From Optical Design to NIIRS and Object Detection: An Integrated Framework for Spatial Image Quality Assessment of Micro-Satellite Constellations

Highlights

What are the main findings?

• An integrated framework was established to connect optical design, NIIRS, and object detection using NEONSAT-1 imagery.
• In the NEONSAT-1 experiments, system MTF and simulated GSD changes were reflected in NIIRS, while detection response varied by scene, product type, and target class.

What are the implications of the main findings?

• NIIRS is useful for estimating detection potential, but it should not be used as a single universal indicator for all target classes.
• The proposed framework provides a baseline for future image quality assessment across a micro-satellite constellation.

Abstract

For micro-satellite constellations, frequent Earth observation alone does not guarantee archive usability; the archive is operationally useful only when the spatial image quality remains adequate for downstream exploitation. This study presents an integrated framework for assessing spatial image quality using NEONSAT-1 imagery by linking optical design analysis, image simulation, GIQE-based NIIRS estimation, and YOLOv8-based object detection within a single workflow. NEONSAT-1 panchromatic (PAN), pan-sharpened (PS), and multispectral (MS) products were analyzed together with controlled simulations of system MTF, altitude-dependent GSD variation, and super-resolution processing. Among the native products, PS imagery showed the highest NIIRS and overall detection performance. In the controlled experiments, higher system MTF increased RER and NIIRS, while lower simulated altitude generally produced finer GSD and higher NIIRS for both PS and PAN products. However, detection performance varied by scene, product type, and target class and did not increase in direct proportion to NIIRS. In the super-resolution case study, ×2 SR provided the most consistent NIIRS improvement, whereas detection responses at higher SR scales were target class dependent. These results suggest that spatial image quality should be evaluated not only through interpretability metrics such as NIIRS but also in relation to practical downstream performance. The proposed framework provides a baseline for future constellation-scale image quality assessment.

1. Introduction

In the New Space era, Earth-observation missions have increasingly shifted from large, single-satellite programs to constellations of micro- and small satellites. Advances in miniaturization and lightweight platform design have substantially improved the operational utility of these systems, enabling high-resolution imaging with short revisit times [1,2,3]. Over the past decade, annual spacecraft launches have increased dramatically, and small satellites (≤1200 kg) now dominate global deployments [4,5]. In 2024 alone, 2,790 small satellites were launched, accounting for 97% of all spacecraft launched that year [5].

The vast volume of optical imagery obtained by these constellations is becoming increasingly valuable when integrated with artificial intelligence (AI). This combination supports a wide range of applications, including multi-sensor data fusion, object detection, and defense and disaster monitoring, for which both spatial detail and temporal revisit are essential [6,7,8]. Consequently, ensuring the consistency, reliability, and operational usability of constellation-scale image archives has become a critical requirement for scientific, policy, and industrial applications.

While national agencies such as ESA and NASA continue to play foundational roles in Earth observation, the New Space era is also characterized by the rapid expansion of commercial satellite operators [9]. In Korea, the NEONSAT program, led by SaTReC at KAIST, was initiated as the country’s first micro-satellite optical constellation mission dedicated to monitoring the Korean Peninsula, with the goal of enabling rapid and accurate responses to national security events and natural disasters [10,11]. As such constellation systems expand, the value of the image archive depends not only on revisit frequency but also on whether the imagery remains consistent and usable for downstream analysis.

Remote sensing datasets are commonly characterized by spatial, spectral, radiometric, and temporal resolutions, while geometric and geolocation accuracies are also critical quality attributes for registration and geodetic fidelity [12,13]. For constellation applications, however, these quality attributes can be viewed as two complementary branches. One branch concerns radiometric, spectral, and geometric consistencies across sensors, which are essential for interoperability, reflectance harmonization, and analysis-ready time-series products [14,15,16]. The other concerns spatial image quality, which is more directly linked to image interpretability and downstream exploitation. In this study, the term image quality is used primarily in the latter sense, namely, the spatial image quality of optical satellite imagery.

Considerable progress has already been made in the former category. At the product level, inter-sensor harmonization efforts, such as NASA’s Harmonized Landsat and Sentinel-2 (HLS) dataset, provide standardized surface-reflectance products through atmospheric correction, co-registration and common gridding, and spectral bandpass adjustment [17,18]. More generally, BRDF normalization has been widely adopted to reduce the apparent reflectance variability caused by illumination-view geometry and surface anisotropy [19,20]. For commercial systems, hierarchical and relative radiometric normalization approaches have also been proposed to improve temporal and inter-sensor consistencies across CubeSat and multi-sensor image archives [21,22,23]. However, these studies primarily address reflectance consistency and interoperability rather than the uniformity of spatial image quality across a constellation.

Within this spatial image quality branch, the National Imagery Interpretability Rating Scale (NIIRS) has been widely used as a practical indicator of image interpretability, and the General Image Quality Equation (GIQE) relates NIIRS to image scale, sharpness, and signal-to-noise characteristics [24,25,26]. Recent studies have examined the applicability of different GIQE formulations and NIIRS-based assessment methods to optical Earth-observation imagery [27,28,29,30,31,32]. Other image quality assessment approaches, including perceptual, structural, and no reference methods, can also provide useful information on visual image degradation. However, the objective of this study is not general perceptual image quality assessment but spatial interpretability assessment linked to measurable imaging factors and downstream object detection. Therefore, GIQE and NIIRS were adopted as the primary framework in this study, while complementary image quality assessment approaches are discussed as future extensions in Section 5.

Nevertheless, most existing efforts remain limited to single sensors, individual products, or post-processing quality assessment and do not provide a continuous framework that connects system design, image formation, image interpretability, and downstream object detection performance at the constellation level.

This limitation is particularly important for micro-satellite constellations. Even when satellites are nominally identical, residual variability can arise from manufacturing tolerances, optical alignment differences, component aging, orbital position, attitude variations, and scene-dependent acquisition conditions [33,34,35]. From the perspective of spatial image quality, such variability mainly affects GSD, edge-based image sharpness represented by the relative edge response (RER), and SNR. These changes can, in turn, lead to differences in image interpretability and downstream usability across the constellation. Therefore, increasing revisit frequency alone does not guarantee uniform usable image quality.

To address this gap, this study presents an integrated framework for spatial image quality assessment using NEONSAT-1 imagery. The framework links mission and orbit conditions, optical analysis, image simulation, GIQE/NIIRS-based assessment, and AI-based object detection within a single workflow. Within this scope, spatial image quality is considered primarily in terms of image interpretability rather than nominal spatial resolution alone. Using NEONSAT-1 imagery and representative simulations, the analysis examines how variations in GSD, RER-derived edge sharpness, SNR, and super-resolution processing are reflected in NIIRS and object detection performance. Although the present analysis is conducted using a reference NEONSAT-1 case, the framework is intended to support future extensions toward constellation-level quality assessment.

This study makes three main contributions. First, it establishes a workflow that connects image formation, NIIRS-based interpretability assessment, and downstream object detection for optical micro-satellite imagery. Second, it quantitatively evaluates how GSD, system MTF-related sharpness, and super-resolution affect NIIRS and YOLOv8-based detection performance in NEONSAT-1 imagery. Third, it shows that improved image interpretability does not necessarily produce proportional gains in detection performance, highlighting the need to evaluate spatial image quality in relation to practical downstream performance rather than through a single image-quality metric alone.

