMatcher), homography estimation, transformation parameter extraction, and image warping.

5.3 CTIS system performance

Land–water classification is a well-established remote sensing task with applications in flood mapping, water resource management, shoreline monitoring, and ecological habitat assessment (McFEETERS, 1996; Xu, 2006). Owing to the strong spectral contrast between water and terrestrial surfaces, HSI is particularly well-suited for this application (Govender et al., 2009; Mancini et al., 2016). Although the classification problem is relatively simple, reliable separation of water and land still depends on preserving the characteristic spectral signatures of both classes. Therefore, a proof-of-principle land–water classification study was conducted to assess whether the reconstructed datacubes retained sufficient spectral information for downstream analyses. To evaluate the quality of the reconstructed datacubes, we performed a simple downstream analysis. As outlined in Section 2, the raw CTIS images are 2D diffraction images that require reconstruction to obtain the hyperspectral datacubes. For this purpose, we employed both CNN-based and EM-based reconstruction algorithms. The CNN and EM methods were both used exclusively for datacube reconstruction and not for subsequent classification analysis. To evaluate the reconstructed datacubes, a proof-of-principle classification study was performed using the partial least-squares discriminant analysis (PLS-DA) method. For this analysis, 23 datacubes containing both water and land regions were manually selected and masked. One of these datacubes and the corresponding masks are shown in Figure 12.

The spectra from these regions were averaged over 5× 5 neighboring pixels to reduce noise, and standard normal variate (SNV) normalization was applied for standardization. The resulting (training) dataset consisted of 213,206 spectra across the 23 images, with 72% representing tree-covered areas and 28% representing water, reflecting the prevalence of land regions in the observed areas and indicating the expected imbalance. Using the normalized spectra, PLS-DA (Brereton and Lloyd, 2014) was performed to evaluate the ability of the reconstructed datacubes to discriminate between land and water classes. PLS-DA is a supervised multivariate classification technique that projects high-dimensional predictor data onto a reduced set of latent components while maximizing the covariance between the predictor variables and associated class labels. This dimensionality reduction enables the method to capture the spectral variation most relevant for class separation and makes it particularly suitable for highly collinear hyperspectral data.

The number of latent components is an important model parameter in PLS-DA as it controls the complexity of the resulting model. Increasing the number of components generally improves the fit to the training data by incorporating additional variance and class-related information. However, excessive numbers of components may introduce overfitting and reduce the model’s ability to generalize to unseen data. Consequently, the optimal number of components is typically determined via cross-validation by balancing prediction accuracy and model robustness.

We trained separate PLS-DA models using the datacubes reconstructed with the two image reconstruction methods, and the corresponding cross-validation accuracy scores are presented in Figure 13. The CNN-based reconstruction achieved a cross-validation accuracy of approximately 0.80 after convergence at approximately 25 components, whereas the EM-based reconstruction converged to approximately 0.58 with 15 components. These results indicate improved class discrimination performance for the CNN-based reconstruction approach than the EM-based approach.

In addition to the accuracy scores, the performances of the two models were evaluated for both the training and test data using the mean-squared error (MSE). Figure 14 presents the MSE curves for the CNN- and EM-based reconstructions as functions of the numbers of latent components included in the PLS-DA model. The plots show the training and test MSE results for both the predictor variables (X-space) and response variables (Y-space), thereby illustrating how model performance changes with increasing model complexity. The lower test MSE values indicate improved predictive performance and better generalization to unseen data. Consistent with the cross-validation accuracy results, the CNN-based reconstruction outperformed the EM-based reconstruction.

Beyond the classification accuracy and MSE, the variable importance in projection (VIP) score was used to evaluate the relative contribution of each spectral band to the PLS-DA model. The VIP score summarizes the importance of each predictor variable by considering both the variance explained by each latent component and the variable’s contribution to the model projection. Equation 3 for a predictor variable j, the VIP score is commonly defined as

V I P j = p ⋅ ∑ a = 1 A S S Y a w j a 2 ‖ w a ‖ 2 ∑ a = 1 A S S Y a , (3)

where p is the total number of predictor variables, A is the number of latent components, S S Y a denotes the amount of response variance explained by component a, and w j a represents the weight of variable j for component a. Variables with VIP scores greater than 1 are generally considered influential for the model prediction (Galindo-Prieto et al., 2014).