The remainder of this paper is organized as follows. Section 2 describes the study data and materials. Section 3 first presents the overall framework and then describes the methodology, including optical design and MTF analysis, image simulation, GIQE/NIIRS-based assessment, object detection, and super-resolution processing. Section 4 presents the experimental results for each module. Section 5 discusses the main implications of the results, together with the scope, limitations, and future extensions of the present study. Finally, Section 6 concludes the paper.

2. Materials and Methods

2.1. NEONSAT Optical System Specifications

NEONSAT (New-Space Earth Observation Satellite Constellation for National Safety) is a pioneering micro-satellite constellation system under development in South Korea. The program aims to deploy eleven micro-satellites, together with an imaging data utilization system, to improve the efficiency and accuracy of national security and disaster response over the Korean Peninsula [36]. The first prototype satellite was successfully launched in April 2024, and the remaining ten satellites are scheduled for launch in two batches, with five planned in 2026 and five in 2027.

The NEONSAT platform was designed to enable cost-effective and rapid constellation deployment with a satellite mass of less than 100 kg. The satellites are planned to operate in two sun-synchronous orbital planes at an altitude of about 500 km (LEO). The optical payload provides panchromatic (PAN) imagery with a spatial resolution of 1 m and multispectral (MS; RGB) imagery with a spatial resolution of 4 m, while satisfying system requirements in terms of modulation transfer function (MTF), signal-to-noise ratio (SNR), and geolocation accuracy [36].

Because this study focuses on spatial image quality, the optical payload specifications of NEONSAT were used as the reference system conditions for the subsequent optical analysis, image simulation, and image quality assessment. The flight model of NEONSAT is shown in Figure 1, and the main system specifications are summarized in

Table 1. NEONSAT system specifications.

Parameters	Value
Orbit	Sun-synchronous
Altitude (km)	500
Weight (kg)	<100
Resolution (m)	PAN: ≤1; MS: ≤4
MTF (%)	PAN: ≥7; MS: ≥20
SNR	PAN and MS: ≥90
Swath (km)	≥9.76 at nadir

.

2.2. NEONSAT Image Archive and Analysis Subsets

This study primarily used NEONSAT-1 imagery acquired from June to December 2024 during the launch and early operation phase (LEOP). The dataset consists of Level-1R PAN and MS products that had undergone radiometric correction. The imagery was collected over geographically diverse regions, including global calibration/validation sites and other areas of interest, and spans a wide range of land-cover and terrain conditions, such as dense forests, agricultural fields, urban areas, coastal zones, and mountainous terrain.

For the analyses presented in this paper, a total of 350 PAN-MS image pairs with associated metadata were selected from the archive. The Level-1R PAN and MS products served as the primary study data, while corresponding pan-sharpened (PS) products were also generated and included in the analysis. These data were used for GIQE coefficient estimation, optical image quality assessment, object-detection analysis, and super-resolution evaluation. For object-detection and super-resolution analyses, image chips were extracted from the selected scenes to improve computational efficiency, yielding approximately 1300 chips in total. To ensure consistency in the spatial image quality analysis, each experiment used a single product level rather than mixing products with different geometric processing levels.

From this archive, several analysis subsets were defined for different experimental purposes. A Boston scene centered on Boston Logan International Airport was used as the primary representative case for product-level comparison and joint NIIRS–detection analysis because it contains multiple object classes, including vehicles, airplanes, storage tanks, and ground-track fields. For controlled MTF sensitivity analysis, both an ideal digital edge target and a real edge target extracted from NEONSAT-1 imagery acquired over Salon-de-Provence, France, were used.

To examine altitude effects across different scene characteristics, three representative PS and PAN scenes were selected from Boston, Paris, and Abu Dhabi. The Paris scene covered Paris–Le Bourget Airport and the surrounding Stade de France area, whereas the Abu Dhabi scene covered Zayed International Airport and the surrounding Ferrari World area. For local object-detection analyses, representative image chips containing these classes were used. For the super-resolution case study, PS chips were used, and qualitative visual comparison and quantitative detection-rate evaluation were focused on four annotated target classes.

This subset design was intended to validate different components of the proposed framework under representative conditions, while keeping NEONSAT-1 imagery as the primary data source throughout the study.

2.3. External Benchmarks and Auxiliary Data

The DIOR (Diverse Object Detection in Optical Remote Sensing Images) dataset was used to train and fine-tune the YOLOv8n object detector [37] for methodological purposes. DIOR is a widely used benchmark in the remote-sensing community and provides large-scale annotated optical imagery acquired by multiple sensors over a broad range of spatial resolutions. In this study, it was used only as an external benchmark for detector training prior to application to NEONSAT-1 imagery.

Second, the US Merced Land Use dataset was used as a controlled benchmark for super-resolution validation [38]. This dataset contains high-resolution aerial imagery covering multiple land-use categories and was used to assess the basic reconstruction performance of the compared super-resolution methods before they were applied to NEONSAT-1 data.

These auxiliary datasets were therefore used only for specific methodological purposes, namely, object-detector training and super-resolution validation, whereas the core image quality analysis and downstream task evaluations were conducted primarily on NEONSAT-1 imagery. The corresponding analysis subsets and auxiliary benchmark datasets are summarized in

Table 2. NEONSAT analysis subsets and auxiliary benchmark datasets used in this study.

Section	Data Subset	Date	Product/Target	Purpose
Section 4.1	Boston	1 June 2024	PS, PAN, MS	Product-level comparison of NIIRS and detection rate
Section 4.2.1	Ideal digital edge target	N/A	Synthetic edge pattern	Controlled sensitivity analysis of MTF, RER, and NIIRS
Section 4.2.1	Salon-de-Provence, France	9 July 2024	Real NEONSAT edge target	Validation of MTF sensitivity using an observed edge target
Section 4.2.1	Boston	1 June 2024	PS	Quantitative comparison of system MTF effects on NIIRS and detection rate
Section 4.2.2	Boston	1 June 2024	PS, PAN	Altitude-dependent simulation across representative scenes
	Paris	10 August 2024
	Abu Dhabi	14 December 2024
Section 4.3.2	Selected PS image chips	1 June 2024	PS + SR outputs	SR scaling analysis; detection-rate evaluation for four target classes *
Section 3.4	DIOR	N/A	Object-detection benchmark	YOLOv8 training and fine-tuning
Section 3.5 and Section 4.3.1	US Merced Land Use	N/A	Runway and intersection subset	Controlled super-resolution validation

* The object detector was configured for four annotated target classes: vehicle, airplane, storage tank, and ground-track field. For the super-resolution case study (Section 4.3.2), SR was visually examined and quantitatively evaluated on representative PS image chips containing these target classes. Boston refers to Boston Logan International Airport; Paris refers to Paris–Le Bourget Airport and the surrounding Stade de France area; Abu Dhabi refers to Zayed International Airport and the surrounding Ferrari World area.

.

3. Methods: Integrated Framework for Spatial Image Quality Assessment and Object Detection Performance

Figure 2 presents the integrated framework used in this study for spatial image quality assessment. The framework was designed to examine how mission and optical conditions are reflected in image formation, image interpretability, and object detection performance. The current scope is limited to validation of the framework components using NEONSAT-1 imagery and representative simulations rather than full constellation-level validation.

Within this framework, optical design analysis and image simulation are used to represent controlled changes in GSD, image sharpness, and SNR. These image characteristics are then evaluated through GIQE-based NIIRS assessment, and their practical implications are further examined through object-detection performance. In the subsequent image quality analysis, image sharpness is quantified by ESF-derived RER, because RER is the edge-based metric directly used in the adopted GIQE/NIIRS assessment. Super-resolution analysis is included as an additional module to test whether apparent improvements in image interpretability are accompanied by improvements in object-detection performance.

The remainder of this section describes the main components of the framework. Section 3.1 presents the optical design analysis and MTF characterization, Section 3.2 describes the image simulation procedure, Section 3.3 details the GIQE-based NIIRS assessment, Section 3.4 presents the object-detection model and evaluation procedure, and Section 3.5 describes the super-resolution analysis.

3.1. Optical Design Analysis and MTF Characterization

Because this study focuses on spatial image quality, the optical specifications of the NEONSAT payload described in Section 2 were used as the baseline optical configuration for the subsequent analysis. Zemax OpticStudio 2024 R2 was used to examine how variations in the baseline optical design affect spot size and the optical modulation transfer function, $MT{F}_{optics}$ [39]. Although imaging performance is influenced by multiple optical design variables, the conic constant $k$ was selected as the primary controlled variable in this study because it directly controls the aspheric departure of a reflective surface. It therefore provides a practical means of analyzing how surface-figure variations affect spot size and $MT{F}_{optics}$ while maintaining the baseline optical configuration [40,41].

Each reflective aspheric surface was modeled as a rotationally symmetric conic defined by the radius of curvature $R$ and the conic constant $k$, while the axial spacing $d$ between elements was fixed to the baseline configuration, unless otherwise stated. The surface sag $z\left(r\right)$ was described using the standard conic sag equation [40,42]:

$$z\left(r\right)={\displaystyle \frac{C{r}^2}{1+\sqrt{1-\left(1+k\right)C^2{r}^2}}}, C={\displaystyle \frac{1}{R}}, r=\sqrt{x^2+y^2}$$

(1)

Under the paraxial approximation, a variation in $k$ changes the sag departure relative to the corresponding reference sphere and thereby induces an optical path difference $W\left(r\right)$. This relationship can be approximated as follows:

$$∆z\left(r\right)\approx {\displaystyle \frac{∆k{r}^4}{8R^3}},W\left(r\right)\approx 2\mathsf{\Delta }z\left(r\right)\approx {\displaystyle \frac{∆k{r}^4}{4R^3}}$$

(2)

The resulting wavefront variation modifies the optical response of the imaging system. $MT{F}_{optics}$ describes the transfer of scene contrast as a function of spatial frequency [41,43]. In this study, $MT{F}_{optics}$ was obtained through Zemax ray tracing and FFT-based analysis and is expressed as the magnitude of the optical transfer function [39,44,45]:

$$MTF\left(f_x, f_y\right)=\left|OTF(f_x, f_y)\right|$$

(3)

Using this procedure, a set of optical cases was selected so that the MTF at Nyquist frequency spanned 5, 11, 15, 18, 20, and 25%. The baseline NEONSAT optical condition corresponded to approximately 18% at Nyquist, and lower and higher cases were selected around this reference for the subsequent image simulation experiments. Figure 3 shows the corresponding $MT{F}_{optics}$ curves over the spatial frequency for the selected optical cases, indicating that lower Nyquist MTF cases exhibit faster attenuation of contrast transfer over the full frequency range. These cases were used to quantify the sensitivity of spatial image quality to MTF-related sharpness. Figure 4 summarizes the optical design workflow used in this study. Based on the reference payload specifications, the conic constant was varied while the remaining baseline configuration was maintained. This procedure generated the set of optical MTF cases used in the subsequent image simulation.

3.2. Image Simulation and System MTF Formulation

To propagate the optical MTF variations derived in Section 3.1 to simulated imagery, the pyBSM (Python-based System Model) imaging chain concept was adopted [46,47], and the corresponding simulation procedure was implemented under the conditions considered here. The purpose of this module was not to reproduce absolute at-sensor radiance through a full radiative-transfer model but to generate relative changes in spatial image quality under varying acquisition conditions. In the present framework, the simulation module was used to examine how selected imaging parameters affect quantitative image-quality metrics and the downstream task response. In this study, altitude-dependent GSD and system MTF variation were treated as the primary variables, while atmospheric, noise, and detector-related conditions were fixed or simplified to maintain a controlled comparison.

The optical MTF, $MT{F}_{optics}$, was obtained from the Zemax optical design analysis described in Section 3.1. The pyBSM imaging chain was then used to incorporate additional spatial degradation components, including detector response, atmospheric effects, and motion-related terms, under the NEONSAT imaging configuration.

NEONSAT orbital conditions were used as the reference mission setting. The satellite altitude and velocity, detector pixel size, focal length, aperture diameter, TDI, and integration time were specified according to the NEONSAT sensor and imaging geometry. Altitude-dependent sampling variation was introduced through the ground sampling distance, GSD:

$$GSD\left(H\right)={\displaystyle \frac{H·p}{f}}$$

(4)

where $H$ is altitude, $p$ is pixel size, and $f$ is focal length. These values are referred to as simulated acquisition GSDs.

To represent the relative radiometric variation in the simulation, an empirical effective attenuation factor and path radiance term were introduced. This empirical formulation was used only to introduce a consistent relative radiometric perturbation across the simulated altitude cases:

$$\tau \left(H\right)=\exp \left(-{\displaystyle \frac{k_a}{H}}\right),L_{sim}\left(H\right)=\tau \left(H\right)·L_{scene}+\left[1-\tau \left(H\right)\right]·L_{path}$$

(5)

where $\tau \left(H\right)$ is an empirical effective attenuation factor, $k_a$ is an empirical coefficient with the same length unit as $H$, $L_{scene}$ denotes the reference scene radiance, and $L_{path}$ is a path radiance term [48]. The attenuation coefficients were initialized using Sentinel-2-based reference information and then applied consistently across the simulated cases.

A simplified altitude-dependent SNR scaling was used to assign a consistent relative SNR condition to each simulated altitude case:

$$SNR\left(H\right)=SN{R}_0\times {\left({\displaystyle \frac{H_0}{H}}\right)}^2$$

(6)

where $SN{R}_0$ denotes the reference SNR derived from NEONSAT calibration information at the reference altitude $H_0$. These assumptions were kept fixed across experiments so that the relative effects of optical sharpness and sampling conditions could be examined consistently.