Figure 15 presents the VIP scores for the CNN- and EM-based reconstructions, illustrating the differences in the importance of individual spectral bands between the two approaches. The VIP analysis provides insight into the wavelength regions that contribute most strongly to discriminating between land and water classes, thereby complementing the performance metrics with an assessment of model interpretability and relevance of spectral features. The VIP scores highlight consistently significant spectral features at approximately 650 nm for both models while showing some variability in importance across the second half of the spectrum. However, this observation by itself is insufficient to conclusively determine a preference for one reconstruction method over the other.

Figure 16a presents the CTIS image 12255 reconstruction using the CNN model, where all spectral bands are averaged to form a 2D image; this image is not part of the training dataset and was selected as an example containing both land and water regions. Figure 16c shows the corresponding EM reconstruction of the same CTIS image, with spectral averaging applied in the same manner. Figures 16b,d present the PLS-DA results based on the CNN and EM reconstructions, respectively. Upon inspection, both models appear to successfully identify water and land and generally map these regions correctly. However, the EM reconstruction exhibits noticeable artifacts, including distinct lines within the datacube. Additionally, upon closer inspection, it is evident that neither model accurately captures the precise waterline; most misclassified pixels are located near the shoreline boundaries. This behavior is consistent with the visual appearance of the scene and may be attributed to mixed pixels and shallow-water regions, where the measured signal contains contributions from both water and underlying terrestrial features. Consequently, the spectral distinction between the two classes is lower in these transition zones, making accurate delineation of the waterline more challenging for the PLS-DA model.

Additional figures, including Figure 16, are provided in Supplementary Appendix A for different CTIS images used as inputs, offering further analysis and validation.

Figure 17 presents a CTIS image series, where Figure 17a shows the autocorrelation algorithm applied to the series and overlaid on a corresponding RGB image of the same region. Figure 17b,c show the results of the PLS-DA models for the same image series based on the CNN and EM reconstruction methods, respectively. For pixels that are covered by multiple CTIS images, the categories shown in Figure 17b,c correspond to the most frequent types present across the overlapping images.

6 Conclusion

In this work, we present a first HAB test flight with a (prototype) snapshot hyperspectral camera based on the CTIS with a spectral range from 600 to 850 nm.

The battery-based power source for the entire imaging system delivered sufficient power throughout the duration of the >5-h flight comprising 1.5 h of ascent, 3.5 h of float time, and descent.

Data were successfully acquired and stored, and 3D hyperspectral datacubes were reconstructed during the flight. The entire system was retrieved without data loss, and the cameras also survived the flight, sustaining only minor damage to the outermost filter of the CTIS optics.

Based on the acquired data, a proof-of-principle land-cover classification was conducted to distinguish landmass from water bodies along the flight track using the reconstructed spectra.

Several improvements are planned for the imaging system in the future, including enhancements to the optics, reconstruction algorithms, and data link, to enable real-time data transmission and deliver high-quality video-rate HSI.

The present study represents a promising initial step in this direction and constitutes the first application of this type of CTIS camera outside a laboratory environment. To the best of the authors’ knowledge, this work also constitutes the first successful integration and deployment of a CTIS camera on a HAB for remote sensing applications.

Generative AI statement

The author(s) declared that generative AI was used in the creation of this manuscript. During the preparation of this work, the authors used ChatGPT to improve the language and readability. After using this tool/service, the authors reviewed and edited the content as needed and take full responsibility for the content of the publication.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/frsen.2026.1843759/full#supplementary-material

Supplementary material Supplementary material at the source

Footnotes

Edited by: Minghua Wang, Nankai University, China

Reviewed by: Gonzalo Cucho-Padin, National Aeronautics and Space Administration, United States

Fatih Adıgüzel, Bitlis Eren University, Türkiye