Starting from $MT{F}_{optics}$ obtained in Section 3.1, the simulation combined optical, detector, and atmospheric contributions to construct the total system modulation transfer function, $MT{F}_{system}$ [49,50]:

$$MT{F}_{system}\left(f\right)=MT{F}_{optics}\left(f\right)·MT{F}_{detector}\left(f\right)·MT{F}_{atmosphere}\left(f\right)$$

(7)

where $MT{F}_{optics}$ was obtained from Zemax; while $MT{F}_{detector}$ and $MT{F}_{atmosphere}$ represent the detector response and atmospheric spatial degradation, respectively. In the pyBSM-based imaging chain, $MT{F}_{detector}$ represents the finite spatial integration effect of the detector element, not detector noise. $MT{F}_{atmosphere}$ was used to represent the atmospheric spatial degradation under the selected NEONSAT geometry and center wavelength condition. The resulting images therefore reflect $MT{F}_{system}$ rather than by $MT{F}_{optics}$ alone. This distinction is important for interpreting the experiments in Section 4.2.

Figure 5 summarizes the simulation workflow, in which the optical MTF cases were combined with GSD sampling, detector response, atmospheric spatial degradation, relative radiometric perturbation, and simplified relative SNR scaling to generate the simulated image cases.

3.3. GIQE-Based NIIRS Assessment

The National Imagery Interpretability Rating Scale (NIIRS) is a widely used indicator of image interpretability, and the General Image Quality Equation (GIQE) provides an empirical relationship between NIIRS and measurable image-quality factors, such as ground sampling distance (GSD), relative edge response (RER), and signal-to-noise ratio (SNR) [24,25,26,27]. In this study, GIQE v4 was adopted as the reference NIIRS model, and NIIRS was used as the primary indicator of spatial image quality for comparison with object-detection performance. This subsection describes how image-specific GSD, RER, and SNR were measured from NEONSAT-1 imagery and used for NIIRS assessment.

3.3.1. Edge-Based Measurement of RER and SNR from Natural Targets

To obtain image-specific RER and SNR, a natural target edge approach was adopted because it enables in-scene measurement without dedicated calibration targets [51,52]. Candidate pixels were first detected using the Canny edge detector, and straight edge features were then identified using a Hough transform [53,54]. These features were screened to retain only reliable candidates by applying constraints on length, contrast, and local homogeneity on both sides of the boundary.

For each accepted feature, an edge spread function (ESF) was constructed by accumulating intensity profiles sampled along the local normal direction and normalizing the resulting response. RER was then computed from the ESF as follows:

$$RER=ER\left(+0.5\right)-ER(-0.5)$$

(8)

where $ER(·)$ denotes the normalized edge response at the specified sub-pixel offset [55]. For each image, the representative RER was defined as the mean RER across all accepted edges.

For each edge, SNR was estimated from the contrast between the bright and dark sides and normalized by the corresponding local noise statistics [56]:

$$SNR={\displaystyle \frac{\mathsf{\Delta }DN}{({\sigma }_{bright}+{\sigma }_{dark})/2}}$$

(9)

For each image, the representative SNR was defined as the mean SNR across all accepted edges. This procedure provided the directly measurable image-quality factors used in the subsequent NIIRS assessment and surrogate fitting.

3.3.2. Reference GIQE v4 Calculation and Surrogate Fitting

Using the measured GSD, RER, and SNR values, NIIRS was first computed using the standard GIQE v4 formulation as follows:

$$NIIRS=c_0+c_1 {log}_{10}\left(GSD\right)+c_2{\mathit{log}}_{10}\left(RER\right)+c_3\left({\displaystyle \frac{G}{SNR}}\right)+c_{4}H$$

(10)

where GSD is the ground sampling distance, RER is the relative edge response, SNR is the signal-to-noise ratio, $G$ is the processing noise-gain term, and $H$ is the edge-overshoot term associated with MTF compensation [25,32]. In GIQE v4, the coefficients associated with GSD and RER are selected according to whether RER is greater than or less than 0.9 [25,32]. In the present study, GIQE v4 was adopted as the reference interpretability model throughout the analysis.

In addition to the reference GIQE v4 calculation, surrogate fitting coefficients were estimated separately for PAN, PS, and MS products. This step was introduced to obtain stable empirical approximations of the reference GIQE v4 response for routine comparison among NEONSAT products under the conditions used in this study. For each image in the NEONSAT-1 analysis archive, GSD, RER, and SNR were measured directly from the image, and the corresponding reference NIIRS value was first computed using GIQE v4. A surrogate model was then fitted separately for PAN, PS, and MS products using least squares regression:

$$NIIR{S}_{sur}=c_0+c_1\ast {\mathit{log}}_{10}\left(GSD\right)+c_2\left(RER\right)+c_3\left(SNR\right)$$

(11)

In this surrogate model, directly measurable image factors were retained explicitly, whereas the effects associated with the processing gain and overshoot terms were absorbed into the fitted coefficients for each product type. The resulting coefficients should therefore be interpreted as a surrogate approximation of the reference GIQE v4 response under the NEONSAT conditions used in this study, not as a replacement for the standard GIQE formulation. A total of 350 NEONSAT-1 image pairs described in Section 2 were used for coefficient estimation. The workflow is summarized in Figure 6, and the fitted coefficients are listed in

Table 3. Surrogate fitting coefficients estimated separately for PAN, PS, and MS products from NEONSAT-1 imagery.

Surrogate Fitting Coefficients
	$c_0$	$c_1$	$c_2$	$c_3$
PAN	7.17	−1.29	3.11	0.06
PS	7.68	−1.43	3.23	0.004
MS	6.57	−1.30	3.92	0.01

.

3.4. Object Detection Performance Evaluation Using YOLOv8

Object detection was included in the present framework to examine whether changes in spatial image quality are reflected in downstream task performance. In this study, object detection performance was used as the final evaluation variable for comparison with NIIRS-based image interpretability. YOLOv8n was used as the object detector because it provides a stable and computationally efficient implementation with fast inference under the available processing environment. The detector configuration was fixed throughout the experiments so that changes in detection performance could be compared with changes in image quality.

YOLOv8n was first trained using the DIOR benchmark dataset described in Section 2 and then applied to NEONSAT-1 imagery [7,37,57]. The use of an external benchmark dataset and a fixed detector configuration was intended to reduce dependence on NEONSAT-1 imagery alone. In the NEONSAT-1 analyses, four annotated classes were considered: vehicle, airplane, storage tank, and ground-track field.

Training and inference were conducted in a Python 3.10/PyTorch 2.2.1 environment with an input image size of 800 × 800 pixels, for 50 epochs, using a batch size of 16 and an initial learning rate of 0.01. During inference, the same trained weights and inference settings were applied to all product, MTF, altitude, and SR cases. This fixed-detector setting was adopted to ensure that changes in detection performance could be attributed primarily to changes in image quality rather than to changes in the detector configuration.

To improve the detection of small objects in high-resolution scenes, SAHI (Slicing Aided Hyper Inference) was used to perform inference on overlapping image slices and merge the resulting predictions [58]. Test-time augmentation (TTA) was additionally applied to improve inference robustness by aggregating predictions from augmented views of the same image [59].

The output of this module consists of detected object classes, bounding boxes, and confidence scores for each NEONSAT-1 image chip. For the analyses in Section 4, the detection rate was defined for the evaluated targets as the ratio of valid detections to the corresponding number of ground-truth targets manually counted in each scene or chip. A detection was regarded as valid when the predicted bounding box corresponded to a ground-truth object and the confidence score exceeded the predefined threshold. The resulting detection rate should therefore be interpreted as a study-specific comparison metric rather than as a standard benchmark metric such as mAP.

Figure 7 summarizes the object detection workflow and presents representative detection examples from NEONSAT-1 image chips used in the subsequent comparison with NIIRS.

3.5. Super-Resolution–Aided Image Quality Enhancement

Super-resolution (SR) was included in the present framework as an optional post-processing module to examine whether apparent gains in spatial detail are accompanied by improvements in image interpretability and object detection performance. In optical satellite imagery, SR aims to reconstruct a higher-resolution image from a lower-resolution observation by compensating for information loss associated with sensor optics, sampling, and noise [60,61,62,63,64].

In this study, three representative SR approaches were considered: bicubic interpolation as a classical deterministic resampling baseline [60], the Efficient Sub-Pixel Convolutional Network (ESPCN) as a lightweight CNN-based SR model [61], and the Adaptive Multi-scale Feature Fusion Network (AMFFN) as a remote-sensing-oriented SR model [62]. Recent remote sensing SR studies have further advanced transformer-based and diffusion-based models, such as TTST and EDiffSR [65,66]. These methods provide important references for advanced SR architecture design. In this study, however, SR was used as an optional module for evaluating NIIRS and object detection response, and the experiments focused on representative baseline methods. All experiments were conducted on a workstation equipped with an NVIDIA RTX 4060 GPU (8 GB VRAM).

To support model selection for the NEONSAT-1 case study, the compared SR methods were first evaluated on the UC Merced Land Use dataset [38]. Runway and intersection classes were selected, and low-resolution/high-resolution pairs were generated by down-sampling reference patches and reconstructing them at a ×2 scale. Reconstruction fidelity was evaluated using PSNR and SSIM, together with model size, FLOPs, and runtime. The comparative benchmark results are reported in Section 4.3.1.

SR was then applied to NEONSAT-1 pan-sharpened (PS) image chips selected for the analysis of NIIRS and object detection performance. For each chip, SR outputs were generated at scale factors of ×2, ×3, and ×4. These outputs were evaluated using the same NIIRS procedure described in Section 3.3. For object detection, the detector was applied to the SR outputs, but both the enlarged visual comparison and the quantitative detection-rate evaluation under SR scaling were focused on four target classes: airplane, vehicle, storage tank, and ground-track field. This design allowed the NIIRS response to be compared with the detection response of the selected targets across SR scales.

4. Results

4.1. Product-Level Comparison of NIIRS and Object Detection on Native NEONSAT Imagery

To establish the baseline relationship between native NEONSAT image quality and downstream object detection performance, a representative Boston scene, acquired on 1 June 2024, was analyzed using the PS, PAN, and MS products. This scene was selected as the primary case for product-level comparison because it contains multiple target classes relevant to the present framework, including vehicles, airplanes, storage tanks, and ground-track fields. NIIRS and object detection performance were evaluated using the procedures described in Section 3.3 and Section 3.4.

Figure 8, Figure 9 and Figure 10 show representative detection examples, detection rates, and NIIRS statistics for the PS, PAN, and MS products, and

Table 4. Detection rates and NIIRS statistics for the PS product.

	Detection Rate (%)	NIIRS Range	NIIRS Average
Vehicle	63.00	3.73–4.72	4.22
Airplane	57.00	4.54–4.93	4.82
Storage Tank	83.00	4.00–4.58	4.33
Ground-Track Field	69.00	3.93–4.44	4.26
Total	71.00	3.73–4.93	4.41

,

Table 5. Detection rates and NIIRS statistics for the PAN product.

	Detection Rate (%)	NIIRS Range	NIIRS Average
Vehicle	58.00	3.60–4.61	4.13
Airplane	36.00	3.95–4.26	4.12
Storage Tank	90.00	3.93–4.34	4.13
Ground-Track Field	12.00	3.71–4.48	4.10
Total	65.00	3.60–4.61	4.12

and

Table 6. Detection rates and NIIRS statistics for the MS product.

	Detection Rate (%)	NIIRS Range	NIIRS Average
Vehicle	N/A	N/A	N/A
Airplane	2.00	2.54–3.33	2.97
Storage Tank	77.00	2.59–3.09	2.89
Ground-Track Field	69.00	2.72–3.15	2.97
Total	57.00	2.54–3.33	2.94

summarize the corresponding quantitative results. At the overall scene level, the PS product showed the highest performance, with an overall detection rate of 71% and a mean NIIRS of 4.41. The PAN product followed with an overall detection rate of 65% and a mean NIIRS of 4.12, whereas the MS product yielded the lowest overall detection rate, 57%, together with substantially lower NIIRS values, with a mean of 2.94. Overall, the ordering of detection performance across the three products was broadly consistent with the relative interpretability indicated by the NIIRS estimates.

The product response differed depending on target type. For small targets, the degradation from PS and PAN to MS was pronounced. Vehicles were detected in the PS and PAN products but not in the MS product, and airplane detection also decreased markedly from PS to PAN and then to MS. By contrast, larger and structurally distinct targets remained detectable across all three products. Storage tanks maintained relatively high detection rates in all products, and ground-track fields also remained detectable, except in the PAN case, where the result appears to have been affected by the limited number of ground-truth instances in this scene. These results indicate that the relationship between product-level NIIRS and object detection was not uniform across all targets but depended on object size and structural distinctiveness.

In the Boston case, products with mean NIIRS values around 4 supported moderate detection of small to medium-sized targets. By contrast, the MS product showed lower NIIRS values and retained utility mainly for larger structures. This native product comparison indicates that object detection response in NEONSAT-1 imagery is target dependent even before controlled variations in MTF, GSD, or super-resolution are introduced.

4.2. Image Quality Analysis as a Function of System MTF

4.2.1. Effects of System MTF on Image Quality and Detection Performance

Using the optical MTF cases and image simulation procedure described in Section 3.1 and Section 3.2, the controlled sharpness experiment evaluated how changes in system MTF affected image interpretability and object detection performance. The selected optical cases spanned Nyquist MTF values of 5, 11, 15, 18, 20, and 25%, with 18% representing the reference NEONSAT optical condition. After propagation through the simulation chain, the corresponding system MTF cases were used for the edge target and Boston scene experiments in this section.

Figure 11 shows the resulting image response for two edge targets: an ideal digital target and the Salon-de-Provence NEONSAT edge target described in Section 2.2. In both cases, increasing the system MTF led to a sharper edge transition and higher local contrast. The visual differences were more apparent in the digital target, whereas the real target showed a more moderate change because scene heterogeneity and residual blur were superimposed on the edge response. Nevertheless, the edge-based analysis responded systematically to the controlled sharpness variation.

As summarized in

Table 7. Quantitative image-quality metrics for the real NEONSAT edge target under the controlled $MT{F}_{system}$ cases.

Optic		System	Image Quality			Note
Conic Coefficient ($\mathit{k}$)	MTF (%)	MTF (%)	RER	SNR	NIIRS
−1.9450	5	2.20	0.32	25.70	3.87
−1.9430	11	4.69	0.34	27.12	3.96
−1.9416	15	6.60	0.36	27.86	4.02
−1.9406	18	7.96	0.37	28.51	4.06	NEONSAT
−1.9400	20	8.74	0.38	28.85	4.08
−1.9380	25	10.95	0.39	29.21	4.11

, when the system MTF increased from 2.20% to 10.95%, RER increased from 0.32 to 0.39 and NIIRS increased from 3.87 to 4.11, whereas SNR changed only slightly, from 25.70 to 29.21. Figure 12 shows the same tendency in compact form, with a smoother monotonic response for the digital target and greater scatter for the observed edge target. These results indicate that controlled changes in system sharpness were consistently reflected in RER and NIIRS, even when the visual differences were modest [24,25,32,55,56].

The same set of system MTF cases was then applied to the Boston PS scene used in Section 4.1 to test whether the controlled sharpness variation was also reflected in downstream task performance. As shown in Figure 13 and summarized in

Table 8. Quantitative image-quality metrics and detection rates for the Boston PS scene under the controlled $MT{F}_{system}$ cases.

Case	$\mathit{M}\mathit{T}{\mathit{F}}_{\mathit{s}\mathit{y}\mathit{s}\mathit{t}\mathit{e}\mathit{m}}$ (%)	GSD (m)	RER	SNR	NIIRS	Detection Rate (%)	Note
(a)	2	0.939	0.305	21.56	3.58	58.2
(b)	5	0.939	0.321	21.52	3.63	59.5
(c)	7	0.939	0.327	21.58	3.65	62.1
(d)	8	0.939	0.332	21.68	3.67	62.1	NEONSAT
(e)	9	0.939	0.338	21.60	3.69	64.7
(f)	11	0.939	0.344	21.46	3.71	66.0

, increasing the system MTF from 2% to 11% increased RER from 0.305 to 0.344 and NIIRS from 3.58 to 3.71, while the detection rate increased from 58.2% to 66.0%. By contrast, SNR remained within a relatively narrow range and did not show a consistent trend across the tested MTF conditions.

Figure 14 shows this relationship in compact form, indicating that the increase in NIIRS was accompanied by an increase in detection rate under this controlled setting. Taken together, these results indicate that, within the tested MTF range, the dominant effect of the controlled optical variation was expressed through edge sharpness and image interpretability rather than through large changes in noise level.

4.2.2. Effects of Altitude on Image Quality

To examine how changes in acquisition geometry affect spatial image quality under a representative operational scenario, altitude-dependent simulations were performed from 500 km, corresponding to the nominal NEONSAT orbit, down to 300 km in 50 km steps. The baseline optical design and detector-related settings were retained throughout the experiment. Altitude-related changes were introduced primarily through simulated acquisition GSD, together with the empirical atmospheric and SNR parameterizations described in Section 3.2. The results should therefore be interpreted as relative trends at the product level under the adopted simulation assumptions. Three representative scenes, Boston, Paris, and Abu Dhabi, were analyzed using both PS and PAN products to compare product-level responses under different scene characteristics.

Figure 15 presents the Boston PS example across the tested altitude range. As the simulated altitude decreased from 500 km to 300 km, GSD decreased from approximately 0.80 m to 0.49 m, and the image chip showed finer spatial detail within a smaller ground extent. The corresponding object detection results exhibited the same directional change. For the main targets in this scene, the detection rate increased from 76% at 500 km to 85% at 300 km, together with an increase in NIIRS.

Figure 16 and

Table 9. Summary of altitude-dependent GSD, NIIRS, and detection rate results for the Boston, Paris, and Abu Dhabi scenes using PS and PAN products.

Study Area	Product Type	Altitude (km)	GSD (m)	Detection Rate (%)	NIIRS Range	NIIRS Average
Boston	PS	300–500	0.49–0.80	63–94	4.44–4.97	4.72
Boston	PAN			63–88	4.15–4.59	4.36
Paris	PS		0.52–0.84	19–50	4.27–4.88	4.54
Paris	PAN			18–50	3.93–4.52	4.23
Abu Dhabi	PS		0.52–0.84	75–87	4.39–4.98	4.66
Abu Dhabi	PAN			75–87	4.01–4.58	4.31

Note: The MS product was not included in the altitude-dependent detection analysis because the product-level comparison showed substantially lower NIIRS and limited detectability for small targets.

summarize the altitude-dependent results for the Boston, Paris, and Abu Dhabi scenes using both PS and PAN products. In both products, lower simulated altitude generally produced finer GSD and higher NIIRS. The PS products showed higher mean NIIRS than the corresponding PAN products in all three scenes. The mean NIIRS values were 4.72 for Boston PS and 4.36 for Boston PAN, 4.54 for Paris PS and 4.23 for Paris PAN, and 4.66 for Abu Dhabi PS and 4.31 for Abu Dhabi PAN.

The detection response also differed by scene and product. In Boston, the detection rate ranged from 63% to 94% for PS and from 63% to 88% for PAN. In Paris, the detection rate ranged from 19% to 50% for PS and from 18% to 50% for PAN. In Abu Dhabi, the detection rate ranged from 75% to 87% for both PS and PAN. Although the Paris scene showed lower absolute detection rates than the Boston and Abu Dhabi scenes, NIIRS still increased consistently as the simulated altitude decreased. This indicates that the altitude-dependent NIIRS trend was preserved, while the absolute detection response remained dependent on scene content and product type.

The supplementary detection examples in Figure A1 and Figure A2 provide visual support for these product-level trends. Figure A1 shows altitude-dependent detection examples for the Paris and Abu Dhabi PS scenes, whereas Figure A2 shows the corresponding PAN examples for Boston, Paris, and Abu Dhabi. Together, these results show that lower simulated altitude generally improved image interpretability, but the practical detection response was not uniform across scenes or products.

4.3. Super-Resolution–Aided Image Quality Enhancement

4.3.1. Controlled Benchmark on UC Merced Land Use Imagery

Before applying super-resolution (SR) to NEONSAT imagery, the compared SR methods were evaluated on the UC Merced Land Use dataset as a controlled benchmark. Following the procedure described in Section 3.5, runway and intersection patches were degraded to generate low-resolution inputs and then reconstructed at a ×2 scale using bicubic interpolation, ESPCN, and AMFFN [60,61,62]. Reconstruction fidelity and computational cost were assessed to compare the basic performance of the candidate methods before their application to NEONSAT data.

Figure 17 shows representative results for the runway and intersection examples. The degraded input and bicubic interpolation produced relatively smooth boundaries, whereas the CNN-based methods restored sharper structures in the zoomed regions. Among the tested methods, AMFFN produced the clearest boundary reconstruction in the visual comparison, while ESPCN showed an intermediate response between bicubic interpolation and AMFFN.

Table 10. Quantitative reconstruction accuracy and computational cost of the candidate SR methods on the UC Merced ×2 benchmark.

	Degraded Input	Bicubic	ESPCN	AMFFN
PSNR	30.70	32.42	33.52	35.44
SSIM	0.8610	0.8986	0.9175	0.9378
Parameters *	N/A	N/A	63,724	298,112
FLOPs (G) *	N/A	N/A	2.10	8.76
Time (ms) *	N/A	0.77	4.59	36.54

* N/A indicates not applicable. Model parameters and FLOPs were not applicable to the degraded input or bicubic interpolation, and runtime was not measured for the degraded input.

summarizes the quantitative results. Overall, AMFFN achieved the highest reconstruction fidelity, whereas ESPCN provided a lighter and faster alternative with lower computational cost. These results support the selection of AMFFN as the primary SR model for the subsequent NEONSAT case study, where the main interest is downstream interpretability rather than computational efficiency alone.

4.3.2. NEONSAT Case Study: NIIRS and Detection Response to SR Scaling

Based on the controlled benchmark in Section 4.3.1, AMFFN was selected as the primary SR model and applied to representative NEONSAT-1 PS image chips. For each chip, SR outputs were generated at scale factors of ×2, ×3, and ×4 and evaluated together with the native PS images using the same NIIRS procedure described in Section 3.3.

For the SR scaling experiment, qualitative visual comparison and quantitative detection rate evaluation were performed for four target classes: airplane, vehicle, storage tank, and ground-track field. Detection was evaluated using the same procedure described in Section 3.4. The image-scale values reported in

Table A1. Effects of SR scaling on image scale, NIIRS, and detection rate for airplane, vehicle, storage tank, and ground track field target classes in NEONSAT-1 PS image chips.

Object	Scale	Image Scale (m)	NIIRS	Detection Rate (%)
Airplane	Native	0.939	4.29	71
Airplane	$\times$2	0.470	4.35	82
Airplane	$\times$3	0.313	4.18	77
Airplane	$\times$4	0.235	4.07	77
Vehicle	Native	0.939	4.19	46
Vehicle	$\times$2	0.470	4.28	55
Vehicle	$\times$3	0.313	4.04	64
Vehicle	$\times$4	0.235	4.12	77
Storage Tank	Native	0.939	4.33	89
Storage Tank	$\times$2	0.470	4.39	93
Storage Tank	$\times$3	0.313	4.24	96
Storage Tank	$\times$4	0.235	4.02	93
Ground Track Field	Native	0.939	4.26	65
Ground Track Field	$\times$2	0.470	4.32	75
Ground Track Field	$\times$3	0.313	4.11	75
Ground Track Field	$\times$4	0.235	4.10	75

Note: Image scale was calculated from the native GSD of 0.939 m and the SR scale factor. NIIRS was calculated using the native GSD of 0.939 m for all SR scales.

were calculated from the native GSD of 0.939 m and the SR scale factor. However, NIIRS was calculated using the native GSD of 0.939 m for all SR scales, so the SR outputs were not treated as images acquired with finer GSD.

Figure 18 presents representative SR outputs and object-detection examples for the four target classes. The examples show that SR processing changed apparent sharpness, local texture, and target boundaries depending on target class and scale factor. In several cases, SR enhanced object boundaries and made target structures more distinct. However, the visual response was not uniform across all targets, indicating that the effect of SR should be evaluated together with object detection results rather than by visual sharpness alone.

Figure 19 summarizes the quantitative NIIRS and detection rate responses to SR scaling, and the corresponding numerical values are provided in Table A1. Across the four target classes, NIIRS did not increase monotonically with SR scale. Instead, the highest NIIRS was obtained at ×2 for all evaluated classes, followed by lower or variable values at ×3 and ×4. This indicates that stronger SR scaling did not necessarily produce higher NIIRS when the native GSD was retained in the NIIRS calculation.

The detection rate response showed a different pattern from NIIRS. Airplane detection increased at ×2 and then slightly decreased at higher scales, whereas vehicle detection continued to increase up to ×4. Storage tank detection reached the highest value at ×3, while ground-track field detection increased at ×2 and remained unchanged at higher scales. These results show that SR scaling affected NIIRS and detection rate differently depending on target class.

Overall, the SR case study indicates that moderate SR can improve image interpretability and object detection response, but the effect is not uniform across target classes. The ×2 SR outputs provided the most consistent improvement in NIIRS, whereas detection rate did not follow a single monotonic pattern across all classes. Therefore, higher SR scale should not be interpreted as directly producing proportional improvement in object detection. Because the number of available samples differed among target classes, these results should be interpreted as a NEONSAT-1 case study rather than as a general benchmark comparison.

5. Discussion

5.1. Sensitivity of Spatial Image Quality to Controlled Sharpness and GSD Variations

The results in Section 4.2.1 and Section 4.2.2 show that, within the tested range, the NIIRS trend was mainly associated with changes in RER and GSD. In the controlled MTF experiment, increases in system MTF were accompanied by systematic increases in RER and NIIRS, while the measured SNR changed only within a narrow range. A similar tendency appeared in the altitude experiment, where lower altitude produced smaller GSD values and higher NIIRS across the tested PS and PAN products.

This result should be interpreted within the scope of the present experiments. The SNR used in this study was derived from local edge contrast and nearby noise statistics in the processed image product rather than from a mission-level radiometric noise model. In addition, simplified relative SNR scaling in the simulation was used to assign consistent relative SNR conditions across the altitude cases. The present results therefore do not imply that SNR is unimportant. They indicate that, under the tested conditions, the local SNR metric varied less than RER and GSD and did not explain the main NIIRS trend.

This interpretation is consistent with the structure of GIQE, in which noise-related terms are included together with GSD and RER. Previous analysis of the GIQE showed that noise can affect image quality by reducing effective image resolution and by masking object features, while the relative contribution of the noise term depends on the imaging and processing conditions [25]. Because detector noise was not independently perturbed over a wide range in the present experiments, the sensitivity of NEONSAT image quality to noise should be examined further using controlled noise simulations and additional constellation imagery.

The comparison between the ideal digital target and the real NEONSAT-1 edge target also clarifies the role of quantitative image-quality metrics. In Figure 11, visual differences among MTF cases are visible but limited, especially for the observed target. By contrast, the edge-based analysis captured a consistent increase in RER and NIIRS as system MTF increased. This suggests that visual inspection is useful for qualitative confirmation, but quantitative metrics are more informative when the difference is small and difficult to distinguish visually.

5.2. NIIRS as a Practical but Target-Dependent Indicator for Detection

The baseline comparison in Section 4.1 showed that the relationship between NIIRS and object detection was not uniform across all target classes. Detection of vehicles and airplanes degraded more strongly than that of storage tanks and ground-track fields as the product changed from PS and PAN to MS. This result demonstrates that the same interpretability framework can be reflected differently across NEONSAT-1 products and target classes.

The altitude experiment further supports this interpretation. Lower altitude generally increased NIIRS for both PS and PAN products, but the detection response differed by scene and product type. The Paris scene showed lower absolute detection rates than the Boston and Abu Dhabi scenes, although NIIRS still increased with decreasing altitude. This indicates that the altitude-dependent NIIRS trend was preserved, while the absolute detection response remained affected by scene content, target characteristics, and product type.

In this sense, NIIRS is useful as a practical indicator of detection potential, but it should not be treated as a single universal threshold for all automated detection tasks. Because NIIRS is defined as a task-based measure of image interpretability, its practical meaning can vary with target class. The same product-level NIIRS range may not yield the same detection response for small objects, structurally distinct objects, or targets embedded in different background conditions. Therefore, image quality requirements for automated exploitation should be linked to target class and task objective rather than expressed only as one common NIIRS threshold.

5.3. Target Class-Dependent Response to Super Resolution

The super-resolution results in Section 4.3.2 show that SR should be interpreted as a target class-dependent image enhancement process rather than as a direct substitute for improved acquisition resolution. In this study, the image-scale values reported in Table A1 were calculated from the native GSD of 0.939 m and the SR scale factor, but NIIRS was calculated using the native GSD of 0.939 m for all SR scales. Therefore, the SR outputs were not treated as images acquired with finer GSD. This setting avoids directly increasing NIIRS by substituting a smaller SR-based image scale into the GIQE calculation.

SR estimates high-frequency spatial details from the input image and learned image priors; therefore, it does not increase the physical information originally captured by the sensor [60,61,62,63,64,65,66,67]. Therefore, its response depends on the training data, input image quality, reconstruction model, and downstream detector. When the input image contains clear target structures and is consistent with the resolution range and appearance of the training data, moderate SR can enhance object boundaries and improve detection response. However, when the input image is blurred, noisy, or structurally ambiguous, stronger SR scaling may amplify unstable textures or artificial boundaries instead of providing physically meaningful target information.

This interpretation is supported by the SR results in this study. NIIRS did not increase monotonically with SR scale. Across the four evaluated target classes, the highest NIIRS was obtained at ×2, followed by lower or variable values at ×3 and ×4. This result suggests that moderate SR improved the edge response used in NIIRS estimation, whereas stronger scaling did not necessarily improve estimated interpretability when native GSD was retained in the calculation.

The visual examples in Figure 18 also show that stronger SR scaling can change detector behavior. In the airplane examples, ×3 and ×4 SR sometimes enhanced internal texture or fragmented boundary patterns, resulting in detection inside the target rather than more stable object detection. In the vehicle example, the ×4 SR output increased the detection of vehicle-like structures in the surrounding background. These cases indicate that aggressive SR can increase detector responses by changing local shape and texture cues, but such responses do not always indicate more reliable object recognition.

The quantitative detection results showed the same target class dependence. Airplane detection improved at ×2 and then slightly decreased at higher scales, whereas vehicle detection continued to increase up to ×4. Storage tank detection reached the highest value at ×3, while ground-track field detection improved at ×2 and remained unchanged at higher scales. These results indicate that the practical benefit of SR depends on the target size, structural distinctiveness, background clutter, input image quality, and detector sensitivity. Therefore, higher SR scale should not be interpreted as producing a proportional improvement in object detection. SR should be evaluated together with detector stability, false positive behavior, and localization quality rather than by visual sharpness or image scale alone.

5.4. Scope, Limitations, and Future Extensions

The present study is confined to a limited but clearly defined scope. It focuses on spatial image quality, primarily in terms of image interpretability, and examines how optical sharpness, GSD, and super-resolution affect NIIRS and object detection performance. The study was therefore not intended to cover all dimensions of image quality but to provide an initial framework for linking image formation, interpretability assessment, and downstream task response in a micro-satellite constellation context.

First, the object-detection analysis is tied to the YOLOv8n-based evaluation setting used in this study. YOLOv8n was selected as a stable and computationally efficient detector, and the same trained weights and inference settings were applied across all product, MTF, altitude, and SR cases. This design allowed changes in detection response to be compared under fixed detector conditions. However, newer detector architectures may improve absolute detection accuracy, inference speed, or computational efficiency. Therefore, the reported detection rates should be interpreted as the response of YOLOv8n under the present experimental setting, not as a general performance limit of all object detection models. Future work should compare multiple detector architectures and model scales using additional NEONSAT scenes and future constellation imagery.

Second, the simulation results should be interpreted under the controlled assumptions used in this study. The simulation module was designed to examine how selected imaging parameters affect NIIRS and downstream task response. Accordingly, system MTF and altitude-dependent GSD were treated as the main variables, whereas atmospheric, noise, and detector-related conditions were fixed or simplified to maintain a controlled comparison. The empirical attenuation and path radiance terms were used only for relative radiometric perturbation, and full band-dependent atmospheric correction or complete radiative transfer simulation was not performed. Detector noise was also not decomposed into separate physical sources or independently varied over a wide range. Future work should therefore include controlled atmospheric, detector noise, and motion-related perturbation experiments to quantify their effects on NIIRS and object detection performance.

Third, the SR analysis should be interpreted as a NEONSAT-1 case study based on representative PS image chips. Although the SR experiment included four target classes, the number of available samples differed among classes. In addition, SR-induced artifacts, boundary changes, and false-positive behavior were examined qualitatively but not fully quantified using standard detection metrics. Future SR analysis should evaluate not only NIIRS and detection rate but also precision, recall, F1 score, mAP, false positive rate, and localization accuracy. More recent remote-sensing SR models, including transformer- based models such as TTST and diffusion-based models such as EDiffSR, should also be compared to determine whether advanced SR methods improve practical image usability and downstream task performance rather than only changing apparent sharpness or image scale [65,66].

Fourth, the present framework should be extended beyond spatial image quality for operational constellation assessment. This study did not perform full radiometric calibration, spectral response characterization, or band-specific radiative transfer modeling. Therefore, the current results should be interpreted primarily in terms of spatial image quality and image interpretability. For future NEONSAT constellation operations, the framework should be linked with absolute radiometric calibration, atmospheric correction, BRDF normalization, and surface reflectance harmonization. These steps are important because images with similar spatial interpretability can still differ in radiance or surface reflectance due to sensor calibration, spectral response differences, illumination and viewing geometry, atmospheric conditions, and surface anisotropy.

Finally, complementary image-quality descriptors beyond GIQE and NIIRS should be examined in future work. Full-reference metrics such as SSIM, FSIM, and GMSD can quantify structural or perceptual image degradation when reliable reference images are available [68,69,70]. Human visual system (HVS)-related approaches, such as the TTP criterion, are also relevant because they connect target task performance, visual perception, and imaging chain degradation [27]. Recent no-reference methods, such as ARNIQA, provide another direction for estimating perceptual image quality without requiring a reference image [71]. Once additional NEONSAT constellation imagery, expert interpretation data, and task annotations become available, these descriptors could be evaluated together with GSD, RER, SNR, NIIRS, and object detection metrics to build a more complete image quality assessment framework.

6. Conclusions

This study presented an integrated framework for spatial image quality assessment in a micro-satellite constellation context by linking optical design analysis, image simulation, NIIRS-based interpretability assessment, and object detection performance using NEONSAT-1 imagery. The results showed that controlled changes in system MTF and simulated GSD were reflected in RER and NIIRS. Native product comparison indicated that PS imagery provided the strongest overall performance, while PAN and MS responses varied depending on target class and object structure.

The altitude-dependent experiment showed that lower simulated altitude generally increased NIIRS for both PS and PAN products. However, the corresponding detection response differed by scene and product type, indicating that higher image interpretability does not necessarily produce a proportional object detection improvement. The super-resolution case study further showed that SR response was target class dependent. The ×2 SR outputs provided the most consistent NIIRS improvement, whereas detection rates at ×3 and ×4 varied by target class and were affected by changes in local texture, boundary representation, and detector response.

These findings support the use of NIIRS as a practical indicator of detection potential, while also demonstrating that spatial image quality should be interpreted jointly with downstream task performance. The proposed framework provides a baseline for future constellation-level assessment and can be extended by incorporating more detailed detector noise modeling, atmospheric effects, radiometric calibration, spectral consistency, BRDF normalization, and additional task-based image quality descriptors.